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-c/Transforms/InstCombine.h"
39 #include "llvm/ADT/APInt.h"
40 #include "llvm/ADT/ArrayRef.h"
41 #include "llvm/ADT/DenseMap.h"
42 #include "llvm/ADT/None.h"
43 #include "llvm/ADT/SmallPtrSet.h"
44 #include "llvm/ADT/SmallVector.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/ADT/TinyPtrVector.h"
47 #include "llvm/Analysis/AliasAnalysis.h"
48 #include "llvm/Analysis/AssumptionCache.h"
49 #include "llvm/Analysis/BasicAliasAnalysis.h"
50 #include "llvm/Analysis/CFG.h"
51 #include "llvm/Analysis/ConstantFolding.h"
52 #include "llvm/Analysis/EHPersonalities.h"
53 #include "llvm/Analysis/GlobalsModRef.h"
54 #include "llvm/Analysis/InstructionSimplify.h"
55 #include "llvm/Analysis/LoopInfo.h"
56 #include "llvm/Analysis/MemoryBuiltins.h"
57 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
58 #include "llvm/Analysis/TargetFolder.h"
59 #include "llvm/Analysis/TargetLibraryInfo.h"
60 #include "llvm/Transforms/Utils/Local.h"
61 #include "llvm/Analysis/ValueTracking.h"
62 #include "llvm/IR/BasicBlock.h"
63 #include "llvm/IR/CFG.h"
64 #include "llvm/IR/Constant.h"
65 #include "llvm/IR/Constants.h"
66 #include "llvm/IR/DIBuilder.h"
67 #include "llvm/IR/DataLayout.h"
68 #include "llvm/IR/DerivedTypes.h"
69 #include "llvm/IR/Dominators.h"
70 #include "llvm/IR/Function.h"
71 #include "llvm/IR/GetElementPtrTypeIterator.h"
72 #include "llvm/IR/IRBuilder.h"
73 #include "llvm/IR/InstrTypes.h"
74 #include "llvm/IR/Instruction.h"
75 #include "llvm/IR/Instructions.h"
76 #include "llvm/IR/IntrinsicInst.h"
77 #include "llvm/IR/Intrinsics.h"
78 #include "llvm/IR/LegacyPassManager.h"
79 #include "llvm/IR/Metadata.h"
80 #include "llvm/IR/Operator.h"
81 #include "llvm/IR/PassManager.h"
82 #include "llvm/IR/PatternMatch.h"
83 #include "llvm/IR/Type.h"
84 #include "llvm/IR/Use.h"
85 #include "llvm/IR/User.h"
86 #include "llvm/IR/Value.h"
87 #include "llvm/IR/ValueHandle.h"
88 #include "llvm/Pass.h"
89 #include "llvm/Support/CBindingWrapping.h"
90 #include "llvm/Support/Casting.h"
91 #include "llvm/Support/CommandLine.h"
92 #include "llvm/Support/Compiler.h"
93 #include "llvm/Support/Debug.h"
94 #include "llvm/Support/DebugCounter.h"
95 #include "llvm/Support/ErrorHandling.h"
96 #include "llvm/Support/KnownBits.h"
97 #include "llvm/Support/raw_ostream.h"
98 #include "llvm/Transforms/InstCombine/InstCombine.h"
99 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h"
100 #include <algorithm>
101 #include <cassert>
102 #include <cstdint>
103 #include <memory>
104 #include <string>
105 #include <utility>
106
107 using namespace llvm;
108 using namespace llvm::PatternMatch;
109
110 #define DEBUG_TYPE "instcombine"
111
112 STATISTIC(NumCombined , "Number of insts combined");
113 STATISTIC(NumConstProp, "Number of constant folds");
114 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
115 STATISTIC(NumSunkInst , "Number of instructions sunk");
116 STATISTIC(NumExpand, "Number of expansions");
117 STATISTIC(NumFactor , "Number of factorizations");
118 STATISTIC(NumReassoc , "Number of reassociations");
119 DEBUG_COUNTER(VisitCounter, "instcombine-visit",
120 "Controls which instructions are visited");
121
122 static cl::opt<bool>
123 EnableExpensiveCombines("expensive-combines",
124 cl::desc("Enable expensive instruction combines"));
125
126 static cl::opt<unsigned>
127 MaxArraySize("instcombine-maxarray-size", cl::init(1024),
128 cl::desc("Maximum array size considered when doing a combine"));
129
130 // FIXME: Remove this flag when it is no longer necessary to convert
131 // llvm.dbg.declare to avoid inaccurate debug info. Setting this to false
132 // increases variable availability at the cost of accuracy. Variables that
133 // cannot be promoted by mem2reg or SROA will be described as living in memory
134 // for their entire lifetime. However, passes like DSE and instcombine can
135 // delete stores to the alloca, leading to misleading and inaccurate debug
136 // information. This flag can be removed when those passes are fixed.
137 static cl::opt<unsigned> ShouldLowerDbgDeclare("instcombine-lower-dbg-declare",
138 cl::Hidden, cl::init(true));
139
EmitGEPOffset(User * GEP)140 Value *InstCombiner::EmitGEPOffset(User *GEP) {
141 return llvm::EmitGEPOffset(&Builder, DL, GEP);
142 }
143
144 /// Return true if it is desirable to convert an integer computation from a
145 /// given bit width to a new bit width.
146 /// We don't want to convert from a legal to an illegal type or from a smaller
147 /// to a larger illegal type. A width of '1' is always treated as a legal type
148 /// because i1 is a fundamental type in IR, and there are many specialized
149 /// optimizations for i1 types. Widths of 8, 16 or 32 are equally treated as
150 /// legal to convert to, in order to open up more combining opportunities.
151 /// NOTE: this treats i8, i16 and i32 specially, due to them being so common
152 /// from frontend languages.
shouldChangeType(unsigned FromWidth,unsigned ToWidth) const153 bool InstCombiner::shouldChangeType(unsigned FromWidth,
154 unsigned ToWidth) const {
155 bool FromLegal = FromWidth == 1 || DL.isLegalInteger(FromWidth);
156 bool ToLegal = ToWidth == 1 || DL.isLegalInteger(ToWidth);
157
158 // Convert to widths of 8, 16 or 32 even if they are not legal types. Only
159 // shrink types, to prevent infinite loops.
160 if (ToWidth < FromWidth && (ToWidth == 8 || ToWidth == 16 || ToWidth == 32))
161 return true;
162
163 // If this is a legal integer from type, and the result would be an illegal
164 // type, don't do the transformation.
165 if (FromLegal && !ToLegal)
166 return false;
167
168 // Otherwise, if both are illegal, do not increase the size of the result. We
169 // do allow things like i160 -> i64, but not i64 -> i160.
170 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
171 return false;
172
173 return true;
174 }
175
176 /// Return true if it is desirable to convert a computation from 'From' to 'To'.
177 /// We don't want to convert from a legal to an illegal type or from a smaller
178 /// to a larger illegal type. i1 is always treated as a legal type because it is
179 /// a fundamental type in IR, and there are many specialized optimizations for
180 /// i1 types.
shouldChangeType(Type * From,Type * To) const181 bool InstCombiner::shouldChangeType(Type *From, Type *To) const {
182 assert(From->isIntegerTy() && To->isIntegerTy());
183
184 unsigned FromWidth = From->getPrimitiveSizeInBits();
185 unsigned ToWidth = To->getPrimitiveSizeInBits();
186 return shouldChangeType(FromWidth, ToWidth);
187 }
188
189 // Return true, if No Signed Wrap should be maintained for I.
190 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
191 // where both B and C should be ConstantInts, results in a constant that does
192 // not overflow. This function only handles the Add and Sub opcodes. For
193 // all other opcodes, the function conservatively returns false.
MaintainNoSignedWrap(BinaryOperator & I,Value * B,Value * C)194 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
195 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
196 if (!OBO || !OBO->hasNoSignedWrap())
197 return false;
198
199 // We reason about Add and Sub Only.
200 Instruction::BinaryOps Opcode = I.getOpcode();
201 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
202 return false;
203
204 const APInt *BVal, *CVal;
205 if (!match(B, m_APInt(BVal)) || !match(C, m_APInt(CVal)))
206 return false;
207
208 bool Overflow = false;
209 if (Opcode == Instruction::Add)
210 (void)BVal->sadd_ov(*CVal, Overflow);
211 else
212 (void)BVal->ssub_ov(*CVal, Overflow);
213
214 return !Overflow;
215 }
216
217 /// Conservatively clears subclassOptionalData after a reassociation or
218 /// commutation. We preserve fast-math flags when applicable as they can be
219 /// preserved.
ClearSubclassDataAfterReassociation(BinaryOperator & I)220 static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
221 FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
222 if (!FPMO) {
223 I.clearSubclassOptionalData();
224 return;
225 }
226
227 FastMathFlags FMF = I.getFastMathFlags();
228 I.clearSubclassOptionalData();
229 I.setFastMathFlags(FMF);
230 }
231
232 /// Combine constant operands of associative operations either before or after a
233 /// cast to eliminate one of the associative operations:
234 /// (op (cast (op X, C2)), C1) --> (cast (op X, op (C1, C2)))
235 /// (op (cast (op X, C2)), C1) --> (op (cast X), op (C1, C2))
simplifyAssocCastAssoc(BinaryOperator * BinOp1)236 static bool simplifyAssocCastAssoc(BinaryOperator *BinOp1) {
237 auto *Cast = dyn_cast<CastInst>(BinOp1->getOperand(0));
238 if (!Cast || !Cast->hasOneUse())
239 return false;
240
241 // TODO: Enhance logic for other casts and remove this check.
242 auto CastOpcode = Cast->getOpcode();
243 if (CastOpcode != Instruction::ZExt)
244 return false;
245
246 // TODO: Enhance logic for other BinOps and remove this check.
247 if (!BinOp1->isBitwiseLogicOp())
248 return false;
249
250 auto AssocOpcode = BinOp1->getOpcode();
251 auto *BinOp2 = dyn_cast<BinaryOperator>(Cast->getOperand(0));
252 if (!BinOp2 || !BinOp2->hasOneUse() || BinOp2->getOpcode() != AssocOpcode)
253 return false;
254
255 Constant *C1, *C2;
256 if (!match(BinOp1->getOperand(1), m_Constant(C1)) ||
257 !match(BinOp2->getOperand(1), m_Constant(C2)))
258 return false;
259
260 // TODO: This assumes a zext cast.
261 // Eg, if it was a trunc, we'd cast C1 to the source type because casting C2
262 // to the destination type might lose bits.
263
264 // Fold the constants together in the destination type:
265 // (op (cast (op X, C2)), C1) --> (op (cast X), FoldedC)
266 Type *DestTy = C1->getType();
267 Constant *CastC2 = ConstantExpr::getCast(CastOpcode, C2, DestTy);
268 Constant *FoldedC = ConstantExpr::get(AssocOpcode, C1, CastC2);
269 Cast->setOperand(0, BinOp2->getOperand(0));
270 BinOp1->setOperand(1, FoldedC);
271 return true;
272 }
273
274 /// This performs a few simplifications for operators that are associative or
275 /// commutative:
276 ///
277 /// Commutative operators:
278 ///
279 /// 1. Order operands such that they are listed from right (least complex) to
280 /// left (most complex). This puts constants before unary operators before
281 /// binary operators.
282 ///
283 /// Associative operators:
284 ///
285 /// 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
286 /// 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
287 ///
288 /// Associative and commutative operators:
289 ///
290 /// 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
291 /// 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
292 /// 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
293 /// if C1 and C2 are constants.
SimplifyAssociativeOrCommutative(BinaryOperator & I)294 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
295 Instruction::BinaryOps Opcode = I.getOpcode();
296 bool Changed = false;
297
298 do {
299 // Order operands such that they are listed from right (least complex) to
300 // left (most complex). This puts constants before unary operators before
301 // binary operators.
302 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
303 getComplexity(I.getOperand(1)))
304 Changed = !I.swapOperands();
305
306 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
307 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
308
309 if (I.isAssociative()) {
310 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
311 if (Op0 && Op0->getOpcode() == Opcode) {
312 Value *A = Op0->getOperand(0);
313 Value *B = Op0->getOperand(1);
314 Value *C = I.getOperand(1);
315
316 // Does "B op C" simplify?
317 if (Value *V = SimplifyBinOp(Opcode, B, C, SQ.getWithInstruction(&I))) {
318 // It simplifies to V. Form "A op V".
319 I.setOperand(0, A);
320 I.setOperand(1, V);
321 // Conservatively clear the optional flags, since they may not be
322 // preserved by the reassociation.
323 if (MaintainNoSignedWrap(I, B, C) &&
324 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
325 // Note: this is only valid because SimplifyBinOp doesn't look at
326 // the operands to Op0.
327 I.clearSubclassOptionalData();
328 I.setHasNoSignedWrap(true);
329 } else {
330 ClearSubclassDataAfterReassociation(I);
331 }
332
333 Changed = true;
334 ++NumReassoc;
335 continue;
336 }
337 }
338
339 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
340 if (Op1 && Op1->getOpcode() == Opcode) {
341 Value *A = I.getOperand(0);
342 Value *B = Op1->getOperand(0);
343 Value *C = Op1->getOperand(1);
344
345 // Does "A op B" simplify?
346 if (Value *V = SimplifyBinOp(Opcode, A, B, SQ.getWithInstruction(&I))) {
347 // It simplifies to V. Form "V op C".
348 I.setOperand(0, V);
349 I.setOperand(1, C);
350 // Conservatively clear the optional flags, since they may not be
351 // preserved by the reassociation.
352 ClearSubclassDataAfterReassociation(I);
353 Changed = true;
354 ++NumReassoc;
355 continue;
356 }
357 }
358 }
359
360 if (I.isAssociative() && I.isCommutative()) {
361 if (simplifyAssocCastAssoc(&I)) {
362 Changed = true;
363 ++NumReassoc;
364 continue;
365 }
366
367 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
368 if (Op0 && Op0->getOpcode() == Opcode) {
369 Value *A = Op0->getOperand(0);
370 Value *B = Op0->getOperand(1);
371 Value *C = I.getOperand(1);
372
373 // Does "C op A" simplify?
374 if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
375 // It simplifies to V. Form "V op B".
376 I.setOperand(0, V);
377 I.setOperand(1, B);
378 // Conservatively clear the optional flags, since they may not be
379 // preserved by the reassociation.
380 ClearSubclassDataAfterReassociation(I);
381 Changed = true;
382 ++NumReassoc;
383 continue;
384 }
385 }
386
387 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
388 if (Op1 && Op1->getOpcode() == Opcode) {
389 Value *A = I.getOperand(0);
390 Value *B = Op1->getOperand(0);
391 Value *C = Op1->getOperand(1);
392
393 // Does "C op A" simplify?
394 if (Value *V = SimplifyBinOp(Opcode, C, A, SQ.getWithInstruction(&I))) {
395 // It simplifies to V. Form "B op V".
396 I.setOperand(0, B);
397 I.setOperand(1, V);
398 // Conservatively clear the optional flags, since they may not be
399 // preserved by the reassociation.
400 ClearSubclassDataAfterReassociation(I);
401 Changed = true;
402 ++NumReassoc;
403 continue;
404 }
405 }
406
407 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
408 // if C1 and C2 are constants.
409 Value *A, *B;
410 Constant *C1, *C2;
411 if (Op0 && Op1 &&
412 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
413 match(Op0, m_OneUse(m_BinOp(m_Value(A), m_Constant(C1)))) &&
414 match(Op1, m_OneUse(m_BinOp(m_Value(B), m_Constant(C2))))) {
415 BinaryOperator *NewBO = BinaryOperator::Create(Opcode, A, B);
416 if (isa<FPMathOperator>(NewBO)) {
417 FastMathFlags Flags = I.getFastMathFlags();
418 Flags &= Op0->getFastMathFlags();
419 Flags &= Op1->getFastMathFlags();
420 NewBO->setFastMathFlags(Flags);
421 }
422 InsertNewInstWith(NewBO, I);
423 NewBO->takeName(Op1);
424 I.setOperand(0, NewBO);
425 I.setOperand(1, ConstantExpr::get(Opcode, C1, C2));
426 // Conservatively clear the optional flags, since they may not be
427 // preserved by the reassociation.
428 ClearSubclassDataAfterReassociation(I);
429
430 Changed = true;
431 continue;
432 }
433 }
434
435 // No further simplifications.
436 return Changed;
437 } while (true);
438 }
439
440 /// Return whether "X LOp (Y ROp Z)" is always equal to
441 /// "(X LOp Y) ROp (X LOp Z)".
leftDistributesOverRight(Instruction::BinaryOps LOp,Instruction::BinaryOps ROp)442 static bool leftDistributesOverRight(Instruction::BinaryOps LOp,
443 Instruction::BinaryOps ROp) {
444 // X & (Y | Z) <--> (X & Y) | (X & Z)
445 // X & (Y ^ Z) <--> (X & Y) ^ (X & Z)
446 if (LOp == Instruction::And)
447 return ROp == Instruction::Or || ROp == Instruction::Xor;
448
449 // X | (Y & Z) <--> (X | Y) & (X | Z)
450 if (LOp == Instruction::Or)
451 return ROp == Instruction::And;
452
453 // X * (Y + Z) <--> (X * Y) + (X * Z)
454 // X * (Y - Z) <--> (X * Y) - (X * Z)
455 if (LOp == Instruction::Mul)
456 return ROp == Instruction::Add || ROp == Instruction::Sub;
457
458 return false;
459 }
460
461 /// Return whether "(X LOp Y) ROp Z" is always equal to
462 /// "(X ROp Z) LOp (Y ROp Z)".
rightDistributesOverLeft(Instruction::BinaryOps LOp,Instruction::BinaryOps ROp)463 static bool rightDistributesOverLeft(Instruction::BinaryOps LOp,
464 Instruction::BinaryOps ROp) {
465 if (Instruction::isCommutative(ROp))
466 return leftDistributesOverRight(ROp, LOp);
467
468 // (X {&|^} Y) >> Z <--> (X >> Z) {&|^} (Y >> Z) for all shifts.
469 return Instruction::isBitwiseLogicOp(LOp) && Instruction::isShift(ROp);
470
471 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
472 // but this requires knowing that the addition does not overflow and other
473 // such subtleties.
474 }
475
476 /// This function returns identity value for given opcode, which can be used to
477 /// factor patterns like (X * 2) + X ==> (X * 2) + (X * 1) ==> X * (2 + 1).
getIdentityValue(Instruction::BinaryOps Opcode,Value * V)478 static Value *getIdentityValue(Instruction::BinaryOps Opcode, Value *V) {
479 if (isa<Constant>(V))
480 return nullptr;
481
482 return ConstantExpr::getBinOpIdentity(Opcode, V->getType());
483 }
484
485 /// This function predicates factorization using distributive laws. By default,
486 /// it just returns the 'Op' inputs. But for special-cases like
487 /// 'add(shl(X, 5), ...)', this function will have TopOpcode == Instruction::Add
488 /// and Op = shl(X, 5). The 'shl' is treated as the more general 'mul X, 32' to
489 /// allow more factorization opportunities.
490 static Instruction::BinaryOps
getBinOpsForFactorization(Instruction::BinaryOps TopOpcode,BinaryOperator * Op,Value * & LHS,Value * & RHS)491 getBinOpsForFactorization(Instruction::BinaryOps TopOpcode, BinaryOperator *Op,
492 Value *&LHS, Value *&RHS) {
493 assert(Op && "Expected a binary operator");
494 LHS = Op->getOperand(0);
495 RHS = Op->getOperand(1);
496 if (TopOpcode == Instruction::Add || TopOpcode == Instruction::Sub) {
497 Constant *C;
498 if (match(Op, m_Shl(m_Value(), m_Constant(C)))) {
499 // X << C --> X * (1 << C)
500 RHS = ConstantExpr::getShl(ConstantInt::get(Op->getType(), 1), C);
501 return Instruction::Mul;
502 }
503 // TODO: We can add other conversions e.g. shr => div etc.
504 }
505 return Op->getOpcode();
506 }
507
508 /// This tries to simplify binary operations by factorizing out common terms
509 /// (e. g. "(A*B)+(A*C)" -> "A*(B+C)").
tryFactorization(BinaryOperator & I,Instruction::BinaryOps InnerOpcode,Value * A,Value * B,Value * C,Value * D)510 Value *InstCombiner::tryFactorization(BinaryOperator &I,
511 Instruction::BinaryOps InnerOpcode,
512 Value *A, Value *B, Value *C, Value *D) {
513 assert(A && B && C && D && "All values must be provided");
514
515 Value *V = nullptr;
516 Value *SimplifiedInst = nullptr;
517 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
518 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
519
520 // Does "X op' Y" always equal "Y op' X"?
521 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
522
523 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
524 if (leftDistributesOverRight(InnerOpcode, TopLevelOpcode))
525 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
526 // commutative case, "(A op' B) op (C op' A)"?
527 if (A == C || (InnerCommutative && A == D)) {
528 if (A != C)
529 std::swap(C, D);
530 // Consider forming "A op' (B op D)".
531 // If "B op D" simplifies then it can be formed with no cost.
532 V = SimplifyBinOp(TopLevelOpcode, B, D, SQ.getWithInstruction(&I));
533 // If "B op D" doesn't simplify then only go on if both of the existing
534 // operations "A op' B" and "C op' D" will be zapped as no longer used.
535 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
536 V = Builder.CreateBinOp(TopLevelOpcode, B, D, RHS->getName());
537 if (V) {
538 SimplifiedInst = Builder.CreateBinOp(InnerOpcode, A, V);
539 }
540 }
541
542 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
543 if (!SimplifiedInst && rightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
544 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
545 // commutative case, "(A op' B) op (B op' D)"?
546 if (B == D || (InnerCommutative && B == C)) {
547 if (B != D)
548 std::swap(C, D);
549 // Consider forming "(A op C) op' B".
550 // If "A op C" simplifies then it can be formed with no cost.
551 V = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
552
553 // If "A op C" doesn't simplify then only go on if both of the existing
554 // operations "A op' B" and "C op' D" will be zapped as no longer used.
555 if (!V && LHS->hasOneUse() && RHS->hasOneUse())
556 V = Builder.CreateBinOp(TopLevelOpcode, A, C, LHS->getName());
557 if (V) {
558 SimplifiedInst = Builder.CreateBinOp(InnerOpcode, V, B);
559 }
560 }
561
562 if (SimplifiedInst) {
563 ++NumFactor;
564 SimplifiedInst->takeName(&I);
565
566 // Check if we can add NSW flag to SimplifiedInst. If so, set NSW flag.
567 // TODO: Check for NUW.
568 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(SimplifiedInst)) {
569 if (isa<OverflowingBinaryOperator>(SimplifiedInst)) {
570 bool HasNSW = false;
571 if (isa<OverflowingBinaryOperator>(&I))
572 HasNSW = I.hasNoSignedWrap();
573
574 if (auto *LOBO = dyn_cast<OverflowingBinaryOperator>(LHS))
575 HasNSW &= LOBO->hasNoSignedWrap();
576
577 if (auto *ROBO = dyn_cast<OverflowingBinaryOperator>(RHS))
578 HasNSW &= ROBO->hasNoSignedWrap();
579
580 // We can propagate 'nsw' if we know that
581 // %Y = mul nsw i16 %X, C
582 // %Z = add nsw i16 %Y, %X
583 // =>
584 // %Z = mul nsw i16 %X, C+1
585 //
586 // iff C+1 isn't INT_MIN
587 const APInt *CInt;
588 if (TopLevelOpcode == Instruction::Add &&
589 InnerOpcode == Instruction::Mul)
590 if (match(V, m_APInt(CInt)) && !CInt->isMinSignedValue())
591 BO->setHasNoSignedWrap(HasNSW);
592 }
593 }
594 }
595 return SimplifiedInst;
596 }
597
598 /// This tries to simplify binary operations which some other binary operation
599 /// distributes over either by factorizing out common terms
600 /// (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this results in
601 /// simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is a win).
602 /// Returns the simplified value, or null if it didn't simplify.
SimplifyUsingDistributiveLaws(BinaryOperator & I)603 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
604 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
605 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
606 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
607 Instruction::BinaryOps TopLevelOpcode = I.getOpcode();
608
609 {
610 // Factorization.
611 Value *A, *B, *C, *D;
612 Instruction::BinaryOps LHSOpcode, RHSOpcode;
613 if (Op0)
614 LHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op0, A, B);
615 if (Op1)
616 RHSOpcode = getBinOpsForFactorization(TopLevelOpcode, Op1, C, D);
617
618 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
619 // a common term.
620 if (Op0 && Op1 && LHSOpcode == RHSOpcode)
621 if (Value *V = tryFactorization(I, LHSOpcode, A, B, C, D))
622 return V;
623
624 // The instruction has the form "(A op' B) op (C)". Try to factorize common
625 // term.
626 if (Op0)
627 if (Value *Ident = getIdentityValue(LHSOpcode, RHS))
628 if (Value *V = tryFactorization(I, LHSOpcode, A, B, RHS, Ident))
629 return V;
630
631 // The instruction has the form "(B) op (C op' D)". Try to factorize common
632 // term.
633 if (Op1)
634 if (Value *Ident = getIdentityValue(RHSOpcode, LHS))
635 if (Value *V = tryFactorization(I, RHSOpcode, LHS, Ident, C, D))
636 return V;
637 }
638
639 // Expansion.
640 if (Op0 && rightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
641 // The instruction has the form "(A op' B) op C". See if expanding it out
642 // to "(A op C) op' (B op C)" results in simplifications.
643 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
644 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
645
646 Value *L = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
647 Value *R = SimplifyBinOp(TopLevelOpcode, B, C, SQ.getWithInstruction(&I));
648
649 // Do "A op C" and "B op C" both simplify?
650 if (L && R) {
651 // They do! Return "L op' R".
652 ++NumExpand;
653 C = Builder.CreateBinOp(InnerOpcode, L, R);
654 C->takeName(&I);
655 return C;
656 }
657
658 // Does "A op C" simplify to the identity value for the inner opcode?
659 if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
660 // They do! Return "B op C".
661 ++NumExpand;
662 C = Builder.CreateBinOp(TopLevelOpcode, B, C);
663 C->takeName(&I);
664 return C;
665 }
666
667 // Does "B op C" simplify to the identity value for the inner opcode?
668 if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
669 // They do! Return "A op C".
670 ++NumExpand;
671 C = Builder.CreateBinOp(TopLevelOpcode, A, C);
672 C->takeName(&I);
673 return C;
674 }
675 }
676
677 if (Op1 && leftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
678 // The instruction has the form "A op (B op' C)". See if expanding it out
679 // to "(A op B) op' (A op C)" results in simplifications.
680 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
681 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
682
683 Value *L = SimplifyBinOp(TopLevelOpcode, A, B, SQ.getWithInstruction(&I));
684 Value *R = SimplifyBinOp(TopLevelOpcode, A, C, SQ.getWithInstruction(&I));
685
686 // Do "A op B" and "A op C" both simplify?
687 if (L && R) {
688 // They do! Return "L op' R".
689 ++NumExpand;
690 A = Builder.CreateBinOp(InnerOpcode, L, R);
691 A->takeName(&I);
692 return A;
693 }
694
695 // Does "A op B" simplify to the identity value for the inner opcode?
696 if (L && L == ConstantExpr::getBinOpIdentity(InnerOpcode, L->getType())) {
697 // They do! Return "A op C".
698 ++NumExpand;
699 A = Builder.CreateBinOp(TopLevelOpcode, A, C);
700 A->takeName(&I);
701 return A;
702 }
703
704 // Does "A op C" simplify to the identity value for the inner opcode?
705 if (R && R == ConstantExpr::getBinOpIdentity(InnerOpcode, R->getType())) {
706 // They do! Return "A op B".
707 ++NumExpand;
708 A = Builder.CreateBinOp(TopLevelOpcode, A, B);
709 A->takeName(&I);
710 return A;
711 }
712 }
713
714 return SimplifySelectsFeedingBinaryOp(I, LHS, RHS);
715 }
716
SimplifySelectsFeedingBinaryOp(BinaryOperator & I,Value * LHS,Value * RHS)717 Value *InstCombiner::SimplifySelectsFeedingBinaryOp(BinaryOperator &I,
718 Value *LHS, Value *RHS) {
719 Instruction::BinaryOps Opcode = I.getOpcode();
720 // (op (select (a, b, c)), (select (a, d, e))) -> (select (a, (op b, d), (op
721 // c, e)))
722 Value *A, *B, *C, *D, *E;
723 Value *SI = nullptr;
724 if (match(LHS, m_Select(m_Value(A), m_Value(B), m_Value(C))) &&
725 match(RHS, m_Select(m_Specific(A), m_Value(D), m_Value(E)))) {
726 bool SelectsHaveOneUse = LHS->hasOneUse() && RHS->hasOneUse();
727 BuilderTy::FastMathFlagGuard Guard(Builder);
728 if (isa<FPMathOperator>(&I))
729 Builder.setFastMathFlags(I.getFastMathFlags());
730
731 Value *V1 = SimplifyBinOp(Opcode, C, E, SQ.getWithInstruction(&I));
732 Value *V2 = SimplifyBinOp(Opcode, B, D, SQ.getWithInstruction(&I));
733 if (V1 && V2)
734 SI = Builder.CreateSelect(A, V2, V1);
735 else if (V2 && SelectsHaveOneUse)
736 SI = Builder.CreateSelect(A, V2, Builder.CreateBinOp(Opcode, C, E));
737 else if (V1 && SelectsHaveOneUse)
738 SI = Builder.CreateSelect(A, Builder.CreateBinOp(Opcode, B, D), V1);
739
740 if (SI)
741 SI->takeName(&I);
742 }
743
744 return SI;
745 }
746
747 /// Given a 'sub' instruction, return the RHS of the instruction if the LHS is a
748 /// constant zero (which is the 'negate' form).
dyn_castNegVal(Value * V) const749 Value *InstCombiner::dyn_castNegVal(Value *V) const {
750 if (BinaryOperator::isNeg(V))
751 return BinaryOperator::getNegArgument(V);
752
753 // Constants can be considered to be negated values if they can be folded.
754 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
755 return ConstantExpr::getNeg(C);
756
757 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
758 if (C->getType()->getElementType()->isIntegerTy())
759 return ConstantExpr::getNeg(C);
760
761 if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
762 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
763 Constant *Elt = CV->getAggregateElement(i);
764 if (!Elt)
765 return nullptr;
766
767 if (isa<UndefValue>(Elt))
768 continue;
769
770 if (!isa<ConstantInt>(Elt))
771 return nullptr;
772 }
773 return ConstantExpr::getNeg(CV);
774 }
775
776 return nullptr;
777 }
778
foldOperationIntoSelectOperand(Instruction & I,Value * SO,InstCombiner::BuilderTy & Builder)779 static Value *foldOperationIntoSelectOperand(Instruction &I, Value *SO,
780 InstCombiner::BuilderTy &Builder) {
781 if (auto *Cast = dyn_cast<CastInst>(&I))
782 return Builder.CreateCast(Cast->getOpcode(), SO, I.getType());
783
784 assert(I.isBinaryOp() && "Unexpected opcode for select folding");
785
786 // Figure out if the constant is the left or the right argument.
787 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
788 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
789
790 if (auto *SOC = dyn_cast<Constant>(SO)) {
791 if (ConstIsRHS)
792 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
793 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
794 }
795
796 Value *Op0 = SO, *Op1 = ConstOperand;
797 if (!ConstIsRHS)
798 std::swap(Op0, Op1);
799
800 auto *BO = cast<BinaryOperator>(&I);
801 Value *RI = Builder.CreateBinOp(BO->getOpcode(), Op0, Op1,
802 SO->getName() + ".op");
803 auto *FPInst = dyn_cast<Instruction>(RI);
804 if (FPInst && isa<FPMathOperator>(FPInst))
805 FPInst->copyFastMathFlags(BO);
806 return RI;
807 }
808
FoldOpIntoSelect(Instruction & Op,SelectInst * SI)809 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
810 // Don't modify shared select instructions.
811 if (!SI->hasOneUse())
812 return nullptr;
813
814 Value *TV = SI->getTrueValue();
815 Value *FV = SI->getFalseValue();
816 if (!(isa<Constant>(TV) || isa<Constant>(FV)))
817 return nullptr;
818
819 // Bool selects with constant operands can be folded to logical ops.
820 if (SI->getType()->isIntOrIntVectorTy(1))
821 return nullptr;
822
823 // If it's a bitcast involving vectors, make sure it has the same number of
824 // elements on both sides.
825 if (auto *BC = dyn_cast<BitCastInst>(&Op)) {
826 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
827 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
828
829 // Verify that either both or neither are vectors.
830 if ((SrcTy == nullptr) != (DestTy == nullptr))
831 return nullptr;
832
833 // If vectors, verify that they have the same number of elements.
834 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
835 return nullptr;
836 }
837
838 // Test if a CmpInst instruction is used exclusively by a select as
839 // part of a minimum or maximum operation. If so, refrain from doing
840 // any other folding. This helps out other analyses which understand
841 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
842 // and CodeGen. And in this case, at least one of the comparison
843 // operands has at least one user besides the compare (the select),
844 // which would often largely negate the benefit of folding anyway.
845 if (auto *CI = dyn_cast<CmpInst>(SI->getCondition())) {
846 if (CI->hasOneUse()) {
847 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
848 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
849 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
850 return nullptr;
851 }
852 }
853
854 Value *NewTV = foldOperationIntoSelectOperand(Op, TV, Builder);
855 Value *NewFV = foldOperationIntoSelectOperand(Op, FV, Builder);
856 return SelectInst::Create(SI->getCondition(), NewTV, NewFV, "", nullptr, SI);
857 }
858
foldOperationIntoPhiValue(BinaryOperator * I,Value * InV,InstCombiner::BuilderTy & Builder)859 static Value *foldOperationIntoPhiValue(BinaryOperator *I, Value *InV,
860 InstCombiner::BuilderTy &Builder) {
861 bool ConstIsRHS = isa<Constant>(I->getOperand(1));
862 Constant *C = cast<Constant>(I->getOperand(ConstIsRHS));
863
864 if (auto *InC = dyn_cast<Constant>(InV)) {
865 if (ConstIsRHS)
866 return ConstantExpr::get(I->getOpcode(), InC, C);
867 return ConstantExpr::get(I->getOpcode(), C, InC);
868 }
869
870 Value *Op0 = InV, *Op1 = C;
871 if (!ConstIsRHS)
872 std::swap(Op0, Op1);
873
874 Value *RI = Builder.CreateBinOp(I->getOpcode(), Op0, Op1, "phitmp");
875 auto *FPInst = dyn_cast<Instruction>(RI);
876 if (FPInst && isa<FPMathOperator>(FPInst))
877 FPInst->copyFastMathFlags(I);
878 return RI;
879 }
880
foldOpIntoPhi(Instruction & I,PHINode * PN)881 Instruction *InstCombiner::foldOpIntoPhi(Instruction &I, PHINode *PN) {
882 unsigned NumPHIValues = PN->getNumIncomingValues();
883 if (NumPHIValues == 0)
884 return nullptr;
885
886 // We normally only transform phis with a single use. However, if a PHI has
887 // multiple uses and they are all the same operation, we can fold *all* of the
888 // uses into the PHI.
889 if (!PN->hasOneUse()) {
890 // Walk the use list for the instruction, comparing them to I.
891 for (User *U : PN->users()) {
892 Instruction *UI = cast<Instruction>(U);
893 if (UI != &I && !I.isIdenticalTo(UI))
894 return nullptr;
895 }
896 // Otherwise, we can replace *all* users with the new PHI we form.
897 }
898
899 // Check to see if all of the operands of the PHI are simple constants
900 // (constantint/constantfp/undef). If there is one non-constant value,
901 // remember the BB it is in. If there is more than one or if *it* is a PHI,
902 // bail out. We don't do arbitrary constant expressions here because moving
903 // their computation can be expensive without a cost model.
904 BasicBlock *NonConstBB = nullptr;
905 for (unsigned i = 0; i != NumPHIValues; ++i) {
906 Value *InVal = PN->getIncomingValue(i);
907 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
908 continue;
909
910 if (isa<PHINode>(InVal)) return nullptr; // Itself a phi.
911 if (NonConstBB) return nullptr; // More than one non-const value.
912
913 NonConstBB = PN->getIncomingBlock(i);
914
915 // If the InVal is an invoke at the end of the pred block, then we can't
916 // insert a computation after it without breaking the edge.
917 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
918 if (II->getParent() == NonConstBB)
919 return nullptr;
920
921 // If the incoming non-constant value is in I's block, we will remove one
922 // instruction, but insert another equivalent one, leading to infinite
923 // instcombine.
924 if (isPotentiallyReachable(I.getParent(), NonConstBB, &DT, LI))
925 return nullptr;
926 }
927
928 // If there is exactly one non-constant value, we can insert a copy of the
929 // operation in that block. However, if this is a critical edge, we would be
930 // inserting the computation on some other paths (e.g. inside a loop). Only
931 // do this if the pred block is unconditionally branching into the phi block.
932 if (NonConstBB != nullptr) {
933 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
934 if (!BI || !BI->isUnconditional()) return nullptr;
935 }
936
937 // Okay, we can do the transformation: create the new PHI node.
938 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
939 InsertNewInstBefore(NewPN, *PN);
940 NewPN->takeName(PN);
941
942 // If we are going to have to insert a new computation, do so right before the
943 // predecessor's terminator.
944 if (NonConstBB)
945 Builder.SetInsertPoint(NonConstBB->getTerminator());
946
947 // Next, add all of the operands to the PHI.
948 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
949 // We only currently try to fold the condition of a select when it is a phi,
950 // not the true/false values.
951 Value *TrueV = SI->getTrueValue();
952 Value *FalseV = SI->getFalseValue();
953 BasicBlock *PhiTransBB = PN->getParent();
954 for (unsigned i = 0; i != NumPHIValues; ++i) {
955 BasicBlock *ThisBB = PN->getIncomingBlock(i);
956 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
957 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
958 Value *InV = nullptr;
959 // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
960 // even if currently isNullValue gives false.
961 Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
962 // For vector constants, we cannot use isNullValue to fold into
963 // FalseVInPred versus TrueVInPred. When we have individual nonzero
964 // elements in the vector, we will incorrectly fold InC to
965 // `TrueVInPred`.
966 if (InC && !isa<ConstantExpr>(InC) && isa<ConstantInt>(InC))
967 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
968 else {
969 // Generate the select in the same block as PN's current incoming block.
970 // Note: ThisBB need not be the NonConstBB because vector constants
971 // which are constants by definition are handled here.
972 // FIXME: This can lead to an increase in IR generation because we might
973 // generate selects for vector constant phi operand, that could not be
974 // folded to TrueVInPred or FalseVInPred as done for ConstantInt. For
975 // non-vector phis, this transformation was always profitable because
976 // the select would be generated exactly once in the NonConstBB.
977 Builder.SetInsertPoint(ThisBB->getTerminator());
978 InV = Builder.CreateSelect(PN->getIncomingValue(i), TrueVInPred,
979 FalseVInPred, "phitmp");
980 }
981 NewPN->addIncoming(InV, ThisBB);
982 }
983 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
984 Constant *C = cast<Constant>(I.getOperand(1));
985 for (unsigned i = 0; i != NumPHIValues; ++i) {
986 Value *InV = nullptr;
987 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
988 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
989 else if (isa<ICmpInst>(CI))
990 InV = Builder.CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
991 C, "phitmp");
992 else
993 InV = Builder.CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
994 C, "phitmp");
995 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
996 }
997 } else if (auto *BO = dyn_cast<BinaryOperator>(&I)) {
998 for (unsigned i = 0; i != NumPHIValues; ++i) {
999 Value *InV = foldOperationIntoPhiValue(BO, PN->getIncomingValue(i),
1000 Builder);
1001 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1002 }
1003 } else {
1004 CastInst *CI = cast<CastInst>(&I);
1005 Type *RetTy = CI->getType();
1006 for (unsigned i = 0; i != NumPHIValues; ++i) {
1007 Value *InV;
1008 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
1009 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1010 else
1011 InV = Builder.CreateCast(CI->getOpcode(), PN->getIncomingValue(i),
1012 I.getType(), "phitmp");
1013 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1014 }
1015 }
1016
1017 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
1018 Instruction *User = cast<Instruction>(*UI++);
1019 if (User == &I) continue;
1020 replaceInstUsesWith(*User, NewPN);
1021 eraseInstFromFunction(*User);
1022 }
1023 return replaceInstUsesWith(I, NewPN);
1024 }
1025
foldBinOpIntoSelectOrPhi(BinaryOperator & I)1026 Instruction *InstCombiner::foldBinOpIntoSelectOrPhi(BinaryOperator &I) {
1027 if (!isa<Constant>(I.getOperand(1)))
1028 return nullptr;
1029
1030 if (auto *Sel = dyn_cast<SelectInst>(I.getOperand(0))) {
1031 if (Instruction *NewSel = FoldOpIntoSelect(I, Sel))
1032 return NewSel;
1033 } else if (auto *PN = dyn_cast<PHINode>(I.getOperand(0))) {
1034 if (Instruction *NewPhi = foldOpIntoPhi(I, PN))
1035 return NewPhi;
1036 }
1037 return nullptr;
1038 }
1039
1040 /// Given a pointer type and a constant offset, determine whether or not there
1041 /// is a sequence of GEP indices into the pointed type that will land us at the
1042 /// specified offset. If so, fill them into NewIndices and return the resultant
1043 /// element type, otherwise return null.
FindElementAtOffset(PointerType * PtrTy,int64_t Offset,SmallVectorImpl<Value * > & NewIndices)1044 Type *InstCombiner::FindElementAtOffset(PointerType *PtrTy, int64_t Offset,
1045 SmallVectorImpl<Value *> &NewIndices) {
1046 Type *Ty = PtrTy->getElementType();
1047 if (!Ty->isSized())
1048 return nullptr;
1049
1050 // Start with the index over the outer type. Note that the type size
1051 // might be zero (even if the offset isn't zero) if the indexed type
1052 // is something like [0 x {int, int}]
1053 Type *IndexTy = DL.getIndexType(PtrTy);
1054 int64_t FirstIdx = 0;
1055 if (int64_t TySize = DL.getTypeAllocSize(Ty)) {
1056 FirstIdx = Offset/TySize;
1057 Offset -= FirstIdx*TySize;
1058
1059 // Handle hosts where % returns negative instead of values [0..TySize).
1060 if (Offset < 0) {
1061 --FirstIdx;
1062 Offset += TySize;
1063 assert(Offset >= 0);
1064 }
1065 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
1066 }
1067
1068 NewIndices.push_back(ConstantInt::get(IndexTy, FirstIdx));
1069
1070 // Index into the types. If we fail, set OrigBase to null.
1071 while (Offset) {
1072 // Indexing into tail padding between struct/array elements.
1073 if (uint64_t(Offset * 8) >= DL.getTypeSizeInBits(Ty))
1074 return nullptr;
1075
1076 if (StructType *STy = dyn_cast<StructType>(Ty)) {
1077 const StructLayout *SL = DL.getStructLayout(STy);
1078 assert(Offset < (int64_t)SL->getSizeInBytes() &&
1079 "Offset must stay within the indexed type");
1080
1081 unsigned Elt = SL->getElementContainingOffset(Offset);
1082 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
1083 Elt));
1084
1085 Offset -= SL->getElementOffset(Elt);
1086 Ty = STy->getElementType(Elt);
1087 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
1088 uint64_t EltSize = DL.getTypeAllocSize(AT->getElementType());
1089 assert(EltSize && "Cannot index into a zero-sized array");
1090 NewIndices.push_back(ConstantInt::get(IndexTy,Offset/EltSize));
1091 Offset %= EltSize;
1092 Ty = AT->getElementType();
1093 } else {
1094 // Otherwise, we can't index into the middle of this atomic type, bail.
1095 return nullptr;
1096 }
1097 }
1098
1099 return Ty;
1100 }
1101
shouldMergeGEPs(GEPOperator & GEP,GEPOperator & Src)1102 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
1103 // If this GEP has only 0 indices, it is the same pointer as
1104 // Src. If Src is not a trivial GEP too, don't combine
1105 // the indices.
1106 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
1107 !Src.hasOneUse())
1108 return false;
1109 return true;
1110 }
1111
1112 /// Return a value X such that Val = X * Scale, or null if none.
1113 /// If the multiplication is known not to overflow, then NoSignedWrap is set.
Descale(Value * Val,APInt Scale,bool & NoSignedWrap)1114 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
1115 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
1116 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
1117 Scale.getBitWidth() && "Scale not compatible with value!");
1118
1119 // If Val is zero or Scale is one then Val = Val * Scale.
1120 if (match(Val, m_Zero()) || Scale == 1) {
1121 NoSignedWrap = true;
1122 return Val;
1123 }
1124
1125 // If Scale is zero then it does not divide Val.
1126 if (Scale.isMinValue())
1127 return nullptr;
1128
1129 // Look through chains of multiplications, searching for a constant that is
1130 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
1131 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
1132 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
1133 // down from Val:
1134 //
1135 // Val = M1 * X || Analysis starts here and works down
1136 // M1 = M2 * Y || Doesn't descend into terms with more
1137 // M2 = Z * 4 \/ than one use
1138 //
1139 // Then to modify a term at the bottom:
1140 //
1141 // Val = M1 * X
1142 // M1 = Z * Y || Replaced M2 with Z
1143 //
1144 // Then to work back up correcting nsw flags.
1145
1146 // Op - the term we are currently analyzing. Starts at Val then drills down.
1147 // Replaced with its descaled value before exiting from the drill down loop.
1148 Value *Op = Val;
1149
1150 // Parent - initially null, but after drilling down notes where Op came from.
1151 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
1152 // 0'th operand of Val.
1153 std::pair<Instruction *, unsigned> Parent;
1154
1155 // Set if the transform requires a descaling at deeper levels that doesn't
1156 // overflow.
1157 bool RequireNoSignedWrap = false;
1158
1159 // Log base 2 of the scale. Negative if not a power of 2.
1160 int32_t logScale = Scale.exactLogBase2();
1161
1162 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
1163 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
1164 // If Op is a constant divisible by Scale then descale to the quotient.
1165 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
1166 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
1167 if (!Remainder.isMinValue())
1168 // Not divisible by Scale.
1169 return nullptr;
1170 // Replace with the quotient in the parent.
1171 Op = ConstantInt::get(CI->getType(), Quotient);
1172 NoSignedWrap = true;
1173 break;
1174 }
1175
1176 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
1177 if (BO->getOpcode() == Instruction::Mul) {
1178 // Multiplication.
1179 NoSignedWrap = BO->hasNoSignedWrap();
1180 if (RequireNoSignedWrap && !NoSignedWrap)
1181 return nullptr;
1182
1183 // There are three cases for multiplication: multiplication by exactly
1184 // the scale, multiplication by a constant different to the scale, and
1185 // multiplication by something else.
1186 Value *LHS = BO->getOperand(0);
1187 Value *RHS = BO->getOperand(1);
1188
1189 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1190 // Multiplication by a constant.
1191 if (CI->getValue() == Scale) {
1192 // Multiplication by exactly the scale, replace the multiplication
1193 // by its left-hand side in the parent.
1194 Op = LHS;
1195 break;
1196 }
1197
1198 // Otherwise drill down into the constant.
1199 if (!Op->hasOneUse())
1200 return nullptr;
1201
1202 Parent = std::make_pair(BO, 1);
1203 continue;
1204 }
1205
1206 // Multiplication by something else. Drill down into the left-hand side
1207 // since that's where the reassociate pass puts the good stuff.
1208 if (!Op->hasOneUse())
1209 return nullptr;
1210
1211 Parent = std::make_pair(BO, 0);
1212 continue;
1213 }
1214
1215 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
1216 isa<ConstantInt>(BO->getOperand(1))) {
1217 // Multiplication by a power of 2.
1218 NoSignedWrap = BO->hasNoSignedWrap();
1219 if (RequireNoSignedWrap && !NoSignedWrap)
1220 return nullptr;
1221
1222 Value *LHS = BO->getOperand(0);
1223 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
1224 getLimitedValue(Scale.getBitWidth());
1225 // Op = LHS << Amt.
1226
1227 if (Amt == logScale) {
1228 // Multiplication by exactly the scale, replace the multiplication
1229 // by its left-hand side in the parent.
1230 Op = LHS;
1231 break;
1232 }
1233 if (Amt < logScale || !Op->hasOneUse())
1234 return nullptr;
1235
1236 // Multiplication by more than the scale. Reduce the multiplying amount
1237 // by the scale in the parent.
1238 Parent = std::make_pair(BO, 1);
1239 Op = ConstantInt::get(BO->getType(), Amt - logScale);
1240 break;
1241 }
1242 }
1243
1244 if (!Op->hasOneUse())
1245 return nullptr;
1246
1247 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
1248 if (Cast->getOpcode() == Instruction::SExt) {
1249 // Op is sign-extended from a smaller type, descale in the smaller type.
1250 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1251 APInt SmallScale = Scale.trunc(SmallSize);
1252 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
1253 // descale Op as (sext Y) * Scale. In order to have
1254 // sext (Y * SmallScale) = (sext Y) * Scale
1255 // some conditions need to hold however: SmallScale must sign-extend to
1256 // Scale and the multiplication Y * SmallScale should not overflow.
1257 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
1258 // SmallScale does not sign-extend to Scale.
1259 return nullptr;
1260 assert(SmallScale.exactLogBase2() == logScale);
1261 // Require that Y * SmallScale must not overflow.
1262 RequireNoSignedWrap = true;
1263
1264 // Drill down through the cast.
1265 Parent = std::make_pair(Cast, 0);
1266 Scale = SmallScale;
1267 continue;
1268 }
1269
1270 if (Cast->getOpcode() == Instruction::Trunc) {
1271 // Op is truncated from a larger type, descale in the larger type.
1272 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
1273 // trunc (Y * sext Scale) = (trunc Y) * Scale
1274 // always holds. However (trunc Y) * Scale may overflow even if
1275 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
1276 // from this point up in the expression (see later).
1277 if (RequireNoSignedWrap)
1278 return nullptr;
1279
1280 // Drill down through the cast.
1281 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
1282 Parent = std::make_pair(Cast, 0);
1283 Scale = Scale.sext(LargeSize);
1284 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
1285 logScale = -1;
1286 assert(Scale.exactLogBase2() == logScale);
1287 continue;
1288 }
1289 }
1290
1291 // Unsupported expression, bail out.
1292 return nullptr;
1293 }
1294
1295 // If Op is zero then Val = Op * Scale.
1296 if (match(Op, m_Zero())) {
1297 NoSignedWrap = true;
1298 return Op;
1299 }
1300
1301 // We know that we can successfully descale, so from here on we can safely
1302 // modify the IR. Op holds the descaled version of the deepest term in the
1303 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1304 // not to overflow.
1305
1306 if (!Parent.first)
1307 // The expression only had one term.
1308 return Op;
1309
1310 // Rewrite the parent using the descaled version of its operand.
1311 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1312 assert(Op != Parent.first->getOperand(Parent.second) &&
1313 "Descaling was a no-op?");
1314 Parent.first->setOperand(Parent.second, Op);
1315 Worklist.Add(Parent.first);
1316
1317 // Now work back up the expression correcting nsw flags. The logic is based
1318 // on the following observation: if X * Y is known not to overflow as a signed
1319 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1320 // then X * Z will not overflow as a signed multiplication either. As we work
1321 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1322 // current level has strictly smaller absolute value than the original.
1323 Instruction *Ancestor = Parent.first;
1324 do {
1325 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1326 // If the multiplication wasn't nsw then we can't say anything about the
1327 // value of the descaled multiplication, and we have to clear nsw flags
1328 // from this point on up.
1329 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1330 NoSignedWrap &= OpNoSignedWrap;
1331 if (NoSignedWrap != OpNoSignedWrap) {
1332 BO->setHasNoSignedWrap(NoSignedWrap);
1333 Worklist.Add(Ancestor);
1334 }
1335 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1336 // The fact that the descaled input to the trunc has smaller absolute
1337 // value than the original input doesn't tell us anything useful about
1338 // the absolute values of the truncations.
1339 NoSignedWrap = false;
1340 }
1341 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1342 "Failed to keep proper track of nsw flags while drilling down?");
1343
1344 if (Ancestor == Val)
1345 // Got to the top, all done!
1346 return Val;
1347
1348 // Move up one level in the expression.
1349 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1350 Ancestor = Ancestor->user_back();
1351 } while (true);
1352 }
1353
foldShuffledBinop(BinaryOperator & Inst)1354 Instruction *InstCombiner::foldShuffledBinop(BinaryOperator &Inst) {
1355 if (!Inst.getType()->isVectorTy()) return nullptr;
1356
1357 // It may not be safe to reorder shuffles and things like div, urem, etc.
1358 // because we may trap when executing those ops on unknown vector elements.
1359 // See PR20059.
1360 if (!isSafeToSpeculativelyExecute(&Inst))
1361 return nullptr;
1362
1363 unsigned VWidth = cast<VectorType>(Inst.getType())->getNumElements();
1364 Value *LHS = Inst.getOperand(0), *RHS = Inst.getOperand(1);
1365 assert(cast<VectorType>(LHS->getType())->getNumElements() == VWidth);
1366 assert(cast<VectorType>(RHS->getType())->getNumElements() == VWidth);
1367
1368 auto createBinOpShuffle = [&](Value *X, Value *Y, Constant *M) {
1369 Value *XY = Builder.CreateBinOp(Inst.getOpcode(), X, Y);
1370 if (auto *BO = dyn_cast<BinaryOperator>(XY))
1371 BO->copyIRFlags(&Inst);
1372 return new ShuffleVectorInst(XY, UndefValue::get(XY->getType()), M);
1373 };
1374
1375 // If both arguments of the binary operation are shuffles that use the same
1376 // mask and shuffle within a single vector, move the shuffle after the binop.
1377 Value *V1, *V2;
1378 Constant *Mask;
1379 if (match(LHS, m_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(Mask))) &&
1380 match(RHS, m_ShuffleVector(m_Value(V2), m_Undef(), m_Specific(Mask))) &&
1381 V1->getType() == V2->getType() &&
1382 (LHS->hasOneUse() || RHS->hasOneUse() || LHS == RHS)) {
1383 // Op(shuffle(V1, Mask), shuffle(V2, Mask)) -> shuffle(Op(V1, V2), Mask)
1384 return createBinOpShuffle(V1, V2, Mask);
1385 }
1386
1387 // If one argument is a shuffle within one vector and the other is a constant,
1388 // try moving the shuffle after the binary operation. This canonicalization
1389 // intends to move shuffles closer to other shuffles and binops closer to
1390 // other binops, so they can be folded. It may also enable demanded elements
1391 // transforms.
1392 Constant *C;
1393 if (match(&Inst, m_c_BinOp(
1394 m_OneUse(m_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(Mask))),
1395 m_Constant(C))) &&
1396 V1->getType() == Inst.getType()) {
1397 // Find constant NewC that has property:
1398 // shuffle(NewC, ShMask) = C
1399 // If such constant does not exist (example: ShMask=<0,0> and C=<1,2>)
1400 // reorder is not possible. A 1-to-1 mapping is not required. Example:
1401 // ShMask = <1,1,2,2> and C = <5,5,6,6> --> NewC = <undef,5,6,undef>
1402 SmallVector<int, 16> ShMask;
1403 ShuffleVectorInst::getShuffleMask(Mask, ShMask);
1404 SmallVector<Constant *, 16>
1405 NewVecC(VWidth, UndefValue::get(C->getType()->getScalarType()));
1406 bool MayChange = true;
1407 for (unsigned I = 0; I < VWidth; ++I) {
1408 if (ShMask[I] >= 0) {
1409 assert(ShMask[I] < (int)VWidth);
1410 Constant *CElt = C->getAggregateElement(I);
1411 Constant *NewCElt = NewVecC[ShMask[I]];
1412 if (!CElt || (!isa<UndefValue>(NewCElt) && NewCElt != CElt)) {
1413 MayChange = false;
1414 break;
1415 }
1416 NewVecC[ShMask[I]] = CElt;
1417 }
1418 }
1419 if (MayChange) {
1420 Constant *NewC = ConstantVector::get(NewVecC);
1421 // It may not be safe to execute a binop on a vector with undef elements
1422 // because the entire instruction can be folded to undef or create poison
1423 // that did not exist in the original code.
1424 bool ConstOp1 = isa<Constant>(Inst.getOperand(1));
1425 if (Inst.isIntDivRem() || (Inst.isShift() && ConstOp1))
1426 NewC = getSafeVectorConstantForBinop(Inst.getOpcode(), NewC, ConstOp1);
1427
1428 // Op(shuffle(V1, Mask), C) -> shuffle(Op(V1, NewC), Mask)
1429 // Op(C, shuffle(V1, Mask)) -> shuffle(Op(NewC, V1), Mask)
1430 Value *NewLHS = isa<Constant>(LHS) ? NewC : V1;
1431 Value *NewRHS = isa<Constant>(LHS) ? V1 : NewC;
1432 return createBinOpShuffle(NewLHS, NewRHS, Mask);
1433 }
1434 }
1435
1436 return nullptr;
1437 }
1438
visitGetElementPtrInst(GetElementPtrInst & GEP)1439 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1440 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1441 Type *GEPType = GEP.getType();
1442 Type *GEPEltType = GEP.getSourceElementType();
1443 if (Value *V = SimplifyGEPInst(GEPEltType, Ops, SQ.getWithInstruction(&GEP)))
1444 return replaceInstUsesWith(GEP, V);
1445
1446 Value *PtrOp = GEP.getOperand(0);
1447
1448 // Eliminate unneeded casts for indices, and replace indices which displace
1449 // by multiples of a zero size type with zero.
1450 bool MadeChange = false;
1451
1452 // Index width may not be the same width as pointer width.
1453 // Data layout chooses the right type based on supported integer types.
1454 Type *NewScalarIndexTy =
1455 DL.getIndexType(GEP.getPointerOperandType()->getScalarType());
1456
1457 gep_type_iterator GTI = gep_type_begin(GEP);
1458 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end(); I != E;
1459 ++I, ++GTI) {
1460 // Skip indices into struct types.
1461 if (GTI.isStruct())
1462 continue;
1463
1464 Type *IndexTy = (*I)->getType();
1465 Type *NewIndexType =
1466 IndexTy->isVectorTy()
1467 ? VectorType::get(NewScalarIndexTy, IndexTy->getVectorNumElements())
1468 : NewScalarIndexTy;
1469
1470 // If the element type has zero size then any index over it is equivalent
1471 // to an index of zero, so replace it with zero if it is not zero already.
1472 Type *EltTy = GTI.getIndexedType();
1473 if (EltTy->isSized() && DL.getTypeAllocSize(EltTy) == 0)
1474 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1475 *I = Constant::getNullValue(NewIndexType);
1476 MadeChange = true;
1477 }
1478
1479 if (IndexTy != NewIndexType) {
1480 // If we are using a wider index than needed for this platform, shrink
1481 // it to what we need. If narrower, sign-extend it to what we need.
1482 // This explicit cast can make subsequent optimizations more obvious.
1483 *I = Builder.CreateIntCast(*I, NewIndexType, true);
1484 MadeChange = true;
1485 }
1486 }
1487 if (MadeChange)
1488 return &GEP;
1489
1490 // Check to see if the inputs to the PHI node are getelementptr instructions.
1491 if (auto *PN = dyn_cast<PHINode>(PtrOp)) {
1492 auto *Op1 = dyn_cast<GetElementPtrInst>(PN->getOperand(0));
1493 if (!Op1)
1494 return nullptr;
1495
1496 // Don't fold a GEP into itself through a PHI node. This can only happen
1497 // through the back-edge of a loop. Folding a GEP into itself means that
1498 // the value of the previous iteration needs to be stored in the meantime,
1499 // thus requiring an additional register variable to be live, but not
1500 // actually achieving anything (the GEP still needs to be executed once per
1501 // loop iteration).
1502 if (Op1 == &GEP)
1503 return nullptr;
1504
1505 int DI = -1;
1506
1507 for (auto I = PN->op_begin()+1, E = PN->op_end(); I !=E; ++I) {
1508 auto *Op2 = dyn_cast<GetElementPtrInst>(*I);
1509 if (!Op2 || Op1->getNumOperands() != Op2->getNumOperands())
1510 return nullptr;
1511
1512 // As for Op1 above, don't try to fold a GEP into itself.
1513 if (Op2 == &GEP)
1514 return nullptr;
1515
1516 // Keep track of the type as we walk the GEP.
1517 Type *CurTy = nullptr;
1518
1519 for (unsigned J = 0, F = Op1->getNumOperands(); J != F; ++J) {
1520 if (Op1->getOperand(J)->getType() != Op2->getOperand(J)->getType())
1521 return nullptr;
1522
1523 if (Op1->getOperand(J) != Op2->getOperand(J)) {
1524 if (DI == -1) {
1525 // We have not seen any differences yet in the GEPs feeding the
1526 // PHI yet, so we record this one if it is allowed to be a
1527 // variable.
1528
1529 // The first two arguments can vary for any GEP, the rest have to be
1530 // static for struct slots
1531 if (J > 1 && CurTy->isStructTy())
1532 return nullptr;
1533
1534 DI = J;
1535 } else {
1536 // The GEP is different by more than one input. While this could be
1537 // extended to support GEPs that vary by more than one variable it
1538 // doesn't make sense since it greatly increases the complexity and
1539 // would result in an R+R+R addressing mode which no backend
1540 // directly supports and would need to be broken into several
1541 // simpler instructions anyway.
1542 return nullptr;
1543 }
1544 }
1545
1546 // Sink down a layer of the type for the next iteration.
1547 if (J > 0) {
1548 if (J == 1) {
1549 CurTy = Op1->getSourceElementType();
1550 } else if (auto *CT = dyn_cast<CompositeType>(CurTy)) {
1551 CurTy = CT->getTypeAtIndex(Op1->getOperand(J));
1552 } else {
1553 CurTy = nullptr;
1554 }
1555 }
1556 }
1557 }
1558
1559 // If not all GEPs are identical we'll have to create a new PHI node.
1560 // Check that the old PHI node has only one use so that it will get
1561 // removed.
1562 if (DI != -1 && !PN->hasOneUse())
1563 return nullptr;
1564
1565 auto *NewGEP = cast<GetElementPtrInst>(Op1->clone());
1566 if (DI == -1) {
1567 // All the GEPs feeding the PHI are identical. Clone one down into our
1568 // BB so that it can be merged with the current GEP.
1569 GEP.getParent()->getInstList().insert(
1570 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1571 } else {
1572 // All the GEPs feeding the PHI differ at a single offset. Clone a GEP
1573 // into the current block so it can be merged, and create a new PHI to
1574 // set that index.
1575 PHINode *NewPN;
1576 {
1577 IRBuilderBase::InsertPointGuard Guard(Builder);
1578 Builder.SetInsertPoint(PN);
1579 NewPN = Builder.CreatePHI(Op1->getOperand(DI)->getType(),
1580 PN->getNumOperands());
1581 }
1582
1583 for (auto &I : PN->operands())
1584 NewPN->addIncoming(cast<GEPOperator>(I)->getOperand(DI),
1585 PN->getIncomingBlock(I));
1586
1587 NewGEP->setOperand(DI, NewPN);
1588 GEP.getParent()->getInstList().insert(
1589 GEP.getParent()->getFirstInsertionPt(), NewGEP);
1590 NewGEP->setOperand(DI, NewPN);
1591 }
1592
1593 GEP.setOperand(0, NewGEP);
1594 PtrOp = NewGEP;
1595 }
1596
1597 // Combine Indices - If the source pointer to this getelementptr instruction
1598 // is a getelementptr instruction, combine the indices of the two
1599 // getelementptr instructions into a single instruction.
1600 if (auto *Src = dyn_cast<GEPOperator>(PtrOp)) {
1601 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1602 return nullptr;
1603
1604 // Try to reassociate loop invariant GEP chains to enable LICM.
1605 if (LI && Src->getNumOperands() == 2 && GEP.getNumOperands() == 2 &&
1606 Src->hasOneUse()) {
1607 if (Loop *L = LI->getLoopFor(GEP.getParent())) {
1608 Value *GO1 = GEP.getOperand(1);
1609 Value *SO1 = Src->getOperand(1);
1610 // Reassociate the two GEPs if SO1 is variant in the loop and GO1 is
1611 // invariant: this breaks the dependence between GEPs and allows LICM
1612 // to hoist the invariant part out of the loop.
1613 if (L->isLoopInvariant(GO1) && !L->isLoopInvariant(SO1)) {
1614 // We have to be careful here.
1615 // We have something like:
1616 // %src = getelementptr <ty>, <ty>* %base, <ty> %idx
1617 // %gep = getelementptr <ty>, <ty>* %src, <ty> %idx2
1618 // If we just swap idx & idx2 then we could inadvertantly
1619 // change %src from a vector to a scalar, or vice versa.
1620 // Cases:
1621 // 1) %base a scalar & idx a scalar & idx2 a vector
1622 // => Swapping idx & idx2 turns %src into a vector type.
1623 // 2) %base a scalar & idx a vector & idx2 a scalar
1624 // => Swapping idx & idx2 turns %src in a scalar type
1625 // 3) %base, %idx, and %idx2 are scalars
1626 // => %src & %gep are scalars
1627 // => swapping idx & idx2 is safe
1628 // 4) %base a vector
1629 // => %src is a vector
1630 // => swapping idx & idx2 is safe.
1631 auto *SO0 = Src->getOperand(0);
1632 auto *SO0Ty = SO0->getType();
1633 if (!isa<VectorType>(GEPType) || // case 3
1634 isa<VectorType>(SO0Ty)) { // case 4
1635 Src->setOperand(1, GO1);
1636 GEP.setOperand(1, SO1);
1637 return &GEP;
1638 } else {
1639 // Case 1 or 2
1640 // -- have to recreate %src & %gep
1641 // put NewSrc at same location as %src
1642 Builder.SetInsertPoint(cast<Instruction>(PtrOp));
1643 auto *NewSrc = cast<GetElementPtrInst>(
1644 Builder.CreateGEP(SO0, GO1, Src->getName()));
1645 NewSrc->setIsInBounds(Src->isInBounds());
1646 auto *NewGEP = GetElementPtrInst::Create(nullptr, NewSrc, {SO1});
1647 NewGEP->setIsInBounds(GEP.isInBounds());
1648 return NewGEP;
1649 }
1650 }
1651 }
1652 }
1653
1654 // Note that if our source is a gep chain itself then we wait for that
1655 // chain to be resolved before we perform this transformation. This
1656 // avoids us creating a TON of code in some cases.
1657 if (auto *SrcGEP = dyn_cast<GEPOperator>(Src->getOperand(0)))
1658 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1659 return nullptr; // Wait until our source is folded to completion.
1660
1661 SmallVector<Value*, 8> Indices;
1662
1663 // Find out whether the last index in the source GEP is a sequential idx.
1664 bool EndsWithSequential = false;
1665 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1666 I != E; ++I)
1667 EndsWithSequential = I.isSequential();
1668
1669 // Can we combine the two pointer arithmetics offsets?
1670 if (EndsWithSequential) {
1671 // Replace: gep (gep %P, long B), long A, ...
1672 // With: T = long A+B; gep %P, T, ...
1673 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1674 Value *GO1 = GEP.getOperand(1);
1675
1676 // If they aren't the same type, then the input hasn't been processed
1677 // by the loop above yet (which canonicalizes sequential index types to
1678 // intptr_t). Just avoid transforming this until the input has been
1679 // normalized.
1680 if (SO1->getType() != GO1->getType())
1681 return nullptr;
1682
1683 Value *Sum =
1684 SimplifyAddInst(GO1, SO1, false, false, SQ.getWithInstruction(&GEP));
1685 // Only do the combine when we are sure the cost after the
1686 // merge is never more than that before the merge.
1687 if (Sum == nullptr)
1688 return nullptr;
1689
1690 // Update the GEP in place if possible.
1691 if (Src->getNumOperands() == 2) {
1692 GEP.setOperand(0, Src->getOperand(0));
1693 GEP.setOperand(1, Sum);
1694 return &GEP;
1695 }
1696 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1697 Indices.push_back(Sum);
1698 Indices.append(GEP.op_begin()+2, GEP.op_end());
1699 } else if (isa<Constant>(*GEP.idx_begin()) &&
1700 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1701 Src->getNumOperands() != 1) {
1702 // Otherwise we can do the fold if the first index of the GEP is a zero
1703 Indices.append(Src->op_begin()+1, Src->op_end());
1704 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1705 }
1706
1707 if (!Indices.empty())
1708 return GEP.isInBounds() && Src->isInBounds()
1709 ? GetElementPtrInst::CreateInBounds(
1710 Src->getSourceElementType(), Src->getOperand(0), Indices,
1711 GEP.getName())
1712 : GetElementPtrInst::Create(Src->getSourceElementType(),
1713 Src->getOperand(0), Indices,
1714 GEP.getName());
1715 }
1716
1717 if (GEP.getNumIndices() == 1) {
1718 unsigned AS = GEP.getPointerAddressSpace();
1719 if (GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
1720 DL.getIndexSizeInBits(AS)) {
1721 uint64_t TyAllocSize = DL.getTypeAllocSize(GEPEltType);
1722
1723 bool Matched = false;
1724 uint64_t C;
1725 Value *V = nullptr;
1726 if (TyAllocSize == 1) {
1727 V = GEP.getOperand(1);
1728 Matched = true;
1729 } else if (match(GEP.getOperand(1),
1730 m_AShr(m_Value(V), m_ConstantInt(C)))) {
1731 if (TyAllocSize == 1ULL << C)
1732 Matched = true;
1733 } else if (match(GEP.getOperand(1),
1734 m_SDiv(m_Value(V), m_ConstantInt(C)))) {
1735 if (TyAllocSize == C)
1736 Matched = true;
1737 }
1738
1739 if (Matched) {
1740 // Canonicalize (gep i8* X, -(ptrtoint Y))
1741 // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
1742 // The GEP pattern is emitted by the SCEV expander for certain kinds of
1743 // pointer arithmetic.
1744 if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
1745 Operator *Index = cast<Operator>(V);
1746 Value *PtrToInt = Builder.CreatePtrToInt(PtrOp, Index->getType());
1747 Value *NewSub = Builder.CreateSub(PtrToInt, Index->getOperand(1));
1748 return CastInst::Create(Instruction::IntToPtr, NewSub, GEPType);
1749 }
1750 // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
1751 // to (bitcast Y)
1752 Value *Y;
1753 if (match(V, m_Sub(m_PtrToInt(m_Value(Y)),
1754 m_PtrToInt(m_Specific(GEP.getOperand(0))))))
1755 return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y, GEPType);
1756 }
1757 }
1758 }
1759
1760 // We do not handle pointer-vector geps here.
1761 if (GEPType->isVectorTy())
1762 return nullptr;
1763
1764 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1765 Value *StrippedPtr = PtrOp->stripPointerCasts();
1766 PointerType *StrippedPtrTy = cast<PointerType>(StrippedPtr->getType());
1767
1768 if (StrippedPtr != PtrOp) {
1769 bool HasZeroPointerIndex = false;
1770 if (auto *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1771 HasZeroPointerIndex = C->isZero();
1772
1773 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1774 // into : GEP [10 x i8]* X, i32 0, ...
1775 //
1776 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1777 // into : GEP i8* X, ...
1778 //
1779 // This occurs when the program declares an array extern like "int X[];"
1780 if (HasZeroPointerIndex) {
1781 if (auto *CATy = dyn_cast<ArrayType>(GEPEltType)) {
1782 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1783 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1784 // -> GEP i8* X, ...
1785 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1786 GetElementPtrInst *Res = GetElementPtrInst::Create(
1787 StrippedPtrTy->getElementType(), StrippedPtr, Idx, GEP.getName());
1788 Res->setIsInBounds(GEP.isInBounds());
1789 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
1790 return Res;
1791 // Insert Res, and create an addrspacecast.
1792 // e.g.,
1793 // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
1794 // ->
1795 // %0 = GEP i8 addrspace(1)* X, ...
1796 // addrspacecast i8 addrspace(1)* %0 to i8*
1797 return new AddrSpaceCastInst(Builder.Insert(Res), GEPType);
1798 }
1799
1800 if (auto *XATy = dyn_cast<ArrayType>(StrippedPtrTy->getElementType())) {
1801 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1802 if (CATy->getElementType() == XATy->getElementType()) {
1803 // -> GEP [10 x i8]* X, i32 0, ...
1804 // At this point, we know that the cast source type is a pointer
1805 // to an array of the same type as the destination pointer
1806 // array. Because the array type is never stepped over (there
1807 // is a leading zero) we can fold the cast into this GEP.
1808 if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
1809 GEP.setOperand(0, StrippedPtr);
1810 GEP.setSourceElementType(XATy);
1811 return &GEP;
1812 }
1813 // Cannot replace the base pointer directly because StrippedPtr's
1814 // address space is different. Instead, create a new GEP followed by
1815 // an addrspacecast.
1816 // e.g.,
1817 // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
1818 // i32 0, ...
1819 // ->
1820 // %0 = GEP [10 x i8] addrspace(1)* X, ...
1821 // addrspacecast i8 addrspace(1)* %0 to i8*
1822 SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
1823 Value *NewGEP = GEP.isInBounds()
1824 ? Builder.CreateInBoundsGEP(
1825 nullptr, StrippedPtr, Idx, GEP.getName())
1826 : Builder.CreateGEP(nullptr, StrippedPtr, Idx,
1827 GEP.getName());
1828 return new AddrSpaceCastInst(NewGEP, GEPType);
1829 }
1830 }
1831 }
1832 } else if (GEP.getNumOperands() == 2) {
1833 // Transform things like:
1834 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1835 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1836 Type *SrcEltTy = StrippedPtrTy->getElementType();
1837 if (SrcEltTy->isArrayTy() &&
1838 DL.getTypeAllocSize(SrcEltTy->getArrayElementType()) ==
1839 DL.getTypeAllocSize(GEPEltType)) {
1840 Type *IdxType = DL.getIndexType(GEPType);
1841 Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
1842 Value *NewGEP =
1843 GEP.isInBounds()
1844 ? Builder.CreateInBoundsGEP(nullptr, StrippedPtr, Idx,
1845 GEP.getName())
1846 : Builder.CreateGEP(nullptr, StrippedPtr, Idx, GEP.getName());
1847
1848 // V and GEP are both pointer types --> BitCast
1849 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP, GEPType);
1850 }
1851
1852 // Transform things like:
1853 // %V = mul i64 %N, 4
1854 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1855 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1856 if (GEPEltType->isSized() && SrcEltTy->isSized()) {
1857 // Check that changing the type amounts to dividing the index by a scale
1858 // factor.
1859 uint64_t ResSize = DL.getTypeAllocSize(GEPEltType);
1860 uint64_t SrcSize = DL.getTypeAllocSize(SrcEltTy);
1861 if (ResSize && SrcSize % ResSize == 0) {
1862 Value *Idx = GEP.getOperand(1);
1863 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1864 uint64_t Scale = SrcSize / ResSize;
1865
1866 // Earlier transforms ensure that the index has the right type
1867 // according to Data Layout, which considerably simplifies the
1868 // logic by eliminating implicit casts.
1869 assert(Idx->getType() == DL.getIndexType(GEPType) &&
1870 "Index type does not match the Data Layout preferences");
1871
1872 bool NSW;
1873 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1874 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1875 // If the multiplication NewIdx * Scale may overflow then the new
1876 // GEP may not be "inbounds".
1877 Value *NewGEP =
1878 GEP.isInBounds() && NSW
1879 ? Builder.CreateInBoundsGEP(nullptr, StrippedPtr, NewIdx,
1880 GEP.getName())
1881 : Builder.CreateGEP(nullptr, StrippedPtr, NewIdx,
1882 GEP.getName());
1883
1884 // The NewGEP must be pointer typed, so must the old one -> BitCast
1885 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1886 GEPType);
1887 }
1888 }
1889 }
1890
1891 // Similarly, transform things like:
1892 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1893 // (where tmp = 8*tmp2) into:
1894 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1895 if (GEPEltType->isSized() && SrcEltTy->isSized() &&
1896 SrcEltTy->isArrayTy()) {
1897 // Check that changing to the array element type amounts to dividing the
1898 // index by a scale factor.
1899 uint64_t ResSize = DL.getTypeAllocSize(GEPEltType);
1900 uint64_t ArrayEltSize =
1901 DL.getTypeAllocSize(SrcEltTy->getArrayElementType());
1902 if (ResSize && ArrayEltSize % ResSize == 0) {
1903 Value *Idx = GEP.getOperand(1);
1904 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1905 uint64_t Scale = ArrayEltSize / ResSize;
1906
1907 // Earlier transforms ensure that the index has the right type
1908 // according to the Data Layout, which considerably simplifies
1909 // the logic by eliminating implicit casts.
1910 assert(Idx->getType() == DL.getIndexType(GEPType) &&
1911 "Index type does not match the Data Layout preferences");
1912
1913 bool NSW;
1914 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1915 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1916 // If the multiplication NewIdx * Scale may overflow then the new
1917 // GEP may not be "inbounds".
1918 Type *IndTy = DL.getIndexType(GEPType);
1919 Value *Off[2] = {Constant::getNullValue(IndTy), NewIdx};
1920
1921 Value *NewGEP = GEP.isInBounds() && NSW
1922 ? Builder.CreateInBoundsGEP(
1923 SrcEltTy, StrippedPtr, Off, GEP.getName())
1924 : Builder.CreateGEP(SrcEltTy, StrippedPtr, Off,
1925 GEP.getName());
1926 // The NewGEP must be pointer typed, so must the old one -> BitCast
1927 return CastInst::CreatePointerBitCastOrAddrSpaceCast(NewGEP,
1928 GEPType);
1929 }
1930 }
1931 }
1932 }
1933 }
1934
1935 // addrspacecast between types is canonicalized as a bitcast, then an
1936 // addrspacecast. To take advantage of the below bitcast + struct GEP, look
1937 // through the addrspacecast.
1938 Value *ASCStrippedPtrOp = PtrOp;
1939 if (auto *ASC = dyn_cast<AddrSpaceCastInst>(PtrOp)) {
1940 // X = bitcast A addrspace(1)* to B addrspace(1)*
1941 // Y = addrspacecast A addrspace(1)* to B addrspace(2)*
1942 // Z = gep Y, <...constant indices...>
1943 // Into an addrspacecasted GEP of the struct.
1944 if (auto *BC = dyn_cast<BitCastInst>(ASC->getOperand(0)))
1945 ASCStrippedPtrOp = BC;
1946 }
1947
1948 if (auto *BCI = dyn_cast<BitCastInst>(ASCStrippedPtrOp)) {
1949 Value *SrcOp = BCI->getOperand(0);
1950 PointerType *SrcType = cast<PointerType>(BCI->getSrcTy());
1951 Type *SrcEltType = SrcType->getElementType();
1952
1953 // GEP directly using the source operand if this GEP is accessing an element
1954 // of a bitcasted pointer to vector or array of the same dimensions:
1955 // gep (bitcast <c x ty>* X to [c x ty]*), Y, Z --> gep X, Y, Z
1956 // gep (bitcast [c x ty]* X to <c x ty>*), Y, Z --> gep X, Y, Z
1957 auto areMatchingArrayAndVecTypes = [](Type *ArrTy, Type *VecTy) {
1958 return ArrTy->getArrayElementType() == VecTy->getVectorElementType() &&
1959 ArrTy->getArrayNumElements() == VecTy->getVectorNumElements();
1960 };
1961 if (GEP.getNumOperands() == 3 &&
1962 ((GEPEltType->isArrayTy() && SrcEltType->isVectorTy() &&
1963 areMatchingArrayAndVecTypes(GEPEltType, SrcEltType)) ||
1964 (GEPEltType->isVectorTy() && SrcEltType->isArrayTy() &&
1965 areMatchingArrayAndVecTypes(SrcEltType, GEPEltType)))) {
1966 GEP.setOperand(0, SrcOp);
1967 GEP.setSourceElementType(SrcEltType);
1968 return &GEP;
1969 }
1970
1971 // See if we can simplify:
1972 // X = bitcast A* to B*
1973 // Y = gep X, <...constant indices...>
1974 // into a gep of the original struct. This is important for SROA and alias
1975 // analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1976 unsigned OffsetBits = DL.getIndexTypeSizeInBits(GEPType);
1977 APInt Offset(OffsetBits, 0);
1978 if (!isa<BitCastInst>(SrcOp) && GEP.accumulateConstantOffset(DL, Offset)) {
1979 // If this GEP instruction doesn't move the pointer, just replace the GEP
1980 // with a bitcast of the real input to the dest type.
1981 if (!Offset) {
1982 // If the bitcast is of an allocation, and the allocation will be
1983 // converted to match the type of the cast, don't touch this.
1984 if (isa<AllocaInst>(SrcOp) || isAllocationFn(SrcOp, &TLI)) {
1985 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1986 if (Instruction *I = visitBitCast(*BCI)) {
1987 if (I != BCI) {
1988 I->takeName(BCI);
1989 BCI->getParent()->getInstList().insert(BCI->getIterator(), I);
1990 replaceInstUsesWith(*BCI, I);
1991 }
1992 return &GEP;
1993 }
1994 }
1995
1996 if (SrcType->getPointerAddressSpace() != GEP.getAddressSpace())
1997 return new AddrSpaceCastInst(SrcOp, GEPType);
1998 return new BitCastInst(SrcOp, GEPType);
1999 }
2000
2001 // Otherwise, if the offset is non-zero, we need to find out if there is a
2002 // field at Offset in 'A's type. If so, we can pull the cast through the
2003 // GEP.
2004 SmallVector<Value*, 8> NewIndices;
2005 if (FindElementAtOffset(SrcType, Offset.getSExtValue(), NewIndices)) {
2006 Value *NGEP =
2007 GEP.isInBounds()
2008 ? Builder.CreateInBoundsGEP(nullptr, SrcOp, NewIndices)
2009 : Builder.CreateGEP(nullptr, SrcOp, NewIndices);
2010
2011 if (NGEP->getType() == GEPType)
2012 return replaceInstUsesWith(GEP, NGEP);
2013 NGEP->takeName(&GEP);
2014
2015 if (NGEP->getType()->getPointerAddressSpace() != GEP.getAddressSpace())
2016 return new AddrSpaceCastInst(NGEP, GEPType);
2017 return new BitCastInst(NGEP, GEPType);
2018 }
2019 }
2020 }
2021
2022 if (!GEP.isInBounds()) {
2023 unsigned IdxWidth =
2024 DL.getIndexSizeInBits(PtrOp->getType()->getPointerAddressSpace());
2025 APInt BasePtrOffset(IdxWidth, 0);
2026 Value *UnderlyingPtrOp =
2027 PtrOp->stripAndAccumulateInBoundsConstantOffsets(DL,
2028 BasePtrOffset);
2029 if (auto *AI = dyn_cast<AllocaInst>(UnderlyingPtrOp)) {
2030 if (GEP.accumulateConstantOffset(DL, BasePtrOffset) &&
2031 BasePtrOffset.isNonNegative()) {
2032 APInt AllocSize(IdxWidth, DL.getTypeAllocSize(AI->getAllocatedType()));
2033 if (BasePtrOffset.ule(AllocSize)) {
2034 return GetElementPtrInst::CreateInBounds(
2035 PtrOp, makeArrayRef(Ops).slice(1), GEP.getName());
2036 }
2037 }
2038 }
2039 }
2040
2041 return nullptr;
2042 }
2043
isNeverEqualToUnescapedAlloc(Value * V,const TargetLibraryInfo * TLI,Instruction * AI)2044 static bool isNeverEqualToUnescapedAlloc(Value *V, const TargetLibraryInfo *TLI,
2045 Instruction *AI) {
2046 if (isa<ConstantPointerNull>(V))
2047 return true;
2048 if (auto *LI = dyn_cast<LoadInst>(V))
2049 return isa<GlobalVariable>(LI->getPointerOperand());
2050 // Two distinct allocations will never be equal.
2051 // We rely on LookThroughBitCast in isAllocLikeFn being false, since looking
2052 // through bitcasts of V can cause
2053 // the result statement below to be true, even when AI and V (ex:
2054 // i8* ->i32* ->i8* of AI) are the same allocations.
2055 return isAllocLikeFn(V, TLI) && V != AI;
2056 }
2057
isAllocSiteRemovable(Instruction * AI,SmallVectorImpl<WeakTrackingVH> & Users,const TargetLibraryInfo * TLI)2058 static bool isAllocSiteRemovable(Instruction *AI,
2059 SmallVectorImpl<WeakTrackingVH> &Users,
2060 const TargetLibraryInfo *TLI) {
2061 SmallVector<Instruction*, 4> Worklist;
2062 Worklist.push_back(AI);
2063
2064 do {
2065 Instruction *PI = Worklist.pop_back_val();
2066 for (User *U : PI->users()) {
2067 Instruction *I = cast<Instruction>(U);
2068 switch (I->getOpcode()) {
2069 default:
2070 // Give up the moment we see something we can't handle.
2071 return false;
2072
2073 case Instruction::AddrSpaceCast:
2074 case Instruction::BitCast:
2075 case Instruction::GetElementPtr:
2076 Users.emplace_back(I);
2077 Worklist.push_back(I);
2078 continue;
2079
2080 case Instruction::ICmp: {
2081 ICmpInst *ICI = cast<ICmpInst>(I);
2082 // We can fold eq/ne comparisons with null to false/true, respectively.
2083 // We also fold comparisons in some conditions provided the alloc has
2084 // not escaped (see isNeverEqualToUnescapedAlloc).
2085 if (!ICI->isEquality())
2086 return false;
2087 unsigned OtherIndex = (ICI->getOperand(0) == PI) ? 1 : 0;
2088 if (!isNeverEqualToUnescapedAlloc(ICI->getOperand(OtherIndex), TLI, AI))
2089 return false;
2090 Users.emplace_back(I);
2091 continue;
2092 }
2093
2094 case Instruction::Call:
2095 // Ignore no-op and store intrinsics.
2096 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2097 switch (II->getIntrinsicID()) {
2098 default:
2099 return false;
2100
2101 case Intrinsic::memmove:
2102 case Intrinsic::memcpy:
2103 case Intrinsic::memset: {
2104 MemIntrinsic *MI = cast<MemIntrinsic>(II);
2105 if (MI->isVolatile() || MI->getRawDest() != PI)
2106 return false;
2107 LLVM_FALLTHROUGH;
2108 }
2109 case Intrinsic::invariant_start:
2110 case Intrinsic::invariant_end:
2111 case Intrinsic::lifetime_start:
2112 case Intrinsic::lifetime_end:
2113 case Intrinsic::objectsize:
2114 Users.emplace_back(I);
2115 continue;
2116 }
2117 }
2118
2119 if (isFreeCall(I, TLI)) {
2120 Users.emplace_back(I);
2121 continue;
2122 }
2123 return false;
2124
2125 case Instruction::Store: {
2126 StoreInst *SI = cast<StoreInst>(I);
2127 if (SI->isVolatile() || SI->getPointerOperand() != PI)
2128 return false;
2129 Users.emplace_back(I);
2130 continue;
2131 }
2132 }
2133 llvm_unreachable("missing a return?");
2134 }
2135 } while (!Worklist.empty());
2136 return true;
2137 }
2138
visitAllocSite(Instruction & MI)2139 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
2140 // If we have a malloc call which is only used in any amount of comparisons
2141 // to null and free calls, delete the calls and replace the comparisons with
2142 // true or false as appropriate.
2143 SmallVector<WeakTrackingVH, 64> Users;
2144
2145 // If we are removing an alloca with a dbg.declare, insert dbg.value calls
2146 // before each store.
2147 TinyPtrVector<DbgInfoIntrinsic *> DIIs;
2148 std::unique_ptr<DIBuilder> DIB;
2149 if (isa<AllocaInst>(MI)) {
2150 DIIs = FindDbgAddrUses(&MI);
2151 DIB.reset(new DIBuilder(*MI.getModule(), /*AllowUnresolved=*/false));
2152 }
2153
2154 if (isAllocSiteRemovable(&MI, Users, &TLI)) {
2155 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2156 // Lowering all @llvm.objectsize calls first because they may
2157 // use a bitcast/GEP of the alloca we are removing.
2158 if (!Users[i])
2159 continue;
2160
2161 Instruction *I = cast<Instruction>(&*Users[i]);
2162
2163 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
2164 if (II->getIntrinsicID() == Intrinsic::objectsize) {
2165 ConstantInt *Result = lowerObjectSizeCall(II, DL, &TLI,
2166 /*MustSucceed=*/true);
2167 replaceInstUsesWith(*I, Result);
2168 eraseInstFromFunction(*I);
2169 Users[i] = nullptr; // Skip examining in the next loop.
2170 }
2171 }
2172 }
2173 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
2174 if (!Users[i])
2175 continue;
2176
2177 Instruction *I = cast<Instruction>(&*Users[i]);
2178
2179 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
2180 replaceInstUsesWith(*C,
2181 ConstantInt::get(Type::getInt1Ty(C->getContext()),
2182 C->isFalseWhenEqual()));
2183 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I) ||
2184 isa<AddrSpaceCastInst>(I)) {
2185 replaceInstUsesWith(*I, UndefValue::get(I->getType()));
2186 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
2187 for (auto *DII : DIIs)
2188 ConvertDebugDeclareToDebugValue(DII, SI, *DIB);
2189 }
2190 eraseInstFromFunction(*I);
2191 }
2192
2193 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
2194 // Replace invoke with a NOP intrinsic to maintain the original CFG
2195 Module *M = II->getModule();
2196 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
2197 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
2198 None, "", II->getParent());
2199 }
2200
2201 for (auto *DII : DIIs)
2202 eraseInstFromFunction(*DII);
2203
2204 return eraseInstFromFunction(MI);
2205 }
2206 return nullptr;
2207 }
2208
2209 /// Move the call to free before a NULL test.
2210 ///
2211 /// Check if this free is accessed after its argument has been test
2212 /// against NULL (property 0).
2213 /// If yes, it is legal to move this call in its predecessor block.
2214 ///
2215 /// The move is performed only if the block containing the call to free
2216 /// will be removed, i.e.:
2217 /// 1. it has only one predecessor P, and P has two successors
2218 /// 2. it contains the call and an unconditional branch
2219 /// 3. its successor is the same as its predecessor's successor
2220 ///
2221 /// The profitability is out-of concern here and this function should
2222 /// be called only if the caller knows this transformation would be
2223 /// profitable (e.g., for code size).
2224 static Instruction *
tryToMoveFreeBeforeNullTest(CallInst & FI)2225 tryToMoveFreeBeforeNullTest(CallInst &FI) {
2226 Value *Op = FI.getArgOperand(0);
2227 BasicBlock *FreeInstrBB = FI.getParent();
2228 BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
2229
2230 // Validate part of constraint #1: Only one predecessor
2231 // FIXME: We can extend the number of predecessor, but in that case, we
2232 // would duplicate the call to free in each predecessor and it may
2233 // not be profitable even for code size.
2234 if (!PredBB)
2235 return nullptr;
2236
2237 // Validate constraint #2: Does this block contains only the call to
2238 // free and an unconditional branch?
2239 // FIXME: We could check if we can speculate everything in the
2240 // predecessor block
2241 if (FreeInstrBB->size() != 2)
2242 return nullptr;
2243 BasicBlock *SuccBB;
2244 if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
2245 return nullptr;
2246
2247 // Validate the rest of constraint #1 by matching on the pred branch.
2248 TerminatorInst *TI = PredBB->getTerminator();
2249 BasicBlock *TrueBB, *FalseBB;
2250 ICmpInst::Predicate Pred;
2251 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
2252 return nullptr;
2253 if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
2254 return nullptr;
2255
2256 // Validate constraint #3: Ensure the null case just falls through.
2257 if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
2258 return nullptr;
2259 assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
2260 "Broken CFG: missing edge from predecessor to successor");
2261
2262 FI.moveBefore(TI);
2263 return &FI;
2264 }
2265
visitFree(CallInst & FI)2266 Instruction *InstCombiner::visitFree(CallInst &FI) {
2267 Value *Op = FI.getArgOperand(0);
2268
2269 // free undef -> unreachable.
2270 if (isa<UndefValue>(Op)) {
2271 // Insert a new store to null because we cannot modify the CFG here.
2272 Builder.CreateStore(ConstantInt::getTrue(FI.getContext()),
2273 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
2274 return eraseInstFromFunction(FI);
2275 }
2276
2277 // If we have 'free null' delete the instruction. This can happen in stl code
2278 // when lots of inlining happens.
2279 if (isa<ConstantPointerNull>(Op))
2280 return eraseInstFromFunction(FI);
2281
2282 // If we optimize for code size, try to move the call to free before the null
2283 // test so that simplify cfg can remove the empty block and dead code
2284 // elimination the branch. I.e., helps to turn something like:
2285 // if (foo) free(foo);
2286 // into
2287 // free(foo);
2288 if (MinimizeSize)
2289 if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
2290 return I;
2291
2292 return nullptr;
2293 }
2294
visitReturnInst(ReturnInst & RI)2295 Instruction *InstCombiner::visitReturnInst(ReturnInst &RI) {
2296 if (RI.getNumOperands() == 0) // ret void
2297 return nullptr;
2298
2299 Value *ResultOp = RI.getOperand(0);
2300 Type *VTy = ResultOp->getType();
2301 if (!VTy->isIntegerTy())
2302 return nullptr;
2303
2304 // There might be assume intrinsics dominating this return that completely
2305 // determine the value. If so, constant fold it.
2306 KnownBits Known = computeKnownBits(ResultOp, 0, &RI);
2307 if (Known.isConstant())
2308 RI.setOperand(0, Constant::getIntegerValue(VTy, Known.getConstant()));
2309
2310 return nullptr;
2311 }
2312
visitBranchInst(BranchInst & BI)2313 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
2314 // Change br (not X), label True, label False to: br X, label False, True
2315 Value *X = nullptr;
2316 BasicBlock *TrueDest;
2317 BasicBlock *FalseDest;
2318 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
2319 !isa<Constant>(X)) {
2320 // Swap Destinations and condition...
2321 BI.setCondition(X);
2322 BI.swapSuccessors();
2323 return &BI;
2324 }
2325
2326 // If the condition is irrelevant, remove the use so that other
2327 // transforms on the condition become more effective.
2328 if (BI.isConditional() && !isa<ConstantInt>(BI.getCondition()) &&
2329 BI.getSuccessor(0) == BI.getSuccessor(1)) {
2330 BI.setCondition(ConstantInt::getFalse(BI.getCondition()->getType()));
2331 return &BI;
2332 }
2333
2334 // Canonicalize, for example, icmp_ne -> icmp_eq or fcmp_one -> fcmp_oeq.
2335 CmpInst::Predicate Pred;
2336 if (match(&BI, m_Br(m_OneUse(m_Cmp(Pred, m_Value(), m_Value())), TrueDest,
2337 FalseDest)) &&
2338 !isCanonicalPredicate(Pred)) {
2339 // Swap destinations and condition.
2340 CmpInst *Cond = cast<CmpInst>(BI.getCondition());
2341 Cond->setPredicate(CmpInst::getInversePredicate(Pred));
2342 BI.swapSuccessors();
2343 Worklist.Add(Cond);
2344 return &BI;
2345 }
2346
2347 return nullptr;
2348 }
2349
visitSwitchInst(SwitchInst & SI)2350 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
2351 Value *Cond = SI.getCondition();
2352 Value *Op0;
2353 ConstantInt *AddRHS;
2354 if (match(Cond, m_Add(m_Value(Op0), m_ConstantInt(AddRHS)))) {
2355 // Change 'switch (X+4) case 1:' into 'switch (X) case -3'.
2356 for (auto Case : SI.cases()) {
2357 Constant *NewCase = ConstantExpr::getSub(Case.getCaseValue(), AddRHS);
2358 assert(isa<ConstantInt>(NewCase) &&
2359 "Result of expression should be constant");
2360 Case.setValue(cast<ConstantInt>(NewCase));
2361 }
2362 SI.setCondition(Op0);
2363 return &SI;
2364 }
2365
2366 KnownBits Known = computeKnownBits(Cond, 0, &SI);
2367 unsigned LeadingKnownZeros = Known.countMinLeadingZeros();
2368 unsigned LeadingKnownOnes = Known.countMinLeadingOnes();
2369
2370 // Compute the number of leading bits we can ignore.
2371 // TODO: A better way to determine this would use ComputeNumSignBits().
2372 for (auto &C : SI.cases()) {
2373 LeadingKnownZeros = std::min(
2374 LeadingKnownZeros, C.getCaseValue()->getValue().countLeadingZeros());
2375 LeadingKnownOnes = std::min(
2376 LeadingKnownOnes, C.getCaseValue()->getValue().countLeadingOnes());
2377 }
2378
2379 unsigned NewWidth = Known.getBitWidth() - std::max(LeadingKnownZeros, LeadingKnownOnes);
2380
2381 // Shrink the condition operand if the new type is smaller than the old type.
2382 // This may produce a non-standard type for the switch, but that's ok because
2383 // the backend should extend back to a legal type for the target.
2384 if (NewWidth > 0 && NewWidth < Known.getBitWidth()) {
2385 IntegerType *Ty = IntegerType::get(SI.getContext(), NewWidth);
2386 Builder.SetInsertPoint(&SI);
2387 Value *NewCond = Builder.CreateTrunc(Cond, Ty, "trunc");
2388 SI.setCondition(NewCond);
2389
2390 for (auto Case : SI.cases()) {
2391 APInt TruncatedCase = Case.getCaseValue()->getValue().trunc(NewWidth);
2392 Case.setValue(ConstantInt::get(SI.getContext(), TruncatedCase));
2393 }
2394 return &SI;
2395 }
2396
2397 return nullptr;
2398 }
2399
visitExtractValueInst(ExtractValueInst & EV)2400 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
2401 Value *Agg = EV.getAggregateOperand();
2402
2403 if (!EV.hasIndices())
2404 return replaceInstUsesWith(EV, Agg);
2405
2406 if (Value *V = SimplifyExtractValueInst(Agg, EV.getIndices(),
2407 SQ.getWithInstruction(&EV)))
2408 return replaceInstUsesWith(EV, V);
2409
2410 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
2411 // We're extracting from an insertvalue instruction, compare the indices
2412 const unsigned *exti, *exte, *insi, *inse;
2413 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
2414 exte = EV.idx_end(), inse = IV->idx_end();
2415 exti != exte && insi != inse;
2416 ++exti, ++insi) {
2417 if (*insi != *exti)
2418 // The insert and extract both reference distinctly different elements.
2419 // This means the extract is not influenced by the insert, and we can
2420 // replace the aggregate operand of the extract with the aggregate
2421 // operand of the insert. i.e., replace
2422 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2423 // %E = extractvalue { i32, { i32 } } %I, 0
2424 // with
2425 // %E = extractvalue { i32, { i32 } } %A, 0
2426 return ExtractValueInst::Create(IV->getAggregateOperand(),
2427 EV.getIndices());
2428 }
2429 if (exti == exte && insi == inse)
2430 // Both iterators are at the end: Index lists are identical. Replace
2431 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2432 // %C = extractvalue { i32, { i32 } } %B, 1, 0
2433 // with "i32 42"
2434 return replaceInstUsesWith(EV, IV->getInsertedValueOperand());
2435 if (exti == exte) {
2436 // The extract list is a prefix of the insert list. i.e. replace
2437 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
2438 // %E = extractvalue { i32, { i32 } } %I, 1
2439 // with
2440 // %X = extractvalue { i32, { i32 } } %A, 1
2441 // %E = insertvalue { i32 } %X, i32 42, 0
2442 // by switching the order of the insert and extract (though the
2443 // insertvalue should be left in, since it may have other uses).
2444 Value *NewEV = Builder.CreateExtractValue(IV->getAggregateOperand(),
2445 EV.getIndices());
2446 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
2447 makeArrayRef(insi, inse));
2448 }
2449 if (insi == inse)
2450 // The insert list is a prefix of the extract list
2451 // We can simply remove the common indices from the extract and make it
2452 // operate on the inserted value instead of the insertvalue result.
2453 // i.e., replace
2454 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
2455 // %E = extractvalue { i32, { i32 } } %I, 1, 0
2456 // with
2457 // %E extractvalue { i32 } { i32 42 }, 0
2458 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
2459 makeArrayRef(exti, exte));
2460 }
2461 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
2462 // We're extracting from an intrinsic, see if we're the only user, which
2463 // allows us to simplify multiple result intrinsics to simpler things that
2464 // just get one value.
2465 if (II->hasOneUse()) {
2466 // Check if we're grabbing the overflow bit or the result of a 'with
2467 // overflow' intrinsic. If it's the latter we can remove the intrinsic
2468 // and replace it with a traditional binary instruction.
2469 switch (II->getIntrinsicID()) {
2470 case Intrinsic::uadd_with_overflow:
2471 case Intrinsic::sadd_with_overflow:
2472 if (*EV.idx_begin() == 0) { // Normal result.
2473 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2474 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2475 eraseInstFromFunction(*II);
2476 return BinaryOperator::CreateAdd(LHS, RHS);
2477 }
2478
2479 // If the normal result of the add is dead, and the RHS is a constant,
2480 // we can transform this into a range comparison.
2481 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
2482 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
2483 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
2484 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
2485 ConstantExpr::getNot(CI));
2486 break;
2487 case Intrinsic::usub_with_overflow:
2488 case Intrinsic::ssub_with_overflow:
2489 if (*EV.idx_begin() == 0) { // Normal result.
2490 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2491 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2492 eraseInstFromFunction(*II);
2493 return BinaryOperator::CreateSub(LHS, RHS);
2494 }
2495 break;
2496 case Intrinsic::umul_with_overflow:
2497 case Intrinsic::smul_with_overflow:
2498 if (*EV.idx_begin() == 0) { // Normal result.
2499 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
2500 replaceInstUsesWith(*II, UndefValue::get(II->getType()));
2501 eraseInstFromFunction(*II);
2502 return BinaryOperator::CreateMul(LHS, RHS);
2503 }
2504 break;
2505 default:
2506 break;
2507 }
2508 }
2509 }
2510 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
2511 // If the (non-volatile) load only has one use, we can rewrite this to a
2512 // load from a GEP. This reduces the size of the load. If a load is used
2513 // only by extractvalue instructions then this either must have been
2514 // optimized before, or it is a struct with padding, in which case we
2515 // don't want to do the transformation as it loses padding knowledge.
2516 if (L->isSimple() && L->hasOneUse()) {
2517 // extractvalue has integer indices, getelementptr has Value*s. Convert.
2518 SmallVector<Value*, 4> Indices;
2519 // Prefix an i32 0 since we need the first element.
2520 Indices.push_back(Builder.getInt32(0));
2521 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
2522 I != E; ++I)
2523 Indices.push_back(Builder.getInt32(*I));
2524
2525 // We need to insert these at the location of the old load, not at that of
2526 // the extractvalue.
2527 Builder.SetInsertPoint(L);
2528 Value *GEP = Builder.CreateInBoundsGEP(L->getType(),
2529 L->getPointerOperand(), Indices);
2530 Instruction *NL = Builder.CreateLoad(GEP);
2531 // Whatever aliasing information we had for the orignal load must also
2532 // hold for the smaller load, so propagate the annotations.
2533 AAMDNodes Nodes;
2534 L->getAAMetadata(Nodes);
2535 NL->setAAMetadata(Nodes);
2536 // Returning the load directly will cause the main loop to insert it in
2537 // the wrong spot, so use replaceInstUsesWith().
2538 return replaceInstUsesWith(EV, NL);
2539 }
2540 // We could simplify extracts from other values. Note that nested extracts may
2541 // already be simplified implicitly by the above: extract (extract (insert) )
2542 // will be translated into extract ( insert ( extract ) ) first and then just
2543 // the value inserted, if appropriate. Similarly for extracts from single-use
2544 // loads: extract (extract (load)) will be translated to extract (load (gep))
2545 // and if again single-use then via load (gep (gep)) to load (gep).
2546 // However, double extracts from e.g. function arguments or return values
2547 // aren't handled yet.
2548 return nullptr;
2549 }
2550
2551 /// Return 'true' if the given typeinfo will match anything.
isCatchAll(EHPersonality Personality,Constant * TypeInfo)2552 static bool isCatchAll(EHPersonality Personality, Constant *TypeInfo) {
2553 switch (Personality) {
2554 case EHPersonality::GNU_C:
2555 case EHPersonality::GNU_C_SjLj:
2556 case EHPersonality::Rust:
2557 // The GCC C EH and Rust personality only exists to support cleanups, so
2558 // it's not clear what the semantics of catch clauses are.
2559 return false;
2560 case EHPersonality::Unknown:
2561 return false;
2562 case EHPersonality::GNU_Ada:
2563 // While __gnat_all_others_value will match any Ada exception, it doesn't
2564 // match foreign exceptions (or didn't, before gcc-4.7).
2565 return false;
2566 case EHPersonality::GNU_CXX:
2567 case EHPersonality::GNU_CXX_SjLj:
2568 case EHPersonality::GNU_ObjC:
2569 case EHPersonality::MSVC_X86SEH:
2570 case EHPersonality::MSVC_Win64SEH:
2571 case EHPersonality::MSVC_CXX:
2572 case EHPersonality::CoreCLR:
2573 case EHPersonality::Wasm_CXX:
2574 return TypeInfo->isNullValue();
2575 }
2576 llvm_unreachable("invalid enum");
2577 }
2578
shorter_filter(const Value * LHS,const Value * RHS)2579 static bool shorter_filter(const Value *LHS, const Value *RHS) {
2580 return
2581 cast<ArrayType>(LHS->getType())->getNumElements()
2582 <
2583 cast<ArrayType>(RHS->getType())->getNumElements();
2584 }
2585
visitLandingPadInst(LandingPadInst & LI)2586 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
2587 // The logic here should be correct for any real-world personality function.
2588 // However if that turns out not to be true, the offending logic can always
2589 // be conditioned on the personality function, like the catch-all logic is.
2590 EHPersonality Personality =
2591 classifyEHPersonality(LI.getParent()->getParent()->getPersonalityFn());
2592
2593 // Simplify the list of clauses, eg by removing repeated catch clauses
2594 // (these are often created by inlining).
2595 bool MakeNewInstruction = false; // If true, recreate using the following:
2596 SmallVector<Constant *, 16> NewClauses; // - Clauses for the new instruction;
2597 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
2598
2599 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
2600 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
2601 bool isLastClause = i + 1 == e;
2602 if (LI.isCatch(i)) {
2603 // A catch clause.
2604 Constant *CatchClause = LI.getClause(i);
2605 Constant *TypeInfo = CatchClause->stripPointerCasts();
2606
2607 // If we already saw this clause, there is no point in having a second
2608 // copy of it.
2609 if (AlreadyCaught.insert(TypeInfo).second) {
2610 // This catch clause was not already seen.
2611 NewClauses.push_back(CatchClause);
2612 } else {
2613 // Repeated catch clause - drop the redundant copy.
2614 MakeNewInstruction = true;
2615 }
2616
2617 // If this is a catch-all then there is no point in keeping any following
2618 // clauses or marking the landingpad as having a cleanup.
2619 if (isCatchAll(Personality, TypeInfo)) {
2620 if (!isLastClause)
2621 MakeNewInstruction = true;
2622 CleanupFlag = false;
2623 break;
2624 }
2625 } else {
2626 // A filter clause. If any of the filter elements were already caught
2627 // then they can be dropped from the filter. It is tempting to try to
2628 // exploit the filter further by saying that any typeinfo that does not
2629 // occur in the filter can't be caught later (and thus can be dropped).
2630 // However this would be wrong, since typeinfos can match without being
2631 // equal (for example if one represents a C++ class, and the other some
2632 // class derived from it).
2633 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
2634 Constant *FilterClause = LI.getClause(i);
2635 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
2636 unsigned NumTypeInfos = FilterType->getNumElements();
2637
2638 // An empty filter catches everything, so there is no point in keeping any
2639 // following clauses or marking the landingpad as having a cleanup. By
2640 // dealing with this case here the following code is made a bit simpler.
2641 if (!NumTypeInfos) {
2642 NewClauses.push_back(FilterClause);
2643 if (!isLastClause)
2644 MakeNewInstruction = true;
2645 CleanupFlag = false;
2646 break;
2647 }
2648
2649 bool MakeNewFilter = false; // If true, make a new filter.
2650 SmallVector<Constant *, 16> NewFilterElts; // New elements.
2651 if (isa<ConstantAggregateZero>(FilterClause)) {
2652 // Not an empty filter - it contains at least one null typeinfo.
2653 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
2654 Constant *TypeInfo =
2655 Constant::getNullValue(FilterType->getElementType());
2656 // If this typeinfo is a catch-all then the filter can never match.
2657 if (isCatchAll(Personality, TypeInfo)) {
2658 // Throw the filter away.
2659 MakeNewInstruction = true;
2660 continue;
2661 }
2662
2663 // There is no point in having multiple copies of this typeinfo, so
2664 // discard all but the first copy if there is more than one.
2665 NewFilterElts.push_back(TypeInfo);
2666 if (NumTypeInfos > 1)
2667 MakeNewFilter = true;
2668 } else {
2669 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
2670 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
2671 NewFilterElts.reserve(NumTypeInfos);
2672
2673 // Remove any filter elements that were already caught or that already
2674 // occurred in the filter. While there, see if any of the elements are
2675 // catch-alls. If so, the filter can be discarded.
2676 bool SawCatchAll = false;
2677 for (unsigned j = 0; j != NumTypeInfos; ++j) {
2678 Constant *Elt = Filter->getOperand(j);
2679 Constant *TypeInfo = Elt->stripPointerCasts();
2680 if (isCatchAll(Personality, TypeInfo)) {
2681 // This element is a catch-all. Bail out, noting this fact.
2682 SawCatchAll = true;
2683 break;
2684 }
2685
2686 // Even if we've seen a type in a catch clause, we don't want to
2687 // remove it from the filter. An unexpected type handler may be
2688 // set up for a call site which throws an exception of the same
2689 // type caught. In order for the exception thrown by the unexpected
2690 // handler to propagate correctly, the filter must be correctly
2691 // described for the call site.
2692 //
2693 // Example:
2694 //
2695 // void unexpected() { throw 1;}
2696 // void foo() throw (int) {
2697 // std::set_unexpected(unexpected);
2698 // try {
2699 // throw 2.0;
2700 // } catch (int i) {}
2701 // }
2702
2703 // There is no point in having multiple copies of the same typeinfo in
2704 // a filter, so only add it if we didn't already.
2705 if (SeenInFilter.insert(TypeInfo).second)
2706 NewFilterElts.push_back(cast<Constant>(Elt));
2707 }
2708 // A filter containing a catch-all cannot match anything by definition.
2709 if (SawCatchAll) {
2710 // Throw the filter away.
2711 MakeNewInstruction = true;
2712 continue;
2713 }
2714
2715 // If we dropped something from the filter, make a new one.
2716 if (NewFilterElts.size() < NumTypeInfos)
2717 MakeNewFilter = true;
2718 }
2719 if (MakeNewFilter) {
2720 FilterType = ArrayType::get(FilterType->getElementType(),
2721 NewFilterElts.size());
2722 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
2723 MakeNewInstruction = true;
2724 }
2725
2726 NewClauses.push_back(FilterClause);
2727
2728 // If the new filter is empty then it will catch everything so there is
2729 // no point in keeping any following clauses or marking the landingpad
2730 // as having a cleanup. The case of the original filter being empty was
2731 // already handled above.
2732 if (MakeNewFilter && !NewFilterElts.size()) {
2733 assert(MakeNewInstruction && "New filter but not a new instruction!");
2734 CleanupFlag = false;
2735 break;
2736 }
2737 }
2738 }
2739
2740 // If several filters occur in a row then reorder them so that the shortest
2741 // filters come first (those with the smallest number of elements). This is
2742 // advantageous because shorter filters are more likely to match, speeding up
2743 // unwinding, but mostly because it increases the effectiveness of the other
2744 // filter optimizations below.
2745 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
2746 unsigned j;
2747 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
2748 for (j = i; j != e; ++j)
2749 if (!isa<ArrayType>(NewClauses[j]->getType()))
2750 break;
2751
2752 // Check whether the filters are already sorted by length. We need to know
2753 // if sorting them is actually going to do anything so that we only make a
2754 // new landingpad instruction if it does.
2755 for (unsigned k = i; k + 1 < j; ++k)
2756 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
2757 // Not sorted, so sort the filters now. Doing an unstable sort would be
2758 // correct too but reordering filters pointlessly might confuse users.
2759 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
2760 shorter_filter);
2761 MakeNewInstruction = true;
2762 break;
2763 }
2764
2765 // Look for the next batch of filters.
2766 i = j + 1;
2767 }
2768
2769 // If typeinfos matched if and only if equal, then the elements of a filter L
2770 // that occurs later than a filter F could be replaced by the intersection of
2771 // the elements of F and L. In reality two typeinfos can match without being
2772 // equal (for example if one represents a C++ class, and the other some class
2773 // derived from it) so it would be wrong to perform this transform in general.
2774 // However the transform is correct and useful if F is a subset of L. In that
2775 // case L can be replaced by F, and thus removed altogether since repeating a
2776 // filter is pointless. So here we look at all pairs of filters F and L where
2777 // L follows F in the list of clauses, and remove L if every element of F is
2778 // an element of L. This can occur when inlining C++ functions with exception
2779 // specifications.
2780 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
2781 // Examine each filter in turn.
2782 Value *Filter = NewClauses[i];
2783 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
2784 if (!FTy)
2785 // Not a filter - skip it.
2786 continue;
2787 unsigned FElts = FTy->getNumElements();
2788 // Examine each filter following this one. Doing this backwards means that
2789 // we don't have to worry about filters disappearing under us when removed.
2790 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
2791 Value *LFilter = NewClauses[j];
2792 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
2793 if (!LTy)
2794 // Not a filter - skip it.
2795 continue;
2796 // If Filter is a subset of LFilter, i.e. every element of Filter is also
2797 // an element of LFilter, then discard LFilter.
2798 SmallVectorImpl<Constant *>::iterator J = NewClauses.begin() + j;
2799 // If Filter is empty then it is a subset of LFilter.
2800 if (!FElts) {
2801 // Discard LFilter.
2802 NewClauses.erase(J);
2803 MakeNewInstruction = true;
2804 // Move on to the next filter.
2805 continue;
2806 }
2807 unsigned LElts = LTy->getNumElements();
2808 // If Filter is longer than LFilter then it cannot be a subset of it.
2809 if (FElts > LElts)
2810 // Move on to the next filter.
2811 continue;
2812 // At this point we know that LFilter has at least one element.
2813 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
2814 // Filter is a subset of LFilter iff Filter contains only zeros (as we
2815 // already know that Filter is not longer than LFilter).
2816 if (isa<ConstantAggregateZero>(Filter)) {
2817 assert(FElts <= LElts && "Should have handled this case earlier!");
2818 // Discard LFilter.
2819 NewClauses.erase(J);
2820 MakeNewInstruction = true;
2821 }
2822 // Move on to the next filter.
2823 continue;
2824 }
2825 ConstantArray *LArray = cast<ConstantArray>(LFilter);
2826 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
2827 // Since Filter is non-empty and contains only zeros, it is a subset of
2828 // LFilter iff LFilter contains a zero.
2829 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
2830 for (unsigned l = 0; l != LElts; ++l)
2831 if (LArray->getOperand(l)->isNullValue()) {
2832 // LFilter contains a zero - discard it.
2833 NewClauses.erase(J);
2834 MakeNewInstruction = true;
2835 break;
2836 }
2837 // Move on to the next filter.
2838 continue;
2839 }
2840 // At this point we know that both filters are ConstantArrays. Loop over
2841 // operands to see whether every element of Filter is also an element of
2842 // LFilter. Since filters tend to be short this is probably faster than
2843 // using a method that scales nicely.
2844 ConstantArray *FArray = cast<ConstantArray>(Filter);
2845 bool AllFound = true;
2846 for (unsigned f = 0; f != FElts; ++f) {
2847 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2848 AllFound = false;
2849 for (unsigned l = 0; l != LElts; ++l) {
2850 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2851 if (LTypeInfo == FTypeInfo) {
2852 AllFound = true;
2853 break;
2854 }
2855 }
2856 if (!AllFound)
2857 break;
2858 }
2859 if (AllFound) {
2860 // Discard LFilter.
2861 NewClauses.erase(J);
2862 MakeNewInstruction = true;
2863 }
2864 // Move on to the next filter.
2865 }
2866 }
2867
2868 // If we changed any of the clauses, replace the old landingpad instruction
2869 // with a new one.
2870 if (MakeNewInstruction) {
2871 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2872 NewClauses.size());
2873 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2874 NLI->addClause(NewClauses[i]);
2875 // A landing pad with no clauses must have the cleanup flag set. It is
2876 // theoretically possible, though highly unlikely, that we eliminated all
2877 // clauses. If so, force the cleanup flag to true.
2878 if (NewClauses.empty())
2879 CleanupFlag = true;
2880 NLI->setCleanup(CleanupFlag);
2881 return NLI;
2882 }
2883
2884 // Even if none of the clauses changed, we may nonetheless have understood
2885 // that the cleanup flag is pointless. Clear it if so.
2886 if (LI.isCleanup() != CleanupFlag) {
2887 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2888 LI.setCleanup(CleanupFlag);
2889 return &LI;
2890 }
2891
2892 return nullptr;
2893 }
2894
2895 /// Try to move the specified instruction from its current block into the
2896 /// beginning of DestBlock, which can only happen if it's safe to move the
2897 /// instruction past all of the instructions between it and the end of its
2898 /// block.
TryToSinkInstruction(Instruction * I,BasicBlock * DestBlock)2899 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2900 assert(I->hasOneUse() && "Invariants didn't hold!");
2901 BasicBlock *SrcBlock = I->getParent();
2902
2903 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2904 if (isa<PHINode>(I) || I->isEHPad() || I->mayHaveSideEffects() ||
2905 isa<TerminatorInst>(I))
2906 return false;
2907
2908 // Do not sink alloca instructions out of the entry block.
2909 if (isa<AllocaInst>(I) && I->getParent() ==
2910 &DestBlock->getParent()->getEntryBlock())
2911 return false;
2912
2913 // Do not sink into catchswitch blocks.
2914 if (isa<CatchSwitchInst>(DestBlock->getTerminator()))
2915 return false;
2916
2917 // Do not sink convergent call instructions.
2918 if (auto *CI = dyn_cast<CallInst>(I)) {
2919 if (CI->isConvergent())
2920 return false;
2921 }
2922 // We can only sink load instructions if there is nothing between the load and
2923 // the end of block that could change the value.
2924 if (I->mayReadFromMemory()) {
2925 for (BasicBlock::iterator Scan = I->getIterator(),
2926 E = I->getParent()->end();
2927 Scan != E; ++Scan)
2928 if (Scan->mayWriteToMemory())
2929 return false;
2930 }
2931 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2932 I->moveBefore(&*InsertPos);
2933 ++NumSunkInst;
2934
2935 // Also sink all related debug uses from the source basic block. Otherwise we
2936 // get debug use before the def.
2937 SmallVector<DbgInfoIntrinsic *, 1> DbgUsers;
2938 findDbgUsers(DbgUsers, I);
2939 for (auto *DII : DbgUsers) {
2940 if (DII->getParent() == SrcBlock) {
2941 DII->moveBefore(&*InsertPos);
2942 LLVM_DEBUG(dbgs() << "SINK: " << *DII << '\n');
2943 }
2944 }
2945 return true;
2946 }
2947
run()2948 bool InstCombiner::run() {
2949 while (!Worklist.isEmpty()) {
2950 Instruction *I = Worklist.RemoveOne();
2951 if (I == nullptr) continue; // skip null values.
2952
2953 // Check to see if we can DCE the instruction.
2954 if (isInstructionTriviallyDead(I, &TLI)) {
2955 LLVM_DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
2956 eraseInstFromFunction(*I);
2957 ++NumDeadInst;
2958 MadeIRChange = true;
2959 continue;
2960 }
2961
2962 if (!DebugCounter::shouldExecute(VisitCounter))
2963 continue;
2964
2965 // Instruction isn't dead, see if we can constant propagate it.
2966 if (!I->use_empty() &&
2967 (I->getNumOperands() == 0 || isa<Constant>(I->getOperand(0)))) {
2968 if (Constant *C = ConstantFoldInstruction(I, DL, &TLI)) {
2969 LLVM_DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I
2970 << '\n');
2971
2972 // Add operands to the worklist.
2973 replaceInstUsesWith(*I, C);
2974 ++NumConstProp;
2975 if (isInstructionTriviallyDead(I, &TLI))
2976 eraseInstFromFunction(*I);
2977 MadeIRChange = true;
2978 continue;
2979 }
2980 }
2981
2982 // In general, it is possible for computeKnownBits to determine all bits in
2983 // a value even when the operands are not all constants.
2984 Type *Ty = I->getType();
2985 if (ExpensiveCombines && !I->use_empty() && Ty->isIntOrIntVectorTy()) {
2986 KnownBits Known = computeKnownBits(I, /*Depth*/0, I);
2987 if (Known.isConstant()) {
2988 Constant *C = ConstantInt::get(Ty, Known.getConstant());
2989 LLVM_DEBUG(dbgs() << "IC: ConstFold (all bits known) to: " << *C
2990 << " from: " << *I << '\n');
2991
2992 // Add operands to the worklist.
2993 replaceInstUsesWith(*I, C);
2994 ++NumConstProp;
2995 if (isInstructionTriviallyDead(I, &TLI))
2996 eraseInstFromFunction(*I);
2997 MadeIRChange = true;
2998 continue;
2999 }
3000 }
3001
3002 // See if we can trivially sink this instruction to a successor basic block.
3003 if (I->hasOneUse()) {
3004 BasicBlock *BB = I->getParent();
3005 Instruction *UserInst = cast<Instruction>(*I->user_begin());
3006 BasicBlock *UserParent;
3007
3008 // Get the block the use occurs in.
3009 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
3010 UserParent = PN->getIncomingBlock(*I->use_begin());
3011 else
3012 UserParent = UserInst->getParent();
3013
3014 if (UserParent != BB) {
3015 bool UserIsSuccessor = false;
3016 // See if the user is one of our successors.
3017 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
3018 if (*SI == UserParent) {
3019 UserIsSuccessor = true;
3020 break;
3021 }
3022
3023 // If the user is one of our immediate successors, and if that successor
3024 // only has us as a predecessors (we'd have to split the critical edge
3025 // otherwise), we can keep going.
3026 if (UserIsSuccessor && UserParent->getUniquePredecessor()) {
3027 // Okay, the CFG is simple enough, try to sink this instruction.
3028 if (TryToSinkInstruction(I, UserParent)) {
3029 LLVM_DEBUG(dbgs() << "IC: Sink: " << *I << '\n');
3030 MadeIRChange = true;
3031 // We'll add uses of the sunk instruction below, but since sinking
3032 // can expose opportunities for it's *operands* add them to the
3033 // worklist
3034 for (Use &U : I->operands())
3035 if (Instruction *OpI = dyn_cast<Instruction>(U.get()))
3036 Worklist.Add(OpI);
3037 }
3038 }
3039 }
3040 }
3041
3042 // Now that we have an instruction, try combining it to simplify it.
3043 Builder.SetInsertPoint(I);
3044 Builder.SetCurrentDebugLocation(I->getDebugLoc());
3045
3046 #ifndef NDEBUG
3047 std::string OrigI;
3048 #endif
3049 LLVM_DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
3050 LLVM_DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
3051
3052 if (Instruction *Result = visit(*I)) {
3053 ++NumCombined;
3054 // Should we replace the old instruction with a new one?
3055 if (Result != I) {
3056 LLVM_DEBUG(dbgs() << "IC: Old = " << *I << '\n'
3057 << " New = " << *Result << '\n');
3058
3059 if (I->getDebugLoc())
3060 Result->setDebugLoc(I->getDebugLoc());
3061 // Everything uses the new instruction now.
3062 I->replaceAllUsesWith(Result);
3063
3064 // Move the name to the new instruction first.
3065 Result->takeName(I);
3066
3067 // Push the new instruction and any users onto the worklist.
3068 Worklist.AddUsersToWorkList(*Result);
3069 Worklist.Add(Result);
3070
3071 // Insert the new instruction into the basic block...
3072 BasicBlock *InstParent = I->getParent();
3073 BasicBlock::iterator InsertPos = I->getIterator();
3074
3075 // If we replace a PHI with something that isn't a PHI, fix up the
3076 // insertion point.
3077 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
3078 InsertPos = InstParent->getFirstInsertionPt();
3079
3080 InstParent->getInstList().insert(InsertPos, Result);
3081
3082 eraseInstFromFunction(*I);
3083 } else {
3084 LLVM_DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
3085 << " New = " << *I << '\n');
3086
3087 // If the instruction was modified, it's possible that it is now dead.
3088 // if so, remove it.
3089 if (isInstructionTriviallyDead(I, &TLI)) {
3090 eraseInstFromFunction(*I);
3091 } else {
3092 Worklist.AddUsersToWorkList(*I);
3093 Worklist.Add(I);
3094 }
3095 }
3096 MadeIRChange = true;
3097 }
3098 }
3099
3100 Worklist.Zap();
3101 return MadeIRChange;
3102 }
3103
3104 /// Walk the function in depth-first order, adding all reachable code to the
3105 /// worklist.
3106 ///
3107 /// This has a couple of tricks to make the code faster and more powerful. In
3108 /// particular, we constant fold and DCE instructions as we go, to avoid adding
3109 /// them to the worklist (this significantly speeds up instcombine on code where
3110 /// many instructions are dead or constant). Additionally, if we find a branch
3111 /// whose condition is a known constant, we only visit the reachable successors.
AddReachableCodeToWorklist(BasicBlock * BB,const DataLayout & DL,SmallPtrSetImpl<BasicBlock * > & Visited,InstCombineWorklist & ICWorklist,const TargetLibraryInfo * TLI)3112 static bool AddReachableCodeToWorklist(BasicBlock *BB, const DataLayout &DL,
3113 SmallPtrSetImpl<BasicBlock *> &Visited,
3114 InstCombineWorklist &ICWorklist,
3115 const TargetLibraryInfo *TLI) {
3116 bool MadeIRChange = false;
3117 SmallVector<BasicBlock*, 256> Worklist;
3118 Worklist.push_back(BB);
3119
3120 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
3121 DenseMap<Constant *, Constant *> FoldedConstants;
3122
3123 do {
3124 BB = Worklist.pop_back_val();
3125
3126 // We have now visited this block! If we've already been here, ignore it.
3127 if (!Visited.insert(BB).second)
3128 continue;
3129
3130 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
3131 Instruction *Inst = &*BBI++;
3132
3133 // DCE instruction if trivially dead.
3134 if (isInstructionTriviallyDead(Inst, TLI)) {
3135 ++NumDeadInst;
3136 LLVM_DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
3137 salvageDebugInfo(*Inst);
3138 Inst->eraseFromParent();
3139 MadeIRChange = true;
3140 continue;
3141 }
3142
3143 // ConstantProp instruction if trivially constant.
3144 if (!Inst->use_empty() &&
3145 (Inst->getNumOperands() == 0 || isa<Constant>(Inst->getOperand(0))))
3146 if (Constant *C = ConstantFoldInstruction(Inst, DL, TLI)) {
3147 LLVM_DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *Inst
3148 << '\n');
3149 Inst->replaceAllUsesWith(C);
3150 ++NumConstProp;
3151 if (isInstructionTriviallyDead(Inst, TLI))
3152 Inst->eraseFromParent();
3153 MadeIRChange = true;
3154 continue;
3155 }
3156
3157 // See if we can constant fold its operands.
3158 for (Use &U : Inst->operands()) {
3159 if (!isa<ConstantVector>(U) && !isa<ConstantExpr>(U))
3160 continue;
3161
3162 auto *C = cast<Constant>(U);
3163 Constant *&FoldRes = FoldedConstants[C];
3164 if (!FoldRes)
3165 FoldRes = ConstantFoldConstant(C, DL, TLI);
3166 if (!FoldRes)
3167 FoldRes = C;
3168
3169 if (FoldRes != C) {
3170 LLVM_DEBUG(dbgs() << "IC: ConstFold operand of: " << *Inst
3171 << "\n Old = " << *C
3172 << "\n New = " << *FoldRes << '\n');
3173 U = FoldRes;
3174 MadeIRChange = true;
3175 }
3176 }
3177
3178 // Skip processing debug intrinsics in InstCombine. Processing these call instructions
3179 // consumes non-trivial amount of time and provides no value for the optimization.
3180 if (!isa<DbgInfoIntrinsic>(Inst))
3181 InstrsForInstCombineWorklist.push_back(Inst);
3182 }
3183
3184 // Recursively visit successors. If this is a branch or switch on a
3185 // constant, only visit the reachable successor.
3186 TerminatorInst *TI = BB->getTerminator();
3187 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3188 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
3189 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
3190 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
3191 Worklist.push_back(ReachableBB);
3192 continue;
3193 }
3194 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3195 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
3196 Worklist.push_back(SI->findCaseValue(Cond)->getCaseSuccessor());
3197 continue;
3198 }
3199 }
3200
3201 for (BasicBlock *SuccBB : TI->successors())
3202 Worklist.push_back(SuccBB);
3203 } while (!Worklist.empty());
3204
3205 // Once we've found all of the instructions to add to instcombine's worklist,
3206 // add them in reverse order. This way instcombine will visit from the top
3207 // of the function down. This jives well with the way that it adds all uses
3208 // of instructions to the worklist after doing a transformation, thus avoiding
3209 // some N^2 behavior in pathological cases.
3210 ICWorklist.AddInitialGroup(InstrsForInstCombineWorklist);
3211
3212 return MadeIRChange;
3213 }
3214
3215 /// Populate the IC worklist from a function, and prune any dead basic
3216 /// blocks discovered in the process.
3217 ///
3218 /// This also does basic constant propagation and other forward fixing to make
3219 /// the combiner itself run much faster.
prepareICWorklistFromFunction(Function & F,const DataLayout & DL,TargetLibraryInfo * TLI,InstCombineWorklist & ICWorklist)3220 static bool prepareICWorklistFromFunction(Function &F, const DataLayout &DL,
3221 TargetLibraryInfo *TLI,
3222 InstCombineWorklist &ICWorklist) {
3223 bool MadeIRChange = false;
3224
3225 // Do a depth-first traversal of the function, populate the worklist with
3226 // the reachable instructions. Ignore blocks that are not reachable. Keep
3227 // track of which blocks we visit.
3228 SmallPtrSet<BasicBlock *, 32> Visited;
3229 MadeIRChange |=
3230 AddReachableCodeToWorklist(&F.front(), DL, Visited, ICWorklist, TLI);
3231
3232 // Do a quick scan over the function. If we find any blocks that are
3233 // unreachable, remove any instructions inside of them. This prevents
3234 // the instcombine code from having to deal with some bad special cases.
3235 for (BasicBlock &BB : F) {
3236 if (Visited.count(&BB))
3237 continue;
3238
3239 unsigned NumDeadInstInBB = removeAllNonTerminatorAndEHPadInstructions(&BB);
3240 MadeIRChange |= NumDeadInstInBB > 0;
3241 NumDeadInst += NumDeadInstInBB;
3242 }
3243
3244 return MadeIRChange;
3245 }
3246
combineInstructionsOverFunction(Function & F,InstCombineWorklist & Worklist,AliasAnalysis * AA,AssumptionCache & AC,TargetLibraryInfo & TLI,DominatorTree & DT,OptimizationRemarkEmitter & ORE,bool ExpensiveCombines=true,LoopInfo * LI=nullptr)3247 static bool combineInstructionsOverFunction(
3248 Function &F, InstCombineWorklist &Worklist, AliasAnalysis *AA,
3249 AssumptionCache &AC, TargetLibraryInfo &TLI, DominatorTree &DT,
3250 OptimizationRemarkEmitter &ORE, bool ExpensiveCombines = true,
3251 LoopInfo *LI = nullptr) {
3252 auto &DL = F.getParent()->getDataLayout();
3253 ExpensiveCombines |= EnableExpensiveCombines;
3254
3255 /// Builder - This is an IRBuilder that automatically inserts new
3256 /// instructions into the worklist when they are created.
3257 IRBuilder<TargetFolder, IRBuilderCallbackInserter> Builder(
3258 F.getContext(), TargetFolder(DL),
3259 IRBuilderCallbackInserter([&Worklist, &AC](Instruction *I) {
3260 Worklist.Add(I);
3261 if (match(I, m_Intrinsic<Intrinsic::assume>()))
3262 AC.registerAssumption(cast<CallInst>(I));
3263 }));
3264
3265 // Lower dbg.declare intrinsics otherwise their value may be clobbered
3266 // by instcombiner.
3267 bool MadeIRChange = false;
3268 if (ShouldLowerDbgDeclare)
3269 MadeIRChange = LowerDbgDeclare(F);
3270
3271 // Iterate while there is work to do.
3272 int Iteration = 0;
3273 while (true) {
3274 ++Iteration;
3275 LLVM_DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
3276 << F.getName() << "\n");
3277
3278 MadeIRChange |= prepareICWorklistFromFunction(F, DL, &TLI, Worklist);
3279
3280 InstCombiner IC(Worklist, Builder, F.optForMinSize(), ExpensiveCombines, AA,
3281 AC, TLI, DT, ORE, DL, LI);
3282 IC.MaxArraySizeForCombine = MaxArraySize;
3283
3284 if (!IC.run())
3285 break;
3286 }
3287
3288 return MadeIRChange || Iteration > 1;
3289 }
3290
run(Function & F,FunctionAnalysisManager & AM)3291 PreservedAnalyses InstCombinePass::run(Function &F,
3292 FunctionAnalysisManager &AM) {
3293 auto &AC = AM.getResult<AssumptionAnalysis>(F);
3294 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
3295 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
3296 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
3297
3298 auto *LI = AM.getCachedResult<LoopAnalysis>(F);
3299
3300 auto *AA = &AM.getResult<AAManager>(F);
3301 if (!combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT, ORE,
3302 ExpensiveCombines, LI))
3303 // No changes, all analyses are preserved.
3304 return PreservedAnalyses::all();
3305
3306 // Mark all the analyses that instcombine updates as preserved.
3307 PreservedAnalyses PA;
3308 PA.preserveSet<CFGAnalyses>();
3309 PA.preserve<AAManager>();
3310 PA.preserve<BasicAA>();
3311 PA.preserve<GlobalsAA>();
3312 return PA;
3313 }
3314
getAnalysisUsage(AnalysisUsage & AU) const3315 void InstructionCombiningPass::getAnalysisUsage(AnalysisUsage &AU) const {
3316 AU.setPreservesCFG();
3317 AU.addRequired<AAResultsWrapperPass>();
3318 AU.addRequired<AssumptionCacheTracker>();
3319 AU.addRequired<TargetLibraryInfoWrapperPass>();
3320 AU.addRequired<DominatorTreeWrapperPass>();
3321 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
3322 AU.addPreserved<DominatorTreeWrapperPass>();
3323 AU.addPreserved<AAResultsWrapperPass>();
3324 AU.addPreserved<BasicAAWrapperPass>();
3325 AU.addPreserved<GlobalsAAWrapperPass>();
3326 }
3327
runOnFunction(Function & F)3328 bool InstructionCombiningPass::runOnFunction(Function &F) {
3329 if (skipFunction(F))
3330 return false;
3331
3332 // Required analyses.
3333 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
3334 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3335 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
3336 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3337 auto &ORE = getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
3338
3339 // Optional analyses.
3340 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
3341 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr;
3342
3343 return combineInstructionsOverFunction(F, Worklist, AA, AC, TLI, DT, ORE,
3344 ExpensiveCombines, LI);
3345 }
3346
3347 char InstructionCombiningPass::ID = 0;
3348
3349 INITIALIZE_PASS_BEGIN(InstructionCombiningPass, "instcombine",
3350 "Combine redundant instructions", false, false)
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)3351 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
3352 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3353 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3354 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
3355 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
3356 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
3357 INITIALIZE_PASS_END(InstructionCombiningPass, "instcombine",
3358 "Combine redundant instructions", false, false)
3359
3360 // Initialization Routines
3361 void llvm::initializeInstCombine(PassRegistry &Registry) {
3362 initializeInstructionCombiningPassPass(Registry);
3363 }
3364
LLVMInitializeInstCombine(LLVMPassRegistryRef R)3365 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
3366 initializeInstructionCombiningPassPass(*unwrap(R));
3367 }
3368
createInstructionCombiningPass(bool ExpensiveCombines)3369 FunctionPass *llvm::createInstructionCombiningPass(bool ExpensiveCombines) {
3370 return new InstructionCombiningPass(ExpensiveCombines);
3371 }
3372
LLVMAddInstructionCombiningPass(LLVMPassManagerRef PM)3373 void LLVMAddInstructionCombiningPass(LLVMPassManagerRef PM) {
3374 unwrap(PM)->add(createInstructionCombiningPass());
3375 }
3376