LLVM  6.0.0svn
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"
38 #include "llvm/ADT/SmallPtrSet.h"
39 #include "llvm/ADT/Statistic.h"
40 #include "llvm/ADT/StringSwitch.h"
44 #include "llvm/Analysis/CFG.h"
49 #include "llvm/Analysis/LoopInfo.h"
53 #include "llvm/IR/CFG.h"
54 #include "llvm/IR/DataLayout.h"
55 #include "llvm/IR/Dominators.h"
57 #include "llvm/IR/IntrinsicInst.h"
58 #include "llvm/IR/PatternMatch.h"
59 #include "llvm/IR/ValueHandle.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/KnownBits.h"
65 #include "llvm/Transforms/Scalar.h"
67 #include <algorithm>
68 #include <climits>
69 using namespace llvm;
70 using namespace llvm::PatternMatch;
71 
72 #define DEBUG_TYPE "instcombine"
73 
74 STATISTIC(NumCombined , "Number of insts combined");
75 STATISTIC(NumConstProp, "Number of constant folds");
76 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
77 STATISTIC(NumSunkInst , "Number of instructions sunk");
78 STATISTIC(NumExpand, "Number of expansions");
79 STATISTIC(NumFactor , "Number of factorizations");
80 STATISTIC(NumReassoc , "Number of reassociations");
81 
82 static cl::opt<bool>
83 EnableExpensiveCombines("expensive-combines",
84  cl::desc("Enable expensive instruction combines"));
85 
86 static cl::opt<unsigned>
87 MaxArraySize("instcombine-maxarray-size", cl::init(1024),
88  cl::desc("Maximum array size considered when doing a combine"));
89 
90 Value *InstCombiner::EmitGEPOffset(User *GEP) {
91  return llvm::EmitGEPOffset(&Builder, DL, GEP);
92 }
93 
94 /// Return true if it is desirable to convert an integer computation from a
95 /// given bit width to a new bit width.
96 /// We don't want to convert from a legal to an illegal type or from a smaller
97 /// to a larger illegal type. A width of '1' is always treated as a legal type
98 /// because i1 is a fundamental type in IR, and there are many specialized
99 /// optimizations for i1 types.
100 bool InstCombiner::shouldChangeType(unsigned FromWidth,
101  unsigned ToWidth) const {
102  bool FromLegal = FromWidth == 1 || DL.isLegalInteger(FromWidth);
103  bool ToLegal = ToWidth == 1 || DL.isLegalInteger(ToWidth);
104 
105  // If this is a legal integer from type, and the result would be an illegal
106  // type, don't do the transformation.
107  if (FromLegal && !ToLegal)
108  return false;
109 
110  // Otherwise, if both are illegal, do not increase the size of the result. We
111  // do allow things like i160 -> i64, but not i64 -> i160.
112  if (!FromLegal && !ToLegal && ToWidth > FromWidth)
113  return false;
114 
115  return true;
116 }
117 
118 /// Return true if it is desirable to convert a computation from 'From' to 'To'.
119 /// We don't want to convert from a legal to an illegal type or from a smaller
120 /// to a larger illegal type. i1 is always treated as a legal type because it is
121 /// a fundamental type in IR, and there are many specialized optimizations for
122 /// i1 types.
123 bool InstCombiner::shouldChangeType(Type *From, Type *To) const {
124  assert(From->isIntegerTy() && To->isIntegerTy());
125 
126  unsigned FromWidth = From->getPrimitiveSizeInBits();
127  unsigned ToWidth = To->getPrimitiveSizeInBits();
128  return shouldChangeType(FromWidth, ToWidth);
129 }
130 
131 // Return true, if No Signed Wrap should be maintained for I.
132 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
133 // where both B and C should be ConstantInts, results in a constant that does
134 // not overflow. This function only handles the Add and Sub opcodes. For
135 // all other opcodes, the function conservatively returns false.
138  if (!OBO || !OBO->hasNoSignedWrap())
139  return false;
140 
141  // We reason about Add and Sub Only.
142  Instruction::BinaryOps Opcode = I.getOpcode();
143  if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
144  return false;
145 
146  const APInt *BVal, *CVal;
147  if (!match(B, m_APInt(BVal)) || !match(C, m_APInt(CVal)))
148  return false;
149 
150  bool Overflow = false;
151  if (Opcode == Instruction::Add)
152  (void)BVal->sadd_ov(*CVal, Overflow);
153  else
154  (void)BVal->ssub_ov(*CVal, Overflow);
155 
156  return !Overflow;
157 }
158 
159 /// Conservatively clears subclassOptionalData after a reassociation or
160 /// commutation. We preserve fast-math flags when applicable as they can be
161 /// preserved.
164  if (!FPMO) {
166  return;
167  }
168 
171  I.setFastMathFlags(FMF);
172 }
173 
174 /// Combine constant operands of associative operations either before or after a
175 /// cast to eliminate one of the associative operations:
176 /// (op (cast (op X, C2)), C1) --> (cast (op X, op (C1, C2)))
177 /// (op (cast (op X, C2)), C1) --> (op (cast X), op (C1, C2))
179  auto *Cast = dyn_cast<CastInst>(BinOp1->getOperand(0));
180  if (!Cast || !Cast->hasOneUse())
181  return false;
182 
183  // TODO: Enhance logic for other casts and remove this check.
184  auto CastOpcode = Cast->getOpcode();
185  if (CastOpcode != Instruction::ZExt)
186  return false;
187 
188  // TODO: Enhance logic for other BinOps and remove this check.
189  if (!BinOp1->isBitwiseLogicOp())
190  return false;
191 
192  auto AssocOpcode = BinOp1->getOpcode();
193  auto *BinOp2 = dyn_cast<BinaryOperator>(Cast->getOperand(0));
194  if (!BinOp2 || !BinOp2->hasOneUse() || BinOp2->getOpcode() != AssocOpcode)
195  return false;
196 
197  Constant *C1, *C2;
198  if (!match(BinOp1->getOperand(1), m_Constant(C1)) ||
199  !match(BinOp2->getOperand(1), m_Constant(C2)))
200  return false;
201 
202  // TODO: This assumes a zext cast.
203  // Eg, if it was a trunc, we'd cast C1 to the source type because casting C2
204  // to the destination type might lose bits.
205 
206  // Fold the constants together in the destination type:
207  // (op (cast (op X, C2)), C1) --> (op (cast X), FoldedC)
208  Type *DestTy = C1->getType();
209  Constant *CastC2 = ConstantExpr::getCast(CastOpcode, C2, DestTy);
210  Constant *FoldedC = ConstantExpr::get(AssocOpcode, C1, CastC2);
211  Cast->setOperand(0, BinOp2->getOperand(0));
212  BinOp1->setOperand(1, FoldedC);
213  return true;
214 }
215 
216 /// This performs a few simplifications for operators that are associative or
217 /// commutative:
218 ///
219 /// Commutative operators:
220 ///
221 /// 1. Order operands such that they are listed from right (least complex) to
222 /// left (most complex). This puts constants before unary operators before
223 /// binary operators.
224 ///
225 /// Associative operators:
226 ///
227 /// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
228 /// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
229 ///
230 /// Associative and commutative operators:
231 ///
232 /// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
233 /// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
234 /// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
235 /// if C1 and C2 are constants.
236 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
237  Instruction::BinaryOps Opcode = I.getOpcode();
238  bool Changed = false;
239 
240  do {
241  // Order operands such that they are listed from right (least complex) to
242  // left (most complex). This puts constants before unary operators before
243  // binary operators.
244  if (I.isCommutative() && getComplexity(I.getOperand(0)) <
246  Changed = !I.swapOperands();
247 
250 
251  if (I.isAssociative()) {
252  // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
253  if (Op0 && Op0->getOpcode() == Opcode) {
254  Value *A = Op0->getOperand(0);
255  Value *B = Op0->getOperand(1);
256  Value *C = I.getOperand(1);
257 
258  // Does "B op C" simplify?
259  if (Value *V = SimplifyBinOp(Opcode, B, C, SQ.getWithInstruction(&I))) {
260  // It simplifies to V. Form "A op V".
261  I.setOperand(0, A);
262  I.setOperand(1, V);
263  // Conservatively clear the optional flags, since they may not be
264  // preserved by the reassociation.
265  if (MaintainNoSignedWrap(I, B, C) &&
266  (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
267  // Note: this is only valid because SimplifyBinOp doesn't look at
268  // the operands to Op0.
270  I.setHasNoSignedWrap(true);
271  } else {
273  }
274 
275  Changed = true;
276  ++NumReassoc;
277  continue;
278  }
279  }
280 
281  // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
282  if (Op1 && Op1->getOpcode() == Opcode) {
283  Value *A = I.getOperand(0);
284  Value *B = Op1->getOperand(0);
285  Value *C = Op1->getOperand(1);
286 
287  // Does "A op B" simplify?
288  if (Value *V = SimplifyBinOp(Opcode, A, B, SQ.getWithInstruction(&I))) {
289  // It simplifies to V. Form "V op C".
290  I.setOperand(0, V);
291  I.setOperand(1, C);
292  // Conservatively clear the optional flags, since they may not be
293  // preserved by the reassociation.
295  Changed = true;
296  ++NumReassoc;
297  continue;
298  }
299  }
300  }
301 
302  if (I.isAssociative() && I.isCommutative()) {
303  if (simplifyAssocCastAssoc(&I)) {
304  Changed = true;
305  ++NumReassoc;
306  continue;
307  }
308 
309  // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
310  if (Op0 && Op0->getOpcode() == Opcode) {
311  Value *A = Op0->getOperand(0);
312  Value *B = Op0->getOperand(1);
313  Value *C = I.getOperand(1);
314 
315  // Does "C op A" simplify?
316  if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
317  // It simplifies to V. Form "V op B".
318  I.setOperand(0, V);
319  I.setOperand(1, B);
320  // Conservatively clear the optional flags, since they may not be
321  // preserved by the reassociation.
323  Changed = true;
324  ++NumReassoc;
325  continue;
326  }
327  }
328 
329  // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
330  if (Op1 && Op1->getOpcode() == Opcode) {
331  Value *A = I.getOperand(0);
332  Value *B = Op1->getOperand(0);
333  Value *C = Op1->getOperand(1);
334 
335  // Does "C op A" simplify?
336  if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
337  // It simplifies to V. Form "B op V".
338  I.setOperand(0, B);
339  I.setOperand(1, V);
340  // Conservatively clear the optional flags, since they may not be
341  // preserved by the reassociation.
343  Changed = true;
344  ++NumReassoc;
345  continue;
346  }
347  }
348 
349  // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
350  // if C1 and C2 are constants.
351  if (Op0 && Op1 &&
352  Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
353  isa<Constant>(Op0->getOperand(1)) &&
354  isa<Constant>(Op1->getOperand(1)) &&
355  Op0->hasOneUse() && Op1->hasOneUse()) {
356  Value *A = Op0->getOperand(0);
357  Constant *C1 = cast<Constant>(Op0->getOperand(1));
358  Value *B = Op1->getOperand(0);
359  Constant *C2 = cast<Constant>(Op1->getOperand(1));
360 
361  Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
362  BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
363  if (isa<FPMathOperator>(New)) {
365  Flags &= Op0->getFastMathFlags();
366  Flags &= Op1->getFastMathFlags();
367  New->setFastMathFlags(Flags);
368  }
369  InsertNewInstWith(New, I);
370  New->takeName(Op1);
371  I.setOperand(0, New);
372  I.setOperand(1, Folded);
373  // Conservatively clear the optional flags, since they may not be
374  // preserved by the reassociation.
376 
377  Changed = true;
378  continue;
379  }
380  }
381 
382  // No further simplifications.
383  return Changed;
384  } while (1);
385 }
386 
387 /// Return whether "X LOp (Y ROp Z)" is always equal to
388 /// "(X LOp Y) ROp (X LOp Z)".
391  switch (LOp) {
392  default:
393  return false;
394 
395  case Instruction::And:
396  // And distributes over Or and Xor.
397  switch (ROp) {
398  default:
399  return false;
400  case Instruction::Or:
401  case Instruction::Xor:
402  return true;
403  }
404 
405  case Instruction::Mul:
406  // Multiplication distributes over addition and subtraction.
407  switch (ROp) {
408  default:
409  return false;
410  case Instruction::Add:
411  case Instruction::Sub:
412  return true;
413  }
414 
415  case Instruction::Or:
416  // Or distributes over And.
417  switch (ROp) {
418  default:
419  return false;
420  case Instruction::And:
421  return true;
422  }
423  }
424 }
425 
426 /// Return whether "(X LOp Y) ROp Z" is always equal to
427 /// "(X ROp Z) LOp (Y ROp Z)".
431  return LeftDistributesOverRight(ROp, LOp);
432 
433  switch (LOp) {
434  default:
435  return false;
436  // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
437  // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
438  // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
439  case Instruction::And:
440  case Instruction::Or:
441  case Instruction::Xor:
442  switch (ROp) {
443  default:
444  return false;
445  case Instruction::Shl:
446  case Instruction::LShr:
447  case Instruction::AShr:
448  return true;
449  }
450  }
451  // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
452  // but this requires knowing that the addition does not overflow and other
453  // such subtleties.
454  return false;
455 }
456 
457 /// This function returns identity value for given opcode, which can be used to
458 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
460  if (isa<Constant>(V))
461  return nullptr;
462 
463  return ConstantExpr::getBinOpIdentity(Opcode, V->getType());
464 }
465 
466 /// This function factors binary ops which can be combined using distributive
467 /// laws. This function tries to transform 'Op' based TopLevelOpcode to enable
468 /// factorization e.g for ADD(SHL(X , 2), MUL(X, 5)), When this function called
469 /// with TopLevelOpcode == Instruction::Add and Op = SHL(X, 2), transforms
470 /// SHL(X, 2) to MUL(X, 4) i.e. returns Instruction::Mul with LHS set to 'X' and
471 /// RHS to 4.
474  BinaryOperator *Op, Value *&LHS, Value *&RHS) {
475  assert(Op && "Expected a binary operator");
476 
477  LHS = Op->getOperand(0);
478  RHS = Op->getOperand(1);
479 
480  switch (TopLevelOpcode) {
481  default:
482  return Op->getOpcode();
483 
484  case Instruction::Add:
485  case Instruction::Sub:
486  if (Op->getOpcode() == Instruction::Shl) {
487  if (Constant *CST = dyn_cast<Constant>(Op->getOperand(1))) {
488  // The multiplier is really 1 << CST.
489  RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), CST);
490  return Instruction::Mul;
491  }
492  }
493  return Op->getOpcode();
494  }
495 
496  // TODO: We can add other conversions e.g. shr => div etc.
497 }
498 
499 /// This tries to simplify binary operations by factorizing out common terms
500 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
501 Value *InstCombiner::tryFactorization(BinaryOperator &I,
502  Instruction::BinaryOps InnerOpcode,
503  Value *A, Value *B, Value *C, Value *D) {
504  assert(A && B && C && D && "All values must be provided");
505 
506  Value *V = nullptr;
507  Value *SimplifiedInst = nullptr;
508  Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
509  Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
510 
511  // Does "X op' Y" always equal "Y op' X"?
512  bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
513 
514  // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
515  if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
516  // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
517  // commutative case, "(A op' B) op (C op' A)"?
518  if (A == C || (InnerCommutative && A == D)) {
519  if (A != C)
520  std::swap(C, D);
521  // Consider forming "A op' (B op D)".
522  // If "B op D" simplifies then it can be formed with no cost.
523  V = SimplifyBinOp(TopLevelOpcode, B, D, SQ.getWithInstruction(&I));
524  // If "B op D" doesn't simplify then only go on if both of the existing
525  // operations "A op' B" and "C op' D" will be zapped as no longer used.
526  if (!V && LHS->hasOneUse() && RHS->hasOneUse())
527  V = Builder.CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
528  if (V) {
529  SimplifiedInst = Builder.CreateBinOp(InnerOpcode, A, V);
530  }
531  }
532 
533  // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
534  if (!SimplifiedInst && RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
535  // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
536  // commutative case, "(A op' B) op (B op' D)"?
537  if (B == D || (InnerCommutative && B == C)) {
538  if (B != D)
539  std::swap(C, D);
540  // Consider forming "(A op C) op' B".
541  // If "A op C" simplifies then it can be formed with no cost.
542  V = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
543 
544  // If "A op C" doesn't simplify then only go on if both of the existing
545  // operations "A op' B" and "C op' D" will be zapped as no longer used.
546  if (!V && LHS->hasOneUse() && RHS->hasOneUse())
547  V = Builder.CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
548  if (V) {
549  SimplifiedInst = Builder.CreateBinOp(InnerOpcode, V, B);
550  }
551  }
552 
553  if (SimplifiedInst) {
554  ++NumFactor;
555  SimplifiedInst->takeName(&I);
556 
557  // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
558  // TODO: Check for NUW.
559  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
560  if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
561  bool HasNSW = false;
562  if (isa<OverflowingBinaryOperator>(&I))
563  HasNSW = I.hasNoSignedWrap();
564 
565  if (auto *LOBO = dyn_cast<OverflowingBinaryOperator>(LHS))
566  HasNSW &= LOBO->hasNoSignedWrap();
567 
568  if (auto *ROBO = dyn_cast<OverflowingBinaryOperator>(RHS))
569  HasNSW &= ROBO->hasNoSignedWrap();
570 
571  // We can propagate 'nsw' if we know that
572  // %Y = mul nsw i16 %X, C
573  // %Z = add nsw i16 %Y, %X
574  // =>
575  // %Z = mul nsw i16 %X, C+1
576  //
577  // iff C+1 isn't INT_MIN
578  const APInt *CInt;
579  if (TopLevelOpcode == Instruction::Add &&
580  InnerOpcode == Instruction::Mul)
581  if (match(V, m_APInt(CInt)) && !CInt->isMinSignedValue())
582  BO->setHasNoSignedWrap(HasNSW);
583  }
584  }
585  }
586  return SimplifiedInst;
587 }
588 
589 /// This tries to simplify binary operations which some other binary operation
590 /// distributes over either by factorizing out common terms
591 /// (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this results in
592 /// simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is a win).
593 /// Returns the simplified value, or null if it didn't simplify.
594 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
595  Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
598  Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
599 
600  {
601  // Factorization.
602  Value *A, *B, *C, *D;
603  Instruction::BinaryOps LHSOpcode, RHSOpcode;
604  if (Op0)
605  LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
606  if (Op1)
607  RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
608 
609  // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
610  // a common term.
611  if (Op0 && Op1 && LHSOpcode == RHSOpcode)
612  if (Value *V = tryFactorization(I, LHSOpcode, A, B, C, D))
613  return V;
614 
615  // The instruction has the form "(A op' B) op (C)". Try to factorize common
616  // term.
617  if (Op0)
618  if (Value *Ident = getIdentityValue(LHSOpcode, RHS))
619  if (Value *V =
620  tryFactorization(I, LHSOpcode, A, B, RHS, Ident))
621  return V;
622 
623  // The instruction has the form "(B) op (C op' D)". Try to factorize common
624  // term.
625  if (Op1)
626  if (Value *Ident = getIdentityValue(RHSOpcode, LHS))
627  if (Value *V =
628  tryFactorization(I, RHSOpcode, LHS, Ident, C, D))
629  return V;
630  }
631 
632  // Expansion.
633  if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
634  // The instruction has the form "(A op' B) op C". See if expanding it out
635  // to "(A op C) op' (B op C)" results in simplifications.
636  Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
637  Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
638 
639  Value *L = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
640  Value *R = SimplifyBinOp(TopLevelOpcode, B, C, SQ.getWithInstruction(&I));
641 
642  // Do "A op C" and "B op C" both simplify?
643  if (L && R) {
644  // They do! Return "L op' R".
645  ++NumExpand;
646  C = Builder.CreateBinOp(InnerOpcode, L, R);
647  C->takeName(&I);
648  return C;
649  }
650 
651  // Does "A op C" simplify to the identity value for the inner opcode?
652  if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
653  // They do! Return "B op C".
654  ++NumExpand;
655  C = Builder.CreateBinOp(TopLevelOpcode, B, C);
656  C->takeName(&I);
657  return C;
658  }
659 
660  // Does "B op C" simplify to the identity value for the inner opcode?
661  if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
662  // They do! Return "A op C".
663  ++NumExpand;
664  C = Builder.CreateBinOp(TopLevelOpcode, A, C);
665  C->takeName(&I);
666  return C;
667  }
668  }
669 
670  if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
671  // The instruction has the form "A op (B op' C)". See if expanding it out
672  // to "(A op B) op' (A op C)" results in simplifications.
673  Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
674  Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
675 
676  Value *L = SimplifyBinOp(TopLevelOpcode, A, B, SQ.getWithInstruction(&I));
677  Value *R = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
678 
679  // Do "A op B" and "A op C" both simplify?
680  if (L && R) {
681  // They do! Return "L op' R".
682  ++NumExpand;
683  A = Builder.CreateBinOp(InnerOpcode, L, R);
684  A->takeName(&I);
685  return A;
686  }
687 
688  // Does "A op B" simplify to the identity value for the inner opcode?
689  if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
690  // They do! Return "A op C".
691  ++NumExpand;
692  A = Builder.CreateBinOp(TopLevelOpcode, A, C);
693  A->takeName(&I);
694  return A;
695  }
696 
697  // Does "A op C" simplify to the identity value for the inner opcode?
698  if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
699  // They do! Return "A op B".
700  ++NumExpand;
701  A = Builder.CreateBinOp(TopLevelOpcode, A, B);
702  A->takeName(&I);
703  return A;
704  }
705  }
706 
707  // (op (select (a, c, b)), (select (a, d, b))) -> (select (a, (op c, d), 0))
708  // (op (select (a, b, c)), (select (a, b, d))) -> (select (a, 0, (op c, d)))
709  if (auto *SI0 = dyn_cast<SelectInst>(LHS)) {
710  if (auto *SI1 = dyn_cast<SelectInst>(RHS)) {
711  if (SI0->getCondition() == SI1->getCondition()) {
712  Value *SI = nullptr;
713  if (Value *V =
714  SimplifyBinOp(TopLevelOpcode, SI0->getFalseValue(),
715  SI1->getFalseValue(), SQ.getWithInstruction(&I)))
716  SI = Builder.CreateSelect(SI0->getCondition(),
717  Builder.CreateBinOp(TopLevelOpcode,
718  SI0->getTrueValue(),
719  SI1->getTrueValue()),
720  V);
721  if (Value *V =
722  SimplifyBinOp(TopLevelOpcode, SI0->getTrueValue(),
723  SI1->getTrueValue(), SQ.getWithInstruction(&I)))
724  SI = Builder.CreateSelect(
725  SI0->getCondition(), V,
726  Builder.CreateBinOp(TopLevelOpcode, SI0->getFalseValue(),
727  SI1->getFalseValue()));
728  if (SI) {
729  SI->takeName(&I);
730  return SI;
731  }
732  }
733  }
734  }
735 
736  return nullptr;
737 }
738 
739 /// Given a 'sub' instruction, return the RHS of the instruction if the LHS is a
740 /// constant zero (which is the 'negate' form).
741 Value *InstCombiner::dyn_castNegVal(Value *V) const {
742  if (BinaryOperator::isNeg(V))
744 
745  // Constants can be considered to be negated values if they can be folded.
746  if (ConstantInt *C = dyn_cast<ConstantInt>(V))
747  return ConstantExpr::getNeg(C);
748 
749  if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
750  if (C->getType()->getElementType()->isIntegerTy())
751  return ConstantExpr::getNeg(C);
752 
753  if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
754  for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
755  Constant *Elt = CV->getAggregateElement(i);
756  if (!Elt)
757  return nullptr;
758 
759  if (isa<UndefValue>(Elt))
760  continue;
761 
762  if (!isa<ConstantInt>(Elt))
763  return nullptr;
764  }
765  return ConstantExpr::getNeg(CV);
766  }
767 
768  return nullptr;
769 }
770 
771 /// Given a 'fsub' instruction, return the RHS of the instruction if the LHS is
772 /// a constant negative zero (which is the 'negate' form).
773 Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
774  if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
776 
777  // Constants can be considered to be negated values if they can be folded.
778  if (ConstantFP *C = dyn_cast<ConstantFP>(V))
779  return ConstantExpr::getFNeg(C);
780 
781  if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
782  if (C->getType()->getElementType()->isFloatingPointTy())
783  return ConstantExpr::getFNeg(C);
784 
785  return nullptr;
786 }
787 
789  InstCombiner::BuilderTy &Builder) {
790  if (auto *Cast = dyn_cast<CastInst>(&I))
791  return Builder.CreateCast(Cast->getOpcode(), SO, I.getType());
792 
793  assert(I.isBinaryOp() && "Unexpected opcode for select folding");
794 
795  // Figure out if the constant is the left or the right argument.
796  bool ConstIsRHS = isa<Constant>(I.getOperand(1));
797  Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
798 
799  if (auto *SOC = dyn_cast<Constant>(SO)) {
800  if (ConstIsRHS)
801  return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
802  return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
803  }
804 
805  Value *Op0 = SO, *Op1 = ConstOperand;
806  if (!ConstIsRHS)
807  std::swap(Op0, Op1);
808 
809  auto *BO = cast<BinaryOperator>(&I);
810  Value *RI = Builder.CreateBinOp(BO->getOpcode(), Op0, Op1,
811  SO->getName() + ".op");
812  auto *FPInst = dyn_cast<Instruction>(RI);
813  if (FPInst && isa<FPMathOperator>(FPInst))
814  FPInst->copyFastMathFlags(BO);
815  return RI;
816 }
817 
818 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
819  // Don't modify shared select instructions.
820  if (!SI->hasOneUse())
821  return nullptr;
822 
823  Value *TV = SI->getTrueValue();
824  Value *FV = SI->getFalseValue();
825  if (!(isa<Constant>(TV) || isa<Constant>(FV)))
826  return nullptr;
827 
828  // Bool selects with constant operands can be folded to logical ops.
829  if (SI->getType()->isIntOrIntVectorTy(1))
830  return nullptr;
831 
832  // If it's a bitcast involving vectors, make sure it has the same number of
833  // elements on both sides.
834  if (auto *BC = dyn_cast<BitCastInst>(&Op)) {
835  VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
836  VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
837 
838  // Verify that either both or neither are vectors.
839  if ((SrcTy == nullptr) != (DestTy == nullptr))
840  return nullptr;
841 
842  // If vectors, verify that they have the same number of elements.
843  if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
844  return nullptr;
845  }
846 
847  // Test if a CmpInst instruction is used exclusively by a select as
848  // part of a minimum or maximum operation. If so, refrain from doing
849  // any other folding. This helps out other analyses which understand
850  // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
851  // and CodeGen. And in this case, at least one of the comparison
852  // operands has at least one user besides the compare (the select),
853  // which would often largely negate the benefit of folding anyway.
854  if (auto *CI = dyn_cast<CmpInst>(SI->getCondition())) {
855  if (CI->hasOneUse()) {
856  Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
857  if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
858  (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
859  return nullptr;
860  }
861  }
862 
863  Value *NewTV = foldOperationIntoSelectOperand(Op, TV, Builder);
864  Value *NewFV = foldOperationIntoSelectOperand(Op, FV, Builder);
865  return SelectInst::Create(SI->getCondition(), NewTV, NewFV, "", nullptr, SI);
866 }
867 
869  InstCombiner::BuilderTy &Builder) {
870  bool ConstIsRHS = isa<Constant>(I->getOperand(1));
871  Constant *C = cast<Constant>(I->getOperand(ConstIsRHS));
872 
873  if (auto *InC = dyn_cast<Constant>(InV)) {
874  if (ConstIsRHS)
875  return ConstantExpr::get(I->getOpcode(), InC, C);
876  return ConstantExpr::get(I->getOpcode(), C, InC);
877  }
878 
879  Value *Op0 = InV, *Op1 = C;
880  if (!ConstIsRHS)
881  std::swap(Op0, Op1);
882 
883  Value *RI = Builder.CreateBinOp(I->getOpcode(), Op0, Op1, "phitmp");
884  auto *FPInst = dyn_cast<Instruction>(RI);
885  if (FPInst && isa<FPMathOperator>(FPInst))
886  FPInst->copyFastMathFlags(I);
887  return RI;
888 }
889 
890 Instruction *InstCombiner::foldOpIntoPhi(Instruction &I, PHINode *PN) {
891  unsigned NumPHIValues = PN->getNumIncomingValues();
892  if (NumPHIValues == 0)
893  return nullptr;
894 
895  // We normally only transform phis with a single use. However, if a PHI has
896  // multiple uses and they are all the same operation, we can fold *all* of the
897  // uses into the PHI.
898  if (!PN->hasOneUse()) {
899  // Walk the use list for the instruction, comparing them to I.
900  for (User *U : PN->users()) {
901  Instruction *UI = cast<Instruction>(U);
902  if (UI != &I && !I.isIdenticalTo(UI))
903  return nullptr;
904  }
905  // Otherwise, we can replace *all* users with the new PHI we form.
906  }
907 
908  // Check to see if all of the operands of the PHI are simple constants
909  // (constantint/constantfp/undef). If there is one non-constant value,
910  // remember the BB it is in. If there is more than one or if *it* is a PHI,
911  // bail out. We don't do arbitrary constant expressions here because moving
912  // their computation can be expensive without a cost model.
913  BasicBlock *NonConstBB = nullptr;
914  for (unsigned i = 0; i != NumPHIValues; ++i) {
915  Value *InVal = PN->getIncomingValue(i);
916  if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
917  continue;
918 
919  if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
920  if (NonConstBB) return nullptr; // More than one non-const value.
921 
922  NonConstBB = PN->getIncomingBlock(i);
923 
924  // If the InVal is an invoke at the end of the pred block, then we can't
925  // insert a computation after it without breaking the edge.
926  if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
927  if (II->getParent() == NonConstBB)
928  return nullptr;
929 
930  // If the incoming non-constant value is in I's block, we will remove one
931  // instruction, but insert another equivalent one, leading to infinite
932  // instcombine.
933  if (isPotentiallyReachable(I.getParent(), NonConstBB, &DT, LI))
934  return nullptr;
935  }
936 
937  // If there is exactly one non-constant value, we can insert a copy of the
938  // operation in that block. However, if this is a critical edge, we would be
939  // inserting the computation on some other paths (e.g. inside a loop). Only
940  // do this if the pred block is unconditionally branching into the phi block.
941  if (NonConstBB != nullptr) {
942  BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
943  if (!BI || !BI->isUnconditional()) return nullptr;
944  }
945 
946  // Okay, we can do the transformation: create the new PHI node.
948  InsertNewInstBefore(NewPN, *PN);
949  NewPN->takeName(PN);
950 
951  // If we are going to have to insert a new computation, do so right before the
952  // predecessor's terminator.
953  if (NonConstBB)
954  Builder.SetInsertPoint(NonConstBB->getTerminator());
955 
956  // Next, add all of the operands to the PHI.
957  if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
958  // We only currently try to fold the condition of a select when it is a phi,
959  // not the true/false values.
960  Value *TrueV = SI->getTrueValue();
961  Value *FalseV = SI->getFalseValue();
962  BasicBlock *PhiTransBB = PN->getParent();
963  for (unsigned i = 0; i != NumPHIValues; ++i) {
964  BasicBlock *ThisBB = PN->getIncomingBlock(i);
965  Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
966  Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
967  Value *InV = nullptr;
968  // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
969  // even if currently isNullValue gives false.
970  Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
971  // For vector constants, we cannot use isNullValue to fold into
972  // FalseVInPred versus TrueVInPred. When we have individual nonzero
973  // elements in the vector, we will incorrectly fold InC to
974  // `TrueVInPred`.
975  if (InC && !isa<ConstantExpr>(InC) && isa<ConstantInt>(InC))
976  InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
977  else {
978  // Generate the select in the same block as PN's current incoming block.
979  // Note: ThisBB need not be the NonConstBB because vector constants
980  // which are constants by definition are handled here.
981  // FIXME: This can lead to an increase in IR generation because we might
982  // generate selects for vector constant phi operand, that could not be
983  // folded to TrueVInPred or FalseVInPred as done for ConstantInt. For
984  // non-vector phis, this transformation was always profitable because
985  // the select would be generated exactly once in the NonConstBB.
986  Builder.SetInsertPoint(ThisBB->getTerminator());
987  InV = Builder.CreateSelect(PN->getIncomingValue(i), TrueVInPred,
988  FalseVInPred, "phitmp");
989  }
990  NewPN->addIncoming(InV, ThisBB);
991  }
992  } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
993  Constant *C = cast<Constant>(I.getOperand(1));
994  for (unsigned i = 0; i != NumPHIValues; ++i) {
995  Value *InV = nullptr;
996  if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
997  InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
998  else if (isa<ICmpInst>(CI))
999  InV = Builder.CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
1000  C, "phitmp");
1001  else
1002  InV = Builder.CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
1003  C, "phitmp");
1004  NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1005  }
1006  } else if (auto *BO = dyn_cast<BinaryOperator>(&I)) {
1007  for (unsigned i = 0; i != NumPHIValues; ++i) {
1009  Builder);
1010  NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1011  }
1012  } else {
1013  CastInst *CI = cast<CastInst>(&I);
1014  Type *RetTy = CI->getType();
1015  for (unsigned i = 0; i != NumPHIValues; ++i) {
1016  Value *InV;
1017  if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
1018  InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1019  else
1020  InV = Builder.CreateCast(CI->getOpcode(), PN->getIncomingValue(i),
1021  I.getType(), "phitmp");
1022  NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1023  }
1024  }
1025 
1026  for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
1027  Instruction *User = cast<Instruction>(*UI++);
1028  if (User == &I) continue;
1029  replaceInstUsesWith(*User, NewPN);
1030  eraseInstFromFunction(*User);
1031  }
1032  return replaceInstUsesWith(I, NewPN);
1033 }
1034 
1035 Instruction *InstCombiner::foldOpWithConstantIntoOperand(BinaryOperator &I) {
1036  assert(isa<Constant>(I.getOperand(1)) && "Unexpected operand type");
1037 
1038  if (auto *Sel = dyn_cast<SelectInst>(I.getOperand(0))) {
1039  if (Instruction *NewSel = FoldOpIntoSelect(I, Sel))
1040  return NewSel;
1041  } else if (auto *PN = dyn_cast<PHINode>(I.getOperand(0))) {
1042  if (Instruction *NewPhi = foldOpIntoPhi(I, PN))
1043  return NewPhi;
1044  }
1045  return nullptr;
1046 }
1047 
1048 /// Given a pointer type and a constant offset, determine whether or not there
1049 /// is a sequence of GEP indices into the pointed type that will land us at the
1050 /// specified offset. If so, fill them into NewIndices and return the resultant
1051 /// element type, otherwise return null.
1052 Type *InstCombiner::FindElementAtOffset(PointerType *PtrTy, int64_t Offset,
1053  SmallVectorImpl<Value *> &NewIndices) {
1054  Type *Ty = PtrTy->getElementType();
1055  if (!Ty->isSized())
1056  return nullptr;
1057 
1058  // Start with the index over the outer type. Note that the type size
1059  // might be zero (even if the offset isn't zero) if the indexed type
1060  // is something like [0 x {int, int}]
1061  Type *IntPtrTy = DL.getIntPtrType(PtrTy);
1062  int64_t FirstIdx = 0;
1063  if (int64_t TySize = DL.getTypeAllocSize(Ty)) {
1064  FirstIdx = Offset/TySize;
1065  Offset -= FirstIdx*TySize;
1066 
1067  // Handle hosts where % returns negative instead of values [0..TySize).
1068  if (Offset < 0) {
1069  --FirstIdx;
1070  Offset += TySize;
1071  assert(Offset >= 0);
1072  }
1073  assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
1074  }
1075 
1076  NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
1077 
1078  // Index into the types. If we fail, set OrigBase to null.
1079  while (Offset) {
1080  // Indexing into tail padding between struct/array elements.
1081  if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty))
1082  return nullptr;
1083 
1084  if (StructType *STy = dyn_cast<StructType>(Ty)) {
1085  const StructLayout *SL = DL.getStructLayout(STy);
1086  assert(Offset < (int64_t)SL->getSizeInBytes() &&
1087  "Offset must stay within the indexed type");
1088 
1089  unsigned Elt = SL->getElementContainingOffset(Offset);
1091  Elt));
1092 
1093  Offset -= SL->getElementOffset(Elt);
1094  Ty = STy->getElementType(Elt);
1095  } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
1096  uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType());
1097  assert(EltSize && "Cannot index into a zero-sized array");
1098  NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
1099  Offset %= EltSize;
1100  Ty = AT->getElementType();
1101  } else {
1102  // Otherwise, we can't index into the middle of this atomic type, bail.
1103  return nullptr;
1104  }
1105  }
1106 
1107  return Ty;
1108 }
1109 
1111  // If this GEP has only 0 indices, it is the same pointer as
1112  // Src. If Src is not a trivial GEP too, don't combine
1113  // the indices.
1114  if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
1115  !Src.hasOneUse())
1116  return false;
1117  return true;
1118 }
1119 
1120 /// Return a value X such that Val = X * Scale, or null if none.
1121 /// If the multiplication is known not to overflow, then NoSignedWrap is set.
1122 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
1123  assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
1124  assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
1125  Scale.getBitWidth() && "Scale not compatible with value!");
1126 
1127  // If Val is zero or Scale is one then Val = Val * Scale.
1128  if (match(Val, m_Zero()) || Scale == 1) {
1129  NoSignedWrap = true;
1130  return Val;
1131  }
1132 
1133  // If Scale is zero then it does not divide Val.
1134  if (Scale.isMinValue())
1135  return nullptr;
1136 
1137  // Look through chains of multiplications, searching for a constant that is
1138  // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
1139  // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
1140  // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
1141  // down from Val:
1142  //
1143  // Val = M1 * X || Analysis starts here and works down
1144  // M1 = M2 * Y || Doesn't descend into terms with more
1145  // M2 = Z * 4 \/ than one use
1146  //
1147  // Then to modify a term at the bottom:
1148  //
1149  // Val = M1 * X
1150  // M1 = Z * Y || Replaced M2 with Z
1151  //
1152  // Then to work back up correcting nsw flags.
1153 
1154  // Op - the term we are currently analyzing. Starts at Val then drills down.
1155  // Replaced with its descaled value before exiting from the drill down loop.
1156  Value *Op = Val;
1157 
1158  // Parent - initially null, but after drilling down notes where Op came from.
1159  // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1160  // 0'th operand of Val.
1161  std::pair<Instruction*, unsigned> Parent;
1162 
1163  // Set if the transform requires a descaling at deeper levels that doesn't
1164  // overflow.
1165  bool RequireNoSignedWrap = false;
1166 
1167  // Log base 2 of the scale. Negative if not a power of 2.
1168  int32_t logScale = Scale.exactLogBase2();
1169 
1170  for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1171 
1172  if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1173  // If Op is a constant divisible by Scale then descale to the quotient.
1174  APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1175  APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1176  if (!Remainder.isMinValue())
1177  // Not divisible by Scale.
1178  return nullptr;
1179  // Replace with the quotient in the parent.
1180  Op = ConstantInt::get(CI->getType(), Quotient);
1181  NoSignedWrap = true;
1182  break;
1183  }
1184 
1185  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1186 
1187  if (BO->getOpcode() == Instruction::Mul) {
1188  // Multiplication.
1189  NoSignedWrap = BO->hasNoSignedWrap();
1190  if (RequireNoSignedWrap && !NoSignedWrap)
1191  return nullptr;
1192 
1193  // There are three cases for multiplication: multiplication by exactly
1194  // the scale, multiplication by a constant different to the scale, and
1195  // multiplication by something else.
1196  Value *LHS = BO->getOperand(0);
1197  Value *RHS = BO->getOperand(1);
1198 
1199  if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1200  // Multiplication by a constant.
1201  if (CI->getValue() == Scale) {
1202  // Multiplication by exactly the scale, replace the multiplication
1203  // by its left-hand side in the parent.
1204  Op = LHS;
1205  break;
1206  }
1207 
1208  // Otherwise drill down into the constant.
1209  if (!Op->hasOneUse())
1210  return nullptr;
1211 
1212  Parent = std::make_pair(BO, 1);
1213  continue;
1214  }
1215 
1216  // Multiplication by something else. Drill down into the left-hand side
1217  // since that's where the reassociate pass puts the good stuff.
1218  if (!Op->hasOneUse())
1219  return nullptr;
1220 
1221  Parent = std::make_pair(BO, 0);
1222  continue;
1223  }
1224 
1225  if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1226  isa<ConstantInt>(BO->getOperand(1))) {
1227  // Multiplication by a power of 2.
1228  NoSignedWrap = BO->hasNoSignedWrap();
1229  if (RequireNoSignedWrap && !NoSignedWrap)
1230  return nullptr;
1231 
1232  Value *LHS = BO->getOperand(0);
1233  int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1234  getLimitedValue(Scale.getBitWidth());
1235  // Op = LHS << Amt.
1236 
1237  if (Amt == logScale) {
1238  // Multiplication by exactly the scale, replace the multiplication
1239  // by its left-hand side in the parent.
1240  Op = LHS;
1241  break;
1242  }
1243  if (Amt < logScale || !Op->hasOneUse())
1244  return nullptr;
1245 
1246  // Multiplication by more than the scale. Reduce the multiplying amount
1247  // by the scale in the parent.
1248  Parent = std::make_pair(BO, 1);
1249  Op = ConstantInt::get(BO->getType(), Amt - logScale);
1250  break;
1251  }
1252  }
1253 
1254  if (!Op->hasOneUse())
1255  return nullptr;
1256 
1257  if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1258  if (Cast->getOpcode() == Instruction::SExt) {
1259  // Op is sign-extended from a smaller type, descale in the smaller type.
1260  unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1261  APInt SmallScale = Scale.trunc(SmallSize);
1262  // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1263  // descale Op as (sext Y) * Scale. In order to have
1264  // sext (Y * SmallScale) = (sext Y) * Scale
1265  // some conditions need to hold however: SmallScale must sign-extend to
1266  // Scale and the multiplication Y * SmallScale should not overflow.
1267  if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1268  // SmallScale does not sign-extend to Scale.
1269  return nullptr;
1270  assert(SmallScale.exactLogBase2() == logScale);
1271  // Require that Y * SmallScale must not overflow.
1272  RequireNoSignedWrap = true;
1273 
1274  // Drill down through the cast.
1275  Parent = std::make_pair(Cast, 0);
1276  Scale = SmallScale;
1277  continue;
1278  }
1279 
1280  if (Cast->getOpcode() == Instruction::Trunc) {
1281  // Op is truncated from a larger type, descale in the larger type.
1282  // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1283  // trunc (Y * sext Scale) = (trunc Y) * Scale
1284  // always holds. However (trunc Y) * Scale may overflow even if
1285  // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1286  // from this point up in the expression (see later).
1287  if (RequireNoSignedWrap)
1288  return nullptr;
1289 
1290  // Drill down through the cast.
1291  unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1292  Parent = std::make_pair(Cast, 0);
1293  Scale = Scale.sext(LargeSize);
1294  if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1295  logScale = -1;
1296  assert(Scale.exactLogBase2() == logScale);
1297  continue;
1298  }
1299  }
1300 
1301  // Unsupported expression, bail out.
1302  return nullptr;
1303  }
1304 
1305  // If Op is zero then Val = Op * Scale.
1306  if (match(Op, m_Zero())) {
1307  NoSignedWrap = true;
1308  return Op;
1309  }
1310 
1311  // We know that we can successfully descale, so from here on we can safely
1312  // modify the IR. Op holds the descaled version of the deepest term in the
1313  // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1314  // not to overflow.
1315 
1316  if (!Parent.first)
1317  // The expression only had one term.
1318  return Op;
1319 
1320  // Rewrite the parent using the descaled version of its operand.
1321  assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1322  assert(Op != Parent.first->getOperand(Parent.second) &&
1323  "Descaling was a no-op?");
1324  Parent.first->setOperand(Parent.second, Op);
1325  Worklist.Add(Parent.first);
1326 
1327  // Now work back up the expression correcting nsw flags. The logic is based
1328  // on the following observation: if X * Y is known not to overflow as a signed
1329  // multiplication, and Y is replaced by a value Z with smaller absolute value,
1330  // then X * Z will not overflow as a signed multiplication either. As we work
1331  // our way up, having NoSignedWrap 'true' means that the descaled value at the
1332  // current level has strictly smaller absolute value than the original.
1333  Instruction *Ancestor = Parent.first;
1334  do {
1335  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1336  // If the multiplication wasn't nsw then we can't say anything about the
1337  // value of the descaled multiplication, and we have to clear nsw flags
1338  // from this point on up.
1339  bool OpNoSignedWrap = BO->hasNoSignedWrap();
1340  NoSignedWrap &= OpNoSignedWrap;
1341  if (NoSignedWrap != OpNoSignedWrap) {
1342  BO->setHasNoSignedWrap(NoSignedWrap);
1343  Worklist.Add(Ancestor);
1344  }
1345  } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1346  // The fact that the descaled input to the trunc has smaller absolute
1347  // value than the original input doesn't tell us anything useful about
1348  // the absolute values of the truncations.
1349  NoSignedWrap = false;
1350  }
1351  assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1352  "Failed to keep proper track of nsw flags while drilling down?");
1353 
1354  if (Ancestor == Val)
1355  // Got to the top, all done!
1356  return Val;
1357 
1358  // Move up one level in the expression.
1359  assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1360  Ancestor = Ancestor->user_back();
1361  } while (1);
1362 }
1363 
1364 /// \brief Creates node of binary operation with the same attributes as the
1365 /// specified one but with other operands.
1368  Value *BO = B.CreateBinOp(Inst.getOpcode(), LHS, RHS);
1369  // If LHS and RHS are constant, BO won't be a binary operator.
1370  if (BinaryOperator *NewBO = dyn_cast<BinaryOperator>(BO))
1371  NewBO->copyIRFlags(&Inst);
1372  return BO;
1373 }
1374 
1375 /// \brief Makes transformation of binary operation specific for vector types.
1376 /// \param Inst Binary operator to transform.
1377 /// \return Pointer to node that must replace the original binary operator, or
1378 /// null pointer if no transformation was made.
1379 Value *InstCombiner::SimplifyVectorOp(BinaryOperator &Inst) {
1380  if (!Inst.getType()->isVectorTy()) return nullptr;
1381 
1382  // It may not be safe to reorder shuffles and things like div, urem, etc.
1383  // because we may trap when executing those ops on unknown vector elements.
1384  // See PR20059.
1385  if (!isSafeToSpeculativelyExecute(&Inst))
1386  return nullptr;
1387 
1388  unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1389  Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1390  assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1391  assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1392 
1393  // If both arguments of the binary operation are shuffles that use the same
1394  // mask and shuffle within a single vector, move the shuffle after the binop:
1395  // Op(shuffle(v1, m), shuffle(v2, m)) -> shuffle(Op(v1, v2), m)
1396  auto *LShuf = dyn_cast<ShuffleVectorInst>(LHS);
1397  auto *RShuf = dyn_cast<ShuffleVectorInst>(RHS);
1398  if (LShuf && RShuf && LShuf->getMask() == RShuf->getMask() &&
1399  isa<UndefValue>(LShuf->getOperand(1)) &&
1400  isa<UndefValue>(RShuf->getOperand(1)) &&
1401  LShuf->getOperand(0)->getType() == RShuf->getOperand(0)->getType()) {
1402  Value *NewBO = CreateBinOpAsGiven(Inst, LShuf->getOperand(0),
1403  RShuf->getOperand(0), Builder);
1404  return Builder.CreateShuffleVector(
1405  NewBO, UndefValue::get(NewBO->getType()), LShuf->getMask());
1406  }
1407 
1408  // If one argument is a shuffle within one vector, the other is a constant,
1409  // try moving the shuffle after the binary operation.
1410  ShuffleVectorInst *Shuffle = nullptr;
1411  Constant *C1 = nullptr;
1412  if (isa<ShuffleVectorInst>(LHS)) Shuffle = cast<ShuffleVectorInst>(LHS);
1413  if (isa<ShuffleVectorInst>(RHS)) Shuffle = cast<ShuffleVectorInst>(RHS);
1414  if (isa<Constant>(LHS)) C1 = cast<Constant>(LHS);
1415  if (isa<Constant>(RHS)) C1 = cast<Constant>(RHS);
1416  if (Shuffle && C1 &&
1417  (isa<ConstantVector>(C1) || isa<ConstantDataVector>(C1)) &&
1418  isa<UndefValue>(Shuffle->getOperand(1)) &&
1419  Shuffle->getType() == Shuffle->getOperand(0)->getType()) {
1420  SmallVector<int, 16> ShMask = Shuffle->getShuffleMask();
1421  // Find constant C2 that has property:
1422  // shuffle(C2, ShMask) = C1
1423  // If such constant does not exist (example: ShMask=<0,0> and C1=<1,2>)
1424  // reorder is not possible.
1425  SmallVector<Constant*, 16> C2M(VWidth,
1427  bool MayChange = true;
1428  for (unsigned I = 0; I < VWidth; ++I) {
1429  if (ShMask[I] >= 0) {
1430  assert(ShMask[I] < (int)VWidth);
1431  if (!isa<UndefValue>(C2M[ShMask[I]])) {
1432  MayChange = false;
1433  break;
1434  }
1435  C2M[ShMask[I]] = C1->getAggregateElement(I);
1436  }
1437  }
1438  if (MayChange) {
1439  Constant *C2 = ConstantVector::get(C2M);
1440  Value *NewLHS = isa<Constant>(LHS) ? C2 : Shuffle->getOperand(0);
1441  Value *NewRHS = isa<Constant>(LHS) ? Shuffle->getOperand(0) : C2;
1442  Value *NewBO = CreateBinOpAsGiven(Inst, NewLHS, NewRHS, Builder);
1443  return Builder.CreateShuffleVector(NewBO,
1444  UndefValue::get(Inst.getType()), Shuffle->getMask());
1445  }
1446  }
1447 
1448  return nullptr;
1449 }
1450 
1452  SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1453 
1454  if (Value *V = SimplifyGEPInst(GEP.getSourceElementType(), Ops,
1455  SQ.getWithInstruction(&GEP)))
1456  return replaceInstUsesWith(GEP, V);
1457 
1458  Value *PtrOp = GEP.getOperand(0);
1459 
1460  // Eliminate unneeded casts for indices, and replace indices which displace
1461  // by multiples of a zero size type with zero.
1462  bool MadeChange = false;
1463  Type *IntPtrTy =
1464  DL.getIntPtrType(GEP.getPointerOperandType()->getScalarType());
1465 
1466  gep_type_iterator GTI = gep_type_begin(GEP);
1467  for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E;
1468  ++I, ++GTI) {
1469  // Skip indices into struct types.
1470  if (GTI.isStruct())
1471  continue;
1472 
1473  // Index type should have the same width as IntPtr
1474  Type *IndexTy = (*I)->getType();
1475  Type *NewIndexType = IndexTy->isVectorTy() ?
1476  VectorType::get(IntPtrTy, IndexTy->getVectorNumElements()) : IntPtrTy;
1477 
1478  // If the element type has zero size then any index over it is equivalent
1479  // to an index of zero, so replace it with zero if it is not zero already.
1480  Type *EltTy = GTI.getIndexedType();
1481  if (EltTy->isSized() && DL.getTypeAllocSize(EltTy) == 0)
1482  if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1483  *I = Constant::getNullValue(NewIndexType);
1484  MadeChange = true;
1485  }
1486 
1487  if (IndexTy != NewIndexType) {
1488  // If we are using a wider index than needed for this platform, shrink
1489  // it to what we need. If narrower, sign-extend it to what we need.
1490  // This explicit cast can make subsequent optimizations more obvious.
1491  *I = Builder.CreateIntCast(*I, NewIndexType, true);
1492  MadeChange = true;
1493  }
1494  }
1495  if (MadeChange)
1496  return &GEP;
1497 
1498  // Check to see if the inputs to the PHI node are getelementptr instructions.
1499  if (PHINode *PN = dyn_cast<PHINode>(PtrOp)) {
1501  if (!Op1)
1502  return nullptr;
1503 
1504  // Don't fold a GEP into itself through a PHI node. This can only happen
1505  // through the back-edge of a loop. Folding a GEP into itself means that
1506  // the value of the previous iteration needs to be stored in the meantime,
1507  // thus requiring an additional register variable to be live, but not
1508  // actually achieving anything (the GEP still needs to be executed once per
1509  // loop iteration).
1510  if (Op1 == &GEP)
1511  return nullptr;
1512 
1513  int DI = -1;
1514 
1515  for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1517  if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1518  return nullptr;
1519 
1520  // As for Op1 above, don't try to fold a GEP into itself.
1521  if (Op2 == &GEP)
1522  return nullptr;
1523 
1524  // Keep track of the type as we walk the GEP.
1525  Type *CurTy = nullptr;
1526 
1527  for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1528  if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1529  return nullptr;
1530 
1531  if (Op1->getOperand(J) != Op2->getOperand(J)) {
1532  if (DI == -1) {
1533  // We have not seen any differences yet in the GEPs feeding the
1534  // PHI yet, so we record this one if it is allowed to be a
1535  // variable.
1536 
1537  // The first two arguments can vary for any GEP, the rest have to be
1538  // static for struct slots
1539  if (J > 1 && CurTy->isStructTy())
1540  return nullptr;
1541 
1542  DI = J;
1543  } else {
1544  // The GEP is different by more than one input. While this could be
1545  // extended to support GEPs that vary by more than one variable it
1546  // doesn't make sense since it greatly increases the complexity and
1547  // would result in an R+R+R addressing mode which no backend
1548  // directly supports and would need to be broken into several
1549  // simpler instructions anyway.
1550  return nullptr;
1551  }
1552  }
1553 
1554  // Sink down a layer of the type for the next iteration.
1555  if (J > 0) {
1556  if (J == 1) {
1557  CurTy = Op1->getSourceElementType();
1558  } else if (CompositeType *CT = dyn_cast<CompositeType>(CurTy)) {
1559  CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1560  } else {
1561  CurTy = nullptr;
1562  }
1563  }
1564  }
1565  }
1566 
1567  // If not all GEPs are identical we'll have to create a new PHI node.
1568  // Check that the old PHI node has only one use so that it will get
1569  // removed.
1570  if (DI != -1 && !PN->hasOneUse())
1571  return nullptr;
1572 
1573  GetElementPtrInst *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1574  if (DI == -1) {
1575  // All the GEPs feeding the PHI are identical. Clone one down into our
1576  // BB so that it can be merged with the current GEP.
1577  GEP.getParent()->getInstList().insert(
1578  GEP.getParent()->getFirstInsertionPt(), NewGEP);
1579  } else {
1580  // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1581  // into the current block so it can be merged, and create a new PHI to
1582  // set that index.
1583  PHINode *NewPN;
1584  {
1585  IRBuilderBase::InsertPointGuard Guard(Builder);
1586  Builder.SetInsertPoint(PN);
1587  NewPN = Builder.CreatePHI(Op1->getOperand(DI)->getType(),
1588  PN->getNumOperands());
1589  }
1590 
1591  for (auto &I : PN->operands())
1592  NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1593  PN->getIncomingBlock(I));
1594 
1595  NewGEP->setOperand(DI, NewPN);
1596  GEP.getParent()->getInstList().insert(
1597  GEP.getParent()->getFirstInsertionPt(), NewGEP);
1598  NewGEP->setOperand(DI, NewPN);
1599  }
1600 
1601  GEP.setOperand(0, NewGEP);
1602  PtrOp = NewGEP;
1603  }
1604 
1605  // Combine Indices - If the source pointer to this getelementptr instruction
1606  // is a getelementptr instruction, combine the indices of the two
1607  // getelementptr instructions into a single instruction.
1608  //
1609  if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1610  if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1611  return nullptr;
1612 
1613  // Note that if our source is a gep chain itself then we wait for that
1614  // chain to be resolved before we perform this transformation. This
1615  // avoids us creating a TON of code in some cases.
1616  if (GEPOperator *SrcGEP =
1617  dyn_cast<GEPOperator>(Src->getOperand(0)))
1618  if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1619  return nullptr; // Wait until our source is folded to completion.
1620 
1621  SmallVector<Value*, 8> Indices;
1622 
1623  // Find out whether the last index in the source GEP is a sequential idx.
1624  bool EndsWithSequential = false;
1625  for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1626  I != E; ++I)
1627  EndsWithSequential = I.isSequential();
1628 
1629  // Can we combine the two pointer arithmetics offsets?
1630  if (EndsWithSequential) {
1631  // Replace: gep (gep %P, long B), long A, ...
1632  // With: T = long A+B; gep %P, T, ...
1633  //
1634  Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1635  Value *GO1 = GEP.getOperand(1);
1636 
1637  // If they aren't the same type, then the input hasn't been processed
1638  // by the loop above yet (which canonicalizes sequential index types to
1639  // intptr_t). Just avoid transforming this until the input has been
1640  // normalized.
1641  if (SO1->getType() != GO1->getType())
1642  return nullptr;
1643 
1644  Value *Sum =
1645  SimplifyAddInst(GO1, SO1, false, false, SQ.getWithInstruction(&GEP));
1646  // Only do the combine when we are sure the cost after the
1647  // merge is never more than that before the merge.
1648  if (Sum == nullptr)
1649  return nullptr;
1650 
1651  // Update the GEP in place if possible.
1652  if (Src->getNumOperands() == 2) {
1653  GEP.setOperand(0, Src->getOperand(0));
1654  GEP.setOperand(1, Sum);
1655  return &GEP;
1656  }
1657  Indices.append(Src->op_begin()+1, Src->op_end()-1);
1658  Indices.push_back(Sum);
1659  Indices.append(GEP.op_begin()+2, GEP.op_end());
1660  } else if (isa<Constant>(*GEP.idx_begin()) &&
1661  cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1662  Src->getNumOperands() != 1) {
1663  // Otherwise we can do the fold if the first index of the GEP is a zero
1664  Indices.append(Src->op_begin()+1, Src->op_end());
1665  Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1666  }
1667 
1668  if (!Indices.empty())
1669  return GEP.isInBounds() && Src->isInBounds()
1671  Src->getSourceElementType(), Src->getOperand(0), Indices,
1672  GEP.getName())
1673  : GetElementPtrInst::Create(Src->getSourceElementType(),
1674  Src->getOperand(0), Indices,
1675  GEP.getName());
1676  }
1677 
1678  if (GEP.getNumIndices() == 1) {
1679  unsigned AS = GEP.getPointerAddressSpace();
1680  if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1681  DL.getPointerSizeInBits(AS)) {
1682  Type *Ty = GEP.getSourceElementType();
1683  uint64_t TyAllocSize = DL.getTypeAllocSize(Ty);
1684 
1685  bool Matched = false;
1686  uint64_t C;
1687  Value *V = nullptr;
1688  if (TyAllocSize == 1) {
1689  V = GEP.getOperand(1);
1690  Matched = true;
1691  } else if (match(GEP.getOperand(1),
1692  m_AShr(m_Value(V), m_ConstantInt(C)))) {
1693  if (TyAllocSize == 1ULL << C)
1694  Matched = true;
1695  } else if (match(GEP.getOperand(1),
1696  m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1697  if (TyAllocSize == C)
1698  Matched = true;
1699  }
1700 
1701  if (Matched) {
1702  // Canonicalize (gep i8* X, -(ptrtoint Y))
1703  // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1704  // The GEP pattern is emitted by the SCEV expander for certain kinds of
1705  // pointer arithmetic.
1706  if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1707  Operator *Index = cast<Operator>(V);
1708  Value *PtrToInt = Builder.CreatePtrToInt(PtrOp, Index->getType());
1709  Value *NewSub = Builder.CreateSub(PtrToInt, Index->getOperand(1));
1710  return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
1711  }
1712  // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1713  // to (bitcast Y)
1714  Value *Y;
1715  if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1716  m_PtrToInt(m_Specific(GEP.getOperand(0)))))) {
1718  GEP.getType());
1719  }
1720  }
1721  }
1722  }
1723 
1724  // We do not handle pointer-vector geps here.
1725  if (GEP.getType()->isVectorTy())
1726  return nullptr;
1727 
1728  // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1729  Value *StrippedPtr = PtrOp->stripPointerCasts();
1730  PointerType *StrippedPtrTy = cast<PointerType>(StrippedPtr->getType());
1731 
1732  if (StrippedPtr != PtrOp) {
1733  bool HasZeroPointerIndex = false;
1734  if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1735  HasZeroPointerIndex = C->isZero();
1736 
1737  // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1738  // into : GEP [10 x i8]* X, i32 0, ...
1739  //
1740  // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1741  // into : GEP i8* X, ...
1742  //
1743  // This occurs when the program declares an array extern like "int X[];"
1744  if (HasZeroPointerIndex) {
1745  if (ArrayType *CATy =
1746  dyn_cast<ArrayType>(GEP.getSourceElementType())) {
1747  // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1748  if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1749  // -> GEP i8* X, ...
1750  SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1752  StrippedPtrTy->getElementType(), StrippedPtr, Idx, GEP.getName());
1753  Res->setIsInBounds(GEP.isInBounds());
1754  if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1755  return Res;
1756  // Insert Res, and create an addrspacecast.
1757  // e.g.,
1758  // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1759  // ->
1760  // %0 = GEP i8 addrspace(1)* X, ...
1761  // addrspacecast i8 addrspace(1)* %0 to i8*
1762  return new AddrSpaceCastInst(Builder.Insert(Res), GEP.getType());
1763  }
1764 
1765  if (ArrayType *XATy =
1766  dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1767  // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1768  if (CATy->getElementType() == XATy->getElementType()) {
1769  // -> GEP [10 x i8]* X, i32 0, ...
1770  // At this point, we know that the cast source type is a pointer
1771  // to an array of the same type as the destination pointer
1772  // array. Because the array type is never stepped over (there
1773  // is a leading zero) we can fold the cast into this GEP.
1774  if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1775  GEP.setOperand(0, StrippedPtr);
1776  GEP.setSourceElementType(XATy);
1777  return &GEP;
1778  }
1779  // Cannot replace the base pointer directly because StrippedPtr's
1780  // address space is different. Instead, create a new GEP followed by
1781  // an addrspacecast.
1782  // e.g.,
1783  // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1784  // i32 0, ...
1785  // ->
1786  // %0 = GEP [10 x i8] addrspace(1)* X, ...
1787  // addrspacecast i8 addrspace(1)* %0 to i8*
1788  SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1789  Value *NewGEP = GEP.isInBounds()
1790  ? Builder.CreateInBoundsGEP(
1791  nullptr, StrippedPtr, Idx, GEP.getName())
1792  : Builder.CreateGEP(nullptr, StrippedPtr, Idx,
1793  GEP.getName());
1794  return new AddrSpaceCastInst(NewGEP, GEP.getType());
1795  }
1796  }
1797  }
1798  } else if (GEP.getNumOperands() == 2) {
1799  // Transform things like:
1800  // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1801  // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1802  Type *SrcElTy = StrippedPtrTy->getElementType();
1803  Type *ResElTy = GEP.getSourceElementType();
1804  if (SrcElTy->isArrayTy() &&
1805  DL.getTypeAllocSize(SrcElTy->getArrayElementType()) ==
1806  DL.getTypeAllocSize(ResElTy)) {
1807  Type *IdxType = DL.getIntPtrType(GEP.getType());
1808  Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1809  Value *NewGEP =
1810  GEP.isInBounds()
1811  ? Builder.CreateInBoundsGEP(nullptr, StrippedPtr, Idx,
1812  GEP.getName())
1813  : Builder.CreateGEP(nullptr, StrippedPtr, Idx, GEP.getName());
1814 
1815  // V and GEP are both pointer types --> BitCast
1817  GEP.getType());
1818  }
1819 
1820  // Transform things like:
1821  // %V = mul i64 %N, 4
1822  // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1823  // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1824  if (ResElTy->isSized() && SrcElTy->isSized()) {
1825  // Check that changing the type amounts to dividing the index by a scale
1826  // factor.
1827  uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1828  uint64_t SrcSize = DL.getTypeAllocSize(SrcElTy);
1829  if (ResSize && SrcSize % ResSize == 0) {
1830  Value *Idx = GEP.getOperand(1);
1831  unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1832  uint64_t Scale = SrcSize / ResSize;
1833 
1834  // Earlier transforms ensure that the index has type IntPtrType, which
1835  // considerably simplifies the logic by eliminating implicit casts.
1836  assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1837  "Index not cast to pointer width?");
1838 
1839  bool NSW;
1840  if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1841  // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1842  // If the multiplication NewIdx * Scale may overflow then the new
1843  // GEP may not be "inbounds".
1844  Value *NewGEP =
1845  GEP.isInBounds() && NSW
1846  ? Builder.CreateInBoundsGEP(nullptr, StrippedPtr, NewIdx,
1847  GEP.getName())
1848  : Builder.CreateGEP(nullptr, StrippedPtr, NewIdx,
1849  GEP.getName());
1850 
1851  // The NewGEP must be pointer typed, so must the old one -> BitCast
1853  GEP.getType());
1854  }
1855  }
1856  }
1857 
1858  // Similarly, transform things like:
1859  // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1860  // (where tmp = 8*tmp2) into:
1861  // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1862  if (ResElTy->isSized() && SrcElTy->isSized() && SrcElTy->isArrayTy()) {
1863  // Check that changing to the array element type amounts to dividing the
1864  // index by a scale factor.
1865  uint64_t ResSize = DL.getTypeAllocSize(ResElTy);
1866  uint64_t ArrayEltSize =
1867  DL.getTypeAllocSize(SrcElTy->getArrayElementType());
1868  if (ResSize && ArrayEltSize % ResSize == 0) {
1869  Value *Idx = GEP.getOperand(1);
1870  unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1871  uint64_t Scale = ArrayEltSize / ResSize;
1872 
1873  // Earlier transforms ensure that the index has type IntPtrType, which
1874  // considerably simplifies the logic by eliminating implicit casts.
1875  assert(Idx->getType() == DL.getIntPtrType(GEP.getType()) &&
1876  "Index not cast to pointer width?");
1877 
1878  bool NSW;
1879  if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1880  // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1881  // If the multiplication NewIdx * Scale may overflow then the new
1882  // GEP may not be "inbounds".
1883  Value *Off[2] = {
1884  Constant::getNullValue(DL.getIntPtrType(GEP.getType())),
1885  NewIdx};
1886 
1887  Value *NewGEP = GEP.isInBounds() && NSW
1888  ? Builder.CreateInBoundsGEP(
1889  SrcElTy, StrippedPtr, Off, GEP.getName())
1890  : Builder.CreateGEP(SrcElTy, StrippedPtr, Off,
1891  GEP.getName());
1892  // The NewGEP must be pointer typed, so must the old one -> BitCast
1894  GEP.getType());
1895  }
1896  }
1897  }
1898  }
1899  }
1900 
1901  // addrspacecast between types is canonicalized as a bitcast, then an
1902  // addrspacecast. To take advantage of the below bitcast + struct GEP, look
1903  // through the addrspacecast.
1904  if (AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
1905  // X = bitcast A addrspace(1)* to B addrspace(1)*
1906  // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
1907  // Z = gep Y, <...constant indices...>
1908  // Into an addrspacecasted GEP of the struct.
1909  if (BitCastInst *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
1910  PtrOp = BC;
1911  }
1912 
1913  /// See if we can simplify:
1914  /// X = bitcast A* to B*
1915  /// Y = gep X, <...constant indices...>
1916  /// into a gep of the original struct. This is important for SROA and alias
1917  /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1918  if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1919  Value *Operand = BCI->getOperand(0);
1920  PointerType *OpType = cast<PointerType>(Operand->getType());
1921  unsigned OffsetBits = DL.getPointerTypeSizeInBits(GEP.getType());
1922  APInt Offset(OffsetBits, 0);
1923  if (!isa<BitCastInst>(Operand) &&
1924  GEP.accumulateConstantOffset(DL, Offset)) {
1925 
1926  // If this GEP instruction doesn't move the pointer, just replace the GEP
1927  // with a bitcast of the real input to the dest type.
1928  if (!Offset) {
1929  // If the bitcast is of an allocation, and the allocation will be
1930  // converted to match the type of the cast, don't touch this.
1931  if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, &TLI)) {
1932  // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1933  if (Instruction *I = visitBitCast(*BCI)) {
1934  if (I != BCI) {
1935  I->takeName(BCI);
1936  BCI->getParent()->getInstList().insert(BCI->getIterator(), I);
1937  replaceInstUsesWith(*BCI, I);
1938  }
1939  return &GEP;
1940  }
1941  }
1942 
1943  if (Operand->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1944  return new AddrSpaceCastInst(Operand, GEP.getType());
1945  return new BitCastInst(Operand, GEP.getType());
1946  }
1947 
1948  // Otherwise, if the offset is non-zero, we need to find out if there is a
1949  // field at Offset in 'A's type. If so, we can pull the cast through the
1950  // GEP.
1951  SmallVector<Value*, 8> NewIndices;
1952  if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
1953  Value *NGEP =
1954  GEP.isInBounds()
1955  ? Builder.CreateInBoundsGEP(nullptr, Operand, NewIndices)
1956  : Builder.CreateGEP(nullptr, Operand, NewIndices);
1957 
1958  if (NGEP->getType() == GEP.getType())
1959  return replaceInstUsesWith(GEP, NGEP);
1960  NGEP->takeName(&GEP);
1961 
1962  if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
1963  return new AddrSpaceCastInst(NGEP, GEP.getType());
1964  return new BitCastInst(NGEP, GEP.getType());
1965  }
1966  }
1967  }
1968 
1969  if (!GEP.isInBounds()) {
1970  unsigned PtrWidth =
1971  DL.getPointerSizeInBits(PtrOp->getType()->getPointerAddressSpace());
1972  APInt BasePtrOffset(PtrWidth, 0);
1973  Value *UnderlyingPtrOp =
1975  BasePtrOffset);
1976  if (auto *AI = dyn_cast<AllocaInst>(UnderlyingPtrOp)) {
1977  if (GEP.accumulateConstantOffset(DL, BasePtrOffset) &&
1978  BasePtrOffset.isNonNegative()) {
1979  APInt AllocSize(PtrWidth, DL.getTypeAllocSize(AI->getAllocatedType()));
1980  if (BasePtrOffset.ule(AllocSize)) {
1982  PtrOp, makeArrayRef(Ops).slice(1), GEP.getName());
1983  }
1984  }
1985  }
1986  }
1987 
1988  return nullptr;
1989 }
1990 
1992  Instruction *AI) {
1993  if (isa<ConstantPointerNull>(V))
1994  return true;
1995  if (auto *LI = dyn_cast<LoadInst>(V))
1996  return isa<GlobalVariable>(LI->getPointerOperand());
1997  // Two distinct allocations will never be equal.
1998  // We rely on LookThroughBitCast in isAllocLikeFn being false, since looking
1999  // through bitcasts of V can cause
2000  // the result statement below to be true, even when AI and V (ex:
2001  // i8* ->i32* ->i8* of AI) are the same allocations.
2002  return isAllocLikeFn(V, TLI) && V != AI;
2003 }
2004 
2007  const TargetLibraryInfo *TLI) {
2009  Worklist.push_back(AI);
2010 
2011  do {
2012  Instruction *PI = Worklist.pop_back_val();
2013  for (User *U : PI->users()) {
2014  Instruction *I = cast<Instruction>(U);
2015  switch (I->getOpcode()) {
2016  default:
2017  // Give up the moment we see something we can't handle.
2018  return false;
2019 
2020  case Instruction::AddrSpaceCast:
2021  case Instruction::BitCast:
2022  case Instruction::GetElementPtr:
2023  Users.emplace_back(I);
2024  Worklist.push_back(I);
2025  continue;
2026 
2027  case Instruction::ICmp: {
2028  ICmpInst *ICI = cast<ICmpInst>(I);
2029  // We can fold eq/ne comparisons with null to false/true, respectively.
2030  // We also fold comparisons in some conditions provided the alloc has
2031  // not escaped (see isNeverEqualToUnescapedAlloc).
2032  if (!ICI->isEquality())
2033  return false;
2034  unsigned OtherIndex = (ICI->getOperand(0) == PI) ? 1 : 0;
2035  if (!isNeverEqualToUnescapedAlloc(ICI->getOperand(OtherIndex), TLI, AI))
2036  return false;
2037  Users.emplace_back(I);
2038  continue;
2039  }
2040 
2041  case Instruction::Call:
2042  // Ignore no-op and store intrinsics.
2043  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2044  switch (II->getIntrinsicID()) {
2045  default:
2046  return false;
2047 
2048  case Intrinsic::memmove:
2049  case Intrinsic::memcpy:
2050  case Intrinsic::memset: {
2051  MemIntrinsic *MI = cast<MemIntrinsic>(II);
2052  if (MI->isVolatile() || MI->getRawDest() != PI)
2053  return false;
2055  }
2056  case Intrinsic::dbg_declare:
2057  case Intrinsic::dbg_value:
2058  case Intrinsic::invariant_start:
2059  case Intrinsic::invariant_end:
2060  case Intrinsic::lifetime_start:
2061  case Intrinsic::lifetime_end:
2062  case Intrinsic::objectsize:
2063  Users.emplace_back(I);
2064  continue;
2065  }
2066  }
2067 
2068  if (isFreeCall(I, TLI)) {
2069  Users.emplace_back(I);
2070  continue;
2071  }
2072  return false;
2073 
2074  case Instruction::Store: {
2075  StoreInst *SI = cast<StoreInst>(I);
2076  if (SI->isVolatile() || SI->getPointerOperand() != PI)
2077  return false;
2078  Users.emplace_back(I);
2079  continue;
2080  }
2081  }
2082  llvm_unreachable("missing a return?");
2083  }
2084  } while (!Worklist.empty());
2085  return true;
2086 }
2087 
2089  // If we have a malloc call which is only used in any amount of comparisons
2090  // to null and free calls, delete the calls and replace the comparisons with
2091  // true or false as appropriate.
2093  if (isAllocSiteRemovable(&MI, Users, &TLI)) {
2094  for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2095  // Lowering all @llvm.objectsize calls first because they may
2096  // use a bitcast/GEP of the alloca we are removing.
2097  if (!Users[i])
2098  continue;
2099 
2100  Instruction *I = cast<Instruction>(&*Users[i]);
2101 
2102  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2103  if (II->getIntrinsicID() == Intrinsic::objectsize) {
2104  ConstantInt *Result = lowerObjectSizeCall(II, DL, &TLI,
2105  /*MustSucceed=*/true);
2106  replaceInstUsesWith(*I, Result);
2107  eraseInstFromFunction(*I);
2108  Users[i] = nullptr; // Skip examining in the next loop.
2109  }
2110  }
2111  }
2112  for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2113  if (!Users[i])
2114  continue;
2115 
2116  Instruction *I = cast<Instruction>(&*Users[i]);
2117 
2118  if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
2119  replaceInstUsesWith(*C,
2121  C->isFalseWhenEqual()));
2122  } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I) ||
2123  isa<AddrSpaceCastInst>(I)) {
2124  replaceInstUsesWith(*I, UndefValue::get(I->getType()));
2125  }
2126  eraseInstFromFunction(*I);
2127  }
2128 
2129  if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
2130  // Replace invoke with a NOP intrinsic to maintain the original CFG
2131  Module *M = II->getModule();
2132  Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
2133  InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
2134  None, "", II->getParent());
2135  }
2136  return eraseInstFromFunction(MI);
2137  }
2138  return nullptr;
2139 }
2140 
2141 /// \brief Move the call to free before a NULL test.
2142 ///
2143 /// Check if this free is accessed after its argument has been test
2144 /// against NULL (property 0).
2145 /// If yes, it is legal to move this call in its predecessor block.
2146 ///
2147 /// The move is performed only if the block containing the call to free
2148 /// will be removed, i.e.:
2149 /// 1. it has only one predecessor P, and P has two successors
2150 /// 2. it contains the call and an unconditional branch
2151 /// 3. its successor is the same as its predecessor's successor
2152 ///
2153 /// The profitability is out-of concern here and this function should
2154 /// be called only if the caller knows this transformation would be
2155 /// profitable (e.g., for code size).
2156 static Instruction *
2158  Value *Op = FI.getArgOperand(0);
2159  BasicBlock *FreeInstrBB = FI.getParent();
2160  BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
2161 
2162  // Validate part of constraint #1: Only one predecessor
2163  // FIXME: We can extend the number of predecessor, but in that case, we
2164  // would duplicate the call to free in each predecessor and it may
2165  // not be profitable even for code size.
2166  if (!PredBB)
2167  return nullptr;
2168 
2169  // Validate constraint #2: Does this block contains only the call to
2170  // free and an unconditional branch?
2171  // FIXME: We could check if we can speculate everything in the
2172  // predecessor block
2173  if (FreeInstrBB->size() != 2)
2174  return nullptr;
2175  BasicBlock *SuccBB;
2176  if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
2177  return nullptr;
2178 
2179  // Validate the rest of constraint #1 by matching on the pred branch.
2180  TerminatorInst *TI = PredBB->getTerminator();
2181  BasicBlock *TrueBB, *FalseBB;
2182  ICmpInst::Predicate Pred;
2183  if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
2184  return nullptr;
2185  if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2186  return nullptr;
2187 
2188  // Validate constraint #3: Ensure the null case just falls through.
2189  if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
2190  return nullptr;
2191  assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
2192  "Broken CFG: missing edge from predecessor to successor");
2193 
2194  FI.moveBefore(TI);
2195  return &FI;
2196 }
2197 
2198 
2200  Value *Op = FI.getArgOperand(0);
2201 
2202  // free undef -> unreachable.
2203  if (isa<UndefValue>(Op)) {
2204  // Insert a new store to null because we cannot modify the CFG here.
2205  Builder.CreateStore(ConstantInt::getTrue(FI.getContext()),
2207  return eraseInstFromFunction(FI);
2208  }
2209 
2210  // If we have 'free null' delete the instruction. This can happen in stl code
2211  // when lots of inlining happens.
2212  if (isa<ConstantPointerNull>(Op))
2213  return eraseInstFromFunction(FI);
2214 
2215  // If we optimize for code size, try to move the call to free before the null
2216  // test so that simplify cfg can remove the empty block and dead code
2217  // elimination the branch. I.e., helps to turn something like:
2218  // if (foo) free(foo);
2219  // into
2220  // free(foo);
2221  if (MinimizeSize)
2223  return I;
2224 
2225  return nullptr;
2226 }
2227 
2229  if (RI.getNumOperands() == 0) // ret void
2230  return nullptr;
2231 
2232  Value *ResultOp = RI.getOperand(0);
2233  Type *VTy = ResultOp->getType();
2234  if (!VTy->isIntegerTy())
2235  return nullptr;
2236 
2237  // There might be assume intrinsics dominating this return that completely
2238  // determine the value. If so, constant fold it.
2239  KnownBits Known = computeKnownBits(ResultOp, 0, &RI);
2240  if (Known.isConstant())
2241  RI.setOperand(0, Constant::getIntegerValue(VTy, Known.getConstant()));
2242 
2243  return nullptr;
2244 }
2245 
2247  // Change br (not X), label True, label False to: br X, label False, True
2248  Value *X = nullptr;
2249  BasicBlock *TrueDest;
2250  BasicBlock *FalseDest;
2251  if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2252  !isa<Constant>(X)) {
2253  // Swap Destinations and condition...
2254  BI.setCondition(X);
2255  BI.swapSuccessors();
2256  return &BI;
2257  }
2258 
2259  // If the condition is irrelevant, remove the use so that other
2260  // transforms on the condition become more effective.
2261  if (BI.isConditional() &&
2262  BI.getSuccessor(0) == BI.getSuccessor(1) &&
2263  !isa<UndefValue>(BI.getCondition())) {
2265  return &BI;
2266  }
2267 
2268  // Canonicalize, for example, icmp_ne -> icmp_eq or fcmp_one -> fcmp_oeq.
2269  CmpInst::Predicate Pred;
2270  if (match(&BI, m_Br(m_OneUse(m_Cmp(Pred, m_Value(), m_Value())), TrueDest,
2271  FalseDest)) &&
2272  !isCanonicalPredicate(Pred)) {
2273  // Swap destinations and condition.
2274  CmpInst *Cond = cast<CmpInst>(BI.getCondition());
2276  BI.swapSuccessors();
2277  Worklist.Add(Cond);
2278  return &BI;
2279  }
2280 
2281  return nullptr;
2282 }
2283 
2285  Value *Cond = SI.getCondition();
2286  Value *Op0;
2287  ConstantInt *AddRHS;
2288  if (match(Cond, m_Add(m_Value(Op0), m_ConstantInt(AddRHS)))) {
2289  // Change 'switch (X+4) case 1:' into 'switch (X) case -3'.
2290  for (auto Case : SI.cases()) {
2291  Constant *NewCase = ConstantExpr::getSub(Case.getCaseValue(), AddRHS);
2292  assert(isa<ConstantInt>(NewCase) &&
2293  "Result of expression should be constant");
2294  Case.setValue(cast<ConstantInt>(NewCase));
2295  }
2296  SI.setCondition(Op0);
2297  return &SI;
2298  }
2299 
2300  KnownBits Known = computeKnownBits(Cond, 0, &SI);
2301  unsigned LeadingKnownZeros = Known.countMinLeadingZeros();
2302  unsigned LeadingKnownOnes = Known.countMinLeadingOnes();
2303 
2304  // Compute the number of leading bits we can ignore.
2305  // TODO: A better way to determine this would use ComputeNumSignBits().
2306  for (auto &C : SI.cases()) {
2307  LeadingKnownZeros = std::min(
2308  LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2309  LeadingKnownOnes = std::min(
2310  LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2311  }
2312 
2313  unsigned NewWidth = Known.getBitWidth() - std::max(LeadingKnownZeros, LeadingKnownOnes);
2314 
2315  // Shrink the condition operand if the new type is smaller than the old type.
2316  // This may produce a non-standard type for the switch, but that's ok because
2317  // the backend should extend back to a legal type for the target.
2318  if (NewWidth > 0 && NewWidth < Known.getBitWidth()) {
2319  IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2320  Builder.SetInsertPoint(&SI);
2321  Value *NewCond = Builder.CreateTrunc(Cond, Ty, "trunc");
2322  SI.setCondition(NewCond);
2323 
2324  for (auto Case : SI.cases()) {
2325  APInt TruncatedCase = Case.getCaseValue()->getValue().trunc(NewWidth);
2326  Case.setValue(ConstantInt::get(SI.getContext(), TruncatedCase));
2327  }
2328  return &SI;
2329  }
2330 
2331  return nullptr;
2332 }
2333 
2335  Value *Agg = EV.getAggregateOperand();
2336 
2337  if (!EV.hasIndices())
2338  return replaceInstUsesWith(EV, Agg);
2339 
2340  if (Value *V = SimplifyExtractValueInst(Agg, EV.getIndices(),
2341  SQ.getWithInstruction(&EV)))
2342  return replaceInstUsesWith(EV, V);
2343 
2344  if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2345  // We're extracting from an insertvalue instruction, compare the indices
2346  const unsigned *exti, *exte, *insi, *inse;
2347  for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2348  exte = EV.idx_end(), inse = IV->idx_end();
2349  exti != exte && insi != inse;
2350  ++exti, ++insi) {
2351  if (*insi != *exti)
2352  // The insert and extract both reference distinctly different elements.
2353  // This means the extract is not influenced by the insert, and we can
2354  // replace the aggregate operand of the extract with the aggregate
2355  // operand of the insert. i.e., replace
2356  // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2357  // %E = extractvalue { i32, { i32 } } %I, 0
2358  // with
2359  // %E = extractvalue { i32, { i32 } } %A, 0
2360  return ExtractValueInst::Create(IV->getAggregateOperand(),
2361  EV.getIndices());
2362  }
2363  if (exti == exte && insi == inse)
2364  // Both iterators are at the end: Index lists are identical. Replace
2365  // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2366  // %C = extractvalue { i32, { i32 } } %B, 1, 0
2367  // with "i32 42"
2368  return replaceInstUsesWith(EV, IV->getInsertedValueOperand());
2369  if (exti == exte) {
2370  // The extract list is a prefix of the insert list. i.e. replace
2371  // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2372  // %E = extractvalue { i32, { i32 } } %I, 1
2373  // with
2374  // %X = extractvalue { i32, { i32 } } %A, 1
2375  // %E = insertvalue { i32 } %X, i32 42, 0
2376  // by switching the order of the insert and extract (though the
2377  // insertvalue should be left in, since it may have other uses).
2378  Value *NewEV = Builder.CreateExtractValue(IV->getAggregateOperand(),
2379  EV.getIndices());
2380  return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2381  makeArrayRef(insi, inse));
2382  }
2383  if (insi == inse)
2384  // The insert list is a prefix of the extract list
2385  // We can simply remove the common indices from the extract and make it
2386  // operate on the inserted value instead of the insertvalue result.
2387  // i.e., replace
2388  // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2389  // %E = extractvalue { i32, { i32 } } %I, 1, 0
2390  // with
2391  // %E extractvalue { i32 } { i32 42 }, 0
2392  return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2393  makeArrayRef(exti, exte));
2394  }
2395  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2396  // We're extracting from an intrinsic, see if we're the only user, which
2397  // allows us to simplify multiple result intrinsics to simpler things that
2398  // just get one value.
2399  if (II->hasOneUse()) {
2400  // Check if we're grabbing the overflow bit or the result of a 'with
2401  // overflow' intrinsic. If it's the latter we can remove the intrinsic
2402  // and replace it with a traditional binary instruction.
2403  switch (II->getIntrinsicID()) {
2404  case Intrinsic::uadd_with_overflow:
2405  case Intrinsic::sadd_with_overflow:
2406  if (*EV.idx_begin() == 0) { // Normal result.
2407  Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2408  replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2409  eraseInstFromFunction(*II);
2410  return BinaryOperator::CreateAdd(LHS, RHS);
2411  }
2412 
2413  // If the normal result of the add is dead, and the RHS is a constant,
2414  // we can transform this into a range comparison.
2415  // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2416  if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2417  if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2418  return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2419  ConstantExpr::getNot(CI));
2420  break;
2421  case Intrinsic::usub_with_overflow:
2422  case Intrinsic::ssub_with_overflow:
2423  if (*EV.idx_begin() == 0) { // Normal result.
2424  Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2425  replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2426  eraseInstFromFunction(*II);
2427  return BinaryOperator::CreateSub(LHS, RHS);
2428  }
2429  break;
2430  case Intrinsic::umul_with_overflow:
2431  case Intrinsic::smul_with_overflow:
2432  if (*EV.idx_begin() == 0) { // Normal result.
2433  Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2434  replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2435  eraseInstFromFunction(*II);
2436  return BinaryOperator::CreateMul(LHS, RHS);
2437  }
2438  break;
2439  default:
2440  break;
2441  }
2442  }
2443  }
2444  if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2445  // If the (non-volatile) load only has one use, we can rewrite this to a
2446  // load from a GEP. This reduces the size of the load. If a load is used
2447  // only by extractvalue instructions then this either must have been
2448  // optimized before, or it is a struct with padding, in which case we
2449  // don't want to do the transformation as it loses padding knowledge.
2450  if (L->isSimple() && L->hasOneUse()) {
2451  // extractvalue has integer indices, getelementptr has Value*s. Convert.
2452  SmallVector<Value*, 4> Indices;
2453  // Prefix an i32 0 since we need the first element.
2454  Indices.push_back(Builder.getInt32(0));
2455  for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2456  I != E; ++I)
2457  Indices.push_back(Builder.getInt32(*I));
2458 
2459  // We need to insert these at the location of the old load, not at that of
2460  // the extractvalue.
2461  Builder.SetInsertPoint(L);
2462  Value *GEP = Builder.CreateInBoundsGEP(L->getType(),
2463  L->getPointerOperand(), Indices);
2464  Instruction *NL = Builder.CreateLoad(GEP);
2465  // Whatever aliasing information we had for the orignal load must also
2466  // hold for the smaller load, so propagate the annotations.
2467  AAMDNodes Nodes;
2468  L->getAAMetadata(Nodes);
2469  NL->setAAMetadata(Nodes);
2470  // Returning the load directly will cause the main loop to insert it in
2471  // the wrong spot, so use replaceInstUsesWith().
2472  return replaceInstUsesWith(EV, NL);
2473  }
2474  // We could simplify extracts from other values. Note that nested extracts may
2475  // already be simplified implicitly by the above: extract (extract (insert) )
2476  // will be translated into extract ( insert ( extract ) ) first and then just
2477  // the value inserted, if appropriate. Similarly for extracts from single-use
2478  // loads: extract (extract (load)) will be translated to extract (load (gep))
2479  // and if again single-use then via load (gep (gep)) to load (gep).
2480  // However, double extracts from e.g. function arguments or return values
2481  // aren't handled yet.
2482  return nullptr;
2483 }
2484 
2485 /// Return 'true' if the given typeinfo will match anything.
2486 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) {
2487  switch (Personality) {
2488  case EHPersonality::GNU_C:
2490  case EHPersonality::Rust:
2491  // The GCC C EH and Rust personality only exists to support cleanups, so
2492  // it's not clear what the semantics of catch clauses are.
2493  return false;
2495  return false;
2497  // While __gnat_all_others_value will match any Ada exception, it doesn't
2498  // match foreign exceptions (or didn't, before gcc-4.7).
2499  return false;
2507  return TypeInfo->isNullValue();
2508  }
2509  llvm_unreachable("invalid enum");
2510 }
2511 
2512 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2513  return
2514  cast<ArrayType>(LHS->getType())->getNumElements()
2515  <
2516  cast<ArrayType>(RHS->getType())->getNumElements();
2517 }
2518 
2520  // The logic here should be correct for any real-world personality function.
2521  // However if that turns out not to be true, the offending logic can always
2522  // be conditioned on the personality function, like the catch-all logic is.
2523  EHPersonality Personality =
2525 
2526  // Simplify the list of clauses, eg by removing repeated catch clauses
2527  // (these are often created by inlining).
2528  bool MakeNewInstruction = false; // If true, recreate using the following:
2529  SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2530  bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2531 
2532  SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2533  for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2534  bool isLastClause = i + 1 == e;
2535  if (LI.isCatch(i)) {
2536  // A catch clause.
2537  Constant *CatchClause = LI.getClause(i);
2538  Constant *TypeInfo = CatchClause->stripPointerCasts();
2539 
2540  // If we already saw this clause, there is no point in having a second
2541  // copy of it.
2542  if (AlreadyCaught.insert(TypeInfo).second) {
2543  // This catch clause was not already seen.
2544  NewClauses.push_back(CatchClause);
2545  } else {
2546  // Repeated catch clause - drop the redundant copy.
2547  MakeNewInstruction = true;
2548  }
2549 
2550  // If this is a catch-all then there is no point in keeping any following
2551  // clauses or marking the landingpad as having a cleanup.
2552  if (isCatchAll(Personality, TypeInfo)) {
2553  if (!isLastClause)
2554  MakeNewInstruction = true;
2555  CleanupFlag = false;
2556  break;
2557  }
2558  } else {
2559  // A filter clause. If any of the filter elements were already caught
2560  // then they can be dropped from the filter. It is tempting to try to
2561  // exploit the filter further by saying that any typeinfo that does not
2562  // occur in the filter can't be caught later (and thus can be dropped).
2563  // However this would be wrong, since typeinfos can match without being
2564  // equal (for example if one represents a C++ class, and the other some
2565  // class derived from it).
2566  assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2567  Constant *FilterClause = LI.getClause(i);
2568  ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2569  unsigned NumTypeInfos = FilterType->getNumElements();
2570 
2571  // An empty filter catches everything, so there is no point in keeping any
2572  // following clauses or marking the landingpad as having a cleanup. By
2573  // dealing with this case here the following code is made a bit simpler.
2574  if (!NumTypeInfos) {
2575  NewClauses.push_back(FilterClause);
2576  if (!isLastClause)
2577  MakeNewInstruction = true;
2578  CleanupFlag = false;
2579  break;
2580  }
2581 
2582  bool MakeNewFilter = false; // If true, make a new filter.
2583  SmallVector<Constant *, 16> NewFilterElts; // New elements.
2584  if (isa<ConstantAggregateZero>(FilterClause)) {
2585  // Not an empty filter - it contains at least one null typeinfo.
2586  assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2587  Constant *TypeInfo =
2588  Constant::getNullValue(FilterType->getElementType());
2589  // If this typeinfo is a catch-all then the filter can never match.
2590  if (isCatchAll(Personality, TypeInfo)) {
2591  // Throw the filter away.
2592  MakeNewInstruction = true;
2593  continue;
2594  }
2595 
2596  // There is no point in having multiple copies of this typeinfo, so
2597  // discard all but the first copy if there is more than one.
2598  NewFilterElts.push_back(TypeInfo);
2599  if (NumTypeInfos > 1)
2600  MakeNewFilter = true;
2601  } else {
2602  ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2603  SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2604  NewFilterElts.reserve(NumTypeInfos);
2605 
2606  // Remove any filter elements that were already caught or that already
2607  // occurred in the filter. While there, see if any of the elements are
2608  // catch-alls. If so, the filter can be discarded.
2609  bool SawCatchAll = false;
2610  for (unsigned j = 0; j != NumTypeInfos; ++j) {
2611  Constant *Elt = Filter->getOperand(j);
2612  Constant *TypeInfo = Elt->stripPointerCasts();
2613  if (isCatchAll(Personality, TypeInfo)) {
2614  // This element is a catch-all. Bail out, noting this fact.
2615  SawCatchAll = true;
2616  break;
2617  }
2618 
2619  // Even if we've seen a type in a catch clause, we don't want to
2620  // remove it from the filter. An unexpected type handler may be
2621  // set up for a call site which throws an exception of the same
2622  // type caught. In order for the exception thrown by the unexpected
2623  // handler to propagate correctly, the filter must be correctly
2624  // described for the call site.
2625  //
2626  // Example:
2627  //
2628  // void unexpected() { throw 1;}
2629  // void foo() throw (int) {
2630  // std::set_unexpected(unexpected);
2631  // try {
2632  // throw 2.0;
2633  // } catch (int i) {}
2634  // }
2635 
2636  // There is no point in having multiple copies of the same typeinfo in
2637  // a filter, so only add it if we didn't already.
2638  if (SeenInFilter.insert(TypeInfo).second)
2639  NewFilterElts.push_back(cast<Constant>(Elt));
2640  }
2641  // A filter containing a catch-all cannot match anything by definition.
2642  if (SawCatchAll) {
2643  // Throw the filter away.
2644  MakeNewInstruction = true;
2645  continue;
2646  }
2647 
2648  // If we dropped something from the filter, make a new one.
2649  if (NewFilterElts.size() < NumTypeInfos)
2650  MakeNewFilter = true;
2651  }
2652  if (MakeNewFilter) {
2653  FilterType = ArrayType::get(FilterType->getElementType(),
2654  NewFilterElts.size());
2655  FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2656  MakeNewInstruction = true;
2657  }
2658 
2659  NewClauses.push_back(FilterClause);
2660 
2661  // If the new filter is empty then it will catch everything so there is
2662  // no point in keeping any following clauses or marking the landingpad
2663  // as having a cleanup. The case of the original filter being empty was
2664  // already handled above.
2665  if (MakeNewFilter && !NewFilterElts.size()) {
2666  assert(MakeNewInstruction && "New filter but not a new instruction!");
2667  CleanupFlag = false;
2668  break;
2669  }
2670  }
2671  }
2672 
2673  // If several filters occur in a row then reorder them so that the shortest
2674  // filters come first (those with the smallest number of elements). This is
2675  // advantageous because shorter filters are more likely to match, speeding up
2676  // unwinding, but mostly because it increases the effectiveness of the other
2677  // filter optimizations below.
2678  for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2679  unsigned j;
2680  // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2681  for (j = i; j != e; ++j)
2682  if (!isa<ArrayType>(NewClauses[j]->getType()))
2683  break;
2684 
2685  // Check whether the filters are already sorted by length. We need to know
2686  // if sorting them is actually going to do anything so that we only make a
2687  // new landingpad instruction if it does.
2688  for (unsigned k = i; k + 1 < j; ++k)
2689  if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2690  // Not sorted, so sort the filters now. Doing an unstable sort would be
2691  // correct too but reordering filters pointlessly might confuse users.
2692  std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2693  shorter_filter);
2694  MakeNewInstruction = true;
2695  break;
2696  }
2697 
2698  // Look for the next batch of filters.
2699  i = j + 1;
2700  }
2701 
2702  // If typeinfos matched if and only if equal, then the elements of a filter L
2703  // that occurs later than a filter F could be replaced by the intersection of
2704  // the elements of F and L. In reality two typeinfos can match without being
2705  // equal (for example if one represents a C++ class, and the other some class
2706  // derived from it) so it would be wrong to perform this transform in general.
2707  // However the transform is correct and useful if F is a subset of L. In that
2708  // case L can be replaced by F, and thus removed altogether since repeating a
2709  // filter is pointless. So here we look at all pairs of filters F and L where
2710  // L follows F in the list of clauses, and remove L if every element of F is
2711  // an element of L. This can occur when inlining C++ functions with exception
2712  // specifications.
2713  for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2714  // Examine each filter in turn.
2715  Value *Filter = NewClauses[i];
2716  ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2717  if (!FTy)
2718  // Not a filter - skip it.
2719  continue;
2720  unsigned FElts = FTy->getNumElements();
2721  // Examine each filter following this one. Doing this backwards means that
2722  // we don't have to worry about filters disappearing under us when removed.
2723  for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2724  Value *LFilter = NewClauses[j];
2725  ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2726  if (!LTy)
2727  // Not a filter - skip it.
2728  continue;
2729  // If Filter is a subset of LFilter, i.e. every element of Filter is also
2730  // an element of LFilter, then discard LFilter.
2731  SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2732  // If Filter is empty then it is a subset of LFilter.
2733  if (!FElts) {
2734  // Discard LFilter.
2735  NewClauses.erase(J);
2736  MakeNewInstruction = true;
2737  // Move on to the next filter.
2738  continue;
2739  }
2740  unsigned LElts = LTy->getNumElements();
2741  // If Filter is longer than LFilter then it cannot be a subset of it.
2742  if (FElts > LElts)
2743  // Move on to the next filter.
2744  continue;
2745  // At this point we know that LFilter has at least one element.
2746  if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2747  // Filter is a subset of LFilter iff Filter contains only zeros (as we
2748  // already know that Filter is not longer than LFilter).
2749  if (isa<ConstantAggregateZero>(Filter)) {
2750  assert(FElts <= LElts && "Should have handled this case earlier!");
2751  // Discard LFilter.
2752  NewClauses.erase(J);
2753  MakeNewInstruction = true;
2754  }
2755  // Move on to the next filter.
2756  continue;
2757  }
2758  ConstantArray *LArray = cast<ConstantArray>(LFilter);
2759  if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2760  // Since Filter is non-empty and contains only zeros, it is a subset of
2761  // LFilter iff LFilter contains a zero.
2762  assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2763  for (unsigned l = 0; l != LElts; ++l)
2764  if (LArray->getOperand(l)->isNullValue()) {
2765  // LFilter contains a zero - discard it.
2766  NewClauses.erase(J);
2767  MakeNewInstruction = true;
2768  break;
2769  }
2770  // Move on to the next filter.
2771  continue;
2772  }
2773  // At this point we know that both filters are ConstantArrays. Loop over
2774  // operands to see whether every element of Filter is also an element of
2775  // LFilter. Since filters tend to be short this is probably faster than
2776  // using a method that scales nicely.
2777  ConstantArray *FArray = cast<ConstantArray>(Filter);
2778  bool AllFound = true;
2779  for (unsigned f = 0; f != FElts; ++f) {
2780  Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2781  AllFound = false;
2782  for (unsigned l = 0; l != LElts; ++l) {
2783  Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2784  if (LTypeInfo == FTypeInfo) {
2785  AllFound = true;
2786  break;
2787  }
2788  }
2789  if (!AllFound)
2790  break;
2791  }
2792  if (AllFound) {
2793  // Discard LFilter.
2794  NewClauses.erase(J);
2795  MakeNewInstruction = true;
2796  }
2797  // Move on to the next filter.
2798  }
2799  }
2800 
2801  // If we changed any of the clauses, replace the old landingpad instruction
2802  // with a new one.
2803  if (MakeNewInstruction) {
2805  NewClauses.size());
2806  for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2807  NLI->addClause(NewClauses[i]);
2808  // A landing pad with no clauses must have the cleanup flag set. It is
2809  // theoretically possible, though highly unlikely, that we eliminated all
2810  // clauses. If so, force the cleanup flag to true.
2811  if (NewClauses.empty())
2812  CleanupFlag = true;
2813  NLI->setCleanup(CleanupFlag);
2814  return NLI;
2815  }
2816 
2817  // Even if none of the clauses changed, we may nonetheless have understood
2818  // that the cleanup flag is pointless. Clear it if so.
2819  if (LI.isCleanup() != CleanupFlag) {
2820  assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2821  LI.setCleanup(CleanupFlag);
2822  return &LI;
2823  }
2824 
2825  return nullptr;
2826 }
2827 
2828 /// Try to move the specified instruction from its current block into the
2829 /// beginning of DestBlock, which can only happen if it's safe to move the
2830 /// instruction past all of the instructions between it and the end of its
2831 /// block.
2832 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2833  assert(I->hasOneUse() && "Invariants didn't hold!");
2834 
2835  // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2836  if (isa<PHINode>(I) || I->isEHPad() || I->mayHaveSideEffects() ||
2837  isa<TerminatorInst>(I))
2838  return false;
2839 
2840  // Do not sink alloca instructions out of the entry block.
2841  if (isa<AllocaInst>(I) && I->getParent() ==
2842  &DestBlock->getParent()->getEntryBlock())
2843  return false;
2844 
2845  // Do not sink into catchswitch blocks.
2846  if (isa<CatchSwitchInst>(DestBlock->getTerminator()))
2847  return false;
2848 
2849  // Do not sink convergent call instructions.
2850  if (auto *CI = dyn_cast<CallInst>(I)) {
2851  if (CI->isConvergent())
2852  return false;
2853  }
2854  // We can only sink load instructions if there is nothing between the load and
2855  // the end of block that could change the value.
2856  if (I->mayReadFromMemory()) {
2857  for (BasicBlock::iterator Scan = I->getIterator(),
2858  E = I->getParent()->end();
2859  Scan != E; ++Scan)
2860  if (Scan->mayWriteToMemory())
2861  return false;
2862  }
2863 
2864  BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2865  I->moveBefore(&*InsertPos);
2866  ++NumSunkInst;
2867  return true;
2868 }
2869 
2871  while (!Worklist.isEmpty()) {
2872  Instruction *I = Worklist.RemoveOne();
2873  if (I == nullptr) continue; // skip null values.
2874 
2875  // Check to see if we can DCE the instruction.
2876  if (isInstructionTriviallyDead(I, &TLI)) {
2877  DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2878  eraseInstFromFunction(*I);
2879  ++NumDeadInst;
2880  MadeIRChange = true;
2881  continue;
2882  }
2883 
2884  // Instruction isn't dead, see if we can constant propagate it.
2885  if (!I->use_empty() &&
2886  (I->getNumOperands() == 0 || isa<Constant>(I->getOperand(0)))) {
2887  if (Constant *C = ConstantFoldInstruction(I, DL, &TLI)) {
2888  DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2889 
2890  // Add operands to the worklist.
2891  replaceInstUsesWith(*I, C);
2892  ++NumConstProp;
2893  if (isInstructionTriviallyDead(I, &TLI))
2894  eraseInstFromFunction(*I);
2895  MadeIRChange = true;
2896  continue;
2897  }
2898  }
2899 
2900  // In general, it is possible for computeKnownBits to determine all bits in
2901  // a value even when the operands are not all constants.
2902  Type *Ty = I->getType();
2903  if (ExpensiveCombines && !I->use_empty() && Ty->isIntOrIntVectorTy()) {
2904  KnownBits Known = computeKnownBits(I, /*Depth*/0, I);
2905  if (Known.isConstant()) {
2906  Constant *C = ConstantInt::get(Ty, Known.getConstant());
2907  DEBUG(dbgs() << "IC: ConstFold (all bits known) to: " << *C <<
2908  " from: " << *I << '\n');
2909 
2910  // Add operands to the worklist.
2911  replaceInstUsesWith(*I, C);
2912  ++NumConstProp;
2913  if (isInstructionTriviallyDead(I, &TLI))
2914  eraseInstFromFunction(*I);
2915  MadeIRChange = true;
2916  continue;
2917  }
2918  }
2919 
2920  // See if we can trivially sink this instruction to a successor basic block.
2921  if (I->hasOneUse()) {
2922  BasicBlock *BB = I->getParent();
2923  Instruction *UserInst = cast<Instruction>(*I->user_begin());
2924  BasicBlock *UserParent;
2925 
2926  // Get the block the use occurs in.
2927  if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2928  UserParent = PN->getIncomingBlock(*I->use_begin());
2929  else
2930  UserParent = UserInst->getParent();
2931 
2932  if (UserParent != BB) {
2933  bool UserIsSuccessor = false;
2934  // See if the user is one of our successors.
2935  for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2936  if (*SI == UserParent) {
2937  UserIsSuccessor = true;
2938  break;
2939  }
2940 
2941  // If the user is one of our immediate successors, and if that successor
2942  // only has us as a predecessors (we'd have to split the critical edge
2943  // otherwise), we can keep going.
2944  if (UserIsSuccessor && UserParent->getUniquePredecessor()) {
2945  // Okay, the CFG is simple enough, try to sink this instruction.
2946  if (TryToSinkInstruction(I, UserParent)) {
2947  DEBUG(dbgs() << "IC: Sink: " << *I << '\n');
2948  MadeIRChange = true;
2949  // We'll add uses of the sunk instruction below, but since sinking
2950  // can expose opportunities for it's *operands* add them to the
2951  // worklist
2952  for (Use &U : I->operands())
2953  if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
2954  Worklist.Add(OpI);
2955  }
2956  }
2957  }
2958  }
2959 
2960  // Now that we have an instruction, try combining it to simplify it.
2961  Builder.SetInsertPoint(I);
2962  Builder.SetCurrentDebugLocation(I->getDebugLoc());
2963 
2964 #ifndef NDEBUG
2965  std::string OrigI;
2966 #endif
2967  DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2968  DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
2969 
2970  if (Instruction *Result = visit(*I)) {
2971  ++NumCombined;
2972  // Should we replace the old instruction with a new one?
2973  if (Result != I) {
2974  DEBUG(dbgs() << "IC: Old = " << *I << '\n'
2975  << " New = " << *Result << '\n');
2976 
2977  if (I->getDebugLoc())
2978  Result->setDebugLoc(I->getDebugLoc());
2979  // Everything uses the new instruction now.
2980  I->replaceAllUsesWith(Result);
2981 
2982  // Move the name to the new instruction first.
2983  Result->takeName(I);
2984 
2985  // Push the new instruction and any users onto the worklist.
2986  Worklist.AddUsersToWorkList(*Result);
2987  Worklist.Add(Result);
2988 
2989  // Insert the new instruction into the basic block...
2990  BasicBlock *InstParent = I->getParent();
2991  BasicBlock::iterator InsertPos = I->getIterator();
2992 
2993  // If we replace a PHI with something that isn't a PHI, fix up the
2994  // insertion point.
2995  if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2996  InsertPos = InstParent->getFirstInsertionPt();
2997 
2998  InstParent->getInstList().insert(InsertPos, Result);
2999 
3000  eraseInstFromFunction(*I);
3001  } else {
3002  DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
3003  << " New = " << *I << '\n');
3004 
3005  // If the instruction was modified, it's possible that it is now dead.
3006  // if so, remove it.
3007  if (isInstructionTriviallyDead(I, &TLI)) {
3008  eraseInstFromFunction(*I);
3009  } else {
3010  Worklist.AddUsersToWorkList(*I);
3011  Worklist.Add(I);
3012  }
3013  }
3014  MadeIRChange = true;
3015  }
3016  }
3017 
3018  Worklist.Zap();
3019  return MadeIRChange;
3020 }
3021 
3022 /// Walk the function in depth-first order, adding all reachable code to the
3023 /// worklist.
3024 ///
3025 /// This has a couple of tricks to make the code faster and more powerful. In
3026 /// particular, we constant fold and DCE instructions as we go, to avoid adding
3027 /// them to the worklist (this significantly speeds up instcombine on code where
3028 /// many instructions are dead or constant). Additionally, if we find a branch
3029 /// whose condition is a known constant, we only visit the reachable successors.
3030 ///
3033  InstCombineWorklist &ICWorklist,
3034  const TargetLibraryInfo *TLI) {
3035  bool MadeIRChange = false;
3037  Worklist.push_back(BB);
3038 
3039  SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
3040  DenseMap<Constant *, Constant *> FoldedConstants;
3041 
3042  do {
3043  BB = Worklist.pop_back_val();
3044 
3045  // We have now visited this block! If we've already been here, ignore it.
3046  if (!Visited.insert(BB).second)
3047  continue;
3048 
3049  for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
3050  Instruction *Inst = &*BBI++;
3051 
3052  // DCE instruction if trivially dead.
3053  if (isInstructionTriviallyDead(Inst, TLI)) {
3054  ++NumDeadInst;
3055  DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
3056  Inst->eraseFromParent();
3057  MadeIRChange = true;
3058  continue;
3059  }
3060 
3061  // ConstantProp instruction if trivially constant.
3062  if (!Inst->use_empty() &&
3063  (Inst->getNumOperands() == 0 || isa<Constant>(Inst->getOperand(0))))
3064  if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
3065  DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
3066  << *Inst << '\n');
3067  Inst->replaceAllUsesWith(C);
3068  ++NumConstProp;
3069  if (isInstructionTriviallyDead(Inst, TLI))
3070  Inst->eraseFromParent();
3071  MadeIRChange = true;
3072  continue;
3073  }
3074 
3075  // See if we can constant fold its operands.
3076  for (Use &U : Inst->operands()) {
3077  if (!isa<ConstantVector>(U) && !isa<ConstantExpr>(U))
3078  continue;
3079 
3080  auto *C = cast<Constant>(U);
3081  Constant *&FoldRes = FoldedConstants[C];
3082  if (!FoldRes)
3083  FoldRes = ConstantFoldConstant(C, DL, TLI);
3084  if (!FoldRes)
3085  FoldRes = C;
3086 
3087  if (FoldRes != C) {
3088  DEBUG(dbgs() << "IC: ConstFold operand of: " << *Inst
3089  << "\n Old = " << *C
3090  << "\n New = " << *FoldRes << '\n');
3091  U = FoldRes;
3092  MadeIRChange = true;
3093  }
3094  }
3095 
3096  // Skip processing debug intrinsics in InstCombine. Processing these call instructions
3097  // consumes non-trivial amount of time and provides no value for the optimization.
3098  if (!isa<DbgInfoIntrinsic>(Inst))
3099  InstrsForInstCombineWorklist.push_back(Inst);
3100  }
3101 
3102  // Recursively visit successors. If this is a branch or switch on a
3103  // constant, only visit the reachable successor.
3104  TerminatorInst *TI = BB->getTerminator();
3105  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3106  if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
3107  bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
3108  BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
3109  Worklist.push_back(ReachableBB);
3110  continue;
3111  }
3112  } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3113  if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
3114  Worklist.push_back(SI->findCaseValue(Cond)->getCaseSuccessor());
3115  continue;
3116  }
3117  }
3118 
3119  for (BasicBlock *SuccBB : TI->successors())
3120  Worklist.push_back(SuccBB);
3121  } while (!Worklist.empty());
3122 
3123  // Once we've found all of the instructions to add to instcombine's worklist,
3124  // add them in reverse order. This way instcombine will visit from the top
3125  // of the function down. This jives well with the way that it adds all uses
3126  // of instructions to the worklist after doing a transformation, thus avoiding
3127  // some N^2 behavior in pathological cases.
3128  ICWorklist.AddInitialGroup(InstrsForInstCombineWorklist);
3129 
3130  return MadeIRChange;
3131 }
3132 
3133 /// \brief Populate the IC worklist from a function, and prune any dead basic
3134 /// blocks discovered in the process.
3135 ///
3136 /// This also does basic constant propagation and other forward fixing to make
3137 /// the combiner itself run much faster.
3139  TargetLibraryInfo *TLI,
3140  InstCombineWorklist &ICWorklist) {
3141  bool MadeIRChange = false;
3142 
3143  // Do a depth-first traversal of the function, populate the worklist with
3144  // the reachable instructions. Ignore blocks that are not reachable. Keep
3145  // track of which blocks we visit.
3147  MadeIRChange |=
3148  AddReachableCodeToWorklist(&F.front(), DL, Visited, ICWorklist, TLI);
3149 
3150  // Do a quick scan over the function. If we find any blocks that are
3151  // unreachable, remove any instructions inside of them. This prevents
3152  // the instcombine code from having to deal with some bad special cases.
3153  for (BasicBlock &BB : F) {
3154  if (Visited.count(&BB))
3155  continue;
3156 
3157  unsigned NumDeadInstInBB = removeAllNonTerminatorAndEHPadInstructions(&BB);
3158  MadeIRChange |= NumDeadInstInBB > 0;
3159  NumDeadInst += NumDeadInstInBB;
3160  }
3161 
3162  return MadeIRChange;
3163 }
3164 
3165 static bool
3167  AliasAnalysis *AA, AssumptionCache &AC,
3168  TargetLibraryInfo &TLI, DominatorTree &DT,
3169  bool ExpensiveCombines = true,
3170  LoopInfo *LI = nullptr) {
3171  auto &DL = F.getParent()->getDataLayout();
3172  ExpensiveCombines |= EnableExpensiveCombines;
3173 
3174  /// Builder - This is an IRBuilder that automatically inserts new
3175  /// instructions into the worklist when they are created.
3177  F.getContext(), TargetFolder(DL),
3178  IRBuilderCallbackInserter([&Worklist, &AC](Instruction *I) {
3179  Worklist.Add(I);
3180 
3181  using namespace llvm::PatternMatch;
3182  if (match(I, m_Intrinsic<Intrinsic::assume>()))
3183  AC.registerAssumption(cast<CallInst>(I));
3184  }));
3185 
3186  // Lower dbg.declare intrinsics otherwise their value may be clobbered
3187  // by instcombiner.
3188  bool MadeIRChange = LowerDbgDeclare(F);
3189 
3190  // Iterate while there is work to do.
3191  int Iteration = 0;
3192  for (;;) {
3193  ++Iteration;
3194  DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
3195  << F.getName() << "\n");
3196 
3197  MadeIRChange |= prepareICWorklistFromFunction(F, DL, &TLI, Worklist);
3198 
3199  InstCombiner IC(Worklist, Builder, F.optForMinSize(), ExpensiveCombines,
3200  AA, AC, TLI, DT, DL, LI);
3202 
3203  if (!IC.run())
3204  break;
3205  }
3206 
3207  return MadeIRChange || Iteration > 1;
3208 }
3209 
3212  auto &AC = AM.getResult<AssumptionAnalysis>(F);
3213  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
3214  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
3215 
3216  auto *LI = AM.getCachedResult<LoopAnalysis>(F);
3217 
3218  // FIXME: The AliasAnalysis is not yet supported in the new pass manager
3219  if (!combineInstructionsOverFunction(F, Worklist, nullptr, AC, TLI, DT,
3220  ExpensiveCombines, LI))
3221  // No changes, all analyses are preserved.
3222  return PreservedAnalyses::all();
3223 
3224  // Mark all the analyses that instcombine updates as preserved.
3225  PreservedAnalyses PA;
3226  PA.preserveSet<CFGAnalyses>();
3227  PA.preserve<AAManager>();
3228  PA.preserve<GlobalsAA>();
3229  return PA;
3230 }
3231 
3233  AU.setPreservesCFG();
3242 }
3243 
3245  if (skipFunction(F))
3246  return false;
3247 
3248  // Required analyses.
3249  auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
3250  auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3251  auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
3252  auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3253 
3254  // Optional analyses.
3255  auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
3256  auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
3257 
3258  return combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT,
3259  ExpensiveCombines, LI);
3260 }
3261 
3264  "Combine redundant instructions", false, false)
3271  "Combine redundant instructions", false, false)
3272 
3273 // Initialization Routines
3276 }
3277 
3280 }
3281 
3283  return new InstructionCombiningPass(ExpensiveCombines);
3284 }
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.
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:735
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:201
uint64_t CallInst * C
Return a value (possibly void), from a function.
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:69
void push_back(const T &Elt)
Definition: SmallVector.h:212
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:109
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...
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:850
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:841
Value * CreateBinOp(Instruction::BinaryOps Opc, Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1108
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:490
static const Value * getFNegArgument(const Value *BinOp)
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:687
void LLVMInitializeInstCombine(LLVMPassRegistryRef R)
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
BinaryOps getOpcode() const
Definition: InstrTypes.h:523
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:63
bool isSized(SmallPtrSetImpl< Type *> *Visited=nullptr) const
Return true if it makes sense to take the size of this type.
Definition: Type.h:262
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
static Value * foldOperationIntoPhiValue(BinaryOperator *I, Value *InV, InstCombiner::BuilderTy &Builder)
static void ClearSubclassDataAfterReassociation(BinaryOperator &I)
Conservatively clears subclassOptionalData after a reassociation or commutation.
match_zero m_Zero()
Match an arbitrary zero/null constant.
Definition: PatternMatch.h:145
This file provides the primary interface to the instcombine pass.
static GetElementPtrInst * Create(Type *PointeeType, Value *Ptr, ArrayRef< Value *> IdxList, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Definition: Instructions.h:863
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
This class represents a function call, abstracting a target machine&#39;s calling convention.
static Constant * getBinOpIdentity(unsigned Opcode, Type *Ty)
Return the identity for the given binary operation, i.e.
Definition: Constants.cpp:2203
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:580
A cache of .assume calls within a function.
static void sdivrem(const APInt &LHS, const APInt &RHS, APInt &Quotient, APInt &Remainder)
Definition: APInt.cpp:1835
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:697
struct LLVMOpaquePassRegistry * LLVMPassRegistryRef
Definition: Types.h:117
void Add(Instruction *I)
Add - Add the specified instruction to the worklist if it isn&#39;t already in it.
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)".
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:818
STATISTIC(NumFunctions, "Total number of functions")
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:232
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:164
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:1832
Hexagon Common GEP
Value * getCondition() const
static Constant * getSub(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2126
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:227
iv Induction Variable Users
Definition: IVUsers.cpp:51
Constant * getClause(unsigned Idx) const
Get the value of the clause at index Idx.
void reserve(size_type N)
Definition: SmallVector.h:380
idx_iterator idx_end() const
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:689
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:156
op_iterator op_begin()
Definition: User.h:214
static Constant * get(ArrayType *T, ArrayRef< Constant *> V)
Definition: Constants.cpp:888
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:1488
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:207
A templated base class for SmallPtrSet which provides the typesafe interface that is common across al...
Definition: SmallPtrSet.h:345
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:252
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
Constant * getMask() const
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:493
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
static const Value * getNegArgument(const Value *BinOp)
Helper functions to extract the unary argument of a NEG, FNEG or NOT operation implemented via Sub...
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:958
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:361
Attribute unwrap(LLVMAttributeRef Attr)
Definition: Attributes.h:195
bool isNonNegative() const
Determine if this APInt Value is non-negative (>= 0)
Definition: APInt.h:362
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:560
ArrayRef< T > makeArrayRef(const T &OneElt)
Construct an ArrayRef from a single element.
Definition: ArrayRef.h:451
bool isFloatingPointTy() const
Return true if this is one of the six floating-point types.
Definition: Type.h:162
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
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:664
The core instruction combiner logic.
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:820
static bool shorter_filter(const Value *LHS, const Value *RHS)
uint64_t getNumElements() const
Definition: DerivedTypes.h:359
Type * getSourceElementType() const
Definition: Instructions.h:934
unsigned getNumClauses() const
Get the number of clauses for this landing pad.
not_match< LHS > m_Not(const LHS &L)
Definition: PatternMatch.h:961
BinaryOp_match< LHS, RHS, Instruction::Add > m_Add(const LHS &L, const RHS &R)
Definition: PatternMatch.h:478
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)
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 Constant * get(unsigned Opcode, Constant *C1, Constant *C2, unsigned Flags=0, Type *OnlyIfReducedTy=nullptr)
get - Return a binary or shift operator constant expression, folding if possible. ...
Definition: Constants.cpp:1711
Instruction * clone() const
Create a copy of &#39;this&#39; instruction that is identical in all ways except the following: ...
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:86
int64_t getSExtValue() const
Get sign extended value.
Definition: APInt.h:1554
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:820
#define F(x, y, z)
Definition: MD5.cpp:55
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)
bool isInBounds() const
Determine whether the GEP has the inbounds flag.
Class to represent array types.
Definition: DerivedTypes.h:369
int32_t exactLogBase2() const
Definition: APInt.h:1767
This class represents a no-op cast from one type to another.
op_iterator idx_begin()
Definition: Instructions.h:962
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:121
TargetFolder - Create constants with target dependent folding.
Definition: TargetFolder.h:32
An instruction for storing to memory.
Definition: Instructions.h:306
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:428
FunctionPass * createInstructionCombiningPass(bool ExpensiveCombines=true)
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:290
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:134
Function * getDeclaration(Module *M, ID id, ArrayRef< Type *> Tys=None)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
Definition: Function.cpp:975
static BinaryOperator * CreateAdd(Value *S1, Value *S2, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp)
Value * getOperand(unsigned i) const
Definition: User.h:154
Class to represent pointers.
Definition: DerivedTypes.h:467
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:946
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:277
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return &#39;this&#39;.
Definition: Type.h:301
const BasicBlock & getEntryBlock() const
Definition: Function.h:564
an instruction for type-safe pointer arithmetic to access elements of arrays and structs ...
Definition: Instructions.h:837
succ_range successors()
Definition: InstrTypes.h:267
OneUse_match< T > m_OneUse(const T &SubPattern)
Definition: PatternMatch.h:63
static Constant * getFNeg(Constant *C)
Definition: Constants.cpp:2103
bool hasAllZeroIndices() const
Return true if all of the indices of this GEP are zeros.
Definition: Operator.h:454
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:404
The landingpad instruction holds all of the information necessary to generate correct exception handl...
static Instruction::BinaryOps getBinOpsForFactorization(Instruction::BinaryOps TopLevelOpcode, BinaryOperator *Op, Value *&LHS, Value *&RHS)
This function factors binary ops which can be combined using distributive laws.
Subclasses of this class are all able to terminate a basic block.
Definition: InstrTypes.h:54
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
apint_match m_APInt(const APInt *&Res)
Match a ConstantInt or splatted ConstantVector, binding the specified pointer to the contained APInt...
Definition: PatternMatch.h:240
const_iterator getFirstInsertionPt() const
Returns an iterator to the first instruction in this block that is suitable for inserting a non-PHI i...
Definition: BasicBlock.cpp:200
BinaryOp_match< LHS, RHS, Instruction::SDiv > m_SDiv(const LHS &L, const RHS &R)
Definition: PatternMatch.h:520
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:217
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:1253
ConstantInt * lowerObjectSizeCall(IntrinsicInst *ObjectSize, const DataLayout &DL, const TargetLibraryInfo *TLI, bool MustSucceed)
Try to turn a call to .objectsize into an integer value of the given Type.
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
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.
This is an important base class in LLVM.
Definition: Constant.h:42
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:116
const APInt & getConstant() const
Returns the value when all bits have a known value.
Definition: KnownBits.h:57
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)".
A manager for alias analyses.
#define A
Definition: LargeTest.cpp:12
ConstantFP - Floating Point Values [float, double].
Definition: Constants.h:264
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:372
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:500
APInt ssub_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1886
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:262
unsigned getPointerAddressSpace() const
Returns the address space of the pointer operand.
Definition: Instructions.h:992
Value * getRawDest() const
specificval_ty m_Specific(const Value *V)
Match if we have a specific specified value.
Definition: PatternMatch.h:358
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:436
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:216
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:860
bool isBinaryOp() const
Definition: Instruction.h:125
Utility class for integer arithmetic operators which may exhibit overflow - Add, Sub, and Mul.
Definition: Operator.h:67
void print(raw_ostream &O, bool IsForDebug=false) const
Implement operator<< on Value.
Definition: AsmWriter.cpp:3459
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:298
op_range operands()
Definition: User.h:222
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:186
self_iterator getIterator()
Definition: ilist_node.h:82
Class to represent integer types.
Definition: DerivedTypes.h:40
Constant Vector Declarations.
Definition: Constants.h:491
The legacy pass manager&#39;s instcombine pass.
Definition: InstCombine.h:44
static Constant * getNot(Constant *C)
Definition: Constants.cpp:2109
void setSourceElementType(Type *Ty)
Definition: Instructions.h:936
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:987
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:1320
const Constant * stripPointerCasts() const
Definition: Constant.h:153
const AMDGPUAS & AS
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs, and aliases.
Definition: Value.cpp:527
iterator erase(const_iterator CI)
Definition: SmallVector.h:449
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
bool runOnFunction(Function &F) override
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass...
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:1193
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:244
neg_match< LHS > m_Neg(const LHS &L)
Match an integer negate.
Definition: PatternMatch.h:984
A function analysis which provides an AssumptionCache.
const InstListType & getInstList() const
Return the underlying instruction list container.
Definition: BasicBlock.h:317
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:176
static PointerType * getInt1PtrTy(LLVMContext &C, unsigned AS=0)
Definition: Type.cpp:216
#define E
Definition: LargeTest.cpp:27
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...
#define B
Definition: LargeTest.cpp:24
iterator end()
Definition: BasicBlock.h:254
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:864
Utility class for floating point operations which can have information about relaxed accuracy require...
Definition: Operator.h:220
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:63
Provides information about what library functions are available for the current target.
static Value * CreateBinOpAsGiven(BinaryOperator &Inst, Value *LHS, Value *RHS, InstCombiner::BuilderTy &B)
Creates node of binary operation with the same attributes as the specified one but with other operand...
A collection of metadata nodes that might be associated with a memory access used by the alias-analys...
Definition: Metadata.h:642
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:385
uint64_t getSizeInBytes() const
Definition: DataLayout.h:501
Instruction * visitFree(CallInst &FI)
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:560
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:278
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:1372
unsigned getNumIncomingValues() const
Return the number of incoming edges.
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:516
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:416
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:939
void setOperand(unsigned i, Value *Val)
Definition: User.h:159
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:923
unsigned getVectorNumElements() const
Definition: DerivedTypes.h:462
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:545
static bool isNeg(const Value *V)
Check if the given Value is a NEG, FNeg, or NOT instruction.
ConstantArray - Constant Array Declarations.
Definition: Constants.h:405
Class for arbitrary precision integers.
Definition: APInt.h:69
bool ule(const APInt &RHS) const
Unsigned less or equal comparison.
Definition: APInt.h:1202
static bool combineInstructionsOverFunction(Function &F, InstCombineWorklist &Worklist, AliasAnalysis *AA, AssumptionCache &AC, TargetLibraryInfo &TLI, DominatorTree &DT, bool ExpensiveCombines=true, LoopInfo *LI=nullptr)
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.
Definition: PatternMatch.h:882
typename SuperClass::iterator iterator
Definition: SmallVector.h:328
iterator_range< user_iterator > users()
Definition: Value.h:395
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:1435
Represents analyses that only rely on functions&#39; control flow.
Definition: PassManager.h:114
unsigned countMinLeadingOnes() const
Returns the minimum number of leading one bits.
Definition: KnownBits.h:153
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:398
Common super class of ArrayType, StructType and VectorType.
Definition: DerivedTypes.h:162
Instruction * visitLandingPadInst(LandingPadInst &LI)
use_iterator use_begin()
Definition: Value.h:334
static Constant * getNeg(Constant *C, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2096
static std::vector< std::string > Flags
Definition: FlagsTest.cpp:8
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 ...
Provides an &#39;InsertHelper&#39; that calls a user-provided callback after performing the default insertion...
Definition: IRBuilder.h:74
bool isVolatile() const
Return true if this is a store to a volatile memory location.
Definition: Instructions.h:339
iterator insert(iterator where, pointer New)
Definition: ilist.h:241
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:280
void registerAssumption(CallInst *CI)
Add an .assume intrinsic to this function&#39;s cache.
uint64_t getElementOffset(unsigned Idx) const
Definition: DataLayout.h:515
void emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:656
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:61
static bool isFNeg(const Value *V, bool IgnoreZeroSign=false)
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:189
Value * getArgOperand(unsigned i) const
getArgOperand/setArgOperand - Return/set the i-th call argument.
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:218
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:108
#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:519
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:706
static ArrayType * get(Type *ElementType, uint64_t NumElements)
This static method is the primary way to construct an ArrayType.
Definition: Type.cpp:568
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:2186
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:174
bool isUnconditional() const
void initializeInstructionCombiningPassPass(PassRegistry &)
bool isMinValue() const
Determine if this is the smallest unsigned value.
Definition: APInt.h:430
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:1476
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:897
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:371
const BasicBlock & front() const
Definition: Function.h:587
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:463
APInt sadd_ov(const APInt &RHS, bool &Overflow) const
Definition: APInt.cpp:1873
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:545
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:293
LLVM Value Representation.
Definition: Value.h:73
Constant * getPersonalityFn() const
Get the personality function associated with this function.
Definition: Function.cpp:1255
This file provides internal interfaces used to implement the InstCombine.
void clearSubclassOptionalData()
Clear the optional flags contained in this value.
Definition: Value.h:471
static VectorType * get(Type *ElementType, unsigned NumElements)
This static method is the primary way to construct an VectorType.
Definition: Type.cpp:593
#define LLVM_FALLTHROUGH
LLVM_FALLTHROUGH - Mark fallthrough cases in switch statements.
Definition: Compiler.h:235
void moveBefore(Instruction *MovePos)
Unlink this instruction from its current basic block and insert it into the basic block that MovePos ...
Definition: Instruction.cpp:88
static Instruction * tryToMoveFreeBeforeNullTest(CallInst &FI)
Move the call to free before a NULL test.
Invoke instruction.
#define DEBUG(X)
Definition: Debug.h:118
Instruction * visitGetElementPtrInst(GetElementPtrInst &GEP)
unsigned countMinLeadingZeros() const
Returns the minimum number of leading zero bits.
Definition: KnownBits.h:148
bool isEHPad() const
Return true if the instruction is a variety of EH-block.
Definition: Instruction.h:503
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:408
Convenience struct for specifying and reasoning about fast-math flags.
Definition: Operator.h:160
unsigned greater than
Definition: InstrTypes.h:883
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:40
A container for analyses that lazily runs them and caches their results.
Type * getArrayElementType() const
Definition: Type.h:362
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:261
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
const TerminatorInst * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:120
static BinaryOperator * CreateMul(Value *S1, Value *S2, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp)
static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src)
static void getShuffleMask(Constant *Mask, SmallVectorImpl< int > &Result)
Convert the input shuffle mask operand to a vector of integers.
VectorType * getType() const
Overload to return most specific vector type.
Value * getPointerOperand()
Definition: Instructions.h:398
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)
Determine which bits of V are known to be either zero or one and return them in the KnownZero/KnownOn...
Instruction * visitAllocSite(Instruction &FI)
bool use_empty() const
Definition: Value.h:322
static Constant * get(ArrayRef< Constant *> V)
Definition: Constants.cpp:984
#define D
Definition: LargeTest.cpp:26
Type * getElementType() const
Definition: DerivedTypes.h:486
bool isStructTy() const
True if this is an instance of StructType.
Definition: Type.h:215
bool isArrayTy() const
True if this is an instance of ArrayType.
Definition: Type.h:218
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:66
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)
Definition: PatternMatch.h:813
Legacy wrapper pass to provide the BasicAAResult object.
gep_type_iterator gep_type_begin(const User *GEP)
user_iterator user_end()
Definition: Value.h:379