1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
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 // Peephole optimize the CFG.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Utils/Local.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/ConstantFolding.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/CFG.h"
26 #include "llvm/IR/ConstantRange.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/GlobalVariable.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Metadata.h"
37 #include "llvm/IR/Module.h"
38 #include "llvm/IR/NoFolder.h"
39 #include "llvm/IR/Operator.h"
40 #include "llvm/IR/PatternMatch.h"
41 #include "llvm/IR/Type.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/ValueMapper.h"
52 using namespace PatternMatch
;
54 #define DEBUG_TYPE "simplifycfg"
56 static cl::opt
<unsigned>
57 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden
, cl::init(1),
58 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
61 DupRet("simplifycfg-dup-ret", cl::Hidden
, cl::init(false),
62 cl::desc("Duplicate return instructions into unconditional branches"));
65 SinkCommon("simplifycfg-sink-common", cl::Hidden
, cl::init(true),
66 cl::desc("Sink common instructions down to the end block"));
68 static cl::opt
<bool> HoistCondStores(
69 "simplifycfg-hoist-cond-stores", cl::Hidden
, cl::init(true),
70 cl::desc("Hoist conditional stores if an unconditional store precedes"));
72 STATISTIC(NumBitMaps
, "Number of switch instructions turned into bitmaps");
73 STATISTIC(NumLinearMaps
, "Number of switch instructions turned into linear mapping");
74 STATISTIC(NumLookupTables
, "Number of switch instructions turned into lookup tables");
75 STATISTIC(NumLookupTablesHoles
, "Number of switch instructions turned into lookup tables (holes checked)");
76 STATISTIC(NumTableCmpReuses
, "Number of reused switch table lookup compares");
77 STATISTIC(NumSinkCommons
, "Number of common instructions sunk down to the end block");
78 STATISTIC(NumSpeculations
, "Number of speculative executed instructions");
81 // The first field contains the value that the switch produces when a certain
82 // case group is selected, and the second field is a vector containing the cases
83 // composing the case group.
84 typedef SmallVector
<std::pair
<Constant
*, SmallVector
<ConstantInt
*, 4>>, 2>
85 SwitchCaseResultVectorTy
;
86 // The first field contains the phi node that generates a result of the switch
87 // and the second field contains the value generated for a certain case in the switch
89 typedef SmallVector
<std::pair
<PHINode
*, Constant
*>, 4> SwitchCaseResultsTy
;
91 /// ValueEqualityComparisonCase - Represents a case of a switch.
92 struct ValueEqualityComparisonCase
{
96 ValueEqualityComparisonCase(ConstantInt
*Value
, BasicBlock
*Dest
)
97 : Value(Value
), Dest(Dest
) {}
99 bool operator<(ValueEqualityComparisonCase RHS
) const {
100 // Comparing pointers is ok as we only rely on the order for uniquing.
101 return Value
< RHS
.Value
;
104 bool operator==(BasicBlock
*RHSDest
) const { return Dest
== RHSDest
; }
107 class SimplifyCFGOpt
{
108 const TargetTransformInfo
&TTI
;
109 unsigned BonusInstThreshold
;
110 const DataLayout
*const DL
;
112 Value
*isValueEqualityComparison(TerminatorInst
*TI
);
113 BasicBlock
*GetValueEqualityComparisonCases(TerminatorInst
*TI
,
114 std::vector
<ValueEqualityComparisonCase
> &Cases
);
115 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst
*TI
,
117 IRBuilder
<> &Builder
);
118 bool FoldValueComparisonIntoPredecessors(TerminatorInst
*TI
,
119 IRBuilder
<> &Builder
);
121 bool SimplifyReturn(ReturnInst
*RI
, IRBuilder
<> &Builder
);
122 bool SimplifyResume(ResumeInst
*RI
, IRBuilder
<> &Builder
);
123 bool SimplifyUnreachable(UnreachableInst
*UI
);
124 bool SimplifySwitch(SwitchInst
*SI
, IRBuilder
<> &Builder
);
125 bool SimplifyIndirectBr(IndirectBrInst
*IBI
);
126 bool SimplifyUncondBranch(BranchInst
*BI
, IRBuilder
<> &Builder
);
127 bool SimplifyCondBranch(BranchInst
*BI
, IRBuilder
<>&Builder
);
130 SimplifyCFGOpt(const TargetTransformInfo
&TTI
, unsigned BonusInstThreshold
,
131 const DataLayout
*DL
, AssumptionCache
*AC
)
132 : TTI(TTI
), BonusInstThreshold(BonusInstThreshold
), DL(DL
), AC(AC
) {}
133 bool run(BasicBlock
*BB
);
137 /// SafeToMergeTerminators - Return true if it is safe to merge these two
138 /// terminator instructions together.
140 static bool SafeToMergeTerminators(TerminatorInst
*SI1
, TerminatorInst
*SI2
) {
141 if (SI1
== SI2
) return false; // Can't merge with self!
143 // It is not safe to merge these two switch instructions if they have a common
144 // successor, and if that successor has a PHI node, and if *that* PHI node has
145 // conflicting incoming values from the two switch blocks.
146 BasicBlock
*SI1BB
= SI1
->getParent();
147 BasicBlock
*SI2BB
= SI2
->getParent();
148 SmallPtrSet
<BasicBlock
*, 16> SI1Succs(succ_begin(SI1BB
), succ_end(SI1BB
));
150 for (succ_iterator I
= succ_begin(SI2BB
), E
= succ_end(SI2BB
); I
!= E
; ++I
)
151 if (SI1Succs
.count(*I
))
152 for (BasicBlock::iterator BBI
= (*I
)->begin();
153 isa
<PHINode
>(BBI
); ++BBI
) {
154 PHINode
*PN
= cast
<PHINode
>(BBI
);
155 if (PN
->getIncomingValueForBlock(SI1BB
) !=
156 PN
->getIncomingValueForBlock(SI2BB
))
163 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
164 /// to merge these two terminator instructions together, where SI1 is an
165 /// unconditional branch. PhiNodes will store all PHI nodes in common
168 static bool isProfitableToFoldUnconditional(BranchInst
*SI1
,
171 SmallVectorImpl
<PHINode
*> &PhiNodes
) {
172 if (SI1
== SI2
) return false; // Can't merge with self!
173 assert(SI1
->isUnconditional() && SI2
->isConditional());
175 // We fold the unconditional branch if we can easily update all PHI nodes in
176 // common successors:
177 // 1> We have a constant incoming value for the conditional branch;
178 // 2> We have "Cond" as the incoming value for the unconditional branch;
179 // 3> SI2->getCondition() and Cond have same operands.
180 CmpInst
*Ci2
= dyn_cast
<CmpInst
>(SI2
->getCondition());
181 if (!Ci2
) return false;
182 if (!(Cond
->getOperand(0) == Ci2
->getOperand(0) &&
183 Cond
->getOperand(1) == Ci2
->getOperand(1)) &&
184 !(Cond
->getOperand(0) == Ci2
->getOperand(1) &&
185 Cond
->getOperand(1) == Ci2
->getOperand(0)))
188 BasicBlock
*SI1BB
= SI1
->getParent();
189 BasicBlock
*SI2BB
= SI2
->getParent();
190 SmallPtrSet
<BasicBlock
*, 16> SI1Succs(succ_begin(SI1BB
), succ_end(SI1BB
));
191 for (succ_iterator I
= succ_begin(SI2BB
), E
= succ_end(SI2BB
); I
!= E
; ++I
)
192 if (SI1Succs
.count(*I
))
193 for (BasicBlock::iterator BBI
= (*I
)->begin();
194 isa
<PHINode
>(BBI
); ++BBI
) {
195 PHINode
*PN
= cast
<PHINode
>(BBI
);
196 if (PN
->getIncomingValueForBlock(SI1BB
) != Cond
||
197 !isa
<ConstantInt
>(PN
->getIncomingValueForBlock(SI2BB
)))
199 PhiNodes
.push_back(PN
);
204 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
205 /// now be entries in it from the 'NewPred' block. The values that will be
206 /// flowing into the PHI nodes will be the same as those coming in from
207 /// ExistPred, an existing predecessor of Succ.
208 static void AddPredecessorToBlock(BasicBlock
*Succ
, BasicBlock
*NewPred
,
209 BasicBlock
*ExistPred
) {
210 if (!isa
<PHINode
>(Succ
->begin())) return; // Quick exit if nothing to do
213 for (BasicBlock::iterator I
= Succ
->begin();
214 (PN
= dyn_cast
<PHINode
>(I
)); ++I
)
215 PN
->addIncoming(PN
->getIncomingValueForBlock(ExistPred
), NewPred
);
218 /// ComputeSpeculationCost - Compute an abstract "cost" of speculating the
219 /// given instruction, which is assumed to be safe to speculate. 1 means
220 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
221 static unsigned ComputeSpeculationCost(const User
*I
, const DataLayout
*DL
) {
222 assert(isSafeToSpeculativelyExecute(I
, DL
) &&
223 "Instruction is not safe to speculatively execute!");
224 switch (Operator::getOpcode(I
)) {
226 // In doubt, be conservative.
228 case Instruction::GetElementPtr
:
229 // GEPs are cheap if all indices are constant.
230 if (!cast
<GEPOperator
>(I
)->hasAllConstantIndices())
233 case Instruction::ExtractValue
:
234 case Instruction::Load
:
235 case Instruction::Add
:
236 case Instruction::Sub
:
237 case Instruction::And
:
238 case Instruction::Or
:
239 case Instruction::Xor
:
240 case Instruction::Shl
:
241 case Instruction::LShr
:
242 case Instruction::AShr
:
243 case Instruction::ICmp
:
244 case Instruction::Trunc
:
245 case Instruction::ZExt
:
246 case Instruction::SExt
:
247 case Instruction::BitCast
:
248 case Instruction::ExtractElement
:
249 case Instruction::InsertElement
:
250 return 1; // These are all cheap.
252 case Instruction::Call
:
253 case Instruction::Select
:
258 /// DominatesMergePoint - If we have a merge point of an "if condition" as
259 /// accepted above, return true if the specified value dominates the block. We
260 /// don't handle the true generality of domination here, just a special case
261 /// which works well enough for us.
263 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
264 /// see if V (which must be an instruction) and its recursive operands
265 /// that do not dominate BB have a combined cost lower than CostRemaining and
266 /// are non-trapping. If both are true, the instruction is inserted into the
267 /// set and true is returned.
269 /// The cost for most non-trapping instructions is defined as 1 except for
270 /// Select whose cost is 2.
272 /// After this function returns, CostRemaining is decreased by the cost of
273 /// V plus its non-dominating operands. If that cost is greater than
274 /// CostRemaining, false is returned and CostRemaining is undefined.
275 static bool DominatesMergePoint(Value
*V
, BasicBlock
*BB
,
276 SmallPtrSetImpl
<Instruction
*> *AggressiveInsts
,
277 unsigned &CostRemaining
,
278 const DataLayout
*DL
) {
279 Instruction
*I
= dyn_cast
<Instruction
>(V
);
281 // Non-instructions all dominate instructions, but not all constantexprs
282 // can be executed unconditionally.
283 if (ConstantExpr
*C
= dyn_cast
<ConstantExpr
>(V
))
288 BasicBlock
*PBB
= I
->getParent();
290 // We don't want to allow weird loops that might have the "if condition" in
291 // the bottom of this block.
292 if (PBB
== BB
) return false;
294 // If this instruction is defined in a block that contains an unconditional
295 // branch to BB, then it must be in the 'conditional' part of the "if
296 // statement". If not, it definitely dominates the region.
297 BranchInst
*BI
= dyn_cast
<BranchInst
>(PBB
->getTerminator());
298 if (!BI
|| BI
->isConditional() || BI
->getSuccessor(0) != BB
)
301 // If we aren't allowing aggressive promotion anymore, then don't consider
302 // instructions in the 'if region'.
303 if (!AggressiveInsts
) return false;
305 // If we have seen this instruction before, don't count it again.
306 if (AggressiveInsts
->count(I
)) return true;
308 // Okay, it looks like the instruction IS in the "condition". Check to
309 // see if it's a cheap instruction to unconditionally compute, and if it
310 // only uses stuff defined outside of the condition. If so, hoist it out.
311 if (!isSafeToSpeculativelyExecute(I
, DL
))
314 unsigned Cost
= ComputeSpeculationCost(I
, DL
);
316 if (Cost
> CostRemaining
)
319 CostRemaining
-= Cost
;
321 // Okay, we can only really hoist these out if their operands do
322 // not take us over the cost threshold.
323 for (User::op_iterator i
= I
->op_begin(), e
= I
->op_end(); i
!= e
; ++i
)
324 if (!DominatesMergePoint(*i
, BB
, AggressiveInsts
, CostRemaining
, DL
))
326 // Okay, it's safe to do this! Remember this instruction.
327 AggressiveInsts
->insert(I
);
331 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
332 /// and PointerNullValue. Return NULL if value is not a constant int.
333 static ConstantInt
*GetConstantInt(Value
*V
, const DataLayout
*DL
) {
334 // Normal constant int.
335 ConstantInt
*CI
= dyn_cast
<ConstantInt
>(V
);
336 if (CI
|| !DL
|| !isa
<Constant
>(V
) || !V
->getType()->isPointerTy())
339 // This is some kind of pointer constant. Turn it into a pointer-sized
340 // ConstantInt if possible.
341 IntegerType
*PtrTy
= cast
<IntegerType
>(DL
->getIntPtrType(V
->getType()));
343 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
344 if (isa
<ConstantPointerNull
>(V
))
345 return ConstantInt::get(PtrTy
, 0);
347 // IntToPtr const int.
348 if (ConstantExpr
*CE
= dyn_cast
<ConstantExpr
>(V
))
349 if (CE
->getOpcode() == Instruction::IntToPtr
)
350 if (ConstantInt
*CI
= dyn_cast
<ConstantInt
>(CE
->getOperand(0))) {
351 // The constant is very likely to have the right type already.
352 if (CI
->getType() == PtrTy
)
355 return cast
<ConstantInt
>
356 (ConstantExpr::getIntegerCast(CI
, PtrTy
, /*isSigned=*/false));
363 /// Given a chain of or (||) or and (&&) comparison of a value against a
364 /// constant, this will try to recover the information required for a switch
366 /// It will depth-first traverse the chain of comparison, seeking for patterns
367 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
368 /// representing the different cases for the switch.
369 /// Note that if the chain is composed of '||' it will build the set of elements
370 /// that matches the comparisons (i.e. any of this value validate the chain)
371 /// while for a chain of '&&' it will build the set elements that make the test
373 struct ConstantComparesGatherer
{
375 Value
*CompValue
; /// Value found for the switch comparison
376 Value
*Extra
; /// Extra clause to be checked before the switch
377 SmallVector
<ConstantInt
*, 8> Vals
; /// Set of integers to match in switch
378 unsigned UsedICmps
; /// Number of comparisons matched in the and/or chain
380 /// Construct and compute the result for the comparison instruction Cond
381 ConstantComparesGatherer(Instruction
*Cond
, const DataLayout
*DL
)
382 : CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
387 ConstantComparesGatherer(const ConstantComparesGatherer
&)
388 LLVM_DELETED_FUNCTION
;
389 ConstantComparesGatherer
&
390 operator=(const ConstantComparesGatherer
&) LLVM_DELETED_FUNCTION
;
394 /// Try to set the current value used for the comparison, it succeeds only if
395 /// it wasn't set before or if the new value is the same as the old one
396 bool setValueOnce(Value
*NewVal
) {
397 if(CompValue
&& CompValue
!= NewVal
) return false;
399 return (CompValue
!= nullptr);
402 /// Try to match Instruction "I" as a comparison against a constant and
403 /// populates the array Vals with the set of values that match (or do not
404 /// match depending on isEQ).
405 /// Return false on failure. On success, the Value the comparison matched
406 /// against is placed in CompValue.
407 /// If CompValue is already set, the function is expected to fail if a match
408 /// is found but the value compared to is different.
409 bool matchInstruction(Instruction
*I
, const DataLayout
*DL
, bool isEQ
) {
410 // If this is an icmp against a constant, handle this as one of the cases.
413 if (!((ICI
= dyn_cast
<ICmpInst
>(I
)) &&
414 (C
= GetConstantInt(I
->getOperand(1), DL
)))) {
421 // Pattern match a special case
422 // (x & ~2^x) == y --> x == y || x == y|2^x
423 // This undoes a transformation done by instcombine to fuse 2 compares.
424 if (ICI
->getPredicate() == (isEQ
? ICmpInst::ICMP_EQ
:ICmpInst::ICMP_NE
)) {
425 if (match(ICI
->getOperand(0),
426 m_And(m_Value(RHSVal
), m_ConstantInt(RHSC
)))) {
427 APInt Not
= ~RHSC
->getValue();
428 if (Not
.isPowerOf2()) {
429 // If we already have a value for the switch, it has to match!
430 if(!setValueOnce(RHSVal
))
434 Vals
.push_back(ConstantInt::get(C
->getContext(),
435 C
->getValue() | Not
));
441 // If we already have a value for the switch, it has to match!
442 if(!setValueOnce(ICI
->getOperand(0)))
447 return ICI
->getOperand(0);
450 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
451 ConstantRange Span
= ConstantRange::makeICmpRegion(ICI
->getPredicate(),
454 // Shift the range if the compare is fed by an add. This is the range
455 // compare idiom as emitted by instcombine.
456 Value
*CandidateVal
= I
->getOperand(0);
457 if(match(I
->getOperand(0), m_Add(m_Value(RHSVal
), m_ConstantInt(RHSC
)))) {
458 Span
= Span
.subtract(RHSC
->getValue());
459 CandidateVal
= RHSVal
;
462 // If this is an and/!= check, then we are looking to build the set of
463 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
466 Span
= Span
.inverse();
468 // If there are a ton of values, we don't want to make a ginormous switch.
469 if (Span
.getSetSize().ugt(8) || Span
.isEmptySet()) {
473 // If we already have a value for the switch, it has to match!
474 if(!setValueOnce(CandidateVal
))
477 // Add all values from the range to the set
478 for (APInt Tmp
= Span
.getLower(); Tmp
!= Span
.getUpper(); ++Tmp
)
479 Vals
.push_back(ConstantInt::get(I
->getContext(), Tmp
));
486 /// gather - Given a potentially 'or'd or 'and'd together collection of icmp
487 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
488 /// the value being compared, and stick the list constants into the Vals
490 /// One "Extra" case is allowed to differ from the other.
491 void gather(Value
*V
, const DataLayout
*DL
) {
492 Instruction
*I
= dyn_cast
<Instruction
>(V
);
493 bool isEQ
= (I
->getOpcode() == Instruction::Or
);
495 // Keep a stack (SmallVector for efficiency) for depth-first traversal
496 SmallVector
<Value
*, 8> DFT
;
501 while(!DFT
.empty()) {
502 V
= DFT
.pop_back_val();
504 if (Instruction
*I
= dyn_cast
<Instruction
>(V
)) {
505 // If it is a || (or && depending on isEQ), process the operands.
506 if (I
->getOpcode() == (isEQ
? Instruction::Or
: Instruction::And
)) {
507 DFT
.push_back(I
->getOperand(1));
508 DFT
.push_back(I
->getOperand(0));
512 // Try to match the current instruction
513 if (matchInstruction(I
, DL
, isEQ
))
514 // Match succeed, continue the loop
518 // One element of the sequence of || (or &&) could not be match as a
519 // comparison against the same value as the others.
520 // We allow only one "Extra" case to be checked before the switch
525 // Failed to parse a proper sequence, abort now
534 static void EraseTerminatorInstAndDCECond(TerminatorInst
*TI
) {
535 Instruction
*Cond
= nullptr;
536 if (SwitchInst
*SI
= dyn_cast
<SwitchInst
>(TI
)) {
537 Cond
= dyn_cast
<Instruction
>(SI
->getCondition());
538 } else if (BranchInst
*BI
= dyn_cast
<BranchInst
>(TI
)) {
539 if (BI
->isConditional())
540 Cond
= dyn_cast
<Instruction
>(BI
->getCondition());
541 } else if (IndirectBrInst
*IBI
= dyn_cast
<IndirectBrInst
>(TI
)) {
542 Cond
= dyn_cast
<Instruction
>(IBI
->getAddress());
545 TI
->eraseFromParent();
546 if (Cond
) RecursivelyDeleteTriviallyDeadInstructions(Cond
);
549 /// isValueEqualityComparison - Return true if the specified terminator checks
550 /// to see if a value is equal to constant integer value.
551 Value
*SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst
*TI
) {
553 if (SwitchInst
*SI
= dyn_cast
<SwitchInst
>(TI
)) {
554 // Do not permit merging of large switch instructions into their
555 // predecessors unless there is only one predecessor.
556 if (SI
->getNumSuccessors()*std::distance(pred_begin(SI
->getParent()),
557 pred_end(SI
->getParent())) <= 128)
558 CV
= SI
->getCondition();
559 } else if (BranchInst
*BI
= dyn_cast
<BranchInst
>(TI
))
560 if (BI
->isConditional() && BI
->getCondition()->hasOneUse())
561 if (ICmpInst
*ICI
= dyn_cast
<ICmpInst
>(BI
->getCondition()))
562 if (ICI
->isEquality() && GetConstantInt(ICI
->getOperand(1), DL
))
563 CV
= ICI
->getOperand(0);
565 // Unwrap any lossless ptrtoint cast.
567 if (PtrToIntInst
*PTII
= dyn_cast
<PtrToIntInst
>(CV
)) {
568 Value
*Ptr
= PTII
->getPointerOperand();
569 if (PTII
->getType() == DL
->getIntPtrType(Ptr
->getType()))
576 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
577 /// decode all of the 'cases' that it represents and return the 'default' block.
578 BasicBlock
*SimplifyCFGOpt::
579 GetValueEqualityComparisonCases(TerminatorInst
*TI
,
580 std::vector
<ValueEqualityComparisonCase
>
582 if (SwitchInst
*SI
= dyn_cast
<SwitchInst
>(TI
)) {
583 Cases
.reserve(SI
->getNumCases());
584 for (SwitchInst::CaseIt i
= SI
->case_begin(), e
= SI
->case_end(); i
!= e
; ++i
)
585 Cases
.push_back(ValueEqualityComparisonCase(i
.getCaseValue(),
586 i
.getCaseSuccessor()));
587 return SI
->getDefaultDest();
590 BranchInst
*BI
= cast
<BranchInst
>(TI
);
591 ICmpInst
*ICI
= cast
<ICmpInst
>(BI
->getCondition());
592 BasicBlock
*Succ
= BI
->getSuccessor(ICI
->getPredicate() == ICmpInst::ICMP_NE
);
593 Cases
.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI
->getOperand(1),
596 return BI
->getSuccessor(ICI
->getPredicate() == ICmpInst::ICMP_EQ
);
600 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
601 /// in the list that match the specified block.
602 static void EliminateBlockCases(BasicBlock
*BB
,
603 std::vector
<ValueEqualityComparisonCase
> &Cases
) {
604 Cases
.erase(std::remove(Cases
.begin(), Cases
.end(), BB
), Cases
.end());
607 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
610 ValuesOverlap(std::vector
<ValueEqualityComparisonCase
> &C1
,
611 std::vector
<ValueEqualityComparisonCase
> &C2
) {
612 std::vector
<ValueEqualityComparisonCase
> *V1
= &C1
, *V2
= &C2
;
614 // Make V1 be smaller than V2.
615 if (V1
->size() > V2
->size())
618 if (V1
->size() == 0) return false;
619 if (V1
->size() == 1) {
621 ConstantInt
*TheVal
= (*V1
)[0].Value
;
622 for (unsigned i
= 0, e
= V2
->size(); i
!= e
; ++i
)
623 if (TheVal
== (*V2
)[i
].Value
)
627 // Otherwise, just sort both lists and compare element by element.
628 array_pod_sort(V1
->begin(), V1
->end());
629 array_pod_sort(V2
->begin(), V2
->end());
630 unsigned i1
= 0, i2
= 0, e1
= V1
->size(), e2
= V2
->size();
631 while (i1
!= e1
&& i2
!= e2
) {
632 if ((*V1
)[i1
].Value
== (*V2
)[i2
].Value
)
634 if ((*V1
)[i1
].Value
< (*V2
)[i2
].Value
)
642 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
643 /// terminator instruction and its block is known to only have a single
644 /// predecessor block, check to see if that predecessor is also a value
645 /// comparison with the same value, and if that comparison determines the
646 /// outcome of this comparison. If so, simplify TI. This does a very limited
647 /// form of jump threading.
648 bool SimplifyCFGOpt::
649 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst
*TI
,
651 IRBuilder
<> &Builder
) {
652 Value
*PredVal
= isValueEqualityComparison(Pred
->getTerminator());
653 if (!PredVal
) return false; // Not a value comparison in predecessor.
655 Value
*ThisVal
= isValueEqualityComparison(TI
);
656 assert(ThisVal
&& "This isn't a value comparison!!");
657 if (ThisVal
!= PredVal
) return false; // Different predicates.
659 // TODO: Preserve branch weight metadata, similarly to how
660 // FoldValueComparisonIntoPredecessors preserves it.
662 // Find out information about when control will move from Pred to TI's block.
663 std::vector
<ValueEqualityComparisonCase
> PredCases
;
664 BasicBlock
*PredDef
= GetValueEqualityComparisonCases(Pred
->getTerminator(),
666 EliminateBlockCases(PredDef
, PredCases
); // Remove default from cases.
668 // Find information about how control leaves this block.
669 std::vector
<ValueEqualityComparisonCase
> ThisCases
;
670 BasicBlock
*ThisDef
= GetValueEqualityComparisonCases(TI
, ThisCases
);
671 EliminateBlockCases(ThisDef
, ThisCases
); // Remove default from cases.
673 // If TI's block is the default block from Pred's comparison, potentially
674 // simplify TI based on this knowledge.
675 if (PredDef
== TI
->getParent()) {
676 // If we are here, we know that the value is none of those cases listed in
677 // PredCases. If there are any cases in ThisCases that are in PredCases, we
679 if (!ValuesOverlap(PredCases
, ThisCases
))
682 if (isa
<BranchInst
>(TI
)) {
683 // Okay, one of the successors of this condbr is dead. Convert it to a
685 assert(ThisCases
.size() == 1 && "Branch can only have one case!");
686 // Insert the new branch.
687 Instruction
*NI
= Builder
.CreateBr(ThisDef
);
690 // Remove PHI node entries for the dead edge.
691 ThisCases
[0].Dest
->removePredecessor(TI
->getParent());
693 DEBUG(dbgs() << "Threading pred instr: " << *Pred
->getTerminator()
694 << "Through successor TI: " << *TI
<< "Leaving: " << *NI
<< "\n");
696 EraseTerminatorInstAndDCECond(TI
);
700 SwitchInst
*SI
= cast
<SwitchInst
>(TI
);
701 // Okay, TI has cases that are statically dead, prune them away.
702 SmallPtrSet
<Constant
*, 16> DeadCases
;
703 for (unsigned i
= 0, e
= PredCases
.size(); i
!= e
; ++i
)
704 DeadCases
.insert(PredCases
[i
].Value
);
706 DEBUG(dbgs() << "Threading pred instr: " << *Pred
->getTerminator()
707 << "Through successor TI: " << *TI
);
709 // Collect branch weights into a vector.
710 SmallVector
<uint32_t, 8> Weights
;
711 MDNode
*MD
= SI
->getMetadata(LLVMContext::MD_prof
);
712 bool HasWeight
= MD
&& (MD
->getNumOperands() == 2 + SI
->getNumCases());
714 for (unsigned MD_i
= 1, MD_e
= MD
->getNumOperands(); MD_i
< MD_e
;
716 ConstantInt
*CI
= mdconst::extract
<ConstantInt
>(MD
->getOperand(MD_i
));
717 Weights
.push_back(CI
->getValue().getZExtValue());
719 for (SwitchInst::CaseIt i
= SI
->case_end(), e
= SI
->case_begin(); i
!= e
;) {
721 if (DeadCases
.count(i
.getCaseValue())) {
723 std::swap(Weights
[i
.getCaseIndex()+1], Weights
.back());
726 i
.getCaseSuccessor()->removePredecessor(TI
->getParent());
730 if (HasWeight
&& Weights
.size() >= 2)
731 SI
->setMetadata(LLVMContext::MD_prof
,
732 MDBuilder(SI
->getParent()->getContext()).
733 createBranchWeights(Weights
));
735 DEBUG(dbgs() << "Leaving: " << *TI
<< "\n");
739 // Otherwise, TI's block must correspond to some matched value. Find out
740 // which value (or set of values) this is.
741 ConstantInt
*TIV
= nullptr;
742 BasicBlock
*TIBB
= TI
->getParent();
743 for (unsigned i
= 0, e
= PredCases
.size(); i
!= e
; ++i
)
744 if (PredCases
[i
].Dest
== TIBB
) {
746 return false; // Cannot handle multiple values coming to this block.
747 TIV
= PredCases
[i
].Value
;
749 assert(TIV
&& "No edge from pred to succ?");
751 // Okay, we found the one constant that our value can be if we get into TI's
752 // BB. Find out which successor will unconditionally be branched to.
753 BasicBlock
*TheRealDest
= nullptr;
754 for (unsigned i
= 0, e
= ThisCases
.size(); i
!= e
; ++i
)
755 if (ThisCases
[i
].Value
== TIV
) {
756 TheRealDest
= ThisCases
[i
].Dest
;
760 // If not handled by any explicit cases, it is handled by the default case.
761 if (!TheRealDest
) TheRealDest
= ThisDef
;
763 // Remove PHI node entries for dead edges.
764 BasicBlock
*CheckEdge
= TheRealDest
;
765 for (succ_iterator SI
= succ_begin(TIBB
), e
= succ_end(TIBB
); SI
!= e
; ++SI
)
766 if (*SI
!= CheckEdge
)
767 (*SI
)->removePredecessor(TIBB
);
771 // Insert the new branch.
772 Instruction
*NI
= Builder
.CreateBr(TheRealDest
);
775 DEBUG(dbgs() << "Threading pred instr: " << *Pred
->getTerminator()
776 << "Through successor TI: " << *TI
<< "Leaving: " << *NI
<< "\n");
778 EraseTerminatorInstAndDCECond(TI
);
783 /// ConstantIntOrdering - This class implements a stable ordering of constant
784 /// integers that does not depend on their address. This is important for
785 /// applications that sort ConstantInt's to ensure uniqueness.
786 struct ConstantIntOrdering
{
787 bool operator()(const ConstantInt
*LHS
, const ConstantInt
*RHS
) const {
788 return LHS
->getValue().ult(RHS
->getValue());
793 static int ConstantIntSortPredicate(ConstantInt
*const *P1
,
794 ConstantInt
*const *P2
) {
795 const ConstantInt
*LHS
= *P1
;
796 const ConstantInt
*RHS
= *P2
;
797 if (LHS
->getValue().ult(RHS
->getValue()))
799 if (LHS
->getValue() == RHS
->getValue())
804 static inline bool HasBranchWeights(const Instruction
* I
) {
805 MDNode
*ProfMD
= I
->getMetadata(LLVMContext::MD_prof
);
806 if (ProfMD
&& ProfMD
->getOperand(0))
807 if (MDString
* MDS
= dyn_cast
<MDString
>(ProfMD
->getOperand(0)))
808 return MDS
->getString().equals("branch_weights");
813 /// Get Weights of a given TerminatorInst, the default weight is at the front
814 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
816 static void GetBranchWeights(TerminatorInst
*TI
,
817 SmallVectorImpl
<uint64_t> &Weights
) {
818 MDNode
*MD
= TI
->getMetadata(LLVMContext::MD_prof
);
820 for (unsigned i
= 1, e
= MD
->getNumOperands(); i
< e
; ++i
) {
821 ConstantInt
*CI
= mdconst::extract
<ConstantInt
>(MD
->getOperand(i
));
822 Weights
.push_back(CI
->getValue().getZExtValue());
825 // If TI is a conditional eq, the default case is the false case,
826 // and the corresponding branch-weight data is at index 2. We swap the
827 // default weight to be the first entry.
828 if (BranchInst
* BI
= dyn_cast
<BranchInst
>(TI
)) {
829 assert(Weights
.size() == 2);
830 ICmpInst
*ICI
= cast
<ICmpInst
>(BI
->getCondition());
831 if (ICI
->getPredicate() == ICmpInst::ICMP_EQ
)
832 std::swap(Weights
.front(), Weights
.back());
836 /// Keep halving the weights until all can fit in uint32_t.
837 static void FitWeights(MutableArrayRef
<uint64_t> Weights
) {
838 uint64_t Max
= *std::max_element(Weights
.begin(), Weights
.end());
839 if (Max
> UINT_MAX
) {
840 unsigned Offset
= 32 - countLeadingZeros(Max
);
841 for (uint64_t &I
: Weights
)
846 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
847 /// equality comparison instruction (either a switch or a branch on "X == c").
848 /// See if any of the predecessors of the terminator block are value comparisons
849 /// on the same value. If so, and if safe to do so, fold them together.
850 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst
*TI
,
851 IRBuilder
<> &Builder
) {
852 BasicBlock
*BB
= TI
->getParent();
853 Value
*CV
= isValueEqualityComparison(TI
); // CondVal
854 assert(CV
&& "Not a comparison?");
855 bool Changed
= false;
857 SmallVector
<BasicBlock
*, 16> Preds(pred_begin(BB
), pred_end(BB
));
858 while (!Preds
.empty()) {
859 BasicBlock
*Pred
= Preds
.pop_back_val();
861 // See if the predecessor is a comparison with the same value.
862 TerminatorInst
*PTI
= Pred
->getTerminator();
863 Value
*PCV
= isValueEqualityComparison(PTI
); // PredCondVal
865 if (PCV
== CV
&& SafeToMergeTerminators(TI
, PTI
)) {
866 // Figure out which 'cases' to copy from SI to PSI.
867 std::vector
<ValueEqualityComparisonCase
> BBCases
;
868 BasicBlock
*BBDefault
= GetValueEqualityComparisonCases(TI
, BBCases
);
870 std::vector
<ValueEqualityComparisonCase
> PredCases
;
871 BasicBlock
*PredDefault
= GetValueEqualityComparisonCases(PTI
, PredCases
);
873 // Based on whether the default edge from PTI goes to BB or not, fill in
874 // PredCases and PredDefault with the new switch cases we would like to
876 SmallVector
<BasicBlock
*, 8> NewSuccessors
;
878 // Update the branch weight metadata along the way
879 SmallVector
<uint64_t, 8> Weights
;
880 bool PredHasWeights
= HasBranchWeights(PTI
);
881 bool SuccHasWeights
= HasBranchWeights(TI
);
883 if (PredHasWeights
) {
884 GetBranchWeights(PTI
, Weights
);
885 // branch-weight metadata is inconsistent here.
886 if (Weights
.size() != 1 + PredCases
.size())
887 PredHasWeights
= SuccHasWeights
= false;
888 } else if (SuccHasWeights
)
889 // If there are no predecessor weights but there are successor weights,
890 // populate Weights with 1, which will later be scaled to the sum of
891 // successor's weights
892 Weights
.assign(1 + PredCases
.size(), 1);
894 SmallVector
<uint64_t, 8> SuccWeights
;
895 if (SuccHasWeights
) {
896 GetBranchWeights(TI
, SuccWeights
);
897 // branch-weight metadata is inconsistent here.
898 if (SuccWeights
.size() != 1 + BBCases
.size())
899 PredHasWeights
= SuccHasWeights
= false;
900 } else if (PredHasWeights
)
901 SuccWeights
.assign(1 + BBCases
.size(), 1);
903 if (PredDefault
== BB
) {
904 // If this is the default destination from PTI, only the edges in TI
905 // that don't occur in PTI, or that branch to BB will be activated.
906 std::set
<ConstantInt
*, ConstantIntOrdering
> PTIHandled
;
907 for (unsigned i
= 0, e
= PredCases
.size(); i
!= e
; ++i
)
908 if (PredCases
[i
].Dest
!= BB
)
909 PTIHandled
.insert(PredCases
[i
].Value
);
911 // The default destination is BB, we don't need explicit targets.
912 std::swap(PredCases
[i
], PredCases
.back());
914 if (PredHasWeights
|| SuccHasWeights
) {
915 // Increase weight for the default case.
916 Weights
[0] += Weights
[i
+1];
917 std::swap(Weights
[i
+1], Weights
.back());
921 PredCases
.pop_back();
925 // Reconstruct the new switch statement we will be building.
926 if (PredDefault
!= BBDefault
) {
927 PredDefault
->removePredecessor(Pred
);
928 PredDefault
= BBDefault
;
929 NewSuccessors
.push_back(BBDefault
);
932 unsigned CasesFromPred
= Weights
.size();
933 uint64_t ValidTotalSuccWeight
= 0;
934 for (unsigned i
= 0, e
= BBCases
.size(); i
!= e
; ++i
)
935 if (!PTIHandled
.count(BBCases
[i
].Value
) &&
936 BBCases
[i
].Dest
!= BBDefault
) {
937 PredCases
.push_back(BBCases
[i
]);
938 NewSuccessors
.push_back(BBCases
[i
].Dest
);
939 if (SuccHasWeights
|| PredHasWeights
) {
940 // The default weight is at index 0, so weight for the ith case
941 // should be at index i+1. Scale the cases from successor by
942 // PredDefaultWeight (Weights[0]).
943 Weights
.push_back(Weights
[0] * SuccWeights
[i
+1]);
944 ValidTotalSuccWeight
+= SuccWeights
[i
+1];
948 if (SuccHasWeights
|| PredHasWeights
) {
949 ValidTotalSuccWeight
+= SuccWeights
[0];
950 // Scale the cases from predecessor by ValidTotalSuccWeight.
951 for (unsigned i
= 1; i
< CasesFromPred
; ++i
)
952 Weights
[i
] *= ValidTotalSuccWeight
;
953 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
954 Weights
[0] *= SuccWeights
[0];
957 // If this is not the default destination from PSI, only the edges
958 // in SI that occur in PSI with a destination of BB will be
960 std::set
<ConstantInt
*, ConstantIntOrdering
> PTIHandled
;
961 std::map
<ConstantInt
*, uint64_t> WeightsForHandled
;
962 for (unsigned i
= 0, e
= PredCases
.size(); i
!= e
; ++i
)
963 if (PredCases
[i
].Dest
== BB
) {
964 PTIHandled
.insert(PredCases
[i
].Value
);
966 if (PredHasWeights
|| SuccHasWeights
) {
967 WeightsForHandled
[PredCases
[i
].Value
] = Weights
[i
+1];
968 std::swap(Weights
[i
+1], Weights
.back());
972 std::swap(PredCases
[i
], PredCases
.back());
973 PredCases
.pop_back();
977 // Okay, now we know which constants were sent to BB from the
978 // predecessor. Figure out where they will all go now.
979 for (unsigned i
= 0, e
= BBCases
.size(); i
!= e
; ++i
)
980 if (PTIHandled
.count(BBCases
[i
].Value
)) {
981 // If this is one we are capable of getting...
982 if (PredHasWeights
|| SuccHasWeights
)
983 Weights
.push_back(WeightsForHandled
[BBCases
[i
].Value
]);
984 PredCases
.push_back(BBCases
[i
]);
985 NewSuccessors
.push_back(BBCases
[i
].Dest
);
986 PTIHandled
.erase(BBCases
[i
].Value
);// This constant is taken care of
989 // If there are any constants vectored to BB that TI doesn't handle,
990 // they must go to the default destination of TI.
991 for (std::set
<ConstantInt
*, ConstantIntOrdering
>::iterator I
=
993 E
= PTIHandled
.end(); I
!= E
; ++I
) {
994 if (PredHasWeights
|| SuccHasWeights
)
995 Weights
.push_back(WeightsForHandled
[*I
]);
996 PredCases
.push_back(ValueEqualityComparisonCase(*I
, BBDefault
));
997 NewSuccessors
.push_back(BBDefault
);
1001 // Okay, at this point, we know which new successor Pred will get. Make
1002 // sure we update the number of entries in the PHI nodes for these
1004 for (unsigned i
= 0, e
= NewSuccessors
.size(); i
!= e
; ++i
)
1005 AddPredecessorToBlock(NewSuccessors
[i
], Pred
, BB
);
1007 Builder
.SetInsertPoint(PTI
);
1008 // Convert pointer to int before we switch.
1009 if (CV
->getType()->isPointerTy()) {
1010 assert(DL
&& "Cannot switch on pointer without DataLayout");
1011 CV
= Builder
.CreatePtrToInt(CV
, DL
->getIntPtrType(CV
->getType()),
1015 // Now that the successors are updated, create the new Switch instruction.
1016 SwitchInst
*NewSI
= Builder
.CreateSwitch(CV
, PredDefault
,
1018 NewSI
->setDebugLoc(PTI
->getDebugLoc());
1019 for (unsigned i
= 0, e
= PredCases
.size(); i
!= e
; ++i
)
1020 NewSI
->addCase(PredCases
[i
].Value
, PredCases
[i
].Dest
);
1022 if (PredHasWeights
|| SuccHasWeights
) {
1023 // Halve the weights if any of them cannot fit in an uint32_t
1024 FitWeights(Weights
);
1026 SmallVector
<uint32_t, 8> MDWeights(Weights
.begin(), Weights
.end());
1028 NewSI
->setMetadata(LLVMContext::MD_prof
,
1029 MDBuilder(BB
->getContext()).
1030 createBranchWeights(MDWeights
));
1033 EraseTerminatorInstAndDCECond(PTI
);
1035 // Okay, last check. If BB is still a successor of PSI, then we must
1036 // have an infinite loop case. If so, add an infinitely looping block
1037 // to handle the case to preserve the behavior of the code.
1038 BasicBlock
*InfLoopBlock
= nullptr;
1039 for (unsigned i
= 0, e
= NewSI
->getNumSuccessors(); i
!= e
; ++i
)
1040 if (NewSI
->getSuccessor(i
) == BB
) {
1041 if (!InfLoopBlock
) {
1042 // Insert it at the end of the function, because it's either code,
1043 // or it won't matter if it's hot. :)
1044 InfLoopBlock
= BasicBlock::Create(BB
->getContext(),
1045 "infloop", BB
->getParent());
1046 BranchInst::Create(InfLoopBlock
, InfLoopBlock
);
1048 NewSI
->setSuccessor(i
, InfLoopBlock
);
1057 // isSafeToHoistInvoke - If we would need to insert a select that uses the
1058 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
1059 // would need to do this), we can't hoist the invoke, as there is nowhere
1060 // to put the select in this case.
1061 static bool isSafeToHoistInvoke(BasicBlock
*BB1
, BasicBlock
*BB2
,
1062 Instruction
*I1
, Instruction
*I2
) {
1063 for (succ_iterator SI
= succ_begin(BB1
), E
= succ_end(BB1
); SI
!= E
; ++SI
) {
1065 for (BasicBlock::iterator BBI
= SI
->begin();
1066 (PN
= dyn_cast
<PHINode
>(BBI
)); ++BBI
) {
1067 Value
*BB1V
= PN
->getIncomingValueForBlock(BB1
);
1068 Value
*BB2V
= PN
->getIncomingValueForBlock(BB2
);
1069 if (BB1V
!= BB2V
&& (BB1V
==I1
|| BB2V
==I2
)) {
1077 static bool passingValueIsAlwaysUndefined(Value
*V
, Instruction
*I
);
1079 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1080 /// BB2, hoist any common code in the two blocks up into the branch block. The
1081 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1082 static bool HoistThenElseCodeToIf(BranchInst
*BI
, const DataLayout
*DL
) {
1083 // This does very trivial matching, with limited scanning, to find identical
1084 // instructions in the two blocks. In particular, we don't want to get into
1085 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1086 // such, we currently just scan for obviously identical instructions in an
1088 BasicBlock
*BB1
= BI
->getSuccessor(0); // The true destination.
1089 BasicBlock
*BB2
= BI
->getSuccessor(1); // The false destination
1091 BasicBlock::iterator BB1_Itr
= BB1
->begin();
1092 BasicBlock::iterator BB2_Itr
= BB2
->begin();
1094 Instruction
*I1
= BB1_Itr
++, *I2
= BB2_Itr
++;
1095 // Skip debug info if it is not identical.
1096 DbgInfoIntrinsic
*DBI1
= dyn_cast
<DbgInfoIntrinsic
>(I1
);
1097 DbgInfoIntrinsic
*DBI2
= dyn_cast
<DbgInfoIntrinsic
>(I2
);
1098 if (!DBI1
|| !DBI2
|| !DBI1
->isIdenticalToWhenDefined(DBI2
)) {
1099 while (isa
<DbgInfoIntrinsic
>(I1
))
1101 while (isa
<DbgInfoIntrinsic
>(I2
))
1104 if (isa
<PHINode
>(I1
) || !I1
->isIdenticalToWhenDefined(I2
) ||
1105 (isa
<InvokeInst
>(I1
) && !isSafeToHoistInvoke(BB1
, BB2
, I1
, I2
)))
1108 BasicBlock
*BIParent
= BI
->getParent();
1110 bool Changed
= false;
1112 // If we are hoisting the terminator instruction, don't move one (making a
1113 // broken BB), instead clone it, and remove BI.
1114 if (isa
<TerminatorInst
>(I1
))
1115 goto HoistTerminator
;
1117 // For a normal instruction, we just move one to right before the branch,
1118 // then replace all uses of the other with the first. Finally, we remove
1119 // the now redundant second instruction.
1120 BIParent
->getInstList().splice(BI
, BB1
->getInstList(), I1
);
1121 if (!I2
->use_empty())
1122 I2
->replaceAllUsesWith(I1
);
1123 I1
->intersectOptionalDataWith(I2
);
1124 unsigned KnownIDs
[] = {
1125 LLVMContext::MD_tbaa
,
1126 LLVMContext::MD_range
,
1127 LLVMContext::MD_fpmath
,
1128 LLVMContext::MD_invariant_load
,
1129 LLVMContext::MD_nonnull
1131 combineMetadata(I1
, I2
, KnownIDs
);
1132 I2
->eraseFromParent();
1137 // Skip debug info if it is not identical.
1138 DbgInfoIntrinsic
*DBI1
= dyn_cast
<DbgInfoIntrinsic
>(I1
);
1139 DbgInfoIntrinsic
*DBI2
= dyn_cast
<DbgInfoIntrinsic
>(I2
);
1140 if (!DBI1
|| !DBI2
|| !DBI1
->isIdenticalToWhenDefined(DBI2
)) {
1141 while (isa
<DbgInfoIntrinsic
>(I1
))
1143 while (isa
<DbgInfoIntrinsic
>(I2
))
1146 } while (I1
->isIdenticalToWhenDefined(I2
));
1151 // It may not be possible to hoist an invoke.
1152 if (isa
<InvokeInst
>(I1
) && !isSafeToHoistInvoke(BB1
, BB2
, I1
, I2
))
1155 for (succ_iterator SI
= succ_begin(BB1
), E
= succ_end(BB1
); SI
!= E
; ++SI
) {
1157 for (BasicBlock::iterator BBI
= SI
->begin();
1158 (PN
= dyn_cast
<PHINode
>(BBI
)); ++BBI
) {
1159 Value
*BB1V
= PN
->getIncomingValueForBlock(BB1
);
1160 Value
*BB2V
= PN
->getIncomingValueForBlock(BB2
);
1164 // Check for passingValueIsAlwaysUndefined here because we would rather
1165 // eliminate undefined control flow then converting it to a select.
1166 if (passingValueIsAlwaysUndefined(BB1V
, PN
) ||
1167 passingValueIsAlwaysUndefined(BB2V
, PN
))
1170 if (isa
<ConstantExpr
>(BB1V
) && !isSafeToSpeculativelyExecute(BB1V
, DL
))
1172 if (isa
<ConstantExpr
>(BB2V
) && !isSafeToSpeculativelyExecute(BB2V
, DL
))
1177 // Okay, it is safe to hoist the terminator.
1178 Instruction
*NT
= I1
->clone();
1179 BIParent
->getInstList().insert(BI
, NT
);
1180 if (!NT
->getType()->isVoidTy()) {
1181 I1
->replaceAllUsesWith(NT
);
1182 I2
->replaceAllUsesWith(NT
);
1186 IRBuilder
<true, NoFolder
> Builder(NT
);
1187 // Hoisting one of the terminators from our successor is a great thing.
1188 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1189 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1190 // nodes, so we insert select instruction to compute the final result.
1191 std::map
<std::pair
<Value
*,Value
*>, SelectInst
*> InsertedSelects
;
1192 for (succ_iterator SI
= succ_begin(BB1
), E
= succ_end(BB1
); SI
!= E
; ++SI
) {
1194 for (BasicBlock::iterator BBI
= SI
->begin();
1195 (PN
= dyn_cast
<PHINode
>(BBI
)); ++BBI
) {
1196 Value
*BB1V
= PN
->getIncomingValueForBlock(BB1
);
1197 Value
*BB2V
= PN
->getIncomingValueForBlock(BB2
);
1198 if (BB1V
== BB2V
) continue;
1200 // These values do not agree. Insert a select instruction before NT
1201 // that determines the right value.
1202 SelectInst
*&SI
= InsertedSelects
[std::make_pair(BB1V
, BB2V
)];
1204 SI
= cast
<SelectInst
>
1205 (Builder
.CreateSelect(BI
->getCondition(), BB1V
, BB2V
,
1206 BB1V
->getName()+"."+BB2V
->getName()));
1208 // Make the PHI node use the select for all incoming values for BB1/BB2
1209 for (unsigned i
= 0, e
= PN
->getNumIncomingValues(); i
!= e
; ++i
)
1210 if (PN
->getIncomingBlock(i
) == BB1
|| PN
->getIncomingBlock(i
) == BB2
)
1211 PN
->setIncomingValue(i
, SI
);
1215 // Update any PHI nodes in our new successors.
1216 for (succ_iterator SI
= succ_begin(BB1
), E
= succ_end(BB1
); SI
!= E
; ++SI
)
1217 AddPredecessorToBlock(*SI
, BIParent
, BB1
);
1219 EraseTerminatorInstAndDCECond(BI
);
1223 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1224 /// check whether BBEnd has only two predecessors and the other predecessor
1225 /// ends with an unconditional branch. If it is true, sink any common code
1226 /// in the two predecessors to BBEnd.
1227 static bool SinkThenElseCodeToEnd(BranchInst
*BI1
) {
1228 assert(BI1
->isUnconditional());
1229 BasicBlock
*BB1
= BI1
->getParent();
1230 BasicBlock
*BBEnd
= BI1
->getSuccessor(0);
1232 // Check that BBEnd has two predecessors and the other predecessor ends with
1233 // an unconditional branch.
1234 pred_iterator PI
= pred_begin(BBEnd
), PE
= pred_end(BBEnd
);
1235 BasicBlock
*Pred0
= *PI
++;
1236 if (PI
== PE
) // Only one predecessor.
1238 BasicBlock
*Pred1
= *PI
++;
1239 if (PI
!= PE
) // More than two predecessors.
1241 BasicBlock
*BB2
= (Pred0
== BB1
) ? Pred1
: Pred0
;
1242 BranchInst
*BI2
= dyn_cast
<BranchInst
>(BB2
->getTerminator());
1243 if (!BI2
|| !BI2
->isUnconditional())
1246 // Gather the PHI nodes in BBEnd.
1247 SmallDenseMap
<std::pair
<Value
*, Value
*>, PHINode
*> JointValueMap
;
1248 Instruction
*FirstNonPhiInBBEnd
= nullptr;
1249 for (BasicBlock::iterator I
= BBEnd
->begin(), E
= BBEnd
->end(); I
!= E
; ++I
) {
1250 if (PHINode
*PN
= dyn_cast
<PHINode
>(I
)) {
1251 Value
*BB1V
= PN
->getIncomingValueForBlock(BB1
);
1252 Value
*BB2V
= PN
->getIncomingValueForBlock(BB2
);
1253 JointValueMap
[std::make_pair(BB1V
, BB2V
)] = PN
;
1255 FirstNonPhiInBBEnd
= &*I
;
1259 if (!FirstNonPhiInBBEnd
)
1262 // This does very trivial matching, with limited scanning, to find identical
1263 // instructions in the two blocks. We scan backward for obviously identical
1264 // instructions in an identical order.
1265 BasicBlock::InstListType::reverse_iterator RI1
= BB1
->getInstList().rbegin(),
1266 RE1
= BB1
->getInstList().rend(),
1267 RI2
= BB2
->getInstList().rbegin(),
1268 RE2
= BB2
->getInstList().rend();
1270 while (RI1
!= RE1
&& isa
<DbgInfoIntrinsic
>(&*RI1
)) ++RI1
;
1273 while (RI2
!= RE2
&& isa
<DbgInfoIntrinsic
>(&*RI2
)) ++RI2
;
1276 // Skip the unconditional branches.
1280 bool Changed
= false;
1281 while (RI1
!= RE1
&& RI2
!= RE2
) {
1283 while (RI1
!= RE1
&& isa
<DbgInfoIntrinsic
>(&*RI1
)) ++RI1
;
1286 while (RI2
!= RE2
&& isa
<DbgInfoIntrinsic
>(&*RI2
)) ++RI2
;
1290 Instruction
*I1
= &*RI1
, *I2
= &*RI2
;
1291 auto InstPair
= std::make_pair(I1
, I2
);
1292 // I1 and I2 should have a single use in the same PHI node, and they
1293 // perform the same operation.
1294 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1295 if (isa
<PHINode
>(I1
) || isa
<PHINode
>(I2
) ||
1296 isa
<TerminatorInst
>(I1
) || isa
<TerminatorInst
>(I2
) ||
1297 isa
<LandingPadInst
>(I1
) || isa
<LandingPadInst
>(I2
) ||
1298 isa
<AllocaInst
>(I1
) || isa
<AllocaInst
>(I2
) ||
1299 I1
->mayHaveSideEffects() || I2
->mayHaveSideEffects() ||
1300 I1
->mayReadOrWriteMemory() || I2
->mayReadOrWriteMemory() ||
1301 !I1
->hasOneUse() || !I2
->hasOneUse() ||
1302 !JointValueMap
.count(InstPair
))
1305 // Check whether we should swap the operands of ICmpInst.
1306 // TODO: Add support of communativity.
1307 ICmpInst
*ICmp1
= dyn_cast
<ICmpInst
>(I1
), *ICmp2
= dyn_cast
<ICmpInst
>(I2
);
1308 bool SwapOpnds
= false;
1309 if (ICmp1
&& ICmp2
&&
1310 ICmp1
->getOperand(0) != ICmp2
->getOperand(0) &&
1311 ICmp1
->getOperand(1) != ICmp2
->getOperand(1) &&
1312 (ICmp1
->getOperand(0) == ICmp2
->getOperand(1) ||
1313 ICmp1
->getOperand(1) == ICmp2
->getOperand(0))) {
1314 ICmp2
->swapOperands();
1317 if (!I1
->isSameOperationAs(I2
)) {
1319 ICmp2
->swapOperands();
1323 // The operands should be either the same or they need to be generated
1324 // with a PHI node after sinking. We only handle the case where there is
1325 // a single pair of different operands.
1326 Value
*DifferentOp1
= nullptr, *DifferentOp2
= nullptr;
1327 unsigned Op1Idx
= ~0U;
1328 for (unsigned I
= 0, E
= I1
->getNumOperands(); I
!= E
; ++I
) {
1329 if (I1
->getOperand(I
) == I2
->getOperand(I
))
1331 // Early exit if we have more-than one pair of different operands or if
1332 // we need a PHI node to replace a constant.
1333 if (Op1Idx
!= ~0U ||
1334 isa
<Constant
>(I1
->getOperand(I
)) ||
1335 isa
<Constant
>(I2
->getOperand(I
))) {
1336 // If we can't sink the instructions, undo the swapping.
1338 ICmp2
->swapOperands();
1341 DifferentOp1
= I1
->getOperand(I
);
1343 DifferentOp2
= I2
->getOperand(I
);
1346 DEBUG(dbgs() << "SINK common instructions " << *I1
<< "\n");
1347 DEBUG(dbgs() << " " << *I2
<< "\n");
1349 // We insert the pair of different operands to JointValueMap and
1350 // remove (I1, I2) from JointValueMap.
1351 if (Op1Idx
!= ~0U) {
1352 auto &NewPN
= JointValueMap
[std::make_pair(DifferentOp1
, DifferentOp2
)];
1355 PHINode::Create(DifferentOp1
->getType(), 2,
1356 DifferentOp1
->getName() + ".sink", BBEnd
->begin());
1357 NewPN
->addIncoming(DifferentOp1
, BB1
);
1358 NewPN
->addIncoming(DifferentOp2
, BB2
);
1359 DEBUG(dbgs() << "Create PHI node " << *NewPN
<< "\n";);
1361 // I1 should use NewPN instead of DifferentOp1.
1362 I1
->setOperand(Op1Idx
, NewPN
);
1364 PHINode
*OldPN
= JointValueMap
[InstPair
];
1365 JointValueMap
.erase(InstPair
);
1367 // We need to update RE1 and RE2 if we are going to sink the first
1368 // instruction in the basic block down.
1369 bool UpdateRE1
= (I1
== BB1
->begin()), UpdateRE2
= (I2
== BB2
->begin());
1370 // Sink the instruction.
1371 BBEnd
->getInstList().splice(FirstNonPhiInBBEnd
, BB1
->getInstList(), I1
);
1372 if (!OldPN
->use_empty())
1373 OldPN
->replaceAllUsesWith(I1
);
1374 OldPN
->eraseFromParent();
1376 if (!I2
->use_empty())
1377 I2
->replaceAllUsesWith(I1
);
1378 I1
->intersectOptionalDataWith(I2
);
1379 // TODO: Use combineMetadata here to preserve what metadata we can
1380 // (analogous to the hoisting case above).
1381 I2
->eraseFromParent();
1384 RE1
= BB1
->getInstList().rend();
1386 RE2
= BB2
->getInstList().rend();
1387 FirstNonPhiInBBEnd
= I1
;
1394 /// \brief Determine if we can hoist sink a sole store instruction out of a
1395 /// conditional block.
1397 /// We are looking for code like the following:
1399 /// store i32 %add, i32* %arrayidx2
1400 /// ... // No other stores or function calls (we could be calling a memory
1401 /// ... // function).
1402 /// %cmp = icmp ult %x, %y
1403 /// br i1 %cmp, label %EndBB, label %ThenBB
1405 /// store i32 %add5, i32* %arrayidx2
1409 /// We are going to transform this into:
1411 /// store i32 %add, i32* %arrayidx2
1413 /// %cmp = icmp ult %x, %y
1414 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1415 /// store i32 %add.add5, i32* %arrayidx2
1418 /// \return The pointer to the value of the previous store if the store can be
1419 /// hoisted into the predecessor block. 0 otherwise.
1420 static Value
*isSafeToSpeculateStore(Instruction
*I
, BasicBlock
*BrBB
,
1421 BasicBlock
*StoreBB
, BasicBlock
*EndBB
) {
1422 StoreInst
*StoreToHoist
= dyn_cast
<StoreInst
>(I
);
1426 // Volatile or atomic.
1427 if (!StoreToHoist
->isSimple())
1430 Value
*StorePtr
= StoreToHoist
->getPointerOperand();
1432 // Look for a store to the same pointer in BrBB.
1433 unsigned MaxNumInstToLookAt
= 10;
1434 for (BasicBlock::reverse_iterator RI
= BrBB
->rbegin(),
1435 RE
= BrBB
->rend(); RI
!= RE
&& (--MaxNumInstToLookAt
); ++RI
) {
1436 Instruction
*CurI
= &*RI
;
1438 // Could be calling an instruction that effects memory like free().
1439 if (CurI
->mayHaveSideEffects() && !isa
<StoreInst
>(CurI
))
1442 StoreInst
*SI
= dyn_cast
<StoreInst
>(CurI
);
1443 // Found the previous store make sure it stores to the same location.
1444 if (SI
&& SI
->getPointerOperand() == StorePtr
)
1445 // Found the previous store, return its value operand.
1446 return SI
->getValueOperand();
1448 return nullptr; // Unknown store.
1454 /// \brief Speculate a conditional basic block flattening the CFG.
1456 /// Note that this is a very risky transform currently. Speculating
1457 /// instructions like this is most often not desirable. Instead, there is an MI
1458 /// pass which can do it with full awareness of the resource constraints.
1459 /// However, some cases are "obvious" and we should do directly. An example of
1460 /// this is speculating a single, reasonably cheap instruction.
1462 /// There is only one distinct advantage to flattening the CFG at the IR level:
1463 /// it makes very common but simplistic optimizations such as are common in
1464 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1465 /// modeling their effects with easier to reason about SSA value graphs.
1468 /// An illustration of this transform is turning this IR:
1471 /// %cmp = icmp ult %x, %y
1472 /// br i1 %cmp, label %EndBB, label %ThenBB
1474 /// %sub = sub %x, %y
1477 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1484 /// %cmp = icmp ult %x, %y
1485 /// %sub = sub %x, %y
1486 /// %cond = select i1 %cmp, 0, %sub
1490 /// \returns true if the conditional block is removed.
1491 static bool SpeculativelyExecuteBB(BranchInst
*BI
, BasicBlock
*ThenBB
,
1492 const DataLayout
*DL
) {
1493 // Be conservative for now. FP select instruction can often be expensive.
1494 Value
*BrCond
= BI
->getCondition();
1495 if (isa
<FCmpInst
>(BrCond
))
1498 BasicBlock
*BB
= BI
->getParent();
1499 BasicBlock
*EndBB
= ThenBB
->getTerminator()->getSuccessor(0);
1501 // If ThenBB is actually on the false edge of the conditional branch, remember
1502 // to swap the select operands later.
1503 bool Invert
= false;
1504 if (ThenBB
!= BI
->getSuccessor(0)) {
1505 assert(ThenBB
== BI
->getSuccessor(1) && "No edge from 'if' block?");
1508 assert(EndBB
== BI
->getSuccessor(!Invert
) && "No edge from to end block");
1510 // Keep a count of how many times instructions are used within CondBB when
1511 // they are candidates for sinking into CondBB. Specifically:
1512 // - They are defined in BB, and
1513 // - They have no side effects, and
1514 // - All of their uses are in CondBB.
1515 SmallDenseMap
<Instruction
*, unsigned, 4> SinkCandidateUseCounts
;
1517 unsigned SpeculationCost
= 0;
1518 Value
*SpeculatedStoreValue
= nullptr;
1519 StoreInst
*SpeculatedStore
= nullptr;
1520 for (BasicBlock::iterator BBI
= ThenBB
->begin(),
1521 BBE
= std::prev(ThenBB
->end());
1522 BBI
!= BBE
; ++BBI
) {
1523 Instruction
*I
= BBI
;
1525 if (isa
<DbgInfoIntrinsic
>(I
))
1528 // Only speculatively execution a single instruction (not counting the
1529 // terminator) for now.
1531 if (SpeculationCost
> 1)
1534 // Don't hoist the instruction if it's unsafe or expensive.
1535 if (!isSafeToSpeculativelyExecute(I
, DL
) &&
1536 !(HoistCondStores
&&
1537 (SpeculatedStoreValue
= isSafeToSpeculateStore(I
, BB
, ThenBB
,
1540 if (!SpeculatedStoreValue
&&
1541 ComputeSpeculationCost(I
, DL
) > PHINodeFoldingThreshold
)
1544 // Store the store speculation candidate.
1545 if (SpeculatedStoreValue
)
1546 SpeculatedStore
= cast
<StoreInst
>(I
);
1548 // Do not hoist the instruction if any of its operands are defined but not
1549 // used in BB. The transformation will prevent the operand from
1550 // being sunk into the use block.
1551 for (User::op_iterator i
= I
->op_begin(), e
= I
->op_end();
1553 Instruction
*OpI
= dyn_cast
<Instruction
>(*i
);
1554 if (!OpI
|| OpI
->getParent() != BB
||
1555 OpI
->mayHaveSideEffects())
1556 continue; // Not a candidate for sinking.
1558 ++SinkCandidateUseCounts
[OpI
];
1562 // Consider any sink candidates which are only used in CondBB as costs for
1563 // speculation. Note, while we iterate over a DenseMap here, we are summing
1564 // and so iteration order isn't significant.
1565 for (SmallDenseMap
<Instruction
*, unsigned, 4>::iterator I
=
1566 SinkCandidateUseCounts
.begin(), E
= SinkCandidateUseCounts
.end();
1568 if (I
->first
->getNumUses() == I
->second
) {
1570 if (SpeculationCost
> 1)
1574 // Check that the PHI nodes can be converted to selects.
1575 bool HaveRewritablePHIs
= false;
1576 for (BasicBlock::iterator I
= EndBB
->begin();
1577 PHINode
*PN
= dyn_cast
<PHINode
>(I
); ++I
) {
1578 Value
*OrigV
= PN
->getIncomingValueForBlock(BB
);
1579 Value
*ThenV
= PN
->getIncomingValueForBlock(ThenBB
);
1581 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1582 // Skip PHIs which are trivial.
1586 // Don't convert to selects if we could remove undefined behavior instead.
1587 if (passingValueIsAlwaysUndefined(OrigV
, PN
) ||
1588 passingValueIsAlwaysUndefined(ThenV
, PN
))
1591 HaveRewritablePHIs
= true;
1592 ConstantExpr
*OrigCE
= dyn_cast
<ConstantExpr
>(OrigV
);
1593 ConstantExpr
*ThenCE
= dyn_cast
<ConstantExpr
>(ThenV
);
1594 if (!OrigCE
&& !ThenCE
)
1595 continue; // Known safe and cheap.
1597 if ((ThenCE
&& !isSafeToSpeculativelyExecute(ThenCE
, DL
)) ||
1598 (OrigCE
&& !isSafeToSpeculativelyExecute(OrigCE
, DL
)))
1600 unsigned OrigCost
= OrigCE
? ComputeSpeculationCost(OrigCE
, DL
) : 0;
1601 unsigned ThenCost
= ThenCE
? ComputeSpeculationCost(ThenCE
, DL
) : 0;
1602 if (OrigCost
+ ThenCost
> 2 * PHINodeFoldingThreshold
)
1605 // Account for the cost of an unfolded ConstantExpr which could end up
1606 // getting expanded into Instructions.
1607 // FIXME: This doesn't account for how many operations are combined in the
1608 // constant expression.
1610 if (SpeculationCost
> 1)
1614 // If there are no PHIs to process, bail early. This helps ensure idempotence
1616 if (!HaveRewritablePHIs
&& !(HoistCondStores
&& SpeculatedStoreValue
))
1619 // If we get here, we can hoist the instruction and if-convert.
1620 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB
<< "\n";);
1622 // Insert a select of the value of the speculated store.
1623 if (SpeculatedStoreValue
) {
1624 IRBuilder
<true, NoFolder
> Builder(BI
);
1625 Value
*TrueV
= SpeculatedStore
->getValueOperand();
1626 Value
*FalseV
= SpeculatedStoreValue
;
1628 std::swap(TrueV
, FalseV
);
1629 Value
*S
= Builder
.CreateSelect(BrCond
, TrueV
, FalseV
, TrueV
->getName() +
1630 "." + FalseV
->getName());
1631 SpeculatedStore
->setOperand(0, S
);
1634 // Hoist the instructions.
1635 BB
->getInstList().splice(BI
, ThenBB
->getInstList(), ThenBB
->begin(),
1636 std::prev(ThenBB
->end()));
1638 // Insert selects and rewrite the PHI operands.
1639 IRBuilder
<true, NoFolder
> Builder(BI
);
1640 for (BasicBlock::iterator I
= EndBB
->begin();
1641 PHINode
*PN
= dyn_cast
<PHINode
>(I
); ++I
) {
1642 unsigned OrigI
= PN
->getBasicBlockIndex(BB
);
1643 unsigned ThenI
= PN
->getBasicBlockIndex(ThenBB
);
1644 Value
*OrigV
= PN
->getIncomingValue(OrigI
);
1645 Value
*ThenV
= PN
->getIncomingValue(ThenI
);
1647 // Skip PHIs which are trivial.
1651 // Create a select whose true value is the speculatively executed value and
1652 // false value is the preexisting value. Swap them if the branch
1653 // destinations were inverted.
1654 Value
*TrueV
= ThenV
, *FalseV
= OrigV
;
1656 std::swap(TrueV
, FalseV
);
1657 Value
*V
= Builder
.CreateSelect(BrCond
, TrueV
, FalseV
,
1658 TrueV
->getName() + "." + FalseV
->getName());
1659 PN
->setIncomingValue(OrigI
, V
);
1660 PN
->setIncomingValue(ThenI
, V
);
1667 /// \returns True if this block contains a CallInst with the NoDuplicate
1669 static bool HasNoDuplicateCall(const BasicBlock
*BB
) {
1670 for (BasicBlock::const_iterator I
= BB
->begin(), E
= BB
->end(); I
!= E
; ++I
) {
1671 const CallInst
*CI
= dyn_cast
<CallInst
>(I
);
1674 if (CI
->cannotDuplicate())
1680 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1681 /// across this block.
1682 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock
*BB
) {
1683 BranchInst
*BI
= cast
<BranchInst
>(BB
->getTerminator());
1686 for (BasicBlock::iterator BBI
= BB
->begin(); &*BBI
!= BI
; ++BBI
) {
1687 if (isa
<DbgInfoIntrinsic
>(BBI
))
1689 if (Size
> 10) return false; // Don't clone large BB's.
1692 // We can only support instructions that do not define values that are
1693 // live outside of the current basic block.
1694 for (User
*U
: BBI
->users()) {
1695 Instruction
*UI
= cast
<Instruction
>(U
);
1696 if (UI
->getParent() != BB
|| isa
<PHINode
>(UI
)) return false;
1699 // Looks ok, continue checking.
1705 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1706 /// that is defined in the same block as the branch and if any PHI entries are
1707 /// constants, thread edges corresponding to that entry to be branches to their
1708 /// ultimate destination.
1709 static bool FoldCondBranchOnPHI(BranchInst
*BI
, const DataLayout
*DL
) {
1710 BasicBlock
*BB
= BI
->getParent();
1711 PHINode
*PN
= dyn_cast
<PHINode
>(BI
->getCondition());
1712 // NOTE: we currently cannot transform this case if the PHI node is used
1713 // outside of the block.
1714 if (!PN
|| PN
->getParent() != BB
|| !PN
->hasOneUse())
1717 // Degenerate case of a single entry PHI.
1718 if (PN
->getNumIncomingValues() == 1) {
1719 FoldSingleEntryPHINodes(PN
->getParent());
1723 // Now we know that this block has multiple preds and two succs.
1724 if (!BlockIsSimpleEnoughToThreadThrough(BB
)) return false;
1726 if (HasNoDuplicateCall(BB
)) return false;
1728 // Okay, this is a simple enough basic block. See if any phi values are
1730 for (unsigned i
= 0, e
= PN
->getNumIncomingValues(); i
!= e
; ++i
) {
1731 ConstantInt
*CB
= dyn_cast
<ConstantInt
>(PN
->getIncomingValue(i
));
1732 if (!CB
|| !CB
->getType()->isIntegerTy(1)) continue;
1734 // Okay, we now know that all edges from PredBB should be revectored to
1735 // branch to RealDest.
1736 BasicBlock
*PredBB
= PN
->getIncomingBlock(i
);
1737 BasicBlock
*RealDest
= BI
->getSuccessor(!CB
->getZExtValue());
1739 if (RealDest
== BB
) continue; // Skip self loops.
1740 // Skip if the predecessor's terminator is an indirect branch.
1741 if (isa
<IndirectBrInst
>(PredBB
->getTerminator())) continue;
1743 // The dest block might have PHI nodes, other predecessors and other
1744 // difficult cases. Instead of being smart about this, just insert a new
1745 // block that jumps to the destination block, effectively splitting
1746 // the edge we are about to create.
1747 BasicBlock
*EdgeBB
= BasicBlock::Create(BB
->getContext(),
1748 RealDest
->getName()+".critedge",
1749 RealDest
->getParent(), RealDest
);
1750 BranchInst::Create(RealDest
, EdgeBB
);
1752 // Update PHI nodes.
1753 AddPredecessorToBlock(RealDest
, EdgeBB
, BB
);
1755 // BB may have instructions that are being threaded over. Clone these
1756 // instructions into EdgeBB. We know that there will be no uses of the
1757 // cloned instructions outside of EdgeBB.
1758 BasicBlock::iterator InsertPt
= EdgeBB
->begin();
1759 DenseMap
<Value
*, Value
*> TranslateMap
; // Track translated values.
1760 for (BasicBlock::iterator BBI
= BB
->begin(); &*BBI
!= BI
; ++BBI
) {
1761 if (PHINode
*PN
= dyn_cast
<PHINode
>(BBI
)) {
1762 TranslateMap
[PN
] = PN
->getIncomingValueForBlock(PredBB
);
1765 // Clone the instruction.
1766 Instruction
*N
= BBI
->clone();
1767 if (BBI
->hasName()) N
->setName(BBI
->getName()+".c");
1769 // Update operands due to translation.
1770 for (User::op_iterator i
= N
->op_begin(), e
= N
->op_end();
1772 DenseMap
<Value
*, Value
*>::iterator PI
= TranslateMap
.find(*i
);
1773 if (PI
!= TranslateMap
.end())
1777 // Check for trivial simplification.
1778 if (Value
*V
= SimplifyInstruction(N
, DL
)) {
1779 TranslateMap
[BBI
] = V
;
1780 delete N
; // Instruction folded away, don't need actual inst
1782 // Insert the new instruction into its new home.
1783 EdgeBB
->getInstList().insert(InsertPt
, N
);
1784 if (!BBI
->use_empty())
1785 TranslateMap
[BBI
] = N
;
1789 // Loop over all of the edges from PredBB to BB, changing them to branch
1790 // to EdgeBB instead.
1791 TerminatorInst
*PredBBTI
= PredBB
->getTerminator();
1792 for (unsigned i
= 0, e
= PredBBTI
->getNumSuccessors(); i
!= e
; ++i
)
1793 if (PredBBTI
->getSuccessor(i
) == BB
) {
1794 BB
->removePredecessor(PredBB
);
1795 PredBBTI
->setSuccessor(i
, EdgeBB
);
1798 // Recurse, simplifying any other constants.
1799 return FoldCondBranchOnPHI(BI
, DL
) | true;
1805 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1806 /// PHI node, see if we can eliminate it.
1807 static bool FoldTwoEntryPHINode(PHINode
*PN
, const DataLayout
*DL
) {
1808 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1809 // statement", which has a very simple dominance structure. Basically, we
1810 // are trying to find the condition that is being branched on, which
1811 // subsequently causes this merge to happen. We really want control
1812 // dependence information for this check, but simplifycfg can't keep it up
1813 // to date, and this catches most of the cases we care about anyway.
1814 BasicBlock
*BB
= PN
->getParent();
1815 BasicBlock
*IfTrue
, *IfFalse
;
1816 Value
*IfCond
= GetIfCondition(BB
, IfTrue
, IfFalse
);
1818 // Don't bother if the branch will be constant folded trivially.
1819 isa
<ConstantInt
>(IfCond
))
1822 // Okay, we found that we can merge this two-entry phi node into a select.
1823 // Doing so would require us to fold *all* two entry phi nodes in this block.
1824 // At some point this becomes non-profitable (particularly if the target
1825 // doesn't support cmov's). Only do this transformation if there are two or
1826 // fewer PHI nodes in this block.
1827 unsigned NumPhis
= 0;
1828 for (BasicBlock::iterator I
= BB
->begin(); isa
<PHINode
>(I
); ++NumPhis
, ++I
)
1832 // Loop over the PHI's seeing if we can promote them all to select
1833 // instructions. While we are at it, keep track of the instructions
1834 // that need to be moved to the dominating block.
1835 SmallPtrSet
<Instruction
*, 4> AggressiveInsts
;
1836 unsigned MaxCostVal0
= PHINodeFoldingThreshold
,
1837 MaxCostVal1
= PHINodeFoldingThreshold
;
1839 for (BasicBlock::iterator II
= BB
->begin(); isa
<PHINode
>(II
);) {
1840 PHINode
*PN
= cast
<PHINode
>(II
++);
1841 if (Value
*V
= SimplifyInstruction(PN
, DL
)) {
1842 PN
->replaceAllUsesWith(V
);
1843 PN
->eraseFromParent();
1847 if (!DominatesMergePoint(PN
->getIncomingValue(0), BB
, &AggressiveInsts
,
1849 !DominatesMergePoint(PN
->getIncomingValue(1), BB
, &AggressiveInsts
,
1854 // If we folded the first phi, PN dangles at this point. Refresh it. If
1855 // we ran out of PHIs then we simplified them all.
1856 PN
= dyn_cast
<PHINode
>(BB
->begin());
1857 if (!PN
) return true;
1859 // Don't fold i1 branches on PHIs which contain binary operators. These can
1860 // often be turned into switches and other things.
1861 if (PN
->getType()->isIntegerTy(1) &&
1862 (isa
<BinaryOperator
>(PN
->getIncomingValue(0)) ||
1863 isa
<BinaryOperator
>(PN
->getIncomingValue(1)) ||
1864 isa
<BinaryOperator
>(IfCond
)))
1867 // If we all PHI nodes are promotable, check to make sure that all
1868 // instructions in the predecessor blocks can be promoted as well. If
1869 // not, we won't be able to get rid of the control flow, so it's not
1870 // worth promoting to select instructions.
1871 BasicBlock
*DomBlock
= nullptr;
1872 BasicBlock
*IfBlock1
= PN
->getIncomingBlock(0);
1873 BasicBlock
*IfBlock2
= PN
->getIncomingBlock(1);
1874 if (cast
<BranchInst
>(IfBlock1
->getTerminator())->isConditional()) {
1877 DomBlock
= *pred_begin(IfBlock1
);
1878 for (BasicBlock::iterator I
= IfBlock1
->begin();!isa
<TerminatorInst
>(I
);++I
)
1879 if (!AggressiveInsts
.count(I
) && !isa
<DbgInfoIntrinsic
>(I
)) {
1880 // This is not an aggressive instruction that we can promote.
1881 // Because of this, we won't be able to get rid of the control
1882 // flow, so the xform is not worth it.
1887 if (cast
<BranchInst
>(IfBlock2
->getTerminator())->isConditional()) {
1890 DomBlock
= *pred_begin(IfBlock2
);
1891 for (BasicBlock::iterator I
= IfBlock2
->begin();!isa
<TerminatorInst
>(I
);++I
)
1892 if (!AggressiveInsts
.count(I
) && !isa
<DbgInfoIntrinsic
>(I
)) {
1893 // This is not an aggressive instruction that we can promote.
1894 // Because of this, we won't be able to get rid of the control
1895 // flow, so the xform is not worth it.
1900 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
<< " T: "
1901 << IfTrue
->getName() << " F: " << IfFalse
->getName() << "\n");
1903 // If we can still promote the PHI nodes after this gauntlet of tests,
1904 // do all of the PHI's now.
1905 Instruction
*InsertPt
= DomBlock
->getTerminator();
1906 IRBuilder
<true, NoFolder
> Builder(InsertPt
);
1908 // Move all 'aggressive' instructions, which are defined in the
1909 // conditional parts of the if's up to the dominating block.
1911 DomBlock
->getInstList().splice(InsertPt
,
1912 IfBlock1
->getInstList(), IfBlock1
->begin(),
1913 IfBlock1
->getTerminator());
1915 DomBlock
->getInstList().splice(InsertPt
,
1916 IfBlock2
->getInstList(), IfBlock2
->begin(),
1917 IfBlock2
->getTerminator());
1919 while (PHINode
*PN
= dyn_cast
<PHINode
>(BB
->begin())) {
1920 // Change the PHI node into a select instruction.
1921 Value
*TrueVal
= PN
->getIncomingValue(PN
->getIncomingBlock(0) == IfFalse
);
1922 Value
*FalseVal
= PN
->getIncomingValue(PN
->getIncomingBlock(0) == IfTrue
);
1925 cast
<SelectInst
>(Builder
.CreateSelect(IfCond
, TrueVal
, FalseVal
, ""));
1926 PN
->replaceAllUsesWith(NV
);
1928 PN
->eraseFromParent();
1931 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1932 // has been flattened. Change DomBlock to jump directly to our new block to
1933 // avoid other simplifycfg's kicking in on the diamond.
1934 TerminatorInst
*OldTI
= DomBlock
->getTerminator();
1935 Builder
.SetInsertPoint(OldTI
);
1936 Builder
.CreateBr(BB
);
1937 OldTI
->eraseFromParent();
1941 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1942 /// to two returning blocks, try to merge them together into one return,
1943 /// introducing a select if the return values disagree.
1944 static bool SimplifyCondBranchToTwoReturns(BranchInst
*BI
,
1945 IRBuilder
<> &Builder
) {
1946 assert(BI
->isConditional() && "Must be a conditional branch");
1947 BasicBlock
*TrueSucc
= BI
->getSuccessor(0);
1948 BasicBlock
*FalseSucc
= BI
->getSuccessor(1);
1949 ReturnInst
*TrueRet
= cast
<ReturnInst
>(TrueSucc
->getTerminator());
1950 ReturnInst
*FalseRet
= cast
<ReturnInst
>(FalseSucc
->getTerminator());
1952 // Check to ensure both blocks are empty (just a return) or optionally empty
1953 // with PHI nodes. If there are other instructions, merging would cause extra
1954 // computation on one path or the other.
1955 if (!TrueSucc
->getFirstNonPHIOrDbg()->isTerminator())
1957 if (!FalseSucc
->getFirstNonPHIOrDbg()->isTerminator())
1960 Builder
.SetInsertPoint(BI
);
1961 // Okay, we found a branch that is going to two return nodes. If
1962 // there is no return value for this function, just change the
1963 // branch into a return.
1964 if (FalseRet
->getNumOperands() == 0) {
1965 TrueSucc
->removePredecessor(BI
->getParent());
1966 FalseSucc
->removePredecessor(BI
->getParent());
1967 Builder
.CreateRetVoid();
1968 EraseTerminatorInstAndDCECond(BI
);
1972 // Otherwise, figure out what the true and false return values are
1973 // so we can insert a new select instruction.
1974 Value
*TrueValue
= TrueRet
->getReturnValue();
1975 Value
*FalseValue
= FalseRet
->getReturnValue();
1977 // Unwrap any PHI nodes in the return blocks.
1978 if (PHINode
*TVPN
= dyn_cast_or_null
<PHINode
>(TrueValue
))
1979 if (TVPN
->getParent() == TrueSucc
)
1980 TrueValue
= TVPN
->getIncomingValueForBlock(BI
->getParent());
1981 if (PHINode
*FVPN
= dyn_cast_or_null
<PHINode
>(FalseValue
))
1982 if (FVPN
->getParent() == FalseSucc
)
1983 FalseValue
= FVPN
->getIncomingValueForBlock(BI
->getParent());
1985 // In order for this transformation to be safe, we must be able to
1986 // unconditionally execute both operands to the return. This is
1987 // normally the case, but we could have a potentially-trapping
1988 // constant expression that prevents this transformation from being
1990 if (ConstantExpr
*TCV
= dyn_cast_or_null
<ConstantExpr
>(TrueValue
))
1993 if (ConstantExpr
*FCV
= dyn_cast_or_null
<ConstantExpr
>(FalseValue
))
1997 // Okay, we collected all the mapped values and checked them for sanity, and
1998 // defined to really do this transformation. First, update the CFG.
1999 TrueSucc
->removePredecessor(BI
->getParent());
2000 FalseSucc
->removePredecessor(BI
->getParent());
2002 // Insert select instructions where needed.
2003 Value
*BrCond
= BI
->getCondition();
2005 // Insert a select if the results differ.
2006 if (TrueValue
== FalseValue
|| isa
<UndefValue
>(FalseValue
)) {
2007 } else if (isa
<UndefValue
>(TrueValue
)) {
2008 TrueValue
= FalseValue
;
2010 TrueValue
= Builder
.CreateSelect(BrCond
, TrueValue
,
2011 FalseValue
, "retval");
2015 Value
*RI
= !TrueValue
?
2016 Builder
.CreateRetVoid() : Builder
.CreateRet(TrueValue
);
2020 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2021 << "\n " << *BI
<< "NewRet = " << *RI
2022 << "TRUEBLOCK: " << *TrueSucc
<< "FALSEBLOCK: "<< *FalseSucc
);
2024 EraseTerminatorInstAndDCECond(BI
);
2029 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
2030 /// probabilities of the branch taking each edge. Fills in the two APInt
2031 /// parameters and return true, or returns false if no or invalid metadata was
2033 static bool ExtractBranchMetadata(BranchInst
*BI
,
2034 uint64_t &ProbTrue
, uint64_t &ProbFalse
) {
2035 assert(BI
->isConditional() &&
2036 "Looking for probabilities on unconditional branch?");
2037 MDNode
*ProfileData
= BI
->getMetadata(LLVMContext::MD_prof
);
2038 if (!ProfileData
|| ProfileData
->getNumOperands() != 3) return false;
2039 ConstantInt
*CITrue
=
2040 mdconst::dyn_extract
<ConstantInt
>(ProfileData
->getOperand(1));
2041 ConstantInt
*CIFalse
=
2042 mdconst::dyn_extract
<ConstantInt
>(ProfileData
->getOperand(2));
2043 if (!CITrue
|| !CIFalse
) return false;
2044 ProbTrue
= CITrue
->getValue().getZExtValue();
2045 ProbFalse
= CIFalse
->getValue().getZExtValue();
2049 /// checkCSEInPredecessor - Return true if the given instruction is available
2050 /// in its predecessor block. If yes, the instruction will be removed.
2052 static bool checkCSEInPredecessor(Instruction
*Inst
, BasicBlock
*PB
) {
2053 if (!isa
<BinaryOperator
>(Inst
) && !isa
<CmpInst
>(Inst
))
2055 for (BasicBlock::iterator I
= PB
->begin(), E
= PB
->end(); I
!= E
; I
++) {
2056 Instruction
*PBI
= &*I
;
2057 // Check whether Inst and PBI generate the same value.
2058 if (Inst
->isIdenticalTo(PBI
)) {
2059 Inst
->replaceAllUsesWith(PBI
);
2060 Inst
->eraseFromParent();
2067 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
2068 /// predecessor branches to us and one of our successors, fold the block into
2069 /// the predecessor and use logical operations to pick the right destination.
2070 bool llvm::FoldBranchToCommonDest(BranchInst
*BI
, const DataLayout
*DL
,
2071 unsigned BonusInstThreshold
) {
2072 BasicBlock
*BB
= BI
->getParent();
2074 Instruction
*Cond
= nullptr;
2075 if (BI
->isConditional())
2076 Cond
= dyn_cast
<Instruction
>(BI
->getCondition());
2078 // For unconditional branch, check for a simple CFG pattern, where
2079 // BB has a single predecessor and BB's successor is also its predecessor's
2080 // successor. If such pattern exisits, check for CSE between BB and its
2082 if (BasicBlock
*PB
= BB
->getSinglePredecessor())
2083 if (BranchInst
*PBI
= dyn_cast
<BranchInst
>(PB
->getTerminator()))
2084 if (PBI
->isConditional() &&
2085 (BI
->getSuccessor(0) == PBI
->getSuccessor(0) ||
2086 BI
->getSuccessor(0) == PBI
->getSuccessor(1))) {
2087 for (BasicBlock::iterator I
= BB
->begin(), E
= BB
->end();
2089 Instruction
*Curr
= I
++;
2090 if (isa
<CmpInst
>(Curr
)) {
2094 // Quit if we can't remove this instruction.
2095 if (!checkCSEInPredecessor(Curr
, PB
))
2104 if (!Cond
|| (!isa
<CmpInst
>(Cond
) && !isa
<BinaryOperator
>(Cond
)) ||
2105 Cond
->getParent() != BB
|| !Cond
->hasOneUse())
2108 // Make sure the instruction after the condition is the cond branch.
2109 BasicBlock::iterator CondIt
= Cond
; ++CondIt
;
2111 // Ignore dbg intrinsics.
2112 while (isa
<DbgInfoIntrinsic
>(CondIt
)) ++CondIt
;
2117 // Only allow this transformation if computing the condition doesn't involve
2118 // too many instructions and these involved instructions can be executed
2119 // unconditionally. We denote all involved instructions except the condition
2120 // as "bonus instructions", and only allow this transformation when the
2121 // number of the bonus instructions does not exceed a certain threshold.
2122 unsigned NumBonusInsts
= 0;
2123 for (auto I
= BB
->begin(); Cond
!= I
; ++I
) {
2124 // Ignore dbg intrinsics.
2125 if (isa
<DbgInfoIntrinsic
>(I
))
2127 if (!I
->hasOneUse() || !isSafeToSpeculativelyExecute(I
, DL
))
2129 // I has only one use and can be executed unconditionally.
2130 Instruction
*User
= dyn_cast
<Instruction
>(I
->user_back());
2131 if (User
== nullptr || User
->getParent() != BB
)
2133 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2134 // to use any other instruction, User must be an instruction between next(I)
2137 // Early exits once we reach the limit.
2138 if (NumBonusInsts
> BonusInstThreshold
)
2142 // Cond is known to be a compare or binary operator. Check to make sure that
2143 // neither operand is a potentially-trapping constant expression.
2144 if (ConstantExpr
*CE
= dyn_cast
<ConstantExpr
>(Cond
->getOperand(0)))
2147 if (ConstantExpr
*CE
= dyn_cast
<ConstantExpr
>(Cond
->getOperand(1)))
2151 // Finally, don't infinitely unroll conditional loops.
2152 BasicBlock
*TrueDest
= BI
->getSuccessor(0);
2153 BasicBlock
*FalseDest
= (BI
->isConditional()) ? BI
->getSuccessor(1) : nullptr;
2154 if (TrueDest
== BB
|| FalseDest
== BB
)
2157 for (pred_iterator PI
= pred_begin(BB
), E
= pred_end(BB
); PI
!= E
; ++PI
) {
2158 BasicBlock
*PredBlock
= *PI
;
2159 BranchInst
*PBI
= dyn_cast
<BranchInst
>(PredBlock
->getTerminator());
2161 // Check that we have two conditional branches. If there is a PHI node in
2162 // the common successor, verify that the same value flows in from both
2164 SmallVector
<PHINode
*, 4> PHIs
;
2165 if (!PBI
|| PBI
->isUnconditional() ||
2166 (BI
->isConditional() &&
2167 !SafeToMergeTerminators(BI
, PBI
)) ||
2168 (!BI
->isConditional() &&
2169 !isProfitableToFoldUnconditional(BI
, PBI
, Cond
, PHIs
)))
2172 // Determine if the two branches share a common destination.
2173 Instruction::BinaryOps Opc
= Instruction::BinaryOpsEnd
;
2174 bool InvertPredCond
= false;
2176 if (BI
->isConditional()) {
2177 if (PBI
->getSuccessor(0) == TrueDest
)
2178 Opc
= Instruction::Or
;
2179 else if (PBI
->getSuccessor(1) == FalseDest
)
2180 Opc
= Instruction::And
;
2181 else if (PBI
->getSuccessor(0) == FalseDest
)
2182 Opc
= Instruction::And
, InvertPredCond
= true;
2183 else if (PBI
->getSuccessor(1) == TrueDest
)
2184 Opc
= Instruction::Or
, InvertPredCond
= true;
2188 if (PBI
->getSuccessor(0) != TrueDest
&& PBI
->getSuccessor(1) != TrueDest
)
2192 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI
<< *BB
);
2193 IRBuilder
<> Builder(PBI
);
2195 // If we need to invert the condition in the pred block to match, do so now.
2196 if (InvertPredCond
) {
2197 Value
*NewCond
= PBI
->getCondition();
2199 if (NewCond
->hasOneUse() && isa
<CmpInst
>(NewCond
)) {
2200 CmpInst
*CI
= cast
<CmpInst
>(NewCond
);
2201 CI
->setPredicate(CI
->getInversePredicate());
2203 NewCond
= Builder
.CreateNot(NewCond
,
2204 PBI
->getCondition()->getName()+".not");
2207 PBI
->setCondition(NewCond
);
2208 PBI
->swapSuccessors();
2211 // If we have bonus instructions, clone them into the predecessor block.
2212 // Note that there may be mutliple predecessor blocks, so we cannot move
2213 // bonus instructions to a predecessor block.
2214 ValueToValueMapTy VMap
; // maps original values to cloned values
2215 // We already make sure Cond is the last instruction before BI. Therefore,
2216 // every instructions before Cond other than DbgInfoIntrinsic are bonus
2218 for (auto BonusInst
= BB
->begin(); Cond
!= BonusInst
; ++BonusInst
) {
2219 if (isa
<DbgInfoIntrinsic
>(BonusInst
))
2221 Instruction
*NewBonusInst
= BonusInst
->clone();
2222 RemapInstruction(NewBonusInst
, VMap
,
2223 RF_NoModuleLevelChanges
| RF_IgnoreMissingEntries
);
2224 VMap
[BonusInst
] = NewBonusInst
;
2226 // If we moved a load, we cannot any longer claim any knowledge about
2227 // its potential value. The previous information might have been valid
2228 // only given the branch precondition.
2229 // For an analogous reason, we must also drop all the metadata whose
2230 // semantics we don't understand.
2231 NewBonusInst
->dropUnknownMetadata(LLVMContext::MD_dbg
);
2233 PredBlock
->getInstList().insert(PBI
, NewBonusInst
);
2234 NewBonusInst
->takeName(BonusInst
);
2235 BonusInst
->setName(BonusInst
->getName() + ".old");
2238 // Clone Cond into the predecessor basic block, and or/and the
2239 // two conditions together.
2240 Instruction
*New
= Cond
->clone();
2241 RemapInstruction(New
, VMap
,
2242 RF_NoModuleLevelChanges
| RF_IgnoreMissingEntries
);
2243 PredBlock
->getInstList().insert(PBI
, New
);
2244 New
->takeName(Cond
);
2245 Cond
->setName(New
->getName() + ".old");
2247 if (BI
->isConditional()) {
2248 Instruction
*NewCond
=
2249 cast
<Instruction
>(Builder
.CreateBinOp(Opc
, PBI
->getCondition(),
2251 PBI
->setCondition(NewCond
);
2253 uint64_t PredTrueWeight
, PredFalseWeight
, SuccTrueWeight
, SuccFalseWeight
;
2254 bool PredHasWeights
= ExtractBranchMetadata(PBI
, PredTrueWeight
,
2256 bool SuccHasWeights
= ExtractBranchMetadata(BI
, SuccTrueWeight
,
2258 SmallVector
<uint64_t, 8> NewWeights
;
2260 if (PBI
->getSuccessor(0) == BB
) {
2261 if (PredHasWeights
&& SuccHasWeights
) {
2262 // PBI: br i1 %x, BB, FalseDest
2263 // BI: br i1 %y, TrueDest, FalseDest
2264 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2265 NewWeights
.push_back(PredTrueWeight
* SuccTrueWeight
);
2266 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2267 // TrueWeight for PBI * FalseWeight for BI.
2268 // We assume that total weights of a BranchInst can fit into 32 bits.
2269 // Therefore, we will not have overflow using 64-bit arithmetic.
2270 NewWeights
.push_back(PredFalseWeight
* (SuccFalseWeight
+
2271 SuccTrueWeight
) + PredTrueWeight
* SuccFalseWeight
);
2273 AddPredecessorToBlock(TrueDest
, PredBlock
, BB
);
2274 PBI
->setSuccessor(0, TrueDest
);
2276 if (PBI
->getSuccessor(1) == BB
) {
2277 if (PredHasWeights
&& SuccHasWeights
) {
2278 // PBI: br i1 %x, TrueDest, BB
2279 // BI: br i1 %y, TrueDest, FalseDest
2280 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2281 // FalseWeight for PBI * TrueWeight for BI.
2282 NewWeights
.push_back(PredTrueWeight
* (SuccFalseWeight
+
2283 SuccTrueWeight
) + PredFalseWeight
* SuccTrueWeight
);
2284 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2285 NewWeights
.push_back(PredFalseWeight
* SuccFalseWeight
);
2287 AddPredecessorToBlock(FalseDest
, PredBlock
, BB
);
2288 PBI
->setSuccessor(1, FalseDest
);
2290 if (NewWeights
.size() == 2) {
2291 // Halve the weights if any of them cannot fit in an uint32_t
2292 FitWeights(NewWeights
);
2294 SmallVector
<uint32_t, 8> MDWeights(NewWeights
.begin(),NewWeights
.end());
2295 PBI
->setMetadata(LLVMContext::MD_prof
,
2296 MDBuilder(BI
->getContext()).
2297 createBranchWeights(MDWeights
));
2299 PBI
->setMetadata(LLVMContext::MD_prof
, nullptr);
2301 // Update PHI nodes in the common successors.
2302 for (unsigned i
= 0, e
= PHIs
.size(); i
!= e
; ++i
) {
2303 ConstantInt
*PBI_C
= cast
<ConstantInt
>(
2304 PHIs
[i
]->getIncomingValueForBlock(PBI
->getParent()));
2305 assert(PBI_C
->getType()->isIntegerTy(1));
2306 Instruction
*MergedCond
= nullptr;
2307 if (PBI
->getSuccessor(0) == TrueDest
) {
2308 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2309 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2310 // is false: !PBI_Cond and BI_Value
2311 Instruction
*NotCond
=
2312 cast
<Instruction
>(Builder
.CreateNot(PBI
->getCondition(),
2315 cast
<Instruction
>(Builder
.CreateBinOp(Instruction::And
,
2320 cast
<Instruction
>(Builder
.CreateBinOp(Instruction::Or
,
2321 PBI
->getCondition(), MergedCond
,
2324 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2325 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2326 // is false: PBI_Cond and BI_Value
2328 cast
<Instruction
>(Builder
.CreateBinOp(Instruction::And
,
2329 PBI
->getCondition(), New
,
2331 if (PBI_C
->isOne()) {
2332 Instruction
*NotCond
=
2333 cast
<Instruction
>(Builder
.CreateNot(PBI
->getCondition(),
2336 cast
<Instruction
>(Builder
.CreateBinOp(Instruction::Or
,
2337 NotCond
, MergedCond
,
2342 PHIs
[i
]->setIncomingValue(PHIs
[i
]->getBasicBlockIndex(PBI
->getParent()),
2345 // Change PBI from Conditional to Unconditional.
2346 BranchInst
*New_PBI
= BranchInst::Create(TrueDest
, PBI
);
2347 EraseTerminatorInstAndDCECond(PBI
);
2351 // TODO: If BB is reachable from all paths through PredBlock, then we
2352 // could replace PBI's branch probabilities with BI's.
2354 // Copy any debug value intrinsics into the end of PredBlock.
2355 for (BasicBlock::iterator I
= BB
->begin(), E
= BB
->end(); I
!= E
; ++I
)
2356 if (isa
<DbgInfoIntrinsic
>(*I
))
2357 I
->clone()->insertBefore(PBI
);
2364 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2365 /// predecessor of another block, this function tries to simplify it. We know
2366 /// that PBI and BI are both conditional branches, and BI is in one of the
2367 /// successor blocks of PBI - PBI branches to BI.
2368 static bool SimplifyCondBranchToCondBranch(BranchInst
*PBI
, BranchInst
*BI
) {
2369 assert(PBI
->isConditional() && BI
->isConditional());
2370 BasicBlock
*BB
= BI
->getParent();
2372 // If this block ends with a branch instruction, and if there is a
2373 // predecessor that ends on a branch of the same condition, make
2374 // this conditional branch redundant.
2375 if (PBI
->getCondition() == BI
->getCondition() &&
2376 PBI
->getSuccessor(0) != PBI
->getSuccessor(1)) {
2377 // Okay, the outcome of this conditional branch is statically
2378 // knowable. If this block had a single pred, handle specially.
2379 if (BB
->getSinglePredecessor()) {
2380 // Turn this into a branch on constant.
2381 bool CondIsTrue
= PBI
->getSuccessor(0) == BB
;
2382 BI
->setCondition(ConstantInt::get(Type::getInt1Ty(BB
->getContext()),
2384 return true; // Nuke the branch on constant.
2387 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2388 // in the constant and simplify the block result. Subsequent passes of
2389 // simplifycfg will thread the block.
2390 if (BlockIsSimpleEnoughToThreadThrough(BB
)) {
2391 pred_iterator PB
= pred_begin(BB
), PE
= pred_end(BB
);
2392 PHINode
*NewPN
= PHINode::Create(Type::getInt1Ty(BB
->getContext()),
2393 std::distance(PB
, PE
),
2394 BI
->getCondition()->getName() + ".pr",
2396 // Okay, we're going to insert the PHI node. Since PBI is not the only
2397 // predecessor, compute the PHI'd conditional value for all of the preds.
2398 // Any predecessor where the condition is not computable we keep symbolic.
2399 for (pred_iterator PI
= PB
; PI
!= PE
; ++PI
) {
2400 BasicBlock
*P
= *PI
;
2401 if ((PBI
= dyn_cast
<BranchInst
>(P
->getTerminator())) &&
2402 PBI
!= BI
&& PBI
->isConditional() &&
2403 PBI
->getCondition() == BI
->getCondition() &&
2404 PBI
->getSuccessor(0) != PBI
->getSuccessor(1)) {
2405 bool CondIsTrue
= PBI
->getSuccessor(0) == BB
;
2406 NewPN
->addIncoming(ConstantInt::get(Type::getInt1Ty(BB
->getContext()),
2409 NewPN
->addIncoming(BI
->getCondition(), P
);
2413 BI
->setCondition(NewPN
);
2418 // If this is a conditional branch in an empty block, and if any
2419 // predecessors are a conditional branch to one of our destinations,
2420 // fold the conditions into logical ops and one cond br.
2421 BasicBlock::iterator BBI
= BB
->begin();
2422 // Ignore dbg intrinsics.
2423 while (isa
<DbgInfoIntrinsic
>(BBI
))
2429 if (ConstantExpr
*CE
= dyn_cast
<ConstantExpr
>(BI
->getCondition()))
2434 if (PBI
->getSuccessor(0) == BI
->getSuccessor(0))
2436 else if (PBI
->getSuccessor(0) == BI
->getSuccessor(1))
2437 PBIOp
= 0, BIOp
= 1;
2438 else if (PBI
->getSuccessor(1) == BI
->getSuccessor(0))
2439 PBIOp
= 1, BIOp
= 0;
2440 else if (PBI
->getSuccessor(1) == BI
->getSuccessor(1))
2445 // Check to make sure that the other destination of this branch
2446 // isn't BB itself. If so, this is an infinite loop that will
2447 // keep getting unwound.
2448 if (PBI
->getSuccessor(PBIOp
) == BB
)
2451 // Do not perform this transformation if it would require
2452 // insertion of a large number of select instructions. For targets
2453 // without predication/cmovs, this is a big pessimization.
2455 // Also do not perform this transformation if any phi node in the common
2456 // destination block can trap when reached by BB or PBB (PR17073). In that
2457 // case, it would be unsafe to hoist the operation into a select instruction.
2459 BasicBlock
*CommonDest
= PBI
->getSuccessor(PBIOp
);
2460 unsigned NumPhis
= 0;
2461 for (BasicBlock::iterator II
= CommonDest
->begin();
2462 isa
<PHINode
>(II
); ++II
, ++NumPhis
) {
2463 if (NumPhis
> 2) // Disable this xform.
2466 PHINode
*PN
= cast
<PHINode
>(II
);
2467 Value
*BIV
= PN
->getIncomingValueForBlock(BB
);
2468 if (ConstantExpr
*CE
= dyn_cast
<ConstantExpr
>(BIV
))
2472 unsigned PBBIdx
= PN
->getBasicBlockIndex(PBI
->getParent());
2473 Value
*PBIV
= PN
->getIncomingValue(PBBIdx
);
2474 if (ConstantExpr
*CE
= dyn_cast
<ConstantExpr
>(PBIV
))
2479 // Finally, if everything is ok, fold the branches to logical ops.
2480 BasicBlock
*OtherDest
= BI
->getSuccessor(BIOp
^ 1);
2482 DEBUG(dbgs() << "FOLDING BRs:" << *PBI
->getParent()
2483 << "AND: " << *BI
->getParent());
2486 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2487 // branch in it, where one edge (OtherDest) goes back to itself but the other
2488 // exits. We don't *know* that the program avoids the infinite loop
2489 // (even though that seems likely). If we do this xform naively, we'll end up
2490 // recursively unpeeling the loop. Since we know that (after the xform is
2491 // done) that the block *is* infinite if reached, we just make it an obviously
2492 // infinite loop with no cond branch.
2493 if (OtherDest
== BB
) {
2494 // Insert it at the end of the function, because it's either code,
2495 // or it won't matter if it's hot. :)
2496 BasicBlock
*InfLoopBlock
= BasicBlock::Create(BB
->getContext(),
2497 "infloop", BB
->getParent());
2498 BranchInst::Create(InfLoopBlock
, InfLoopBlock
);
2499 OtherDest
= InfLoopBlock
;
2502 DEBUG(dbgs() << *PBI
->getParent()->getParent());
2504 // BI may have other predecessors. Because of this, we leave
2505 // it alone, but modify PBI.
2507 // Make sure we get to CommonDest on True&True directions.
2508 Value
*PBICond
= PBI
->getCondition();
2509 IRBuilder
<true, NoFolder
> Builder(PBI
);
2511 PBICond
= Builder
.CreateNot(PBICond
, PBICond
->getName()+".not");
2513 Value
*BICond
= BI
->getCondition();
2515 BICond
= Builder
.CreateNot(BICond
, BICond
->getName()+".not");
2517 // Merge the conditions.
2518 Value
*Cond
= Builder
.CreateOr(PBICond
, BICond
, "brmerge");
2520 // Modify PBI to branch on the new condition to the new dests.
2521 PBI
->setCondition(Cond
);
2522 PBI
->setSuccessor(0, CommonDest
);
2523 PBI
->setSuccessor(1, OtherDest
);
2525 // Update branch weight for PBI.
2526 uint64_t PredTrueWeight
, PredFalseWeight
, SuccTrueWeight
, SuccFalseWeight
;
2527 bool PredHasWeights
= ExtractBranchMetadata(PBI
, PredTrueWeight
,
2529 bool SuccHasWeights
= ExtractBranchMetadata(BI
, SuccTrueWeight
,
2531 if (PredHasWeights
&& SuccHasWeights
) {
2532 uint64_t PredCommon
= PBIOp
? PredFalseWeight
: PredTrueWeight
;
2533 uint64_t PredOther
= PBIOp
?PredTrueWeight
: PredFalseWeight
;
2534 uint64_t SuccCommon
= BIOp
? SuccFalseWeight
: SuccTrueWeight
;
2535 uint64_t SuccOther
= BIOp
? SuccTrueWeight
: SuccFalseWeight
;
2536 // The weight to CommonDest should be PredCommon * SuccTotal +
2537 // PredOther * SuccCommon.
2538 // The weight to OtherDest should be PredOther * SuccOther.
2539 SmallVector
<uint64_t, 2> NewWeights
;
2540 NewWeights
.push_back(PredCommon
* (SuccCommon
+ SuccOther
) +
2541 PredOther
* SuccCommon
);
2542 NewWeights
.push_back(PredOther
* SuccOther
);
2543 // Halve the weights if any of them cannot fit in an uint32_t
2544 FitWeights(NewWeights
);
2546 SmallVector
<uint32_t, 2> MDWeights(NewWeights
.begin(),NewWeights
.end());
2547 PBI
->setMetadata(LLVMContext::MD_prof
,
2548 MDBuilder(BI
->getContext()).
2549 createBranchWeights(MDWeights
));
2552 // OtherDest may have phi nodes. If so, add an entry from PBI's
2553 // block that are identical to the entries for BI's block.
2554 AddPredecessorToBlock(OtherDest
, PBI
->getParent(), BB
);
2556 // We know that the CommonDest already had an edge from PBI to
2557 // it. If it has PHIs though, the PHIs may have different
2558 // entries for BB and PBI's BB. If so, insert a select to make
2561 for (BasicBlock::iterator II
= CommonDest
->begin();
2562 (PN
= dyn_cast
<PHINode
>(II
)); ++II
) {
2563 Value
*BIV
= PN
->getIncomingValueForBlock(BB
);
2564 unsigned PBBIdx
= PN
->getBasicBlockIndex(PBI
->getParent());
2565 Value
*PBIV
= PN
->getIncomingValue(PBBIdx
);
2567 // Insert a select in PBI to pick the right value.
2568 Value
*NV
= cast
<SelectInst
>
2569 (Builder
.CreateSelect(PBICond
, PBIV
, BIV
, PBIV
->getName()+".mux"));
2570 PN
->setIncomingValue(PBBIdx
, NV
);
2574 DEBUG(dbgs() << "INTO: " << *PBI
->getParent());
2575 DEBUG(dbgs() << *PBI
->getParent()->getParent());
2577 // This basic block is probably dead. We know it has at least
2578 // one fewer predecessor.
2582 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2583 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2584 // Takes care of updating the successors and removing the old terminator.
2585 // Also makes sure not to introduce new successors by assuming that edges to
2586 // non-successor TrueBBs and FalseBBs aren't reachable.
2587 static bool SimplifyTerminatorOnSelect(TerminatorInst
*OldTerm
, Value
*Cond
,
2588 BasicBlock
*TrueBB
, BasicBlock
*FalseBB
,
2589 uint32_t TrueWeight
,
2590 uint32_t FalseWeight
){
2591 // Remove any superfluous successor edges from the CFG.
2592 // First, figure out which successors to preserve.
2593 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2595 BasicBlock
*KeepEdge1
= TrueBB
;
2596 BasicBlock
*KeepEdge2
= TrueBB
!= FalseBB
? FalseBB
: nullptr;
2598 // Then remove the rest.
2599 for (unsigned I
= 0, E
= OldTerm
->getNumSuccessors(); I
!= E
; ++I
) {
2600 BasicBlock
*Succ
= OldTerm
->getSuccessor(I
);
2601 // Make sure only to keep exactly one copy of each edge.
2602 if (Succ
== KeepEdge1
)
2603 KeepEdge1
= nullptr;
2604 else if (Succ
== KeepEdge2
)
2605 KeepEdge2
= nullptr;
2607 Succ
->removePredecessor(OldTerm
->getParent());
2610 IRBuilder
<> Builder(OldTerm
);
2611 Builder
.SetCurrentDebugLocation(OldTerm
->getDebugLoc());
2613 // Insert an appropriate new terminator.
2614 if (!KeepEdge1
&& !KeepEdge2
) {
2615 if (TrueBB
== FalseBB
)
2616 // We were only looking for one successor, and it was present.
2617 // Create an unconditional branch to it.
2618 Builder
.CreateBr(TrueBB
);
2620 // We found both of the successors we were looking for.
2621 // Create a conditional branch sharing the condition of the select.
2622 BranchInst
*NewBI
= Builder
.CreateCondBr(Cond
, TrueBB
, FalseBB
);
2623 if (TrueWeight
!= FalseWeight
)
2624 NewBI
->setMetadata(LLVMContext::MD_prof
,
2625 MDBuilder(OldTerm
->getContext()).
2626 createBranchWeights(TrueWeight
, FalseWeight
));
2628 } else if (KeepEdge1
&& (KeepEdge2
|| TrueBB
== FalseBB
)) {
2629 // Neither of the selected blocks were successors, so this
2630 // terminator must be unreachable.
2631 new UnreachableInst(OldTerm
->getContext(), OldTerm
);
2633 // One of the selected values was a successor, but the other wasn't.
2634 // Insert an unconditional branch to the one that was found;
2635 // the edge to the one that wasn't must be unreachable.
2637 // Only TrueBB was found.
2638 Builder
.CreateBr(TrueBB
);
2640 // Only FalseBB was found.
2641 Builder
.CreateBr(FalseBB
);
2644 EraseTerminatorInstAndDCECond(OldTerm
);
2648 // SimplifySwitchOnSelect - Replaces
2649 // (switch (select cond, X, Y)) on constant X, Y
2650 // with a branch - conditional if X and Y lead to distinct BBs,
2651 // unconditional otherwise.
2652 static bool SimplifySwitchOnSelect(SwitchInst
*SI
, SelectInst
*Select
) {
2653 // Check for constant integer values in the select.
2654 ConstantInt
*TrueVal
= dyn_cast
<ConstantInt
>(Select
->getTrueValue());
2655 ConstantInt
*FalseVal
= dyn_cast
<ConstantInt
>(Select
->getFalseValue());
2656 if (!TrueVal
|| !FalseVal
)
2659 // Find the relevant condition and destinations.
2660 Value
*Condition
= Select
->getCondition();
2661 BasicBlock
*TrueBB
= SI
->findCaseValue(TrueVal
).getCaseSuccessor();
2662 BasicBlock
*FalseBB
= SI
->findCaseValue(FalseVal
).getCaseSuccessor();
2664 // Get weight for TrueBB and FalseBB.
2665 uint32_t TrueWeight
= 0, FalseWeight
= 0;
2666 SmallVector
<uint64_t, 8> Weights
;
2667 bool HasWeights
= HasBranchWeights(SI
);
2669 GetBranchWeights(SI
, Weights
);
2670 if (Weights
.size() == 1 + SI
->getNumCases()) {
2671 TrueWeight
= (uint32_t)Weights
[SI
->findCaseValue(TrueVal
).
2672 getSuccessorIndex()];
2673 FalseWeight
= (uint32_t)Weights
[SI
->findCaseValue(FalseVal
).
2674 getSuccessorIndex()];
2678 // Perform the actual simplification.
2679 return SimplifyTerminatorOnSelect(SI
, Condition
, TrueBB
, FalseBB
,
2680 TrueWeight
, FalseWeight
);
2683 // SimplifyIndirectBrOnSelect - Replaces
2684 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2685 // blockaddress(@fn, BlockB)))
2687 // (br cond, BlockA, BlockB).
2688 static bool SimplifyIndirectBrOnSelect(IndirectBrInst
*IBI
, SelectInst
*SI
) {
2689 // Check that both operands of the select are block addresses.
2690 BlockAddress
*TBA
= dyn_cast
<BlockAddress
>(SI
->getTrueValue());
2691 BlockAddress
*FBA
= dyn_cast
<BlockAddress
>(SI
->getFalseValue());
2695 // Extract the actual blocks.
2696 BasicBlock
*TrueBB
= TBA
->getBasicBlock();
2697 BasicBlock
*FalseBB
= FBA
->getBasicBlock();
2699 // Perform the actual simplification.
2700 return SimplifyTerminatorOnSelect(IBI
, SI
->getCondition(), TrueBB
, FalseBB
,
2704 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2705 /// instruction (a seteq/setne with a constant) as the only instruction in a
2706 /// block that ends with an uncond branch. We are looking for a very specific
2707 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2708 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2709 /// default value goes to an uncond block with a seteq in it, we get something
2712 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2714 /// %tmp = icmp eq i8 %A, 92
2717 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2719 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2720 /// the PHI, merging the third icmp into the switch.
2721 static bool TryToSimplifyUncondBranchWithICmpInIt(
2722 ICmpInst
*ICI
, IRBuilder
<> &Builder
, const TargetTransformInfo
&TTI
,
2723 unsigned BonusInstThreshold
, const DataLayout
*DL
, AssumptionCache
*AC
) {
2724 BasicBlock
*BB
= ICI
->getParent();
2726 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2728 if (isa
<PHINode
>(BB
->begin()) || !ICI
->hasOneUse()) return false;
2730 Value
*V
= ICI
->getOperand(0);
2731 ConstantInt
*Cst
= cast
<ConstantInt
>(ICI
->getOperand(1));
2733 // The pattern we're looking for is where our only predecessor is a switch on
2734 // 'V' and this block is the default case for the switch. In this case we can
2735 // fold the compared value into the switch to simplify things.
2736 BasicBlock
*Pred
= BB
->getSinglePredecessor();
2737 if (!Pred
|| !isa
<SwitchInst
>(Pred
->getTerminator())) return false;
2739 SwitchInst
*SI
= cast
<SwitchInst
>(Pred
->getTerminator());
2740 if (SI
->getCondition() != V
)
2743 // If BB is reachable on a non-default case, then we simply know the value of
2744 // V in this block. Substitute it and constant fold the icmp instruction
2746 if (SI
->getDefaultDest() != BB
) {
2747 ConstantInt
*VVal
= SI
->findCaseDest(BB
);
2748 assert(VVal
&& "Should have a unique destination value");
2749 ICI
->setOperand(0, VVal
);
2751 if (Value
*V
= SimplifyInstruction(ICI
, DL
)) {
2752 ICI
->replaceAllUsesWith(V
);
2753 ICI
->eraseFromParent();
2755 // BB is now empty, so it is likely to simplify away.
2756 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
2759 // Ok, the block is reachable from the default dest. If the constant we're
2760 // comparing exists in one of the other edges, then we can constant fold ICI
2762 if (SI
->findCaseValue(Cst
) != SI
->case_default()) {
2764 if (ICI
->getPredicate() == ICmpInst::ICMP_EQ
)
2765 V
= ConstantInt::getFalse(BB
->getContext());
2767 V
= ConstantInt::getTrue(BB
->getContext());
2769 ICI
->replaceAllUsesWith(V
);
2770 ICI
->eraseFromParent();
2771 // BB is now empty, so it is likely to simplify away.
2772 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
2775 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2777 BasicBlock
*SuccBlock
= BB
->getTerminator()->getSuccessor(0);
2778 PHINode
*PHIUse
= dyn_cast
<PHINode
>(ICI
->user_back());
2779 if (PHIUse
== nullptr || PHIUse
!= &SuccBlock
->front() ||
2780 isa
<PHINode
>(++BasicBlock::iterator(PHIUse
)))
2783 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2785 Constant
*DefaultCst
= ConstantInt::getTrue(BB
->getContext());
2786 Constant
*NewCst
= ConstantInt::getFalse(BB
->getContext());
2788 if (ICI
->getPredicate() == ICmpInst::ICMP_EQ
)
2789 std::swap(DefaultCst
, NewCst
);
2791 // Replace ICI (which is used by the PHI for the default value) with true or
2792 // false depending on if it is EQ or NE.
2793 ICI
->replaceAllUsesWith(DefaultCst
);
2794 ICI
->eraseFromParent();
2796 // Okay, the switch goes to this block on a default value. Add an edge from
2797 // the switch to the merge point on the compared value.
2798 BasicBlock
*NewBB
= BasicBlock::Create(BB
->getContext(), "switch.edge",
2799 BB
->getParent(), BB
);
2800 SmallVector
<uint64_t, 8> Weights
;
2801 bool HasWeights
= HasBranchWeights(SI
);
2803 GetBranchWeights(SI
, Weights
);
2804 if (Weights
.size() == 1 + SI
->getNumCases()) {
2805 // Split weight for default case to case for "Cst".
2806 Weights
[0] = (Weights
[0]+1) >> 1;
2807 Weights
.push_back(Weights
[0]);
2809 SmallVector
<uint32_t, 8> MDWeights(Weights
.begin(), Weights
.end());
2810 SI
->setMetadata(LLVMContext::MD_prof
,
2811 MDBuilder(SI
->getContext()).
2812 createBranchWeights(MDWeights
));
2815 SI
->addCase(Cst
, NewBB
);
2817 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2818 Builder
.SetInsertPoint(NewBB
);
2819 Builder
.SetCurrentDebugLocation(SI
->getDebugLoc());
2820 Builder
.CreateBr(SuccBlock
);
2821 PHIUse
->addIncoming(NewCst
, NewBB
);
2825 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2826 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2827 /// fold it into a switch instruction if so.
2828 static bool SimplifyBranchOnICmpChain(BranchInst
*BI
, const DataLayout
*DL
,
2829 IRBuilder
<> &Builder
) {
2830 Instruction
*Cond
= dyn_cast
<Instruction
>(BI
->getCondition());
2831 if (!Cond
) return false;
2833 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2834 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2835 // 'setne's and'ed together, collect them.
2837 // Try to gather values from a chain of and/or to be turned into a switch
2838 ConstantComparesGatherer
ConstantCompare(Cond
, DL
);
2839 // Unpack the result
2840 SmallVectorImpl
<ConstantInt
*> &Values
= ConstantCompare
.Vals
;
2841 Value
*CompVal
= ConstantCompare
.CompValue
;
2842 unsigned UsedICmps
= ConstantCompare
.UsedICmps
;
2843 Value
*ExtraCase
= ConstantCompare
.Extra
;
2845 // If we didn't have a multiply compared value, fail.
2846 if (!CompVal
) return false;
2848 // Avoid turning single icmps into a switch.
2852 bool TrueWhenEqual
= (Cond
->getOpcode() == Instruction::Or
);
2854 // There might be duplicate constants in the list, which the switch
2855 // instruction can't handle, remove them now.
2856 array_pod_sort(Values
.begin(), Values
.end(), ConstantIntSortPredicate
);
2857 Values
.erase(std::unique(Values
.begin(), Values
.end()), Values
.end());
2859 // If Extra was used, we require at least two switch values to do the
2860 // transformation. A switch with one value is just an cond branch.
2861 if (ExtraCase
&& Values
.size() < 2) return false;
2863 // TODO: Preserve branch weight metadata, similarly to how
2864 // FoldValueComparisonIntoPredecessors preserves it.
2866 // Figure out which block is which destination.
2867 BasicBlock
*DefaultBB
= BI
->getSuccessor(1);
2868 BasicBlock
*EdgeBB
= BI
->getSuccessor(0);
2869 if (!TrueWhenEqual
) std::swap(DefaultBB
, EdgeBB
);
2871 BasicBlock
*BB
= BI
->getParent();
2873 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values
.size()
2874 << " cases into SWITCH. BB is:\n" << *BB
);
2876 // If there are any extra values that couldn't be folded into the switch
2877 // then we evaluate them with an explicit branch first. Split the block
2878 // right before the condbr to handle it.
2880 BasicBlock
*NewBB
= BB
->splitBasicBlock(BI
, "switch.early.test");
2881 // Remove the uncond branch added to the old block.
2882 TerminatorInst
*OldTI
= BB
->getTerminator();
2883 Builder
.SetInsertPoint(OldTI
);
2886 Builder
.CreateCondBr(ExtraCase
, EdgeBB
, NewBB
);
2888 Builder
.CreateCondBr(ExtraCase
, NewBB
, EdgeBB
);
2890 OldTI
->eraseFromParent();
2892 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2893 // for the edge we just added.
2894 AddPredecessorToBlock(EdgeBB
, BB
, NewBB
);
2896 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2897 << "\nEXTRABB = " << *BB
);
2901 Builder
.SetInsertPoint(BI
);
2902 // Convert pointer to int before we switch.
2903 if (CompVal
->getType()->isPointerTy()) {
2904 assert(DL
&& "Cannot switch on pointer without DataLayout");
2905 CompVal
= Builder
.CreatePtrToInt(CompVal
,
2906 DL
->getIntPtrType(CompVal
->getType()),
2910 // Create the new switch instruction now.
2911 SwitchInst
*New
= Builder
.CreateSwitch(CompVal
, DefaultBB
, Values
.size());
2913 // Add all of the 'cases' to the switch instruction.
2914 for (unsigned i
= 0, e
= Values
.size(); i
!= e
; ++i
)
2915 New
->addCase(Values
[i
], EdgeBB
);
2917 // We added edges from PI to the EdgeBB. As such, if there were any
2918 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2919 // the number of edges added.
2920 for (BasicBlock::iterator BBI
= EdgeBB
->begin();
2921 isa
<PHINode
>(BBI
); ++BBI
) {
2922 PHINode
*PN
= cast
<PHINode
>(BBI
);
2923 Value
*InVal
= PN
->getIncomingValueForBlock(BB
);
2924 for (unsigned i
= 0, e
= Values
.size()-1; i
!= e
; ++i
)
2925 PN
->addIncoming(InVal
, BB
);
2928 // Erase the old branch instruction.
2929 EraseTerminatorInstAndDCECond(BI
);
2931 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB
<< '\n');
2935 bool SimplifyCFGOpt::SimplifyResume(ResumeInst
*RI
, IRBuilder
<> &Builder
) {
2936 // If this is a trivial landing pad that just continues unwinding the caught
2937 // exception then zap the landing pad, turning its invokes into calls.
2938 BasicBlock
*BB
= RI
->getParent();
2939 LandingPadInst
*LPInst
= dyn_cast
<LandingPadInst
>(BB
->getFirstNonPHI());
2940 if (RI
->getValue() != LPInst
)
2941 // Not a landing pad, or the resume is not unwinding the exception that
2942 // caused control to branch here.
2945 // Check that there are no other instructions except for debug intrinsics.
2946 BasicBlock::iterator I
= LPInst
, E
= RI
;
2948 if (!isa
<DbgInfoIntrinsic
>(I
))
2951 // Turn all invokes that unwind here into calls and delete the basic block.
2952 bool InvokeRequiresTableEntry
= false;
2953 bool Changed
= false;
2954 for (pred_iterator PI
= pred_begin(BB
), PE
= pred_end(BB
); PI
!= PE
;) {
2955 InvokeInst
*II
= cast
<InvokeInst
>((*PI
++)->getTerminator());
2957 if (II
->hasFnAttr(Attribute::UWTable
)) {
2958 // Don't remove an `invoke' instruction if the ABI requires an entry into
2960 InvokeRequiresTableEntry
= true;
2964 SmallVector
<Value
*, 8> Args(II
->op_begin(), II
->op_end() - 3);
2966 // Insert a call instruction before the invoke.
2967 CallInst
*Call
= CallInst::Create(II
->getCalledValue(), Args
, "", II
);
2969 Call
->setCallingConv(II
->getCallingConv());
2970 Call
->setAttributes(II
->getAttributes());
2971 Call
->setDebugLoc(II
->getDebugLoc());
2973 // Anything that used the value produced by the invoke instruction now uses
2974 // the value produced by the call instruction. Note that we do this even
2975 // for void functions and calls with no uses so that the callgraph edge is
2977 II
->replaceAllUsesWith(Call
);
2978 BB
->removePredecessor(II
->getParent());
2980 // Insert a branch to the normal destination right before the invoke.
2981 BranchInst::Create(II
->getNormalDest(), II
);
2983 // Finally, delete the invoke instruction!
2984 II
->eraseFromParent();
2988 if (!InvokeRequiresTableEntry
)
2989 // The landingpad is now unreachable. Zap it.
2990 BB
->eraseFromParent();
2995 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst
*RI
, IRBuilder
<> &Builder
) {
2996 BasicBlock
*BB
= RI
->getParent();
2997 if (!BB
->getFirstNonPHIOrDbg()->isTerminator()) return false;
2999 // Find predecessors that end with branches.
3000 SmallVector
<BasicBlock
*, 8> UncondBranchPreds
;
3001 SmallVector
<BranchInst
*, 8> CondBranchPreds
;
3002 for (pred_iterator PI
= pred_begin(BB
), E
= pred_end(BB
); PI
!= E
; ++PI
) {
3003 BasicBlock
*P
= *PI
;
3004 TerminatorInst
*PTI
= P
->getTerminator();
3005 if (BranchInst
*BI
= dyn_cast
<BranchInst
>(PTI
)) {
3006 if (BI
->isUnconditional())
3007 UncondBranchPreds
.push_back(P
);
3009 CondBranchPreds
.push_back(BI
);
3013 // If we found some, do the transformation!
3014 if (!UncondBranchPreds
.empty() && DupRet
) {
3015 while (!UncondBranchPreds
.empty()) {
3016 BasicBlock
*Pred
= UncondBranchPreds
.pop_back_val();
3017 DEBUG(dbgs() << "FOLDING: " << *BB
3018 << "INTO UNCOND BRANCH PRED: " << *Pred
);
3019 (void)FoldReturnIntoUncondBranch(RI
, BB
, Pred
);
3022 // If we eliminated all predecessors of the block, delete the block now.
3024 // We know there are no successors, so just nuke the block.
3025 BB
->eraseFromParent();
3030 // Check out all of the conditional branches going to this return
3031 // instruction. If any of them just select between returns, change the
3032 // branch itself into a select/return pair.
3033 while (!CondBranchPreds
.empty()) {
3034 BranchInst
*BI
= CondBranchPreds
.pop_back_val();
3036 // Check to see if the non-BB successor is also a return block.
3037 if (isa
<ReturnInst
>(BI
->getSuccessor(0)->getTerminator()) &&
3038 isa
<ReturnInst
>(BI
->getSuccessor(1)->getTerminator()) &&
3039 SimplifyCondBranchToTwoReturns(BI
, Builder
))
3045 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst
*UI
) {
3046 BasicBlock
*BB
= UI
->getParent();
3048 bool Changed
= false;
3050 // If there are any instructions immediately before the unreachable that can
3051 // be removed, do so.
3052 while (UI
!= BB
->begin()) {
3053 BasicBlock::iterator BBI
= UI
;
3055 // Do not delete instructions that can have side effects which might cause
3056 // the unreachable to not be reachable; specifically, calls and volatile
3057 // operations may have this effect.
3058 if (isa
<CallInst
>(BBI
) && !isa
<DbgInfoIntrinsic
>(BBI
)) break;
3060 if (BBI
->mayHaveSideEffects()) {
3061 if (StoreInst
*SI
= dyn_cast
<StoreInst
>(BBI
)) {
3062 if (SI
->isVolatile())
3064 } else if (LoadInst
*LI
= dyn_cast
<LoadInst
>(BBI
)) {
3065 if (LI
->isVolatile())
3067 } else if (AtomicRMWInst
*RMWI
= dyn_cast
<AtomicRMWInst
>(BBI
)) {
3068 if (RMWI
->isVolatile())
3070 } else if (AtomicCmpXchgInst
*CXI
= dyn_cast
<AtomicCmpXchgInst
>(BBI
)) {
3071 if (CXI
->isVolatile())
3073 } else if (!isa
<FenceInst
>(BBI
) && !isa
<VAArgInst
>(BBI
) &&
3074 !isa
<LandingPadInst
>(BBI
)) {
3077 // Note that deleting LandingPad's here is in fact okay, although it
3078 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3079 // all the predecessors of this block will be the unwind edges of Invokes,
3080 // and we can therefore guarantee this block will be erased.
3083 // Delete this instruction (any uses are guaranteed to be dead)
3084 if (!BBI
->use_empty())
3085 BBI
->replaceAllUsesWith(UndefValue::get(BBI
->getType()));
3086 BBI
->eraseFromParent();
3090 // If the unreachable instruction is the first in the block, take a gander
3091 // at all of the predecessors of this instruction, and simplify them.
3092 if (&BB
->front() != UI
) return Changed
;
3094 SmallVector
<BasicBlock
*, 8> Preds(pred_begin(BB
), pred_end(BB
));
3095 for (unsigned i
= 0, e
= Preds
.size(); i
!= e
; ++i
) {
3096 TerminatorInst
*TI
= Preds
[i
]->getTerminator();
3097 IRBuilder
<> Builder(TI
);
3098 if (BranchInst
*BI
= dyn_cast
<BranchInst
>(TI
)) {
3099 if (BI
->isUnconditional()) {
3100 if (BI
->getSuccessor(0) == BB
) {
3101 new UnreachableInst(TI
->getContext(), TI
);
3102 TI
->eraseFromParent();
3106 if (BI
->getSuccessor(0) == BB
) {
3107 Builder
.CreateBr(BI
->getSuccessor(1));
3108 EraseTerminatorInstAndDCECond(BI
);
3109 } else if (BI
->getSuccessor(1) == BB
) {
3110 Builder
.CreateBr(BI
->getSuccessor(0));
3111 EraseTerminatorInstAndDCECond(BI
);
3115 } else if (SwitchInst
*SI
= dyn_cast
<SwitchInst
>(TI
)) {
3116 for (SwitchInst::CaseIt i
= SI
->case_begin(), e
= SI
->case_end();
3118 if (i
.getCaseSuccessor() == BB
) {
3119 BB
->removePredecessor(SI
->getParent());
3124 // If the default value is unreachable, figure out the most popular
3125 // destination and make it the default.
3126 if (SI
->getDefaultDest() == BB
) {
3127 std::map
<BasicBlock
*, std::pair
<unsigned, unsigned> > Popularity
;
3128 for (SwitchInst::CaseIt i
= SI
->case_begin(), e
= SI
->case_end();
3130 std::pair
<unsigned, unsigned> &entry
=
3131 Popularity
[i
.getCaseSuccessor()];
3132 if (entry
.first
== 0) {
3134 entry
.second
= i
.getCaseIndex();
3140 // Find the most popular block.
3141 unsigned MaxPop
= 0;
3142 unsigned MaxIndex
= 0;
3143 BasicBlock
*MaxBlock
= nullptr;
3144 for (std::map
<BasicBlock
*, std::pair
<unsigned, unsigned> >::iterator
3145 I
= Popularity
.begin(), E
= Popularity
.end(); I
!= E
; ++I
) {
3146 if (I
->second
.first
> MaxPop
||
3147 (I
->second
.first
== MaxPop
&& MaxIndex
> I
->second
.second
)) {
3148 MaxPop
= I
->second
.first
;
3149 MaxIndex
= I
->second
.second
;
3150 MaxBlock
= I
->first
;
3154 // Make this the new default, allowing us to delete any explicit
3156 SI
->setDefaultDest(MaxBlock
);
3159 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3161 if (isa
<PHINode
>(MaxBlock
->begin()))
3162 for (unsigned i
= 0; i
!= MaxPop
-1; ++i
)
3163 MaxBlock
->removePredecessor(SI
->getParent());
3165 for (SwitchInst::CaseIt i
= SI
->case_begin(), e
= SI
->case_end();
3167 if (i
.getCaseSuccessor() == MaxBlock
) {
3173 } else if (InvokeInst
*II
= dyn_cast
<InvokeInst
>(TI
)) {
3174 if (II
->getUnwindDest() == BB
) {
3175 // Convert the invoke to a call instruction. This would be a good
3176 // place to note that the call does not throw though.
3177 BranchInst
*BI
= Builder
.CreateBr(II
->getNormalDest());
3178 II
->removeFromParent(); // Take out of symbol table
3180 // Insert the call now...
3181 SmallVector
<Value
*, 8> Args(II
->op_begin(), II
->op_end()-3);
3182 Builder
.SetInsertPoint(BI
);
3183 CallInst
*CI
= Builder
.CreateCall(II
->getCalledValue(),
3184 Args
, II
->getName());
3185 CI
->setCallingConv(II
->getCallingConv());
3186 CI
->setAttributes(II
->getAttributes());
3187 // If the invoke produced a value, the call does now instead.
3188 II
->replaceAllUsesWith(CI
);
3195 // If this block is now dead, remove it.
3196 if (pred_empty(BB
) &&
3197 BB
!= &BB
->getParent()->getEntryBlock()) {
3198 // We know there are no successors, so just nuke the block.
3199 BB
->eraseFromParent();
3206 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3207 /// integer range comparison into a sub, an icmp and a branch.
3208 static bool TurnSwitchRangeIntoICmp(SwitchInst
*SI
, IRBuilder
<> &Builder
) {
3209 assert(SI
->getNumCases() > 1 && "Degenerate switch?");
3211 // Make sure all cases point to the same destination and gather the values.
3212 SmallVector
<ConstantInt
*, 16> Cases
;
3213 SwitchInst::CaseIt I
= SI
->case_begin();
3214 Cases
.push_back(I
.getCaseValue());
3215 SwitchInst::CaseIt PrevI
= I
++;
3216 for (SwitchInst::CaseIt E
= SI
->case_end(); I
!= E
; PrevI
= I
++) {
3217 if (PrevI
.getCaseSuccessor() != I
.getCaseSuccessor())
3219 Cases
.push_back(I
.getCaseValue());
3221 assert(Cases
.size() == SI
->getNumCases() && "Not all cases gathered");
3223 // Sort the case values, then check if they form a range we can transform.
3224 array_pod_sort(Cases
.begin(), Cases
.end(), ConstantIntSortPredicate
);
3225 for (unsigned I
= 1, E
= Cases
.size(); I
!= E
; ++I
) {
3226 if (Cases
[I
-1]->getValue() != Cases
[I
]->getValue()+1)
3230 Constant
*Offset
= ConstantExpr::getNeg(Cases
.back());
3231 Constant
*NumCases
= ConstantInt::get(Offset
->getType(), SI
->getNumCases());
3233 Value
*Sub
= SI
->getCondition();
3234 if (!Offset
->isNullValue())
3235 Sub
= Builder
.CreateAdd(Sub
, Offset
, Sub
->getName()+".off");
3237 // If NumCases overflowed, then all possible values jump to the successor.
3238 if (NumCases
->isNullValue() && SI
->getNumCases() != 0)
3239 Cmp
= ConstantInt::getTrue(SI
->getContext());
3241 Cmp
= Builder
.CreateICmpULT(Sub
, NumCases
, "switch");
3242 BranchInst
*NewBI
= Builder
.CreateCondBr(
3243 Cmp
, SI
->case_begin().getCaseSuccessor(), SI
->getDefaultDest());
3245 // Update weight for the newly-created conditional branch.
3246 SmallVector
<uint64_t, 8> Weights
;
3247 bool HasWeights
= HasBranchWeights(SI
);
3249 GetBranchWeights(SI
, Weights
);
3250 if (Weights
.size() == 1 + SI
->getNumCases()) {
3251 // Combine all weights for the cases to be the true weight of NewBI.
3252 // We assume that the sum of all weights for a Terminator can fit into 32
3254 uint32_t NewTrueWeight
= 0;
3255 for (unsigned I
= 1, E
= Weights
.size(); I
!= E
; ++I
)
3256 NewTrueWeight
+= (uint32_t)Weights
[I
];
3257 NewBI
->setMetadata(LLVMContext::MD_prof
,
3258 MDBuilder(SI
->getContext()).
3259 createBranchWeights(NewTrueWeight
,
3260 (uint32_t)Weights
[0]));
3264 // Prune obsolete incoming values off the successor's PHI nodes.
3265 for (BasicBlock::iterator BBI
= SI
->case_begin().getCaseSuccessor()->begin();
3266 isa
<PHINode
>(BBI
); ++BBI
) {
3267 for (unsigned I
= 0, E
= SI
->getNumCases()-1; I
!= E
; ++I
)
3268 cast
<PHINode
>(BBI
)->removeIncomingValue(SI
->getParent());
3270 SI
->eraseFromParent();
3275 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3276 /// and use it to remove dead cases.
3277 static bool EliminateDeadSwitchCases(SwitchInst
*SI
, const DataLayout
*DL
,
3278 AssumptionCache
*AC
) {
3279 Value
*Cond
= SI
->getCondition();
3280 unsigned Bits
= Cond
->getType()->getIntegerBitWidth();
3281 APInt
KnownZero(Bits
, 0), KnownOne(Bits
, 0);
3282 computeKnownBits(Cond
, KnownZero
, KnownOne
, DL
, 0, AC
, SI
);
3284 // Gather dead cases.
3285 SmallVector
<ConstantInt
*, 8> DeadCases
;
3286 for (SwitchInst::CaseIt I
= SI
->case_begin(), E
= SI
->case_end(); I
!= E
; ++I
) {
3287 if ((I
.getCaseValue()->getValue() & KnownZero
) != 0 ||
3288 (I
.getCaseValue()->getValue() & KnownOne
) != KnownOne
) {
3289 DeadCases
.push_back(I
.getCaseValue());
3290 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3291 << I
.getCaseValue() << "' is dead.\n");
3295 SmallVector
<uint64_t, 8> Weights
;
3296 bool HasWeight
= HasBranchWeights(SI
);
3298 GetBranchWeights(SI
, Weights
);
3299 HasWeight
= (Weights
.size() == 1 + SI
->getNumCases());
3302 // Remove dead cases from the switch.
3303 for (unsigned I
= 0, E
= DeadCases
.size(); I
!= E
; ++I
) {
3304 SwitchInst::CaseIt Case
= SI
->findCaseValue(DeadCases
[I
]);
3305 assert(Case
!= SI
->case_default() &&
3306 "Case was not found. Probably mistake in DeadCases forming.");
3308 std::swap(Weights
[Case
.getCaseIndex()+1], Weights
.back());
3312 // Prune unused values from PHI nodes.
3313 Case
.getCaseSuccessor()->removePredecessor(SI
->getParent());
3314 SI
->removeCase(Case
);
3316 if (HasWeight
&& Weights
.size() >= 2) {
3317 SmallVector
<uint32_t, 8> MDWeights(Weights
.begin(), Weights
.end());
3318 SI
->setMetadata(LLVMContext::MD_prof
,
3319 MDBuilder(SI
->getParent()->getContext()).
3320 createBranchWeights(MDWeights
));
3323 return !DeadCases
.empty();
3326 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3327 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3328 /// by an unconditional branch), look at the phi node for BB in the successor
3329 /// block and see if the incoming value is equal to CaseValue. If so, return
3330 /// the phi node, and set PhiIndex to BB's index in the phi node.
3331 static PHINode
*FindPHIForConditionForwarding(ConstantInt
*CaseValue
,
3334 if (BB
->getFirstNonPHIOrDbg() != BB
->getTerminator())
3335 return nullptr; // BB must be empty to be a candidate for simplification.
3336 if (!BB
->getSinglePredecessor())
3337 return nullptr; // BB must be dominated by the switch.
3339 BranchInst
*Branch
= dyn_cast
<BranchInst
>(BB
->getTerminator());
3340 if (!Branch
|| !Branch
->isUnconditional())
3341 return nullptr; // Terminator must be unconditional branch.
3343 BasicBlock
*Succ
= Branch
->getSuccessor(0);
3345 BasicBlock::iterator I
= Succ
->begin();
3346 while (PHINode
*PHI
= dyn_cast
<PHINode
>(I
++)) {
3347 int Idx
= PHI
->getBasicBlockIndex(BB
);
3348 assert(Idx
>= 0 && "PHI has no entry for predecessor?");
3350 Value
*InValue
= PHI
->getIncomingValue(Idx
);
3351 if (InValue
!= CaseValue
) continue;
3360 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3361 /// instruction to a phi node dominated by the switch, if that would mean that
3362 /// some of the destination blocks of the switch can be folded away.
3363 /// Returns true if a change is made.
3364 static bool ForwardSwitchConditionToPHI(SwitchInst
*SI
) {
3365 typedef DenseMap
<PHINode
*, SmallVector
<int,4> > ForwardingNodesMap
;
3366 ForwardingNodesMap ForwardingNodes
;
3368 for (SwitchInst::CaseIt I
= SI
->case_begin(), E
= SI
->case_end(); I
!= E
; ++I
) {
3369 ConstantInt
*CaseValue
= I
.getCaseValue();
3370 BasicBlock
*CaseDest
= I
.getCaseSuccessor();
3373 PHINode
*PHI
= FindPHIForConditionForwarding(CaseValue
, CaseDest
,
3377 ForwardingNodes
[PHI
].push_back(PhiIndex
);
3380 bool Changed
= false;
3382 for (ForwardingNodesMap::iterator I
= ForwardingNodes
.begin(),
3383 E
= ForwardingNodes
.end(); I
!= E
; ++I
) {
3384 PHINode
*Phi
= I
->first
;
3385 SmallVectorImpl
<int> &Indexes
= I
->second
;
3387 if (Indexes
.size() < 2) continue;
3389 for (size_t I
= 0, E
= Indexes
.size(); I
!= E
; ++I
)
3390 Phi
->setIncomingValue(Indexes
[I
], SI
->getCondition());
3397 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3398 /// initializing an array of constants like C.
3399 static bool ValidLookupTableConstant(Constant
*C
) {
3400 if (C
->isThreadDependent())
3402 if (C
->isDLLImportDependent())
3405 if (ConstantExpr
*CE
= dyn_cast
<ConstantExpr
>(C
))
3406 return CE
->isGEPWithNoNotionalOverIndexing();
3408 return isa
<ConstantFP
>(C
) ||
3409 isa
<ConstantInt
>(C
) ||
3410 isa
<ConstantPointerNull
>(C
) ||
3411 isa
<GlobalValue
>(C
) ||
3415 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3416 /// its constant value in ConstantPool, returning 0 if it's not there.
3417 static Constant
*LookupConstant(Value
*V
,
3418 const SmallDenseMap
<Value
*, Constant
*>& ConstantPool
) {
3419 if (Constant
*C
= dyn_cast
<Constant
>(V
))
3421 return ConstantPool
.lookup(V
);
3424 /// ConstantFold - Try to fold instruction I into a constant. This works for
3425 /// simple instructions such as binary operations where both operands are
3426 /// constant or can be replaced by constants from the ConstantPool. Returns the
3427 /// resulting constant on success, 0 otherwise.
3429 ConstantFold(Instruction
*I
,
3430 const SmallDenseMap
<Value
*, Constant
*> &ConstantPool
,
3431 const DataLayout
*DL
) {
3432 if (SelectInst
*Select
= dyn_cast
<SelectInst
>(I
)) {
3433 Constant
*A
= LookupConstant(Select
->getCondition(), ConstantPool
);
3436 if (A
->isAllOnesValue())
3437 return LookupConstant(Select
->getTrueValue(), ConstantPool
);
3438 if (A
->isNullValue())
3439 return LookupConstant(Select
->getFalseValue(), ConstantPool
);
3443 SmallVector
<Constant
*, 4> COps
;
3444 for (unsigned N
= 0, E
= I
->getNumOperands(); N
!= E
; ++N
) {
3445 if (Constant
*A
= LookupConstant(I
->getOperand(N
), ConstantPool
))
3451 if (CmpInst
*Cmp
= dyn_cast
<CmpInst
>(I
))
3452 return ConstantFoldCompareInstOperands(Cmp
->getPredicate(), COps
[0],
3455 return ConstantFoldInstOperands(I
->getOpcode(), I
->getType(), COps
, DL
);
3458 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3459 /// at the common destination basic block, *CommonDest, for one of the case
3460 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3461 /// case), of a switch instruction SI.
3463 GetCaseResults(SwitchInst
*SI
,
3464 ConstantInt
*CaseVal
,
3465 BasicBlock
*CaseDest
,
3466 BasicBlock
**CommonDest
,
3467 SmallVectorImpl
<std::pair
<PHINode
*, Constant
*> > &Res
,
3468 const DataLayout
*DL
) {
3469 // The block from which we enter the common destination.
3470 BasicBlock
*Pred
= SI
->getParent();
3472 // If CaseDest is empty except for some side-effect free instructions through
3473 // which we can constant-propagate the CaseVal, continue to its successor.
3474 SmallDenseMap
<Value
*, Constant
*> ConstantPool
;
3475 ConstantPool
.insert(std::make_pair(SI
->getCondition(), CaseVal
));
3476 for (BasicBlock::iterator I
= CaseDest
->begin(), E
= CaseDest
->end(); I
!= E
;
3478 if (TerminatorInst
*T
= dyn_cast
<TerminatorInst
>(I
)) {
3479 // If the terminator is a simple branch, continue to the next block.
3480 if (T
->getNumSuccessors() != 1)
3483 CaseDest
= T
->getSuccessor(0);
3484 } else if (isa
<DbgInfoIntrinsic
>(I
)) {
3485 // Skip debug intrinsic.
3487 } else if (Constant
*C
= ConstantFold(I
, ConstantPool
, DL
)) {
3488 // Instruction is side-effect free and constant.
3490 // If the instruction has uses outside this block or a phi node slot for
3491 // the block, it is not safe to bypass the instruction since it would then
3492 // no longer dominate all its uses.
3493 for (auto &Use
: I
->uses()) {
3494 User
*User
= Use
.getUser();
3495 if (Instruction
*I
= dyn_cast
<Instruction
>(User
))
3496 if (I
->getParent() == CaseDest
)
3498 if (PHINode
*Phi
= dyn_cast
<PHINode
>(User
))
3499 if (Phi
->getIncomingBlock(Use
) == CaseDest
)
3504 ConstantPool
.insert(std::make_pair(I
, C
));
3510 // If we did not have a CommonDest before, use the current one.
3512 *CommonDest
= CaseDest
;
3513 // If the destination isn't the common one, abort.
3514 if (CaseDest
!= *CommonDest
)
3517 // Get the values for this case from phi nodes in the destination block.
3518 BasicBlock::iterator I
= (*CommonDest
)->begin();
3519 while (PHINode
*PHI
= dyn_cast
<PHINode
>(I
++)) {
3520 int Idx
= PHI
->getBasicBlockIndex(Pred
);
3524 Constant
*ConstVal
= LookupConstant(PHI
->getIncomingValue(Idx
),
3529 // Be conservative about which kinds of constants we support.
3530 if (!ValidLookupTableConstant(ConstVal
))
3533 Res
.push_back(std::make_pair(PHI
, ConstVal
));
3536 return Res
.size() > 0;
3539 // MapCaseToResult - Helper function used to
3540 // add CaseVal to the list of cases that generate Result.
3541 static void MapCaseToResult(ConstantInt
*CaseVal
,
3542 SwitchCaseResultVectorTy
&UniqueResults
,
3544 for (auto &I
: UniqueResults
) {
3545 if (I
.first
== Result
) {
3546 I
.second
.push_back(CaseVal
);
3550 UniqueResults
.push_back(std::make_pair(Result
,
3551 SmallVector
<ConstantInt
*, 4>(1, CaseVal
)));
3554 // InitializeUniqueCases - Helper function that initializes a map containing
3555 // results for the PHI node of the common destination block for a switch
3556 // instruction. Returns false if multiple PHI nodes have been found or if
3557 // there is not a common destination block for the switch.
3558 static bool InitializeUniqueCases(
3559 SwitchInst
*SI
, const DataLayout
*DL
, PHINode
*&PHI
,
3560 BasicBlock
*&CommonDest
,
3561 SwitchCaseResultVectorTy
&UniqueResults
,
3562 Constant
*&DefaultResult
) {
3563 for (auto &I
: SI
->cases()) {
3564 ConstantInt
*CaseVal
= I
.getCaseValue();
3566 // Resulting value at phi nodes for this case value.
3567 SwitchCaseResultsTy Results
;
3568 if (!GetCaseResults(SI
, CaseVal
, I
.getCaseSuccessor(), &CommonDest
, Results
,
3572 // Only one value per case is permitted
3573 if (Results
.size() > 1)
3575 MapCaseToResult(CaseVal
, UniqueResults
, Results
.begin()->second
);
3577 // Check the PHI consistency.
3579 PHI
= Results
[0].first
;
3580 else if (PHI
!= Results
[0].first
)
3583 // Find the default result value.
3584 SmallVector
<std::pair
<PHINode
*, Constant
*>, 1> DefaultResults
;
3585 BasicBlock
*DefaultDest
= SI
->getDefaultDest();
3586 GetCaseResults(SI
, nullptr, SI
->getDefaultDest(), &CommonDest
, DefaultResults
,
3588 // If the default value is not found abort unless the default destination
3591 DefaultResults
.size() == 1 ? DefaultResults
.begin()->second
: nullptr;
3592 if ((!DefaultResult
&&
3593 !isa
<UnreachableInst
>(DefaultDest
->getFirstNonPHIOrDbg())))
3599 // ConvertTwoCaseSwitch - Helper function that checks if it is possible to
3600 // transform a switch with only two cases (or two cases + default)
3601 // that produces a result into a value select.
3604 // case 10: %0 = icmp eq i32 %a, 10
3605 // return 10; %1 = select i1 %0, i32 10, i32 4
3606 // case 20: ----> %2 = icmp eq i32 %a, 20
3607 // return 2; %3 = select i1 %2, i32 2, i32 %1
3612 ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy
&ResultVector
,
3613 Constant
*DefaultResult
, Value
*Condition
,
3614 IRBuilder
<> &Builder
) {
3615 assert(ResultVector
.size() == 2 &&
3616 "We should have exactly two unique results at this point");
3617 // If we are selecting between only two cases transform into a simple
3618 // select or a two-way select if default is possible.
3619 if (ResultVector
[0].second
.size() == 1 &&
3620 ResultVector
[1].second
.size() == 1) {
3621 ConstantInt
*const FirstCase
= ResultVector
[0].second
[0];
3622 ConstantInt
*const SecondCase
= ResultVector
[1].second
[0];
3624 bool DefaultCanTrigger
= DefaultResult
;
3625 Value
*SelectValue
= ResultVector
[1].first
;
3626 if (DefaultCanTrigger
) {
3627 Value
*const ValueCompare
=
3628 Builder
.CreateICmpEQ(Condition
, SecondCase
, "switch.selectcmp");
3629 SelectValue
= Builder
.CreateSelect(ValueCompare
, ResultVector
[1].first
,
3630 DefaultResult
, "switch.select");
3632 Value
*const ValueCompare
=
3633 Builder
.CreateICmpEQ(Condition
, FirstCase
, "switch.selectcmp");
3634 return Builder
.CreateSelect(ValueCompare
, ResultVector
[0].first
, SelectValue
,
3641 // RemoveSwitchAfterSelectConversion - Helper function to cleanup a switch
3642 // instruction that has been converted into a select, fixing up PHI nodes and
3644 static void RemoveSwitchAfterSelectConversion(SwitchInst
*SI
, PHINode
*PHI
,
3646 IRBuilder
<> &Builder
) {
3647 BasicBlock
*SelectBB
= SI
->getParent();
3648 while (PHI
->getBasicBlockIndex(SelectBB
) >= 0)
3649 PHI
->removeIncomingValue(SelectBB
);
3650 PHI
->addIncoming(SelectValue
, SelectBB
);
3652 Builder
.CreateBr(PHI
->getParent());
3654 // Remove the switch.
3655 for (unsigned i
= 0, e
= SI
->getNumSuccessors(); i
< e
; ++i
) {
3656 BasicBlock
*Succ
= SI
->getSuccessor(i
);
3658 if (Succ
== PHI
->getParent())
3660 Succ
->removePredecessor(SelectBB
);
3662 SI
->eraseFromParent();
3665 /// SwitchToSelect - If the switch is only used to initialize one or more
3666 /// phi nodes in a common successor block with only two different
3667 /// constant values, replace the switch with select.
3668 static bool SwitchToSelect(SwitchInst
*SI
, IRBuilder
<> &Builder
,
3669 const DataLayout
*DL
, AssumptionCache
*AC
) {
3670 Value
*const Cond
= SI
->getCondition();
3671 PHINode
*PHI
= nullptr;
3672 BasicBlock
*CommonDest
= nullptr;
3673 Constant
*DefaultResult
;
3674 SwitchCaseResultVectorTy UniqueResults
;
3675 // Collect all the cases that will deliver the same value from the switch.
3676 if (!InitializeUniqueCases(SI
, DL
, PHI
, CommonDest
, UniqueResults
,
3679 // Selects choose between maximum two values.
3680 if (UniqueResults
.size() != 2)
3682 assert(PHI
!= nullptr && "PHI for value select not found");
3684 Builder
.SetInsertPoint(SI
);
3685 Value
*SelectValue
= ConvertTwoCaseSwitch(
3687 DefaultResult
, Cond
, Builder
);
3689 RemoveSwitchAfterSelectConversion(SI
, PHI
, SelectValue
, Builder
);
3692 // The switch couldn't be converted into a select.
3697 /// SwitchLookupTable - This class represents a lookup table that can be used
3698 /// to replace a switch.
3699 class SwitchLookupTable
{
3701 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3702 /// with the contents of Values, using DefaultValue to fill any holes in the
3704 SwitchLookupTable(Module
&M
,
3706 ConstantInt
*Offset
,
3707 const SmallVectorImpl
<std::pair
<ConstantInt
*, Constant
*> >& Values
,
3708 Constant
*DefaultValue
,
3709 const DataLayout
*DL
);
3711 /// BuildLookup - Build instructions with Builder to retrieve the value at
3712 /// the position given by Index in the lookup table.
3713 Value
*BuildLookup(Value
*Index
, IRBuilder
<> &Builder
);
3715 /// WouldFitInRegister - Return true if a table with TableSize elements of
3716 /// type ElementType would fit in a target-legal register.
3717 static bool WouldFitInRegister(const DataLayout
*DL
,
3719 const Type
*ElementType
);
3722 // Depending on the contents of the table, it can be represented in
3725 // For tables where each element contains the same value, we just have to
3726 // store that single value and return it for each lookup.
3729 // For tables where there is a linear relationship between table index
3730 // and values. We calculate the result with a simple multiplication
3731 // and addition instead of a table lookup.
3734 // For small tables with integer elements, we can pack them into a bitmap
3735 // that fits into a target-legal register. Values are retrieved by
3736 // shift and mask operations.
3739 // The table is stored as an array of values. Values are retrieved by load
3740 // instructions from the table.
3744 // For SingleValueKind, this is the single value.
3745 Constant
*SingleValue
;
3747 // For BitMapKind, this is the bitmap.
3748 ConstantInt
*BitMap
;
3749 IntegerType
*BitMapElementTy
;
3751 // For LinearMapKind, these are the constants used to derive the value.
3752 ConstantInt
*LinearOffset
;
3753 ConstantInt
*LinearMultiplier
;
3755 // For ArrayKind, this is the array.
3756 GlobalVariable
*Array
;
3760 SwitchLookupTable::SwitchLookupTable(Module
&M
,
3762 ConstantInt
*Offset
,
3763 const SmallVectorImpl
<std::pair
<ConstantInt
*, Constant
*> >& Values
,
3764 Constant
*DefaultValue
,
3765 const DataLayout
*DL
)
3766 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
3767 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
3768 assert(Values
.size() && "Can't build lookup table without values!");
3769 assert(TableSize
>= Values
.size() && "Can't fit values in table!");
3771 // If all values in the table are equal, this is that value.
3772 SingleValue
= Values
.begin()->second
;
3774 Type
*ValueType
= Values
.begin()->second
->getType();
3776 // Build up the table contents.
3777 SmallVector
<Constant
*, 64> TableContents(TableSize
);
3778 for (size_t I
= 0, E
= Values
.size(); I
!= E
; ++I
) {
3779 ConstantInt
*CaseVal
= Values
[I
].first
;
3780 Constant
*CaseRes
= Values
[I
].second
;
3781 assert(CaseRes
->getType() == ValueType
);
3783 uint64_t Idx
= (CaseVal
->getValue() - Offset
->getValue())
3785 TableContents
[Idx
] = CaseRes
;
3787 if (CaseRes
!= SingleValue
)
3788 SingleValue
= nullptr;
3791 // Fill in any holes in the table with the default result.
3792 if (Values
.size() < TableSize
) {
3793 assert(DefaultValue
&&
3794 "Need a default value to fill the lookup table holes.");
3795 assert(DefaultValue
->getType() == ValueType
);
3796 for (uint64_t I
= 0; I
< TableSize
; ++I
) {
3797 if (!TableContents
[I
])
3798 TableContents
[I
] = DefaultValue
;
3801 if (DefaultValue
!= SingleValue
)
3802 SingleValue
= nullptr;
3805 // If each element in the table contains the same value, we only need to store
3806 // that single value.
3808 Kind
= SingleValueKind
;
3812 // Check if we can derive the value with a linear transformation from the
3814 if (isa
<IntegerType
>(ValueType
)) {
3815 bool LinearMappingPossible
= true;
3818 assert(TableSize
>= 2 && "Should be a SingleValue table.");
3819 // Check if there is the same distance between two consecutive values.
3820 for (uint64_t I
= 0; I
< TableSize
; ++I
) {
3821 ConstantInt
*ConstVal
= dyn_cast
<ConstantInt
>(TableContents
[I
]);
3823 // This is an undef. We could deal with it, but undefs in lookup tables
3824 // are very seldom. It's probably not worth the additional complexity.
3825 LinearMappingPossible
= false;
3828 APInt Val
= ConstVal
->getValue();
3830 APInt Dist
= Val
- PrevVal
;
3833 } else if (Dist
!= DistToPrev
) {
3834 LinearMappingPossible
= false;
3840 if (LinearMappingPossible
) {
3841 LinearOffset
= cast
<ConstantInt
>(TableContents
[0]);
3842 LinearMultiplier
= ConstantInt::get(M
.getContext(), DistToPrev
);
3843 Kind
= LinearMapKind
;
3849 // If the type is integer and the table fits in a register, build a bitmap.
3850 if (WouldFitInRegister(DL
, TableSize
, ValueType
)) {
3851 IntegerType
*IT
= cast
<IntegerType
>(ValueType
);
3852 APInt
TableInt(TableSize
* IT
->getBitWidth(), 0);
3853 for (uint64_t I
= TableSize
; I
> 0; --I
) {
3854 TableInt
<<= IT
->getBitWidth();
3855 // Insert values into the bitmap. Undef values are set to zero.
3856 if (!isa
<UndefValue
>(TableContents
[I
- 1])) {
3857 ConstantInt
*Val
= cast
<ConstantInt
>(TableContents
[I
- 1]);
3858 TableInt
|= Val
->getValue().zext(TableInt
.getBitWidth());
3861 BitMap
= ConstantInt::get(M
.getContext(), TableInt
);
3862 BitMapElementTy
= IT
;
3868 // Store the table in an array.
3869 ArrayType
*ArrayTy
= ArrayType::get(ValueType
, TableSize
);
3870 Constant
*Initializer
= ConstantArray::get(ArrayTy
, TableContents
);
3872 Array
= new GlobalVariable(M
, ArrayTy
, /*constant=*/ true,
3873 GlobalVariable::PrivateLinkage
,
3876 Array
->setUnnamedAddr(true);
3880 Value
*SwitchLookupTable::BuildLookup(Value
*Index
, IRBuilder
<> &Builder
) {
3882 case SingleValueKind
:
3884 case LinearMapKind
: {
3885 // Derive the result value from the input value.
3886 Value
*Result
= Builder
.CreateIntCast(Index
, LinearMultiplier
->getType(),
3887 false, "switch.idx.cast");
3888 if (!LinearMultiplier
->isOne())
3889 Result
= Builder
.CreateMul(Result
, LinearMultiplier
, "switch.idx.mult");
3890 if (!LinearOffset
->isZero())
3891 Result
= Builder
.CreateAdd(Result
, LinearOffset
, "switch.offset");
3895 // Type of the bitmap (e.g. i59).
3896 IntegerType
*MapTy
= BitMap
->getType();
3898 // Cast Index to the same type as the bitmap.
3899 // Note: The Index is <= the number of elements in the table, so
3900 // truncating it to the width of the bitmask is safe.
3901 Value
*ShiftAmt
= Builder
.CreateZExtOrTrunc(Index
, MapTy
, "switch.cast");
3903 // Multiply the shift amount by the element width.
3904 ShiftAmt
= Builder
.CreateMul(ShiftAmt
,
3905 ConstantInt::get(MapTy
, BitMapElementTy
->getBitWidth()),
3909 Value
*DownShifted
= Builder
.CreateLShr(BitMap
, ShiftAmt
,
3910 "switch.downshift");
3912 return Builder
.CreateTrunc(DownShifted
, BitMapElementTy
,
3916 // Make sure the table index will not overflow when treated as signed.
3917 IntegerType
*IT
= cast
<IntegerType
>(Index
->getType());
3918 uint64_t TableSize
= Array
->getInitializer()->getType()
3919 ->getArrayNumElements();
3920 if (TableSize
> (1ULL << (IT
->getBitWidth() - 1)))
3921 Index
= Builder
.CreateZExt(Index
,
3922 IntegerType::get(IT
->getContext(),
3923 IT
->getBitWidth() + 1),
3924 "switch.tableidx.zext");
3926 Value
*GEPIndices
[] = { Builder
.getInt32(0), Index
};
3927 Value
*GEP
= Builder
.CreateInBoundsGEP(Array
, GEPIndices
,
3929 return Builder
.CreateLoad(GEP
, "switch.load");
3932 llvm_unreachable("Unknown lookup table kind!");
3935 bool SwitchLookupTable::WouldFitInRegister(const DataLayout
*DL
,
3937 const Type
*ElementType
) {
3940 const IntegerType
*IT
= dyn_cast
<IntegerType
>(ElementType
);
3943 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3944 // are <= 15, we could try to narrow the type.
3946 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3947 if (TableSize
>= UINT_MAX
/IT
->getBitWidth())
3949 return DL
->fitsInLegalInteger(TableSize
* IT
->getBitWidth());
3952 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3953 /// for this switch, based on the number of cases, size of the table and the
3954 /// types of the results.
3955 static bool ShouldBuildLookupTable(SwitchInst
*SI
,
3957 const TargetTransformInfo
&TTI
,
3958 const DataLayout
*DL
,
3959 const SmallDenseMap
<PHINode
*, Type
*>& ResultTypes
) {
3960 if (SI
->getNumCases() > TableSize
|| TableSize
>= UINT64_MAX
/ 10)
3961 return false; // TableSize overflowed, or mul below might overflow.
3963 bool AllTablesFitInRegister
= true;
3964 bool HasIllegalType
= false;
3965 for (const auto &I
: ResultTypes
) {
3966 Type
*Ty
= I
.second
;
3968 // Saturate this flag to true.
3969 HasIllegalType
= HasIllegalType
|| !TTI
.isTypeLegal(Ty
);
3971 // Saturate this flag to false.
3972 AllTablesFitInRegister
= AllTablesFitInRegister
&&
3973 SwitchLookupTable::WouldFitInRegister(DL
, TableSize
, Ty
);
3975 // If both flags saturate, we're done. NOTE: This *only* works with
3976 // saturating flags, and all flags have to saturate first due to the
3977 // non-deterministic behavior of iterating over a dense map.
3978 if (HasIllegalType
&& !AllTablesFitInRegister
)
3982 // If each table would fit in a register, we should build it anyway.
3983 if (AllTablesFitInRegister
)
3986 // Don't build a table that doesn't fit in-register if it has illegal types.
3990 // The table density should be at least 40%. This is the same criterion as for
3991 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3992 // FIXME: Find the best cut-off.
3993 return SI
->getNumCases() * 10 >= TableSize
* 4;
3996 /// Try to reuse the switch table index compare. Following pattern:
3998 /// if (idx < tablesize)
3999 /// r = table[idx]; // table does not contain default_value
4001 /// r = default_value;
4002 /// if (r != default_value)
4005 /// Is optimized to:
4007 /// cond = idx < tablesize;
4011 /// r = default_value;
4015 /// Jump threading will then eliminate the second if(cond).
4016 static void reuseTableCompare(User
*PhiUser
, BasicBlock
*PhiBlock
,
4017 BranchInst
*RangeCheckBranch
, Constant
*DefaultValue
,
4018 const SmallVectorImpl
<std::pair
<ConstantInt
*, Constant
*> >& Values
) {
4020 ICmpInst
*CmpInst
= dyn_cast
<ICmpInst
>(PhiUser
);
4024 // We require that the compare is in the same block as the phi so that jump
4025 // threading can do its work afterwards.
4026 if (CmpInst
->getParent() != PhiBlock
)
4029 Constant
*CmpOp1
= dyn_cast
<Constant
>(CmpInst
->getOperand(1));
4033 Value
*RangeCmp
= RangeCheckBranch
->getCondition();
4034 Constant
*TrueConst
= ConstantInt::getTrue(RangeCmp
->getType());
4035 Constant
*FalseConst
= ConstantInt::getFalse(RangeCmp
->getType());
4037 // Check if the compare with the default value is constant true or false.
4038 Constant
*DefaultConst
= ConstantExpr::getICmp(CmpInst
->getPredicate(),
4039 DefaultValue
, CmpOp1
, true);
4040 if (DefaultConst
!= TrueConst
&& DefaultConst
!= FalseConst
)
4043 // Check if the compare with the case values is distinct from the default
4045 for (auto ValuePair
: Values
) {
4046 Constant
*CaseConst
= ConstantExpr::getICmp(CmpInst
->getPredicate(),
4047 ValuePair
.second
, CmpOp1
, true);
4048 if (!CaseConst
|| CaseConst
== DefaultConst
)
4050 assert((CaseConst
== TrueConst
|| CaseConst
== FalseConst
) &&
4051 "Expect true or false as compare result.");
4054 // Check if the branch instruction dominates the phi node. It's a simple
4055 // dominance check, but sufficient for our needs.
4056 // Although this check is invariant in the calling loops, it's better to do it
4057 // at this late stage. Practically we do it at most once for a switch.
4058 BasicBlock
*BranchBlock
= RangeCheckBranch
->getParent();
4059 for (auto PI
= pred_begin(PhiBlock
), E
= pred_end(PhiBlock
); PI
!= E
; ++PI
) {
4060 BasicBlock
*Pred
= *PI
;
4061 if (Pred
!= BranchBlock
&& Pred
->getUniquePredecessor() != BranchBlock
)
4065 if (DefaultConst
== FalseConst
) {
4066 // The compare yields the same result. We can replace it.
4067 CmpInst
->replaceAllUsesWith(RangeCmp
);
4068 ++NumTableCmpReuses
;
4070 // The compare yields the same result, just inverted. We can replace it.
4071 Value
*InvertedTableCmp
= BinaryOperator::CreateXor(RangeCmp
,
4072 ConstantInt::get(RangeCmp
->getType(), 1), "inverted.cmp",
4074 CmpInst
->replaceAllUsesWith(InvertedTableCmp
);
4075 ++NumTableCmpReuses
;
4079 /// SwitchToLookupTable - If the switch is only used to initialize one or more
4080 /// phi nodes in a common successor block with different constant values,
4081 /// replace the switch with lookup tables.
4082 static bool SwitchToLookupTable(SwitchInst
*SI
,
4083 IRBuilder
<> &Builder
,
4084 const TargetTransformInfo
&TTI
,
4085 const DataLayout
* DL
) {
4086 assert(SI
->getNumCases() > 1 && "Degenerate switch?");
4088 // Only build lookup table when we have a target that supports it.
4089 if (!TTI
.shouldBuildLookupTables())
4092 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
4093 // split off a dense part and build a lookup table for that.
4095 // FIXME: This creates arrays of GEPs to constant strings, which means each
4096 // GEP needs a runtime relocation in PIC code. We should just build one big
4097 // string and lookup indices into that.
4099 // Ignore switches with less than three cases. Lookup tables will not make them
4100 // faster, so we don't analyze them.
4101 if (SI
->getNumCases() < 3)
4104 // Figure out the corresponding result for each case value and phi node in the
4105 // common destination, as well as the the min and max case values.
4106 assert(SI
->case_begin() != SI
->case_end());
4107 SwitchInst::CaseIt CI
= SI
->case_begin();
4108 ConstantInt
*MinCaseVal
= CI
.getCaseValue();
4109 ConstantInt
*MaxCaseVal
= CI
.getCaseValue();
4111 BasicBlock
*CommonDest
= nullptr;
4112 typedef SmallVector
<std::pair
<ConstantInt
*, Constant
*>, 4> ResultListTy
;
4113 SmallDenseMap
<PHINode
*, ResultListTy
> ResultLists
;
4114 SmallDenseMap
<PHINode
*, Constant
*> DefaultResults
;
4115 SmallDenseMap
<PHINode
*, Type
*> ResultTypes
;
4116 SmallVector
<PHINode
*, 4> PHIs
;
4118 for (SwitchInst::CaseIt E
= SI
->case_end(); CI
!= E
; ++CI
) {
4119 ConstantInt
*CaseVal
= CI
.getCaseValue();
4120 if (CaseVal
->getValue().slt(MinCaseVal
->getValue()))
4121 MinCaseVal
= CaseVal
;
4122 if (CaseVal
->getValue().sgt(MaxCaseVal
->getValue()))
4123 MaxCaseVal
= CaseVal
;
4125 // Resulting value at phi nodes for this case value.
4126 typedef SmallVector
<std::pair
<PHINode
*, Constant
*>, 4> ResultsTy
;
4128 if (!GetCaseResults(SI
, CaseVal
, CI
.getCaseSuccessor(), &CommonDest
,
4132 // Append the result from this case to the list for each phi.
4133 for (const auto &I
: Results
) {
4134 PHINode
*PHI
= I
.first
;
4135 Constant
*Value
= I
.second
;
4136 if (!ResultLists
.count(PHI
))
4137 PHIs
.push_back(PHI
);
4138 ResultLists
[PHI
].push_back(std::make_pair(CaseVal
, Value
));
4142 // Keep track of the result types.
4143 for (PHINode
*PHI
: PHIs
) {
4144 ResultTypes
[PHI
] = ResultLists
[PHI
][0].second
->getType();
4147 uint64_t NumResults
= ResultLists
[PHIs
[0]].size();
4148 APInt RangeSpread
= MaxCaseVal
->getValue() - MinCaseVal
->getValue();
4149 uint64_t TableSize
= RangeSpread
.getLimitedValue() + 1;
4150 bool TableHasHoles
= (NumResults
< TableSize
);
4152 // If the table has holes, we need a constant result for the default case
4153 // or a bitmask that fits in a register.
4154 SmallVector
<std::pair
<PHINode
*, Constant
*>, 4> DefaultResultsList
;
4155 bool HasDefaultResults
= GetCaseResults(SI
, nullptr, SI
->getDefaultDest(),
4156 &CommonDest
, DefaultResultsList
, DL
);
4158 bool NeedMask
= (TableHasHoles
&& !HasDefaultResults
);
4160 // As an extra penalty for the validity test we require more cases.
4161 if (SI
->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
4163 if (!(DL
&& DL
->fitsInLegalInteger(TableSize
)))
4167 for (const auto &I
: DefaultResultsList
) {
4168 PHINode
*PHI
= I
.first
;
4169 Constant
*Result
= I
.second
;
4170 DefaultResults
[PHI
] = Result
;
4173 if (!ShouldBuildLookupTable(SI
, TableSize
, TTI
, DL
, ResultTypes
))
4176 // Create the BB that does the lookups.
4177 Module
&Mod
= *CommonDest
->getParent()->getParent();
4178 BasicBlock
*LookupBB
= BasicBlock::Create(Mod
.getContext(),
4180 CommonDest
->getParent(),
4183 // Compute the table index value.
4184 Builder
.SetInsertPoint(SI
);
4185 Value
*TableIndex
= Builder
.CreateSub(SI
->getCondition(), MinCaseVal
,
4188 // Compute the maximum table size representable by the integer type we are
4190 unsigned CaseSize
= MinCaseVal
->getType()->getPrimitiveSizeInBits();
4191 uint64_t MaxTableSize
= CaseSize
> 63 ? UINT64_MAX
: 1ULL << CaseSize
;
4192 assert(MaxTableSize
>= TableSize
&&
4193 "It is impossible for a switch to have more entries than the max "
4194 "representable value of its input integer type's size.");
4196 // If we have a fully covered lookup table, unconditionally branch to the
4197 // lookup table BB. Otherwise, check if the condition value is within the case
4198 // range. If it is so, branch to the new BB. Otherwise branch to SI's default
4200 BranchInst
*RangeCheckBranch
= nullptr;
4202 const bool GeneratingCoveredLookupTable
= MaxTableSize
== TableSize
;
4203 if (GeneratingCoveredLookupTable
) {
4204 Builder
.CreateBr(LookupBB
);
4205 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
4206 // do not delete PHINodes here.
4207 SI
->getDefaultDest()->removePredecessor(SI
->getParent(),
4208 true/*DontDeleteUselessPHIs*/);
4210 Value
*Cmp
= Builder
.CreateICmpULT(TableIndex
, ConstantInt::get(
4211 MinCaseVal
->getType(), TableSize
));
4212 RangeCheckBranch
= Builder
.CreateCondBr(Cmp
, LookupBB
, SI
->getDefaultDest());
4215 // Populate the BB that does the lookups.
4216 Builder
.SetInsertPoint(LookupBB
);
4219 // Before doing the lookup we do the hole check.
4220 // The LookupBB is therefore re-purposed to do the hole check
4221 // and we create a new LookupBB.
4222 BasicBlock
*MaskBB
= LookupBB
;
4223 MaskBB
->setName("switch.hole_check");
4224 LookupBB
= BasicBlock::Create(Mod
.getContext(),
4226 CommonDest
->getParent(),
4229 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
4230 // unnecessary illegal types.
4231 uint64_t TableSizePowOf2
= NextPowerOf2(std::max(7ULL, TableSize
- 1ULL));
4232 APInt
MaskInt(TableSizePowOf2
, 0);
4233 APInt
One(TableSizePowOf2
, 1);
4234 // Build bitmask; fill in a 1 bit for every case.
4235 const ResultListTy
&ResultList
= ResultLists
[PHIs
[0]];
4236 for (size_t I
= 0, E
= ResultList
.size(); I
!= E
; ++I
) {
4237 uint64_t Idx
= (ResultList
[I
].first
->getValue() -
4238 MinCaseVal
->getValue()).getLimitedValue();
4239 MaskInt
|= One
<< Idx
;
4241 ConstantInt
*TableMask
= ConstantInt::get(Mod
.getContext(), MaskInt
);
4243 // Get the TableIndex'th bit of the bitmask.
4244 // If this bit is 0 (meaning hole) jump to the default destination,
4245 // else continue with table lookup.
4246 IntegerType
*MapTy
= TableMask
->getType();
4247 Value
*MaskIndex
= Builder
.CreateZExtOrTrunc(TableIndex
, MapTy
,
4248 "switch.maskindex");
4249 Value
*Shifted
= Builder
.CreateLShr(TableMask
, MaskIndex
,
4251 Value
*LoBit
= Builder
.CreateTrunc(Shifted
,
4252 Type::getInt1Ty(Mod
.getContext()),
4254 Builder
.CreateCondBr(LoBit
, LookupBB
, SI
->getDefaultDest());
4256 Builder
.SetInsertPoint(LookupBB
);
4257 AddPredecessorToBlock(SI
->getDefaultDest(), MaskBB
, SI
->getParent());
4260 bool ReturnedEarly
= false;
4261 for (size_t I
= 0, E
= PHIs
.size(); I
!= E
; ++I
) {
4262 PHINode
*PHI
= PHIs
[I
];
4263 const ResultListTy
&ResultList
= ResultLists
[PHI
];
4265 // If using a bitmask, use any value to fill the lookup table holes.
4266 Constant
*DV
= NeedMask
? ResultLists
[PHI
][0].second
: DefaultResults
[PHI
];
4267 SwitchLookupTable
Table(Mod
, TableSize
, MinCaseVal
, ResultList
, DV
, DL
);
4269 Value
*Result
= Table
.BuildLookup(TableIndex
, Builder
);
4271 // If the result is used to return immediately from the function, we want to
4272 // do that right here.
4273 if (PHI
->hasOneUse() && isa
<ReturnInst
>(*PHI
->user_begin()) &&
4274 PHI
->user_back() == CommonDest
->getFirstNonPHIOrDbg()) {
4275 Builder
.CreateRet(Result
);
4276 ReturnedEarly
= true;
4280 // Do a small peephole optimization: re-use the switch table compare if
4282 if (!TableHasHoles
&& HasDefaultResults
&& RangeCheckBranch
) {
4283 BasicBlock
*PhiBlock
= PHI
->getParent();
4284 // Search for compare instructions which use the phi.
4285 for (auto *User
: PHI
->users()) {
4286 reuseTableCompare(User
, PhiBlock
, RangeCheckBranch
, DV
, ResultList
);
4290 PHI
->addIncoming(Result
, LookupBB
);
4294 Builder
.CreateBr(CommonDest
);
4296 // Remove the switch.
4297 for (unsigned i
= 0, e
= SI
->getNumSuccessors(); i
< e
; ++i
) {
4298 BasicBlock
*Succ
= SI
->getSuccessor(i
);
4300 if (Succ
== SI
->getDefaultDest())
4302 Succ
->removePredecessor(SI
->getParent());
4304 SI
->eraseFromParent();
4308 ++NumLookupTablesHoles
;
4312 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst
*SI
, IRBuilder
<> &Builder
) {
4313 BasicBlock
*BB
= SI
->getParent();
4315 if (isValueEqualityComparison(SI
)) {
4316 // If we only have one predecessor, and if it is a branch on this value,
4317 // see if that predecessor totally determines the outcome of this switch.
4318 if (BasicBlock
*OnlyPred
= BB
->getSinglePredecessor())
4319 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI
, OnlyPred
, Builder
))
4320 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4322 Value
*Cond
= SI
->getCondition();
4323 if (SelectInst
*Select
= dyn_cast
<SelectInst
>(Cond
))
4324 if (SimplifySwitchOnSelect(SI
, Select
))
4325 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4327 // If the block only contains the switch, see if we can fold the block
4328 // away into any preds.
4329 BasicBlock::iterator BBI
= BB
->begin();
4330 // Ignore dbg intrinsics.
4331 while (isa
<DbgInfoIntrinsic
>(BBI
))
4334 if (FoldValueComparisonIntoPredecessors(SI
, Builder
))
4335 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4338 // Try to transform the switch into an icmp and a branch.
4339 if (TurnSwitchRangeIntoICmp(SI
, Builder
))
4340 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4342 // Remove unreachable cases.
4343 if (EliminateDeadSwitchCases(SI
, DL
, AC
))
4344 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4346 if (SwitchToSelect(SI
, Builder
, DL
, AC
))
4347 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4349 if (ForwardSwitchConditionToPHI(SI
))
4350 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4352 if (SwitchToLookupTable(SI
, Builder
, TTI
, DL
))
4353 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4358 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst
*IBI
) {
4359 BasicBlock
*BB
= IBI
->getParent();
4360 bool Changed
= false;
4362 // Eliminate redundant destinations.
4363 SmallPtrSet
<Value
*, 8> Succs
;
4364 for (unsigned i
= 0, e
= IBI
->getNumDestinations(); i
!= e
; ++i
) {
4365 BasicBlock
*Dest
= IBI
->getDestination(i
);
4366 if (!Dest
->hasAddressTaken() || !Succs
.insert(Dest
).second
) {
4367 Dest
->removePredecessor(BB
);
4368 IBI
->removeDestination(i
);
4374 if (IBI
->getNumDestinations() == 0) {
4375 // If the indirectbr has no successors, change it to unreachable.
4376 new UnreachableInst(IBI
->getContext(), IBI
);
4377 EraseTerminatorInstAndDCECond(IBI
);
4381 if (IBI
->getNumDestinations() == 1) {
4382 // If the indirectbr has one successor, change it to a direct branch.
4383 BranchInst::Create(IBI
->getDestination(0), IBI
);
4384 EraseTerminatorInstAndDCECond(IBI
);
4388 if (SelectInst
*SI
= dyn_cast
<SelectInst
>(IBI
->getAddress())) {
4389 if (SimplifyIndirectBrOnSelect(IBI
, SI
))
4390 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4395 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst
*BI
, IRBuilder
<> &Builder
){
4396 BasicBlock
*BB
= BI
->getParent();
4398 if (SinkCommon
&& SinkThenElseCodeToEnd(BI
))
4401 // If the Terminator is the only non-phi instruction, simplify the block.
4402 BasicBlock::iterator I
= BB
->getFirstNonPHIOrDbg();
4403 if (I
->isTerminator() && BB
!= &BB
->getParent()->getEntryBlock() &&
4404 TryToSimplifyUncondBranchFromEmptyBlock(BB
))
4407 // If the only instruction in the block is a seteq/setne comparison
4408 // against a constant, try to simplify the block.
4409 if (ICmpInst
*ICI
= dyn_cast
<ICmpInst
>(I
))
4410 if (ICI
->isEquality() && isa
<ConstantInt
>(ICI
->getOperand(1))) {
4411 for (++I
; isa
<DbgInfoIntrinsic
>(I
); ++I
)
4413 if (I
->isTerminator() &&
4414 TryToSimplifyUncondBranchWithICmpInIt(ICI
, Builder
, TTI
,
4415 BonusInstThreshold
, DL
, AC
))
4419 // If this basic block is ONLY a compare and a branch, and if a predecessor
4420 // branches to us and our successor, fold the comparison into the
4421 // predecessor and use logical operations to update the incoming value
4422 // for PHI nodes in common successor.
4423 if (FoldBranchToCommonDest(BI
, DL
, BonusInstThreshold
))
4424 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4429 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst
*BI
, IRBuilder
<> &Builder
) {
4430 BasicBlock
*BB
= BI
->getParent();
4432 // Conditional branch
4433 if (isValueEqualityComparison(BI
)) {
4434 // If we only have one predecessor, and if it is a branch on this value,
4435 // see if that predecessor totally determines the outcome of this
4437 if (BasicBlock
*OnlyPred
= BB
->getSinglePredecessor())
4438 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI
, OnlyPred
, Builder
))
4439 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4441 // This block must be empty, except for the setcond inst, if it exists.
4442 // Ignore dbg intrinsics.
4443 BasicBlock::iterator I
= BB
->begin();
4444 // Ignore dbg intrinsics.
4445 while (isa
<DbgInfoIntrinsic
>(I
))
4448 if (FoldValueComparisonIntoPredecessors(BI
, Builder
))
4449 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4450 } else if (&*I
== cast
<Instruction
>(BI
->getCondition())){
4452 // Ignore dbg intrinsics.
4453 while (isa
<DbgInfoIntrinsic
>(I
))
4455 if (&*I
== BI
&& FoldValueComparisonIntoPredecessors(BI
, Builder
))
4456 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4460 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4461 if (SimplifyBranchOnICmpChain(BI
, DL
, Builder
))
4464 // If this basic block is ONLY a compare and a branch, and if a predecessor
4465 // branches to us and one of our successors, fold the comparison into the
4466 // predecessor and use logical operations to pick the right destination.
4467 if (FoldBranchToCommonDest(BI
, DL
, BonusInstThreshold
))
4468 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4470 // We have a conditional branch to two blocks that are only reachable
4471 // from BI. We know that the condbr dominates the two blocks, so see if
4472 // there is any identical code in the "then" and "else" blocks. If so, we
4473 // can hoist it up to the branching block.
4474 if (BI
->getSuccessor(0)->getSinglePredecessor()) {
4475 if (BI
->getSuccessor(1)->getSinglePredecessor()) {
4476 if (HoistThenElseCodeToIf(BI
, DL
))
4477 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4479 // If Successor #1 has multiple preds, we may be able to conditionally
4480 // execute Successor #0 if it branches to Successor #1.
4481 TerminatorInst
*Succ0TI
= BI
->getSuccessor(0)->getTerminator();
4482 if (Succ0TI
->getNumSuccessors() == 1 &&
4483 Succ0TI
->getSuccessor(0) == BI
->getSuccessor(1))
4484 if (SpeculativelyExecuteBB(BI
, BI
->getSuccessor(0), DL
))
4485 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4487 } else if (BI
->getSuccessor(1)->getSinglePredecessor()) {
4488 // If Successor #0 has multiple preds, we may be able to conditionally
4489 // execute Successor #1 if it branches to Successor #0.
4490 TerminatorInst
*Succ1TI
= BI
->getSuccessor(1)->getTerminator();
4491 if (Succ1TI
->getNumSuccessors() == 1 &&
4492 Succ1TI
->getSuccessor(0) == BI
->getSuccessor(0))
4493 if (SpeculativelyExecuteBB(BI
, BI
->getSuccessor(1), DL
))
4494 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4497 // If this is a branch on a phi node in the current block, thread control
4498 // through this block if any PHI node entries are constants.
4499 if (PHINode
*PN
= dyn_cast
<PHINode
>(BI
->getCondition()))
4500 if (PN
->getParent() == BI
->getParent())
4501 if (FoldCondBranchOnPHI(BI
, DL
))
4502 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4504 // Scan predecessor blocks for conditional branches.
4505 for (pred_iterator PI
= pred_begin(BB
), E
= pred_end(BB
); PI
!= E
; ++PI
)
4506 if (BranchInst
*PBI
= dyn_cast
<BranchInst
>((*PI
)->getTerminator()))
4507 if (PBI
!= BI
&& PBI
->isConditional())
4508 if (SimplifyCondBranchToCondBranch(PBI
, BI
))
4509 return SimplifyCFG(BB
, TTI
, BonusInstThreshold
, DL
, AC
) | true;
4514 /// Check if passing a value to an instruction will cause undefined behavior.
4515 static bool passingValueIsAlwaysUndefined(Value
*V
, Instruction
*I
) {
4516 Constant
*C
= dyn_cast
<Constant
>(V
);
4523 if (C
->isNullValue()) {
4524 // Only look at the first use, avoid hurting compile time with long uselists
4525 User
*Use
= *I
->user_begin();
4527 // Now make sure that there are no instructions in between that can alter
4528 // control flow (eg. calls)
4529 for (BasicBlock::iterator i
= ++BasicBlock::iterator(I
); &*i
!= Use
; ++i
)
4530 if (i
== I
->getParent()->end() || i
->mayHaveSideEffects())
4533 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4534 if (GetElementPtrInst
*GEP
= dyn_cast
<GetElementPtrInst
>(Use
))
4535 if (GEP
->getPointerOperand() == I
)
4536 return passingValueIsAlwaysUndefined(V
, GEP
);
4538 // Look through bitcasts.
4539 if (BitCastInst
*BC
= dyn_cast
<BitCastInst
>(Use
))
4540 return passingValueIsAlwaysUndefined(V
, BC
);
4542 // Load from null is undefined.
4543 if (LoadInst
*LI
= dyn_cast
<LoadInst
>(Use
))
4544 if (!LI
->isVolatile())
4545 return LI
->getPointerAddressSpace() == 0;
4547 // Store to null is undefined.
4548 if (StoreInst
*SI
= dyn_cast
<StoreInst
>(Use
))
4549 if (!SI
->isVolatile())
4550 return SI
->getPointerAddressSpace() == 0 && SI
->getPointerOperand() == I
;
4555 /// If BB has an incoming value that will always trigger undefined behavior
4556 /// (eg. null pointer dereference), remove the branch leading here.
4557 static bool removeUndefIntroducingPredecessor(BasicBlock
*BB
) {
4558 for (BasicBlock::iterator i
= BB
->begin();
4559 PHINode
*PHI
= dyn_cast
<PHINode
>(i
); ++i
)
4560 for (unsigned i
= 0, e
= PHI
->getNumIncomingValues(); i
!= e
; ++i
)
4561 if (passingValueIsAlwaysUndefined(PHI
->getIncomingValue(i
), PHI
)) {
4562 TerminatorInst
*T
= PHI
->getIncomingBlock(i
)->getTerminator();
4563 IRBuilder
<> Builder(T
);
4564 if (BranchInst
*BI
= dyn_cast
<BranchInst
>(T
)) {
4565 BB
->removePredecessor(PHI
->getIncomingBlock(i
));
4566 // Turn uncoditional branches into unreachables and remove the dead
4567 // destination from conditional branches.
4568 if (BI
->isUnconditional())
4569 Builder
.CreateUnreachable();
4571 Builder
.CreateBr(BI
->getSuccessor(0) == BB
? BI
->getSuccessor(1) :
4572 BI
->getSuccessor(0));
4573 BI
->eraseFromParent();
4576 // TODO: SwitchInst.
4582 bool SimplifyCFGOpt::run(BasicBlock
*BB
) {
4583 bool Changed
= false;
4585 assert(BB
&& BB
->getParent() && "Block not embedded in function!");
4586 assert(BB
->getTerminator() && "Degenerate basic block encountered!");
4588 // Remove basic blocks that have no predecessors (except the entry block)...
4589 // or that just have themself as a predecessor. These are unreachable.
4590 if ((pred_empty(BB
) &&
4591 BB
!= &BB
->getParent()->getEntryBlock()) ||
4592 BB
->getSinglePredecessor() == BB
) {
4593 DEBUG(dbgs() << "Removing BB: \n" << *BB
);
4594 DeleteDeadBlock(BB
);
4598 // Check to see if we can constant propagate this terminator instruction
4600 Changed
|= ConstantFoldTerminator(BB
, true);
4602 // Check for and eliminate duplicate PHI nodes in this block.
4603 Changed
|= EliminateDuplicatePHINodes(BB
);
4605 // Check for and remove branches that will always cause undefined behavior.
4606 Changed
|= removeUndefIntroducingPredecessor(BB
);
4608 // Merge basic blocks into their predecessor if there is only one distinct
4609 // pred, and if there is only one distinct successor of the predecessor, and
4610 // if there are no PHI nodes.
4612 if (MergeBlockIntoPredecessor(BB
))
4615 IRBuilder
<> Builder(BB
);
4617 // If there is a trivial two-entry PHI node in this basic block, and we can
4618 // eliminate it, do so now.
4619 if (PHINode
*PN
= dyn_cast
<PHINode
>(BB
->begin()))
4620 if (PN
->getNumIncomingValues() == 2)
4621 Changed
|= FoldTwoEntryPHINode(PN
, DL
);
4623 Builder
.SetInsertPoint(BB
->getTerminator());
4624 if (BranchInst
*BI
= dyn_cast
<BranchInst
>(BB
->getTerminator())) {
4625 if (BI
->isUnconditional()) {
4626 if (SimplifyUncondBranch(BI
, Builder
)) return true;
4628 if (SimplifyCondBranch(BI
, Builder
)) return true;
4630 } else if (ReturnInst
*RI
= dyn_cast
<ReturnInst
>(BB
->getTerminator())) {
4631 if (SimplifyReturn(RI
, Builder
)) return true;
4632 } else if (ResumeInst
*RI
= dyn_cast
<ResumeInst
>(BB
->getTerminator())) {
4633 if (SimplifyResume(RI
, Builder
)) return true;
4634 } else if (SwitchInst
*SI
= dyn_cast
<SwitchInst
>(BB
->getTerminator())) {
4635 if (SimplifySwitch(SI
, Builder
)) return true;
4636 } else if (UnreachableInst
*UI
=
4637 dyn_cast
<UnreachableInst
>(BB
->getTerminator())) {
4638 if (SimplifyUnreachable(UI
)) return true;
4639 } else if (IndirectBrInst
*IBI
=
4640 dyn_cast
<IndirectBrInst
>(BB
->getTerminator())) {
4641 if (SimplifyIndirectBr(IBI
)) return true;
4647 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4648 /// example, it adjusts branches to branches to eliminate the extra hop, it
4649 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4650 /// of the CFG. It returns true if a modification was made.
4652 bool llvm::SimplifyCFG(BasicBlock
*BB
, const TargetTransformInfo
&TTI
,
4653 unsigned BonusInstThreshold
, const DataLayout
*DL
,
4654 AssumptionCache
*AC
) {
4655 return SimplifyCFGOpt(TTI
, BonusInstThreshold
, DL
, AC
).run(BB
);