LLVM  8.0.0svn
InstructionCombining.cpp
Go to the documentation of this file.
1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
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 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
13 //
14 // This pass combines things like:
15 // %Y = add i32 %X, 1
16 // %Z = add i32 %Y, 1
17 // into:
18 // %Z = add i32 %X, 2
19 //
20 // This is a simple worklist driven algorithm.
21 //
22 // This pass guarantees that the following canonicalizations are performed on
23 // the program:
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
31 // shifts.
32 // ... etc.
33 //
34 //===----------------------------------------------------------------------===//
35 
36 #include "InstCombineInternal.h"
37 #include "llvm-c/Initialization.h"
39 #include "llvm/ADT/APInt.h"
40 #include "llvm/ADT/ArrayRef.h"
41 #include "llvm/ADT/DenseMap.h"
42 #include "llvm/ADT/None.h"
43 #include "llvm/ADT/SmallPtrSet.h"
44 #include "llvm/ADT/SmallVector.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/ADT/TinyPtrVector.h"
50 #include "llvm/Analysis/CFG.h"
55 #include "llvm/Analysis/LoopInfo.h"
61 #include "llvm/IR/BasicBlock.h"
62 #include "llvm/IR/CFG.h"
63 #include "llvm/IR/Constant.h"
64 #include "llvm/IR/Constants.h"
65 #include "llvm/IR/DIBuilder.h"
66 #include "llvm/IR/DataLayout.h"
67 #include "llvm/IR/DerivedTypes.h"
68 #include "llvm/IR/Dominators.h"
69 #include "llvm/IR/Function.h"
71 #include "llvm/IR/IRBuilder.h"
72 #include "llvm/IR/InstrTypes.h"
73 #include "llvm/IR/Instruction.h"
74 #include "llvm/IR/Instructions.h"
75 #include "llvm/IR/IntrinsicInst.h"
76 #include "llvm/IR/Intrinsics.h"
78 #include "llvm/IR/Metadata.h"
79 #include "llvm/IR/Operator.h"
80 #include "llvm/IR/PassManager.h"
81 #include "llvm/IR/PatternMatch.h"
82 #include "llvm/IR/Type.h"
83 #include "llvm/IR/Use.h"
84 #include "llvm/IR/User.h"
85 #include "llvm/IR/Value.h"
86 #include "llvm/IR/ValueHandle.h"
87 #include "llvm/Pass.h"
89 #include "llvm/Support/Casting.h"
91 #include "llvm/Support/Compiler.h"
92 #include "llvm/Support/Debug.h"
95 #include "llvm/Support/KnownBits.h"
100 #include <algorithm>
101 #include <cassert>
102 #include <cstdint>
103 #include <memory>
104 #include <string>
105 #include <utility>
106 
107 using namespace llvm;
108 using namespace llvm::PatternMatch;
109 
110 #define DEBUG_TYPE "instcombine"
111 
112 STATISTIC(NumCombined , "Number of insts combined");
113 STATISTIC(NumConstProp, "Number of constant folds");
114 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
115 STATISTIC(NumSunkInst , "Number of instructions sunk");
116 STATISTIC(NumExpand, "Number of expansions");
117 STATISTIC(NumFactor , "Number of factorizations");
118 STATISTIC(NumReassoc , "Number of reassociations");
119 DEBUG_COUNTER(VisitCounter, "instcombine-visit",
120  "Controls which instructions are visited");
121 
122 static cl::opt<bool>
123 EnableCodeSinking("instcombine-code-sinking", cl::desc("Enable code sinking"),
124  cl::init(true));
125 
126 static cl::opt<bool>
127 EnableExpensiveCombines("expensive-combines",
128  cl::desc("Enable expensive instruction combines"));
129 
130 static cl::opt<unsigned>
131 MaxArraySize("instcombine-maxarray-size", cl::init(1024),
132  cl::desc("Maximum array size considered when doing a combine"));
133 
134 // FIXME: Remove this flag when it is no longer necessary to convert
135 // llvm.dbg.declare to avoid inaccurate debug info. Setting this to false
136 // increases variable availability at the cost of accuracy. Variables that
137 // cannot be promoted by mem2reg or SROA will be described as living in memory
138 // for their entire lifetime. However, passes like DSE and instcombine can
139 // delete stores to the alloca, leading to misleading and inaccurate debug
140 // information. This flag can be removed when those passes are fixed.
141 static cl::opt<unsigned> ShouldLowerDbgDeclare("instcombine-lower-dbg-declare",
142  cl::Hidden, cl::init(true));
143 
144 Value *InstCombiner::EmitGEPOffset(User *GEP) {
145  return llvm::EmitGEPOffset(&Builder, DL, GEP);
146 }
147 
148 /// Return true if it is desirable to convert an integer computation from a
149 /// given bit width to a new bit width.
150 /// We don't want to convert from a legal to an illegal type or from a smaller
151 /// to a larger illegal type. A width of '1' is always treated as a legal type
152 /// because i1 is a fundamental type in IR, and there are many specialized
153 /// optimizations for i1 types. Widths of 8, 16 or 32 are equally treated as
154 /// legal to convert to, in order to open up more combining opportunities.
155 /// NOTE: this treats i8, i16 and i32 specially, due to them being so common
156 /// from frontend languages.
157 bool InstCombiner::shouldChangeType(unsigned FromWidth,
158  unsigned ToWidth) const {
159  bool FromLegal = FromWidth == 1 || DL.isLegalInteger(FromWidth);
160  bool ToLegal = ToWidth == 1 || DL.isLegalInteger(ToWidth);
161 
162  // Convert to widths of 8, 16 or 32 even if they are not legal types. Only
163  // shrink types, to prevent infinite loops.
164  if (ToWidth < FromWidth && (ToWidth == 8 || ToWidth == 16 || ToWidth == 32))
165  return true;
166 
167  // If this is a legal integer from type, and the result would be an illegal
168  // type, don't do the transformation.
169  if (FromLegal && !ToLegal)
170  return false;
171 
172  // Otherwise, if both are illegal, do not increase the size of the result. We
173  // do allow things like i160 -> i64, but not i64 -> i160.
174  if (!FromLegal && !ToLegal && ToWidth > FromWidth)
175  return false;
176 
177  return true;
178 }
179 
180 /// Return true if it is desirable to convert a computation from 'From' to 'To'.
181 /// We don't want to convert from a legal to an illegal type or from a smaller
182 /// to a larger illegal type. i1 is always treated as a legal type because it is
183 /// a fundamental type in IR, and there are many specialized optimizations for
184 /// i1 types.
185 bool InstCombiner::shouldChangeType(Type *From, Type *To) const {
186  assert(From->isIntegerTy() && To->isIntegerTy());
187 
188  unsigned FromWidth = From->getPrimitiveSizeInBits();
189  unsigned ToWidth = To->getPrimitiveSizeInBits();
190  return shouldChangeType(FromWidth, ToWidth);
191 }
192 
193 // Return true, if No Signed Wrap should be maintained for I.
194 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
195 // where both B and C should be ConstantInts, results in a constant that does
196 // not overflow. This function only handles the Add and Sub opcodes. For
197 // all other opcodes, the function conservatively returns false.
200  if (!OBO || !OBO->hasNoSignedWrap())
201  return false;
202 
203  // We reason about Add and Sub Only.
204  Instruction::BinaryOps Opcode = I.getOpcode();
205  if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
206  return false;
207 
208  const APInt *BVal, *CVal;
209  if (!match(B, m_APInt(BVal)) || !match(C, m_APInt(CVal)))
210  return false;
211 
212  bool Overflow = false;
213  if (Opcode == Instruction::Add)
214  (void)BVal->sadd_ov(*CVal, Overflow);
215  else
216  (void)BVal->ssub_ov(*CVal, Overflow);
217 
218  return !Overflow;
219 }
220 
221 /// Conservatively clears subclassOptionalData after a reassociation or
222 /// commutation. We preserve fast-math flags when applicable as they can be
223 /// preserved.
226  if (!FPMO) {
228  return;
229  }
230 
233  I.setFastMathFlags(FMF);
234 }
235 
236 /// Combine constant operands of associative operations either before or after a
237 /// cast to eliminate one of the associative operations:
238 /// (op (cast (op X, C2)), C1) --> (cast (op X, op (C1, C2)))
239 /// (op (cast (op X, C2)), C1) --> (op (cast X), op (C1, C2))
241  auto *Cast = dyn_cast<CastInst>(BinOp1->getOperand(0));
242  if (!Cast || !Cast->hasOneUse())
243  return false;
244 
245  // TODO: Enhance logic for other casts and remove this check.
246  auto CastOpcode = Cast->getOpcode();
247  if (CastOpcode != Instruction::ZExt)
248  return false;
249 
250  // TODO: Enhance logic for other BinOps and remove this check.
251  if (!BinOp1->isBitwiseLogicOp())
252  return false;
253 
254  auto AssocOpcode = BinOp1->getOpcode();
255  auto *BinOp2 = dyn_cast<BinaryOperator>(Cast->getOperand(0));
256  if (!BinOp2 || !BinOp2->hasOneUse() || BinOp2->getOpcode() != AssocOpcode)
257  return false;
258 
259  Constant *C1, *C2;
260  if (!match(BinOp1->getOperand(1), m_Constant(C1)) ||
261  !match(BinOp2->getOperand(1), m_Constant(C2)))
262  return false;
263 
264  // TODO: This assumes a zext cast.
265  // Eg, if it was a trunc, we'd cast C1 to the source type because casting C2
266  // to the destination type might lose bits.
267 
268  // Fold the constants together in the destination type:
269  // (op (cast (op X, C2)), C1) --> (op (cast X), FoldedC)
270  Type *DestTy = C1->getType();
271  Constant *CastC2 = ConstantExpr::getCast(CastOpcode, C2, DestTy);
272  Constant *FoldedC = ConstantExpr::get(AssocOpcode, C1, CastC2);
273  Cast->setOperand(0, BinOp2->getOperand(0));
274  BinOp1->setOperand(1, FoldedC);
275  return true;
276 }
277 
278 /// This performs a few simplifications for operators that are associative or
279 /// commutative:
280 ///
281 /// Commutative operators:
282 ///
283 /// 1. Order operands such that they are listed from right (least complex) to
284 /// left (most complex). This puts constants before unary operators before
285 /// binary operators.
286 ///
287 /// Associative operators:
288 ///
289 /// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
290 /// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
291 ///
292 /// Associative and commutative operators:
293 ///
294 /// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
295 /// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
296 /// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
297 /// if C1 and C2 are constants.
298 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
299  Instruction::BinaryOps Opcode = I.getOpcode();
300  bool Changed = false;
301 
302  do {
303  // Order operands such that they are listed from right (least complex) to
304  // left (most complex). This puts constants before unary operators before
305  // binary operators.
306  if (I.isCommutative() && getComplexity(I.getOperand(0)) <
308  Changed = !I.swapOperands();
309 
312 
313  if (I.isAssociative()) {
314  // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
315  if (Op0 && Op0->getOpcode() == Opcode) {
316  Value *A = Op0->getOperand(0);
317  Value *B = Op0->getOperand(1);
318  Value *C = I.getOperand(1);
319 
320  // Does "B op C" simplify?
321  if (Value *V = SimplifyBinOp(Opcode, B, C, SQ.getWithInstruction(&I))) {
322  // It simplifies to V. Form "A op V".
323  I.setOperand(0, A);
324  I.setOperand(1, V);
325  // Conservatively clear the optional flags, since they may not be
326  // preserved by the reassociation.
327  if (MaintainNoSignedWrap(I, B, C) &&
328  (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
329  // Note: this is only valid because SimplifyBinOp doesn't look at
330  // the operands to Op0.
332  I.setHasNoSignedWrap(true);
333  } else {
335  }
336 
337  Changed = true;
338  ++NumReassoc;
339  continue;
340  }
341  }
342 
343  // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
344  if (Op1 && Op1->getOpcode() == Opcode) {
345  Value *A = I.getOperand(0);
346  Value *B = Op1->getOperand(0);
347  Value *C = Op1->getOperand(1);
348 
349  // Does "A op B" simplify?
350  if (Value *V = SimplifyBinOp(Opcode, A, B, SQ.getWithInstruction(&I))) {
351  // It simplifies to V. Form "V op C".
352  I.setOperand(0, V);
353  I.setOperand(1, C);
354  // Conservatively clear the optional flags, since they may not be
355  // preserved by the reassociation.
357  Changed = true;
358  ++NumReassoc;
359  continue;
360  }
361  }
362  }
363 
364  if (I.isAssociative() && I.isCommutative()) {
365  if (simplifyAssocCastAssoc(&I)) {
366  Changed = true;
367  ++NumReassoc;
368  continue;
369  }
370 
371  // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
372  if (Op0 && Op0->getOpcode() == Opcode) {
373  Value *A = Op0->getOperand(0);
374  Value *B = Op0->getOperand(1);
375  Value *C = I.getOperand(1);
376 
377  // Does "C op A" simplify?
378  if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
379  // It simplifies to V. Form "V op B".
380  I.setOperand(0, V);
381  I.setOperand(1, B);
382  // Conservatively clear the optional flags, since they may not be
383  // preserved by the reassociation.
385  Changed = true;
386  ++NumReassoc;
387  continue;
388  }
389  }
390 
391  // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
392  if (Op1 && Op1->getOpcode() == Opcode) {
393  Value *A = I.getOperand(0);
394  Value *B = Op1->getOperand(0);
395  Value *C = Op1->getOperand(1);
396 
397  // Does "C op A" simplify?
398  if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
399  // It simplifies to V. Form "B op V".
400  I.setOperand(0, B);
401  I.setOperand(1, V);
402  // Conservatively clear the optional flags, since they may not be
403  // preserved by the reassociation.
405  Changed = true;
406  ++NumReassoc;
407  continue;
408  }
409  }
410 
411  // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
412  // if C1 and C2 are constants.
413  Value *A, *B;
414  Constant *C1, *C2;
415  if (Op0 && Op1 &&
416  Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
417  match(Op0, m_OneUse(m_BinOp(m_Value(A), m_Constant(C1)))) &&
418  match(Op1, m_OneUse(m_BinOp(m_Value(B), m_Constant(C2))))) {
419  BinaryOperator *NewBO = BinaryOperator::Create(Opcode, A, B);
420  if (isa<FPMathOperator>(NewBO)) {
421  FastMathFlags Flags = I.getFastMathFlags();
422  Flags &= Op0->getFastMathFlags();
423  Flags &= Op1->getFastMathFlags();
424  NewBO->setFastMathFlags(Flags);
425  }
426  InsertNewInstWith(NewBO, I);
427  NewBO->takeName(Op1);
428  I.setOperand(0, NewBO);
429  I.setOperand(1, ConstantExpr::get(Opcode, C1, C2));
430  // Conservatively clear the optional flags, since they may not be
431  // preserved by the reassociation.
433 
434  Changed = true;
435  continue;
436  }
437  }
438 
439  // No further simplifications.
440  return Changed;
441  } while (true);
442 }
443 
444 /// Return whether "X LOp (Y ROp Z)" is always equal to
445 /// "(X LOp Y) ROp (X LOp Z)".
448  // X & (Y | Z) <--> (X & Y) | (X & Z)
449  // X & (Y ^ Z) <--> (X & Y) ^ (X & Z)
450  if (LOp == Instruction::And)
451  return ROp == Instruction::Or || ROp == Instruction::Xor;
452 
453  // X | (Y & Z) <--> (X | Y) & (X | Z)
454  if (LOp == Instruction::Or)
455  return ROp == Instruction::And;
456 
457  // X * (Y + Z) <--> (X * Y) + (X * Z)
458  // X * (Y - Z) <--> (X * Y) - (X * Z)
459  if (LOp == Instruction::Mul)
460  return ROp == Instruction::Add || ROp == Instruction::Sub;
461 
462  return false;
463 }
464 
465 /// Return whether "(X LOp Y) ROp Z" is always equal to
466 /// "(X ROp Z) LOp (Y ROp Z)".
470  return leftDistributesOverRight(ROp, LOp);
471 
472  // (X {&|^} Y) >> Z <--> (X >> Z) {&|^} (Y >> Z) for all shifts.
474 
475  // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
476  // but this requires knowing that the addition does not overflow and other
477  // such subtleties.
478 }
479 
480 /// This function returns identity value for given opcode, which can be used to
481 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
483  if (isa<Constant>(V))
484  return nullptr;
485 
486  return ConstantExpr::getBinOpIdentity(Opcode, V->getType());
487 }
488 
489 /// This function predicates factorization using distributive laws. By default,
490 /// it just returns the 'Op' inputs. But for special-cases like
491 /// 'add(shl(X, 5), ...)', this function will have TopOpcode == Instruction::Add
492 /// and Op = shl(X, 5). The 'shl' is treated as the more general 'mul X, 32' to
493 /// allow more factorization opportunities.
496  Value *&LHS, Value *&RHS) {
497  assert(Op && "Expected a binary operator");
498  LHS = Op->getOperand(0);
499  RHS = Op->getOperand(1);
500  if (TopOpcode == Instruction::Add || TopOpcode == Instruction::Sub) {
501  Constant *C;
502  if (match(Op, m_Shl(m_Value(), m_Constant(C)))) {
503  // X << C --> X * (1 << C)
504  RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), C);
505  return Instruction::Mul;
506  }
507  // TODO: We can add other conversions e.g. shr => div etc.
508  }
509  return Op->getOpcode();
510 }
511 
512 /// This tries to simplify binary operations by factorizing out common terms
513 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
514 Value *InstCombiner::tryFactorization(BinaryOperator &I,
515  Instruction::BinaryOps InnerOpcode,
516  Value *A, Value *B, Value *C, Value *D) {
517  assert(A && B && C && D && "All values must be provided");
518 
519  Value *V = nullptr;
520  Value *SimplifiedInst = nullptr;
521  Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
522  Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
523 
524  // Does "X op' Y" always equal "Y op' X"?
525  bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
526 
527  // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
528  if (leftDistributesOverRight(InnerOpcode, TopLevelOpcode))
529  // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
530  // commutative case, "(A op' B) op (C op' A)"?
531  if (A == C || (InnerCommutative && A == D)) {
532  if (A != C)
533  std::swap(C, D);
534  // Consider forming "A op' (B op D)".
535  // If "B op D" simplifies then it can be formed with no cost.
536  V = SimplifyBinOp(TopLevelOpcode, B, D, SQ.getWithInstruction(&I));
537  // If "B op D" doesn't simplify then only go on if both of the existing
538  // operations "A op' B" and "C op' D" will be zapped as no longer used.
539  if (!V && LHS->hasOneUse() && RHS->hasOneUse())
540  V = Builder.CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
541  if (V) {
542  SimplifiedInst = Builder.CreateBinOp(InnerOpcode, A, V);
543  }
544  }
545 
546  // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
547  if (!SimplifiedInst && rightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
548  // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
549  // commutative case, "(A op' B) op (B op' D)"?
550  if (B == D || (InnerCommutative && B == C)) {
551  if (B != D)
552  std::swap(C, D);
553  // Consider forming "(A op C) op' B".
554  // If "A op C" simplifies then it can be formed with no cost.
555  V = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
556 
557  // If "A op C" doesn't simplify then only go on if both of the existing
558  // operations "A op' B" and "C op' D" will be zapped as no longer used.
559  if (!V && LHS->hasOneUse() && RHS->hasOneUse())
560  V = Builder.CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
561  if (V) {
562  SimplifiedInst = Builder.CreateBinOp(InnerOpcode, V, B);
563  }
564  }
565 
566  if (SimplifiedInst) {
567  ++NumFactor;
568  SimplifiedInst->takeName(&I);
569 
570  // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
571  // TODO: Check for NUW.
572  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
573  if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
574  bool HasNSW = false;
575  if (isa<OverflowingBinaryOperator>(&I))
576  HasNSW = I.hasNoSignedWrap();
577 
578  if (auto *LOBO = dyn_cast<OverflowingBinaryOperator>(LHS))
579  HasNSW &= LOBO->hasNoSignedWrap();
580 
581  if (auto *ROBO = dyn_cast<OverflowingBinaryOperator>(RHS))
582  HasNSW &= ROBO->hasNoSignedWrap();
583 
584  // We can propagate 'nsw' if we know that
585  // %Y = mul nsw i16 %X, C
586  // %Z = add nsw i16 %Y, %X
587  // =>
588  // %Z = mul nsw i16 %X, C+1
589  //
590  // iff C+1 isn't INT_MIN
591  const APInt *CInt;
592  if (TopLevelOpcode == Instruction::Add &&
593  InnerOpcode == Instruction::Mul)
594  if (match(V, m_APInt(CInt)) && !CInt->isMinSignedValue())
595  BO->setHasNoSignedWrap(HasNSW);
596  }
597  }
598  }
599  return SimplifiedInst;
600 }
601 
602 /// This tries to simplify binary operations which some other binary operation
603 /// distributes over either by factorizing out common terms
604 /// (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this results in
605 /// simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is a win).
606 /// Returns the simplified value, or null if it didn't simplify.
607 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
608  Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
611  Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
612 
613  {
614  // Factorization.
615  Value *A, *B, *C, *D;
616  Instruction::BinaryOps LHSOpcode, RHSOpcode;
617  if (Op0)
618  LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
619  if (Op1)
620  RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
621 
622  // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
623  // a common term.
624  if (Op0 && Op1 && LHSOpcode == RHSOpcode)
625  if (Value *V = tryFactorization(I, LHSOpcode, A, B, C, D))
626  return V;
627 
628  // The instruction has the form "(A op' B) op (C)". Try to factorize common
629  // term.
630  if (Op0)
631  if (Value *Ident = getIdentityValue(LHSOpcode, RHS))
632  if (Value *V = tryFactorization(I, LHSOpcode, A, B, RHS, Ident))
633  return V;
634 
635  // The instruction has the form "(B) op (C op' D)". Try to factorize common
636  // term.
637  if (Op1)
638  if (Value *Ident = getIdentityValue(RHSOpcode, LHS))
639  if (Value *V = tryFactorization(I, RHSOpcode, LHS, Ident, C, D))
640  return V;
641  }
642 
643  // Expansion.
644  if (Op0 && rightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
645  // The instruction has the form "(A op' B) op C". See if expanding it out
646  // to "(A op C) op' (B op C)" results in simplifications.
647  Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
648  Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
649 
650  Value *L = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
651  Value *R = SimplifyBinOp(TopLevelOpcode, B, C, SQ.getWithInstruction(&I));
652 
653  // Do "A op C" and "B op C" both simplify?
654  if (L && R) {
655  // They do! Return "L op' R".
656  ++NumExpand;
657  C = Builder.CreateBinOp(InnerOpcode, L, R);
658  C->takeName(&I);
659  return C;
660  }
661 
662  // Does "A op C" simplify to the identity value for the inner opcode?
663  if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
664  // They do! Return "B op C".
665  ++NumExpand;
666  C = Builder.CreateBinOp(TopLevelOpcode, B, C);
667  C->takeName(&I);
668  return C;
669  }
670 
671  // Does "B op C" simplify to the identity value for the inner opcode?
672  if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
673  // They do! Return "A op C".
674  ++NumExpand;
675  C = Builder.CreateBinOp(TopLevelOpcode, A, C);
676  C->takeName(&I);
677  return C;
678  }
679  }
680 
681  if (Op1 && leftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
682  // The instruction has the form "A op (B op' C)". See if expanding it out
683  // to "(A op B) op' (A op C)" results in simplifications.
684  Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
685  Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
686 
687  Value *L = SimplifyBinOp(TopLevelOpcode, A, B, SQ.getWithInstruction(&I));
688  Value *R = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
689 
690  // Do "A op B" and "A op C" both simplify?
691  if (L && R) {
692  // They do! Return "L op' R".
693  ++NumExpand;
694  A = Builder.CreateBinOp(InnerOpcode, L, R);
695  A->takeName(&I);
696  return A;
697  }
698 
699  // Does "A op B" simplify to the identity value for the inner opcode?
700  if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
701  // They do! Return "A op C".
702  ++NumExpand;
703  A = Builder.CreateBinOp(TopLevelOpcode, A, C);
704  A->takeName(&I);
705  return A;
706  }
707 
708  // Does "A op C" simplify to the identity value for the inner opcode?
709  if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
710  // They do! Return "A op B".
711  ++NumExpand;
712  A = Builder.CreateBinOp(TopLevelOpcode, A, B);
713  A->takeName(&I);
714  return A;
715  }
716  }
717 
718  return SimplifySelectsFeedingBinaryOp(I, LHS, RHS);
719 }
720 
721 Value *InstCombiner::SimplifySelectsFeedingBinaryOp(BinaryOperator &I,
722  Value *LHS, Value *RHS) {
723  Instruction::BinaryOps Opcode = I.getOpcode();
724  // (op (select (a, b, c)), (select (a, d, e))) -> (select (a, (op b, d), (op
725  // c, e)))
726  Value *A, *B, *C, *D, *E;
727  Value *SI = nullptr;
728  if (match(LHS, m_Select(m_Value(A), m_Value(B), m_Value(C))) &&
729  match(RHS, m_Select(m_Specific(A), m_Value(D), m_Value(E)))) {
730  bool SelectsHaveOneUse = LHS->hasOneUse() && RHS->hasOneUse();
731  BuilderTy::FastMathFlagGuard Guard(Builder);
732  if (isa<FPMathOperator>(&I))
733  Builder.setFastMathFlags(I.getFastMathFlags());
734 
735  Value *V1 = SimplifyBinOp(Opcode, C, E, SQ.getWithInstruction(&I));
736  Value *V2 = SimplifyBinOp(Opcode, B, D, SQ.getWithInstruction(&I));
737  if (V1 && V2)
738  SI = Builder.CreateSelect(A, V2, V1);
739  else if (V2 && SelectsHaveOneUse)
740  SI = Builder.CreateSelect(A, V2, Builder.CreateBinOp(Opcode, C, E));
741  else if (V1 && SelectsHaveOneUse)
742  SI = Builder.CreateSelect(A, Builder.CreateBinOp(Opcode, B, D), V1);
743 
744  if (SI)
745  SI->takeName(&I);
746  }
747 
748  return SI;
749 }
750 
751 /// Given a 'sub' instruction, return the RHS of the instruction if the LHS is a
752 /// constant zero (which is the 'negate' form).
753 Value *InstCombiner::dyn_castNegVal(Value *V) const {
754  Value *NegV;
755  if (match(V, m_Neg(m_Value(NegV))))
756  return NegV;
757 
758  // Constants can be considered to be negated values if they can be folded.
759  if (ConstantInt *C = dyn_cast<ConstantInt>(V))
760  return ConstantExpr::getNeg(C);
761 
762  if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
763  if (C->getType()->getElementType()->isIntegerTy())
764  return ConstantExpr::getNeg(C);
765 
766  if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
767  for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
768  Constant *Elt = CV->getAggregateElement(i);
769  if (!Elt)
770  return nullptr;
771 
772  if (isa<UndefValue>(Elt))
773  continue;
774 
775  if (!isa<ConstantInt>(Elt))
776  return nullptr;
777  }
778  return ConstantExpr::getNeg(CV);
779  }
780 
781  return nullptr;
782 }
783 
785  InstCombiner::BuilderTy &Builder) {
786  if (auto *Cast = dyn_cast<CastInst>(&I))
787  return Builder.CreateCast(Cast->getOpcode(), SO, I.getType());
788 
789  assert(I.isBinaryOp() && "Unexpected opcode for select folding");
790 
791  // Figure out if the constant is the left or the right argument.
792  bool ConstIsRHS = isa<Constant>(I.getOperand(1));
793  Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
794 
795  if (auto *SOC = dyn_cast<Constant>(SO)) {
796  if (ConstIsRHS)
797  return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
798  return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
799  }
800 
801  Value *Op0 = SO, *Op1 = ConstOperand;
802  if (!ConstIsRHS)
803  std::swap(Op0, Op1);
804 
805  auto *BO = cast<BinaryOperator>(&I);
806  Value *RI = Builder.CreateBinOp(BO->getOpcode(), Op0, Op1,
807  SO->getName() + ".op");
808  auto *FPInst = dyn_cast<Instruction>(RI);
809  if (FPInst && isa<FPMathOperator>(FPInst))
810  FPInst->copyFastMathFlags(BO);
811  return RI;
812 }
813 
814 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
815  // Don't modify shared select instructions.
816  if (!SI->hasOneUse())
817  return nullptr;
818 
819  Value *TV = SI->getTrueValue();
820  Value *FV = SI->getFalseValue();
821  if (!(isa<Constant>(TV) || isa<Constant>(FV)))
822  return nullptr;
823 
824  // Bool selects with constant operands can be folded to logical ops.
825  if (SI->getType()->isIntOrIntVectorTy(1))
826  return nullptr;
827 
828  // If it's a bitcast involving vectors, make sure it has the same number of
829  // elements on both sides.
830  if (auto *BC = dyn_cast<BitCastInst>(&Op)) {
831  VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
832  VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
833 
834  // Verify that either both or neither are vectors.
835  if ((SrcTy == nullptr) != (DestTy == nullptr))
836  return nullptr;
837 
838  // If vectors, verify that they have the same number of elements.
839  if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
840  return nullptr;
841  }
842 
843  // Test if a CmpInst instruction is used exclusively by a select as
844  // part of a minimum or maximum operation. If so, refrain from doing
845  // any other folding. This helps out other analyses which understand
846  // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
847  // and CodeGen. And in this case, at least one of the comparison
848  // operands has at least one user besides the compare (the select),
849  // which would often largely negate the benefit of folding anyway.
850  if (auto *CI = dyn_cast<CmpInst>(SI->getCondition())) {
851  if (CI->hasOneUse()) {
852  Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
853  if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
854  (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
855  return nullptr;
856  }
857  }
858 
859  Value *NewTV = foldOperationIntoSelectOperand(Op, TV, Builder);
860  Value *NewFV = foldOperationIntoSelectOperand(Op, FV, Builder);
861  return SelectInst::Create(SI->getCondition(), NewTV, NewFV, "", nullptr, SI);
862 }
863 
865  InstCombiner::BuilderTy &Builder) {
866  bool ConstIsRHS = isa<Constant>(I->getOperand(1));
867  Constant *C = cast<Constant>(I->getOperand(ConstIsRHS));
868 
869  if (auto *InC = dyn_cast<Constant>(InV)) {
870  if (ConstIsRHS)
871  return ConstantExpr::get(I->getOpcode(), InC, C);
872  return ConstantExpr::get(I->getOpcode(), C, InC);
873  }
874 
875  Value *Op0 = InV, *Op1 = C;
876  if (!ConstIsRHS)
877  std::swap(Op0, Op1);
878 
879  Value *RI = Builder.CreateBinOp(I->getOpcode(), Op0, Op1, "phitmp");
880  auto *FPInst = dyn_cast<Instruction>(RI);
881  if (FPInst && isa<FPMathOperator>(FPInst))
882  FPInst->copyFastMathFlags(I);
883  return RI;
884 }
885 
886 Instruction *InstCombiner::foldOpIntoPhi(Instruction &I, PHINode *PN) {
887  unsigned NumPHIValues = PN->getNumIncomingValues();
888  if (NumPHIValues == 0)
889  return nullptr;
890 
891  // We normally only transform phis with a single use. However, if a PHI has
892  // multiple uses and they are all the same operation, we can fold *all* of the
893  // uses into the PHI.
894  if (!PN->hasOneUse()) {
895  // Walk the use list for the instruction, comparing them to I.
896  for (User *U : PN->users()) {
897  Instruction *UI = cast<Instruction>(U);
898  if (UI != &I && !I.isIdenticalTo(UI))
899  return nullptr;
900  }
901  // Otherwise, we can replace *all* users with the new PHI we form.
902  }
903 
904  // Check to see if all of the operands of the PHI are simple constants
905  // (constantint/constantfp/undef). If there is one non-constant value,
906  // remember the BB it is in. If there is more than one or if *it* is a PHI,
907  // bail out. We don't do arbitrary constant expressions here because moving
908  // their computation can be expensive without a cost model.
909  BasicBlock *NonConstBB = nullptr;
910  for (unsigned i = 0; i != NumPHIValues; ++i) {
911  Value *InVal = PN->getIncomingValue(i);
912  if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
913  continue;
914 
915  if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
916  if (NonConstBB) return nullptr; // More than one non-const value.
917 
918  NonConstBB = PN->getIncomingBlock(i);
919 
920  // If the InVal is an invoke at the end of the pred block, then we can't
921  // insert a computation after it without breaking the edge.
922  if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
923  if (II->getParent() == NonConstBB)
924  return nullptr;
925 
926  // If the incoming non-constant value is in I's block, we will remove one
927  // instruction, but insert another equivalent one, leading to infinite
928  // instcombine.
929  if (isPotentiallyReachable(I.getParent(), NonConstBB, &DT, LI))
930  return nullptr;
931  }
932 
933  // If there is exactly one non-constant value, we can insert a copy of the
934  // operation in that block. However, if this is a critical edge, we would be
935  // inserting the computation on some other paths (e.g. inside a loop). Only
936  // do this if the pred block is unconditionally branching into the phi block.
937  if (NonConstBB != nullptr) {
938  BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
939  if (!BI || !BI->isUnconditional()) return nullptr;
940  }
941 
942  // Okay, we can do the transformation: create the new PHI node.
944  InsertNewInstBefore(NewPN, *PN);
945  NewPN->takeName(PN);
946 
947  // If we are going to have to insert a new computation, do so right before the
948  // predecessor's terminator.
949  if (NonConstBB)
950  Builder.SetInsertPoint(NonConstBB->getTerminator());
951 
952  // Next, add all of the operands to the PHI.
953  if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
954  // We only currently try to fold the condition of a select when it is a phi,
955  // not the true/false values.
956  Value *TrueV = SI->getTrueValue();
957  Value *FalseV = SI->getFalseValue();
958  BasicBlock *PhiTransBB = PN->getParent();
959  for (unsigned i = 0; i != NumPHIValues; ++i) {
960  BasicBlock *ThisBB = PN->getIncomingBlock(i);
961  Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
962  Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
963  Value *InV = nullptr;
964  // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
965  // even if currently isNullValue gives false.
966  Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
967  // For vector constants, we cannot use isNullValue to fold into
968  // FalseVInPred versus TrueVInPred. When we have individual nonzero
969  // elements in the vector, we will incorrectly fold InC to
970  // `TrueVInPred`.
971  if (InC && !isa<ConstantExpr>(InC) && isa<ConstantInt>(InC))
972  InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
973  else {
974  // Generate the select in the same block as PN's current incoming block.
975  // Note: ThisBB need not be the NonConstBB because vector constants
976  // which are constants by definition are handled here.
977  // FIXME: This can lead to an increase in IR generation because we might
978  // generate selects for vector constant phi operand, that could not be
979  // folded to TrueVInPred or FalseVInPred as done for ConstantInt. For
980  // non-vector phis, this transformation was always profitable because
981  // the select would be generated exactly once in the NonConstBB.
982  Builder.SetInsertPoint(ThisBB->getTerminator());
983  InV = Builder.CreateSelect(PN->getIncomingValue(i), TrueVInPred,
984  FalseVInPred, "phitmp");
985  }
986  NewPN->addIncoming(InV, ThisBB);
987  }
988  } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
989  Constant *C = cast<Constant>(I.getOperand(1));
990  for (unsigned i = 0; i != NumPHIValues; ++i) {
991  Value *InV = nullptr;
992  if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
993  InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
994  else if (isa<ICmpInst>(CI))
995  InV = Builder.CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
996  C, "phitmp");
997  else
998  InV = Builder.CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
999  C, "phitmp");
1000  NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1001  }
1002  } else if (auto *BO = dyn_cast<BinaryOperator>(&I)) {
1003  for (unsigned i = 0; i != NumPHIValues; ++i) {
1005  Builder);
1006  NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1007  }
1008  } else {
1009  CastInst *CI = cast<CastInst>(&I);
1010  Type *RetTy = CI->getType();
1011  for (unsigned i = 0; i != NumPHIValues; ++i) {
1012  Value *InV;
1013  if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
1014  InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1015  else
1016  InV = Builder.CreateCast(CI->getOpcode(), PN->getIncomingValue(i),
1017  I.getType(), "phitmp");
1018  NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1019  }
1020  }
1021 
1022  for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
1023  Instruction *User = cast<Instruction>(*UI++);
1024  if (User == &I) continue;
1025  replaceInstUsesWith(*User, NewPN);
1026  eraseInstFromFunction(*User);
1027  }
1028  return replaceInstUsesWith(I, NewPN);
1029 }
1030 
1031 Instruction *InstCombiner::foldBinOpIntoSelectOrPhi(BinaryOperator &I) {
1032  if (!isa<Constant>(I.getOperand(1)))
1033  return nullptr;
1034 
1035  if (auto *Sel = dyn_cast<SelectInst>(I.getOperand(0))) {
1036  if (Instruction *NewSel = FoldOpIntoSelect(I, Sel))
1037  return NewSel;
1038  } else if (auto *PN = dyn_cast<PHINode>(I.getOperand(0))) {
1039  if (Instruction *NewPhi = foldOpIntoPhi(I, PN))
1040  return NewPhi;
1041  }
1042  return nullptr;
1043 }
1044 
1045 /// Given a pointer type and a constant offset, determine whether or not there
1046 /// is a sequence of GEP indices into the pointed type that will land us at the
1047 /// specified offset. If so, fill them into NewIndices and return the resultant
1048 /// element type, otherwise return null.
1049 Type *InstCombiner::FindElementAtOffset(PointerType *PtrTy, int64_t Offset,
1050  SmallVectorImpl<Value *> &NewIndices) {
1051  Type *Ty = PtrTy->getElementType();
1052  if (!Ty->isSized())
1053  return nullptr;
1054 
1055  // Start with the index over the outer type. Note that the type size
1056  // might be zero (even if the offset isn't zero) if the indexed type
1057  // is something like [0 x {int, int}]
1058  Type *IndexTy = DL.getIndexType(PtrTy);
1059  int64_t FirstIdx = 0;
1060  if (int64_t TySize = DL.getTypeAllocSize(Ty)) {
1061  FirstIdx = Offset/TySize;
1062  Offset -= FirstIdx*TySize;
1063 
1064  // Handle hosts where % returns negative instead of values [0..TySize).
1065  if (Offset < 0) {
1066  --FirstIdx;
1067  Offset += TySize;
1068  assert(Offset >= 0);
1069  }
1070  assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
1071  }
1072 
1073  NewIndices.push_back(ConstantInt::get(IndexTy, FirstIdx));
1074 
1075  // Index into the types. If we fail, set OrigBase to null.
1076  while (Offset) {
1077  // Indexing into tail padding between struct/array elements.
1078  if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty))
1079  return nullptr;
1080 
1081  if (StructType *STy = dyn_cast<StructType>(Ty)) {
1082  const StructLayout *SL = DL.getStructLayout(STy);
1083  assert(Offset < (int64_t)SL->getSizeInBytes() &&
1084  "Offset must stay within the indexed type");
1085 
1086  unsigned Elt = SL->getElementContainingOffset(Offset);
1088  Elt));
1089 
1090  Offset -= SL->getElementOffset(Elt);
1091  Ty = STy->getElementType(Elt);
1092  } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
1093  uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType());
1094  assert(EltSize && "Cannot index into a zero-sized array");
1095  NewIndices.push_back(ConstantInt::get(IndexTy,Offset/EltSize));
1096  Offset %= EltSize;
1097  Ty = AT->getElementType();
1098  } else {
1099  // Otherwise, we can't index into the middle of this atomic type, bail.
1100  return nullptr;
1101  }
1102  }
1103 
1104  return Ty;
1105 }
1106 
1108  // If this GEP has only 0 indices, it is the same pointer as
1109  // Src. If Src is not a trivial GEP too, don't combine
1110  // the indices.
1111  if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
1112  !Src.hasOneUse())
1113  return false;
1114  return true;
1115 }
1116 
1117 /// Return a value X such that Val = X * Scale, or null if none.
1118 /// If the multiplication is known not to overflow, then NoSignedWrap is set.
1119 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
1120  assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
1121  assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
1122  Scale.getBitWidth() && "Scale not compatible with value!");
1123 
1124  // If Val is zero or Scale is one then Val = Val * Scale.
1125  if (match(Val, m_Zero()) || Scale == 1) {
1126  NoSignedWrap = true;
1127  return Val;
1128  }
1129 
1130  // If Scale is zero then it does not divide Val.
1131  if (Scale.isMinValue())
1132  return nullptr;
1133 
1134  // Look through chains of multiplications, searching for a constant that is
1135  // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
1136  // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
1137  // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
1138  // down from Val:
1139  //
1140  // Val = M1 * X || Analysis starts here and works down
1141  // M1 = M2 * Y || Doesn't descend into terms with more
1142  // M2 = Z * 4 \/ than one use
1143  //
1144  // Then to modify a term at the bottom:
1145  //
1146  // Val = M1 * X
1147  // M1 = Z * Y || Replaced M2 with Z
1148  //
1149  // Then to work back up correcting nsw flags.
1150 
1151  // Op - the term we are currently analyzing. Starts at Val then drills down.
1152  // Replaced with its descaled value before exiting from the drill down loop.
1153  Value *Op = Val;
1154 
1155  // Parent - initially null, but after drilling down notes where Op came from.
1156  // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1157  // 0'th operand of Val.
1158  std::pair<Instruction *, unsigned> Parent;
1159 
1160  // Set if the transform requires a descaling at deeper levels that doesn't
1161  // overflow.
1162  bool RequireNoSignedWrap = false;
1163 
1164  // Log base 2 of the scale. Negative if not a power of 2.
1165  int32_t logScale = Scale.exactLogBase2();
1166 
1167  for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1168  if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1169  // If Op is a constant divisible by Scale then descale to the quotient.
1170  APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1171  APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1172  if (!Remainder.isMinValue())
1173  // Not divisible by Scale.
1174  return nullptr;
1175  // Replace with the quotient in the parent.
1176  Op = ConstantInt::get(CI->getType(), Quotient);
1177  NoSignedWrap = true;
1178  break;
1179  }
1180 
1181  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1182  if (BO->getOpcode() == Instruction::Mul) {
1183  // Multiplication.
1184  NoSignedWrap = BO->hasNoSignedWrap();
1185  if (RequireNoSignedWrap && !NoSignedWrap)
1186  return nullptr;
1187 
1188  // There are three cases for multiplication: multiplication by exactly
1189  // the scale, multiplication by a constant different to the scale, and
1190  // multiplication by something else.
1191  Value *LHS = BO->getOperand(0);
1192  Value *RHS = BO->getOperand(1);
1193 
1194  if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1195  // Multiplication by a constant.
1196  if (CI->getValue() == Scale) {
1197  // Multiplication by exactly the scale, replace the multiplication
1198  // by its left-hand side in the parent.
1199  Op = LHS;
1200  break;
1201  }
1202 
1203  // Otherwise drill down into the constant.
1204  if (!Op->hasOneUse())
1205  return nullptr;
1206 
1207  Parent = std::make_pair(BO, 1);
1208  continue;
1209  }
1210 
1211  // Multiplication by something else. Drill down into the left-hand side
1212  // since that's where the reassociate pass puts the good stuff.
1213  if (!Op->hasOneUse())
1214  return nullptr;
1215 
1216  Parent = std::make_pair(BO, 0);
1217  continue;
1218  }
1219 
1220  if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1221  isa<ConstantInt>(BO->getOperand(1))) {
1222  // Multiplication by a power of 2.
1223  NoSignedWrap = BO->hasNoSignedWrap();
1224  if (RequireNoSignedWrap && !NoSignedWrap)
1225  return nullptr;
1226 
1227  Value *LHS = BO->getOperand(0);
1228  int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1229  getLimitedValue(Scale.getBitWidth());
1230  // Op = LHS << Amt.
1231 
1232  if (Amt == logScale) {
1233  // Multiplication by exactly the scale, replace the multiplication
1234  // by its left-hand side in the parent.
1235  Op = LHS;
1236  break;
1237  }
1238  if (Amt < logScale || !Op->hasOneUse())
1239  return nullptr;
1240 
1241  // Multiplication by more than the scale. Reduce the multiplying amount
1242  // by the scale in the parent.
1243  Parent = std::make_pair(BO, 1);
1244  Op = ConstantInt::get(BO->getType(), Amt - logScale);
1245  break;
1246  }
1247  }
1248 
1249  if (!Op->hasOneUse())
1250  return nullptr;
1251 
1252  if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1253  if (Cast->getOpcode() == Instruction::SExt) {
1254  // Op is sign-extended from a smaller type, descale in the smaller type.
1255  unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1256  APInt SmallScale = Scale.trunc(SmallSize);
1257  // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1258  // descale Op as (sext Y) * Scale. In order to have
1259  // sext (Y * SmallScale) = (sext Y) * Scale
1260  // some conditions need to hold however: SmallScale must sign-extend to
1261  // Scale and the multiplication Y * SmallScale should not overflow.
1262  if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1263  // SmallScale does not sign-extend to Scale.
1264  return nullptr;
1265  assert(SmallScale.exactLogBase2() == logScale);
1266  // Require that Y * SmallScale must not overflow.
1267  RequireNoSignedWrap = true;
1268 
1269  // Drill down through the cast.
1270  Parent = std::make_pair(Cast, 0);
1271  Scale = SmallScale;
1272  continue;
1273  }
1274 
1275  if (Cast->getOpcode() == Instruction::Trunc) {
1276  // Op is truncated from a larger type, descale in the larger type.
1277  // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1278  // trunc (Y * sext Scale) = (trunc Y) * Scale
1279  // always holds. However (trunc Y) * Scale may overflow even if
1280  // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1281  // from this point up in the expression (see later).
1282  if (RequireNoSignedWrap)
1283  return nullptr;
1284 
1285  // Drill down through the cast.
1286  unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1287  Parent = std::make_pair(Cast, 0);
1288  Scale = Scale.sext(LargeSize);
1289  if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1290  logScale = -1;
1291  assert(Scale.exactLogBase2() == logScale);
1292  continue;
1293  }
1294  }
1295 
1296  // Unsupported expression, bail out.
1297  return nullptr;
1298  }
1299 
1300  // If Op is zero then Val = Op * Scale.
1301  if (match(Op, m_Zero())) {
1302  NoSignedWrap = true;
1303  return Op;
1304  }
1305 
1306  // We know that we can successfully descale, so from here on we can safely
1307  // modify the IR. Op holds the descaled version of the deepest term in the
1308  // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1309  // not to overflow.
1310 
1311  if (!Parent.first)
1312  // The expression only had one term.
1313  return Op;
1314 
1315  // Rewrite the parent using the descaled version of its operand.
1316  assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1317  assert(Op != Parent.first->getOperand(Parent.second) &&
1318  "Descaling was a no-op?");
1319  Parent.first->setOperand(Parent.second, Op);
1320  Worklist.Add(Parent.first);
1321 
1322  // Now work back up the expression correcting nsw flags. The logic is based
1323  // on the following observation: if X * Y is known not to overflow as a signed
1324  // multiplication, and Y is replaced by a value Z with smaller absolute value,
1325  // then X * Z will not overflow as a signed multiplication either. As we work
1326  // our way up, having NoSignedWrap 'true' means that the descaled value at the
1327  // current level has strictly smaller absolute value than the original.
1328  Instruction *Ancestor = Parent.first;
1329  do {
1330  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1331  // If the multiplication wasn't nsw then we can't say anything about the
1332  // value of the descaled multiplication, and we have to clear nsw flags
1333  // from this point on up.
1334  bool OpNoSignedWrap = BO->hasNoSignedWrap();
1335  NoSignedWrap &= OpNoSignedWrap;
1336  if (NoSignedWrap != OpNoSignedWrap) {
1337  BO->setHasNoSignedWrap(NoSignedWrap);
1338  Worklist.Add(Ancestor);
1339  }
1340  } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1341  // The fact that the descaled input to the trunc has smaller absolute
1342  // value than the original input doesn't tell us anything useful about
1343  // the absolute values of the truncations.
1344  NoSignedWrap = false;
1345  }
1346  assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1347  "Failed to keep proper track of nsw flags while drilling down?");
1348 
1349  if (Ancestor == Val)
1350  // Got to the top, all done!
1351  return Val;
1352 
1353  // Move up one level in the expression.
1354  assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1355  Ancestor = Ancestor->user_back();
1356  } while (true);
1357 }
1358 
1359 Instruction *InstCombiner::foldVectorBinop(BinaryOperator &Inst) {
1360  if (!Inst.getType()->isVectorTy()) return nullptr;
1361 
1362  BinaryOperator::BinaryOps Opcode = Inst.getOpcode();
1363  unsigned NumElts = cast<VectorType>(Inst.getType())->getNumElements();
1364  Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1365  assert(cast<VectorType>(LHS->getType())->getNumElements() == NumElts);
1366  assert(cast<VectorType>(RHS->getType())->getNumElements() == NumElts);
1367 
1368  // If both operands of the binop are vector concatenations, then perform the
1369  // narrow binop on each pair of the source operands followed by concatenation
1370  // of the results.
1371  Value *L0, *L1, *R0, *R1;
1372  Constant *Mask;
1373  if (match(LHS, m_ShuffleVector(m_Value(L0), m_Value(L1), m_Constant(Mask))) &&
1374  match(RHS, m_ShuffleVector(m_Value(R0), m_Value(R1), m_Specific(Mask))) &&
1375  LHS->hasOneUse() && RHS->hasOneUse() &&
1376  cast<ShuffleVectorInst>(LHS)->isConcat()) {
1377  // This transform does not have the speculative execution constraint as
1378  // below because the shuffle is a concatenation. The new binops are
1379  // operating on exactly the same elements as the existing binop.
1380  // TODO: We could ease the mask requirement to allow different undef lanes,
1381  // but that requires an analysis of the binop-with-undef output value.
1382  Value *NewBO0 = Builder.CreateBinOp(Opcode, L0, R0);
1383  if (auto *BO = dyn_cast<BinaryOperator>(NewBO0))
1384  BO->copyIRFlags(&Inst);
1385  Value *NewBO1 = Builder.CreateBinOp(Opcode, L1, R1);
1386  if (auto *BO = dyn_cast<BinaryOperator>(NewBO1))
1387  BO->copyIRFlags(&Inst);
1388  return new ShuffleVectorInst(NewBO0, NewBO1, Mask);
1389  }
1390 
1391  // It may not be safe to reorder shuffles and things like div, urem, etc.
1392  // because we may trap when executing those ops on unknown vector elements.
1393  // See PR20059.
1394  if (!isSafeToSpeculativelyExecute(&Inst))
1395  return nullptr;
1396 
1397  auto createBinOpShuffle = [&](Value *X, Value *Y, Constant *M) {
1398  Value *XY = Builder.CreateBinOp(Opcode, X, Y);
1399  if (auto *BO = dyn_cast<BinaryOperator>(XY))
1400  BO->copyIRFlags(&Inst);
1401  return new ShuffleVectorInst(XY, UndefValue::get(XY->getType()), M);
1402  };
1403 
1404  // If both arguments of the binary operation are shuffles that use the same
1405  // mask and shuffle within a single vector, move the shuffle after the binop.
1406  Value *V1, *V2;
1407  if (match(LHS, m_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(Mask))) &&
1408  match(RHS, m_ShuffleVector(m_Value(V2), m_Undef(), m_Specific(Mask))) &&
1409  V1->getType() == V2->getType() &&
1410  (LHS->hasOneUse() || RHS->hasOneUse() || LHS == RHS)) {
1411  // Op(shuffle(V1, Mask), shuffle(V2, Mask)) -> shuffle(Op(V1, V2), Mask)
1412  return createBinOpShuffle(V1, V2, Mask);
1413  }
1414 
1415  // If one argument is a shuffle within one vector and the other is a constant,
1416  // try moving the shuffle after the binary operation. This canonicalization
1417  // intends to move shuffles closer to other shuffles and binops closer to
1418  // other binops, so they can be folded. It may also enable demanded elements
1419  // transforms.
1420  Constant *C;
1421  if (match(&Inst, m_c_BinOp(
1423  m_Constant(C))) &&
1424  V1->getType()->getVectorNumElements() <= NumElts) {
1425  assert(Inst.getType()->getScalarType() == V1->getType()->getScalarType() &&
1426  "Shuffle should not change scalar type");
1427  unsigned V1Width = V1->getType()->getVectorNumElements();
1428  // Find constant NewC that has property:
1429  // shuffle(NewC, ShMask) = C
1430  // If such constant does not exist (example: ShMask=<0,0> and C=<1,2>)
1431  // reorder is not possible. A 1-to-1 mapping is not required. Example:
1432  // ShMask = <1,1,2,2> and C = <5,5,6,6> --> NewC = <undef,5,6,undef>
1433  SmallVector<int, 16> ShMask;
1434  ShuffleVectorInst::getShuffleMask(Mask, ShMask);
1436  NewVecC(V1Width, UndefValue::get(C->getType()->getScalarType()));
1437  bool MayChange = true;
1438  for (unsigned I = 0; I < NumElts; ++I) {
1439  if (ShMask[I] >= 0) {
1440  assert(ShMask[I] < (int)NumElts && "Not expecting narrowing shuffle");
1441  Constant *CElt = C->getAggregateElement(I);
1442  Constant *NewCElt = NewVecC[ShMask[I]];
1443  // Bail out if:
1444  // 1. The constant vector contains a constant expression.
1445  // 2. The shuffle needs an element of the constant vector that can't
1446  // be mapped to a new constant vector.
1447  // 3. This is a widening shuffle that copies elements of V1 into the
1448  // extended elements (extending with undef is allowed).
1449  if (!CElt || (!isa<UndefValue>(NewCElt) && NewCElt != CElt) ||
1450  I >= V1Width) {
1451  MayChange = false;
1452  break;
1453  }
1454  NewVecC[ShMask[I]] = CElt;
1455  }
1456  }
1457  if (MayChange) {
1458  Constant *NewC = ConstantVector::get(NewVecC);
1459  // It may not be safe to execute a binop on a vector with undef elements
1460  // because the entire instruction can be folded to undef or create poison
1461  // that did not exist in the original code.
1462  bool ConstOp1 = isa<Constant>(Inst.getOperand(1));
1463  if (Inst.isIntDivRem() || (Inst.isShift() && ConstOp1))
1464  NewC = getSafeVectorConstantForBinop(Opcode, NewC, ConstOp1);
1465 
1466  // Op(shuffle(V1, Mask), C) -> shuffle(Op(V1, NewC), Mask)
1467  // Op(C, shuffle(V1, Mask)) -> shuffle(Op(NewC, V1), Mask)
1468  Value *NewLHS = isa<Constant>(LHS) ? NewC : V1;
1469  Value *NewRHS = isa<Constant>(LHS) ? V1 : NewC;
1470  return createBinOpShuffle(NewLHS, NewRHS, Mask);
1471  }
1472  }
1473 
1474  return nullptr;
1475 }
1476 
1477 /// Try to narrow the width of a binop if at least 1 operand is an extend of
1478 /// of a value. This requires a potentially expensive known bits check to make
1479 /// sure the narrow op does not overflow.
1480 Instruction *InstCombiner::narrowMathIfNoOverflow(BinaryOperator &BO) {
1481  // We need at least one extended operand.
1482  Value *Op0 = BO.getOperand(0), *Op1 = BO.getOperand(1);
1483 
1484  // If this is a sub, we swap the operands since we always want an extension
1485  // on the RHS. The LHS can be an extension or a constant.
1486  if (BO.getOpcode() == Instruction::Sub)
1487  std::swap(Op0, Op1);
1488 
1489  Value *X;
1490  bool IsSext = match(Op0, m_SExt(m_Value(X)));
1491  if (!IsSext && !match(Op0, m_ZExt(m_Value(X))))
1492  return nullptr;
1493 
1494  // If both operands are the same extension from the same source type and we
1495  // can eliminate at least one (hasOneUse), this might work.
1496  CastInst::CastOps CastOpc = IsSext ? Instruction::SExt : Instruction::ZExt;
1497  Value *Y;
1498  if (!(match(Op1, m_ZExtOrSExt(m_Value(Y))) && X->getType() == Y->getType() &&
1499  cast<Operator>(Op1)->getOpcode() == CastOpc &&
1500  (Op0->hasOneUse() || Op1->hasOneUse()))) {
1501  // If that did not match, see if we have a suitable constant operand.
1502  // Truncating and extending must produce the same constant.
1503  Constant *WideC;
1504  if (!Op0->hasOneUse() || !match(Op1, m_Constant(WideC)))
1505  return nullptr;
1506  Constant *NarrowC = ConstantExpr::getTrunc(WideC, X->getType());
1507  if (ConstantExpr::getCast(CastOpc, NarrowC, BO.getType()) != WideC)
1508  return nullptr;
1509  Y = NarrowC;
1510  }
1511 
1512  // Swap back now that we found our operands.
1513  if (BO.getOpcode() == Instruction::Sub)
1514  std::swap(X, Y);
1515 
1516  // Both operands have narrow versions. Last step: the math must not overflow
1517  // in the narrow width.
1518  if (!willNotOverflow(BO.getOpcode(), X, Y, BO, IsSext))
1519  return nullptr;
1520 
1521  // bo (ext X), (ext Y) --> ext (bo X, Y)
1522  // bo (ext X), C --> ext (bo X, C')
1523  Value *NarrowBO = Builder.CreateBinOp(BO.getOpcode(), X, Y, "narrow");
1524  if (auto *NewBinOp = dyn_cast<BinaryOperator>(NarrowBO)) {
1525  if (IsSext)
1526  NewBinOp->setHasNoSignedWrap();
1527  else
1528  NewBinOp->setHasNoUnsignedWrap();
1529  }
1530  return CastInst::Create(CastOpc, NarrowBO, BO.getType());
1531 }
1532 
1534  SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1535  Type *GEPType = GEP.getType();
1536  Type *GEPEltType = GEP.getSourceElementType();
1537  if (Value *V = SimplifyGEPInst(GEPEltType, Ops, SQ.getWithInstruction(&GEP)))
1538  return replaceInstUsesWith(GEP, V);
1539 
1540  Value *PtrOp = GEP.getOperand(0);
1541 
1542  // Eliminate unneeded casts for indices, and replace indices which displace
1543  // by multiples of a zero size type with zero.
1544  bool MadeChange = false;
1545 
1546  // Index width may not be the same width as pointer width.
1547  // Data layout chooses the right type based on supported integer types.
1548  Type *NewScalarIndexTy =
1549  DL.getIndexType(GEP.getPointerOperandType()->getScalarType());
1550 
1551  gep_type_iterator GTI = gep_type_begin(GEP);
1552  for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E;
1553  ++I, ++GTI) {
1554  // Skip indices into struct types.
1555  if (GTI.isStruct())
1556  continue;
1557 
1558  Type *IndexTy = (*I)->getType();
1559  Type *NewIndexType =
1560  IndexTy->isVectorTy()
1561  ? VectorType::get(NewScalarIndexTy, IndexTy->getVectorNumElements())
1562  : NewScalarIndexTy;
1563 
1564  // If the element type has zero size then any index over it is equivalent
1565  // to an index of zero, so replace it with zero if it is not zero already.
1566  Type *EltTy = GTI.getIndexedType();
1567  if (EltTy->isSized() && DL.getTypeAllocSize(EltTy) == 0)
1568  if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1569  *I = Constant::getNullValue(NewIndexType);
1570  MadeChange = true;
1571  }
1572 
1573  if (IndexTy != NewIndexType) {
1574  // If we are using a wider index than needed for this platform, shrink
1575  // it to what we need. If narrower, sign-extend it to what we need.
1576  // This explicit cast can make subsequent optimizations more obvious.
1577  *I = Builder.CreateIntCast(*I, NewIndexType, true);
1578  MadeChange = true;
1579  }
1580  }
1581  if (MadeChange)
1582  return &GEP;
1583 
1584  // Check to see if the inputs to the PHI node are getelementptr instructions.
1585  if (auto *PN = dyn_cast<PHINode>(PtrOp)) {
1586  auto *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1587  if (!Op1)
1588  return nullptr;
1589 
1590  // Don't fold a GEP into itself through a PHI node. This can only happen
1591  // through the back-edge of a loop. Folding a GEP into itself means that
1592  // the value of the previous iteration needs to be stored in the meantime,
1593  // thus requiring an additional register variable to be live, but not
1594  // actually achieving anything (the GEP still needs to be executed once per
1595  // loop iteration).
1596  if (Op1 == &GEP)
1597  return nullptr;
1598 
1599  int DI = -1;
1600 
1601  for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1602  auto *Op2 = dyn_cast<GetElementPtrInst>(*I);
1603  if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1604  return nullptr;
1605 
1606  // As for Op1 above, don't try to fold a GEP into itself.
1607  if (Op2 == &GEP)
1608  return nullptr;
1609 
1610  // Keep track of the type as we walk the GEP.
1611  Type *CurTy = nullptr;
1612 
1613  for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1614  if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1615  return nullptr;
1616 
1617  if (Op1->getOperand(J) != Op2->getOperand(J)) {
1618  if (DI == -1) {
1619  // We have not seen any differences yet in the GEPs feeding the
1620  // PHI yet, so we record this one if it is allowed to be a
1621  // variable.
1622 
1623  // The first two arguments can vary for any GEP, the rest have to be
1624  // static for struct slots
1625  if (J > 1 && CurTy->isStructTy())
1626  return nullptr;
1627 
1628  DI = J;
1629  } else {
1630  // The GEP is different by more than one input. While this could be
1631  // extended to support GEPs that vary by more than one variable it
1632  // doesn't make sense since it greatly increases the complexity and
1633  // would result in an R+R+R addressing mode which no backend
1634  // directly supports and would need to be broken into several
1635  // simpler instructions anyway.
1636  return nullptr;
1637  }
1638  }
1639 
1640  // Sink down a layer of the type for the next iteration.
1641  if (J > 0) {
1642  if (J == 1) {
1643  CurTy = Op1->getSourceElementType();
1644  } else if (auto *CT = dyn_cast<CompositeType>(CurTy)) {
1645  CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1646  } else {
1647  CurTy = nullptr;
1648  }
1649  }
1650  }
1651  }
1652 
1653  // If not all GEPs are identical we'll have to create a new PHI node.
1654  // Check that the old PHI node has only one use so that it will get
1655  // removed.
1656  if (DI != -1 && !PN->hasOneUse())
1657  return nullptr;
1658 
1659  auto *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1660  if (DI == -1) {
1661  // All the GEPs feeding the PHI are identical. Clone one down into our
1662  // BB so that it can be merged with the current GEP.
1663  GEP.getParent()->getInstList().insert(
1664  GEP.getParent()->getFirstInsertionPt(), NewGEP);
1665  } else {
1666  // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1667  // into the current block so it can be merged, and create a new PHI to
1668  // set that index.
1669  PHINode *NewPN;
1670  {
1671  IRBuilderBase::InsertPointGuard Guard(Builder);
1672  Builder.SetInsertPoint(PN);
1673  NewPN = Builder.CreatePHI(Op1->getOperand(DI)->getType(),
1674  PN->getNumOperands());
1675  }
1676 
1677  for (auto &I : PN->operands())
1678  NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1679  PN->getIncomingBlock(I));
1680 
1681  NewGEP->setOperand(DI, NewPN);
1682  GEP.getParent()->getInstList().insert(
1683  GEP.getParent()->getFirstInsertionPt(), NewGEP);
1684  NewGEP->setOperand(DI, NewPN);
1685  }
1686 
1687  GEP.setOperand(0, NewGEP);
1688  PtrOp = NewGEP;
1689  }
1690 
1691  // Combine Indices - If the source pointer to this getelementptr instruction
1692  // is a getelementptr instruction, combine the indices of the two
1693  // getelementptr instructions into a single instruction.
1694  if (auto *Src = dyn_cast<GEPOperator>(PtrOp)) {
1695  if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1696  return nullptr;
1697 
1698  // Try to reassociate loop invariant GEP chains to enable LICM.
1699  if (LI && Src->getNumOperands() == 2 && GEP.getNumOperands() == 2 &&
1700  Src->hasOneUse()) {
1701  if (Loop *L = LI->getLoopFor(GEP.getParent())) {
1702  Value *GO1 = GEP.getOperand(1);
1703  Value *SO1 = Src->getOperand(1);
1704  // Reassociate the two GEPs if SO1 is variant in the loop and GO1 is
1705  // invariant: this breaks the dependence between GEPs and allows LICM
1706  // to hoist the invariant part out of the loop.
1707  if (L->isLoopInvariant(GO1) && !L->isLoopInvariant(SO1)) {
1708  // We have to be careful here.
1709  // We have something like:
1710  // %src = getelementptr <ty>, <ty>* %base, <ty> %idx
1711  // %gep = getelementptr <ty>, <ty>* %src, <ty> %idx2
1712  // If we just swap idx & idx2 then we could inadvertantly
1713  // change %src from a vector to a scalar, or vice versa.
1714  // Cases:
1715  // 1) %base a scalar & idx a scalar & idx2 a vector
1716  // => Swapping idx & idx2 turns %src into a vector type.
1717  // 2) %base a scalar & idx a vector & idx2 a scalar
1718  // => Swapping idx & idx2 turns %src in a scalar type
1719  // 3) %base, %idx, and %idx2 are scalars
1720  // => %src & %gep are scalars
1721  // => swapping idx & idx2 is safe
1722  // 4) %base a vector
1723  // => %src is a vector
1724  // => swapping idx & idx2 is safe.
1725  auto *SO0 = Src->getOperand(0);
1726  auto *SO0Ty = SO0->getType();
1727  if (!isa<VectorType>(GEPType) || // case 3
1728  isa<VectorType>(SO0Ty)) { // case 4
1729  Src->setOperand(1, GO1);
1730  GEP.setOperand(1, SO1);
1731  return &GEP;
1732  } else {
1733  // Case 1 or 2
1734  // -- have to recreate %src & %gep
1735  // put NewSrc at same location as %src
1736  Builder.SetInsertPoint(cast<Instruction>(PtrOp));
1737  auto *NewSrc = cast<GetElementPtrInst>(
1738  Builder.CreateGEP(SO0, GO1, Src->getName()));
1739  NewSrc->setIsInBounds(Src->isInBounds());
1740  auto *NewGEP = GetElementPtrInst::Create(nullptr, NewSrc, {SO1});
1741  NewGEP->setIsInBounds(GEP.isInBounds());
1742  return NewGEP;
1743  }
1744  }
1745  }
1746  }
1747 
1748  // Note that if our source is a gep chain itself then we wait for that
1749  // chain to be resolved before we perform this transformation. This
1750  // avoids us creating a TON of code in some cases.
1751  if (auto *SrcGEP = dyn_cast<GEPOperator>(Src->getOperand(0)))
1752  if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1753  return nullptr; // Wait until our source is folded to completion.
1754 
1755  SmallVector<Value*, 8> Indices;
1756 
1757  // Find out whether the last index in the source GEP is a sequential idx.
1758  bool EndsWithSequential = false;
1759  for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1760  I != E; ++I)
1761  EndsWithSequential = I.isSequential();
1762 
1763  // Can we combine the two pointer arithmetics offsets?
1764  if (EndsWithSequential) {
1765  // Replace: gep (gep %P, long B), long A, ...
1766  // With: T = long A+B; gep %P, T, ...
1767  Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1768  Value *GO1 = GEP.getOperand(1);
1769 
1770  // If they aren't the same type, then the input hasn't been processed
1771  // by the loop above yet (which canonicalizes sequential index types to
1772  // intptr_t). Just avoid transforming this until the input has been
1773  // normalized.
1774  if (SO1->getType() != GO1->getType())
1775  return nullptr;
1776 
1777  Value *Sum =
1778  SimplifyAddInst(GO1, SO1, false, false, SQ.getWithInstruction(&GEP));
1779  // Only do the combine when we are sure the cost after the
1780  // merge is never more than that before the merge.
1781  if (Sum == nullptr)
1782  return nullptr;
1783 
1784  // Update the GEP in place if possible.
1785  if (Src->getNumOperands() == 2) {
1786  GEP.setOperand(0, Src->getOperand(0));
1787  GEP.setOperand(1, Sum);
1788  return &GEP;
1789  }
1790  Indices.append(Src->op_begin()+1, Src->op_end()-1);
1791  Indices.push_back(Sum);
1792  Indices.append(GEP.op_begin()+2, GEP.op_end());
1793  } else if (isa<Constant>(*GEP.idx_begin()) &&
1794  cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1795  Src->getNumOperands() != 1) {
1796  // Otherwise we can do the fold if the first index of the GEP is a zero
1797  Indices.append(Src->op_begin()+1, Src->op_end());
1798  Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1799  }
1800 
1801  if (!Indices.empty())
1802  return GEP.isInBounds() && Src->isInBounds()
1804  Src->getSourceElementType(), Src->getOperand(0), Indices,
1805  GEP.getName())
1806  : GetElementPtrInst::Create(Src->getSourceElementType(),
1807  Src->getOperand(0), Indices,
1808  GEP.getName());
1809  }
1810 
1811  if (GEP.getNumIndices() == 1) {
1812  unsigned AS = GEP.getPointerAddressSpace();
1813  if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1814  DL.getIndexSizeInBits(AS)) {
1815  uint64_t TyAllocSize = DL.getTypeAllocSize(GEPEltType);
1816 
1817  bool Matched = false;
1818  uint64_t C;
1819  Value *V = nullptr;
1820  if (TyAllocSize == 1) {
1821  V = GEP.getOperand(1);
1822  Matched = true;
1823  } else if (match(GEP.getOperand(1),
1824  m_AShr(m_Value(V), m_ConstantInt(C)))) {
1825  if (TyAllocSize == 1ULL << C)
1826  Matched = true;
1827  } else if (match(GEP.getOperand(1),
1828  m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1829  if (TyAllocSize == C)
1830  Matched = true;
1831  }
1832 
1833  if (Matched) {
1834  // Canonicalize (gep i8* X, -(ptrtoint Y))
1835  // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1836  // The GEP pattern is emitted by the SCEV expander for certain kinds of
1837  // pointer arithmetic.
1838  if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1839  Operator *Index = cast<Operator>(V);
1840  Value *PtrToInt = Builder.CreatePtrToInt(PtrOp, Index->getType());
1841  Value *NewSub = Builder.CreateSub(PtrToInt, Index->getOperand(1));
1842  return CastInst::Create(Instruction::IntToPtr, NewSub, GEPType);
1843  }
1844  // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1845  // to (bitcast Y)
1846  Value *Y;
1847  if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1848  m_PtrToInt(m_Specific(GEP.getOperand(0))))))
1850  }
1851  }
1852  }
1853 
1854  // We do not handle pointer-vector geps here.
1855  if (GEPType->isVectorTy())
1856  return nullptr;
1857 
1858  // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1859  Value *StrippedPtr = PtrOp->stripPointerCasts();
1860  PointerType *StrippedPtrTy = cast<PointerType>(StrippedPtr->getType());
1861 
1862  if (StrippedPtr != PtrOp) {
1863  bool HasZeroPointerIndex = false;
1864  if (auto *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1865  HasZeroPointerIndex = C->isZero();
1866 
1867  // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1868  // into : GEP [10 x i8]* X, i32 0, ...
1869  //
1870  // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1871  // into : GEP i8* X, ...
1872  //
1873  // This occurs when the program declares an array extern like "int X[];"
1874  if (HasZeroPointerIndex) {
1875  if (auto *CATy = dyn_cast<ArrayType>(GEPEltType)) {
1876  // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1877  if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1878  // -> GEP i8* X, ...
1879  SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1881  StrippedPtrTy->getElementType(), StrippedPtr, Idx, GEP.getName());
1882  Res->setIsInBounds(GEP.isInBounds());
1883  if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1884  return Res;
1885  // Insert Res, and create an addrspacecast.
1886  // e.g.,
1887  // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1888  // ->
1889  // %0 = GEP i8 addrspace(1)* X, ...
1890  // addrspacecast i8 addrspace(1)* %0 to i8*
1891  return new AddrSpaceCastInst(Builder.Insert(Res), GEPType);
1892  }
1893 
1894  if (auto *XATy = dyn_cast<ArrayType>(StrippedPtrTy->getElementType())) {
1895  // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1896  if (CATy->getElementType() == XATy->getElementType()) {
1897  // -> GEP [10 x i8]* X, i32 0, ...
1898  // At this point, we know that the cast source type is a pointer
1899  // to an array of the same type as the destination pointer
1900  // array. Because the array type is never stepped over (there
1901  // is a leading zero) we can fold the cast into this GEP.
1902  if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1903  GEP.setOperand(0, StrippedPtr);
1904  GEP.setSourceElementType(XATy);
1905  return &GEP;
1906  }
1907  // Cannot replace the base pointer directly because StrippedPtr's
1908  // address space is different. Instead, create a new GEP followed by
1909  // an addrspacecast.
1910  // e.g.,
1911  // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1912  // i32 0, ...
1913  // ->
1914  // %0 = GEP [10 x i8] addrspace(1)* X, ...
1915  // addrspacecast i8 addrspace(1)* %0 to i8*
1916  SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1917  Value *NewGEP = GEP.isInBounds()
1918  ? Builder.CreateInBoundsGEP(
1919  nullptr, StrippedPtr, Idx, GEP.getName())
1920  : Builder.CreateGEP(nullptr, StrippedPtr, Idx,
1921  GEP.getName());
1922  return new AddrSpaceCastInst(NewGEP, GEPType);
1923  }
1924  }
1925  }
1926  } else if (GEP.getNumOperands() == 2) {
1927  // Transform things like:
1928  // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1929  // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1930  Type *SrcEltTy = StrippedPtrTy->getElementType();
1931  if (SrcEltTy->isArrayTy() &&
1932  DL.getTypeAllocSize(SrcEltTy->getArrayElementType()) ==
1933  DL.getTypeAllocSize(GEPEltType)) {
1934  Type *IdxType = DL.getIndexType(GEPType);
1935  Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1936  Value *NewGEP =
1937  GEP.isInBounds()
1938  ? Builder.CreateInBoundsGEP(nullptr, StrippedPtr, Idx,
1939  GEP.getName())
1940  : Builder.CreateGEP(nullptr, StrippedPtr, Idx, GEP.getName());
1941 
1942  // V and GEP are both pointer types --> BitCast
1943  return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, GEPType);
1944  }
1945 
1946  // Transform things like:
1947  // %V = mul i64 %N, 4
1948  // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1949  // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1950  if (GEPEltType->isSized() && SrcEltTy->isSized()) {
1951  // Check that changing the type amounts to dividing the index by a scale
1952  // factor.
1953  uint64_t ResSize = DL.getTypeAllocSize(GEPEltType);
1954  uint64_t SrcSize = DL.getTypeAllocSize(SrcEltTy);
1955  if (ResSize && SrcSize % ResSize == 0) {
1956  Value *Idx = GEP.getOperand(1);
1957  unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1958  uint64_t Scale = SrcSize / ResSize;
1959 
1960  // Earlier transforms ensure that the index has the right type
1961  // according to Data Layout, which considerably simplifies the
1962  // logic by eliminating implicit casts.
1963  assert(Idx->getType() == DL.getIndexType(GEPType) &&
1964  "Index type does not match the Data Layout preferences");
1965 
1966  bool NSW;
1967  if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1968  // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1969  // If the multiplication NewIdx * Scale may overflow then the new
1970  // GEP may not be "inbounds".
1971  Value *NewGEP =
1972  GEP.isInBounds() && NSW
1973  ? Builder.CreateInBoundsGEP(nullptr, StrippedPtr, NewIdx,
1974  GEP.getName())
1975  : Builder.CreateGEP(nullptr, StrippedPtr, NewIdx,
1976  GEP.getName());
1977 
1978  // The NewGEP must be pointer typed, so must the old one -> BitCast
1980  GEPType);
1981  }
1982  }
1983  }
1984 
1985  // Similarly, transform things like:
1986  // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1987  // (where tmp = 8*tmp2) into:
1988  // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1989  if (GEPEltType->isSized() && SrcEltTy->isSized() &&
1990  SrcEltTy->isArrayTy()) {
1991  // Check that changing to the array element type amounts to dividing the
1992  // index by a scale factor.
1993  uint64_t ResSize = DL.getTypeAllocSize(GEPEltType);
1994  uint64_t ArrayEltSize =
1995  DL.getTypeAllocSize(SrcEltTy->getArrayElementType());
1996  if (ResSize && ArrayEltSize % ResSize == 0) {
1997  Value *Idx = GEP.getOperand(1);
1998  unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1999  uint64_t Scale = ArrayEltSize / ResSize;
2000 
2001  // Earlier transforms ensure that the index has the right type
2002  // according to the Data Layout, which considerably simplifies
2003  // the logic by eliminating implicit casts.
2004  assert(Idx->getType() == DL.getIndexType(GEPType) &&
2005  "Index type does not match the Data Layout preferences");
2006 
2007  bool NSW;
2008  if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
2009  // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
2010  // If the multiplication NewIdx * Scale may overflow then the new
2011  // GEP may not be "inbounds".
2012  Type *IndTy = DL.getIndexType(GEPType);
2013  Value *Off[2] = {Constant::getNullValue(IndTy), NewIdx};
2014 
2015  Value *NewGEP = GEP.isInBounds() && NSW
2016  ? Builder.CreateInBoundsGEP(
2017  SrcEltTy, StrippedPtr, Off, GEP.getName())
2018  : Builder.CreateGEP(SrcEltTy, StrippedPtr, Off,
2019  GEP.getName());
2020  // The NewGEP must be pointer typed, so must the old one -> BitCast
2022  GEPType);
2023  }
2024  }
2025  }
2026  }
2027  }
2028 
2029  // addrspacecast between types is canonicalized as a bitcast, then an
2030  // addrspacecast. To take advantage of the below bitcast + struct GEP, look
2031  // through the addrspacecast.
2032  Value *ASCStrippedPtrOp = PtrOp;
2033  if (auto *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
2034  // X = bitcast A addrspace(1)* to B addrspace(1)*
2035  // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
2036  // Z = gep Y, <...constant indices...>
2037  // Into an addrspacecasted GEP of the struct.
2038  if (auto *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
2039  ASCStrippedPtrOp = BC;
2040  }
2041 
2042  if (auto *BCI = dyn_cast<BitCastInst>(ASCStrippedPtrOp)) {
2043  Value *SrcOp = BCI->getOperand(0);
2044  PointerType *SrcType = cast<PointerType>(BCI->getSrcTy());
2045  Type *SrcEltType = SrcType->getElementType();
2046 
2047  // GEP directly using the source operand if this GEP is accessing an element
2048  // of a bitcasted pointer to vector or array of the same dimensions:
2049  // gep (bitcast <c x ty>* X to [c x ty]*), Y, Z --> gep X, Y, Z
2050  // gep (bitcast [c x ty]* X to <c x ty>*), Y, Z --> gep X, Y, Z
2051  auto areMatchingArrayAndVecTypes = [](Type *ArrTy, Type *VecTy) {
2052  return ArrTy->getArrayElementType() == VecTy->getVectorElementType() &&
2053  ArrTy->getArrayNumElements() == VecTy->getVectorNumElements();
2054  };
2055  if (GEP.getNumOperands() == 3 &&
2056  ((GEPEltType->isArrayTy() && SrcEltType->isVectorTy() &&
2057  areMatchingArrayAndVecTypes(GEPEltType, SrcEltType)) ||
2058  (GEPEltType->isVectorTy() && SrcEltType->isArrayTy() &&
2059  areMatchingArrayAndVecTypes(SrcEltType, GEPEltType)))) {
2060 
2061  // Create a new GEP here, as using `setOperand()` followed by
2062  // `setSourceElementType()` won't actually update the type of the
2063  // existing GEP Value. Causing issues if this Value is accessed when
2064  // constructing an AddrSpaceCastInst
2065  Value *NGEP =
2066  GEP.isInBounds()
2067  ? Builder.CreateInBoundsGEP(nullptr, SrcOp, {Ops[1], Ops[2]})
2068  : Builder.CreateGEP(nullptr, SrcOp, {Ops[1], Ops[2]});
2069  NGEP->takeName(&GEP);
2070 
2071  // Preserve GEP address space to satisfy users
2072  if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
2073  return new AddrSpaceCastInst(NGEP, GEPType);
2074 
2075  return replaceInstUsesWith(GEP, NGEP);
2076  }
2077 
2078  // See if we can simplify:
2079  // X = bitcast A* to B*
2080  // Y = gep X, <...constant indices...>
2081  // into a gep of the original struct. This is important for SROA and alias
2082  // analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
2083  unsigned OffsetBits = DL.getIndexTypeSizeInBits(GEPType);
2084  APInt Offset(OffsetBits, 0);
2085  if (!isa<BitCastInst>(SrcOp) && GEP.accumulateConstantOffset(DL, Offset)) {
2086  // If this GEP instruction doesn't move the pointer, just replace the GEP
2087  // with a bitcast of the real input to the dest type.
2088  if (!Offset) {
2089  // If the bitcast is of an allocation, and the allocation will be
2090  // converted to match the type of the cast, don't touch this.
2091  if (isa<AllocaInst>(SrcOp) || isAllocationFn(SrcOp, &TLI)) {
2092  // See if the bitcast simplifies, if so, don't nuke this GEP yet.
2093  if (Instruction *I = visitBitCast(*BCI)) {
2094  if (I != BCI) {
2095  I->takeName(BCI);
2096  BCI->getParent()->getInstList().insert(BCI->getIterator(), I);
2097  replaceInstUsesWith(*BCI, I);
2098  }
2099  return &GEP;
2100  }
2101  }
2102 
2103  if (SrcType->getPointerAddressSpace() != GEP.getAddressSpace())
2104  return new AddrSpaceCastInst(SrcOp, GEPType);
2105  return new BitCastInst(SrcOp, GEPType);
2106  }
2107 
2108  // Otherwise, if the offset is non-zero, we need to find out if there is a
2109  // field at Offset in 'A's type. If so, we can pull the cast through the
2110  // GEP.
2111  SmallVector<Value*, 8> NewIndices;
2112  if (FindElementAtOffset(SrcType, Offset.getSExtValue(), NewIndices)) {
2113  Value *NGEP =
2114  GEP.isInBounds()
2115  ? Builder.CreateInBoundsGEP(nullptr, SrcOp, NewIndices)
2116  : Builder.CreateGEP(nullptr, SrcOp, NewIndices);
2117 
2118  if (NGEP->getType() == GEPType)
2119  return replaceInstUsesWith(GEP, NGEP);
2120  NGEP->takeName(&GEP);
2121 
2122  if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
2123  return new AddrSpaceCastInst(NGEP, GEPType);
2124  return new BitCastInst(NGEP, GEPType);
2125  }
2126  }
2127  }
2128 
2129  if (!GEP.isInBounds()) {
2130  unsigned IdxWidth =
2131  DL.getIndexSizeInBits(PtrOp->getType()->getPointerAddressSpace());
2132  APInt BasePtrOffset(IdxWidth, 0);
2133  Value *UnderlyingPtrOp =
2135  BasePtrOffset);
2136  if (auto *AI = dyn_cast<AllocaInst>(UnderlyingPtrOp)) {
2137  if (GEP.accumulateConstantOffset(DL, BasePtrOffset) &&
2138  BasePtrOffset.isNonNegative()) {
2139  APInt AllocSize(IdxWidth, DL.getTypeAllocSize(AI->getAllocatedType()));
2140  if (BasePtrOffset.ule(AllocSize)) {
2142  PtrOp, makeArrayRef(Ops).slice(1), GEP.getName());
2143  }
2144  }
2145  }
2146  }
2147 
2148  return nullptr;
2149 }
2150 
2152  Instruction *AI) {
2153  if (isa<ConstantPointerNull>(V))
2154  return true;
2155  if (auto *LI = dyn_cast<LoadInst>(V))
2156  return isa<GlobalVariable>(LI->getPointerOperand());
2157  // Two distinct allocations will never be equal.
2158  // We rely on LookThroughBitCast in isAllocLikeFn being false, since looking
2159  // through bitcasts of V can cause
2160  // the result statement below to be true, even when AI and V (ex:
2161  // i8* ->i32* ->i8* of AI) are the same allocations.
2162  return isAllocLikeFn(V, TLI) && V != AI;
2163 }
2164 
2167  const TargetLibraryInfo *TLI) {
2169  Worklist.push_back(AI);
2170 
2171  do {
2172  Instruction *PI = Worklist.pop_back_val();
2173  for (User *U : PI->users()) {
2174  Instruction *I = cast<Instruction>(U);
2175  switch (I->getOpcode()) {
2176  default:
2177  // Give up the moment we see something we can't handle.
2178  return false;
2179 
2180  case Instruction::AddrSpaceCast:
2181  case Instruction::BitCast:
2182  case Instruction::GetElementPtr:
2183  Users.emplace_back(I);
2184  Worklist.push_back(I);
2185  continue;
2186 
2187  case Instruction::ICmp: {
2188  ICmpInst *ICI = cast<ICmpInst>(I);
2189  // We can fold eq/ne comparisons with null to false/true, respectively.
2190  // We also fold comparisons in some conditions provided the alloc has
2191  // not escaped (see isNeverEqualToUnescapedAlloc).
2192  if (!ICI->isEquality())
2193  return false;
2194  unsigned OtherIndex = (ICI->getOperand(0) == PI) ? 1 : 0;
2195  if (!isNeverEqualToUnescapedAlloc(ICI->getOperand(OtherIndex), TLI, AI))
2196  return false;
2197  Users.emplace_back(I);
2198  continue;
2199  }
2200 
2201  case Instruction::Call:
2202  // Ignore no-op and store intrinsics.
2203  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2204  switch (II->getIntrinsicID()) {
2205  default:
2206  return false;
2207 
2208  case Intrinsic::memmove:
2209  case Intrinsic::memcpy:
2210  case Intrinsic::memset: {
2211  MemIntrinsic *MI = cast<MemIntrinsic>(II);
2212  if (MI->isVolatile() || MI->getRawDest() != PI)
2213  return false;
2215  }
2216  case Intrinsic::invariant_start:
2217  case Intrinsic::invariant_end:
2218  case Intrinsic::lifetime_start:
2219  case Intrinsic::lifetime_end:
2220  case Intrinsic::objectsize:
2221  Users.emplace_back(I);
2222  continue;
2223  }
2224  }
2225 
2226  if (isFreeCall(I, TLI)) {
2227  Users.emplace_back(I);
2228  continue;
2229  }
2230  return false;
2231 
2232  case Instruction::Store: {
2233  StoreInst *SI = cast<StoreInst>(I);
2234  if (SI->isVolatile() || SI->getPointerOperand() != PI)
2235  return false;
2236  Users.emplace_back(I);
2237  continue;
2238  }
2239  }
2240  llvm_unreachable("missing a return?");
2241  }
2242  } while (!Worklist.empty());
2243  return true;
2244 }
2245 
2247  // If we have a malloc call which is only used in any amount of comparisons
2248  // to null and free calls, delete the calls and replace the comparisons with
2249  // true or false as appropriate.
2251 
2252  // If we are removing an alloca with a dbg.declare, insert dbg.value calls
2253  // before each store.
2255  std::unique_ptr<DIBuilder> DIB;
2256  if (isa<AllocaInst>(MI)) {
2257  DIIs = FindDbgAddrUses(&MI);
2258  DIB.reset(new DIBuilder(*MI.getModule(), /*AllowUnresolved=*/false));
2259  }
2260 
2261  if (isAllocSiteRemovable(&MI, Users, &TLI)) {
2262  for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2263  // Lowering all @llvm.objectsize calls first because they may
2264  // use a bitcast/GEP of the alloca we are removing.
2265  if (!Users[i])
2266  continue;
2267 
2268  Instruction *I = cast<Instruction>(&*Users[i]);
2269 
2270  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2271  if (II->getIntrinsicID() == Intrinsic::objectsize) {
2272  ConstantInt *Result = lowerObjectSizeCall(II, DL, &TLI,
2273  /*MustSucceed=*/true);
2274  replaceInstUsesWith(*I, Result);
2275  eraseInstFromFunction(*I);
2276  Users[i] = nullptr; // Skip examining in the next loop.
2277  }
2278  }
2279  }
2280  for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2281  if (!Users[i])
2282  continue;
2283 
2284  Instruction *I = cast<Instruction>(&*Users[i]);
2285 
2286  if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
2287  replaceInstUsesWith(*C,
2289  C->isFalseWhenEqual()));
2290  } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I) ||
2291  isa<AddrSpaceCastInst>(I)) {
2292  replaceInstUsesWith(*I, UndefValue::get(I->getType()));
2293  } else if (auto *SI = dyn_cast<StoreInst>(I)) {
2294  for (auto *DII : DIIs)
2295  ConvertDebugDeclareToDebugValue(DII, SI, *DIB);
2296  }
2297  eraseInstFromFunction(*I);
2298  }
2299 
2300  if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
2301  // Replace invoke with a NOP intrinsic to maintain the original CFG
2302  Module *M = II->getModule();
2303  Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
2304  InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
2305  None, "", II->getParent());
2306  }
2307 
2308  for (auto *DII : DIIs)
2309  eraseInstFromFunction(*DII);
2310 
2311  return eraseInstFromFunction(MI);
2312  }
2313  return nullptr;
2314 }
2315 
2316 /// Move the call to free before a NULL test.
2317 ///
2318 /// Check if this free is accessed after its argument has been test
2319 /// against NULL (property 0).
2320 /// If yes, it is legal to move this call in its predecessor block.
2321 ///
2322 /// The move is performed only if the block containing the call to free
2323 /// will be removed, i.e.:
2324 /// 1. it has only one predecessor P, and P has two successors
2325 /// 2. it contains the call, noops, and an unconditional branch
2326 /// 3. its successor is the same as its predecessor's successor
2327 ///
2328 /// The profitability is out-of concern here and this function should
2329 /// be called only if the caller knows this transformation would be
2330 /// profitable (e.g., for code size).
2332  const DataLayout &DL) {
2333  Value *Op = FI.getArgOperand(0);
2334  BasicBlock *FreeInstrBB = FI.getParent();
2335  BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
2336 
2337  // Validate part of constraint #1: Only one predecessor
2338  // FIXME: We can extend the number of predecessor, but in that case, we
2339  // would duplicate the call to free in each predecessor and it may
2340  // not be profitable even for code size.
2341  if (!PredBB)
2342  return nullptr;
2343 
2344  // Validate constraint #2: Does this block contains only the call to
2345  // free, noops, and an unconditional branch?
2346  BasicBlock *SuccBB;
2347  Instruction *FreeInstrBBTerminator = FreeInstrBB->getTerminator();
2348  if (!match(FreeInstrBBTerminator, m_UnconditionalBr(SuccBB)))
2349  return nullptr;
2350 
2351  // If there are only 2 instructions in the block, at this point,
2352  // this is the call to free and unconditional.
2353  // If there are more than 2 instructions, check that they are noops
2354  // i.e., they won't hurt the performance of the generated code.
2355  if (FreeInstrBB->size() != 2) {
2356  for (const Instruction &Inst : *FreeInstrBB) {
2357  if (&Inst == &FI || &Inst == FreeInstrBBTerminator)
2358  continue;
2359  auto *Cast = dyn_cast<CastInst>(&Inst);
2360  if (!Cast || !Cast->isNoopCast(DL))
2361  return nullptr;
2362  }
2363  }
2364  // Validate the rest of constraint #1 by matching on the pred branch.
2365  Instruction *TI = PredBB->getTerminator();
2366  BasicBlock *TrueBB, *FalseBB;
2367  ICmpInst::Predicate Pred;
2368  if (!match(TI, m_Br(m_ICmp(Pred,
2369  m_CombineOr(m_Specific(Op),
2370  m_Specific(Op->stripPointerCasts())),
2371  m_Zero()),
2372  TrueBB, FalseBB)))
2373  return nullptr;
2374  if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2375  return nullptr;
2376 
2377  // Validate constraint #3: Ensure the null case just falls through.
2378  if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
2379  return nullptr;
2380  assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
2381  "Broken CFG: missing edge from predecessor to successor");
2382 
2383  // At this point, we know that everything in FreeInstrBB can be moved
2384  // before TI.
2385  for (BasicBlock::iterator It = FreeInstrBB->begin(), End = FreeInstrBB->end();
2386  It != End;) {
2387  Instruction &Instr = *It++;
2388  if (&Instr == FreeInstrBBTerminator)
2389  break;
2390  Instr.moveBefore(TI);
2391  }
2392  assert(FreeInstrBB->size() == 1 &&
2393  "Only the branch instruction should remain");
2394  return &FI;
2395 }
2396 
2398  Value *Op = FI.getArgOperand(0);
2399 
2400  // free undef -> unreachable.
2401  if (isa<UndefValue>(Op)) {
2402  // Insert a new store to null because we cannot modify the CFG here.
2403  Builder.CreateStore(ConstantInt::getTrue(FI.getContext()),
2405  return eraseInstFromFunction(FI);
2406  }
2407 
2408  // If we have 'free null' delete the instruction. This can happen in stl code
2409  // when lots of inlining happens.
2410  if (isa<ConstantPointerNull>(Op))
2411  return eraseInstFromFunction(FI);
2412 
2413  // If we optimize for code size, try to move the call to free before the null
2414  // test so that simplify cfg can remove the empty block and dead code
2415  // elimination the branch. I.e., helps to turn something like:
2416  // if (foo) free(foo);
2417  // into
2418  // free(foo);
2419  if (MinimizeSize)
2420  if (Instruction *I = tryToMoveFreeBeforeNullTest(FI, DL))
2421  return I;
2422 
2423  return nullptr;
2424 }
2425 
2427  if (RI.getNumOperands() == 0) // ret void
2428  return nullptr;
2429 
2430  Value *ResultOp = RI.getOperand(0);
2431  Type *VTy = ResultOp->getType();
2432  if (!VTy->isIntegerTy())
2433  return nullptr;
2434 
2435  // There might be assume intrinsics dominating this return that completely
2436  // determine the value. If so, constant fold it.
2437  KnownBits Known = computeKnownBits(ResultOp, 0, &RI);
2438  if (Known.isConstant())
2439  RI.setOperand(0, Constant::getIntegerValue(VTy, Known.getConstant()));
2440 
2441  return nullptr;
2442 }
2443 
2445  // Change br (not X), label True, label False to: br X, label False, True
2446  Value *X = nullptr;
2447  BasicBlock *TrueDest;
2448  BasicBlock *FalseDest;
2449  if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2450  !isa<Constant>(X)) {
2451  // Swap Destinations and condition...
2452  BI.setCondition(X);
2453  BI.swapSuccessors();
2454  return &BI;
2455  }
2456 
2457  // If the condition is irrelevant, remove the use so that other
2458  // transforms on the condition become more effective.
2459  if (BI.isConditional() && !isa<ConstantInt>(BI.getCondition()) &&
2460  BI.getSuccessor(0) == BI.getSuccessor(1)) {
2462  return &BI;
2463  }
2464 
2465  // Canonicalize, for example, icmp_ne -> icmp_eq or fcmp_one -> fcmp_oeq.
2466  CmpInst::Predicate Pred;
2467  if (match(&BI, m_Br(m_OneUse(m_Cmp(Pred, m_Value(), m_Value())), TrueDest,
2468  FalseDest)) &&
2469  !isCanonicalPredicate(Pred)) {
2470  // Swap destinations and condition.
2471  CmpInst *Cond = cast<CmpInst>(BI.getCondition());
2473  BI.swapSuccessors();
2474  Worklist.Add(Cond);
2475  return &BI;
2476  }
2477 
2478  return nullptr;
2479 }
2480 
2482  Value *Cond = SI.getCondition();
2483  Value *Op0;
2484  ConstantInt *AddRHS;
2485  if (match(Cond, m_Add(m_Value(Op0), m_ConstantInt(AddRHS)))) {
2486  // Change 'switch (X+4) case 1:' into 'switch (X) case -3'.
2487  for (auto Case : SI.cases()) {
2488  Constant *NewCase = ConstantExpr::getSub(Case.getCaseValue(), AddRHS);
2489  assert(isa<ConstantInt>(NewCase) &&
2490  "Result of expression should be constant");
2491  Case.setValue(cast<ConstantInt>(NewCase));
2492  }
2493  SI.setCondition(Op0);
2494  return &SI;
2495  }
2496 
2497  KnownBits Known = computeKnownBits(Cond, 0, &SI);
2498  unsigned LeadingKnownZeros = Known.countMinLeadingZeros();
2499  unsigned LeadingKnownOnes = Known.countMinLeadingOnes();
2500 
2501  // Compute the number of leading bits we can ignore.
2502  // TODO: A better way to determine this would use ComputeNumSignBits().
2503  for (auto &C : SI.cases()) {
2504  LeadingKnownZeros = std::min(
2505  LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2506  LeadingKnownOnes = std::min(
2507  LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2508  }
2509 
2510  unsigned NewWidth = Known.getBitWidth() - std::max(LeadingKnownZeros, LeadingKnownOnes);
2511 
2512  // Shrink the condition operand if the new type is smaller than the old type.
2513  // But do not shrink to a non-standard type, because backend can't generate
2514  // good code for that yet.
2515  // TODO: We can make it aggressive again after fixing PR39569.
2516  if (NewWidth > 0 && NewWidth < Known.getBitWidth() &&
2517  shouldChangeType(Known.getBitWidth(), NewWidth)) {
2518  IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2519  Builder.SetInsertPoint(&SI);
2520  Value *NewCond = Builder.CreateTrunc(Cond, Ty, "trunc");
2521  SI.setCondition(NewCond);
2522 
2523  for (auto Case : SI.cases()) {
2524  APInt TruncatedCase = Case.getCaseValue()->getValue().trunc(NewWidth);
2525  Case.setValue(ConstantInt::get(SI.getContext(), TruncatedCase));
2526  }
2527  return &SI;
2528  }
2529 
2530  return nullptr;
2531 }
2532 
2534  Value *Agg = EV.getAggregateOperand();
2535 
2536  if (!EV.hasIndices())
2537  return replaceInstUsesWith(EV, Agg);
2538 
2539  if (Value *V = SimplifyExtractValueInst(Agg, EV.getIndices(),
2540  SQ.getWithInstruction(&EV)))
2541  return replaceInstUsesWith(EV, V);
2542 
2543  if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2544  // We're extracting from an insertvalue instruction, compare the indices
2545  const unsigned *exti, *exte, *insi, *inse;
2546  for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2547  exte = EV.idx_end(), inse = IV->idx_end();
2548  exti != exte && insi != inse;
2549  ++exti, ++insi) {
2550  if (*insi != *exti)
2551  // The insert and extract both reference distinctly different elements.
2552  // This means the extract is not influenced by the insert, and we can
2553  // replace the aggregate operand of the extract with the aggregate
2554  // operand of the insert. i.e., replace
2555  // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2556  // %E = extractvalue { i32, { i32 } } %I, 0
2557  // with
2558  // %E = extractvalue { i32, { i32 } } %A, 0
2559  return ExtractValueInst::Create(IV->getAggregateOperand(),
2560  EV.getIndices());
2561  }
2562  if (exti == exte && insi == inse)
2563  // Both iterators are at the end: Index lists are identical. Replace
2564  // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2565  // %C = extractvalue { i32, { i32 } } %B, 1, 0
2566  // with "i32 42"
2567  return replaceInstUsesWith(EV, IV->getInsertedValueOperand());
2568  if (exti == exte) {
2569  // The extract list is a prefix of the insert list. i.e. replace
2570  // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2571  // %E = extractvalue { i32, { i32 } } %I, 1
2572  // with
2573  // %X = extractvalue { i32, { i32 } } %A, 1
2574  // %E = insertvalue { i32 } %X, i32 42, 0
2575  // by switching the order of the insert and extract (though the
2576  // insertvalue should be left in, since it may have other uses).
2577  Value *NewEV = Builder.CreateExtractValue(IV->getAggregateOperand(),
2578  EV.getIndices());
2579  return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2580  makeArrayRef(insi, inse));
2581  }
2582  if (insi == inse)
2583  // The insert list is a prefix of the extract list
2584  // We can simply remove the common indices from the extract and make it
2585  // operate on the inserted value instead of the insertvalue result.
2586  // i.e., replace
2587  // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2588  // %E = extractvalue { i32, { i32 } } %I, 1, 0
2589  // with
2590  // %E extractvalue { i32 } { i32 42 }, 0
2591  return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2592  makeArrayRef(exti, exte));
2593  }
2594  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2595  // We're extracting from an intrinsic, see if we're the only user, which
2596  // allows us to simplify multiple result intrinsics to simpler things that
2597  // just get one value.
2598  if (II->hasOneUse()) {
2599  // Check if we're grabbing the overflow bit or the result of a 'with
2600  // overflow' intrinsic. If it's the latter we can remove the intrinsic
2601  // and replace it with a traditional binary instruction.
2602  switch (II->getIntrinsicID()) {
2603  case Intrinsic::uadd_with_overflow:
2604  case Intrinsic::sadd_with_overflow:
2605  if (*EV.idx_begin() == 0) { // Normal result.
2606  Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2607  replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2608  eraseInstFromFunction(*II);
2609  return BinaryOperator::CreateAdd(LHS, RHS);
2610  }
2611 
2612  // If the normal result of the add is dead, and the RHS is a constant,
2613  // we can transform this into a range comparison.
2614  // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2615  if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2616  if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2617  return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2618  ConstantExpr::getNot(CI));
2619  break;
2620  case Intrinsic::usub_with_overflow:
2621  case Intrinsic::ssub_with_overflow:
2622  if (*EV.idx_begin() == 0) { // Normal result.
2623  Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2624  replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2625  eraseInstFromFunction(*II);
2626  return BinaryOperator::CreateSub(LHS, RHS);
2627  }
2628  break;
2629  case Intrinsic::umul_with_overflow:
2630  case Intrinsic::smul_with_overflow:
2631  if (*EV.idx_begin() == 0) { // Normal result.
2632  Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2633  replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2634  eraseInstFromFunction(*II);
2635  return BinaryOperator::CreateMul(LHS, RHS);
2636  }
2637  break;
2638  default:
2639  break;
2640  }
2641  }
2642  }
2643  if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2644  // If the (non-volatile) load only has one use, we can rewrite this to a
2645  // load from a GEP. This reduces the size of the load. If a load is used
2646  // only by extractvalue instructions then this either must have been
2647  // optimized before, or it is a struct with padding, in which case we
2648  // don't want to do the transformation as it loses padding knowledge.
2649  if (L->isSimple() && L->hasOneUse()) {
2650  // extractvalue has integer indices, getelementptr has Value*s. Convert.
2651  SmallVector<Value*, 4> Indices;
2652  // Prefix an i32 0 since we need the first element.
2653  Indices.push_back(Builder.getInt32(0));
2654  for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2655  I != E; ++I)
2656  Indices.push_back(Builder.getInt32(*I));
2657 
2658  // We need to insert these at the location of the old load, not at that of
2659  // the extractvalue.
2660  Builder.SetInsertPoint(L);
2661  Value *GEP = Builder.CreateInBoundsGEP(L->getType(),
2662  L->getPointerOperand(), Indices);
2663  Instruction *NL = Builder.CreateLoad(GEP);
2664  // Whatever aliasing information we had for the orignal load must also
2665  // hold for the smaller load, so propagate the annotations.
2666  AAMDNodes Nodes;
2667  L->getAAMetadata(Nodes);
2668  NL->setAAMetadata(Nodes);
2669  // Returning the load directly will cause the main loop to insert it in
2670  // the wrong spot, so use replaceInstUsesWith().
2671  return replaceInstUsesWith(EV, NL);
2672  }
2673  // We could simplify extracts from other values. Note that nested extracts may
2674  // already be simplified implicitly by the above: extract (extract (insert) )
2675  // will be translated into extract ( insert ( extract ) ) first and then just
2676  // the value inserted, if appropriate. Similarly for extracts from single-use
2677  // loads: extract (extract (load)) will be translated to extract (load (gep))
2678  // and if again single-use then via load (gep (gep)) to load (gep).
2679  // However, double extracts from e.g. function arguments or return values
2680  // aren't handled yet.
2681  return nullptr;
2682 }
2683 
2684 /// Return 'true' if the given typeinfo will match anything.
2685 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) {
2686  switch (Personality) {
2687  case EHPersonality::GNU_C:
2689  case EHPersonality::Rust:
2690  // The GCC C EH and Rust personality only exists to support cleanups, so
2691  // it's not clear what the semantics of catch clauses are.
2692  return false;
2694  return false;
2696  // While __gnat_all_others_value will match any Ada exception, it doesn't
2697  // match foreign exceptions (or didn't, before gcc-4.7).
2698  return false;
2707  return TypeInfo->isNullValue();
2708  }
2709  llvm_unreachable("invalid enum");
2710 }
2711 
2712 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2713  return
2714  cast<ArrayType>(LHS->getType())->getNumElements()
2715  <
2716  cast<ArrayType>(RHS->getType())->getNumElements();
2717 }
2718 
2720  // The logic here should be correct for any real-world personality function.
2721  // However if that turns out not to be true, the offending logic can always
2722  // be conditioned on the personality function, like the catch-all logic is.
2723  EHPersonality Personality =
2725 
2726  // Simplify the list of clauses, eg by removing repeated catch clauses
2727  // (these are often created by inlining).
2728  bool MakeNewInstruction = false; // If true, recreate using the following:
2729  SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2730  bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2731 
2732  SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2733  for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2734  bool isLastClause = i + 1 == e;
2735  if (LI.isCatch(i)) {
2736  // A catch clause.
2737  Constant *CatchClause = LI.getClause(i);
2738  Constant *TypeInfo = CatchClause->stripPointerCasts();
2739 
2740  // If we already saw this clause, there is no point in having a second
2741  // copy of it.
2742  if (AlreadyCaught.insert(TypeInfo).second) {
2743  // This catch clause was not already seen.
2744  NewClauses.push_back(CatchClause);
2745  } else {
2746  // Repeated catch clause - drop the redundant copy.
2747  MakeNewInstruction = true;
2748  }
2749 
2750  // If this is a catch-all then there is no point in keeping any following
2751  // clauses or marking the landingpad as having a cleanup.
2752  if (isCatchAll(Personality, TypeInfo)) {
2753  if (!isLastClause)
2754  MakeNewInstruction = true;
2755  CleanupFlag = false;
2756  break;
2757  }
2758  } else {
2759  // A filter clause. If any of the filter elements were already caught
2760  // then they can be dropped from the filter. It is tempting to try to
2761  // exploit the filter further by saying that any typeinfo that does not
2762  // occur in the filter can't be caught later (and thus can be dropped).
2763  // However this would be wrong, since typeinfos can match without being
2764  // equal (for example if one represents a C++ class, and the other some
2765  // class derived from it).
2766  assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2767  Constant *FilterClause = LI.getClause(i);
2768  ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2769  unsigned NumTypeInfos = FilterType->getNumElements();
2770 
2771  // An empty filter catches everything, so there is no point in keeping any
2772  // following clauses or marking the landingpad as having a cleanup. By
2773  // dealing with this case here the following code is made a bit simpler.
2774  if (!NumTypeInfos) {
2775  NewClauses.push_back(FilterClause);
2776  if (!isLastClause)
2777  MakeNewInstruction = true;
2778  CleanupFlag = false;
2779  break;
2780  }
2781 
2782  bool MakeNewFilter = false; // If true, make a new filter.
2783  SmallVector<Constant *, 16> NewFilterElts; // New elements.
2784  if (isa<ConstantAggregateZero>(FilterClause)) {
2785  // Not an empty filter - it contains at least one null typeinfo.
2786  assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2787  Constant *TypeInfo =
2788  Constant::getNullValue(FilterType->getElementType());
2789  // If this typeinfo is a catch-all then the filter can never match.
2790  if (isCatchAll(Personality, TypeInfo)) {
2791  // Throw the filter away.
2792  MakeNewInstruction = true;
2793  continue;
2794  }
2795 
2796  // There is no point in having multiple copies of this typeinfo, so
2797  // discard all but the first copy if there is more than one.
2798  NewFilterElts.push_back(TypeInfo);
2799  if (NumTypeInfos > 1)
2800  MakeNewFilter = true;
2801  } else {
2802  ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2803  SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2804  NewFilterElts.reserve(NumTypeInfos);
2805 
2806  // Remove any filter elements that were already caught or that already
2807  // occurred in the filter. While there, see if any of the elements are
2808  // catch-alls. If so, the filter can be discarded.
2809  bool SawCatchAll = false;
2810  for (unsigned j = 0; j != NumTypeInfos; ++j) {
2811  Constant *Elt = Filter->getOperand(j);
2812  Constant *TypeInfo = Elt->stripPointerCasts();
2813  if (isCatchAll(Personality, TypeInfo)) {
2814  // This element is a catch-all. Bail out, noting this fact.
2815  SawCatchAll = true;
2816  break;
2817  }
2818 
2819  // Even if we've seen a type in a catch clause, we don't want to
2820  // remove it from the filter. An unexpected type handler may be
2821  // set up for a call site which throws an exception of the same
2822  // type caught. In order for the exception thrown by the unexpected
2823  // handler to propagate correctly, the filter must be correctly
2824  // described for the call site.
2825  //
2826  // Example:
2827  //
2828  // void unexpected() { throw 1;}
2829  // void foo() throw (int) {
2830  // std::set_unexpected(unexpected);
2831  // try {
2832  // throw 2.0;
2833  // } catch (int i) {}
2834  // }
2835 
2836  // There is no point in having multiple copies of the same typeinfo in
2837  // a filter, so only add it if we didn't already.
2838  if (SeenInFilter.insert(TypeInfo).second)
2839  NewFilterElts.push_back(cast<Constant>(Elt));
2840  }
2841  // A filter containing a catch-all cannot match anything by definition.
2842  if (SawCatchAll) {
2843  // Throw the filter away.
2844  MakeNewInstruction = true;
2845  continue;
2846  }
2847 
2848  // If we dropped something from the filter, make a new one.
2849  if (NewFilterElts.size() < NumTypeInfos)
2850  MakeNewFilter = true;
2851  }
2852  if (MakeNewFilter) {
2853  FilterType = ArrayType::get(FilterType->getElementType(),
2854  NewFilterElts.size());
2855  FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2856  MakeNewInstruction = true;
2857  }
2858 
2859  NewClauses.push_back(FilterClause);
2860 
2861  // If the new filter is empty then it will catch everything so there is
2862  // no point in keeping any following clauses or marking the landingpad
2863  // as having a cleanup. The case of the original filter being empty was
2864  // already handled above.
2865  if (MakeNewFilter && !NewFilterElts.size()) {
2866  assert(MakeNewInstruction && "New filter but not a new instruction!");
2867  CleanupFlag = false;
2868  break;
2869  }
2870  }
2871  }
2872 
2873  // If several filters occur in a row then reorder them so that the shortest
2874  // filters come first (those with the smallest number of elements). This is
2875  // advantageous because shorter filters are more likely to match, speeding up
2876  // unwinding, but mostly because it increases the effectiveness of the other
2877  // filter optimizations below.
2878  for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2879  unsigned j;
2880  // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2881  for (j = i; j != e; ++j)
2882  if (!isa<ArrayType>(NewClauses[j]->getType()))
2883  break;
2884 
2885  // Check whether the filters are already sorted by length. We need to know
2886  // if sorting them is actually going to do anything so that we only make a
2887  // new landingpad instruction if it does.
2888  for (unsigned k = i; k + 1 < j; ++k)
2889  if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2890  // Not sorted, so sort the filters now. Doing an unstable sort would be
2891  // correct too but reordering filters pointlessly might confuse users.
2892  std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2893  shorter_filter);
2894  MakeNewInstruction = true;
2895  break;
2896  }
2897 
2898  // Look for the next batch of filters.
2899  i = j + 1;
2900  }
2901 
2902  // If typeinfos matched if and only if equal, then the elements of a filter L
2903  // that occurs later than a filter F could be replaced by the intersection of
2904  // the elements of F and L. In reality two typeinfos can match without being
2905  // equal (for example if one represents a C++ class, and the other some class
2906  // derived from it) so it would be wrong to perform this transform in general.
2907  // However the transform is correct and useful if F is a subset of L. In that
2908  // case L can be replaced by F, and thus removed altogether since repeating a
2909  // filter is pointless. So here we look at all pairs of filters F and L where
2910  // L follows F in the list of clauses, and remove L if every element of F is
2911  // an element of L. This can occur when inlining C++ functions with exception
2912  // specifications.
2913  for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2914  // Examine each filter in turn.
2915  Value *Filter = NewClauses[i];
2916  ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2917  if (!FTy)
2918  // Not a filter - skip it.
2919  continue;
2920  unsigned FElts = FTy->getNumElements();
2921  // Examine each filter following this one. Doing this backwards means that
2922  // we don't have to worry about filters disappearing under us when removed.
2923  for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2924  Value *LFilter = NewClauses[j];
2925  ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2926  if (!LTy)
2927  // Not a filter - skip it.
2928  continue;
2929  // If Filter is a subset of LFilter, i.e. every element of Filter is also
2930  // an element of LFilter, then discard LFilter.
2931  SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2932  // If Filter is empty then it is a subset of LFilter.
2933  if (!FElts) {
2934  // Discard LFilter.
2935  NewClauses.erase(J);
2936  MakeNewInstruction = true;
2937  // Move on to the next filter.
2938  continue;
2939  }
2940  unsigned LElts = LTy->getNumElements();
2941  // If Filter is longer than LFilter then it cannot be a subset of it.
2942  if (FElts > LElts)
2943  // Move on to the next filter.
2944  continue;
2945  // At this point we know that LFilter has at least one element.
2946  if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2947  // Filter is a subset of LFilter iff Filter contains only zeros (as we
2948  // already know that Filter is not longer than LFilter).
2949  if (isa<ConstantAggregateZero>(Filter)) {
2950  assert(FElts <= LElts && "Should have handled this case earlier!");
2951  // Discard LFilter.
2952  NewClauses.erase(J);
2953  MakeNewInstruction = true;
2954  }
2955  // Move on to the next filter.
2956  continue;
2957  }
2958  ConstantArray *LArray = cast<ConstantArray>(LFilter);
2959  if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2960  // Since Filter is non-empty and contains only zeros, it is a subset of
2961  // LFilter iff LFilter contains a zero.
2962  assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2963  for (unsigned l = 0; l != LElts; ++l)
2964  if (LArray->getOperand(l)->isNullValue()) {
2965  // LFilter contains a zero - discard it.
2966  NewClauses.erase(J);
2967  MakeNewInstruction = true;
2968  break;
2969  }
2970  // Move on to the next filter.
2971  continue;
2972  }
2973  // At this point we know that both filters are ConstantArrays. Loop over
2974  // operands to see whether every element of Filter is also an element of
2975  // LFilter. Since filters tend to be short this is probably faster than
2976  // using a method that scales nicely.
2977  ConstantArray *FArray = cast<ConstantArray>(Filter);
2978  bool AllFound = true;
2979  for (unsigned f = 0; f != FElts; ++f) {
2980  Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2981  AllFound = false;
2982  for (unsigned l = 0; l != LElts; ++l) {
2983  Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2984  if (LTypeInfo == FTypeInfo) {
2985  AllFound = true;
2986  break;
2987  }
2988  }
2989  if (!AllFound)
2990  break;
2991  }
2992  if (AllFound) {
2993  // Discard LFilter.
2994  NewClauses.erase(J);
2995  MakeNewInstruction = true;
2996  }
2997  // Move on to the next filter.
2998  }
2999  }
3000 
3001  // If we changed any of the clauses, replace the old landingpad instruction
3002  // with a new one.
3003  if (MakeNewInstruction) {
3005  NewClauses.size());
3006  for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
3007  NLI->addClause(NewClauses[i]);
3008  // A landing pad with no clauses must have the cleanup flag set. It is
3009  // theoretically possible, though highly unlikely, that we eliminated all
3010  // clauses. If so, force the cleanup flag to true.
3011  if (NewClauses.empty())
3012  CleanupFlag = true;
3013  NLI->setCleanup(CleanupFlag);
3014  return NLI;
3015  }
3016 
3017  // Even if none of the clauses changed, we may nonetheless have understood
3018  // that the cleanup flag is pointless. Clear it if so.
3019  if (LI.isCleanup() != CleanupFlag) {
3020  assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
3021  LI.setCleanup(CleanupFlag);
3022  return &LI;
3023  }
3024 
3025  return nullptr;
3026 }
3027 
3028 /// Try to move the specified instruction from its current block into the
3029 /// beginning of DestBlock, which can only happen if it's safe to move the
3030 /// instruction past all of the instructions between it and the end of its
3031 /// block.
3032 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
3033  assert(I->hasOneUse() && "Invariants didn't hold!");
3034  BasicBlock *SrcBlock = I->getParent();
3035 
3036  // Cannot move control-flow-involving, volatile loads, vaarg, etc.
3037  if (isa<PHINode>(I) || I->isEHPad() || I->mayHaveSideEffects() ||
3038  I->isTerminator())
3039  return false;
3040 
3041  // Do not sink alloca instructions out of the entry block.
3042  if (isa<AllocaInst>(I) && I->getParent() ==
3043  &DestBlock->getParent()->getEntryBlock())
3044  return false;
3045 
3046  // Do not sink into catchswitch blocks.
3047  if (isa<CatchSwitchInst>(DestBlock->getTerminator()))
3048  return false;
3049 
3050  // Do not sink convergent call instructions.
3051  if (auto *CI = dyn_cast<CallInst>(I)) {
3052  if (CI->isConvergent())
3053  return false;
3054  }
3055  // We can only sink load instructions if there is nothing between the load and
3056  // the end of block that could change the value.
3057  if (I->mayReadFromMemory()) {
3058  for (BasicBlock::iterator Scan = I->getIterator(),
3059  E = I->getParent()->end();
3060  Scan != E; ++Scan)
3061  if (Scan->mayWriteToMemory())
3062  return false;
3063  }
3064  BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
3065  I->moveBefore(&*InsertPos);
3066  ++NumSunkInst;
3067 
3068  // Also sink all related debug uses from the source basic block. Otherwise we
3069  // get debug use before the def.
3071  findDbgUsers(DbgUsers, I);
3072  for (auto *DII : DbgUsers) {
3073  if (DII->getParent() == SrcBlock) {
3074  DII->moveBefore(&*InsertPos);
3075  LLVM_DEBUG(dbgs() << "SINK: " << *DII << '\n');
3076  }
3077  }
3078  return true;
3079 }
3080 
3082  while (!Worklist.isEmpty()) {
3083  Instruction *I = Worklist.RemoveOne();
3084  if (I == nullptr) continue; // skip null values.
3085 
3086  // Check to see if we can DCE the instruction.
3087  if (isInstructionTriviallyDead(I, &TLI)) {
3088  LLVM_DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
3089  eraseInstFromFunction(*I);
3090  ++NumDeadInst;
3091  MadeIRChange = true;
3092  continue;
3093  }
3094 
3095  if (!DebugCounter::shouldExecute(VisitCounter))
3096  continue;
3097 
3098  // Instruction isn't dead, see if we can constant propagate it.
3099  if (!I->use_empty() &&
3100  (I->getNumOperands() == 0 || isa<Constant>(I->getOperand(0)))) {
3101  if (Constant *C = ConstantFoldInstruction(I, DL, &TLI)) {
3102  LLVM_DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I
3103  << '\n');
3104 
3105  // Add operands to the worklist.
3106  replaceInstUsesWith(*I, C);
3107  ++NumConstProp;
3108  if (isInstructionTriviallyDead(I, &TLI))
3109  eraseInstFromFunction(*I);
3110  MadeIRChange = true;
3111  continue;
3112  }
3113  }
3114 
3115  // In general, it is possible for computeKnownBits to determine all bits in
3116  // a value even when the operands are not all constants.
3117  Type *Ty = I->getType();
3118  if (ExpensiveCombines && !I->use_empty() && Ty->isIntOrIntVectorTy()) {
3119  KnownBits Known = computeKnownBits(I, /*Depth*/0, I);
3120  if (Known.isConstant()) {
3121  Constant *C = ConstantInt::get(Ty, Known.getConstant());
3122  LLVM_DEBUG(dbgs() << "IC: ConstFold (all bits known) to: " << *C
3123  << " from: " << *I << '\n');
3124 
3125  // Add operands to the worklist.
3126  replaceInstUsesWith(*I, C);
3127  ++NumConstProp;
3128  if (isInstructionTriviallyDead(I, &TLI))
3129  eraseInstFromFunction(*I);
3130  MadeIRChange = true;
3131  continue;
3132  }
3133  }
3134 
3135  // See if we can trivially sink this instruction to a successor basic block.
3136  if (EnableCodeSinking && I->hasOneUse()) {
3137  BasicBlock *BB = I->getParent();
3138  Instruction *UserInst = cast<Instruction>(*I->user_begin());
3139  BasicBlock *UserParent;
3140 
3141  // Get the block the use occurs in.
3142  if (PHINode *PN = dyn_cast<PHINode>(UserInst))
3143  UserParent = PN->getIncomingBlock(*I->use_begin());
3144  else
3145  UserParent = UserInst->getParent();
3146 
3147  if (UserParent != BB) {
3148  bool UserIsSuccessor = false;
3149  // See if the user is one of our successors.
3150  for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
3151  if (*SI == UserParent) {
3152  UserIsSuccessor = true;
3153  break;
3154  }
3155 
3156  // If the user is one of our immediate successors, and if that successor
3157  // only has us as a predecessors (we'd have to split the critical edge
3158  // otherwise), we can keep going.
3159  if (UserIsSuccessor && UserParent->getUniquePredecessor()) {
3160  // Okay, the CFG is simple enough, try to sink this instruction.
3161  if (TryToSinkInstruction(I, UserParent)) {
3162  LLVM_DEBUG(dbgs() << "IC: Sink: " << *I << '\n');
3163  MadeIRChange = true;
3164  // We'll add uses of the sunk instruction below, but since sinking
3165  // can expose opportunities for it's *operands* add them to the
3166  // worklist
3167  for (Use &U : I->operands())
3168  if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
3169  Worklist.Add(OpI);
3170  }
3171  }
3172  }
3173  }
3174 
3175  // Now that we have an instruction, try combining it to simplify it.
3176  Builder.SetInsertPoint(I);
3177  Builder.SetCurrentDebugLocation(I->getDebugLoc());
3178 
3179 #ifndef NDEBUG
3180  std::string OrigI;
3181 #endif
3182  LLVM_DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
3183  LLVM_DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
3184 
3185  if (Instruction *Result = visit(*I)) {
3186  ++NumCombined;
3187  // Should we replace the old instruction with a new one?
3188  if (Result != I) {
3189  LLVM_DEBUG(dbgs() << "IC: Old = " << *I << '\n'
3190  << " New = " << *Result << '\n');
3191 
3192  if (I->getDebugLoc())
3193  Result->setDebugLoc(I->getDebugLoc());
3194  // Everything uses the new instruction now.
3195  I->replaceAllUsesWith(Result);
3196 
3197  // Move the name to the new instruction first.
3198  Result->takeName(I);
3199 
3200  // Push the new instruction and any users onto the worklist.
3201  Worklist.AddUsersToWorkList(*Result);
3202  Worklist.Add(Result);
3203 
3204  // Insert the new instruction into the basic block...
3205  BasicBlock *InstParent = I->getParent();
3206  BasicBlock::iterator InsertPos = I->getIterator();
3207 
3208  // If we replace a PHI with something that isn't a PHI, fix up the
3209  // insertion point.
3210  if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
3211  InsertPos = InstParent->getFirstInsertionPt();
3212 
3213  InstParent->getInstList().insert(InsertPos, Result);
3214 
3215  eraseInstFromFunction(*I);
3216  } else {
3217  LLVM_DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
3218  << " New = " << *I << '\n');
3219 
3220  // If the instruction was modified, it's possible that it is now dead.
3221  // if so, remove it.
3222  if (isInstructionTriviallyDead(I, &TLI)) {
3223  eraseInstFromFunction(*I);
3224  } else {
3225  Worklist.AddUsersToWorkList(*I);
3226  Worklist.Add(I);
3227  }
3228  }
3229  MadeIRChange = true;
3230  }
3231  }
3232 
3233  Worklist.Zap();
3234  return MadeIRChange;
3235 }
3236 
3237 /// Walk the function in depth-first order, adding all reachable code to the
3238 /// worklist.
3239 ///
3240 /// This has a couple of tricks to make the code faster and more powerful. In
3241 /// particular, we constant fold and DCE instructions as we go, to avoid adding
3242 /// them to the worklist (this significantly speeds up instcombine on code where
3243 /// many instructions are dead or constant). Additionally, if we find a branch
3244 /// whose condition is a known constant, we only visit the reachable successors.
3247  InstCombineWorklist &ICWorklist,
3248  const TargetLibraryInfo *TLI) {
3249  bool MadeIRChange = false;
3251  Worklist.push_back(BB);
3252 
3253  SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
3254  DenseMap<Constant *, Constant *> FoldedConstants;
3255 
3256  do {
3257  BB = Worklist.pop_back_val();
3258 
3259  // We have now visited this block! If we've already been here, ignore it.
3260  if (!Visited.insert(BB).second)
3261  continue;
3262 
3263  for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
3264  Instruction *Inst = &*BBI++;
3265 
3266  // DCE instruction if trivially dead.
3267  if (isInstructionTriviallyDead(Inst, TLI)) {
3268  ++NumDeadInst;
3269  LLVM_DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
3270  salvageDebugInfo(*Inst);
3271  Inst->eraseFromParent();
3272  MadeIRChange = true;
3273  continue;
3274  }
3275 
3276  // ConstantProp instruction if trivially constant.
3277  if (!Inst->use_empty() &&
3278  (Inst->getNumOperands() == 0 || isa<Constant>(Inst->getOperand(0))))
3279  if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
3280  LLVM_DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *Inst
3281  << '\n');
3282  Inst->replaceAllUsesWith(C);
3283  ++NumConstProp;
3284  if (isInstructionTriviallyDead(Inst, TLI))
3285  Inst->eraseFromParent();
3286  MadeIRChange = true;
3287  continue;
3288  }
3289 
3290  // See if we can constant fold its operands.
3291  for (Use &U : Inst->operands()) {
3292  if (!isa<ConstantVector>(U) && !isa<ConstantExpr>(U))
3293  continue;
3294 
3295  auto *C = cast<Constant>(U);
3296  Constant *&FoldRes = FoldedConstants[C];
3297  if (!FoldRes)
3298  FoldRes = ConstantFoldConstant(C, DL, TLI);
3299  if (!FoldRes)
3300  FoldRes = C;
3301 
3302  if (FoldRes != C) {
3303  LLVM_DEBUG(dbgs() << "IC: ConstFold operand of: " << *Inst
3304  << "\n Old = " << *C
3305  << "\n New = " << *FoldRes << '\n');
3306  U = FoldRes;
3307  MadeIRChange = true;
3308  }
3309  }
3310 
3311  // Skip processing debug intrinsics in InstCombine. Processing these call instructions
3312  // consumes non-trivial amount of time and provides no value for the optimization.
3313  if (!isa<DbgInfoIntrinsic>(Inst))
3314  InstrsForInstCombineWorklist.push_back(Inst);
3315  }
3316 
3317  // Recursively visit successors. If this is a branch or switch on a
3318  // constant, only visit the reachable successor.
3319  Instruction *TI = BB->getTerminator();
3320  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3321  if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
3322  bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
3323  BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
3324  Worklist.push_back(ReachableBB);
3325  continue;
3326  }
3327  } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3328  if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
3329  Worklist.push_back(SI->findCaseValue(Cond)->getCaseSuccessor());
3330  continue;
3331  }
3332  }
3333 
3334  for (BasicBlock *SuccBB : successors(TI))
3335  Worklist.push_back(SuccBB);
3336  } while (!Worklist.empty());
3337 
3338  // Once we've found all of the instructions to add to instcombine's worklist,
3339  // add them in reverse order. This way instcombine will visit from the top
3340  // of the function down. This jives well with the way that it adds all uses
3341  // of instructions to the worklist after doing a transformation, thus avoiding
3342  // some N^2 behavior in pathological cases.
3343  ICWorklist.AddInitialGroup(InstrsForInstCombineWorklist);
3344 
3345  return MadeIRChange;
3346 }
3347 
3348 /// Populate the IC worklist from a function, and prune any dead basic
3349 /// blocks discovered in the process.
3350 ///
3351 /// This also does basic constant propagation and other forward fixing to make
3352 /// the combiner itself run much faster.
3354  TargetLibraryInfo *TLI,
3355  InstCombineWorklist &ICWorklist) {
3356  bool MadeIRChange = false;
3357 
3358  // Do a depth-first traversal of the function, populate the worklist with
3359  // the reachable instructions. Ignore blocks that are not reachable. Keep
3360  // track of which blocks we visit.
3362  MadeIRChange |=
3363  AddReachableCodeToWorklist(&F.front(), DL, Visited, ICWorklist, TLI);
3364 
3365  // Do a quick scan over the function. If we find any blocks that are
3366  // unreachable, remove any instructions inside of them. This prevents
3367  // the instcombine code from having to deal with some bad special cases.
3368  for (BasicBlock &BB : F) {
3369  if (Visited.count(&BB))
3370  continue;
3371 
3372  unsigned NumDeadInstInBB = removeAllNonTerminatorAndEHPadInstructions(&BB);
3373  MadeIRChange |= NumDeadInstInBB > 0;
3374  NumDeadInst += NumDeadInstInBB;
3375  }
3376 
3377  return MadeIRChange;
3378 }
3379 
3381  Function &F, InstCombineWorklist &Worklist, AliasAnalysis *AA,
3383  OptimizationRemarkEmitter &ORE, bool ExpensiveCombines = true,
3384  LoopInfo *LI = nullptr) {
3385  auto &DL = F.getParent()->getDataLayout();
3386  ExpensiveCombines |= EnableExpensiveCombines;
3387 
3388  /// Builder - This is an IRBuilder that automatically inserts new
3389  /// instructions into the worklist when they are created.
3391  F.getContext(), TargetFolder(DL),
3392  IRBuilderCallbackInserter([&Worklist, &AC](Instruction *I) {
3393  Worklist.Add(I);
3394  if (match(I, m_Intrinsic<Intrinsic::assume>()))
3395  AC.registerAssumption(cast<CallInst>(I));
3396  }));
3397 
3398  // Lower dbg.declare intrinsics otherwise their value may be clobbered
3399  // by instcombiner.
3400  bool MadeIRChange = false;
3402  MadeIRChange = LowerDbgDeclare(F);
3403 
3404  // Iterate while there is work to do.
3405  int Iteration = 0;
3406  while (true) {
3407  ++Iteration;
3408  LLVM_DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
3409  << F.getName() << "\n");
3410 
3411  MadeIRChange |= prepareICWorklistFromFunction(F, DL, &TLI, Worklist);
3412 
3413  InstCombiner IC(Worklist, Builder, F.optForMinSize(), ExpensiveCombines, AA,
3414  AC, TLI, DT, ORE, DL, LI);
3416 
3417  if (!IC.run())
3418  break;
3419  }
3420 
3421  return MadeIRChange || Iteration > 1;
3422 }
3423 
3426  auto &AC = AM.getResult<AssumptionAnalysis>(F);
3427  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
3428  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
3430 
3431  auto *LI = AM.getCachedResult<LoopAnalysis>(F);
3432 
3433  auto *AA = &AM.getResult<AAManager>(F);
3434  if (!combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT, ORE,
3435  ExpensiveCombines, LI))
3436  // No changes, all analyses are preserved.
3437  return PreservedAnalyses::all();
3438 
3439  // Mark all the analyses that instcombine updates as preserved.
3440  PreservedAnalyses PA;
3441  PA.preserveSet<CFGAnalyses>();
3442  PA.preserve<AAManager>();
3443  PA.preserve<BasicAA>();
3444  PA.preserve<GlobalsAA>();
3445  return PA;
3446 }
3447 
3449  AU.setPreservesCFG();
3459 }
3460 
3462  if (skipFunction(F))
3463  return false;
3464 
3465  // Required analyses.
3466  auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
3467  auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3468  auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
3469  auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3470  auto &ORE = getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
3471 
3472  // Optional analyses.
3473  auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
3474  auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
3475 
3476  return combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT, ORE,
3477  ExpensiveCombines, LI);
3478 }
3479 
3481 
3483  "Combine redundant instructions", false, false)
3491  "Combine redundant instructions", false, false)
3492 
3493 // Initialization Routines
3496 }
3497 
3500 }
3501 
3503  return new InstructionCombiningPass(ExpensiveCombines);
3504 }
3505 
3508 }
Legacy wrapper pass to provide the GlobalsAAResult object.
static unsigned getBitWidth(Type *Ty, const DataLayout &DL)
Returns the bitwidth of the given scalar or pointer type.
Type * getVectorElementType() const
Definition: Type.h:371
const NoneType None
Definition: None.h:24
A vector constant whose element type is a simple 1/2/4/8-byte integer or float/double, and whose elements are just simple data values (i.e.
Definition: Constants.h:758
Value * EmitGEPOffset(IRBuilderTy *Builder, const DataLayout &DL, User *GEP, bool NoAssumptions=false)
Given a getelementptr instruction/constantexpr, emit the code necessary to compute the offset from th...
Definition: Local.h:29
uint64_t CallInst * C
Return a value (possibly void), from a function.
void computeKnownBits(const Value *V, KnownBits &Known, const DataLayout &DL, unsigned Depth=0, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, OptimizationRemarkEmitter *ORE=nullptr, bool UseInstrInfo=true)
Determine which bits of V are known to be either zero or one and return them in the KnownZero/KnownOn...
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:68
void push_back(const T &Elt)
Definition: SmallVector.h:218
bool isAllocationFn(const Value *V, const TargetLibraryInfo *TLI, bool LookThroughBitCast=false)
Tests if a value is a call or invoke to a library function that allocates or reallocates memory (eith...
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:111
static bool prepareICWorklistFromFunction(Function &F, const DataLayout &DL, TargetLibraryInfo *TLI, InstCombineWorklist &ICWorklist)
Populate the IC worklist from a function, and prune any dead basic blocks discovered in the process...
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:584
void copyFastMathFlags(FastMathFlags FMF)
Convenience function for transferring all fast-math flag values to this instruction, which must be an operator which supports these flags.
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:72
This class is the base class for the comparison instructions.
Definition: InstrTypes.h:631
class_match< UndefValue > m_Undef()
Match an arbitrary undef constant.
Definition: PatternMatch.h:88
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
void setFastMathFlags(FastMathFlags FMF)
Convenience function for setting multiple fast-math flags on this instruction, which must be an opera...
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
static IntegerType * getInt1Ty(LLVMContext &C)
Definition: Type.cpp:173
class_match< CmpInst > m_Cmp()
Matches any compare instruction and ignore it.
Definition: PatternMatch.h:80
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
This instruction extracts a struct member or array element value from an aggregate value...
APInt sext(unsigned width) const
Sign extend to a new width.
Definition: APInt.cpp:834
Value * CreateBinOp(Instruction::BinaryOps Opc, Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1258
ThreeOps_match< Cond, LHS, RHS, Instruction::Select > m_Select(const Cond &C, const LHS &L, const RHS &R)
Matches SelectInst.
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
iterator_range< CaseIt > cases()
Iteration adapter for range-for loops.
BinaryOp_match< LHS, RHS, Instruction::Sub > m_Sub(const LHS &L, const RHS &R)
Definition: PatternMatch.h:651
is_zero m_Zero()
Match any null constant or a vector with all elements equal to 0.
Definition: PatternMatch.h:373
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:770
void LLVMInitializeInstCombine(LLVMPassRegistryRef R)
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
BinaryOps getOpcode() const
Definition: InstrTypes.h:311
void swapSuccessors()
Swap the successors of this branch instruction.
static bool isAllocSiteRemovable(Instruction *AI, SmallVectorImpl< WeakTrackingVH > &Users, const TargetLibraryInfo *TLI)
This is the interface for a simple mod/ref and alias analysis over globals.
A Module instance is used to store all the information related to an LLVM module. ...
Definition: Module.h:64
bool isSized(SmallPtrSetImpl< Type *> *Visited=nullptr) const
Return true if it makes sense to take the size of this type.
Definition: Type.h:265
#define LLVM_FALLTHROUGH
Definition: Compiler.h:86
static Value * foldOperationIntoPhiValue(BinaryOperator *I, Value *InV, InstCombiner::BuilderTy &Builder)
static void ClearSubclassDataAfterReassociation(BinaryOperator &I)
Conservatively clears subclassOptionalData after a reassociation or commutation.
static GetElementPtrInst * Create(Type *PointeeType, Value *Ptr, ArrayRef< Value *> IdxList, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Definition: Instructions.h:869
br_match m_UnconditionalBr(BasicBlock *&Succ)
A global registry used in conjunction with static constructors to make pluggable components (like tar...
Definition: Registry.h:45
void findDbgUsers(SmallVectorImpl< DbgVariableIntrinsic *> &DbgInsts, Value *V)
Finds the debug info intrinsics describing a value.
Definition: Local.cpp:1525
This class represents a function call, abstracting a target machine&#39;s calling convention.
This file contains the declarations for metadata subclasses.
An immutable pass that tracks lazily created AssumptionCache objects.
Value * getCondition() const
class_match< Constant > m_Constant()
Match an arbitrary Constant and ignore it.
Definition: PatternMatch.h:91
static bool isCanonicalPredicate(CmpInst::Predicate Pred)
Predicate canonicalization reduces the number of patterns that need to be matched by other transforms...
gep_type_iterator gep_type_end(const User *GEP)
const Value * getTrueValue() const
BinaryOp_match< LHS, RHS, Instruction::AShr > m_AShr(const LHS &L, const RHS &R)
Definition: PatternMatch.h:776
A cache of @llvm.assume calls within a function.
bool salvageDebugInfo(Instruction &I)
Assuming the instruction I is going to be deleted, attempt to salvage debug users of I by writing the...
Definition: Local.cpp:1609
static void sdivrem(const APInt &LHS, const APInt &RHS, APInt &Quotient, APInt &Remainder)
Definition: APInt.cpp:1837
This instruction constructs a fixed permutation of two input vectors.
static SelectInst * Create(Value *C, Value *S1, Value *S2, const Twine &NameStr="", Instruction *InsertBefore=nullptr, Instruction *MDFrom=nullptr)
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:714
bool isTerminator() const
Definition: Instruction.h:129
struct LLVMOpaquePassRegistry * LLVMPassRegistryRef
Definition: Types.h:131
void Add(Instruction *I)
Add - Add the specified instruction to the worklist if it isn&#39;t already in it.
BasicBlock * getSuccessor(unsigned i) const
static bool isEquality(Predicate P)
Return true if this predicate is either EQ or NE.
APInt trunc(unsigned width) const
Truncate to new width.
Definition: APInt.cpp:811
STATISTIC(NumFunctions, "Total number of functions")
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:231
F(f)
ThreeOps_match< V1_t, V2_t, Mask_t, Instruction::ShuffleVector > m_ShuffleVector(const V1_t &v1, const V2_t &v2, const Mask_t &m)
Matches ShuffleVectorInst.
unsigned getPointerAddressSpace() const
Get the address space of this pointer or pointer vector type.
Definition: DerivedTypes.h:503
An instruction for reading from memory.
Definition: Instructions.h:168
uint64_t MaxArraySizeForCombine
Maximum size of array considered when transforming.
static Constant * getCompare(unsigned short pred, Constant *C1, Constant *C2, bool OnlyIfReduced=false)
Return an ICmp or FCmp comparison operator constant expression.
Definition: Constants.cpp:1933
Hexagon Common GEP
Value * getCondition() const
static Constant * getSub(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2227
const Instruction * 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:138
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:230
iv Induction Variable Users
Definition: IVUsers.cpp:52
static bool willNotOverflow(IntrinsicInst *II, LazyValueInfo *LVI)
Constant * getClause(unsigned Idx) const
Get the value of the clause at index Idx.
This defines the Use class.
void reserve(size_type N)
Definition: SmallVector.h:376
idx_iterator idx_end() const
TinyPtrVector - This class is specialized for cases where there are normally 0 or 1 element in a vect...
Definition: TinyPtrVector.h:31
AnyBinaryOp_match< LHS, RHS, true > m_c_BinOp(const LHS &L, const RHS &R)
Matches a BinaryOperator with LHS and RHS in either order.
unsigned getBitWidth() const
Get the bit width of this value.
Definition: KnownBits.h:40
static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C)
const Value * DoPHITranslation(const BasicBlock *CurBB, const BasicBlock *PredBB) const
Translate PHI node to its predecessor from the given basic block.
Definition: Value.cpp:706
bool hasNoSignedWrap() const
Determine whether the no signed wrap flag is set.
static bool isBitwiseLogicOp(unsigned Opcode)
Determine if the Opcode is and/or/xor.
Definition: Instruction.h:173
op_iterator op_begin()
Definition: User.h:230
static Constant * get(ArrayType *T, ArrayRef< Constant *> V)
Definition: Constants.cpp:960
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
static LandingPadInst * Create(Type *RetTy, unsigned NumReservedClauses, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Constructors - NumReservedClauses is a hint for the number of incoming clauses that this landingpad w...
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1503
LLVMContext & getContext() const
Return the LLVMContext in which this type was uniqued.
Definition: Type.h:130
const CallInst * isFreeCall(const Value *I, const TargetLibraryInfo *TLI)
isFreeCall - Returns non-null if the value is a call to the builtin free()
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:268
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:263
bool isIdenticalTo(const Instruction *I) const
Return true if the specified instruction is exactly identical to the current one. ...
bool swapOperands()
Exchange the two operands to this instruction.
static GCMetadataPrinterRegistry::Add< OcamlGCMetadataPrinter > Y("ocaml", "ocaml 3.10-compatible collector")
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
static Instruction * tryToMoveFreeBeforeNullTest(CallInst &FI, const DataLayout &DL)
Move the call to free before a NULL test.
AnalysisUsage & addRequired()
ArrayRef< unsigned > getIndices() const
Value * SimplifyExtractValueInst(Value *Agg, ArrayRef< unsigned > Idxs, const SimplifyQuery &Q)
Given operands for an ExtractValueInst, fold the result or return null.
Used to lazily calculate structure layout information for a target machine, based on the DataLayout s...
Definition: DataLayout.h:521
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
This class represents a conversion between pointers from one address space to another.
static unsigned getComplexity(Value *V)
Assign a complexity or rank value to LLVM Values.
This class represents the LLVM &#39;select&#39; instruction.
Predicate getInversePredicate() const
For example, EQ -> NE, UGT -> ULE, SLT -> SGE, OEQ -> UNE, UGT -> OLE, OLT -> UGE, etc.
Definition: InstrTypes.h:740
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:364
Attribute unwrap(LLVMAttributeRef Attr)
Definition: Attributes.h:195
bool isNonNegative() const
Determine if this APInt Value is non-negative (>= 0)
Definition: APInt.h:369
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:348
ArrayRef< T > makeArrayRef(const T &OneElt)
Construct an ArrayRef from a single element.
Definition: ArrayRef.h:451
uint64_t getArrayNumElements() const
Definition: DerivedTypes.h:388
Class to represent struct types.
Definition: DerivedTypes.h:201
Value * SimplifyGEPInst(Type *SrcTy, ArrayRef< Value *> Ops, const SimplifyQuery &Q)
Given operands for a GetElementPtrInst, fold the result or return null.
A Use represents the edge between a Value definition and its users.
Definition: Use.h:56
static Optional< unsigned > getOpcode(ArrayRef< VPValue *> Values)
Returns the opcode of Values or ~0 if they do not all agree.
Definition: VPlanSLP.cpp:197
void setIsInBounds(bool b=true)
Set or clear the inbounds flag on this GEP instruction.
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:197
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:743
The core instruction combiner logic.
void LLVMAddInstructionCombiningPass(LLVMPassManagerRef PM)
See llvm::createInstructionCombiningPass function.
static cl::opt< bool > EnableExpensiveCombines("expensive-combines", cl::desc("Enable expensive instruction combines"))
Analysis pass that exposes the LoopInfo for a function.
Definition: LoopInfo.h:939
This file provides an implementation of debug counters.
static bool shorter_filter(const Value *LHS, const Value *RHS)
uint64_t getNumElements() const
Definition: DerivedTypes.h:359
Type * getSourceElementType() const
Definition: Instructions.h:940
unsigned getNumClauses() const
Get the number of clauses for this landing pad.
This file implements a class to represent arbitrary precision integral constant values and operations...
BinaryOp_match< LHS, RHS, Instruction::Add > m_Add(const LHS &L, const RHS &R)
Definition: PatternMatch.h:639
Instruction * visitReturnInst(ReturnInst &RI)
Interval::succ_iterator succ_begin(Interval *I)
succ_begin/succ_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:103
Instruction * visitBranchInst(BranchInst &BI)
bool isBitwiseLogicOp() const
Return true if this is and/or/xor.
Definition: Instruction.h:178
unsigned getNumIndices() const
Constant * ConstantFoldConstant(const Constant *C, const DataLayout &DL, const TargetLibraryInfo *TLI=nullptr)
ConstantFoldConstant - Attempt to fold the constant using the specified DataLayout.
static Value * getIdentityValue(Instruction::BinaryOps Opcode, Value *V)
This function returns identity value for given opcode, which can be used to factor patterns like (X *...
bool isNullValue() const
Return true if this is the value that would be returned by getNullValue.
Definition: Constants.cpp:85
int64_t getSExtValue() const
Get sign extended value.
Definition: APInt.h:1569
void setCleanup(bool V)
Indicate that this landingpad instruction is a cleanup.
FastMathFlags getFastMathFlags() const
Convenience function for getting all the fast-math flags, which must be an operator which supports th...
Instruction::CastOps getOpcode() const
Return the opcode of this CastInst.
Definition: InstrTypes.h:601
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
static bool isNeverEqualToUnescapedAlloc(Value *V, const TargetLibraryInfo *TLI, Instruction *AI)
match_combine_or< LTy, RTy > m_CombineOr(const LTy &L, const RTy &R)
Combine two pattern matchers matching L || R.
Definition: PatternMatch.h:125
bool isInBounds() const
Determine whether the GEP has the inbounds flag.
CastClass_match< OpTy, Instruction::ZExt > m_ZExt(const OpTy &Op)
Matches ZExt.
Class to represent array types.
Definition: DerivedTypes.h:369
static Constant * getSafeVectorConstantForBinop(BinaryOperator::BinaryOps Opcode, Constant *In, bool IsRHSConstant)
Some binary operators require special handling to avoid poison and undefined behavior.
int32_t exactLogBase2() const
Definition: APInt.h:1782
This class represents a no-op cast from one type to another.
op_iterator idx_begin()
Definition: Instructions.h:968
class_match< ConstantInt > m_ConstantInt()
Match an arbitrary ConstantInt and ignore it.
Definition: PatternMatch.h:83
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:126
TargetFolder - Create constants with target dependent folding.
Definition: TargetFolder.h:32
An instruction for storing to memory.
Definition: Instructions.h:310
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:203
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:439
FunctionPass * createInstructionCombiningPass(bool ExpensiveCombines=true)
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:301
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:145
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:1021
static BinaryOperator * CreateAdd(Value *S1, Value *S2, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp)
Value * getOperand(unsigned i) const
Definition: User.h:170
static Instruction::BinaryOps getBinOpsForFactorization(Instruction::BinaryOps TopOpcode, BinaryOperator *Op, Value *&LHS, Value *&RHS)
This function predicates factorization using distributive laws.
Class to represent pointers.
Definition: DerivedTypes.h:467
Interval::succ_iterator succ_end(Interval *I)
Definition: Interval.h:106
unsigned getAddressSpace() const
Returns the address space of this instruction&#39;s pointer type.
Definition: Instructions.h:952
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Constant * getAggregateElement(unsigned Elt) const
For aggregates (struct/array/vector) return the constant that corresponds to the specified element if...
Definition: Constants.cpp:338
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return &#39;this&#39;.
Definition: Type.h:304
Value * getOperand(unsigned i_nocapture) const
const BasicBlock & getEntryBlock() const
Definition: Function.h:640
an instruction for type-safe pointer arithmetic to access elements of arrays and structs ...
Definition: Instructions.h:843
OneUse_match< T > m_OneUse(const T &SubPattern)
Definition: PatternMatch.h:63
bool hasAllZeroIndices() const
Return true if all of the indices of this GEP are zeros.
Definition: Operator.h:501
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:419
The landingpad instruction holds all of the information necessary to generate correct exception handl...
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:154
apint_match m_APInt(const APInt *&Res)
Match a ConstantInt or splatted ConstantVector, binding the specified pointer to the contained APInt...
Definition: PatternMatch.h:177
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:219
BinaryOp_match< LHS, RHS, Instruction::SDiv > m_SDiv(const LHS &L, const RHS &R)
Definition: PatternMatch.h:716
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:236
bool run()
Run the combiner over the entire worklist until it is empty.
void setAAMetadata(const AAMDNodes &N)
Sets the metadata on this instruction from the AAMDNodes structure.
Definition: Metadata.cpp:1265
ConstantInt * lowerObjectSizeCall(IntrinsicInst *ObjectSize, const DataLayout &DL, const TargetLibraryInfo *TLI, bool MustSucceed)
Try to turn a call to @llvm.objectsize into an integer value of the given Type.
LLVM Basic Block Representation.
Definition: BasicBlock.h:58
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
Conditional or Unconditional Branch instruction.
static ExtractValueInst * Create(Value *Agg, ArrayRef< unsigned > Idxs, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo)
Return &#39;true&#39; if the given typeinfo will match anything.
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:129
const APInt & getConstant() const
Returns the value when all bits have a known value.
Definition: KnownBits.h:57
This file contains the declarations for the subclasses of Constant, which represent the different fla...
This file provides the primary interface to the instcombine pass.
A manager for alias analyses.
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:371
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:562
APInt ssub_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1888
Constant * ConstantFoldInstruction(Instruction *I, const DataLayout &DL, const TargetLibraryInfo *TLI=nullptr)
ConstantFoldInstruction - Try to constant fold the specified instruction.
size_t size() const
Definition: BasicBlock.h:273
unsigned getPointerAddressSpace() const
Returns the address space of the pointer operand.
Definition: Instructions.h:998
specificval_ty m_Specific(const Value *V)
Match if we have a specific specified value.
Definition: PatternMatch.h:499
void AddInitialGroup(ArrayRef< Instruction *> List)
AddInitialGroup - Add the specified batch of stuff in reverse order.
EHPersonality classifyEHPersonality(const Value *Pers)
See if the given exception handling personality function is one that we understand.
Value * SimplifyAddInst(Value *LHS, Value *RHS, bool isNSW, bool isNUW, const SimplifyQuery &Q)
Given operands for an Add, fold the result or return null.
brc_match< Cond_t > m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F)
bool isAssociative() const LLVM_READONLY
Return true if the instruction is associative:
bool isMinSignedValue() const
Determine if this is the smallest signed value.
Definition: APInt.h:443
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:232
BinaryOp_match< LHS, RHS, Instruction::Shl > m_Shl(const LHS &L, const RHS &R)
Definition: PatternMatch.h:764
bool isConstant() const
Returns true if we know the value of all bits.
Definition: KnownBits.h:50
This instruction compares its operands according to the predicate given to the constructor.
Analysis pass providing a never-invalidated alias analysis result.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:641
bool isBinaryOp() const
Definition: Instruction.h:131
Utility class for integer operators which may exhibit overflow - Add, Sub, Mul, and Shl...
Definition: Operator.h:67
void print(raw_ostream &O, bool IsForDebug=false) const
Implement operator<< on Value.
Definition: AsmWriter.cpp:4122
static Constant * getBinOpIdentity(unsigned Opcode, Type *Ty, bool AllowRHSConstant=false)
Return the identity constant for a binary opcode.
Definition: Constants.cpp:2304
match_combine_or< CastClass_match< OpTy, Instruction::ZExt >, CastClass_match< OpTy, Instruction::SExt > > m_ZExtOrSExt(const OpTy &Op)
void addClause(Constant *ClauseVal)
Add a catch or filter clause to the landing pad.
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:285
op_range operands()
Definition: User.h:238
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
static CastInst * CreatePointerBitCastOrAddrSpaceCast(Value *S, Type *Ty, const Twine &Name, BasicBlock *InsertAtEnd)
Create a BitCast or an AddrSpaceCast cast instruction.
unsigned getAddressSpace() const
Return the address space of the Pointer type.
Definition: DerivedTypes.h:495
bool isPotentiallyReachable(const Instruction *From, const Instruction *To, const DominatorTree *DT=nullptr, const LoopInfo *LI=nullptr)
Determine whether instruction &#39;To&#39; is reachable from &#39;From&#39;, returning true if uncertain.
Definition: CFG.cpp:187
self_iterator getIterator()
Definition: ilist_node.h:82
class_match< BinaryOperator > m_BinOp()
Match an arbitrary binary operation and ignore it.
Definition: PatternMatch.h:75
static bool shouldExecute(unsigned CounterName)
Definition: DebugCounter.h:74
Class to represent integer types.
Definition: DerivedTypes.h:40
Constant Vector Declarations.
Definition: Constants.h:496
The legacy pass manager&#39;s instcombine pass.
Definition: InstCombine.h:43
static Constant * getNot(Constant *C)
Definition: Constants.cpp:2210
void setSourceElementType(Type *Ty)
Definition: Instructions.h:942
static cl::opt< unsigned > MaxArraySize("instcombine-maxarray-size", cl::init(1024), cl::desc("Maximum array size considered when doing a combine"))
const Value * getCondition() const
LLVMContext & getContext() const
getContext - Return a reference to the LLVMContext associated with this function. ...
Definition: Function.cpp:194
Type * getPointerOperandType() const
Method to return the pointer operand as a PointerType.
Definition: Instructions.h:993
InstCombineWorklist - This is the worklist management logic for InstCombine.
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1392
const Constant * stripPointerCasts() const
Definition: Constant.h:173
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs, and aliases.
Definition: Value.cpp:539
iterator erase(const_iterator CI)
Definition: SmallVector.h:445
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:160
bool runOnFunction(Function &F) override
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass...
size_t size() const
Definition: SmallVector.h:53
static wasm::ValType getType(const TargetRegisterClass *RC)
Value * getIncomingValue(unsigned i) const
Return incoming value number x.
bool LowerDbgDeclare(Function &F)
Lowers llvm.dbg.declare intrinsics into appropriate set of llvm.dbg.value intrinsics.
Definition: Local.cpp:1391
bool isVolatile() const
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
static InvokeInst * Create(Value *Func, BasicBlock *IfNormal, BasicBlock *IfException, ArrayRef< Value *> Args, const Twine &NameStr, Instruction *InsertBefore=nullptr)
bool isFilter(unsigned Idx) const
Return &#39;true&#39; if the clause and index Idx is a filter clause.
static Constant * getIntegerValue(Type *Ty, const APInt &V)
Return the value for an integer or pointer constant, or a vector thereof, with the given scalar value...
Definition: Constants.cpp:305
CastClass_match< OpTy, Instruction::SExt > m_SExt(const OpTy &Op)
Matches SExt.
A function analysis which provides an AssumptionCache.
const InstListType & getInstList() const
Return the underlying instruction list container.
Definition: BasicBlock.h:328
static IntegerType * get(LLVMContext &C, unsigned NumBits)
This static method is the primary way of constructing an IntegerType.
Definition: Type.cpp:240
void setHasNoSignedWrap(bool b=true)
Set or clear the nsw flag on this instruction, which must be an operator which supports this flag...
This is the common base class for memset/memcpy/memmove.
Iterator for intrusive lists based on ilist_node.
unsigned getNumOperands() const
Definition: User.h:192
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
static bool simplifyAssocCastAssoc(BinaryOperator *BinOp1)
Combine constant operands of associative operations either before or after a cast to eliminate one of...
TinyPtrVector< DbgVariableIntrinsic * > FindDbgAddrUses(Value *V)
Finds all intrinsics declaring local variables as living in the memory that &#39;V&#39; points to...
Definition: Local.cpp:1491
BlockVerifier::State From
iterator end()
Definition: BasicBlock.h:265
unsigned getScalarSizeInBits() const LLVM_READONLY
If this is a vector type, return the getPrimitiveSizeInBits value for the element type...
Definition: Type.cpp:130
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:847
Utility class for floating point operations which can have information about relaxed accuracy require...
Definition: Operator.h:240
This is a utility class that provides an abstraction for the common functionality between Instruction...
Definition: Operator.h:31
Instruction * user_back()
Specialize the methods defined in Value, as we know that an instruction can only be used by other ins...
Definition: Instruction.h:64
Provides information about what library functions are available for the current target.
static cl::opt< bool > EnableCodeSinking("instcombine-code-sinking", cl::desc("Enable code sinking"), cl::init(true))
A collection of metadata nodes that might be associated with a memory access used by the alias-analys...
Definition: Metadata.h:644
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:381
uint64_t getSizeInBytes() const
Definition: DataLayout.h:529
Instruction * visitFree(CallInst &FI)
static Constant * getTrunc(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1614
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:621
bool isConditional() const
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...
void setPreservesCFG()
This function should be called by the pass, iff they do not:
Definition: Pass.cpp:286
void initializeInstCombine(PassRegistry &)
Initialize all passes linked into the InstCombine library.
unsigned removeAllNonTerminatorAndEHPadInstructions(BasicBlock *BB)
Remove all instructions from a basic block other than it&#39;s terminator and any present EH pad instruct...
Definition: Local.cpp:1893
unsigned getNumIncomingValues() const
Return the number of incoming edges.
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:577
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:478
static bool AddReachableCodeToWorklist(BasicBlock *BB, const DataLayout &DL, SmallPtrSetImpl< BasicBlock *> &Visited, InstCombineWorklist &ICWorklist, const TargetLibraryInfo *TLI)
Walk the function in depth-first order, adding all reachable code to the worklist.
void setPredicate(Predicate P)
Set the predicate for this instruction to the specified value.
Definition: InstrTypes.h:721
void setOperand(unsigned i, Value *Val)
Definition: User.h:175
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:133
BinaryOp_match< cst_pred_ty< is_zero_int >, ValTy, Instruction::Sub > m_Neg(const ValTy &V)
Matches a &#39;Neg&#39; as &#39;sub 0, V&#39;.
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:941
unsigned getVectorNumElements() const
Definition: DerivedTypes.h:462
struct LLVMOpaquePassManager * LLVMPassManagerRef
Definition: Types.h:128
static Value * foldOperationIntoSelectOperand(Instruction &I, Value *SO, InstCombiner::BuilderTy &Builder)
Class to represent vector types.
Definition: DerivedTypes.h:393
const Value * stripAndAccumulateInBoundsConstantOffsets(const DataLayout &DL, APInt &Offset) const
Accumulate offsets from stripInBoundsConstantOffsets().
Definition: Value.cpp:557
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:56
ConstantArray - Constant Array Declarations.
Definition: Constants.h:410
Class for arbitrary precision integers.
Definition: APInt.h:70
bool ule(const APInt &RHS) const
Unsigned less or equal comparison.
Definition: APInt.h:1217
bool isCleanup() const
Return &#39;true&#39; if this landingpad instruction is a cleanup.
static BinaryOperator * Create(BinaryOps Op, Value *S1, Value *S2, const Twine &Name=Twine(), Instruction *InsertBefore=nullptr)
Construct a binary instruction, given the opcode and the two operands.
CastClass_match< OpTy, Instruction::PtrToInt > m_PtrToInt(const OpTy &Op)
Matches PtrToInt.
typename SuperClass::iterator iterator
Definition: SmallVector.h:327
iterator_range< user_iterator > users()
Definition: Value.h:400
bool hasNoSignedWrap() const
Test whether this operation is known to never undergo signed overflow, aka the nsw property...
Definition: Operator.h:96
Instruction * visitSwitchInst(SwitchInst &SI)
static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock)
Try to move the specified instruction from its current block into the beginning of DestBlock...
Instruction * visitExtractValueInst(ExtractValueInst &EV)
static Constant * getCast(unsigned ops, Constant *C, Type *Ty, bool OnlyIfReduced=false)
Convenience function for getting a Cast operation.
Definition: Constants.cpp:1507
Represents analyses that only rely on functions&#39; control flow.
Definition: PassManager.h:115
unsigned countMinLeadingOnes() const
Returns the minimum number of leading one bits.
Definition: KnownBits.h:151
const Value * getFalseValue() const
void append(in_iter in_start, in_iter in_end)
Add the specified range to the end of the SmallVector.
Definition: SmallVector.h:394
Instruction * visitLandingPadInst(LandingPadInst &LI)
use_iterator use_begin()
Definition: Value.h:339
static Constant * getNeg(Constant *C, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2197
Analysis pass providing a never-invalidated alias analysis result.
static bool leftDistributesOverRight(Instruction::BinaryOps LOp, Instruction::BinaryOps ROp)
Return whether "X LOp (Y ROp Z)" is always equal to "(X LOp Y) ROp (X LOp Z)".
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 ...
This file provides various utilities for inspecting and working with the control flow graph in LLVM I...
Provides an &#39;InsertHelper&#39; that calls a user-provided callback after performing the default insertion...
Definition: IRBuilder.h:73
bool isVolatile() const
Return true if this is a store to a volatile memory location.
Definition: Instructions.h:343
iterator insert(iterator where, pointer New)
Definition: ilist.h:228
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:311
void registerAssumption(CallInst *CI)
Add an @llvm.assume intrinsic to this function&#39;s cache.
static bool combineInstructionsOverFunction(Function &F, InstCombineWorklist &Worklist, AliasAnalysis *AA, AssumptionCache &AC, TargetLibraryInfo &TLI, DominatorTree &DT, OptimizationRemarkEmitter &ORE, bool ExpensiveCombines=true, LoopInfo *LI=nullptr)
uint64_t getElementOffset(unsigned Idx) const
Definition: DataLayout.h:543
void emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:652
static IntegerType * getInt32Ty(LLVMContext &C)
Definition: Type.cpp:176
void setCondition(Value *V)
bool accumulateConstantOffset(const DataLayout &DL, APInt &Offset) const
Accumulate the constant address offset of this GEP if possible.
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:56
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:459
static InsertValueInst * Create(Value *Agg, Value *Val, ArrayRef< unsigned > Idxs, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
void preserveSet()
Mark an analysis set as preserved.
Definition: PassManager.h:190
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:224
BasicBlock * getIncomingBlock(unsigned i) const
Return incoming basic block number i.
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:107
#define I(x, y, z)
Definition: MD5.cpp:58
bool isCatch(unsigned Idx) const
Return &#39;true&#39; if the clause and index Idx is a catch clause.
bool mayReadFromMemory() const
Return true if this instruction may read memory.
bool optForMinSize() const
Optimize this function for minimum size (-Oz).
Definition: Function.h:595
bool isAllocLikeFn(const Value *V, const TargetLibraryInfo *TLI, bool LookThroughBitCast=false)
Tests if a value is a call or invoke to a library function that allocates memory (either malloc...
PassT::Result * getCachedResult(IRUnitT &IR) const
Get the cached result of an analysis pass for a given IR unit.
Definition: PassManager.h:789
static ArrayType * get(Type *ElementType, uint64_t NumElements)
This static method is the primary way to construct an ArrayType.
Definition: Type.cpp:581
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
idx_iterator idx_begin() const
static Constant * getShl(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2287
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:175
DEBUG_COUNTER(VisitCounter, "instcombine-visit", "Controls which instructions are visited")
bool isUnconditional() const
void initializeInstructionCombiningPassPass(PassRegistry &)
static cl::opt< unsigned > ShouldLowerDbgDeclare("instcombine-lower-dbg-declare", cl::Hidden, cl::init(true))
bool isMinValue() const
Determine if this is the smallest unsigned value.
Definition: APInt.h:437
void setCondition(Value *V)
Analysis pass providing the TargetLibraryInfo.
Multiway switch.
Value * CreateCast(Instruction::CastOps Op, Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1678
Value * SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, const SimplifyQuery &Q)
Given operands for a BinaryOperator, fold the result or return null.
static GetElementPtrInst * CreateInBounds(Value *Ptr, ArrayRef< Value *> IdxList, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Create an "inbounds" getelementptr.
Definition: Instructions.h:903
INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine", "Combine redundant instructions", false, false) INITIALIZE_PASS_END(InstructionCombiningPass
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
user_iterator user_begin()
Definition: Value.h:376
const BasicBlock & front() const
Definition: Function.h:663
SmallVector< int, 16 > getShuffleMask() const
Value * getArgOperand(unsigned i) const
getArgOperand/setArgOperand - Return/set the i-th call argument.
bool isSafeToSpeculativelyExecute(const Value *V, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr)
Return true if the instruction does not have any effects besides calculating the result and does not ...
A raw_ostream that writes to an std::string.
Definition: raw_ostream.h:483
APInt sadd_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1875
unsigned getPrimitiveSizeInBits() const LLVM_READONLY
Return the basic size of this type if it is a primitive type.
Definition: Type.cpp:115
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:566
bool isInstructionTriviallyDead(Instruction *I, const TargetLibraryInfo *TLI=nullptr)
Return true if the result produced by the instruction is not used, and the instruction has no side ef...
Definition: Local.cpp:348
LLVM Value Representation.
Definition: Value.h:73
Constant * getPersonalityFn() const
Get the personality function associated with this function.
Definition: Function.cpp:1302
This file provides internal interfaces used to implement the InstCombine.
void clearSubclassOptionalData()
Clear the optional flags contained in this value.
Definition: Value.h:476
succ_range successors(Instruction *I)
Definition: CFG.h:262
static VectorType * get(Type *ElementType, unsigned NumElements)
This static method is the primary way to construct an VectorType.
Definition: Type.cpp:606
OptimizationRemarkEmitter legacy analysis pass.
std::underlying_type< E >::type Mask()
Get a bitmask with 1s in all places up to the high-order bit of E&#39;s largest value.
Definition: BitmaskEnum.h:81
void moveBefore(Instruction *MovePos)
Unlink this instruction from its current basic block and insert it into the basic block that MovePos ...
Definition: Instruction.cpp:87
void ConvertDebugDeclareToDebugValue(DbgVariableIntrinsic *DII, StoreInst *SI, DIBuilder &Builder)
===---------------------------------------------------------------——===// Dbg Intrinsic utilities ...
Definition: Local.cpp:1268
Invoke instruction.
Instruction * visitGetElementPtrInst(GetElementPtrInst &GEP)
unsigned countMinLeadingZeros() const
Returns the minimum number of leading zero bits.
Definition: KnownBits.h:146
bool isEHPad() const
Return true if the instruction is a variety of EH-block.
Definition: Instruction.h:573
Type * getElementType() const
Definition: DerivedTypes.h:360
IRTranslator LLVM IR MI
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:413
Convenience struct for specifying and reasoning about fast-math flags.
Definition: Operator.h:160
unsigned greater than
Definition: InstrTypes.h:664
bool isIntDivRem() const
Definition: Instruction.h:132
This is the interface for LLVM&#39;s primary stateless and local alias analysis.
inst_range instructions(Function *F)
Definition: InstIterator.h:134
PassRegistry - This class manages the registration and intitialization of the pass subsystem as appli...
Definition: PassRegistry.h:39
A container for analyses that lazily runs them and caches their results.
Type * getArrayElementType() const
Definition: Type.h:365
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:260
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
This header defines various interfaces for pass management in LLVM.
static BinaryOperator * CreateMul(Value *S1, Value *S2, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp)
static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src)
#define LLVM_DEBUG(X)
Definition: Debug.h:123
Value * getPointerOperand()
Definition: Instructions.h:402
Instruction * visitAllocSite(Instruction &FI)
The optimization diagnostic interface.
Value * getRawDest() const
bool use_empty() const
Definition: Value.h:323
static Constant * get(ArrayRef< Constant *> V)
Definition: Constants.cpp:1056
Type * getElementType() const
Definition: DerivedTypes.h:486
static bool rightDistributesOverLeft(Instruction::BinaryOps LOp, Instruction::BinaryOps ROp)
Return whether "(X LOp Y) ROp Z" is always equal to "(X ROp Z) LOp (Y ROp Z)".
BinaryOp_match< ValTy, cst_pred_ty< is_all_ones >, Instruction::Xor, true > m_Not(const ValTy &V)
Matches a &#39;Not&#39; as &#39;xor V, -1&#39; or &#39;xor -1, V&#39;.
bool isStructTy() const
True if this is an instance of StructType.
Definition: Type.h:218
bool isArrayTy() const
True if this is an instance of ArrayType.
Definition: Type.h:221
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:67
This instruction inserts a struct field of array element value into an aggregate value.
CmpClass_match< LHS, RHS, ICmpInst, ICmpInst::Predicate > m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R)
Legacy wrapper pass to provide the BasicAAResult object.
gep_type_iterator gep_type_begin(const User *GEP)
static Constant * get(unsigned Opcode, Constant *C1, unsigned Flags=0, Type *OnlyIfReducedTy=nullptr)
get - Return a unary operator constant expression, folding if possible.
Definition: Constants.cpp:1783
user_iterator user_end()
Definition: Value.h:384