LLVM  6.0.0svn
ScalarEvolutionExpander.cpp
Go to the documentation of this file.
1 //===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis ------------===//
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 contains the implementation of the scalar evolution expander,
11 // which is used to generate the code corresponding to a given scalar evolution
12 // expression.
13 //
14 //===----------------------------------------------------------------------===//
15 
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/Analysis/LoopInfo.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/Dominators.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/LLVMContext.h"
26 #include "llvm/IR/Module.h"
27 #include "llvm/IR/PatternMatch.h"
28 #include "llvm/Support/Debug.h"
30 
31 using namespace llvm;
32 using namespace PatternMatch;
33 
34 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
35 /// reusing an existing cast if a suitable one exists, moving an existing
36 /// cast if a suitable one exists but isn't in the right place, or
37 /// creating a new one.
38 Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty,
41  // This function must be called with the builder having a valid insertion
42  // point. It doesn't need to be the actual IP where the uses of the returned
43  // cast will be added, but it must dominate such IP.
44  // We use this precondition to produce a cast that will dominate all its
45  // uses. In particular, this is crucial for the case where the builder's
46  // insertion point *is* the point where we were asked to put the cast.
47  // Since we don't know the builder's insertion point is actually
48  // where the uses will be added (only that it dominates it), we are
49  // not allowed to move it.
50  BasicBlock::iterator BIP = Builder.GetInsertPoint();
51 
52  Instruction *Ret = nullptr;
53 
54  // Check to see if there is already a cast!
55  for (User *U : V->users())
56  if (U->getType() == Ty)
57  if (CastInst *CI = dyn_cast<CastInst>(U))
58  if (CI->getOpcode() == Op) {
59  // If the cast isn't where we want it, create a new cast at IP.
60  // Likewise, do not reuse a cast at BIP because it must dominate
61  // instructions that might be inserted before BIP.
62  if (BasicBlock::iterator(CI) != IP || BIP == IP) {
63  // Create a new cast, and leave the old cast in place in case
64  // it is being used as an insert point. Clear its operand
65  // so that it doesn't hold anything live.
66  Ret = CastInst::Create(Op, V, Ty, "", &*IP);
67  Ret->takeName(CI);
68  CI->replaceAllUsesWith(Ret);
69  CI->setOperand(0, UndefValue::get(V->getType()));
70  break;
71  }
72  Ret = CI;
73  break;
74  }
75 
76  // Create a new cast.
77  if (!Ret)
78  Ret = CastInst::Create(Op, V, Ty, V->getName(), &*IP);
79 
80  // We assert at the end of the function since IP might point to an
81  // instruction with different dominance properties than a cast
82  // (an invoke for example) and not dominate BIP (but the cast does).
83  assert(SE.DT.dominates(Ret, &*BIP));
84 
85  rememberInstruction(Ret);
86  return Ret;
87 }
88 
90  BasicBlock *MustDominate) {
92  if (auto *II = dyn_cast<InvokeInst>(I))
93  IP = II->getNormalDest()->begin();
94 
95  while (isa<PHINode>(IP))
96  ++IP;
97 
98  if (isa<FuncletPadInst>(IP) || isa<LandingPadInst>(IP)) {
99  ++IP;
100  } else if (isa<CatchSwitchInst>(IP)) {
101  IP = MustDominate->getFirstInsertionPt();
102  } else {
103  assert(!IP->isEHPad() && "unexpected eh pad!");
104  }
105 
106  return IP;
107 }
108 
109 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
110 /// which must be possible with a noop cast, doing what we can to share
111 /// the casts.
112 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) {
113  Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
114  assert((Op == Instruction::BitCast ||
115  Op == Instruction::PtrToInt ||
116  Op == Instruction::IntToPtr) &&
117  "InsertNoopCastOfTo cannot perform non-noop casts!");
118  assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
119  "InsertNoopCastOfTo cannot change sizes!");
120 
121  // Short-circuit unnecessary bitcasts.
122  if (Op == Instruction::BitCast) {
123  if (V->getType() == Ty)
124  return V;
125  if (CastInst *CI = dyn_cast<CastInst>(V)) {
126  if (CI->getOperand(0)->getType() == Ty)
127  return CI->getOperand(0);
128  }
129  }
130  // Short-circuit unnecessary inttoptr<->ptrtoint casts.
131  if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
132  SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
133  if (CastInst *CI = dyn_cast<CastInst>(V))
134  if ((CI->getOpcode() == Instruction::PtrToInt ||
135  CI->getOpcode() == Instruction::IntToPtr) &&
136  SE.getTypeSizeInBits(CI->getType()) ==
137  SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
138  return CI->getOperand(0);
139  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
140  if ((CE->getOpcode() == Instruction::PtrToInt ||
141  CE->getOpcode() == Instruction::IntToPtr) &&
142  SE.getTypeSizeInBits(CE->getType()) ==
143  SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
144  return CE->getOperand(0);
145  }
146 
147  // Fold a cast of a constant.
148  if (Constant *C = dyn_cast<Constant>(V))
149  return ConstantExpr::getCast(Op, C, Ty);
150 
151  // Cast the argument at the beginning of the entry block, after
152  // any bitcasts of other arguments.
153  if (Argument *A = dyn_cast<Argument>(V)) {
154  BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
155  while ((isa<BitCastInst>(IP) &&
156  isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
157  cast<BitCastInst>(IP)->getOperand(0) != A) ||
158  isa<DbgInfoIntrinsic>(IP))
159  ++IP;
160  return ReuseOrCreateCast(A, Ty, Op, IP);
161  }
162 
163  // Cast the instruction immediately after the instruction.
164  Instruction *I = cast<Instruction>(V);
165  BasicBlock::iterator IP = findInsertPointAfter(I, Builder.GetInsertBlock());
166  return ReuseOrCreateCast(I, Ty, Op, IP);
167 }
168 
169 /// InsertBinop - Insert the specified binary operator, doing a small amount
170 /// of work to avoid inserting an obviously redundant operation.
171 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
172  Value *LHS, Value *RHS) {
173  // Fold a binop with constant operands.
174  if (Constant *CLHS = dyn_cast<Constant>(LHS))
175  if (Constant *CRHS = dyn_cast<Constant>(RHS))
176  return ConstantExpr::get(Opcode, CLHS, CRHS);
177 
178  // Do a quick scan to see if we have this binop nearby. If so, reuse it.
179  unsigned ScanLimit = 6;
180  BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
181  // Scanning starts from the last instruction before the insertion point.
182  BasicBlock::iterator IP = Builder.GetInsertPoint();
183  if (IP != BlockBegin) {
184  --IP;
185  for (; ScanLimit; --IP, --ScanLimit) {
186  // Don't count dbg.value against the ScanLimit, to avoid perturbing the
187  // generated code.
188  if (isa<DbgInfoIntrinsic>(IP))
189  ScanLimit++;
190  if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
191  IP->getOperand(1) == RHS)
192  return &*IP;
193  if (IP == BlockBegin) break;
194  }
195  }
196 
197  // Save the original insertion point so we can restore it when we're done.
198  DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc();
199  SCEVInsertPointGuard Guard(Builder, this);
200 
201  // Move the insertion point out of as many loops as we can.
202  while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) {
203  if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
204  BasicBlock *Preheader = L->getLoopPreheader();
205  if (!Preheader) break;
206 
207  // Ok, move up a level.
208  Builder.SetInsertPoint(Preheader->getTerminator());
209  }
210 
211  // If we haven't found this binop, insert it.
212  Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS));
213  BO->setDebugLoc(Loc);
214  rememberInstruction(BO);
215 
216  return BO;
217 }
218 
219 /// FactorOutConstant - Test if S is divisible by Factor, using signed
220 /// division. If so, update S with Factor divided out and return true.
221 /// S need not be evenly divisible if a reasonable remainder can be
222 /// computed.
223 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
224 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
225 /// check to see if the divide was folded.
226 static bool FactorOutConstant(const SCEV *&S, const SCEV *&Remainder,
227  const SCEV *Factor, ScalarEvolution &SE,
228  const DataLayout &DL) {
229  // Everything is divisible by one.
230  if (Factor->isOne())
231  return true;
232 
233  // x/x == 1.
234  if (S == Factor) {
235  S = SE.getConstant(S->getType(), 1);
236  return true;
237  }
238 
239  // For a Constant, check for a multiple of the given factor.
240  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
241  // 0/x == 0.
242  if (C->isZero())
243  return true;
244  // Check for divisibility.
245  if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
246  ConstantInt *CI =
247  ConstantInt::get(SE.getContext(), C->getAPInt().sdiv(FC->getAPInt()));
248  // If the quotient is zero and the remainder is non-zero, reject
249  // the value at this scale. It will be considered for subsequent
250  // smaller scales.
251  if (!CI->isZero()) {
252  const SCEV *Div = SE.getConstant(CI);
253  S = Div;
254  Remainder = SE.getAddExpr(
255  Remainder, SE.getConstant(C->getAPInt().srem(FC->getAPInt())));
256  return true;
257  }
258  }
259  }
260 
261  // In a Mul, check if there is a constant operand which is a multiple
262  // of the given factor.
263  if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
264  // Size is known, check if there is a constant operand which is a multiple
265  // of the given factor. If so, we can factor it.
266  const SCEVConstant *FC = cast<SCEVConstant>(Factor);
267  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
268  if (!C->getAPInt().srem(FC->getAPInt())) {
269  SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
270  NewMulOps[0] = SE.getConstant(C->getAPInt().sdiv(FC->getAPInt()));
271  S = SE.getMulExpr(NewMulOps);
272  return true;
273  }
274  }
275 
276  // In an AddRec, check if both start and step are divisible.
277  if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
278  const SCEV *Step = A->getStepRecurrence(SE);
279  const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
280  if (!FactorOutConstant(Step, StepRem, Factor, SE, DL))
281  return false;
282  if (!StepRem->isZero())
283  return false;
284  const SCEV *Start = A->getStart();
285  if (!FactorOutConstant(Start, Remainder, Factor, SE, DL))
286  return false;
287  S = SE.getAddRecExpr(Start, Step, A->getLoop(),
288  A->getNoWrapFlags(SCEV::FlagNW));
289  return true;
290  }
291 
292  return false;
293 }
294 
295 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
296 /// is the number of SCEVAddRecExprs present, which are kept at the end of
297 /// the list.
298 ///
300  Type *Ty,
301  ScalarEvolution &SE) {
302  unsigned NumAddRecs = 0;
303  for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
304  ++NumAddRecs;
305  // Group Ops into non-addrecs and addrecs.
306  SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
307  SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
308  // Let ScalarEvolution sort and simplify the non-addrecs list.
309  const SCEV *Sum = NoAddRecs.empty() ?
310  SE.getConstant(Ty, 0) :
311  SE.getAddExpr(NoAddRecs);
312  // If it returned an add, use the operands. Otherwise it simplified
313  // the sum into a single value, so just use that.
314  Ops.clear();
315  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
316  Ops.append(Add->op_begin(), Add->op_end());
317  else if (!Sum->isZero())
318  Ops.push_back(Sum);
319  // Then append the addrecs.
320  Ops.append(AddRecs.begin(), AddRecs.end());
321 }
322 
323 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
324 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
325 /// This helps expose more opportunities for folding parts of the expressions
326 /// into GEP indices.
327 ///
329  Type *Ty,
330  ScalarEvolution &SE) {
331  // Find the addrecs.
333  for (unsigned i = 0, e = Ops.size(); i != e; ++i)
334  while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
335  const SCEV *Start = A->getStart();
336  if (Start->isZero()) break;
337  const SCEV *Zero = SE.getConstant(Ty, 0);
338  AddRecs.push_back(SE.getAddRecExpr(Zero,
339  A->getStepRecurrence(SE),
340  A->getLoop(),
341  A->getNoWrapFlags(SCEV::FlagNW)));
342  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
343  Ops[i] = Zero;
344  Ops.append(Add->op_begin(), Add->op_end());
345  e += Add->getNumOperands();
346  } else {
347  Ops[i] = Start;
348  }
349  }
350  if (!AddRecs.empty()) {
351  // Add the addrecs onto the end of the list.
352  Ops.append(AddRecs.begin(), AddRecs.end());
353  // Resort the operand list, moving any constants to the front.
354  SimplifyAddOperands(Ops, Ty, SE);
355  }
356 }
357 
358 /// expandAddToGEP - Expand an addition expression with a pointer type into
359 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
360 /// BasicAliasAnalysis and other passes analyze the result. See the rules
361 /// for getelementptr vs. inttoptr in
362 /// http://llvm.org/docs/LangRef.html#pointeraliasing
363 /// for details.
364 ///
365 /// Design note: The correctness of using getelementptr here depends on
366 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
367 /// they may introduce pointer arithmetic which may not be safely converted
368 /// into getelementptr.
369 ///
370 /// Design note: It might seem desirable for this function to be more
371 /// loop-aware. If some of the indices are loop-invariant while others
372 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
373 /// loop-invariant portions of the overall computation outside the loop.
374 /// However, there are a few reasons this is not done here. Hoisting simple
375 /// arithmetic is a low-level optimization that often isn't very
376 /// important until late in the optimization process. In fact, passes
377 /// like InstructionCombining will combine GEPs, even if it means
378 /// pushing loop-invariant computation down into loops, so even if the
379 /// GEPs were split here, the work would quickly be undone. The
380 /// LoopStrengthReduction pass, which is usually run quite late (and
381 /// after the last InstructionCombining pass), takes care of hoisting
382 /// loop-invariant portions of expressions, after considering what
383 /// can be folded using target addressing modes.
384 ///
385 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
386  const SCEV *const *op_end,
387  PointerType *PTy,
388  Type *Ty,
389  Value *V) {
390  Type *OriginalElTy = PTy->getElementType();
391  Type *ElTy = OriginalElTy;
392  SmallVector<Value *, 4> GepIndices;
393  SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
394  bool AnyNonZeroIndices = false;
395 
396  // Split AddRecs up into parts as either of the parts may be usable
397  // without the other.
398  SplitAddRecs(Ops, Ty, SE);
399 
400  Type *IntPtrTy = DL.getIntPtrType(PTy);
401 
402  // Descend down the pointer's type and attempt to convert the other
403  // operands into GEP indices, at each level. The first index in a GEP
404  // indexes into the array implied by the pointer operand; the rest of
405  // the indices index into the element or field type selected by the
406  // preceding index.
407  for (;;) {
408  // If the scale size is not 0, attempt to factor out a scale for
409  // array indexing.
411  if (ElTy->isSized()) {
412  const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy);
413  if (!ElSize->isZero()) {
415  for (const SCEV *Op : Ops) {
416  const SCEV *Remainder = SE.getConstant(Ty, 0);
417  if (FactorOutConstant(Op, Remainder, ElSize, SE, DL)) {
418  // Op now has ElSize factored out.
419  ScaledOps.push_back(Op);
420  if (!Remainder->isZero())
421  NewOps.push_back(Remainder);
422  AnyNonZeroIndices = true;
423  } else {
424  // The operand was not divisible, so add it to the list of operands
425  // we'll scan next iteration.
426  NewOps.push_back(Op);
427  }
428  }
429  // If we made any changes, update Ops.
430  if (!ScaledOps.empty()) {
431  Ops = NewOps;
432  SimplifyAddOperands(Ops, Ty, SE);
433  }
434  }
435  }
436 
437  // Record the scaled array index for this level of the type. If
438  // we didn't find any operands that could be factored, tentatively
439  // assume that element zero was selected (since the zero offset
440  // would obviously be folded away).
441  Value *Scaled = ScaledOps.empty() ?
443  expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
444  GepIndices.push_back(Scaled);
445 
446  // Collect struct field index operands.
447  while (StructType *STy = dyn_cast<StructType>(ElTy)) {
448  bool FoundFieldNo = false;
449  // An empty struct has no fields.
450  if (STy->getNumElements() == 0) break;
451  // Field offsets are known. See if a constant offset falls within any of
452  // the struct fields.
453  if (Ops.empty())
454  break;
455  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
456  if (SE.getTypeSizeInBits(C->getType()) <= 64) {
457  const StructLayout &SL = *DL.getStructLayout(STy);
458  uint64_t FullOffset = C->getValue()->getZExtValue();
459  if (FullOffset < SL.getSizeInBytes()) {
460  unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
461  GepIndices.push_back(
463  ElTy = STy->getTypeAtIndex(ElIdx);
464  Ops[0] =
465  SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
466  AnyNonZeroIndices = true;
467  FoundFieldNo = true;
468  }
469  }
470  // If no struct field offsets were found, tentatively assume that
471  // field zero was selected (since the zero offset would obviously
472  // be folded away).
473  if (!FoundFieldNo) {
474  ElTy = STy->getTypeAtIndex(0u);
475  GepIndices.push_back(
477  }
478  }
479 
480  if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
481  ElTy = ATy->getElementType();
482  else
483  break;
484  }
485 
486  // If none of the operands were convertible to proper GEP indices, cast
487  // the base to i8* and do an ugly getelementptr with that. It's still
488  // better than ptrtoint+arithmetic+inttoptr at least.
489  if (!AnyNonZeroIndices) {
490  // Cast the base to i8*.
491  V = InsertNoopCastOfTo(V,
493 
494  assert(!isa<Instruction>(V) ||
495  SE.DT.dominates(cast<Instruction>(V), &*Builder.GetInsertPoint()));
496 
497  // Expand the operands for a plain byte offset.
498  Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
499 
500  // Fold a GEP with constant operands.
501  if (Constant *CLHS = dyn_cast<Constant>(V))
502  if (Constant *CRHS = dyn_cast<Constant>(Idx))
504  CLHS, CRHS);
505 
506  // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
507  unsigned ScanLimit = 6;
508  BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
509  // Scanning starts from the last instruction before the insertion point.
510  BasicBlock::iterator IP = Builder.GetInsertPoint();
511  if (IP != BlockBegin) {
512  --IP;
513  for (; ScanLimit; --IP, --ScanLimit) {
514  // Don't count dbg.value against the ScanLimit, to avoid perturbing the
515  // generated code.
516  if (isa<DbgInfoIntrinsic>(IP))
517  ScanLimit++;
518  if (IP->getOpcode() == Instruction::GetElementPtr &&
519  IP->getOperand(0) == V && IP->getOperand(1) == Idx)
520  return &*IP;
521  if (IP == BlockBegin) break;
522  }
523  }
524 
525  // Save the original insertion point so we can restore it when we're done.
526  SCEVInsertPointGuard Guard(Builder, this);
527 
528  // Move the insertion point out of as many loops as we can.
529  while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) {
530  if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
531  BasicBlock *Preheader = L->getLoopPreheader();
532  if (!Preheader) break;
533 
534  // Ok, move up a level.
535  Builder.SetInsertPoint(Preheader->getTerminator());
536  }
537 
538  // Emit a GEP.
539  Value *GEP = Builder.CreateGEP(Builder.getInt8Ty(), V, Idx, "uglygep");
540  rememberInstruction(GEP);
541 
542  return GEP;
543  }
544 
545  {
546  SCEVInsertPointGuard Guard(Builder, this);
547 
548  // Move the insertion point out of as many loops as we can.
549  while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) {
550  if (!L->isLoopInvariant(V)) break;
551 
552  bool AnyIndexNotLoopInvariant = any_of(
553  GepIndices, [L](Value *Op) { return !L->isLoopInvariant(Op); });
554 
555  if (AnyIndexNotLoopInvariant)
556  break;
557 
558  BasicBlock *Preheader = L->getLoopPreheader();
559  if (!Preheader) break;
560 
561  // Ok, move up a level.
562  Builder.SetInsertPoint(Preheader->getTerminator());
563  }
564 
565  // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
566  // because ScalarEvolution may have changed the address arithmetic to
567  // compute a value which is beyond the end of the allocated object.
568  Value *Casted = V;
569  if (V->getType() != PTy)
570  Casted = InsertNoopCastOfTo(Casted, PTy);
571  Value *GEP = Builder.CreateGEP(OriginalElTy, Casted, GepIndices, "scevgep");
572  Ops.push_back(SE.getUnknown(GEP));
573  rememberInstruction(GEP);
574  }
575 
576  return expand(SE.getAddExpr(Ops));
577 }
578 
579 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
580 /// SCEV expansion. If they are nested, this is the most nested. If they are
581 /// neighboring, pick the later.
582 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
583  DominatorTree &DT) {
584  if (!A) return B;
585  if (!B) return A;
586  if (A->contains(B)) return B;
587  if (B->contains(A)) return A;
588  if (DT.dominates(A->getHeader(), B->getHeader())) return B;
589  if (DT.dominates(B->getHeader(), A->getHeader())) return A;
590  return A; // Arbitrarily break the tie.
591 }
592 
593 /// getRelevantLoop - Get the most relevant loop associated with the given
594 /// expression, according to PickMostRelevantLoop.
595 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
596  // Test whether we've already computed the most relevant loop for this SCEV.
597  auto Pair = RelevantLoops.insert(std::make_pair(S, nullptr));
598  if (!Pair.second)
599  return Pair.first->second;
600 
601  if (isa<SCEVConstant>(S))
602  // A constant has no relevant loops.
603  return nullptr;
604  if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
605  if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
606  return Pair.first->second = SE.LI.getLoopFor(I->getParent());
607  // A non-instruction has no relevant loops.
608  return nullptr;
609  }
610  if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
611  const Loop *L = nullptr;
612  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
613  L = AR->getLoop();
614  for (const SCEV *Op : N->operands())
615  L = PickMostRelevantLoop(L, getRelevantLoop(Op), SE.DT);
616  return RelevantLoops[N] = L;
617  }
618  if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
619  const Loop *Result = getRelevantLoop(C->getOperand());
620  return RelevantLoops[C] = Result;
621  }
622  if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
623  const Loop *Result = PickMostRelevantLoop(
624  getRelevantLoop(D->getLHS()), getRelevantLoop(D->getRHS()), SE.DT);
625  return RelevantLoops[D] = Result;
626  }
627  llvm_unreachable("Unexpected SCEV type!");
628 }
629 
630 namespace {
631 
632 /// LoopCompare - Compare loops by PickMostRelevantLoop.
633 class LoopCompare {
634  DominatorTree &DT;
635 public:
636  explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
637 
638  bool operator()(std::pair<const Loop *, const SCEV *> LHS,
639  std::pair<const Loop *, const SCEV *> RHS) const {
640  // Keep pointer operands sorted at the end.
641  if (LHS.second->getType()->isPointerTy() !=
642  RHS.second->getType()->isPointerTy())
643  return LHS.second->getType()->isPointerTy();
644 
645  // Compare loops with PickMostRelevantLoop.
646  if (LHS.first != RHS.first)
647  return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
648 
649  // If one operand is a non-constant negative and the other is not,
650  // put the non-constant negative on the right so that a sub can
651  // be used instead of a negate and add.
652  if (LHS.second->isNonConstantNegative()) {
653  if (!RHS.second->isNonConstantNegative())
654  return false;
655  } else if (RHS.second->isNonConstantNegative())
656  return true;
657 
658  // Otherwise they are equivalent according to this comparison.
659  return false;
660  }
661 };
662 
663 }
664 
665 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
666  Type *Ty = SE.getEffectiveSCEVType(S->getType());
667 
668  // Collect all the add operands in a loop, along with their associated loops.
669  // Iterate in reverse so that constants are emitted last, all else equal, and
670  // so that pointer operands are inserted first, which the code below relies on
671  // to form more involved GEPs.
673  for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
674  E(S->op_begin()); I != E; ++I)
675  OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
676 
677  // Sort by loop. Use a stable sort so that constants follow non-constants and
678  // pointer operands precede non-pointer operands.
679  std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(SE.DT));
680 
681  // Emit instructions to add all the operands. Hoist as much as possible
682  // out of loops, and form meaningful getelementptrs where possible.
683  Value *Sum = nullptr;
684  for (auto I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E;) {
685  const Loop *CurLoop = I->first;
686  const SCEV *Op = I->second;
687  if (!Sum) {
688  // This is the first operand. Just expand it.
689  Sum = expand(Op);
690  ++I;
691  } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
692  // The running sum expression is a pointer. Try to form a getelementptr
693  // at this level with that as the base.
695  for (; I != E && I->first == CurLoop; ++I) {
696  // If the operand is SCEVUnknown and not instructions, peek through
697  // it, to enable more of it to be folded into the GEP.
698  const SCEV *X = I->second;
699  if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
700  if (!isa<Instruction>(U->getValue()))
701  X = SE.getSCEV(U->getValue());
702  NewOps.push_back(X);
703  }
704  Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
705  } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
706  // The running sum is an integer, and there's a pointer at this level.
707  // Try to form a getelementptr. If the running sum is instructions,
708  // use a SCEVUnknown to avoid re-analyzing them.
710  NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
711  SE.getSCEV(Sum));
712  for (++I; I != E && I->first == CurLoop; ++I)
713  NewOps.push_back(I->second);
714  Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
715  } else if (Op->isNonConstantNegative()) {
716  // Instead of doing a negate and add, just do a subtract.
717  Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
718  Sum = InsertNoopCastOfTo(Sum, Ty);
719  Sum = InsertBinop(Instruction::Sub, Sum, W);
720  ++I;
721  } else {
722  // A simple add.
723  Value *W = expandCodeFor(Op, Ty);
724  Sum = InsertNoopCastOfTo(Sum, Ty);
725  // Canonicalize a constant to the RHS.
726  if (isa<Constant>(Sum)) std::swap(Sum, W);
727  Sum = InsertBinop(Instruction::Add, Sum, W);
728  ++I;
729  }
730  }
731 
732  return Sum;
733 }
734 
735 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
736  Type *Ty = SE.getEffectiveSCEVType(S->getType());
737 
738  // Collect all the mul operands in a loop, along with their associated loops.
739  // Iterate in reverse so that constants are emitted last, all else equal.
741  for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
742  E(S->op_begin()); I != E; ++I)
743  OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
744 
745  // Sort by loop. Use a stable sort so that constants follow non-constants.
746  std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(SE.DT));
747 
748  // Emit instructions to mul all the operands. Hoist as much as possible
749  // out of loops.
750  Value *Prod = nullptr;
751  auto I = OpsAndLoops.begin();
752 
753  // Expand the calculation of X pow N in the following manner:
754  // Let N = P1 + P2 + ... + PK, where all P are powers of 2. Then:
755  // X pow N = (X pow P1) * (X pow P2) * ... * (X pow PK).
756  const auto ExpandOpBinPowN = [this, &I, &OpsAndLoops, &Ty]() {
757  auto E = I;
758  // Calculate how many times the same operand from the same loop is included
759  // into this power.
760  uint64_t Exponent = 0;
761  const uint64_t MaxExponent = UINT64_MAX >> 1;
762  // No one sane will ever try to calculate such huge exponents, but if we
763  // need this, we stop on UINT64_MAX / 2 because we need to exit the loop
764  // below when the power of 2 exceeds our Exponent, and we want it to be
765  // 1u << 31 at most to not deal with unsigned overflow.
766  while (E != OpsAndLoops.end() && *I == *E && Exponent != MaxExponent) {
767  ++Exponent;
768  ++E;
769  }
770  assert(Exponent > 0 && "Trying to calculate a zeroth exponent of operand?");
771 
772  // Calculate powers with exponents 1, 2, 4, 8 etc. and include those of them
773  // that are needed into the result.
774  Value *P = expandCodeFor(I->second, Ty);
775  Value *Result = nullptr;
776  if (Exponent & 1)
777  Result = P;
778  for (uint64_t BinExp = 2; BinExp <= Exponent; BinExp <<= 1) {
779  P = InsertBinop(Instruction::Mul, P, P);
780  if (Exponent & BinExp)
781  Result = Result ? InsertBinop(Instruction::Mul, Result, P) : P;
782  }
783 
784  I = E;
785  assert(Result && "Nothing was expanded?");
786  return Result;
787  };
788 
789  while (I != OpsAndLoops.end()) {
790  if (!Prod) {
791  // This is the first operand. Just expand it.
792  Prod = ExpandOpBinPowN();
793  } else if (I->second->isAllOnesValue()) {
794  // Instead of doing a multiply by negative one, just do a negate.
795  Prod = InsertNoopCastOfTo(Prod, Ty);
796  Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
797  ++I;
798  } else {
799  // A simple mul.
800  Value *W = ExpandOpBinPowN();
801  Prod = InsertNoopCastOfTo(Prod, Ty);
802  // Canonicalize a constant to the RHS.
803  if (isa<Constant>(Prod)) std::swap(Prod, W);
804  const APInt *RHS;
805  if (match(W, m_Power2(RHS))) {
806  // Canonicalize Prod*(1<<C) to Prod<<C.
807  assert(!Ty->isVectorTy() && "vector types are not SCEVable");
808  Prod = InsertBinop(Instruction::Shl, Prod,
809  ConstantInt::get(Ty, RHS->logBase2()));
810  } else {
811  Prod = InsertBinop(Instruction::Mul, Prod, W);
812  }
813  }
814  }
815 
816  return Prod;
817 }
818 
819 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
820  Type *Ty = SE.getEffectiveSCEVType(S->getType());
821 
822  Value *LHS = expandCodeFor(S->getLHS(), Ty);
823  if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
824  const APInt &RHS = SC->getAPInt();
825  if (RHS.isPowerOf2())
826  return InsertBinop(Instruction::LShr, LHS,
827  ConstantInt::get(Ty, RHS.logBase2()));
828  }
829 
830  Value *RHS = expandCodeFor(S->getRHS(), Ty);
831  return InsertBinop(Instruction::UDiv, LHS, RHS);
832 }
833 
834 /// Move parts of Base into Rest to leave Base with the minimal
835 /// expression that provides a pointer operand suitable for a
836 /// GEP expansion.
837 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
838  ScalarEvolution &SE) {
839  while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
840  Base = A->getStart();
841  Rest = SE.getAddExpr(Rest,
842  SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
843  A->getStepRecurrence(SE),
844  A->getLoop(),
845  A->getNoWrapFlags(SCEV::FlagNW)));
846  }
847  if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
848  Base = A->getOperand(A->getNumOperands()-1);
849  SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
850  NewAddOps.back() = Rest;
851  Rest = SE.getAddExpr(NewAddOps);
852  ExposePointerBase(Base, Rest, SE);
853  }
854 }
855 
856 /// Determine if this is a well-behaved chain of instructions leading back to
857 /// the PHI. If so, it may be reused by expanded expressions.
858 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV,
859  const Loop *L) {
860  if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
861  (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV)))
862  return false;
863  // If any of the operands don't dominate the insert position, bail.
864  // Addrec operands are always loop-invariant, so this can only happen
865  // if there are instructions which haven't been hoisted.
866  if (L == IVIncInsertLoop) {
867  for (User::op_iterator OI = IncV->op_begin()+1,
868  OE = IncV->op_end(); OI != OE; ++OI)
869  if (Instruction *OInst = dyn_cast<Instruction>(OI))
870  if (!SE.DT.dominates(OInst, IVIncInsertPos))
871  return false;
872  }
873  // Advance to the next instruction.
874  IncV = dyn_cast<Instruction>(IncV->getOperand(0));
875  if (!IncV)
876  return false;
877 
878  if (IncV->mayHaveSideEffects())
879  return false;
880 
881  if (IncV != PN)
882  return true;
883 
884  return isNormalAddRecExprPHI(PN, IncV, L);
885 }
886 
887 /// getIVIncOperand returns an induction variable increment's induction
888 /// variable operand.
889 ///
890 /// If allowScale is set, any type of GEP is allowed as long as the nonIV
891 /// operands dominate InsertPos.
892 ///
893 /// If allowScale is not set, ensure that a GEP increment conforms to one of the
894 /// simple patterns generated by getAddRecExprPHILiterally and
895 /// expandAddtoGEP. If the pattern isn't recognized, return NULL.
897  Instruction *InsertPos,
898  bool allowScale) {
899  if (IncV == InsertPos)
900  return nullptr;
901 
902  switch (IncV->getOpcode()) {
903  default:
904  return nullptr;
905  // Check for a simple Add/Sub or GEP of a loop invariant step.
906  case Instruction::Add:
907  case Instruction::Sub: {
908  Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1));
909  if (!OInst || SE.DT.dominates(OInst, InsertPos))
910  return dyn_cast<Instruction>(IncV->getOperand(0));
911  return nullptr;
912  }
913  case Instruction::BitCast:
914  return dyn_cast<Instruction>(IncV->getOperand(0));
915  case Instruction::GetElementPtr:
916  for (auto I = IncV->op_begin() + 1, E = IncV->op_end(); I != E; ++I) {
917  if (isa<Constant>(*I))
918  continue;
919  if (Instruction *OInst = dyn_cast<Instruction>(*I)) {
920  if (!SE.DT.dominates(OInst, InsertPos))
921  return nullptr;
922  }
923  if (allowScale) {
924  // allow any kind of GEP as long as it can be hoisted.
925  continue;
926  }
927  // This must be a pointer addition of constants (pretty), which is already
928  // handled, or some number of address-size elements (ugly). Ugly geps
929  // have 2 operands. i1* is used by the expander to represent an
930  // address-size element.
931  if (IncV->getNumOperands() != 2)
932  return nullptr;
933  unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace();
934  if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS)
935  && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS))
936  return nullptr;
937  break;
938  }
939  return dyn_cast<Instruction>(IncV->getOperand(0));
940  }
941 }
942 
943 /// If the insert point of the current builder or any of the builders on the
944 /// stack of saved builders has 'I' as its insert point, update it to point to
945 /// the instruction after 'I'. This is intended to be used when the instruction
946 /// 'I' is being moved. If this fixup is not done and 'I' is moved to a
947 /// different block, the inconsistent insert point (with a mismatched
948 /// Instruction and Block) can lead to an instruction being inserted in a block
949 /// other than its parent.
950 void SCEVExpander::fixupInsertPoints(Instruction *I) {
951  BasicBlock::iterator It(*I);
952  BasicBlock::iterator NewInsertPt = std::next(It);
953  if (Builder.GetInsertPoint() == It)
954  Builder.SetInsertPoint(&*NewInsertPt);
955  for (auto *InsertPtGuard : InsertPointGuards)
956  if (InsertPtGuard->GetInsertPoint() == It)
957  InsertPtGuard->SetInsertPoint(NewInsertPt);
958 }
959 
960 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make
961 /// it available to other uses in this loop. Recursively hoist any operands,
962 /// until we reach a value that dominates InsertPos.
964  if (SE.DT.dominates(IncV, InsertPos))
965  return true;
966 
967  // InsertPos must itself dominate IncV so that IncV's new position satisfies
968  // its existing users.
969  if (isa<PHINode>(InsertPos) ||
970  !SE.DT.dominates(InsertPos->getParent(), IncV->getParent()))
971  return false;
972 
973  if (!SE.LI.movementPreservesLCSSAForm(IncV, InsertPos))
974  return false;
975 
976  // Check that the chain of IV operands leading back to Phi can be hoisted.
978  for(;;) {
979  Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true);
980  if (!Oper)
981  return false;
982  // IncV is safe to hoist.
983  IVIncs.push_back(IncV);
984  IncV = Oper;
985  if (SE.DT.dominates(IncV, InsertPos))
986  break;
987  }
988  for (auto I = IVIncs.rbegin(), E = IVIncs.rend(); I != E; ++I) {
989  fixupInsertPoints(*I);
990  (*I)->moveBefore(InsertPos);
991  }
992  return true;
993 }
994 
995 /// Determine if this cyclic phi is in a form that would have been generated by
996 /// LSR. We don't care if the phi was actually expanded in this pass, as long
997 /// as it is in a low-cost form, for example, no implied multiplication. This
998 /// should match any patterns generated by getAddRecExprPHILiterally and
999 /// expandAddtoGEP.
1000 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV,
1001  const Loop *L) {
1002  for(Instruction *IVOper = IncV;
1003  (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(),
1004  /*allowScale=*/false));) {
1005  if (IVOper == PN)
1006  return true;
1007  }
1008  return false;
1009 }
1010 
1011 /// expandIVInc - Expand an IV increment at Builder's current InsertPos.
1012 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may
1013 /// need to materialize IV increments elsewhere to handle difficult situations.
1014 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
1015  Type *ExpandTy, Type *IntTy,
1016  bool useSubtract) {
1017  Value *IncV;
1018  // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
1019  if (ExpandTy->isPointerTy()) {
1020  PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
1021  // If the step isn't constant, don't use an implicitly scaled GEP, because
1022  // that would require a multiply inside the loop.
1023  if (!isa<ConstantInt>(StepV))
1024  GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
1025  GEPPtrTy->getAddressSpace());
1026  const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
1027  IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
1028  if (IncV->getType() != PN->getType()) {
1029  IncV = Builder.CreateBitCast(IncV, PN->getType());
1030  rememberInstruction(IncV);
1031  }
1032  } else {
1033  IncV = useSubtract ?
1034  Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
1035  Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
1036  rememberInstruction(IncV);
1037  }
1038  return IncV;
1039 }
1040 
1041 /// \brief Hoist the addrec instruction chain rooted in the loop phi above the
1042 /// position. This routine assumes that this is possible (has been checked).
1043 void SCEVExpander::hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist,
1044  Instruction *Pos, PHINode *LoopPhi) {
1045  do {
1046  if (DT->dominates(InstToHoist, Pos))
1047  break;
1048  // Make sure the increment is where we want it. But don't move it
1049  // down past a potential existing post-inc user.
1050  fixupInsertPoints(InstToHoist);
1051  InstToHoist->moveBefore(Pos);
1052  Pos = InstToHoist;
1053  InstToHoist = cast<Instruction>(InstToHoist->getOperand(0));
1054  } while (InstToHoist != LoopPhi);
1055 }
1056 
1057 /// \brief Check whether we can cheaply express the requested SCEV in terms of
1058 /// the available PHI SCEV by truncation and/or inversion of the step.
1060  const SCEVAddRecExpr *Phi,
1061  const SCEVAddRecExpr *Requested,
1062  bool &InvertStep) {
1063  Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType());
1064  Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType());
1065 
1066  if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth())
1067  return false;
1068 
1069  // Try truncate it if necessary.
1070  Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy));
1071  if (!Phi)
1072  return false;
1073 
1074  // Check whether truncation will help.
1075  if (Phi == Requested) {
1076  InvertStep = false;
1077  return true;
1078  }
1079 
1080  // Check whether inverting will help: {R,+,-1} == R - {0,+,1}.
1081  if (SE.getAddExpr(Requested->getStart(),
1082  SE.getNegativeSCEV(Requested)) == Phi) {
1083  InvertStep = true;
1084  return true;
1085  }
1086 
1087  return false;
1088 }
1089 
1090 static bool IsIncrementNSW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) {
1091  if (!isa<IntegerType>(AR->getType()))
1092  return false;
1093 
1094  unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth();
1095  Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2);
1096  const SCEV *Step = AR->getStepRecurrence(SE);
1097  const SCEV *OpAfterExtend = SE.getAddExpr(SE.getSignExtendExpr(Step, WideTy),
1098  SE.getSignExtendExpr(AR, WideTy));
1099  const SCEV *ExtendAfterOp =
1100  SE.getSignExtendExpr(SE.getAddExpr(AR, Step), WideTy);
1101  return ExtendAfterOp == OpAfterExtend;
1102 }
1103 
1104 static bool IsIncrementNUW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) {
1105  if (!isa<IntegerType>(AR->getType()))
1106  return false;
1107 
1108  unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth();
1109  Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2);
1110  const SCEV *Step = AR->getStepRecurrence(SE);
1111  const SCEV *OpAfterExtend = SE.getAddExpr(SE.getZeroExtendExpr(Step, WideTy),
1112  SE.getZeroExtendExpr(AR, WideTy));
1113  const SCEV *ExtendAfterOp =
1114  SE.getZeroExtendExpr(SE.getAddExpr(AR, Step), WideTy);
1115  return ExtendAfterOp == OpAfterExtend;
1116 }
1117 
1118 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
1119 /// the base addrec, which is the addrec without any non-loop-dominating
1120 /// values, and return the PHI.
1121 PHINode *
1122 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
1123  const Loop *L,
1124  Type *ExpandTy,
1125  Type *IntTy,
1126  Type *&TruncTy,
1127  bool &InvertStep) {
1128  assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
1129 
1130  // Reuse a previously-inserted PHI, if present.
1131  BasicBlock *LatchBlock = L->getLoopLatch();
1132  if (LatchBlock) {
1133  PHINode *AddRecPhiMatch = nullptr;
1134  Instruction *IncV = nullptr;
1135  TruncTy = nullptr;
1136  InvertStep = false;
1137 
1138  // Only try partially matching scevs that need truncation and/or
1139  // step-inversion if we know this loop is outside the current loop.
1140  bool TryNonMatchingSCEV =
1141  IVIncInsertLoop &&
1142  SE.DT.properlyDominates(LatchBlock, IVIncInsertLoop->getHeader());
1143 
1144  for (auto &I : *L->getHeader()) {
1145  auto *PN = dyn_cast<PHINode>(&I);
1146  if (!PN || !SE.isSCEVable(PN->getType()))
1147  continue;
1148 
1149  const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PN));
1150  if (!PhiSCEV)
1151  continue;
1152 
1153  bool IsMatchingSCEV = PhiSCEV == Normalized;
1154  // We only handle truncation and inversion of phi recurrences for the
1155  // expanded expression if the expanded expression's loop dominates the
1156  // loop we insert to. Check now, so we can bail out early.
1157  if (!IsMatchingSCEV && !TryNonMatchingSCEV)
1158  continue;
1159 
1160  Instruction *TempIncV =
1161  cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
1162 
1163  // Check whether we can reuse this PHI node.
1164  if (LSRMode) {
1165  if (!isExpandedAddRecExprPHI(PN, TempIncV, L))
1166  continue;
1167  if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos))
1168  continue;
1169  } else {
1170  if (!isNormalAddRecExprPHI(PN, TempIncV, L))
1171  continue;
1172  }
1173 
1174  // Stop if we have found an exact match SCEV.
1175  if (IsMatchingSCEV) {
1176  IncV = TempIncV;
1177  TruncTy = nullptr;
1178  InvertStep = false;
1179  AddRecPhiMatch = PN;
1180  break;
1181  }
1182 
1183  // Try whether the phi can be translated into the requested form
1184  // (truncated and/or offset by a constant).
1185  if ((!TruncTy || InvertStep) &&
1186  canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) {
1187  // Record the phi node. But don't stop we might find an exact match
1188  // later.
1189  AddRecPhiMatch = PN;
1190  IncV = TempIncV;
1191  TruncTy = SE.getEffectiveSCEVType(Normalized->getType());
1192  }
1193  }
1194 
1195  if (AddRecPhiMatch) {
1196  // Potentially, move the increment. We have made sure in
1197  // isExpandedAddRecExprPHI or hoistIVInc that this is possible.
1198  if (L == IVIncInsertLoop)
1199  hoistBeforePos(&SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch);
1200 
1201  // Ok, the add recurrence looks usable.
1202  // Remember this PHI, even in post-inc mode.
1203  InsertedValues.insert(AddRecPhiMatch);
1204  // Remember the increment.
1205  rememberInstruction(IncV);
1206  return AddRecPhiMatch;
1207  }
1208  }
1209 
1210  // Save the original insertion point so we can restore it when we're done.
1211  SCEVInsertPointGuard Guard(Builder, this);
1212 
1213  // Another AddRec may need to be recursively expanded below. For example, if
1214  // this AddRec is quadratic, the StepV may itself be an AddRec in this
1215  // loop. Remove this loop from the PostIncLoops set before expanding such
1216  // AddRecs. Otherwise, we cannot find a valid position for the step
1217  // (i.e. StepV can never dominate its loop header). Ideally, we could do
1218  // SavedIncLoops.swap(PostIncLoops), but we generally have a single element,
1219  // so it's not worth implementing SmallPtrSet::swap.
1220  PostIncLoopSet SavedPostIncLoops = PostIncLoops;
1221  PostIncLoops.clear();
1222 
1223  // Expand code for the start value into the loop preheader.
1224  assert(L->getLoopPreheader() &&
1225  "Can't expand add recurrences without a loop preheader!");
1226  Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
1228 
1229  // StartV must have been be inserted into L's preheader to dominate the new
1230  // phi.
1231  assert(!isa<Instruction>(StartV) ||
1232  SE.DT.properlyDominates(cast<Instruction>(StartV)->getParent(),
1233  L->getHeader()));
1234 
1235  // Expand code for the step value. Do this before creating the PHI so that PHI
1236  // reuse code doesn't see an incomplete PHI.
1237  const SCEV *Step = Normalized->getStepRecurrence(SE);
1238  // If the stride is negative, insert a sub instead of an add for the increment
1239  // (unless it's a constant, because subtracts of constants are canonicalized
1240  // to adds).
1241  bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1242  if (useSubtract)
1243  Step = SE.getNegativeSCEV(Step);
1244  // Expand the step somewhere that dominates the loop header.
1245  Value *StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front());
1246 
1247  // The no-wrap behavior proved by IsIncrement(NUW|NSW) is only applicable if
1248  // we actually do emit an addition. It does not apply if we emit a
1249  // subtraction.
1250  bool IncrementIsNUW = !useSubtract && IsIncrementNUW(SE, Normalized);
1251  bool IncrementIsNSW = !useSubtract && IsIncrementNSW(SE, Normalized);
1252 
1253  // Create the PHI.
1254  BasicBlock *Header = L->getHeader();
1255  Builder.SetInsertPoint(Header, Header->begin());
1256  pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1257  PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
1258  Twine(IVName) + ".iv");
1259  rememberInstruction(PN);
1260 
1261  // Create the step instructions and populate the PHI.
1262  for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1263  BasicBlock *Pred = *HPI;
1264 
1265  // Add a start value.
1266  if (!L->contains(Pred)) {
1267  PN->addIncoming(StartV, Pred);
1268  continue;
1269  }
1270 
1271  // Create a step value and add it to the PHI.
1272  // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the
1273  // instructions at IVIncInsertPos.
1274  Instruction *InsertPos = L == IVIncInsertLoop ?
1275  IVIncInsertPos : Pred->getTerminator();
1276  Builder.SetInsertPoint(InsertPos);
1277  Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1278 
1279  if (isa<OverflowingBinaryOperator>(IncV)) {
1280  if (IncrementIsNUW)
1281  cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap();
1282  if (IncrementIsNSW)
1283  cast<BinaryOperator>(IncV)->setHasNoSignedWrap();
1284  }
1285  PN->addIncoming(IncV, Pred);
1286  }
1287 
1288  // After expanding subexpressions, restore the PostIncLoops set so the caller
1289  // can ensure that IVIncrement dominates the current uses.
1290  PostIncLoops = SavedPostIncLoops;
1291 
1292  // Remember this PHI, even in post-inc mode.
1293  InsertedValues.insert(PN);
1294 
1295  return PN;
1296 }
1297 
1298 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1299  Type *STy = S->getType();
1300  Type *IntTy = SE.getEffectiveSCEVType(STy);
1301  const Loop *L = S->getLoop();
1302 
1303  // Determine a normalized form of this expression, which is the expression
1304  // before any post-inc adjustment is made.
1305  const SCEVAddRecExpr *Normalized = S;
1306  if (PostIncLoops.count(L)) {
1308  Loops.insert(L);
1309  Normalized = cast<SCEVAddRecExpr>(normalizeForPostIncUse(S, Loops, SE));
1310  }
1311 
1312  // Strip off any non-loop-dominating component from the addrec start.
1313  const SCEV *Start = Normalized->getStart();
1314  const SCEV *PostLoopOffset = nullptr;
1315  if (!SE.properlyDominates(Start, L->getHeader())) {
1316  PostLoopOffset = Start;
1317  Start = SE.getConstant(Normalized->getType(), 0);
1318  Normalized = cast<SCEVAddRecExpr>(
1319  SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1320  Normalized->getLoop(),
1321  Normalized->getNoWrapFlags(SCEV::FlagNW)));
1322  }
1323 
1324  // Strip off any non-loop-dominating component from the addrec step.
1325  const SCEV *Step = Normalized->getStepRecurrence(SE);
1326  const SCEV *PostLoopScale = nullptr;
1327  if (!SE.dominates(Step, L->getHeader())) {
1328  PostLoopScale = Step;
1329  Step = SE.getConstant(Normalized->getType(), 1);
1330  if (!Start->isZero()) {
1331  // The normalization below assumes that Start is constant zero, so if
1332  // it isn't re-associate Start to PostLoopOffset.
1333  assert(!PostLoopOffset && "Start not-null but PostLoopOffset set?");
1334  PostLoopOffset = Start;
1335  Start = SE.getConstant(Normalized->getType(), 0);
1336  }
1337  Normalized =
1338  cast<SCEVAddRecExpr>(SE.getAddRecExpr(
1339  Start, Step, Normalized->getLoop(),
1340  Normalized->getNoWrapFlags(SCEV::FlagNW)));
1341  }
1342 
1343  // Expand the core addrec. If we need post-loop scaling, force it to
1344  // expand to an integer type to avoid the need for additional casting.
1345  Type *ExpandTy = PostLoopScale ? IntTy : STy;
1346  // We can't use a pointer type for the addrec if the pointer type is
1347  // non-integral.
1348  Type *AddRecPHIExpandTy =
1349  DL.isNonIntegralPointerType(STy) ? Normalized->getType() : ExpandTy;
1350 
1351  // In some cases, we decide to reuse an existing phi node but need to truncate
1352  // it and/or invert the step.
1353  Type *TruncTy = nullptr;
1354  bool InvertStep = false;
1355  PHINode *PN = getAddRecExprPHILiterally(Normalized, L, AddRecPHIExpandTy,
1356  IntTy, TruncTy, InvertStep);
1357 
1358  // Accommodate post-inc mode, if necessary.
1359  Value *Result;
1360  if (!PostIncLoops.count(L))
1361  Result = PN;
1362  else {
1363  // In PostInc mode, use the post-incremented value.
1364  BasicBlock *LatchBlock = L->getLoopLatch();
1365  assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1366  Result = PN->getIncomingValueForBlock(LatchBlock);
1367 
1368  // For an expansion to use the postinc form, the client must call
1369  // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
1370  // or dominated by IVIncInsertPos.
1371  if (isa<Instruction>(Result) &&
1372  !SE.DT.dominates(cast<Instruction>(Result),
1373  &*Builder.GetInsertPoint())) {
1374  // The induction variable's postinc expansion does not dominate this use.
1375  // IVUsers tries to prevent this case, so it is rare. However, it can
1376  // happen when an IVUser outside the loop is not dominated by the latch
1377  // block. Adjusting IVIncInsertPos before expansion begins cannot handle
1378  // all cases. Consider a phi outide whose operand is replaced during
1379  // expansion with the value of the postinc user. Without fundamentally
1380  // changing the way postinc users are tracked, the only remedy is
1381  // inserting an extra IV increment. StepV might fold into PostLoopOffset,
1382  // but hopefully expandCodeFor handles that.
1383  bool useSubtract =
1384  !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1385  if (useSubtract)
1386  Step = SE.getNegativeSCEV(Step);
1387  Value *StepV;
1388  {
1389  // Expand the step somewhere that dominates the loop header.
1390  SCEVInsertPointGuard Guard(Builder, this);
1391  StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front());
1392  }
1393  Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1394  }
1395  }
1396 
1397  // We have decided to reuse an induction variable of a dominating loop. Apply
1398  // truncation and/or invertion of the step.
1399  if (TruncTy) {
1400  Type *ResTy = Result->getType();
1401  // Normalize the result type.
1402  if (ResTy != SE.getEffectiveSCEVType(ResTy))
1403  Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy));
1404  // Truncate the result.
1405  if (TruncTy != Result->getType()) {
1406  Result = Builder.CreateTrunc(Result, TruncTy);
1407  rememberInstruction(Result);
1408  }
1409  // Invert the result.
1410  if (InvertStep) {
1411  Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy),
1412  Result);
1413  rememberInstruction(Result);
1414  }
1415  }
1416 
1417  // Re-apply any non-loop-dominating scale.
1418  if (PostLoopScale) {
1419  assert(S->isAffine() && "Can't linearly scale non-affine recurrences.");
1420  Result = InsertNoopCastOfTo(Result, IntTy);
1421  Result = Builder.CreateMul(Result,
1422  expandCodeFor(PostLoopScale, IntTy));
1423  rememberInstruction(Result);
1424  }
1425 
1426  // Re-apply any non-loop-dominating offset.
1427  if (PostLoopOffset) {
1428  if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1429  if (Result->getType()->isIntegerTy()) {
1430  Value *Base = expandCodeFor(PostLoopOffset, ExpandTy);
1431  const SCEV *const OffsetArray[1] = {SE.getUnknown(Result)};
1432  Result = expandAddToGEP(OffsetArray, OffsetArray + 1, PTy, IntTy, Base);
1433  } else {
1434  const SCEV *const OffsetArray[1] = {PostLoopOffset};
1435  Result =
1436  expandAddToGEP(OffsetArray, OffsetArray + 1, PTy, IntTy, Result);
1437  }
1438  } else {
1439  Result = InsertNoopCastOfTo(Result, IntTy);
1440  Result = Builder.CreateAdd(Result,
1441  expandCodeFor(PostLoopOffset, IntTy));
1442  rememberInstruction(Result);
1443  }
1444  }
1445 
1446  return Result;
1447 }
1448 
1449 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1450  if (!CanonicalMode) return expandAddRecExprLiterally(S);
1451 
1452  Type *Ty = SE.getEffectiveSCEVType(S->getType());
1453  const Loop *L = S->getLoop();
1454 
1455  // First check for an existing canonical IV in a suitable type.
1456  PHINode *CanonicalIV = nullptr;
1457  if (PHINode *PN = L->getCanonicalInductionVariable())
1458  if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1459  CanonicalIV = PN;
1460 
1461  // Rewrite an AddRec in terms of the canonical induction variable, if
1462  // its type is more narrow.
1463  if (CanonicalIV &&
1464  SE.getTypeSizeInBits(CanonicalIV->getType()) >
1465  SE.getTypeSizeInBits(Ty)) {
1467  for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1468  NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1469  Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1471  BasicBlock::iterator NewInsertPt =
1472  findInsertPointAfter(cast<Instruction>(V), Builder.GetInsertBlock());
1473  V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), nullptr,
1474  &*NewInsertPt);
1475  return V;
1476  }
1477 
1478  // {X,+,F} --> X + {0,+,F}
1479  if (!S->getStart()->isZero()) {
1480  SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1481  NewOps[0] = SE.getConstant(Ty, 0);
1482  const SCEV *Rest = SE.getAddRecExpr(NewOps, L,
1484 
1485  // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1486  // comments on expandAddToGEP for details.
1487  const SCEV *Base = S->getStart();
1488  const SCEV *RestArray[1] = { Rest };
1489  // Dig into the expression to find the pointer base for a GEP.
1490  ExposePointerBase(Base, RestArray[0], SE);
1491  // If we found a pointer, expand the AddRec with a GEP.
1492  if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1493  // Make sure the Base isn't something exotic, such as a multiplied
1494  // or divided pointer value. In those cases, the result type isn't
1495  // actually a pointer type.
1496  if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1497  Value *StartV = expand(Base);
1498  assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1499  return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1500  }
1501  }
1502 
1503  // Just do a normal add. Pre-expand the operands to suppress folding.
1504  //
1505  // The LHS and RHS values are factored out of the expand call to make the
1506  // output independent of the argument evaluation order.
1507  const SCEV *AddExprLHS = SE.getUnknown(expand(S->getStart()));
1508  const SCEV *AddExprRHS = SE.getUnknown(expand(Rest));
1509  return expand(SE.getAddExpr(AddExprLHS, AddExprRHS));
1510  }
1511 
1512  // If we don't yet have a canonical IV, create one.
1513  if (!CanonicalIV) {
1514  // Create and insert the PHI node for the induction variable in the
1515  // specified loop.
1516  BasicBlock *Header = L->getHeader();
1517  pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1518  CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1519  &Header->front());
1520  rememberInstruction(CanonicalIV);
1521 
1522  SmallSet<BasicBlock *, 4> PredSeen;
1523  Constant *One = ConstantInt::get(Ty, 1);
1524  for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1525  BasicBlock *HP = *HPI;
1526  if (!PredSeen.insert(HP).second) {
1527  // There must be an incoming value for each predecessor, even the
1528  // duplicates!
1529  CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP);
1530  continue;
1531  }
1532 
1533  if (L->contains(HP)) {
1534  // Insert a unit add instruction right before the terminator
1535  // corresponding to the back-edge.
1536  Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1537  "indvar.next",
1538  HP->getTerminator());
1539  Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1540  rememberInstruction(Add);
1541  CanonicalIV->addIncoming(Add, HP);
1542  } else {
1543  CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1544  }
1545  }
1546  }
1547 
1548  // {0,+,1} --> Insert a canonical induction variable into the loop!
1549  if (S->isAffine() && S->getOperand(1)->isOne()) {
1550  assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1551  "IVs with types different from the canonical IV should "
1552  "already have been handled!");
1553  return CanonicalIV;
1554  }
1555 
1556  // {0,+,F} --> {0,+,1} * F
1557 
1558  // If this is a simple linear addrec, emit it now as a special case.
1559  if (S->isAffine()) // {0,+,F} --> i*F
1560  return
1561  expand(SE.getTruncateOrNoop(
1562  SE.getMulExpr(SE.getUnknown(CanonicalIV),
1563  SE.getNoopOrAnyExtend(S->getOperand(1),
1564  CanonicalIV->getType())),
1565  Ty));
1566 
1567  // If this is a chain of recurrences, turn it into a closed form, using the
1568  // folders, then expandCodeFor the closed form. This allows the folders to
1569  // simplify the expression without having to build a bunch of special code
1570  // into this folder.
1571  const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1572 
1573  // Promote S up to the canonical IV type, if the cast is foldable.
1574  const SCEV *NewS = S;
1575  const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1576  if (isa<SCEVAddRecExpr>(Ext))
1577  NewS = Ext;
1578 
1579  const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1580  //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1581 
1582  // Truncate the result down to the original type, if needed.
1583  const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1584  return expand(T);
1585 }
1586 
1587 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1588  Type *Ty = SE.getEffectiveSCEVType(S->getType());
1589  Value *V = expandCodeFor(S->getOperand(),
1590  SE.getEffectiveSCEVType(S->getOperand()->getType()));
1591  Value *I = Builder.CreateTrunc(V, Ty);
1592  rememberInstruction(I);
1593  return I;
1594 }
1595 
1596 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1597  Type *Ty = SE.getEffectiveSCEVType(S->getType());
1598  Value *V = expandCodeFor(S->getOperand(),
1599  SE.getEffectiveSCEVType(S->getOperand()->getType()));
1600  Value *I = Builder.CreateZExt(V, Ty);
1601  rememberInstruction(I);
1602  return I;
1603 }
1604 
1605 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1606  Type *Ty = SE.getEffectiveSCEVType(S->getType());
1607  Value *V = expandCodeFor(S->getOperand(),
1608  SE.getEffectiveSCEVType(S->getOperand()->getType()));
1609  Value *I = Builder.CreateSExt(V, Ty);
1610  rememberInstruction(I);
1611  return I;
1612 }
1613 
1614 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1615  Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1616  Type *Ty = LHS->getType();
1617  for (int i = S->getNumOperands()-2; i >= 0; --i) {
1618  // In the case of mixed integer and pointer types, do the
1619  // rest of the comparisons as integer.
1620  if (S->getOperand(i)->getType() != Ty) {
1621  Ty = SE.getEffectiveSCEVType(Ty);
1622  LHS = InsertNoopCastOfTo(LHS, Ty);
1623  }
1624  Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1625  Value *ICmp = Builder.CreateICmpSGT(LHS, RHS);
1626  rememberInstruction(ICmp);
1627  Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1628  rememberInstruction(Sel);
1629  LHS = Sel;
1630  }
1631  // In the case of mixed integer and pointer types, cast the
1632  // final result back to the pointer type.
1633  if (LHS->getType() != S->getType())
1634  LHS = InsertNoopCastOfTo(LHS, S->getType());
1635  return LHS;
1636 }
1637 
1638 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1639  Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1640  Type *Ty = LHS->getType();
1641  for (int i = S->getNumOperands()-2; i >= 0; --i) {
1642  // In the case of mixed integer and pointer types, do the
1643  // rest of the comparisons as integer.
1644  if (S->getOperand(i)->getType() != Ty) {
1645  Ty = SE.getEffectiveSCEVType(Ty);
1646  LHS = InsertNoopCastOfTo(LHS, Ty);
1647  }
1648  Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1649  Value *ICmp = Builder.CreateICmpUGT(LHS, RHS);
1650  rememberInstruction(ICmp);
1651  Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1652  rememberInstruction(Sel);
1653  LHS = Sel;
1654  }
1655  // In the case of mixed integer and pointer types, cast the
1656  // final result back to the pointer type.
1657  if (LHS->getType() != S->getType())
1658  LHS = InsertNoopCastOfTo(LHS, S->getType());
1659  return LHS;
1660 }
1661 
1663  Instruction *IP) {
1664  setInsertPoint(IP);
1665  return expandCodeFor(SH, Ty);
1666 }
1667 
1669  // Expand the code for this SCEV.
1670  Value *V = expand(SH);
1671  if (Ty) {
1672  assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1673  "non-trivial casts should be done with the SCEVs directly!");
1674  V = InsertNoopCastOfTo(V, Ty);
1675  }
1676  return V;
1677 }
1678 
1679 ScalarEvolution::ValueOffsetPair
1680 SCEVExpander::FindValueInExprValueMap(const SCEV *S,
1681  const Instruction *InsertPt) {
1682  SetVector<ScalarEvolution::ValueOffsetPair> *Set = SE.getSCEVValues(S);
1683  // If the expansion is not in CanonicalMode, and the SCEV contains any
1684  // sub scAddRecExpr type SCEV, it is required to expand the SCEV literally.
1685  if (CanonicalMode || !SE.containsAddRecurrence(S)) {
1686  // If S is scConstant, it may be worse to reuse an existing Value.
1687  if (S->getSCEVType() != scConstant && Set) {
1688  // Choose a Value from the set which dominates the insertPt.
1689  // insertPt should be inside the Value's parent loop so as not to break
1690  // the LCSSA form.
1691  for (auto const &VOPair : *Set) {
1692  Value *V = VOPair.first;
1693  ConstantInt *Offset = VOPair.second;
1694  Instruction *EntInst = nullptr;
1695  if (V && isa<Instruction>(V) && (EntInst = cast<Instruction>(V)) &&
1696  S->getType() == V->getType() &&
1697  EntInst->getFunction() == InsertPt->getFunction() &&
1698  SE.DT.dominates(EntInst, InsertPt) &&
1699  (SE.LI.getLoopFor(EntInst->getParent()) == nullptr ||
1700  SE.LI.getLoopFor(EntInst->getParent())->contains(InsertPt)))
1701  return {V, Offset};
1702  }
1703  }
1704  }
1705  return {nullptr, nullptr};
1706 }
1707 
1708 // The expansion of SCEV will either reuse a previous Value in ExprValueMap,
1709 // or expand the SCEV literally. Specifically, if the expansion is in LSRMode,
1710 // and the SCEV contains any sub scAddRecExpr type SCEV, it will be expanded
1711 // literally, to prevent LSR's transformed SCEV from being reverted. Otherwise,
1712 // the expansion will try to reuse Value from ExprValueMap, and only when it
1713 // fails, expand the SCEV literally.
1714 Value *SCEVExpander::expand(const SCEV *S) {
1715  // Compute an insertion point for this SCEV object. Hoist the instructions
1716  // as far out in the loop nest as possible.
1717  Instruction *InsertPt = &*Builder.GetInsertPoint();
1718  for (Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock());;
1719  L = L->getParentLoop())
1720  if (SE.isLoopInvariant(S, L)) {
1721  if (!L) break;
1722  if (BasicBlock *Preheader = L->getLoopPreheader())
1723  InsertPt = Preheader->getTerminator();
1724  else {
1725  // LSR sets the insertion point for AddRec start/step values to the
1726  // block start to simplify value reuse, even though it's an invalid
1727  // position. SCEVExpander must correct for this in all cases.
1728  InsertPt = &*L->getHeader()->getFirstInsertionPt();
1729  }
1730  } else {
1731  // If the SCEV is computable at this level, insert it into the header
1732  // after the PHIs (and after any other instructions that we've inserted
1733  // there) so that it is guaranteed to dominate any user inside the loop.
1734  if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1735  InsertPt = &*L->getHeader()->getFirstInsertionPt();
1736  while (InsertPt->getIterator() != Builder.GetInsertPoint() &&
1737  (isInsertedInstruction(InsertPt) ||
1738  isa<DbgInfoIntrinsic>(InsertPt))) {
1739  InsertPt = &*std::next(InsertPt->getIterator());
1740  }
1741  break;
1742  }
1743 
1744  // Check to see if we already expanded this here.
1745  auto I = InsertedExpressions.find(std::make_pair(S, InsertPt));
1746  if (I != InsertedExpressions.end())
1747  return I->second;
1748 
1749  SCEVInsertPointGuard Guard(Builder, this);
1750  Builder.SetInsertPoint(InsertPt);
1751 
1752  // Expand the expression into instructions.
1753  ScalarEvolution::ValueOffsetPair VO = FindValueInExprValueMap(S, InsertPt);
1754  Value *V = VO.first;
1755 
1756  if (!V)
1757  V = visit(S);
1758  else if (VO.second) {
1759  if (PointerType *Vty = dyn_cast<PointerType>(V->getType())) {
1760  Type *Ety = Vty->getPointerElementType();
1761  int64_t Offset = VO.second->getSExtValue();
1762  int64_t ESize = SE.getTypeSizeInBits(Ety);
1763  if ((Offset * 8) % ESize == 0) {
1764  ConstantInt *Idx =
1765  ConstantInt::getSigned(VO.second->getType(), -(Offset * 8) / ESize);
1766  V = Builder.CreateGEP(Ety, V, Idx, "scevgep");
1767  } else {
1768  ConstantInt *Idx =
1769  ConstantInt::getSigned(VO.second->getType(), -Offset);
1770  unsigned AS = Vty->getAddressSpace();
1771  V = Builder.CreateBitCast(V, Type::getInt8PtrTy(SE.getContext(), AS));
1772  V = Builder.CreateGEP(Type::getInt8Ty(SE.getContext()), V, Idx,
1773  "uglygep");
1774  V = Builder.CreateBitCast(V, Vty);
1775  }
1776  } else {
1777  V = Builder.CreateSub(V, VO.second);
1778  }
1779  }
1780  // Remember the expanded value for this SCEV at this location.
1781  //
1782  // This is independent of PostIncLoops. The mapped value simply materializes
1783  // the expression at this insertion point. If the mapped value happened to be
1784  // a postinc expansion, it could be reused by a non-postinc user, but only if
1785  // its insertion point was already at the head of the loop.
1786  InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1787  return V;
1788 }
1789 
1790 void SCEVExpander::rememberInstruction(Value *I) {
1791  if (!PostIncLoops.empty())
1792  InsertedPostIncValues.insert(I);
1793  else
1794  InsertedValues.insert(I);
1795 }
1796 
1797 /// getOrInsertCanonicalInductionVariable - This method returns the
1798 /// canonical induction variable of the specified type for the specified
1799 /// loop (inserting one if there is none). A canonical induction variable
1800 /// starts at zero and steps by one on each iteration.
1801 PHINode *
1803  Type *Ty) {
1804  assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1805 
1806  // Build a SCEV for {0,+,1}<L>.
1807  // Conservatively use FlagAnyWrap for now.
1808  const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1809  SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1810 
1811  // Emit code for it.
1812  SCEVInsertPointGuard Guard(Builder, this);
1813  PHINode *V =
1814  cast<PHINode>(expandCodeFor(H, nullptr, &L->getHeader()->front()));
1815 
1816  return V;
1817 }
1818 
1819 /// replaceCongruentIVs - Check for congruent phis in this loop header and
1820 /// replace them with their most canonical representative. Return the number of
1821 /// phis eliminated.
1822 ///
1823 /// This does not depend on any SCEVExpander state but should be used in
1824 /// the same context that SCEVExpander is used.
1825 unsigned
1828  const TargetTransformInfo *TTI) {
1829  // Find integer phis in order of increasing width.
1831  for (auto &I : *L->getHeader()) {
1832  if (auto *PN = dyn_cast<PHINode>(&I))
1833  Phis.push_back(PN);
1834  else
1835  break;
1836  }
1837 
1838  if (TTI)
1839  std::sort(Phis.begin(), Phis.end(), [](Value *LHS, Value *RHS) {
1840  // Put pointers at the back and make sure pointer < pointer = false.
1841  if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy())
1842  return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy();
1843  return RHS->getType()->getPrimitiveSizeInBits() <
1844  LHS->getType()->getPrimitiveSizeInBits();
1845  });
1846 
1847  unsigned NumElim = 0;
1849  // Process phis from wide to narrow. Map wide phis to their truncation
1850  // so narrow phis can reuse them.
1851  for (PHINode *Phi : Phis) {
1852  auto SimplifyPHINode = [&](PHINode *PN) -> Value * {
1853  if (Value *V = SimplifyInstruction(PN, {DL, &SE.TLI, &SE.DT, &SE.AC}))
1854  return V;
1855  if (!SE.isSCEVable(PN->getType()))
1856  return nullptr;
1857  auto *Const = dyn_cast<SCEVConstant>(SE.getSCEV(PN));
1858  if (!Const)
1859  return nullptr;
1860  return Const->getValue();
1861  };
1862 
1863  // Fold constant phis. They may be congruent to other constant phis and
1864  // would confuse the logic below that expects proper IVs.
1865  if (Value *V = SimplifyPHINode(Phi)) {
1866  if (V->getType() != Phi->getType())
1867  continue;
1868  Phi->replaceAllUsesWith(V);
1869  DeadInsts.emplace_back(Phi);
1870  ++NumElim;
1872  << "INDVARS: Eliminated constant iv: " << *Phi << '\n');
1873  continue;
1874  }
1875 
1876  if (!SE.isSCEVable(Phi->getType()))
1877  continue;
1878 
1879  PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)];
1880  if (!OrigPhiRef) {
1881  OrigPhiRef = Phi;
1882  if (Phi->getType()->isIntegerTy() && TTI &&
1883  TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) {
1884  // This phi can be freely truncated to the narrowest phi type. Map the
1885  // truncated expression to it so it will be reused for narrow types.
1886  const SCEV *TruncExpr =
1887  SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType());
1888  ExprToIVMap[TruncExpr] = Phi;
1889  }
1890  continue;
1891  }
1892 
1893  // Replacing a pointer phi with an integer phi or vice-versa doesn't make
1894  // sense.
1895  if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy())
1896  continue;
1897 
1898  if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1899  Instruction *OrigInc = dyn_cast<Instruction>(
1900  OrigPhiRef->getIncomingValueForBlock(LatchBlock));
1901  Instruction *IsomorphicInc =
1902  dyn_cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1903 
1904  if (OrigInc && IsomorphicInc) {
1905  // If this phi has the same width but is more canonical, replace the
1906  // original with it. As part of the "more canonical" determination,
1907  // respect a prior decision to use an IV chain.
1908  if (OrigPhiRef->getType() == Phi->getType() &&
1909  !(ChainedPhis.count(Phi) ||
1910  isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L)) &&
1911  (ChainedPhis.count(Phi) ||
1912  isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) {
1913  std::swap(OrigPhiRef, Phi);
1914  std::swap(OrigInc, IsomorphicInc);
1915  }
1916  // Replacing the congruent phi is sufficient because acyclic
1917  // redundancy elimination, CSE/GVN, should handle the
1918  // rest. However, once SCEV proves that a phi is congruent,
1919  // it's often the head of an IV user cycle that is isomorphic
1920  // with the original phi. It's worth eagerly cleaning up the
1921  // common case of a single IV increment so that DeleteDeadPHIs
1922  // can remove cycles that had postinc uses.
1923  const SCEV *TruncExpr =
1924  SE.getTruncateOrNoop(SE.getSCEV(OrigInc), IsomorphicInc->getType());
1925  if (OrigInc != IsomorphicInc &&
1926  TruncExpr == SE.getSCEV(IsomorphicInc) &&
1927  SE.LI.replacementPreservesLCSSAForm(IsomorphicInc, OrigInc) &&
1928  hoistIVInc(OrigInc, IsomorphicInc)) {
1930  dbgs() << "INDVARS: Eliminated congruent iv.inc: "
1931  << *IsomorphicInc << '\n');
1932  Value *NewInc = OrigInc;
1933  if (OrigInc->getType() != IsomorphicInc->getType()) {
1934  Instruction *IP = nullptr;
1935  if (PHINode *PN = dyn_cast<PHINode>(OrigInc))
1936  IP = &*PN->getParent()->getFirstInsertionPt();
1937  else
1938  IP = OrigInc->getNextNode();
1939 
1940  IRBuilder<> Builder(IP);
1941  Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc());
1942  NewInc = Builder.CreateTruncOrBitCast(
1943  OrigInc, IsomorphicInc->getType(), IVName);
1944  }
1945  IsomorphicInc->replaceAllUsesWith(NewInc);
1946  DeadInsts.emplace_back(IsomorphicInc);
1947  }
1948  }
1949  }
1950  DEBUG_WITH_TYPE(DebugType, dbgs() << "INDVARS: Eliminated congruent iv: "
1951  << *Phi << '\n');
1952  ++NumElim;
1953  Value *NewIV = OrigPhiRef;
1954  if (OrigPhiRef->getType() != Phi->getType()) {
1955  IRBuilder<> Builder(&*L->getHeader()->getFirstInsertionPt());
1956  Builder.SetCurrentDebugLocation(Phi->getDebugLoc());
1957  NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName);
1958  }
1959  Phi->replaceAllUsesWith(NewIV);
1960  DeadInsts.emplace_back(Phi);
1961  }
1962  return NumElim;
1963 }
1964 
1966  const Instruction *At, Loop *L) {
1968  getRelatedExistingExpansion(S, At, L);
1969  if (VO && VO.getValue().second == nullptr)
1970  return VO.getValue().first;
1971  return nullptr;
1972 }
1973 
1976  Loop *L) {
1977  using namespace llvm::PatternMatch;
1978 
1979  SmallVector<BasicBlock *, 4> ExitingBlocks;
1980  L->getExitingBlocks(ExitingBlocks);
1981 
1982  // Look for suitable value in simple conditions at the loop exits.
1983  for (BasicBlock *BB : ExitingBlocks) {
1984  ICmpInst::Predicate Pred;
1985  Instruction *LHS, *RHS;
1986  BasicBlock *TrueBB, *FalseBB;
1987 
1988  if (!match(BB->getTerminator(),
1989  m_Br(m_ICmp(Pred, m_Instruction(LHS), m_Instruction(RHS)),
1990  TrueBB, FalseBB)))
1991  continue;
1992 
1993  if (SE.getSCEV(LHS) == S && SE.DT.dominates(LHS, At))
1994  return ScalarEvolution::ValueOffsetPair(LHS, nullptr);
1995 
1996  if (SE.getSCEV(RHS) == S && SE.DT.dominates(RHS, At))
1997  return ScalarEvolution::ValueOffsetPair(RHS, nullptr);
1998  }
1999 
2000  // Use expand's logic which is used for reusing a previous Value in
2001  // ExprValueMap.
2002  ScalarEvolution::ValueOffsetPair VO = FindValueInExprValueMap(S, At);
2003  if (VO.first)
2004  return VO;
2005 
2006  // There is potential to make this significantly smarter, but this simple
2007  // heuristic already gets some interesting cases.
2008 
2009  // Can not find suitable value.
2010  return None;
2011 }
2012 
2013 bool SCEVExpander::isHighCostExpansionHelper(
2014  const SCEV *S, Loop *L, const Instruction *At,
2015  SmallPtrSetImpl<const SCEV *> &Processed) {
2016 
2017  // If we can find an existing value for this scev available at the point "At"
2018  // then consider the expression cheap.
2019  if (At && getRelatedExistingExpansion(S, At, L))
2020  return false;
2021 
2022  // Zero/One operand expressions
2023  switch (S->getSCEVType()) {
2024  case scUnknown:
2025  case scConstant:
2026  return false;
2027  case scTruncate:
2028  return isHighCostExpansionHelper(cast<SCEVTruncateExpr>(S)->getOperand(),
2029  L, At, Processed);
2030  case scZeroExtend:
2031  return isHighCostExpansionHelper(cast<SCEVZeroExtendExpr>(S)->getOperand(),
2032  L, At, Processed);
2033  case scSignExtend:
2034  return isHighCostExpansionHelper(cast<SCEVSignExtendExpr>(S)->getOperand(),
2035  L, At, Processed);
2036  }
2037 
2038  if (!Processed.insert(S).second)
2039  return false;
2040 
2041  if (auto *UDivExpr = dyn_cast<SCEVUDivExpr>(S)) {
2042  // If the divisor is a power of two and the SCEV type fits in a native
2043  // integer, consider the division cheap irrespective of whether it occurs in
2044  // the user code since it can be lowered into a right shift.
2045  if (auto *SC = dyn_cast<SCEVConstant>(UDivExpr->getRHS()))
2046  if (SC->getAPInt().isPowerOf2()) {
2047  const DataLayout &DL =
2049  unsigned Width = cast<IntegerType>(UDivExpr->getType())->getBitWidth();
2050  return DL.isIllegalInteger(Width);
2051  }
2052 
2053  // UDivExpr is very likely a UDiv that ScalarEvolution's HowFarToZero or
2054  // HowManyLessThans produced to compute a precise expression, rather than a
2055  // UDiv from the user's code. If we can't find a UDiv in the code with some
2056  // simple searching, assume the former consider UDivExpr expensive to
2057  // compute.
2058  BasicBlock *ExitingBB = L->getExitingBlock();
2059  if (!ExitingBB)
2060  return true;
2061 
2062  // At the beginning of this function we already tried to find existing value
2063  // for plain 'S'. Now try to lookup 'S + 1' since it is common pattern
2064  // involving division. This is just a simple search heuristic.
2065  if (!At)
2066  At = &ExitingBB->back();
2067  if (!getRelatedExistingExpansion(
2068  SE.getAddExpr(S, SE.getConstant(S->getType(), 1)), At, L))
2069  return true;
2070  }
2071 
2072  // HowManyLessThans uses a Max expression whenever the loop is not guarded by
2073  // the exit condition.
2074  if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
2075  return true;
2076 
2077  // Recurse past nary expressions, which commonly occur in the
2078  // BackedgeTakenCount. They may already exist in program code, and if not,
2079  // they are not too expensive rematerialize.
2080  if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(S)) {
2081  for (auto *Op : NAry->operands())
2082  if (isHighCostExpansionHelper(Op, L, At, Processed))
2083  return true;
2084  }
2085 
2086  // If we haven't recognized an expensive SCEV pattern, assume it's an
2087  // expression produced by program code.
2088  return false;
2089 }
2090 
2092  Instruction *IP) {
2093  assert(IP);
2094  switch (Pred->getKind()) {
2096  return expandUnionPredicate(cast<SCEVUnionPredicate>(Pred), IP);
2098  return expandEqualPredicate(cast<SCEVEqualPredicate>(Pred), IP);
2099  case SCEVPredicate::P_Wrap: {
2100  auto *AddRecPred = cast<SCEVWrapPredicate>(Pred);
2101  return expandWrapPredicate(AddRecPred, IP);
2102  }
2103  }
2104  llvm_unreachable("Unknown SCEV predicate type");
2105 }
2106 
2108  Instruction *IP) {
2109  Value *Expr0 = expandCodeFor(Pred->getLHS(), Pred->getLHS()->getType(), IP);
2110  Value *Expr1 = expandCodeFor(Pred->getRHS(), Pred->getRHS()->getType(), IP);
2111 
2112  Builder.SetInsertPoint(IP);
2113  auto *I = Builder.CreateICmpNE(Expr0, Expr1, "ident.check");
2114  return I;
2115 }
2116 
2118  Instruction *Loc, bool Signed) {
2119  assert(AR->isAffine() && "Cannot generate RT check for "
2120  "non-affine expression");
2121 
2122  SCEVUnionPredicate Pred;
2123  const SCEV *ExitCount =
2124  SE.getPredicatedBackedgeTakenCount(AR->getLoop(), Pred);
2125 
2126  assert(ExitCount != SE.getCouldNotCompute() && "Invalid loop count");
2127 
2128  const SCEV *Step = AR->getStepRecurrence(SE);
2129  const SCEV *Start = AR->getStart();
2130 
2131  unsigned SrcBits = SE.getTypeSizeInBits(ExitCount->getType());
2132  unsigned DstBits = SE.getTypeSizeInBits(AR->getType());
2133 
2134  // The expression {Start,+,Step} has nusw/nssw if
2135  // Step < 0, Start - |Step| * Backedge <= Start
2136  // Step >= 0, Start + |Step| * Backedge > Start
2137  // and |Step| * Backedge doesn't unsigned overflow.
2138 
2139  IntegerType *CountTy = IntegerType::get(Loc->getContext(), SrcBits);
2140  Builder.SetInsertPoint(Loc);
2141  Value *TripCountVal = expandCodeFor(ExitCount, CountTy, Loc);
2142 
2143  IntegerType *Ty =
2144  IntegerType::get(Loc->getContext(), SE.getTypeSizeInBits(AR->getType()));
2145 
2146  Value *StepValue = expandCodeFor(Step, Ty, Loc);
2147  Value *NegStepValue = expandCodeFor(SE.getNegativeSCEV(Step), Ty, Loc);
2148  Value *StartValue = expandCodeFor(Start, Ty, Loc);
2149 
2150  ConstantInt *Zero =
2152 
2153  Builder.SetInsertPoint(Loc);
2154  // Compute |Step|
2155  Value *StepCompare = Builder.CreateICmp(ICmpInst::ICMP_SLT, StepValue, Zero);
2156  Value *AbsStep = Builder.CreateSelect(StepCompare, NegStepValue, StepValue);
2157 
2158  // Get the backedge taken count and truncate or extended to the AR type.
2159  Value *TruncTripCount = Builder.CreateZExtOrTrunc(TripCountVal, Ty);
2160  auto *MulF = Intrinsic::getDeclaration(Loc->getModule(),
2161  Intrinsic::umul_with_overflow, Ty);
2162 
2163  // Compute |Step| * Backedge
2164  CallInst *Mul = Builder.CreateCall(MulF, {AbsStep, TruncTripCount}, "mul");
2165  Value *MulV = Builder.CreateExtractValue(Mul, 0, "mul.result");
2166  Value *OfMul = Builder.CreateExtractValue(Mul, 1, "mul.overflow");
2167 
2168  // Compute:
2169  // Start + |Step| * Backedge < Start
2170  // Start - |Step| * Backedge > Start
2171  Value *Add = Builder.CreateAdd(StartValue, MulV);
2172  Value *Sub = Builder.CreateSub(StartValue, MulV);
2173 
2174  Value *EndCompareGT = Builder.CreateICmp(
2175  Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT, Sub, StartValue);
2176 
2177  Value *EndCompareLT = Builder.CreateICmp(
2178  Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT, Add, StartValue);
2179 
2180  // Select the answer based on the sign of Step.
2181  Value *EndCheck =
2182  Builder.CreateSelect(StepCompare, EndCompareGT, EndCompareLT);
2183 
2184  // If the backedge taken count type is larger than the AR type,
2185  // check that we don't drop any bits by truncating it. If we are
2186  // droping bits, then we have overflow (unless the step is zero).
2187  if (SE.getTypeSizeInBits(CountTy) > SE.getTypeSizeInBits(Ty)) {
2188  auto MaxVal = APInt::getMaxValue(DstBits).zext(SrcBits);
2189  auto *BackedgeCheck =
2190  Builder.CreateICmp(ICmpInst::ICMP_UGT, TripCountVal,
2191  ConstantInt::get(Loc->getContext(), MaxVal));
2192  BackedgeCheck = Builder.CreateAnd(
2193  BackedgeCheck, Builder.CreateICmp(ICmpInst::ICMP_NE, StepValue, Zero));
2194 
2195  EndCheck = Builder.CreateOr(EndCheck, BackedgeCheck);
2196  }
2197 
2198  EndCheck = Builder.CreateOr(EndCheck, OfMul);
2199  return EndCheck;
2200 }
2201 
2203  Instruction *IP) {
2204  const auto *A = cast<SCEVAddRecExpr>(Pred->getExpr());
2205  Value *NSSWCheck = nullptr, *NUSWCheck = nullptr;
2206 
2207  // Add a check for NUSW
2209  NUSWCheck = generateOverflowCheck(A, IP, false);
2210 
2211  // Add a check for NSSW
2213  NSSWCheck = generateOverflowCheck(A, IP, true);
2214 
2215  if (NUSWCheck && NSSWCheck)
2216  return Builder.CreateOr(NUSWCheck, NSSWCheck);
2217 
2218  if (NUSWCheck)
2219  return NUSWCheck;
2220 
2221  if (NSSWCheck)
2222  return NSSWCheck;
2223 
2224  return ConstantInt::getFalse(IP->getContext());
2225 }
2226 
2228  Instruction *IP) {
2229  auto *BoolType = IntegerType::get(IP->getContext(), 1);
2230  Value *Check = ConstantInt::getNullValue(BoolType);
2231 
2232  // Loop over all checks in this set.
2233  for (auto Pred : Union->getPredicates()) {
2234  auto *NextCheck = expandCodeForPredicate(Pred, IP);
2235  Builder.SetInsertPoint(IP);
2236  Check = Builder.CreateOr(Check, NextCheck);
2237  }
2238 
2239  return Check;
2240 }
2241 
2242 namespace {
2243 // Search for a SCEV subexpression that is not safe to expand. Any expression
2244 // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely
2245 // UDiv expressions. We don't know if the UDiv is derived from an IR divide
2246 // instruction, but the important thing is that we prove the denominator is
2247 // nonzero before expansion.
2248 //
2249 // IVUsers already checks that IV-derived expressions are safe. So this check is
2250 // only needed when the expression includes some subexpression that is not IV
2251 // derived.
2252 //
2253 // Currently, we only allow division by a nonzero constant here. If this is
2254 // inadequate, we could easily allow division by SCEVUnknown by using
2255 // ValueTracking to check isKnownNonZero().
2256 //
2257 // We cannot generally expand recurrences unless the step dominates the loop
2258 // header. The expander handles the special case of affine recurrences by
2259 // scaling the recurrence outside the loop, but this technique isn't generally
2260 // applicable. Expanding a nested recurrence outside a loop requires computing
2261 // binomial coefficients. This could be done, but the recurrence has to be in a
2262 // perfectly reduced form, which can't be guaranteed.
2263 struct SCEVFindUnsafe {
2264  ScalarEvolution &SE;
2265  bool IsUnsafe;
2266 
2267  SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {}
2268 
2269  bool follow(const SCEV *S) {
2270  if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2271  const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS());
2272  if (!SC || SC->getValue()->isZero()) {
2273  IsUnsafe = true;
2274  return false;
2275  }
2276  }
2277  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2278  const SCEV *Step = AR->getStepRecurrence(SE);
2279  if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) {
2280  IsUnsafe = true;
2281  return false;
2282  }
2283  }
2284  return true;
2285  }
2286  bool isDone() const { return IsUnsafe; }
2287 };
2288 }
2289 
2290 namespace llvm {
2291 bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) {
2292  SCEVFindUnsafe Search(SE);
2293  visitAll(S, Search);
2294  return !Search.IsUnsafe;
2295 }
2296 }
static unsigned getBitWidth(Type *Ty, const DataLayout &DL)
Returns the bitwidth of the given scalar or pointer type.
const SCEV * getTruncateOrNoop(const SCEV *V, Type *Ty)
Return a SCEV corresponding to a conversion of the input value to the specified type.
static bool Check(DecodeStatus &Out, DecodeStatus In)
const NoneType None
Definition: None.h:24
uint64_t CallInst * C
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:109
bool hoistIVInc(Instruction *IncV, Instruction *InsertPos)
Utility for hoisting an IV increment.
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:523
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
static IntegerType * getInt1Ty(LLVMContext &C)
Definition: Type.cpp:173
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
BlockT * getLoopLatch() const
If there is a single latch block for this loop, return it.
Definition: LoopInfoImpl.h:157
Value * getExactExistingExpansion(const SCEV *S, const Instruction *At, Loop *L)
Try to find existing LLVM IR value for S available at the point At.
This class represents an incoming formal argument to a Function.
Definition: Argument.h:30
NodeTy * getNextNode()
Get the next node, or nullptr for the list tail.
Definition: ilist_node.h:289
const SCEV * getConstant(ConstantInt *V)
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
Type * getEffectiveSCEVType(Type *Ty) const
Return a type with the same bitwidth as the given type and which represents how SCEV will treat the g...
static Constant * getGetElementPtr(Type *Ty, Constant *C, ArrayRef< Constant *> IdxList, bool InBounds=false, Optional< unsigned > InRangeIndex=None, Type *OnlyIfReducedTy=nullptr)
Getelementptr form.
Definition: Constants.h:1115
bool isSized(SmallPtrSetImpl< Type *> *Visited=nullptr) const
Return true if it makes sense to take the size of this type.
Definition: Type.h:262
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
const SCEV * normalizeForPostIncUse(const SCEV *S, const PostIncLoopSet &Loops, ScalarEvolution &SE)
Normalize S to be post-increment for all loops present in Loops.
The main scalar evolution driver.
APInt zext(unsigned width) const
Zero extend to a new width.
Definition: APInt.cpp:865
bool isZero() const
Return true if the expression is a constant zero.
This class represents a function call, abstracting a target machine&#39;s calling convention.
BlockT * getLoopPreheader() const
If there is a preheader for this loop, return it.
Definition: LoopInfoImpl.h:106
static PointerType * get(Type *ElementType, unsigned AddressSpace)
This constructs a pointer to an object of the specified type in a numbered address space...
Definition: Type.cpp:617
unsigned less than
Definition: InstrTypes.h:885
Optional< ScalarEvolution::ValueOffsetPair > getRelatedExistingExpansion(const SCEV *S, const Instruction *At, Loop *L)
Try to find the ValueOffsetPair for S.
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:697
This class represents a truncation of an integer value to a smaller integer value.
Value * expandWrapPredicate(const SCEVWrapPredicate *P, Instruction *Loc)
A specialized variant of expandCodeForPredicate, handling the case when we are expanding code for a S...
A debug info location.
Definition: DebugLoc.h:34
const SCEV * getOperand() const
Hexagon Common GEP
static void SimplifyAddOperands(SmallVectorImpl< const SCEV *> &Ops, Type *Ty, ScalarEvolution &SE)
SimplifyAddOperands - Sort and simplify a list of add operands.
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:227
op_iterator op_begin()
Definition: User.h:214
unsigned getElementContainingOffset(uint64_t Offset) const
Given a valid byte offset into the structure, returns the structure index that contains it...
Definition: DataLayout.cpp:84
This is the base class for unary cast operator classes.
return AArch64::GPR64RegClass contains(Reg)
LLVMContext & getContext() const
Return the LLVMContext in which this type was uniqued.
Definition: Type.h:130
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:207
A templated base class for SmallPtrSet which provides the typesafe interface that is common across al...
Definition: SmallPtrSet.h:336
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:252
#define DEBUG_WITH_TYPE(TYPE, X)
DEBUG_WITH_TYPE macro - This macro should be used by passes to emit debug information.
Definition: Debug.h:64
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
Used to lazily calculate structure layout information for a target machine, based on the DataLayout s...
Definition: DataLayout.h:493
Hexagon Hardware Loops
Value * expandCodeForPredicate(const SCEVPredicate *Pred, Instruction *Loc)
Generates a code sequence that evaluates this predicate.
Type * getPointerElementType() const
Definition: Type.h:373
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:361
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:81
static const Loop * PickMostRelevantLoop(const Loop *A, const Loop *B, DominatorTree &DT)
PickMostRelevantLoop - Given two loops pick the one that&#39;s most relevant for SCEV expansion...
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:560
DebugType
Definition: COFF.h:640
Class to represent struct types.
Definition: DerivedTypes.h:201
A Use represents the edge between a Value definition and its users.
Definition: Use.h:56
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:42
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:197
LLVMContext & getContext() const
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:664
This node represents multiplication of some number of SCEVs.
Value * generateOverflowCheck(const SCEVAddRecExpr *AR, Instruction *Loc, bool Signed)
Generates code that evaluates if the AR expression will overflow.
const APInt & getAPInt() const
BlockT * getHeader() const
Definition: LoopInfo.h:107
ConstantInt * getValue() const
static Constant * get(unsigned Opcode, Constant *C1, Constant *C2, unsigned Flags=0, Type *OnlyIfReducedTy=nullptr)
get - Return a binary or shift operator constant expression, folding if possible. ...
Definition: Constants.cpp:1711
A constant value that is initialized with an expression using other constant values.
Definition: Constants.h:862
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
bool isTruncateFree(Type *Ty1, Type *Ty2) const
Return true if it&#39;s free to truncate a value of type Ty1 to type Ty2.
This node represents a polynomial recurrence on the trip count of the specified loop.
const T & getValue() const LLVM_LVALUE_FUNCTION
Definition: Optional.h:129
Class to represent array types.
Definition: DerivedTypes.h:369
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
op_iterator op_begin() const
void SetCurrentDebugLocation(DebugLoc L)
Set location information used by debugging information.
Definition: IRBuilder.h:152
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:428
cst_pred_ty< is_power2 > m_Power2()
Match an integer or vector power of 2.
Definition: PatternMatch.h:330
const SCEV * getStepRecurrence(ScalarEvolution &SE) const
Constructs and returns the recurrence indicating how much this expression steps by.
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:290
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:140
Function * getDeclaration(Module *M, ID id, ArrayRef< Type *> Tys=None)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
Definition: Function.cpp:975
const SCEV * getAddRecExpr(const SCEV *Start, const SCEV *Step, const Loop *L, SCEV::NoWrapFlags Flags)
Get an add recurrence expression for the specified loop.
static BinaryOperator * CreateAdd(Value *S1, Value *S2, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp)
Value * getOperand(unsigned i) const
Definition: User.h:154
Class to represent pointers.
Definition: DerivedTypes.h:467
#define P(N)
This means that we are dealing with an entirely unknown SCEV value, and only represent it as its LLVM...
const SCEV * getOperand(unsigned i) const
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
const_iterator getFirstInsertionPt() const
Returns an iterator to the first instruction in this block that is suitable for inserting a non-PHI i...
Definition: BasicBlock.cpp:200
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:281
SCEVPredicateKind getKind() const
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
This class represents a binary unsigned division operation.
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:42
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:116
Value * getIncomingValueForBlock(const BasicBlock *BB) const
SmallSet - This maintains a set of unique values, optimizing for the case when the set is small (less...
Definition: SmallSet.h:36
bool isPointerTy() const
True if this is an instance of PointerType.
Definition: Type.h:221
const Instruction & front() const
Definition: BasicBlock.h:264
#define H(x, y, z)
Definition: MD5.cpp:57
const SCEV * getExpr() const override
Implementation of the SCEVPredicate interface.
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
Definition: SmallPtrSet.h:363
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:504
const SCEV * getAddExpr(SmallVectorImpl< const SCEV *> &Ops, SCEV::NoWrapFlags Flags=SCEV::FlagAnyWrap, unsigned Depth=0)
Get a canonical add expression, or something simpler if possible.
const SCEV * getLHS() const
Interval::pred_iterator pred_begin(Interval *I)
pred_begin/pred_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:113
brc_match< Cond_t > m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F)
op_iterator op_end()
Definition: User.h:216
bool any_of(R &&Range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:823
const Instruction & back() const
Definition: BasicBlock.h:266
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:860
static bool FactorOutConstant(const SCEV *&S, const SCEV *&Remainder, const SCEV *Factor, ScalarEvolution &SE, const DataLayout &DL)
FactorOutConstant - Test if S is divisible by Factor, using signed division.
Value * expandCodeFor(const SCEV *SH, Type *Ty, Instruction *I)
Insert code to directly compute the specified SCEV expression into the program.
Interval::pred_iterator pred_end(Interval *I)
Definition: Interval.h:116
bool isAffine() const
Return true if this represents an expression A + B*x where A and B are loop invariant values...
unsigned getAddressSpace() const
Return the address space of the Pointer type.
Definition: DerivedTypes.h:495
self_iterator getIterator()
Definition: ilist_node.h:82
Class to represent integer types.
Definition: DerivedTypes.h:40
std::pair< NoneType, bool > insert(const T &V)
insert - Insert an element into the set if it isn&#39;t already there.
Definition: SmallSet.h:81
static Expected< BitVector > expand(StringRef S, StringRef Original)
Definition: GlobPattern.cpp:28
const Function * getFunction() const
Return the function this instruction belongs to.
Definition: Instruction.cpp:61
const SCEV * getLHS() const
Returns the left hand side of the equality.
void getExitingBlocks(SmallVectorImpl< BlockT *> &ExitingBlocks) const
Return all blocks inside the loop that have successors outside of the loop.
Definition: LoopInfoImpl.h:35
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1320
const AMDGPUAS & AS
const SCEV * getRHS() const
Returns the right hand side of the equality.
static PointerType * getInt8PtrTy(LLVMContext &C, unsigned AS=0)
Definition: Type.cpp:220
const SCEV * getMulExpr(SmallVectorImpl< const SCEV *> &Ops, SCEV::NoWrapFlags Flags=SCEV::FlagAnyWrap, unsigned Depth=0)
Get a canonical multiply expression, or something simpler if possible.
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
static void SplitAddRecs(SmallVectorImpl< const SCEV *> &Ops, Type *Ty, ScalarEvolution &SE)
SplitAddRecs - Flatten a list of add operands, moving addrec start values out to the top level...
signed greater than
Definition: InstrTypes.h:887
This class represents an assumption made using SCEV expressions which can be checked at run-time...
static bool IsIncrementNSW(ScalarEvolution &SE, const SCEVAddRecExpr *AR)
static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, ScalarEvolution &SE)
Move parts of Base into Rest to leave Base with the minimal expression that provides a pointer operan...
static IntegerType * get(LLVMContext &C, unsigned NumBits)
This static method is the primary way of constructing an IntegerType.
Definition: Type.cpp:240
bool contains(const LoopT *L) const
Return true if the specified loop is contained within in this loop.
Definition: LoopInfo.h:117
unsigned getSCEVType() const
bool isNonConstantNegative() const
Return true if the specified scev is negated, but not a constant.
unsigned getNumOperands() const
Definition: User.h:176
static PointerType * getInt1PtrTy(LLVMContext &C, unsigned AS=0)
Definition: Type.cpp:216
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
Type * getType() const
Return the LLVM type of this SCEV expression.
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
PHINode * getOrInsertCanonicalInductionVariable(const Loop *L, Type *Ty)
This method returns the canonical induction variable of the specified type for the specified loop (in...
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:864
bool dominates(const Instruction *Def, const Use &U) const
Return true if Def dominates a use in User.
Definition: Dominators.cpp:239
Module.h This file contains the declarations for the Module class.
signed less than
Definition: InstrTypes.h:889
uint64_t getSizeInBytes() const
Definition: DataLayout.h:501
CHAIN = SC CHAIN, Imm128 - System call.
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:560
static ConstantInt * getSigned(IntegerType *Ty, int64_t V)
Return a ConstantInt with the specified value for the specified type.
Definition: Constants.cpp:574
static PHINode * Create(Type *Ty, unsigned NumReservedValues, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Constructors - NumReservedValues is a hint for the number of incoming edges that this phi node will h...
unsigned logBase2() const
Definition: APInt.h:1727
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:923
PHINode * getCanonicalInductionVariable() const
Check to see if the loop has a canonical induction variable: an integer recurrence that starts at 0 a...
Definition: LoopInfo.cpp:111
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:57
Class for arbitrary precision integers.
Definition: APInt.h:69
This node represents an addition of some number of SCEVs.
static BasicBlock::iterator findInsertPointAfter(Instruction *I, BasicBlock *MustDominate)
bool isPowerOf2() const
Check if this APInt&#39;s value is a power of two greater than zero.
Definition: APInt.h:457
This class represents a signed maximum selection.
iterator_range< user_iterator > users()
Definition: Value.h:395
InstListType::iterator iterator
Instruction iterators...
Definition: BasicBlock.h:91
Value * expandUnionPredicate(const SCEVUnionPredicate *Pred, Instruction *Loc)
A specialized variant of expandCodeForPredicate, handling the case when we are expanding code for a S...
static Constant * getCast(unsigned ops, Constant *C, Type *Ty, bool OnlyIfReduced=false)
Convenience function for getting a Cast operation.
Definition: Constants.cpp:1435
void visitAll(const SCEV *Root, SV &Visitor)
Use SCEVTraversal to visit all nodes in the given expression tree.
static APInt getMaxValue(unsigned numBits)
Gets maximum unsigned value of APInt for specific bit width.
Definition: APInt.h:523
void append(in_iter in_start, in_iter in_end)
Add the specified range to the end of the SmallVector.
Definition: SmallVector.h:398
This class represents a zero extension of a small integer value to a larger integer value...
Value * CreateTruncOrBitCast(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1468
static bool IsIncrementNUW(ScalarEvolution &SE, const SCEVAddRecExpr *AR)
LoopT * getParentLoop() const
Definition: LoopInfo.h:108
static CastInst * Create(Instruction::CastOps, Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Provides a way to construct any of the CastInst subclasses using an opcode instead of the subclass&#39;s ...
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:120
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:284
uint64_t getElementOffset(unsigned Idx) const
Definition: DataLayout.h:515
void emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:656
This class represents an analyzed expression in the program.
static IntegerType * getInt32Ty(LLVMContext &C)
Definition: Type.cpp:176
unsigned getIntegerBitWidth() const
Definition: DerivedTypes.h:97
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:61
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:404
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:218
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:108
static Instruction::CastOps getCastOpcode(const Value *Val, bool SrcIsSigned, Type *Ty, bool DstIsSigned)
Returns the opcode necessary to cast Val into Ty using usual casting rules.
#define I(x, y, z)
Definition: MD5.cpp:58
#define N
bool isZero() const
This is just a convenience method to make client code smaller for a common code.
Definition: Constants.h:193
This class represents a sign extension of a small integer value to a larger integer value...
This class represents an unsigned maximum selection.
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
Instruction * getIVIncOperand(Instruction *IncV, Instruction *InsertPos, bool allowScale)
Return the induction variable increment&#39;s IV operand.
const SCEV * getRHS() const
unsigned replaceCongruentIVs(Loop *L, const DominatorTree *DT, SmallVectorImpl< WeakTrackingVH > &DeadInsts, const TargetTransformInfo *TTI=nullptr)
replace congruent phis with their most canonical representative.
const SmallVectorImpl< const SCEVPredicate * > & getPredicates() const
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
This class represents a composition of other SCEV predicates, and is the class that most clients will...
bool isOne() const
Return true if the expression is a constant one.
unsigned getPrimitiveSizeInBits() const LLVM_READONLY
Return the basic size of this type if it is a primitive type.
Definition: Type.cpp:115
const SCEV * getNegativeSCEV(const SCEV *V, SCEV::NoWrapFlags Flags=SCEV::FlagAnyWrap)
Return the SCEV object corresponding to -V.
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:545
Value * expandEqualPredicate(const SCEVEqualPredicate *Pred, Instruction *Loc)
A specialized variant of expandCodeForPredicate, handling the case when we are expanding code for a S...
LLVM Value Representation.
Definition: Value.h:73
A vector that has set insertion semantics.
Definition: SetVector.h:41
void moveBefore(Instruction *MovePos)
Unlink this instruction from its current basic block and insert it into the basic block that MovePos ...
Definition: Instruction.cpp:88
static Value * SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q)
See if we can fold the given phi. If not, returns null.
bool dominates(const SCEV *S, const BasicBlock *BB)
Return true if elements that makes up the given SCEV dominate the specified basic block...
unsigned greater than
Definition: InstrTypes.h:883
void sort(Policy policy, RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp=Comparator())
Definition: Parallel.h:199
This pass exposes codegen information to IR-level passes.
bool isIllegalInteger(uint64_t Width) const
Definition: DataLayout.h:245
static APInt getNullValue(unsigned numBits)
Get the &#39;0&#39; value.
Definition: APInt.h:562
const TerminatorInst * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:120
This node is a base class providing common functionality for n&#39;ary operators.
This class represents an assumption made on an AddRec expression.
bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE)
Return true if the given expression is safe to expand in the sense that all materialized values are s...
Value * SimplifyInstruction(Instruction *I, const SimplifyQuery &Q, OptimizationRemarkEmitter *ORE=nullptr)
See if we can compute a simplified version of this instruction.
BlockT * getExitingBlock() const
If getExitingBlocks would return exactly one block, return that block.
Definition: LoopInfoImpl.h:50
NoWrapFlags getNoWrapFlags(NoWrapFlags Mask=NoWrapMask) const
const SCEV * getZeroExtendExpr(const SCEV *Op, Type *Ty, unsigned Depth=0)
This class represents an assumption that two SCEV expressions are equal, and this can be checked at r...
static IntegerType * getInt8Ty(LLVMContext &C)
Definition: Type.cpp:174
IncrementWrapFlags getFlags() const
Returns the set assumed no overflow flags.
Type * getElementType() const
Definition: DerivedTypes.h:486
bind_ty< Instruction > m_Instruction(Instruction *&I)
Match an instruction, capturing it if we match.
Definition: PatternMatch.h:359
const SCEV * getSignExtendExpr(const SCEV *Op, Type *Ty, unsigned Depth=0)
const BasicBlock * getParent() const
Definition: Instruction.h:66
static bool canBeCheaplyTransformed(ScalarEvolution &SE, const SCEVAddRecExpr *Phi, const SCEVAddRecExpr *Requested, bool &InvertStep)
Check whether we can cheaply express the requested SCEV in terms of the available PHI SCEV by truncat...
This class represents a constant integer value.
CmpClass_match< LHS, RHS, ICmpInst, ICmpInst::Predicate > m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R)
Definition: PatternMatch.h:837