1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
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
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "InstCombine.h"
39 #include "llvm/IntrinsicInst.h"
40 #include "llvm/LLVMContext.h"
41 #include "llvm/DerivedTypes.h"
42 #include "llvm/GlobalVariable.h"
43 #include "llvm/Operator.h"
44 #include "llvm/Analysis/ConstantFolding.h"
45 #include "llvm/Analysis/InstructionSimplify.h"
46 #include "llvm/Analysis/MemoryBuiltins.h"
47 #include "llvm/Target/TargetData.h"
48 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/GetElementPtrTypeIterator.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Support/PatternMatch.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/Statistic.h"
57 #include "llvm/ADT/STLExtras.h"
61 using namespace llvm::PatternMatch;
63 STATISTIC(NumCombined , "Number of insts combined");
64 STATISTIC(NumConstProp, "Number of constant folds");
65 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
66 STATISTIC(NumSunkInst , "Number of instructions sunk");
69 char InstCombiner::ID = 0;
70 static RegisterPass<InstCombiner>
71 X("instcombine", "Combine redundant instructions");
73 void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
74 AU.addPreservedID(LCSSAID);
79 /// ShouldChangeType - Return true if it is desirable to convert a computation
80 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
81 /// type for example, or from a smaller to a larger illegal type.
82 bool InstCombiner::ShouldChangeType(const Type *From, const Type *To) const {
83 assert(isa<IntegerType>(From) && isa<IntegerType>(To));
85 // If we don't have TD, we don't know if the source/dest are legal.
86 if (!TD) return false;
88 unsigned FromWidth = From->getPrimitiveSizeInBits();
89 unsigned ToWidth = To->getPrimitiveSizeInBits();
90 bool FromLegal = TD->isLegalInteger(FromWidth);
91 bool ToLegal = TD->isLegalInteger(ToWidth);
93 // If this is a legal integer from type, and the result would be an illegal
94 // type, don't do the transformation.
95 if (FromLegal && !ToLegal)
98 // Otherwise, if both are illegal, do not increase the size of the result. We
99 // do allow things like i160 -> i64, but not i64 -> i160.
100 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
106 /// getBitCastOperand - If the specified operand is a CastInst, a constant
107 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
108 /// operand value, otherwise return null.
110 // FIXME: Value::stripPointerCasts
111 static Value *getBitCastOperand(Value *V) {
112 if (Operator *O = dyn_cast<Operator>(V)) {
113 if (O->getOpcode() == Instruction::BitCast)
114 return O->getOperand(0);
115 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
116 if (GEP->hasAllZeroIndices())
117 return GEP->getPointerOperand();
124 // SimplifyCommutative - This performs a few simplifications for commutative
127 // 1. Order operands such that they are listed from right (least complex) to
128 // left (most complex). This puts constants before unary operators before
131 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
132 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
134 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
135 bool Changed = false;
136 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
137 Changed = !I.swapOperands();
139 if (!I.isAssociative()) return Changed;
141 Instruction::BinaryOps Opcode = I.getOpcode();
142 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
143 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
144 if (isa<Constant>(I.getOperand(1))) {
145 Constant *Folded = ConstantExpr::get(I.getOpcode(),
146 cast<Constant>(I.getOperand(1)),
147 cast<Constant>(Op->getOperand(1)));
148 I.setOperand(0, Op->getOperand(0));
149 I.setOperand(1, Folded);
153 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1)))
154 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
155 Op->hasOneUse() && Op1->hasOneUse()) {
156 Constant *C1 = cast<Constant>(Op->getOperand(1));
157 Constant *C2 = cast<Constant>(Op1->getOperand(1));
159 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
160 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
161 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
165 I.setOperand(0, New);
166 I.setOperand(1, Folded);
173 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
174 // if the LHS is a constant zero (which is the 'negate' form).
176 Value *InstCombiner::dyn_castNegVal(Value *V) const {
177 if (BinaryOperator::isNeg(V))
178 return BinaryOperator::getNegArgument(V);
180 // Constants can be considered to be negated values if they can be folded.
181 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
182 return ConstantExpr::getNeg(C);
184 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
185 if (C->getType()->getElementType()->isInteger())
186 return ConstantExpr::getNeg(C);
191 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
192 // instruction if the LHS is a constant negative zero (which is the 'negate'
195 Value *InstCombiner::dyn_castFNegVal(Value *V) const {
196 if (BinaryOperator::isFNeg(V))
197 return BinaryOperator::getFNegArgument(V);
199 // Constants can be considered to be negated values if they can be folded.
200 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
201 return ConstantExpr::getFNeg(C);
203 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
204 if (C->getType()->getElementType()->isFloatingPoint())
205 return ConstantExpr::getFNeg(C);
210 /// isFreeToInvert - Return true if the specified value is free to invert (apply
211 /// ~ to). This happens in cases where the ~ can be eliminated.
212 static inline bool isFreeToInvert(Value *V) {
214 if (BinaryOperator::isNot(V))
217 // Constants can be considered to be not'ed values.
218 if (isa<ConstantInt>(V))
221 // Compares can be inverted if they have a single use.
222 if (CmpInst *CI = dyn_cast<CmpInst>(V))
223 return CI->hasOneUse();
228 static inline Value *dyn_castNotVal(Value *V) {
229 // If this is not(not(x)) don't return that this is a not: we want the two
230 // not's to be folded first.
231 if (BinaryOperator::isNot(V)) {
232 Value *Operand = BinaryOperator::getNotArgument(V);
233 if (!isFreeToInvert(Operand))
237 // Constants can be considered to be not'ed values...
238 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
239 return ConstantInt::get(C->getType(), ~C->getValue());
245 /// AddOne - Add one to a ConstantInt.
246 static Constant *AddOne(Constant *C) {
247 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
249 /// SubOne - Subtract one from a ConstantInt.
250 static Constant *SubOne(ConstantInt *C) {
251 return ConstantInt::get(C->getContext(), C->getValue()-1);
255 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
257 if (CastInst *CI = dyn_cast<CastInst>(&I))
258 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
260 // Figure out if the constant is the left or the right argument.
261 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
262 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
264 if (Constant *SOC = dyn_cast<Constant>(SO)) {
266 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
267 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
270 Value *Op0 = SO, *Op1 = ConstOperand;
274 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
275 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
276 SO->getName()+".op");
277 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
278 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
279 SO->getName()+".cmp");
280 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
281 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
282 SO->getName()+".cmp");
283 llvm_unreachable("Unknown binary instruction type!");
286 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
287 // constant as the other operand, try to fold the binary operator into the
288 // select arguments. This also works for Cast instructions, which obviously do
289 // not have a second operand.
290 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
291 // Don't modify shared select instructions
292 if (!SI->hasOneUse()) return 0;
293 Value *TV = SI->getOperand(1);
294 Value *FV = SI->getOperand(2);
296 if (isa<Constant>(TV) || isa<Constant>(FV)) {
297 // Bool selects with constant operands can be folded to logical ops.
298 if (SI->getType() == Type::getInt1Ty(SI->getContext())) return 0;
300 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
301 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
303 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
310 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
311 /// has a PHI node as operand #0, see if we can fold the instruction into the
312 /// PHI (which is only possible if all operands to the PHI are constants).
314 /// If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
315 /// that would normally be unprofitable because they strongly encourage jump
317 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I,
318 bool AllowAggressive) {
319 AllowAggressive = false;
320 PHINode *PN = cast<PHINode>(I.getOperand(0));
321 unsigned NumPHIValues = PN->getNumIncomingValues();
322 if (NumPHIValues == 0 ||
323 // We normally only transform phis with a single use, unless we're trying
324 // hard to make jump threading happen.
325 (!PN->hasOneUse() && !AllowAggressive))
329 // Check to see if all of the operands of the PHI are simple constants
330 // (constantint/constantfp/undef). If there is one non-constant value,
331 // remember the BB it is in. If there is more than one or if *it* is a PHI,
332 // bail out. We don't do arbitrary constant expressions here because moving
333 // their computation can be expensive without a cost model.
334 BasicBlock *NonConstBB = 0;
335 for (unsigned i = 0; i != NumPHIValues; ++i)
336 if (!isa<Constant>(PN->getIncomingValue(i)) ||
337 isa<ConstantExpr>(PN->getIncomingValue(i))) {
338 if (NonConstBB) return 0; // More than one non-const value.
339 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
340 NonConstBB = PN->getIncomingBlock(i);
342 // If the incoming non-constant value is in I's block, we have an infinite
344 if (NonConstBB == I.getParent())
348 // If there is exactly one non-constant value, we can insert a copy of the
349 // operation in that block. However, if this is a critical edge, we would be
350 // inserting the computation one some other paths (e.g. inside a loop). Only
351 // do this if the pred block is unconditionally branching into the phi block.
352 if (NonConstBB != 0 && !AllowAggressive) {
353 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
354 if (!BI || !BI->isUnconditional()) return 0;
357 // Okay, we can do the transformation: create the new PHI node.
358 PHINode *NewPN = PHINode::Create(I.getType(), "");
359 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
360 InsertNewInstBefore(NewPN, *PN);
363 // Next, add all of the operands to the PHI.
364 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
365 // We only currently try to fold the condition of a select when it is a phi,
366 // not the true/false values.
367 Value *TrueV = SI->getTrueValue();
368 Value *FalseV = SI->getFalseValue();
369 BasicBlock *PhiTransBB = PN->getParent();
370 for (unsigned i = 0; i != NumPHIValues; ++i) {
371 BasicBlock *ThisBB = PN->getIncomingBlock(i);
372 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
373 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
375 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
376 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
378 assert(PN->getIncomingBlock(i) == NonConstBB);
379 InV = SelectInst::Create(PN->getIncomingValue(i), TrueVInPred,
381 "phitmp", NonConstBB->getTerminator());
382 Worklist.Add(cast<Instruction>(InV));
384 NewPN->addIncoming(InV, ThisBB);
386 } else if (I.getNumOperands() == 2) {
387 Constant *C = cast<Constant>(I.getOperand(1));
388 for (unsigned i = 0; i != NumPHIValues; ++i) {
390 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
391 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
392 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
394 InV = ConstantExpr::get(I.getOpcode(), InC, C);
396 assert(PN->getIncomingBlock(i) == NonConstBB);
397 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
398 InV = BinaryOperator::Create(BO->getOpcode(),
399 PN->getIncomingValue(i), C, "phitmp",
400 NonConstBB->getTerminator());
401 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
402 InV = CmpInst::Create(CI->getOpcode(),
404 PN->getIncomingValue(i), C, "phitmp",
405 NonConstBB->getTerminator());
407 llvm_unreachable("Unknown binop!");
409 Worklist.Add(cast<Instruction>(InV));
411 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
414 CastInst *CI = cast<CastInst>(&I);
415 const Type *RetTy = CI->getType();
416 for (unsigned i = 0; i != NumPHIValues; ++i) {
418 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
419 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
421 assert(PN->getIncomingBlock(i) == NonConstBB);
422 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
423 I.getType(), "phitmp",
424 NonConstBB->getTerminator());
425 Worklist.Add(cast<Instruction>(InV));
427 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
430 return ReplaceInstUsesWith(I, NewPN);
434 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
435 /// are carefully arranged to allow folding of expressions such as:
437 /// (A < B) | (A > B) --> (A != B)
439 /// Note that this is only valid if the first and second predicates have the
440 /// same sign. Is illegal to do: (A u< B) | (A s> B)
442 /// Three bits are used to represent the condition, as follows:
447 /// <=> Value Definition
448 /// 000 0 Always false
455 /// 111 7 Always true
457 static unsigned getICmpCode(const ICmpInst *ICI) {
458 switch (ICI->getPredicate()) {
460 case ICmpInst::ICMP_UGT: return 1; // 001
461 case ICmpInst::ICMP_SGT: return 1; // 001
462 case ICmpInst::ICMP_EQ: return 2; // 010
463 case ICmpInst::ICMP_UGE: return 3; // 011
464 case ICmpInst::ICMP_SGE: return 3; // 011
465 case ICmpInst::ICMP_ULT: return 4; // 100
466 case ICmpInst::ICMP_SLT: return 4; // 100
467 case ICmpInst::ICMP_NE: return 5; // 101
468 case ICmpInst::ICMP_ULE: return 6; // 110
469 case ICmpInst::ICMP_SLE: return 6; // 110
472 llvm_unreachable("Invalid ICmp predicate!");
477 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
478 /// predicate into a three bit mask. It also returns whether it is an ordered
479 /// predicate by reference.
480 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
483 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
484 case FCmpInst::FCMP_UNO: return 0; // 000
485 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
486 case FCmpInst::FCMP_UGT: return 1; // 001
487 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
488 case FCmpInst::FCMP_UEQ: return 2; // 010
489 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
490 case FCmpInst::FCMP_UGE: return 3; // 011
491 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
492 case FCmpInst::FCMP_ULT: return 4; // 100
493 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
494 case FCmpInst::FCMP_UNE: return 5; // 101
495 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
496 case FCmpInst::FCMP_ULE: return 6; // 110
499 // Not expecting FCMP_FALSE and FCMP_TRUE;
500 llvm_unreachable("Unexpected FCmp predicate!");
505 /// getICmpValue - This is the complement of getICmpCode, which turns an
506 /// opcode and two operands into either a constant true or false, or a brand
507 /// new ICmp instruction. The sign is passed in to determine which kind
508 /// of predicate to use in the new icmp instruction.
509 static Value *getICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS) {
511 default: assert(0 && "Illegal ICmp code!");
513 return ConstantInt::getFalse(LHS->getContext());
516 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
517 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
519 return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
522 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
523 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
526 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
527 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
529 return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
532 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
533 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
535 return ConstantInt::getTrue(LHS->getContext());
539 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
540 /// opcode and two operands into either a FCmp instruction. isordered is passed
541 /// in to determine which kind of predicate to use in the new fcmp instruction.
542 static Value *getFCmpValue(bool isordered, unsigned code,
543 Value *LHS, Value *RHS) {
545 default: llvm_unreachable("Illegal FCmp code!");
548 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
550 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
553 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
555 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
558 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
560 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
563 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
565 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
568 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
570 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
573 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
575 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
578 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
580 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
581 case 7: return ConstantInt::getTrue(LHS->getContext());
585 /// PredicatesFoldable - Return true if both predicates match sign or if at
586 /// least one of them is an equality comparison (which is signless).
587 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
588 return (CmpInst::isSigned(p1) == CmpInst::isSigned(p2)) ||
589 (CmpInst::isSigned(p1) && ICmpInst::isEquality(p2)) ||
590 (CmpInst::isSigned(p2) && ICmpInst::isEquality(p1));
593 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
594 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
595 // guaranteed to be a binary operator.
596 Instruction *InstCombiner::OptAndOp(Instruction *Op,
599 BinaryOperator &TheAnd) {
600 Value *X = Op->getOperand(0);
601 Constant *Together = 0;
603 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
605 switch (Op->getOpcode()) {
606 case Instruction::Xor:
607 if (Op->hasOneUse()) {
608 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
609 Value *And = Builder->CreateAnd(X, AndRHS);
611 return BinaryOperator::CreateXor(And, Together);
614 case Instruction::Or:
615 if (Together == AndRHS) // (X | C) & C --> C
616 return ReplaceInstUsesWith(TheAnd, AndRHS);
618 if (Op->hasOneUse() && Together != OpRHS) {
619 // (X | C1) & C2 --> (X | (C1&C2)) & C2
620 Value *Or = Builder->CreateOr(X, Together);
622 return BinaryOperator::CreateAnd(Or, AndRHS);
625 case Instruction::Add:
626 if (Op->hasOneUse()) {
627 // Adding a one to a single bit bit-field should be turned into an XOR
628 // of the bit. First thing to check is to see if this AND is with a
629 // single bit constant.
630 const APInt &AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
632 // If there is only one bit set.
633 if (AndRHSV.isPowerOf2()) {
634 // Ok, at this point, we know that we are masking the result of the
635 // ADD down to exactly one bit. If the constant we are adding has
636 // no bits set below this bit, then we can eliminate the ADD.
637 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
639 // Check to see if any bits below the one bit set in AndRHSV are set.
640 if ((AddRHS & (AndRHSV-1)) == 0) {
641 // If not, the only thing that can effect the output of the AND is
642 // the bit specified by AndRHSV. If that bit is set, the effect of
643 // the XOR is to toggle the bit. If it is clear, then the ADD has
645 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
646 TheAnd.setOperand(0, X);
649 // Pull the XOR out of the AND.
650 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
651 NewAnd->takeName(Op);
652 return BinaryOperator::CreateXor(NewAnd, AndRHS);
659 case Instruction::Shl: {
660 // We know that the AND will not produce any of the bits shifted in, so if
661 // the anded constant includes them, clear them now!
663 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
664 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
665 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
666 ConstantInt *CI = ConstantInt::get(AndRHS->getContext(),
667 AndRHS->getValue() & ShlMask);
669 if (CI->getValue() == ShlMask) {
670 // Masking out bits that the shift already masks
671 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
672 } else if (CI != AndRHS) { // Reducing bits set in and.
673 TheAnd.setOperand(1, CI);
678 case Instruction::LShr: {
679 // We know that the AND will not produce any of the bits shifted in, so if
680 // the anded constant includes them, clear them now! This only applies to
681 // unsigned shifts, because a signed shr may bring in set bits!
683 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
684 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
685 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
686 ConstantInt *CI = ConstantInt::get(Op->getContext(),
687 AndRHS->getValue() & ShrMask);
689 if (CI->getValue() == ShrMask) {
690 // Masking out bits that the shift already masks.
691 return ReplaceInstUsesWith(TheAnd, Op);
692 } else if (CI != AndRHS) {
693 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
698 case Instruction::AShr:
700 // See if this is shifting in some sign extension, then masking it out
702 if (Op->hasOneUse()) {
703 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
704 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
705 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
706 Constant *C = ConstantInt::get(Op->getContext(),
707 AndRHS->getValue() & ShrMask);
708 if (C == AndRHS) { // Masking out bits shifted in.
709 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
710 // Make the argument unsigned.
711 Value *ShVal = Op->getOperand(0);
712 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
713 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
722 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
723 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
724 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
725 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
726 /// insert new instructions.
727 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
728 bool isSigned, bool Inside,
730 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
731 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
732 "Lo is not <= Hi in range emission code!");
735 if (Lo == Hi) // Trivially false.
736 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
738 // V >= Min && V < Hi --> V < Hi
739 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
740 ICmpInst::Predicate pred = (isSigned ?
741 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
742 return new ICmpInst(pred, V, Hi);
745 // Emit V-Lo <u Hi-Lo
746 Constant *NegLo = ConstantExpr::getNeg(Lo);
747 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
748 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
749 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
752 if (Lo == Hi) // Trivially true.
753 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
755 // V < Min || V >= Hi -> V > Hi-1
756 Hi = SubOne(cast<ConstantInt>(Hi));
757 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
758 ICmpInst::Predicate pred = (isSigned ?
759 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
760 return new ICmpInst(pred, V, Hi);
763 // Emit V-Lo >u Hi-1-Lo
764 // Note that Hi has already had one subtracted from it, above.
765 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
766 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
767 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
768 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
771 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
772 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
773 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
774 // not, since all 1s are not contiguous.
775 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
776 const APInt& V = Val->getValue();
777 uint32_t BitWidth = Val->getType()->getBitWidth();
778 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
780 // look for the first zero bit after the run of ones
781 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
782 // look for the first non-zero bit
783 ME = V.getActiveBits();
787 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
788 /// where isSub determines whether the operator is a sub. If we can fold one of
789 /// the following xforms:
791 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
792 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
793 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
795 /// return (A +/- B).
797 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
798 ConstantInt *Mask, bool isSub,
800 Instruction *LHSI = dyn_cast<Instruction>(LHS);
801 if (!LHSI || LHSI->getNumOperands() != 2 ||
802 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
804 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
806 switch (LHSI->getOpcode()) {
808 case Instruction::And:
809 if (ConstantExpr::getAnd(N, Mask) == Mask) {
810 // If the AndRHS is a power of two minus one (0+1+), this is simple.
811 if ((Mask->getValue().countLeadingZeros() +
812 Mask->getValue().countPopulation()) ==
813 Mask->getValue().getBitWidth())
816 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
817 // part, we don't need any explicit masks to take them out of A. If that
818 // is all N is, ignore it.
819 uint32_t MB = 0, ME = 0;
820 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
821 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
822 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
823 if (MaskedValueIsZero(RHS, Mask))
828 case Instruction::Or:
829 case Instruction::Xor:
830 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
831 if ((Mask->getValue().countLeadingZeros() +
832 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
833 && ConstantExpr::getAnd(N, Mask)->isNullValue())
839 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
840 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
843 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
844 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
845 ICmpInst *LHS, ICmpInst *RHS) {
846 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
848 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
849 if (PredicatesFoldable(LHSCC, RHSCC)) {
850 if (LHS->getOperand(0) == RHS->getOperand(1) &&
851 LHS->getOperand(1) == RHS->getOperand(0))
853 if (LHS->getOperand(0) == RHS->getOperand(0) &&
854 LHS->getOperand(1) == RHS->getOperand(1)) {
855 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
856 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
857 bool isSigned = LHS->isSigned() || RHS->isSigned();
858 Value *RV = getICmpValue(isSigned, Code, Op0, Op1);
859 if (Instruction *I = dyn_cast<Instruction>(RV))
861 // Otherwise, it's a constant boolean value.
862 return ReplaceInstUsesWith(I, RV);
866 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
867 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
868 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
869 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
870 if (LHSCst == 0 || RHSCst == 0) return 0;
872 if (LHSCst == RHSCst && LHSCC == RHSCC) {
873 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
874 // where C is a power of 2
875 if (LHSCC == ICmpInst::ICMP_ULT &&
876 LHSCst->getValue().isPowerOf2()) {
877 Value *NewOr = Builder->CreateOr(Val, Val2);
878 return new ICmpInst(LHSCC, NewOr, LHSCst);
881 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
882 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
883 Value *NewOr = Builder->CreateOr(Val, Val2);
884 return new ICmpInst(LHSCC, NewOr, LHSCst);
888 // From here on, we only handle:
889 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
890 if (Val != Val2) return 0;
892 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
893 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
894 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
895 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
896 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
899 // We can't fold (ugt x, C) & (sgt x, C2).
900 if (!PredicatesFoldable(LHSCC, RHSCC))
903 // Ensure that the larger constant is on the RHS.
905 if (CmpInst::isSigned(LHSCC) ||
906 (ICmpInst::isEquality(LHSCC) &&
907 CmpInst::isSigned(RHSCC)))
908 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
910 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
914 std::swap(LHSCst, RHSCst);
915 std::swap(LHSCC, RHSCC);
918 // At this point, we know we have have two icmp instructions
919 // comparing a value against two constants and and'ing the result
920 // together. Because of the above check, we know that we only have
921 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
922 // (from the icmp folding check above), that the two constants
923 // are not equal and that the larger constant is on the RHS
924 assert(LHSCst != RHSCst && "Compares not folded above?");
927 default: llvm_unreachable("Unknown integer condition code!");
928 case ICmpInst::ICMP_EQ:
930 default: llvm_unreachable("Unknown integer condition code!");
931 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
932 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
933 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
934 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
935 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
936 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
937 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
938 return ReplaceInstUsesWith(I, LHS);
940 case ICmpInst::ICMP_NE:
942 default: llvm_unreachable("Unknown integer condition code!");
943 case ICmpInst::ICMP_ULT:
944 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
945 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
946 break; // (X != 13 & X u< 15) -> no change
947 case ICmpInst::ICMP_SLT:
948 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
949 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
950 break; // (X != 13 & X s< 15) -> no change
951 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
952 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
953 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
954 return ReplaceInstUsesWith(I, RHS);
955 case ICmpInst::ICMP_NE:
956 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
957 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
958 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
959 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
960 ConstantInt::get(Add->getType(), 1));
962 break; // (X != 13 & X != 15) -> no change
965 case ICmpInst::ICMP_ULT:
967 default: llvm_unreachable("Unknown integer condition code!");
968 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
969 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
970 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
971 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
973 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
974 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
975 return ReplaceInstUsesWith(I, LHS);
976 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
980 case ICmpInst::ICMP_SLT:
982 default: llvm_unreachable("Unknown integer condition code!");
983 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
984 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
985 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
986 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
988 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
989 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
990 return ReplaceInstUsesWith(I, LHS);
991 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
995 case ICmpInst::ICMP_UGT:
997 default: llvm_unreachable("Unknown integer condition code!");
998 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
999 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
1000 return ReplaceInstUsesWith(I, RHS);
1001 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
1003 case ICmpInst::ICMP_NE:
1004 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
1005 return new ICmpInst(LHSCC, Val, RHSCst);
1006 break; // (X u> 13 & X != 15) -> no change
1007 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
1008 return InsertRangeTest(Val, AddOne(LHSCst),
1009 RHSCst, false, true, I);
1010 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
1014 case ICmpInst::ICMP_SGT:
1016 default: llvm_unreachable("Unknown integer condition code!");
1017 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
1018 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
1019 return ReplaceInstUsesWith(I, RHS);
1020 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
1022 case ICmpInst::ICMP_NE:
1023 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
1024 return new ICmpInst(LHSCC, Val, RHSCst);
1025 break; // (X s> 13 & X != 15) -> no change
1026 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
1027 return InsertRangeTest(Val, AddOne(LHSCst),
1028 RHSCst, true, true, I);
1029 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
1038 Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
1041 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
1042 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
1043 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
1044 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1045 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1046 // If either of the constants are nans, then the whole thing returns
1048 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1049 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1050 return new FCmpInst(FCmpInst::FCMP_ORD,
1051 LHS->getOperand(0), RHS->getOperand(0));
1054 // Handle vector zeros. This occurs because the canonical form of
1055 // "fcmp ord x,x" is "fcmp ord x, 0".
1056 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1057 isa<ConstantAggregateZero>(RHS->getOperand(1)))
1058 return new FCmpInst(FCmpInst::FCMP_ORD,
1059 LHS->getOperand(0), RHS->getOperand(0));
1063 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1064 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1065 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1068 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1069 // Swap RHS operands to match LHS.
1070 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1071 std::swap(Op1LHS, Op1RHS);
1074 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
1075 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
1077 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
1079 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
1080 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1081 if (Op0CC == FCmpInst::FCMP_TRUE)
1082 return ReplaceInstUsesWith(I, RHS);
1083 if (Op1CC == FCmpInst::FCMP_TRUE)
1084 return ReplaceInstUsesWith(I, LHS);
1088 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
1089 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
1091 std::swap(LHS, RHS);
1092 std::swap(Op0Pred, Op1Pred);
1093 std::swap(Op0Ordered, Op1Ordered);
1096 // uno && ueq -> uno && (uno || eq) -> ueq
1097 // ord && olt -> ord && (ord && lt) -> olt
1098 if (Op0Ordered == Op1Ordered)
1099 return ReplaceInstUsesWith(I, RHS);
1101 // uno && oeq -> uno && (ord && eq) -> false
1102 // uno && ord -> false
1104 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1105 // ord && ueq -> ord && (uno || eq) -> oeq
1106 return cast<Instruction>(getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS));
1114 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1115 bool Changed = SimplifyCommutative(I);
1116 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1118 if (Value *V = SimplifyAndInst(Op0, Op1, TD))
1119 return ReplaceInstUsesWith(I, V);
1121 // See if we can simplify any instructions used by the instruction whose sole
1122 // purpose is to compute bits we don't care about.
1123 if (SimplifyDemandedInstructionBits(I))
1126 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1127 const APInt &AndRHSMask = AndRHS->getValue();
1128 APInt NotAndRHS(~AndRHSMask);
1130 // Optimize a variety of ((val OP C1) & C2) combinations...
1131 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1132 Value *Op0LHS = Op0I->getOperand(0);
1133 Value *Op0RHS = Op0I->getOperand(1);
1134 switch (Op0I->getOpcode()) {
1136 case Instruction::Xor:
1137 case Instruction::Or:
1138 // If the mask is only needed on one incoming arm, push it up.
1139 if (!Op0I->hasOneUse()) break;
1141 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
1142 // Not masking anything out for the LHS, move to RHS.
1143 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
1144 Op0RHS->getName()+".masked");
1145 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
1147 if (!isa<Constant>(Op0RHS) &&
1148 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
1149 // Not masking anything out for the RHS, move to LHS.
1150 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
1151 Op0LHS->getName()+".masked");
1152 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
1156 case Instruction::Add:
1157 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
1158 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1159 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1160 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
1161 return BinaryOperator::CreateAnd(V, AndRHS);
1162 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
1163 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
1166 case Instruction::Sub:
1167 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
1168 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1169 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1170 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
1171 return BinaryOperator::CreateAnd(V, AndRHS);
1173 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
1174 // has 1's for all bits that the subtraction with A might affect.
1175 if (Op0I->hasOneUse()) {
1176 uint32_t BitWidth = AndRHSMask.getBitWidth();
1177 uint32_t Zeros = AndRHSMask.countLeadingZeros();
1178 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
1180 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
1181 if (!(A && A->isZero()) && // avoid infinite recursion.
1182 MaskedValueIsZero(Op0LHS, Mask)) {
1183 Value *NewNeg = Builder->CreateNeg(Op0RHS);
1184 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
1189 case Instruction::Shl:
1190 case Instruction::LShr:
1191 // (1 << x) & 1 --> zext(x == 0)
1192 // (1 >> x) & 1 --> zext(x == 0)
1193 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
1195 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
1196 return new ZExtInst(NewICmp, I.getType());
1201 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1202 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1204 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
1205 // If this is an integer truncation or change from signed-to-unsigned, and
1206 // if the source is an and/or with immediate, transform it. This
1207 // frequently occurs for bitfield accesses.
1208 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
1209 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
1210 CastOp->getNumOperands() == 2)
1211 if (ConstantInt *AndCI =dyn_cast<ConstantInt>(CastOp->getOperand(1))){
1212 if (CastOp->getOpcode() == Instruction::And) {
1213 // Change: and (cast (and X, C1) to T), C2
1214 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
1215 // This will fold the two constants together, which may allow
1216 // other simplifications.
1217 Value *NewCast = Builder->CreateTruncOrBitCast(
1218 CastOp->getOperand(0), I.getType(),
1219 CastOp->getName()+".shrunk");
1220 // trunc_or_bitcast(C1)&C2
1221 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
1222 C3 = ConstantExpr::getAnd(C3, AndRHS);
1223 return BinaryOperator::CreateAnd(NewCast, C3);
1224 } else if (CastOp->getOpcode() == Instruction::Or) {
1225 // Change: and (cast (or X, C1) to T), C2
1226 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
1227 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
1228 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
1230 return ReplaceInstUsesWith(I, AndRHS);
1236 // Try to fold constant and into select arguments.
1237 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1238 if (Instruction *R = FoldOpIntoSelect(I, SI))
1240 if (isa<PHINode>(Op0))
1241 if (Instruction *NV = FoldOpIntoPhi(I))
1246 // (~A & ~B) == (~(A | B)) - De Morgan's Law
1247 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1248 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1249 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1250 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
1251 I.getName()+".demorgan");
1252 return BinaryOperator::CreateNot(Or);
1256 Value *A = 0, *B = 0, *C = 0, *D = 0;
1257 // (A|B) & ~(A&B) -> A^B
1258 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1259 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1260 ((A == C && B == D) || (A == D && B == C)))
1261 return BinaryOperator::CreateXor(A, B);
1263 // ~(A&B) & (A|B) -> A^B
1264 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1265 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1266 ((A == C && B == D) || (A == D && B == C)))
1267 return BinaryOperator::CreateXor(A, B);
1269 if (Op0->hasOneUse() &&
1270 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
1271 if (A == Op1) { // (A^B)&A -> A&(A^B)
1272 I.swapOperands(); // Simplify below
1273 std::swap(Op0, Op1);
1274 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
1275 cast<BinaryOperator>(Op0)->swapOperands();
1276 I.swapOperands(); // Simplify below
1277 std::swap(Op0, Op1);
1281 if (Op1->hasOneUse() &&
1282 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
1283 if (B == Op0) { // B&(A^B) -> B&(B^A)
1284 cast<BinaryOperator>(Op1)->swapOperands();
1287 if (A == Op0) // A&(A^B) -> A & ~B
1288 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
1291 // (A&((~A)|B)) -> A&B
1292 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
1293 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
1294 return BinaryOperator::CreateAnd(A, Op1);
1295 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
1296 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
1297 return BinaryOperator::CreateAnd(A, Op0);
1300 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1))
1301 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
1302 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
1305 // fold (and (cast A), (cast B)) -> (cast (and A, B))
1306 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
1307 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
1308 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
1309 const Type *SrcTy = Op0C->getOperand(0)->getType();
1310 if (SrcTy == Op1C->getOperand(0)->getType() &&
1311 SrcTy->isIntOrIntVector() &&
1312 // Only do this if the casts both really cause code to be generated.
1313 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
1315 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
1317 Value *NewOp = Builder->CreateAnd(Op0C->getOperand(0),
1318 Op1C->getOperand(0), I.getName());
1319 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1323 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
1324 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1325 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1326 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1327 SI0->getOperand(1) == SI1->getOperand(1) &&
1328 (SI0->hasOneUse() || SI1->hasOneUse())) {
1330 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
1332 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1333 SI1->getOperand(1));
1337 // If and'ing two fcmp, try combine them into one.
1338 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
1339 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1340 if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
1344 return Changed ? &I : 0;
1347 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
1348 /// capable of providing pieces of a bswap. The subexpression provides pieces
1349 /// of a bswap if it is proven that each of the non-zero bytes in the output of
1350 /// the expression came from the corresponding "byte swapped" byte in some other
1351 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
1352 /// we know that the expression deposits the low byte of %X into the high byte
1353 /// of the bswap result and that all other bytes are zero. This expression is
1354 /// accepted, the high byte of ByteValues is set to X to indicate a correct
1357 /// This function returns true if the match was unsuccessful and false if so.
1358 /// On entry to the function the "OverallLeftShift" is a signed integer value
1359 /// indicating the number of bytes that the subexpression is later shifted. For
1360 /// example, if the expression is later right shifted by 16 bits, the
1361 /// OverallLeftShift value would be -2 on entry. This is used to specify which
1362 /// byte of ByteValues is actually being set.
1364 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
1365 /// byte is masked to zero by a user. For example, in (X & 255), X will be
1366 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
1367 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
1368 /// always in the local (OverallLeftShift) coordinate space.
1370 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
1371 SmallVector<Value*, 8> &ByteValues) {
1372 if (Instruction *I = dyn_cast<Instruction>(V)) {
1373 // If this is an or instruction, it may be an inner node of the bswap.
1374 if (I->getOpcode() == Instruction::Or) {
1375 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1377 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
1381 // If this is a logical shift by a constant multiple of 8, recurse with
1382 // OverallLeftShift and ByteMask adjusted.
1383 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
1385 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
1386 // Ensure the shift amount is defined and of a byte value.
1387 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
1390 unsigned ByteShift = ShAmt >> 3;
1391 if (I->getOpcode() == Instruction::Shl) {
1392 // X << 2 -> collect(X, +2)
1393 OverallLeftShift += ByteShift;
1394 ByteMask >>= ByteShift;
1396 // X >>u 2 -> collect(X, -2)
1397 OverallLeftShift -= ByteShift;
1398 ByteMask <<= ByteShift;
1399 ByteMask &= (~0U >> (32-ByteValues.size()));
1402 if (OverallLeftShift >= (int)ByteValues.size()) return true;
1403 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
1405 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1409 // If this is a logical 'and' with a mask that clears bytes, clear the
1410 // corresponding bytes in ByteMask.
1411 if (I->getOpcode() == Instruction::And &&
1412 isa<ConstantInt>(I->getOperand(1))) {
1413 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
1414 unsigned NumBytes = ByteValues.size();
1415 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
1416 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
1418 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
1419 // If this byte is masked out by a later operation, we don't care what
1421 if ((ByteMask & (1 << i)) == 0)
1424 // If the AndMask is all zeros for this byte, clear the bit.
1425 APInt MaskB = AndMask & Byte;
1427 ByteMask &= ~(1U << i);
1431 // If the AndMask is not all ones for this byte, it's not a bytezap.
1435 // Otherwise, this byte is kept.
1438 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1443 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
1444 // the input value to the bswap. Some observations: 1) if more than one byte
1445 // is demanded from this input, then it could not be successfully assembled
1446 // into a byteswap. At least one of the two bytes would not be aligned with
1447 // their ultimate destination.
1448 if (!isPowerOf2_32(ByteMask)) return true;
1449 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
1451 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
1452 // is demanded, it needs to go into byte 0 of the result. This means that the
1453 // byte needs to be shifted until it lands in the right byte bucket. The
1454 // shift amount depends on the position: if the byte is coming from the high
1455 // part of the value (e.g. byte 3) then it must be shifted right. If from the
1456 // low part, it must be shifted left.
1457 unsigned DestByteNo = InputByteNo + OverallLeftShift;
1458 if (InputByteNo < ByteValues.size()/2) {
1459 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1462 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1466 // If the destination byte value is already defined, the values are or'd
1467 // together, which isn't a bswap (unless it's an or of the same bits).
1468 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
1470 ByteValues[DestByteNo] = V;
1474 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
1475 /// If so, insert the new bswap intrinsic and return it.
1476 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
1477 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
1478 if (!ITy || ITy->getBitWidth() % 16 ||
1479 // ByteMask only allows up to 32-byte values.
1480 ITy->getBitWidth() > 32*8)
1481 return 0; // Can only bswap pairs of bytes. Can't do vectors.
1483 /// ByteValues - For each byte of the result, we keep track of which value
1484 /// defines each byte.
1485 SmallVector<Value*, 8> ByteValues;
1486 ByteValues.resize(ITy->getBitWidth()/8);
1488 // Try to find all the pieces corresponding to the bswap.
1489 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
1490 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
1493 // Check to see if all of the bytes come from the same value.
1494 Value *V = ByteValues[0];
1495 if (V == 0) return 0; // Didn't find a byte? Must be zero.
1497 // Check to make sure that all of the bytes come from the same value.
1498 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
1499 if (ByteValues[i] != V)
1501 const Type *Tys[] = { ITy };
1502 Module *M = I.getParent()->getParent()->getParent();
1503 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
1504 return CallInst::Create(F, V);
1507 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
1508 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
1509 /// we can simplify this expression to "cond ? C : D or B".
1510 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
1511 Value *C, Value *D) {
1512 // If A is not a select of -1/0, this cannot match.
1514 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
1517 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
1518 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
1519 return SelectInst::Create(Cond, C, B);
1520 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
1521 return SelectInst::Create(Cond, C, B);
1522 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
1523 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
1524 return SelectInst::Create(Cond, C, D);
1525 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
1526 return SelectInst::Create(Cond, C, D);
1530 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
1531 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
1532 ICmpInst *LHS, ICmpInst *RHS) {
1533 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
1535 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
1536 if (PredicatesFoldable(LHSCC, RHSCC)) {
1537 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1538 LHS->getOperand(1) == RHS->getOperand(0))
1539 LHS->swapOperands();
1540 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1541 LHS->getOperand(1) == RHS->getOperand(1)) {
1542 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1543 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
1544 bool isSigned = LHS->isSigned() || RHS->isSigned();
1545 Value *RV = getICmpValue(isSigned, Code, Op0, Op1);
1546 if (Instruction *I = dyn_cast<Instruction>(RV))
1548 // Otherwise, it's a constant boolean value.
1549 return ReplaceInstUsesWith(I, RV);
1553 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
1554 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
1555 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
1556 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
1557 if (LHSCst == 0 || RHSCst == 0) return 0;
1559 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
1560 if (LHSCst == RHSCst && LHSCC == RHSCC &&
1561 LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
1562 Value *NewOr = Builder->CreateOr(Val, Val2);
1563 return new ICmpInst(LHSCC, NewOr, LHSCst);
1566 // From here on, we only handle:
1567 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
1568 if (Val != Val2) return 0;
1570 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
1571 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
1572 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
1573 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
1574 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
1577 // We can't fold (ugt x, C) | (sgt x, C2).
1578 if (!PredicatesFoldable(LHSCC, RHSCC))
1581 // Ensure that the larger constant is on the RHS.
1583 if (CmpInst::isSigned(LHSCC) ||
1584 (ICmpInst::isEquality(LHSCC) &&
1585 CmpInst::isSigned(RHSCC)))
1586 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1588 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1591 std::swap(LHS, RHS);
1592 std::swap(LHSCst, RHSCst);
1593 std::swap(LHSCC, RHSCC);
1596 // At this point, we know we have have two icmp instructions
1597 // comparing a value against two constants and or'ing the result
1598 // together. Because of the above check, we know that we only have
1599 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
1600 // icmp folding check above), that the two constants are not
1602 assert(LHSCst != RHSCst && "Compares not folded above?");
1605 default: llvm_unreachable("Unknown integer condition code!");
1606 case ICmpInst::ICMP_EQ:
1608 default: llvm_unreachable("Unknown integer condition code!");
1609 case ICmpInst::ICMP_EQ:
1610 if (LHSCst == SubOne(RHSCst)) {
1611 // (X == 13 | X == 14) -> X-13 <u 2
1612 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1613 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1614 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
1615 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
1617 break; // (X == 13 | X == 15) -> no change
1618 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
1619 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
1621 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
1622 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
1623 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
1624 return ReplaceInstUsesWith(I, RHS);
1627 case ICmpInst::ICMP_NE:
1629 default: llvm_unreachable("Unknown integer condition code!");
1630 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
1631 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
1632 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
1633 return ReplaceInstUsesWith(I, LHS);
1634 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
1635 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
1636 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
1637 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1640 case ICmpInst::ICMP_ULT:
1642 default: llvm_unreachable("Unknown integer condition code!");
1643 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
1645 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
1646 // If RHSCst is [us]MAXINT, it is always false. Not handling
1647 // this can cause overflow.
1648 if (RHSCst->isMaxValue(false))
1649 return ReplaceInstUsesWith(I, LHS);
1650 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
1652 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
1654 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
1655 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
1656 return ReplaceInstUsesWith(I, RHS);
1657 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
1661 case ICmpInst::ICMP_SLT:
1663 default: llvm_unreachable("Unknown integer condition code!");
1664 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
1666 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
1667 // If RHSCst is [us]MAXINT, it is always false. Not handling
1668 // this can cause overflow.
1669 if (RHSCst->isMaxValue(true))
1670 return ReplaceInstUsesWith(I, LHS);
1671 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
1673 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
1675 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
1676 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
1677 return ReplaceInstUsesWith(I, RHS);
1678 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
1682 case ICmpInst::ICMP_UGT:
1684 default: llvm_unreachable("Unknown integer condition code!");
1685 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
1686 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
1687 return ReplaceInstUsesWith(I, LHS);
1688 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
1690 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
1691 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
1692 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1693 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
1697 case ICmpInst::ICMP_SGT:
1699 default: llvm_unreachable("Unknown integer condition code!");
1700 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
1701 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
1702 return ReplaceInstUsesWith(I, LHS);
1703 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
1705 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
1706 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
1707 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1708 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
1716 Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
1718 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
1719 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
1720 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
1721 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1722 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1723 // If either of the constants are nans, then the whole thing returns
1725 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1726 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1728 // Otherwise, no need to compare the two constants, compare the
1730 return new FCmpInst(FCmpInst::FCMP_UNO,
1731 LHS->getOperand(0), RHS->getOperand(0));
1734 // Handle vector zeros. This occurs because the canonical form of
1735 // "fcmp uno x,x" is "fcmp uno x, 0".
1736 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1737 isa<ConstantAggregateZero>(RHS->getOperand(1)))
1738 return new FCmpInst(FCmpInst::FCMP_UNO,
1739 LHS->getOperand(0), RHS->getOperand(0));
1744 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1745 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1746 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1748 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1749 // Swap RHS operands to match LHS.
1750 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1751 std::swap(Op1LHS, Op1RHS);
1753 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
1754 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
1756 return new FCmpInst((FCmpInst::Predicate)Op0CC,
1758 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
1759 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1760 if (Op0CC == FCmpInst::FCMP_FALSE)
1761 return ReplaceInstUsesWith(I, RHS);
1762 if (Op1CC == FCmpInst::FCMP_FALSE)
1763 return ReplaceInstUsesWith(I, LHS);
1766 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
1767 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
1768 if (Op0Ordered == Op1Ordered) {
1769 // If both are ordered or unordered, return a new fcmp with
1770 // or'ed predicates.
1771 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS);
1772 if (Instruction *I = dyn_cast<Instruction>(RV))
1774 // Otherwise, it's a constant boolean value...
1775 return ReplaceInstUsesWith(I, RV);
1781 /// FoldOrWithConstants - This helper function folds:
1783 /// ((A | B) & C1) | (B & C2)
1789 /// when the XOR of the two constants is "all ones" (-1).
1790 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
1791 Value *A, Value *B, Value *C) {
1792 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
1796 ConstantInt *CI2 = 0;
1797 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
1799 APInt Xor = CI1->getValue() ^ CI2->getValue();
1800 if (!Xor.isAllOnesValue()) return 0;
1802 if (V1 == A || V1 == B) {
1803 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
1804 return BinaryOperator::CreateOr(NewOp, V1);
1810 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
1811 bool Changed = SimplifyCommutative(I);
1812 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1814 if (Value *V = SimplifyOrInst(Op0, Op1, TD))
1815 return ReplaceInstUsesWith(I, V);
1818 // See if we can simplify any instructions used by the instruction whose sole
1819 // purpose is to compute bits we don't care about.
1820 if (SimplifyDemandedInstructionBits(I))
1823 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1824 ConstantInt *C1 = 0; Value *X = 0;
1825 // (X & C1) | C2 --> (X | C2) & (C1|C2)
1826 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
1828 Value *Or = Builder->CreateOr(X, RHS);
1830 return BinaryOperator::CreateAnd(Or,
1831 ConstantInt::get(I.getContext(),
1832 RHS->getValue() | C1->getValue()));
1835 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
1836 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
1838 Value *Or = Builder->CreateOr(X, RHS);
1840 return BinaryOperator::CreateXor(Or,
1841 ConstantInt::get(I.getContext(),
1842 C1->getValue() & ~RHS->getValue()));
1845 // Try to fold constant and into select arguments.
1846 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1847 if (Instruction *R = FoldOpIntoSelect(I, SI))
1849 if (isa<PHINode>(Op0))
1850 if (Instruction *NV = FoldOpIntoPhi(I))
1854 Value *A = 0, *B = 0;
1855 ConstantInt *C1 = 0, *C2 = 0;
1857 // (A | B) | C and A | (B | C) -> bswap if possible.
1858 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
1859 if (match(Op0, m_Or(m_Value(), m_Value())) ||
1860 match(Op1, m_Or(m_Value(), m_Value())) ||
1861 (match(Op0, m_Shift(m_Value(), m_Value())) &&
1862 match(Op1, m_Shift(m_Value(), m_Value())))) {
1863 if (Instruction *BSwap = MatchBSwap(I))
1867 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
1868 if (Op0->hasOneUse() &&
1869 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1870 MaskedValueIsZero(Op1, C1->getValue())) {
1871 Value *NOr = Builder->CreateOr(A, Op1);
1873 return BinaryOperator::CreateXor(NOr, C1);
1876 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
1877 if (Op1->hasOneUse() &&
1878 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1879 MaskedValueIsZero(Op0, C1->getValue())) {
1880 Value *NOr = Builder->CreateOr(A, Op0);
1882 return BinaryOperator::CreateXor(NOr, C1);
1886 Value *C = 0, *D = 0;
1887 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1888 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1889 Value *V1 = 0, *V2 = 0, *V3 = 0;
1890 C1 = dyn_cast<ConstantInt>(C);
1891 C2 = dyn_cast<ConstantInt>(D);
1892 if (C1 && C2) { // (A & C1)|(B & C2)
1893 // If we have: ((V + N) & C1) | (V & C2)
1894 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1895 // replace with V+N.
1896 if (C1->getValue() == ~C2->getValue()) {
1897 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
1898 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1899 // Add commutes, try both ways.
1900 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
1901 return ReplaceInstUsesWith(I, A);
1902 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
1903 return ReplaceInstUsesWith(I, A);
1905 // Or commutes, try both ways.
1906 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
1907 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1908 // Add commutes, try both ways.
1909 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
1910 return ReplaceInstUsesWith(I, B);
1911 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
1912 return ReplaceInstUsesWith(I, B);
1916 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
1917 // iff (C1&C2) == 0 and (N&~C1) == 0
1918 if ((C1->getValue() & C2->getValue()) == 0) {
1919 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
1920 ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N)
1921 (V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V)
1922 return BinaryOperator::CreateAnd(A,
1923 ConstantInt::get(A->getContext(),
1924 C1->getValue()|C2->getValue()));
1925 // Or commutes, try both ways.
1926 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
1927 ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N)
1928 (V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V)
1929 return BinaryOperator::CreateAnd(B,
1930 ConstantInt::get(B->getContext(),
1931 C1->getValue()|C2->getValue()));
1935 // Check to see if we have any common things being and'ed. If so, find the
1936 // terms for V1 & (V2|V3).
1937 if (Op0->hasOneUse() || Op1->hasOneUse()) {
1939 if (A == B) // (A & C)|(A & D) == A & (C|D)
1940 V1 = A, V2 = C, V3 = D;
1941 else if (A == D) // (A & C)|(B & A) == A & (B|C)
1942 V1 = A, V2 = B, V3 = C;
1943 else if (C == B) // (A & C)|(C & D) == C & (A|D)
1944 V1 = C, V2 = A, V3 = D;
1945 else if (C == D) // (A & C)|(B & C) == C & (A|B)
1946 V1 = C, V2 = A, V3 = B;
1949 Value *Or = Builder->CreateOr(V2, V3, "tmp");
1950 return BinaryOperator::CreateAnd(V1, Or);
1954 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
1955 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
1957 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
1959 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
1961 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
1964 // ((A&~B)|(~A&B)) -> A^B
1965 if ((match(C, m_Not(m_Specific(D))) &&
1966 match(B, m_Not(m_Specific(A)))))
1967 return BinaryOperator::CreateXor(A, D);
1968 // ((~B&A)|(~A&B)) -> A^B
1969 if ((match(A, m_Not(m_Specific(D))) &&
1970 match(B, m_Not(m_Specific(C)))))
1971 return BinaryOperator::CreateXor(C, D);
1972 // ((A&~B)|(B&~A)) -> A^B
1973 if ((match(C, m_Not(m_Specific(B))) &&
1974 match(D, m_Not(m_Specific(A)))))
1975 return BinaryOperator::CreateXor(A, B);
1976 // ((~B&A)|(B&~A)) -> A^B
1977 if ((match(A, m_Not(m_Specific(B))) &&
1978 match(D, m_Not(m_Specific(C)))))
1979 return BinaryOperator::CreateXor(C, B);
1982 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
1983 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1984 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1985 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1986 SI0->getOperand(1) == SI1->getOperand(1) &&
1987 (SI0->hasOneUse() || SI1->hasOneUse())) {
1988 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
1990 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1991 SI1->getOperand(1));
1995 // ((A|B)&1)|(B&-2) -> (A&1) | B
1996 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
1997 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
1998 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
1999 if (Ret) return Ret;
2001 // (B&-2)|((A|B)&1) -> (A&1) | B
2002 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
2003 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
2004 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
2005 if (Ret) return Ret;
2008 // (~A | ~B) == (~(A & B)) - De Morgan's Law
2009 if (Value *Op0NotVal = dyn_castNotVal(Op0))
2010 if (Value *Op1NotVal = dyn_castNotVal(Op1))
2011 if (Op0->hasOneUse() && Op1->hasOneUse()) {
2012 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
2013 I.getName()+".demorgan");
2014 return BinaryOperator::CreateNot(And);
2017 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2018 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2019 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
2022 // fold (or (cast A), (cast B)) -> (cast (or A, B))
2023 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2024 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2025 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
2026 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
2027 !isa<ICmpInst>(Op1C->getOperand(0))) {
2028 const Type *SrcTy = Op0C->getOperand(0)->getType();
2029 if (SrcTy == Op1C->getOperand(0)->getType() &&
2030 SrcTy->isIntOrIntVector() &&
2031 // Only do this if the casts both really cause code to be
2033 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
2035 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
2037 Value *NewOp = Builder->CreateOr(Op0C->getOperand(0),
2038 Op1C->getOperand(0), I.getName());
2039 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2046 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
2047 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
2048 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2049 if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
2053 return Changed ? &I : 0;
2056 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2057 bool Changed = SimplifyCommutative(I);
2058 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2060 if (isa<UndefValue>(Op1)) {
2061 if (isa<UndefValue>(Op0))
2062 // Handle undef ^ undef -> 0 special case. This is a common
2064 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2065 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
2070 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2072 // See if we can simplify any instructions used by the instruction whose sole
2073 // purpose is to compute bits we don't care about.
2074 if (SimplifyDemandedInstructionBits(I))
2076 if (isa<VectorType>(I.getType()))
2077 if (isa<ConstantAggregateZero>(Op1))
2078 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
2080 // Is this a ~ operation?
2081 if (Value *NotOp = dyn_castNotVal(&I)) {
2082 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
2083 if (Op0I->getOpcode() == Instruction::And ||
2084 Op0I->getOpcode() == Instruction::Or) {
2085 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
2086 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
2087 if (dyn_castNotVal(Op0I->getOperand(1)))
2088 Op0I->swapOperands();
2089 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2091 Builder->CreateNot(Op0I->getOperand(1),
2092 Op0I->getOperand(1)->getName()+".not");
2093 if (Op0I->getOpcode() == Instruction::And)
2094 return BinaryOperator::CreateOr(Op0NotVal, NotY);
2095 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
2098 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
2099 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
2100 if (isFreeToInvert(Op0I->getOperand(0)) &&
2101 isFreeToInvert(Op0I->getOperand(1))) {
2103 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
2105 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
2106 if (Op0I->getOpcode() == Instruction::And)
2107 return BinaryOperator::CreateOr(NotX, NotY);
2108 return BinaryOperator::CreateAnd(NotX, NotY);
2115 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2116 if (RHS->isOne() && Op0->hasOneUse()) {
2117 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
2118 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
2119 return new ICmpInst(ICI->getInversePredicate(),
2120 ICI->getOperand(0), ICI->getOperand(1));
2122 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
2123 return new FCmpInst(FCI->getInversePredicate(),
2124 FCI->getOperand(0), FCI->getOperand(1));
2127 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
2128 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2129 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
2130 if (CI->hasOneUse() && Op0C->hasOneUse()) {
2131 Instruction::CastOps Opcode = Op0C->getOpcode();
2132 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
2133 (RHS == ConstantExpr::getCast(Opcode,
2134 ConstantInt::getTrue(I.getContext()),
2135 Op0C->getDestTy()))) {
2136 CI->setPredicate(CI->getInversePredicate());
2137 return CastInst::Create(Opcode, CI, Op0C->getType());
2143 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2144 // ~(c-X) == X-c-1 == X+(-c-1)
2145 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2146 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2147 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2148 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2149 ConstantInt::get(I.getType(), 1));
2150 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
2153 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2154 if (Op0I->getOpcode() == Instruction::Add) {
2155 // ~(X-c) --> (-c-1)-X
2156 if (RHS->isAllOnesValue()) {
2157 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2158 return BinaryOperator::CreateSub(
2159 ConstantExpr::getSub(NegOp0CI,
2160 ConstantInt::get(I.getType(), 1)),
2161 Op0I->getOperand(0));
2162 } else if (RHS->getValue().isSignBit()) {
2163 // (X + C) ^ signbit -> (X + C + signbit)
2164 Constant *C = ConstantInt::get(I.getContext(),
2165 RHS->getValue() + Op0CI->getValue());
2166 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
2169 } else if (Op0I->getOpcode() == Instruction::Or) {
2170 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
2171 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
2172 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
2173 // Anything in both C1 and C2 is known to be zero, remove it from
2175 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
2176 NewRHS = ConstantExpr::getAnd(NewRHS,
2177 ConstantExpr::getNot(CommonBits));
2179 I.setOperand(0, Op0I->getOperand(0));
2180 I.setOperand(1, NewRHS);
2187 // Try to fold constant and into select arguments.
2188 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2189 if (Instruction *R = FoldOpIntoSelect(I, SI))
2191 if (isa<PHINode>(Op0))
2192 if (Instruction *NV = FoldOpIntoPhi(I))
2196 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
2198 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2200 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
2202 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2205 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
2208 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2209 if (A == Op0) { // B^(B|A) == (A|B)^B
2210 Op1I->swapOperands();
2212 std::swap(Op0, Op1);
2213 } else if (B == Op0) { // B^(A|B) == (A|B)^B
2214 I.swapOperands(); // Simplified below.
2215 std::swap(Op0, Op1);
2217 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
2218 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
2219 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
2220 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
2221 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
2223 if (A == Op0) { // A^(A&B) -> A^(B&A)
2224 Op1I->swapOperands();
2227 if (B == Op0) { // A^(B&A) -> (B&A)^A
2228 I.swapOperands(); // Simplified below.
2229 std::swap(Op0, Op1);
2234 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
2237 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2238 Op0I->hasOneUse()) {
2239 if (A == Op1) // (B|A)^B == (A|B)^B
2241 if (B == Op1) // (A|B)^B == A & ~B
2242 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
2243 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
2244 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
2245 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
2246 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
2247 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2249 if (A == Op1) // (A&B)^A -> (B&A)^A
2251 if (B == Op1 && // (B&A)^A == ~B & A
2252 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
2253 return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
2258 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
2259 if (Op0I && Op1I && Op0I->isShift() &&
2260 Op0I->getOpcode() == Op1I->getOpcode() &&
2261 Op0I->getOperand(1) == Op1I->getOperand(1) &&
2262 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
2264 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
2266 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
2267 Op1I->getOperand(1));
2271 Value *A, *B, *C, *D;
2272 // (A & B)^(A | B) -> A ^ B
2273 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2274 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
2275 if ((A == C && B == D) || (A == D && B == C))
2276 return BinaryOperator::CreateXor(A, B);
2278 // (A | B)^(A & B) -> A ^ B
2279 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2280 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
2281 if ((A == C && B == D) || (A == D && B == C))
2282 return BinaryOperator::CreateXor(A, B);
2286 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
2287 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2288 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
2289 // (X & Y)^(X & Y) -> (Y^Z) & X
2290 Value *X = 0, *Y = 0, *Z = 0;
2292 X = A, Y = B, Z = D;
2294 X = A, Y = B, Z = C;
2296 X = B, Y = A, Z = D;
2298 X = B, Y = A, Z = C;
2301 Value *NewOp = Builder->CreateXor(Y, Z, Op0->getName());
2302 return BinaryOperator::CreateAnd(NewOp, X);
2307 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2308 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2309 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2310 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2311 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2312 LHS->getOperand(1) == RHS->getOperand(0))
2313 LHS->swapOperands();
2314 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2315 LHS->getOperand(1) == RHS->getOperand(1)) {
2316 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2317 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2318 bool isSigned = LHS->isSigned() || RHS->isSigned();
2319 Value *RV = getICmpValue(isSigned, Code, Op0, Op1);
2320 if (Instruction *I = dyn_cast<Instruction>(RV))
2322 // Otherwise, it's a constant boolean value.
2323 return ReplaceInstUsesWith(I, RV);
2327 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
2328 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2329 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2330 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
2331 const Type *SrcTy = Op0C->getOperand(0)->getType();
2332 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
2333 // Only do this if the casts both really cause code to be generated.
2334 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
2336 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
2338 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
2339 Op1C->getOperand(0), I.getName());
2340 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2345 return Changed ? &I : 0;
2349 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
2350 return commonShiftTransforms(I);
2353 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
2354 return commonShiftTransforms(I);
2357 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
2358 if (Instruction *R = commonShiftTransforms(I))
2361 Value *Op0 = I.getOperand(0);
2363 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
2364 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2365 if (CSI->isAllOnesValue())
2366 return ReplaceInstUsesWith(I, CSI);
2368 // See if we can turn a signed shr into an unsigned shr.
2369 if (MaskedValueIsZero(Op0,
2370 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
2371 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
2373 // Arithmetic shifting an all-sign-bit value is a no-op.
2374 unsigned NumSignBits = ComputeNumSignBits(Op0);
2375 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
2376 return ReplaceInstUsesWith(I, Op0);
2381 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
2382 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
2383 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2385 // shl X, 0 == X and shr X, 0 == X
2386 // shl 0, X == 0 and shr 0, X == 0
2387 if (Op1 == Constant::getNullValue(Op1->getType()) ||
2388 Op0 == Constant::getNullValue(Op0->getType()))
2389 return ReplaceInstUsesWith(I, Op0);
2391 if (isa<UndefValue>(Op0)) {
2392 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
2393 return ReplaceInstUsesWith(I, Op0);
2394 else // undef << X -> 0, undef >>u X -> 0
2395 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2397 if (isa<UndefValue>(Op1)) {
2398 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
2399 return ReplaceInstUsesWith(I, Op0);
2400 else // X << undef, X >>u undef -> 0
2401 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2404 // See if we can fold away this shift.
2405 if (SimplifyDemandedInstructionBits(I))
2408 // Try to fold constant and into select arguments.
2409 if (isa<Constant>(Op0))
2410 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2411 if (Instruction *R = FoldOpIntoSelect(I, SI))
2414 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
2415 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
2420 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
2421 BinaryOperator &I) {
2422 bool isLeftShift = I.getOpcode() == Instruction::Shl;
2424 // See if we can simplify any instructions used by the instruction whose sole
2425 // purpose is to compute bits we don't care about.
2426 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
2428 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
2431 if (Op1->uge(TypeBits)) {
2432 if (I.getOpcode() != Instruction::AShr)
2433 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
2435 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
2440 // ((X*C1) << C2) == (X * (C1 << C2))
2441 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
2442 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
2443 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
2444 return BinaryOperator::CreateMul(BO->getOperand(0),
2445 ConstantExpr::getShl(BOOp, Op1));
2447 // Try to fold constant and into select arguments.
2448 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2449 if (Instruction *R = FoldOpIntoSelect(I, SI))
2451 if (isa<PHINode>(Op0))
2452 if (Instruction *NV = FoldOpIntoPhi(I))
2455 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
2456 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
2457 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
2458 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
2459 // place. Don't try to do this transformation in this case. Also, we
2460 // require that the input operand is a shift-by-constant so that we have
2461 // confidence that the shifts will get folded together. We could do this
2462 // xform in more cases, but it is unlikely to be profitable.
2463 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
2464 isa<ConstantInt>(TrOp->getOperand(1))) {
2465 // Okay, we'll do this xform. Make the shift of shift.
2466 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
2467 // (shift2 (shift1 & 0x00FF), c2)
2468 Value *NSh = Builder->CreateBinOp(I.getOpcode(), TrOp, ShAmt,I.getName());
2470 // For logical shifts, the truncation has the effect of making the high
2471 // part of the register be zeros. Emulate this by inserting an AND to
2472 // clear the top bits as needed. This 'and' will usually be zapped by
2473 // other xforms later if dead.
2474 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
2475 unsigned DstSize = TI->getType()->getScalarSizeInBits();
2476 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
2478 // The mask we constructed says what the trunc would do if occurring
2479 // between the shifts. We want to know the effect *after* the second
2480 // shift. We know that it is a logical shift by a constant, so adjust the
2481 // mask as appropriate.
2482 if (I.getOpcode() == Instruction::Shl)
2483 MaskV <<= Op1->getZExtValue();
2485 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
2486 MaskV = MaskV.lshr(Op1->getZExtValue());
2490 Value *And = Builder->CreateAnd(NSh,
2491 ConstantInt::get(I.getContext(), MaskV),
2494 // Return the value truncated to the interesting size.
2495 return new TruncInst(And, I.getType());
2499 if (Op0->hasOneUse()) {
2500 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
2501 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
2504 switch (Op0BO->getOpcode()) {
2506 case Instruction::Add:
2507 case Instruction::And:
2508 case Instruction::Or:
2509 case Instruction::Xor: {
2510 // These operators commute.
2511 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
2512 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
2513 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
2514 m_Specific(Op1)))) {
2515 Value *YS = // (Y << C)
2516 Builder->CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
2518 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), YS, V1,
2519 Op0BO->getOperand(1)->getName());
2520 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
2521 return BinaryOperator::CreateAnd(X, ConstantInt::get(I.getContext(),
2522 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
2525 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
2526 Value *Op0BOOp1 = Op0BO->getOperand(1);
2527 if (isLeftShift && Op0BOOp1->hasOneUse() &&
2529 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
2530 m_ConstantInt(CC))) &&
2531 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
2532 Value *YS = // (Y << C)
2533 Builder->CreateShl(Op0BO->getOperand(0), Op1,
2536 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
2537 V1->getName()+".mask");
2538 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
2543 case Instruction::Sub: {
2544 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
2545 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
2546 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
2547 m_Specific(Op1)))) {
2548 Value *YS = // (Y << C)
2549 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
2551 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), V1, YS,
2552 Op0BO->getOperand(0)->getName());
2553 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
2554 return BinaryOperator::CreateAnd(X, ConstantInt::get(I.getContext(),
2555 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
2558 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
2559 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
2560 match(Op0BO->getOperand(0),
2561 m_And(m_Shr(m_Value(V1), m_Value(V2)),
2562 m_ConstantInt(CC))) && V2 == Op1 &&
2563 cast<BinaryOperator>(Op0BO->getOperand(0))
2564 ->getOperand(0)->hasOneUse()) {
2565 Value *YS = // (Y << C)
2566 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
2568 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
2569 V1->getName()+".mask");
2571 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
2579 // If the operand is an bitwise operator with a constant RHS, and the
2580 // shift is the only use, we can pull it out of the shift.
2581 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
2582 bool isValid = true; // Valid only for And, Or, Xor
2583 bool highBitSet = false; // Transform if high bit of constant set?
2585 switch (Op0BO->getOpcode()) {
2586 default: isValid = false; break; // Do not perform transform!
2587 case Instruction::Add:
2588 isValid = isLeftShift;
2590 case Instruction::Or:
2591 case Instruction::Xor:
2594 case Instruction::And:
2599 // If this is a signed shift right, and the high bit is modified
2600 // by the logical operation, do not perform the transformation.
2601 // The highBitSet boolean indicates the value of the high bit of
2602 // the constant which would cause it to be modified for this
2605 if (isValid && I.getOpcode() == Instruction::AShr)
2606 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
2609 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
2612 Builder->CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
2613 NewShift->takeName(Op0BO);
2615 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
2622 // Find out if this is a shift of a shift by a constant.
2623 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
2624 if (ShiftOp && !ShiftOp->isShift())
2627 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
2628 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
2629 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
2630 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
2631 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
2632 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
2633 Value *X = ShiftOp->getOperand(0);
2635 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
2637 const IntegerType *Ty = cast<IntegerType>(I.getType());
2639 // Check for (X << c1) << c2 and (X >> c1) >> c2
2640 if (I.getOpcode() == ShiftOp->getOpcode()) {
2641 // If this is oversized composite shift, then unsigned shifts get 0, ashr
2643 if (AmtSum >= TypeBits) {
2644 if (I.getOpcode() != Instruction::AShr)
2645 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2646 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
2649 return BinaryOperator::Create(I.getOpcode(), X,
2650 ConstantInt::get(Ty, AmtSum));
2653 if (ShiftOp->getOpcode() == Instruction::LShr &&
2654 I.getOpcode() == Instruction::AShr) {
2655 if (AmtSum >= TypeBits)
2656 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2658 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
2659 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
2662 if (ShiftOp->getOpcode() == Instruction::AShr &&
2663 I.getOpcode() == Instruction::LShr) {
2664 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
2665 if (AmtSum >= TypeBits)
2666 AmtSum = TypeBits-1;
2668 Value *Shift = Builder->CreateAShr(X, ConstantInt::get(Ty, AmtSum));
2670 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
2671 return BinaryOperator::CreateAnd(Shift,
2672 ConstantInt::get(I.getContext(), Mask));
2675 // Okay, if we get here, one shift must be left, and the other shift must be
2676 // right. See if the amounts are equal.
2677 if (ShiftAmt1 == ShiftAmt2) {
2678 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
2679 if (I.getOpcode() == Instruction::Shl) {
2680 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
2681 return BinaryOperator::CreateAnd(X,
2682 ConstantInt::get(I.getContext(),Mask));
2684 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
2685 if (I.getOpcode() == Instruction::LShr) {
2686 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
2687 return BinaryOperator::CreateAnd(X,
2688 ConstantInt::get(I.getContext(), Mask));
2690 // We can simplify ((X << C) >>s C) into a trunc + sext.
2691 // NOTE: we could do this for any C, but that would make 'unusual' integer
2692 // types. For now, just stick to ones well-supported by the code
2694 const Type *SExtType = 0;
2695 switch (Ty->getBitWidth() - ShiftAmt1) {
2702 SExtType = IntegerType::get(I.getContext(),
2703 Ty->getBitWidth() - ShiftAmt1);
2708 return new SExtInst(Builder->CreateTrunc(X, SExtType, "sext"), Ty);
2709 // Otherwise, we can't handle it yet.
2710 } else if (ShiftAmt1 < ShiftAmt2) {
2711 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
2713 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
2714 if (I.getOpcode() == Instruction::Shl) {
2715 assert(ShiftOp->getOpcode() == Instruction::LShr ||
2716 ShiftOp->getOpcode() == Instruction::AShr);
2717 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
2719 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
2720 return BinaryOperator::CreateAnd(Shift,
2721 ConstantInt::get(I.getContext(),Mask));
2724 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
2725 if (I.getOpcode() == Instruction::LShr) {
2726 assert(ShiftOp->getOpcode() == Instruction::Shl);
2727 Value *Shift = Builder->CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
2729 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
2730 return BinaryOperator::CreateAnd(Shift,
2731 ConstantInt::get(I.getContext(),Mask));
2734 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
2736 assert(ShiftAmt2 < ShiftAmt1);
2737 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
2739 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
2740 if (I.getOpcode() == Instruction::Shl) {
2741 assert(ShiftOp->getOpcode() == Instruction::LShr ||
2742 ShiftOp->getOpcode() == Instruction::AShr);
2743 Value *Shift = Builder->CreateBinOp(ShiftOp->getOpcode(), X,
2744 ConstantInt::get(Ty, ShiftDiff));
2746 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
2747 return BinaryOperator::CreateAnd(Shift,
2748 ConstantInt::get(I.getContext(),Mask));
2751 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
2752 if (I.getOpcode() == Instruction::LShr) {
2753 assert(ShiftOp->getOpcode() == Instruction::Shl);
2754 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
2756 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
2757 return BinaryOperator::CreateAnd(Shift,
2758 ConstantInt::get(I.getContext(),Mask));
2761 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
2769 /// FindElementAtOffset - Given a type and a constant offset, determine whether
2770 /// or not there is a sequence of GEP indices into the type that will land us at
2771 /// the specified offset. If so, fill them into NewIndices and return the
2772 /// resultant element type, otherwise return null.
2773 const Type *InstCombiner::FindElementAtOffset(const Type *Ty, int64_t Offset,
2774 SmallVectorImpl<Value*> &NewIndices) {
2776 if (!Ty->isSized()) return 0;
2778 // Start with the index over the outer type. Note that the type size
2779 // might be zero (even if the offset isn't zero) if the indexed type
2780 // is something like [0 x {int, int}]
2781 const Type *IntPtrTy = TD->getIntPtrType(Ty->getContext());
2782 int64_t FirstIdx = 0;
2783 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
2784 FirstIdx = Offset/TySize;
2785 Offset -= FirstIdx*TySize;
2787 // Handle hosts where % returns negative instead of values [0..TySize).
2791 assert(Offset >= 0);
2793 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
2796 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
2798 // Index into the types. If we fail, set OrigBase to null.
2800 // Indexing into tail padding between struct/array elements.
2801 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
2804 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
2805 const StructLayout *SL = TD->getStructLayout(STy);
2806 assert(Offset < (int64_t)SL->getSizeInBytes() &&
2807 "Offset must stay within the indexed type");
2809 unsigned Elt = SL->getElementContainingOffset(Offset);
2810 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
2813 Offset -= SL->getElementOffset(Elt);
2814 Ty = STy->getElementType(Elt);
2815 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
2816 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
2817 assert(EltSize && "Cannot index into a zero-sized array");
2818 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
2820 Ty = AT->getElementType();
2822 // Otherwise, we can't index into the middle of this atomic type, bail.
2832 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
2833 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
2835 if (Value *V = SimplifyGEPInst(&Ops[0], Ops.size(), TD))
2836 return ReplaceInstUsesWith(GEP, V);
2838 Value *PtrOp = GEP.getOperand(0);
2840 if (isa<UndefValue>(GEP.getOperand(0)))
2841 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
2843 // Eliminate unneeded casts for indices.
2845 bool MadeChange = false;
2846 unsigned PtrSize = TD->getPointerSizeInBits();
2848 gep_type_iterator GTI = gep_type_begin(GEP);
2849 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
2850 I != E; ++I, ++GTI) {
2851 if (!isa<SequentialType>(*GTI)) continue;
2853 // If we are using a wider index than needed for this platform, shrink it
2854 // to what we need. If narrower, sign-extend it to what we need. This
2855 // explicit cast can make subsequent optimizations more obvious.
2856 unsigned OpBits = cast<IntegerType>((*I)->getType())->getBitWidth();
2857 if (OpBits == PtrSize)
2860 *I = Builder->CreateIntCast(*I, TD->getIntPtrType(GEP.getContext()),true);
2863 if (MadeChange) return &GEP;
2866 // Combine Indices - If the source pointer to this getelementptr instruction
2867 // is a getelementptr instruction, combine the indices of the two
2868 // getelementptr instructions into a single instruction.
2870 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
2871 // Note that if our source is a gep chain itself that we wait for that
2872 // chain to be resolved before we perform this transformation. This
2873 // avoids us creating a TON of code in some cases.
2875 if (GetElementPtrInst *SrcGEP =
2876 dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
2877 if (SrcGEP->getNumOperands() == 2)
2878 return 0; // Wait until our source is folded to completion.
2880 SmallVector<Value*, 8> Indices;
2882 // Find out whether the last index in the source GEP is a sequential idx.
2883 bool EndsWithSequential = false;
2884 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
2886 EndsWithSequential = !isa<StructType>(*I);
2888 // Can we combine the two pointer arithmetics offsets?
2889 if (EndsWithSequential) {
2890 // Replace: gep (gep %P, long B), long A, ...
2891 // With: T = long A+B; gep %P, T, ...
2894 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
2895 Value *GO1 = GEP.getOperand(1);
2896 if (SO1 == Constant::getNullValue(SO1->getType())) {
2898 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
2901 // If they aren't the same type, then the input hasn't been processed
2902 // by the loop above yet (which canonicalizes sequential index types to
2903 // intptr_t). Just avoid transforming this until the input has been
2905 if (SO1->getType() != GO1->getType())
2907 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
2910 // Update the GEP in place if possible.
2911 if (Src->getNumOperands() == 2) {
2912 GEP.setOperand(0, Src->getOperand(0));
2913 GEP.setOperand(1, Sum);
2916 Indices.append(Src->op_begin()+1, Src->op_end()-1);
2917 Indices.push_back(Sum);
2918 Indices.append(GEP.op_begin()+2, GEP.op_end());
2919 } else if (isa<Constant>(*GEP.idx_begin()) &&
2920 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
2921 Src->getNumOperands() != 1) {
2922 // Otherwise we can do the fold if the first index of the GEP is a zero
2923 Indices.append(Src->op_begin()+1, Src->op_end());
2924 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
2927 if (!Indices.empty())
2928 return (cast<GEPOperator>(&GEP)->isInBounds() &&
2929 Src->isInBounds()) ?
2930 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(),
2931 Indices.end(), GEP.getName()) :
2932 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
2933 Indices.end(), GEP.getName());
2936 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
2937 if (Value *X = getBitCastOperand(PtrOp)) {
2938 assert(isa<PointerType>(X->getType()) && "Must be cast from pointer");
2940 // If the input bitcast is actually "bitcast(bitcast(x))", then we don't
2941 // want to change the gep until the bitcasts are eliminated.
2942 if (getBitCastOperand(X)) {
2943 Worklist.AddValue(PtrOp);
2947 bool HasZeroPointerIndex = false;
2948 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
2949 HasZeroPointerIndex = C->isZero();
2951 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
2952 // into : GEP [10 x i8]* X, i32 0, ...
2954 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
2955 // into : GEP i8* X, ...
2957 // This occurs when the program declares an array extern like "int X[];"
2958 if (HasZeroPointerIndex) {
2959 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
2960 const PointerType *XTy = cast<PointerType>(X->getType());
2961 if (const ArrayType *CATy =
2962 dyn_cast<ArrayType>(CPTy->getElementType())) {
2963 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
2964 if (CATy->getElementType() == XTy->getElementType()) {
2965 // -> GEP i8* X, ...
2966 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
2967 return cast<GEPOperator>(&GEP)->isInBounds() ?
2968 GetElementPtrInst::CreateInBounds(X, Indices.begin(), Indices.end(),
2970 GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
2974 if (const ArrayType *XATy = dyn_cast<ArrayType>(XTy->getElementType())){
2975 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
2976 if (CATy->getElementType() == XATy->getElementType()) {
2977 // -> GEP [10 x i8]* X, i32 0, ...
2978 // At this point, we know that the cast source type is a pointer
2979 // to an array of the same type as the destination pointer
2980 // array. Because the array type is never stepped over (there
2981 // is a leading zero) we can fold the cast into this GEP.
2982 GEP.setOperand(0, X);
2987 } else if (GEP.getNumOperands() == 2) {
2988 // Transform things like:
2989 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
2990 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
2991 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
2992 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
2993 if (TD && isa<ArrayType>(SrcElTy) &&
2994 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
2995 TD->getTypeAllocSize(ResElTy)) {
2997 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
2998 Idx[1] = GEP.getOperand(1);
2999 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
3000 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
3001 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
3002 // V and GEP are both pointer types --> BitCast
3003 return new BitCastInst(NewGEP, GEP.getType());
3006 // Transform things like:
3007 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
3008 // (where tmp = 8*tmp2) into:
3009 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
3011 if (TD && isa<ArrayType>(SrcElTy) &&
3012 ResElTy == Type::getInt8Ty(GEP.getContext())) {
3013 uint64_t ArrayEltSize =
3014 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
3016 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
3017 // allow either a mul, shift, or constant here.
3019 ConstantInt *Scale = 0;
3020 if (ArrayEltSize == 1) {
3021 NewIdx = GEP.getOperand(1);
3022 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
3023 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
3024 NewIdx = ConstantInt::get(CI->getType(), 1);
3026 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
3027 if (Inst->getOpcode() == Instruction::Shl &&
3028 isa<ConstantInt>(Inst->getOperand(1))) {
3029 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
3030 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
3031 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
3033 NewIdx = Inst->getOperand(0);
3034 } else if (Inst->getOpcode() == Instruction::Mul &&
3035 isa<ConstantInt>(Inst->getOperand(1))) {
3036 Scale = cast<ConstantInt>(Inst->getOperand(1));
3037 NewIdx = Inst->getOperand(0);
3041 // If the index will be to exactly the right offset with the scale taken
3042 // out, perform the transformation. Note, we don't know whether Scale is
3043 // signed or not. We'll use unsigned version of division/modulo
3044 // operation after making sure Scale doesn't have the sign bit set.
3045 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
3046 Scale->getZExtValue() % ArrayEltSize == 0) {
3047 Scale = ConstantInt::get(Scale->getType(),
3048 Scale->getZExtValue() / ArrayEltSize);
3049 if (Scale->getZExtValue() != 1) {
3050 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
3052 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale");
3055 // Insert the new GEP instruction.
3057 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
3059 Value *NewGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
3060 Builder->CreateInBoundsGEP(X, Idx, Idx + 2, GEP.getName()) :
3061 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
3062 // The NewGEP must be pointer typed, so must the old one -> BitCast
3063 return new BitCastInst(NewGEP, GEP.getType());
3069 /// See if we can simplify:
3070 /// X = bitcast A* to B*
3071 /// Y = gep X, <...constant indices...>
3072 /// into a gep of the original struct. This is important for SROA and alias
3073 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
3074 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
3076 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
3077 // Determine how much the GEP moves the pointer. We are guaranteed to get
3078 // a constant back from EmitGEPOffset.
3079 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP));
3080 int64_t Offset = OffsetV->getSExtValue();
3082 // If this GEP instruction doesn't move the pointer, just replace the GEP
3083 // with a bitcast of the real input to the dest type.
3085 // If the bitcast is of an allocation, and the allocation will be
3086 // converted to match the type of the cast, don't touch this.
3087 if (isa<AllocaInst>(BCI->getOperand(0)) ||
3088 isMalloc(BCI->getOperand(0))) {
3089 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
3090 if (Instruction *I = visitBitCast(*BCI)) {
3093 BCI->getParent()->getInstList().insert(BCI, I);
3094 ReplaceInstUsesWith(*BCI, I);
3099 return new BitCastInst(BCI->getOperand(0), GEP.getType());
3102 // Otherwise, if the offset is non-zero, we need to find out if there is a
3103 // field at Offset in 'A's type. If so, we can pull the cast through the
3105 SmallVector<Value*, 8> NewIndices;
3107 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
3108 if (FindElementAtOffset(InTy, Offset, NewIndices)) {
3109 Value *NGEP = cast<GEPOperator>(&GEP)->isInBounds() ?
3110 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(),
3112 Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(),
3115 if (NGEP->getType() == GEP.getType())
3116 return ReplaceInstUsesWith(GEP, NGEP);
3117 NGEP->takeName(&GEP);
3118 return new BitCastInst(NGEP, GEP.getType());
3126 Instruction *InstCombiner::visitFree(Instruction &FI) {
3127 Value *Op = FI.getOperand(1);
3129 // free undef -> unreachable.
3130 if (isa<UndefValue>(Op)) {
3131 // Insert a new store to null because we cannot modify the CFG here.
3132 new StoreInst(ConstantInt::getTrue(FI.getContext()),
3133 UndefValue::get(Type::getInt1PtrTy(FI.getContext())), &FI);
3134 return EraseInstFromFunction(FI);
3137 // If we have 'free null' delete the instruction. This can happen in stl code
3138 // when lots of inlining happens.
3139 if (isa<ConstantPointerNull>(Op))
3140 return EraseInstFromFunction(FI);
3142 // If we have a malloc call whose only use is a free call, delete both.
3144 if (CallInst* CI = extractMallocCallFromBitCast(Op)) {
3145 if (Op->hasOneUse() && CI->hasOneUse()) {
3146 EraseInstFromFunction(FI);
3147 EraseInstFromFunction(*CI);
3148 return EraseInstFromFunction(*cast<Instruction>(Op));
3151 // Op is a call to malloc
3152 if (Op->hasOneUse()) {
3153 EraseInstFromFunction(FI);
3154 return EraseInstFromFunction(*cast<Instruction>(Op));
3164 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
3165 // Change br (not X), label True, label False to: br X, label False, True
3167 BasicBlock *TrueDest;
3168 BasicBlock *FalseDest;
3169 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
3170 !isa<Constant>(X)) {
3171 // Swap Destinations and condition...
3173 BI.setSuccessor(0, FalseDest);
3174 BI.setSuccessor(1, TrueDest);
3178 // Cannonicalize fcmp_one -> fcmp_oeq
3179 FCmpInst::Predicate FPred; Value *Y;
3180 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
3181 TrueDest, FalseDest)) &&
3182 BI.getCondition()->hasOneUse())
3183 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
3184 FPred == FCmpInst::FCMP_OGE) {
3185 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
3186 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
3188 // Swap Destinations and condition.
3189 BI.setSuccessor(0, FalseDest);
3190 BI.setSuccessor(1, TrueDest);
3195 // Cannonicalize icmp_ne -> icmp_eq
3196 ICmpInst::Predicate IPred;
3197 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
3198 TrueDest, FalseDest)) &&
3199 BI.getCondition()->hasOneUse())
3200 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
3201 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
3202 IPred == ICmpInst::ICMP_SGE) {
3203 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
3204 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
3205 // Swap Destinations and condition.
3206 BI.setSuccessor(0, FalseDest);
3207 BI.setSuccessor(1, TrueDest);
3215 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
3216 Value *Cond = SI.getCondition();
3217 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
3218 if (I->getOpcode() == Instruction::Add)
3219 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
3220 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
3221 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
3223 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
3225 SI.setOperand(0, I->getOperand(0));
3233 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
3234 Value *Agg = EV.getAggregateOperand();
3236 if (!EV.hasIndices())
3237 return ReplaceInstUsesWith(EV, Agg);
3239 if (Constant *C = dyn_cast<Constant>(Agg)) {
3240 if (isa<UndefValue>(C))
3241 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
3243 if (isa<ConstantAggregateZero>(C))
3244 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
3246 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
3247 // Extract the element indexed by the first index out of the constant
3248 Value *V = C->getOperand(*EV.idx_begin());
3249 if (EV.getNumIndices() > 1)
3250 // Extract the remaining indices out of the constant indexed by the
3252 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
3254 return ReplaceInstUsesWith(EV, V);
3256 return 0; // Can't handle other constants
3258 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
3259 // We're extracting from an insertvalue instruction, compare the indices
3260 const unsigned *exti, *exte, *insi, *inse;
3261 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
3262 exte = EV.idx_end(), inse = IV->idx_end();
3263 exti != exte && insi != inse;
3266 // The insert and extract both reference distinctly different elements.
3267 // This means the extract is not influenced by the insert, and we can
3268 // replace the aggregate operand of the extract with the aggregate
3269 // operand of the insert. i.e., replace
3270 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
3271 // %E = extractvalue { i32, { i32 } } %I, 0
3273 // %E = extractvalue { i32, { i32 } } %A, 0
3274 return ExtractValueInst::Create(IV->getAggregateOperand(),
3275 EV.idx_begin(), EV.idx_end());
3277 if (exti == exte && insi == inse)
3278 // Both iterators are at the end: Index lists are identical. Replace
3279 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
3280 // %C = extractvalue { i32, { i32 } } %B, 1, 0
3282 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
3284 // The extract list is a prefix of the insert list. i.e. replace
3285 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
3286 // %E = extractvalue { i32, { i32 } } %I, 1
3288 // %X = extractvalue { i32, { i32 } } %A, 1
3289 // %E = insertvalue { i32 } %X, i32 42, 0
3290 // by switching the order of the insert and extract (though the
3291 // insertvalue should be left in, since it may have other uses).
3292 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
3293 EV.idx_begin(), EV.idx_end());
3294 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
3298 // The insert list is a prefix of the extract list
3299 // We can simply remove the common indices from the extract and make it
3300 // operate on the inserted value instead of the insertvalue result.
3302 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
3303 // %E = extractvalue { i32, { i32 } } %I, 1, 0
3305 // %E extractvalue { i32 } { i32 42 }, 0
3306 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
3309 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
3310 // We're extracting from an intrinsic, see if we're the only user, which
3311 // allows us to simplify multiple result intrinsics to simpler things that
3312 // just get one value..
3313 if (II->hasOneUse()) {
3314 // Check if we're grabbing the overflow bit or the result of a 'with
3315 // overflow' intrinsic. If it's the latter we can remove the intrinsic
3316 // and replace it with a traditional binary instruction.
3317 switch (II->getIntrinsicID()) {
3318 case Intrinsic::uadd_with_overflow:
3319 case Intrinsic::sadd_with_overflow:
3320 if (*EV.idx_begin() == 0) { // Normal result.
3321 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
3322 II->replaceAllUsesWith(UndefValue::get(II->getType()));
3323 EraseInstFromFunction(*II);
3324 return BinaryOperator::CreateAdd(LHS, RHS);
3327 case Intrinsic::usub_with_overflow:
3328 case Intrinsic::ssub_with_overflow:
3329 if (*EV.idx_begin() == 0) { // Normal result.
3330 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
3331 II->replaceAllUsesWith(UndefValue::get(II->getType()));
3332 EraseInstFromFunction(*II);
3333 return BinaryOperator::CreateSub(LHS, RHS);
3336 case Intrinsic::umul_with_overflow:
3337 case Intrinsic::smul_with_overflow:
3338 if (*EV.idx_begin() == 0) { // Normal result.
3339 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2);
3340 II->replaceAllUsesWith(UndefValue::get(II->getType()));
3341 EraseInstFromFunction(*II);
3342 return BinaryOperator::CreateMul(LHS, RHS);
3350 // Can't simplify extracts from other values. Note that nested extracts are
3351 // already simplified implicitely by the above (extract ( extract (insert) )
3352 // will be translated into extract ( insert ( extract ) ) first and then just
3353 // the value inserted, if appropriate).
3360 /// TryToSinkInstruction - Try to move the specified instruction from its
3361 /// current block into the beginning of DestBlock, which can only happen if it's
3362 /// safe to move the instruction past all of the instructions between it and the
3363 /// end of its block.
3364 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
3365 assert(I->hasOneUse() && "Invariants didn't hold!");
3367 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
3368 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
3371 // Do not sink alloca instructions out of the entry block.
3372 if (isa<AllocaInst>(I) && I->getParent() ==
3373 &DestBlock->getParent()->getEntryBlock())
3376 // We can only sink load instructions if there is nothing between the load and
3377 // the end of block that could change the value.
3378 if (I->mayReadFromMemory()) {
3379 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
3381 if (Scan->mayWriteToMemory())
3385 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
3387 I->moveBefore(InsertPos);
3393 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
3394 /// all reachable code to the worklist.
3396 /// This has a couple of tricks to make the code faster and more powerful. In
3397 /// particular, we constant fold and DCE instructions as we go, to avoid adding
3398 /// them to the worklist (this significantly speeds up instcombine on code where
3399 /// many instructions are dead or constant). Additionally, if we find a branch
3400 /// whose condition is a known constant, we only visit the reachable successors.
3402 static bool AddReachableCodeToWorklist(BasicBlock *BB,
3403 SmallPtrSet<BasicBlock*, 64> &Visited,
3405 const TargetData *TD) {
3406 bool MadeIRChange = false;
3407 SmallVector<BasicBlock*, 256> Worklist;
3408 Worklist.push_back(BB);
3410 std::vector<Instruction*> InstrsForInstCombineWorklist;
3411 InstrsForInstCombineWorklist.reserve(128);
3413 SmallPtrSet<ConstantExpr*, 64> FoldedConstants;
3415 while (!Worklist.empty()) {
3416 BB = Worklist.back();
3417 Worklist.pop_back();
3419 // We have now visited this block! If we've already been here, ignore it.
3420 if (!Visited.insert(BB)) continue;
3422 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
3423 Instruction *Inst = BBI++;
3425 // DCE instruction if trivially dead.
3426 if (isInstructionTriviallyDead(Inst)) {
3428 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
3429 Inst->eraseFromParent();
3433 // ConstantProp instruction if trivially constant.
3434 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
3435 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
3436 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
3438 Inst->replaceAllUsesWith(C);
3440 Inst->eraseFromParent();
3447 // See if we can constant fold its operands.
3448 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
3450 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
3451 if (CE == 0) continue;
3453 // If we already folded this constant, don't try again.
3454 if (!FoldedConstants.insert(CE))
3457 Constant *NewC = ConstantFoldConstantExpression(CE, TD);
3458 if (NewC && NewC != CE) {
3460 MadeIRChange = true;
3466 InstrsForInstCombineWorklist.push_back(Inst);
3469 // Recursively visit successors. If this is a branch or switch on a
3470 // constant, only visit the reachable successor.
3471 TerminatorInst *TI = BB->getTerminator();
3472 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3473 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
3474 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
3475 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
3476 Worklist.push_back(ReachableBB);
3479 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3480 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
3481 // See if this is an explicit destination.
3482 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
3483 if (SI->getCaseValue(i) == Cond) {
3484 BasicBlock *ReachableBB = SI->getSuccessor(i);
3485 Worklist.push_back(ReachableBB);
3489 // Otherwise it is the default destination.
3490 Worklist.push_back(SI->getSuccessor(0));
3495 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
3496 Worklist.push_back(TI->getSuccessor(i));
3499 // Once we've found all of the instructions to add to instcombine's worklist,
3500 // add them in reverse order. This way instcombine will visit from the top
3501 // of the function down. This jives well with the way that it adds all uses
3502 // of instructions to the worklist after doing a transformation, thus avoiding
3503 // some N^2 behavior in pathological cases.
3504 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
3505 InstrsForInstCombineWorklist.size());
3507 return MadeIRChange;
3510 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
3511 MadeIRChange = false;
3513 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
3514 << F.getNameStr() << "\n");
3517 // Do a depth-first traversal of the function, populate the worklist with
3518 // the reachable instructions. Ignore blocks that are not reachable. Keep
3519 // track of which blocks we visit.
3520 SmallPtrSet<BasicBlock*, 64> Visited;
3521 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
3523 // Do a quick scan over the function. If we find any blocks that are
3524 // unreachable, remove any instructions inside of them. This prevents
3525 // the instcombine code from having to deal with some bad special cases.
3526 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
3527 if (!Visited.count(BB)) {
3528 Instruction *Term = BB->getTerminator();
3529 while (Term != BB->begin()) { // Remove instrs bottom-up
3530 BasicBlock::iterator I = Term; --I;
3532 DEBUG(errs() << "IC: DCE: " << *I << '\n');
3533 // A debug intrinsic shouldn't force another iteration if we weren't
3534 // going to do one without it.
3535 if (!isa<DbgInfoIntrinsic>(I)) {
3537 MadeIRChange = true;
3540 // If I is not void type then replaceAllUsesWith undef.
3541 // This allows ValueHandlers and custom metadata to adjust itself.
3542 if (!I->getType()->isVoidTy())
3543 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3544 I->eraseFromParent();
3549 while (!Worklist.isEmpty()) {
3550 Instruction *I = Worklist.RemoveOne();
3551 if (I == 0) continue; // skip null values.
3553 // Check to see if we can DCE the instruction.
3554 if (isInstructionTriviallyDead(I)) {
3555 DEBUG(errs() << "IC: DCE: " << *I << '\n');
3556 EraseInstFromFunction(*I);
3558 MadeIRChange = true;
3562 // Instruction isn't dead, see if we can constant propagate it.
3563 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
3564 if (Constant *C = ConstantFoldInstruction(I, TD)) {
3565 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
3567 // Add operands to the worklist.
3568 ReplaceInstUsesWith(*I, C);
3570 EraseInstFromFunction(*I);
3571 MadeIRChange = true;
3575 // See if we can trivially sink this instruction to a successor basic block.
3576 if (I->hasOneUse()) {
3577 BasicBlock *BB = I->getParent();
3578 Instruction *UserInst = cast<Instruction>(I->use_back());
3579 BasicBlock *UserParent;
3581 // Get the block the use occurs in.
3582 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
3583 UserParent = PN->getIncomingBlock(I->use_begin().getUse());
3585 UserParent = UserInst->getParent();
3587 if (UserParent != BB) {
3588 bool UserIsSuccessor = false;
3589 // See if the user is one of our successors.
3590 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
3591 if (*SI == UserParent) {
3592 UserIsSuccessor = true;
3596 // If the user is one of our immediate successors, and if that successor
3597 // only has us as a predecessors (we'd have to split the critical edge
3598 // otherwise), we can keep going.
3599 if (UserIsSuccessor && UserParent->getSinglePredecessor())
3600 // Okay, the CFG is simple enough, try to sink this instruction.
3601 MadeIRChange |= TryToSinkInstruction(I, UserParent);
3605 // Now that we have an instruction, try combining it to simplify it.
3606 Builder->SetInsertPoint(I->getParent(), I);
3611 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
3612 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n');
3614 if (Instruction *Result = visit(*I)) {
3616 // Should we replace the old instruction with a new one?
3618 DEBUG(errs() << "IC: Old = " << *I << '\n'
3619 << " New = " << *Result << '\n');
3621 // Everything uses the new instruction now.
3622 I->replaceAllUsesWith(Result);
3624 // Push the new instruction and any users onto the worklist.
3625 Worklist.Add(Result);
3626 Worklist.AddUsersToWorkList(*Result);
3628 // Move the name to the new instruction first.
3629 Result->takeName(I);
3631 // Insert the new instruction into the basic block...
3632 BasicBlock *InstParent = I->getParent();
3633 BasicBlock::iterator InsertPos = I;
3635 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
3636 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
3639 InstParent->getInstList().insert(InsertPos, Result);
3641 EraseInstFromFunction(*I);
3644 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
3645 << " New = " << *I << '\n');
3648 // If the instruction was modified, it's possible that it is now dead.
3649 // if so, remove it.
3650 if (isInstructionTriviallyDead(I)) {
3651 EraseInstFromFunction(*I);
3654 Worklist.AddUsersToWorkList(*I);
3657 MadeIRChange = true;
3662 return MadeIRChange;
3666 bool InstCombiner::runOnFunction(Function &F) {
3667 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
3668 TD = getAnalysisIfAvailable<TargetData>();
3671 /// Builder - This is an IRBuilder that automatically inserts new
3672 /// instructions into the worklist when they are created.
3673 IRBuilder<true, TargetFolder, InstCombineIRInserter>
3674 TheBuilder(F.getContext(), TargetFolder(TD),
3675 InstCombineIRInserter(Worklist));
3676 Builder = &TheBuilder;
3678 bool EverMadeChange = false;
3680 // Iterate while there is work to do.
3681 unsigned Iteration = 0;
3682 while (DoOneIteration(F, Iteration++))
3683 EverMadeChange = true;
3686 return EverMadeChange;
3689 FunctionPass *llvm::createInstructionCombiningPass() {
3690 return new InstCombiner();