1	//===- InstCombinePHI.cpp -------------------------------------------------===//
2	//
3	// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4	// See https://llvm.org/LICENSE.txt for license information.
5	// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6	//
7	//===----------------------------------------------------------------------===//
8	//
9	// This file implements the visitPHINode function.
10	//
11	//===----------------------------------------------------------------------===//
12
13	#include "InstCombineInternal.h"
14	#include "llvm/ADT/STLExtras.h"
15	#include "llvm/ADT/SmallPtrSet.h"
16	#include "llvm/ADT/Statistic.h"
17	#include "llvm/Analysis/InstructionSimplify.h"
18	#include "llvm/Analysis/ValueTracking.h"
19	#include "llvm/IR/PatternMatch.h"
20	#include "llvm/Support/CommandLine.h"
21	#include "llvm/Transforms/InstCombine/InstCombiner.h"
22	#include "llvm/Transforms/Utils/Local.h"
23	#include <optional>
24
25	using namespace llvm;
26	using namespace llvm::PatternMatch;
27
28	#define DEBUG_TYPE "instcombine"
29
30	static cl::opt<unsigned>
31	MaxNumPhis("instcombine-max-num-phis", cl::init(Val: 512),
32	cl::desc("Maximum number phis to handle in intptr/ptrint folding"));
33
34	STATISTIC(NumPHIsOfInsertValues,
35	"Number of phi-of-insertvalue turned into insertvalue-of-phis");
36	STATISTIC(NumPHIsOfExtractValues,
37	"Number of phi-of-extractvalue turned into extractvalue-of-phi");
38	STATISTIC(NumPHICSEs, "Number of PHI's that got CSE'd");
39
40	/// The PHI arguments will be folded into a single operation with a PHI node
41	/// as input. The debug location of the single operation will be the merged
42	/// locations of the original PHI node arguments.
43	void InstCombinerImpl::PHIArgMergedDebugLoc(Instruction *Inst, PHINode &PN) {
44	auto *FirstInst = cast<Instruction>(Val: PN.getIncomingValue(i: 0));
45	Inst->setDebugLoc(FirstInst->getDebugLoc());
46	// We do not expect a CallInst here, otherwise, N-way merging of DebugLoc
47	// will be inefficient.
48	assert(!isa<CallInst>(Inst));
49
50	for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) {
51	auto *I = cast<Instruction>(Val: V);
52	Inst->applyMergedLocation(LocA: Inst->getDebugLoc(), LocB: I->getDebugLoc());
53	}
54	}
55
56	// Replace Integer typed PHI PN if the PHI's value is used as a pointer value.
57	// If there is an existing pointer typed PHI that produces the same value as PN,
58	// replace PN and the IntToPtr operation with it. Otherwise, synthesize a new
59	// PHI node:
60	//
61	// Case-1:
62	// bb1:
63	// int_init = PtrToInt(ptr_init)
64	// br label %bb2
65	// bb2:
66	// int_val = PHI([int_init, %bb1], [int_val_inc, %bb2]
67	// ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2]
68	// ptr_val2 = IntToPtr(int_val)
69	// ...
70	// use(ptr_val2)
71	// ptr_val_inc = ...
72	// inc_val_inc = PtrToInt(ptr_val_inc)
73	//
74	// ==>
75	// bb1:
76	// br label %bb2
77	// bb2:
78	// ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2]
79	// ...
80	// use(ptr_val)
81	// ptr_val_inc = ...
82	//
83	// Case-2:
84	// bb1:
85	// int_ptr = BitCast(ptr_ptr)
86	// int_init = Load(int_ptr)
87	// br label %bb2
88	// bb2:
89	// int_val = PHI([int_init, %bb1], [int_val_inc, %bb2]
90	// ptr_val2 = IntToPtr(int_val)
91	// ...
92	// use(ptr_val2)
93	// ptr_val_inc = ...
94	// inc_val_inc = PtrToInt(ptr_val_inc)
95	// ==>
96	// bb1:
97	// ptr_init = Load(ptr_ptr)
98	// br label %bb2
99	// bb2:
100	// ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2]
101	// ...
102	// use(ptr_val)
103	// ptr_val_inc = ...
104	// ...
105	//
106	bool InstCombinerImpl::foldIntegerTypedPHI(PHINode &PN) {
107	if (!PN.getType()->isIntegerTy())
108	return false;
109	if (!PN.hasOneUse())
110	return false;
111
112	auto *IntToPtr = dyn_cast<IntToPtrInst>(Val: PN.user_back());
113	if (!IntToPtr)
114	return false;
115
116	// Check if the pointer is actually used as pointer:
117	auto HasPointerUse = [](Instruction *IIP) {
118	for (User *U : IIP->users()) {
119	Value *Ptr = nullptr;
120	if (LoadInst *LoadI = dyn_cast<LoadInst>(Val: U)) {
121	Ptr = LoadI->getPointerOperand();
122	} else if (StoreInst *SI = dyn_cast<StoreInst>(Val: U)) {
123	Ptr = SI->getPointerOperand();
124	} else if (GetElementPtrInst *GI = dyn_cast<GetElementPtrInst>(Val: U)) {
125	Ptr = GI->getPointerOperand();
126	}
127
128	if (Ptr && Ptr == IIP)
129	return true;
130	}
131	return false;
132	};
133
134	if (!HasPointerUse(IntToPtr))
135	return false;
136
137	if (DL.getPointerSizeInBits(AS: IntToPtr->getAddressSpace()) !=
138	DL.getTypeSizeInBits(Ty: IntToPtr->getOperand(i_nocapture: 0)->getType()))
139	return false;
140
141	SmallVector<Value *, 4> AvailablePtrVals;
142	for (auto Incoming : zip(t: PN.blocks(), u: PN.incoming_values())) {
143	BasicBlock *BB = std::get<0>(t&: Incoming);
144	Value *Arg = std::get<1>(t&: Incoming);
145
146	// First look backward:
147	if (auto *PI = dyn_cast<PtrToIntInst>(Val: Arg)) {
148	AvailablePtrVals.emplace_back(Args: PI->getOperand(i_nocapture: 0));
149	continue;
150	}
151
152	// Next look forward:
153	Value *ArgIntToPtr = nullptr;
154	for (User *U : Arg->users()) {
155	if (isa<IntToPtrInst>(Val: U) && U->getType() == IntToPtr->getType() &&
156	(DT.dominates(Def: cast<Instruction>(Val: U), BB) ||
157	cast<Instruction>(Val: U)->getParent() == BB)) {
158	ArgIntToPtr = U;
159	break;
160	}
161	}
162
163	if (ArgIntToPtr) {
164	AvailablePtrVals.emplace_back(Args&: ArgIntToPtr);
165	continue;
166	}
167
168	// If Arg is defined by a PHI, allow it. This will also create
169	// more opportunities iteratively.
170	if (isa<PHINode>(Val: Arg)) {
171	AvailablePtrVals.emplace_back(Args&: Arg);
172	continue;
173	}
174
175	// For a single use integer load:
176	auto *LoadI = dyn_cast<LoadInst>(Val: Arg);
177	if (!LoadI)
178	return false;
179
180	if (!LoadI->hasOneUse())
181	return false;
182
183	// Push the integer typed Load instruction into the available
184	// value set, and fix it up later when the pointer typed PHI
185	// is synthesized.
186	AvailablePtrVals.emplace_back(Args&: LoadI);
187	}
188
189	// Now search for a matching PHI
190	auto *BB = PN.getParent();
191	assert(AvailablePtrVals.size() == PN.getNumIncomingValues() &&
192	"Not enough available ptr typed incoming values");
193	PHINode *MatchingPtrPHI = nullptr;
194	unsigned NumPhis = 0;
195	for (PHINode &PtrPHI : BB->phis()) {
196	// FIXME: consider handling this in AggressiveInstCombine
197	if (NumPhis++ > MaxNumPhis)
198	return false;
199	if (&PtrPHI == &PN || PtrPHI.getType() != IntToPtr->getType())
200	continue;
201	if (any_of(Range: zip(t: PN.blocks(), u&: AvailablePtrVals),
202	P: [&](const auto &BlockAndValue) {
203	BasicBlock *BB = std::get<0>(BlockAndValue);
204	Value *V = std::get<1>(BlockAndValue);
205	return PtrPHI.getIncomingValueForBlock(BB) != V;
206	}))
207	continue;
208	MatchingPtrPHI = &PtrPHI;
209	break;
210	}
211
212	if (MatchingPtrPHI) {
213	assert(MatchingPtrPHI->getType() == IntToPtr->getType() &&
214	"Phi's Type does not match with IntToPtr");
215	// Explicitly replace the inttoptr (rather than inserting a ptrtoint) here,
216	// to make sure another transform can't undo it in the meantime.
217	replaceInstUsesWith(I&: *IntToPtr, V: MatchingPtrPHI);
218	eraseInstFromFunction(I&: *IntToPtr);
219	eraseInstFromFunction(I&: PN);
220	return true;
221	}
222
223	// If it requires a conversion for every PHI operand, do not do it.
224	if (all_of(Range&: AvailablePtrVals, P: [&](Value *V) {
225	return (V->getType() != IntToPtr->getType()) || isa<IntToPtrInst>(Val: V);
226	}))
227	return false;
228
229	// If any of the operand that requires casting is a terminator
230	// instruction, do not do it. Similarly, do not do the transform if the value
231	// is PHI in a block with no insertion point, for example, a catchswitch
232	// block, since we will not be able to insert a cast after the PHI.
233	if (any_of(Range&: AvailablePtrVals, P: [&](Value *V) {
234	if (V->getType() == IntToPtr->getType())
235	return false;
236	auto *Inst = dyn_cast<Instruction>(Val: V);
237	if (!Inst)
238	return false;
239	if (Inst->isTerminator())
240	return true;
241	auto *BB = Inst->getParent();
242	if (isa<PHINode>(Val: Inst) && BB->getFirstInsertionPt() == BB->end())
243	return true;
244	return false;
245	}))
246	return false;
247
248	PHINode *NewPtrPHI = PHINode::Create(
249	Ty: IntToPtr->getType(), NumReservedValues: PN.getNumIncomingValues(), NameStr: PN.getName() + ".ptr");
250
251	InsertNewInstBefore(New: NewPtrPHI, Old: PN.getIterator());
252	SmallDenseMap<Value *, Instruction *> Casts;
253	for (auto Incoming : zip(t: PN.blocks(), u&: AvailablePtrVals)) {
254	auto *IncomingBB = std::get<0>(t&: Incoming);
255	auto *IncomingVal = std::get<1>(t&: Incoming);
256
257	if (IncomingVal->getType() == IntToPtr->getType()) {
258	NewPtrPHI->addIncoming(V: IncomingVal, BB: IncomingBB);
259	continue;
260	}
261
262	#ifndef NDEBUG
263	LoadInst *LoadI = dyn_cast<LoadInst>(Val: IncomingVal);
264	assert((isa<PHINode>(IncomingVal) ||
265	IncomingVal->getType()->isPointerTy() ||
266	(LoadI && LoadI->hasOneUse())) &&
267	"Can not replace LoadInst with multiple uses");
268	#endif
269	// Need to insert a BitCast.
270	// For an integer Load instruction with a single use, the load + IntToPtr
271	// cast will be simplified into a pointer load:
272	// %v = load i64, i64* %a.ip, align 8
273	// %v.cast = inttoptr i64 %v to float **
274	// ==>
275	// %v.ptrp = bitcast i64 * %a.ip to float **
276	// %v.cast = load float *, float ** %v.ptrp, align 8
277	Instruction *&CI = Casts[IncomingVal];
278	if (!CI) {
279	CI = CastInst::CreateBitOrPointerCast(S: IncomingVal, Ty: IntToPtr->getType(),
280	Name: IncomingVal->getName() + ".ptr");
281	if (auto *IncomingI = dyn_cast<Instruction>(Val: IncomingVal)) {
282	BasicBlock::iterator InsertPos(IncomingI);
283	InsertPos++;
284	BasicBlock *BB = IncomingI->getParent();
285	if (isa<PHINode>(Val: IncomingI))
286	InsertPos = BB->getFirstInsertionPt();
287	assert(InsertPos != BB->end() && "should have checked above");
288	InsertNewInstBefore(New: CI, Old: InsertPos);
289	} else {
290	auto *InsertBB = &IncomingBB->getParent()->getEntryBlock();
291	InsertNewInstBefore(New: CI, Old: InsertBB->getFirstInsertionPt());
292	}
293	}
294	NewPtrPHI->addIncoming(V: CI, BB: IncomingBB);
295	}
296
297	// Explicitly replace the inttoptr (rather than inserting a ptrtoint) here,
298	// to make sure another transform can't undo it in the meantime.
299	replaceInstUsesWith(I&: *IntToPtr, V: NewPtrPHI);
300	eraseInstFromFunction(I&: *IntToPtr);
301	eraseInstFromFunction(I&: PN);
302	return true;
303	}
304
305	// Remove RoundTrip IntToPtr/PtrToInt Cast on PHI-Operand and
306	// fold Phi-operand to bitcast.
307	Instruction *InstCombinerImpl::foldPHIArgIntToPtrToPHI(PHINode &PN) {
308	// convert ptr2int ( phi[ int2ptr(ptr2int(x))] ) --> ptr2int ( phi [ x ] )
309	// Make sure all uses of phi are ptr2int.
310	if (!all_of(Range: PN.users(), P: [](User *U) { return isa<PtrToIntInst>(Val: U); }))
311	return nullptr;
312
313	// Iterating over all operands to check presence of target pointers for
314	// optimization.
315	bool OperandWithRoundTripCast = false;
316	for (unsigned OpNum = 0; OpNum != PN.getNumIncomingValues(); ++OpNum) {
317	if (auto *NewOp =
318	simplifyIntToPtrRoundTripCast(Val: PN.getIncomingValue(i: OpNum))) {
319	replaceOperand(I&: PN, OpNum, V: NewOp);
320	OperandWithRoundTripCast = true;
321	}
322	}
323	if (!OperandWithRoundTripCast)
324	return nullptr;
325	return &PN;
326	}
327
328	/// If we have something like phi [insertvalue(a,b,0), insertvalue(c,d,0)],
329	/// turn this into a phi[a,c] and phi[b,d] and a single insertvalue.
330	Instruction *
331	InstCombinerImpl::foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN) {
332	auto *FirstIVI = cast<InsertValueInst>(Val: PN.getIncomingValue(i: 0));
333
334	// Scan to see if all operands are `insertvalue`'s with the same indicies,
335	// and all have a single use.
336	for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) {
337	auto *I = dyn_cast<InsertValueInst>(Val: V);
338	if (!I || !I->hasOneUser() || I->getIndices() != FirstIVI->getIndices())
339	return nullptr;
340	}
341
342	// For each operand of an `insertvalue`
343	std::array<PHINode *, 2> NewOperands;
344	for (int OpIdx : {0, 1}) {
345	auto *&NewOperand = NewOperands[OpIdx];
346	// Create a new PHI node to receive the values the operand has in each
347	// incoming basic block.
348	NewOperand = PHINode::Create(
349	Ty: FirstIVI->getOperand(i_nocapture: OpIdx)->getType(), NumReservedValues: PN.getNumIncomingValues(),
350	NameStr: FirstIVI->getOperand(i_nocapture: OpIdx)->getName() + ".pn");
351	// And populate each operand's PHI with said values.
352	for (auto Incoming : zip(t: PN.blocks(), u: PN.incoming_values()))
353	NewOperand->addIncoming(
354	V: cast<InsertValueInst>(Val&: std::get<1>(t&: Incoming))->getOperand(i_nocapture: OpIdx),
355	BB: std::get<0>(t&: Incoming));
356	InsertNewInstBefore(New: NewOperand, Old: PN.getIterator());
357	}
358
359	// And finally, create `insertvalue` over the newly-formed PHI nodes.
360	auto *NewIVI = InsertValueInst::Create(Agg: NewOperands[0], Val: NewOperands[1],
361	Idxs: FirstIVI->getIndices(), NameStr: PN.getName());
362
363	PHIArgMergedDebugLoc(Inst: NewIVI, PN);
364	++NumPHIsOfInsertValues;
365	return NewIVI;
366	}
367
368	/// If we have something like phi [extractvalue(a,0), extractvalue(b,0)],
369	/// turn this into a phi[a,b] and a single extractvalue.
370	Instruction *
371	InstCombinerImpl::foldPHIArgExtractValueInstructionIntoPHI(PHINode &PN) {
372	auto *FirstEVI = cast<ExtractValueInst>(Val: PN.getIncomingValue(i: 0));
373
374	// Scan to see if all operands are `extractvalue`'s with the same indicies,
375	// and all have a single use.
376	for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) {
377	auto *I = dyn_cast<ExtractValueInst>(Val: V);
378	if (!I || !I->hasOneUser() || I->getIndices() != FirstEVI->getIndices() ||
379	I->getAggregateOperand()->getType() !=
380	FirstEVI->getAggregateOperand()->getType())
381	return nullptr;
382	}
383
384	// Create a new PHI node to receive the values the aggregate operand has
385	// in each incoming basic block.
386	auto *NewAggregateOperand = PHINode::Create(
387	Ty: FirstEVI->getAggregateOperand()->getType(), NumReservedValues: PN.getNumIncomingValues(),
388	NameStr: FirstEVI->getAggregateOperand()->getName() + ".pn");
389	// And populate the PHI with said values.
390	for (auto Incoming : zip(t: PN.blocks(), u: PN.incoming_values()))
391	NewAggregateOperand->addIncoming(
392	V: cast<ExtractValueInst>(Val&: std::get<1>(t&: Incoming))->getAggregateOperand(),
393	BB: std::get<0>(t&: Incoming));
394	InsertNewInstBefore(New: NewAggregateOperand, Old: PN.getIterator());
395
396	// And finally, create `extractvalue` over the newly-formed PHI nodes.
397	auto *NewEVI = ExtractValueInst::Create(Agg: NewAggregateOperand,
398	Idxs: FirstEVI->getIndices(), NameStr: PN.getName());
399
400	PHIArgMergedDebugLoc(Inst: NewEVI, PN);
401	++NumPHIsOfExtractValues;
402	return NewEVI;
403	}
404
405	/// If we have something like phi [add (a,b), add(a,c)] and if a/b/c and the
406	/// adds all have a single user, turn this into a phi and a single binop.
407	Instruction *InstCombinerImpl::foldPHIArgBinOpIntoPHI(PHINode &PN) {
408	Instruction *FirstInst = cast<Instruction>(Val: PN.getIncomingValue(i: 0));
409	assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
410	unsigned Opc = FirstInst->getOpcode();
411	Value *LHSVal = FirstInst->getOperand(i: 0);
412	Value *RHSVal = FirstInst->getOperand(i: 1);
413
414	Type *LHSType = LHSVal->getType();
415	Type *RHSType = RHSVal->getType();
416
417	// Scan to see if all operands are the same opcode, and all have one user.
418	for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) {
419	Instruction *I = dyn_cast<Instruction>(Val: V);
420	if (!I || I->getOpcode() != Opc || !I->hasOneUser() ||
421	// Verify type of the LHS matches so we don't fold cmp's of different
422	// types.
423	I->getOperand(i: 0)->getType() != LHSType ||
424	I->getOperand(i: 1)->getType() != RHSType)
425	return nullptr;
426
427	// If they are CmpInst instructions, check their predicates
428	if (CmpInst *CI = dyn_cast<CmpInst>(Val: I))
429	if (CI->getPredicate() != cast<CmpInst>(Val: FirstInst)->getPredicate())
430	return nullptr;
431
432	// Keep track of which operand needs a phi node.
433	if (I->getOperand(i: 0) != LHSVal) LHSVal = nullptr;
434	if (I->getOperand(i: 1) != RHSVal) RHSVal = nullptr;
435	}
436
437	// If both LHS and RHS would need a PHI, don't do this transformation,
438	// because it would increase the number of PHIs entering the block,
439	// which leads to higher register pressure. This is especially
440	// bad when the PHIs are in the header of a loop.
441	if (!LHSVal && !RHSVal)
442	return nullptr;
443
444	// Otherwise, this is safe to transform!
445
446	Value *InLHS = FirstInst->getOperand(i: 0);
447	Value *InRHS = FirstInst->getOperand(i: 1);
448	PHINode *NewLHS = nullptr, *NewRHS = nullptr;
449	if (!LHSVal) {
450	NewLHS = PHINode::Create(Ty: LHSType, NumReservedValues: PN.getNumIncomingValues(),
451	NameStr: FirstInst->getOperand(i: 0)->getName() + ".pn");
452	NewLHS->addIncoming(V: InLHS, BB: PN.getIncomingBlock(i: 0));
453	InsertNewInstBefore(New: NewLHS, Old: PN.getIterator());
454	LHSVal = NewLHS;
455	}
456
457	if (!RHSVal) {
458	NewRHS = PHINode::Create(Ty: RHSType, NumReservedValues: PN.getNumIncomingValues(),
459	NameStr: FirstInst->getOperand(i: 1)->getName() + ".pn");
460	NewRHS->addIncoming(V: InRHS, BB: PN.getIncomingBlock(i: 0));
461	InsertNewInstBefore(New: NewRHS, Old: PN.getIterator());
462	RHSVal = NewRHS;
463	}
464
465	// Add all operands to the new PHIs.
466	if (NewLHS || NewRHS) {
467	for (auto Incoming : drop_begin(RangeOrContainer: zip(t: PN.blocks(), u: PN.incoming_values()))) {
468	BasicBlock *InBB = std::get<0>(t&: Incoming);
469	Value *InVal = std::get<1>(t&: Incoming);
470	Instruction *InInst = cast<Instruction>(Val: InVal);
471	if (NewLHS) {
472	Value *NewInLHS = InInst->getOperand(i: 0);
473	NewLHS->addIncoming(V: NewInLHS, BB: InBB);
474	}
475	if (NewRHS) {
476	Value *NewInRHS = InInst->getOperand(i: 1);
477	NewRHS->addIncoming(V: NewInRHS, BB: InBB);
478	}
479	}
480	}
481
482	if (CmpInst *CIOp = dyn_cast<CmpInst>(Val: FirstInst)) {
483	CmpInst *NewCI = CmpInst::Create(Op: CIOp->getOpcode(), Pred: CIOp->getPredicate(),
484	S1: LHSVal, S2: RHSVal);
485	PHIArgMergedDebugLoc(Inst: NewCI, PN);
486	return NewCI;
487	}
488
489	BinaryOperator *BinOp = cast<BinaryOperator>(Val: FirstInst);
490	BinaryOperator *NewBinOp =
491	BinaryOperator::Create(Op: BinOp->getOpcode(), S1: LHSVal, S2: RHSVal);
492
493	NewBinOp->copyIRFlags(V: PN.getIncomingValue(i: 0));
494
495	for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values()))
496	NewBinOp->andIRFlags(V);
497
498	PHIArgMergedDebugLoc(Inst: NewBinOp, PN);
499	return NewBinOp;
500	}
501
502	Instruction *InstCombinerImpl::foldPHIArgGEPIntoPHI(PHINode &PN) {
503	GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(Val: PN.getIncomingValue(i: 0));
504
505	SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
506	FirstInst->op_end());
507	// This is true if all GEP bases are allocas and if all indices into them are
508	// constants.
509	bool AllBasePointersAreAllocas = true;
510
511	// We don't want to replace this phi if the replacement would require
512	// more than one phi, which leads to higher register pressure. This is
513	// especially bad when the PHIs are in the header of a loop.
514	bool NeededPhi = false;
515
516	bool AllInBounds = true;
517
518	// Scan to see if all operands are the same opcode, and all have one user.
519	for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) {
520	GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Val: V);
521	if (!GEP || !GEP->hasOneUser() ||
522	GEP->getSourceElementType() != FirstInst->getSourceElementType() ||
523	GEP->getNumOperands() != FirstInst->getNumOperands())
524	return nullptr;
525
526	AllInBounds &= GEP->isInBounds();
527
528	// Keep track of whether or not all GEPs are of alloca pointers.
529	if (AllBasePointersAreAllocas &&
530	(!isa<AllocaInst>(Val: GEP->getOperand(i_nocapture: 0)) ||
531	!GEP->hasAllConstantIndices()))
532	AllBasePointersAreAllocas = false;
533
534	// Compare the operand lists.
535	for (unsigned Op = 0, E = FirstInst->getNumOperands(); Op != E; ++Op) {
536	if (FirstInst->getOperand(i_nocapture: Op) == GEP->getOperand(i_nocapture: Op))
537	continue;
538
539	// Don't merge two GEPs when two operands differ (introducing phi nodes)
540	// if one of the PHIs has a constant for the index. The index may be
541	// substantially cheaper to compute for the constants, so making it a
542	// variable index could pessimize the path. This also handles the case
543	// for struct indices, which must always be constant.
544	if (isa<ConstantInt>(Val: FirstInst->getOperand(i_nocapture: Op)) ||
545	isa<ConstantInt>(Val: GEP->getOperand(i_nocapture: Op)))
546	return nullptr;
547
548	if (FirstInst->getOperand(i_nocapture: Op)->getType() !=
549	GEP->getOperand(i_nocapture: Op)->getType())
550	return nullptr;
551
552	// If we already needed a PHI for an earlier operand, and another operand
553	// also requires a PHI, we'd be introducing more PHIs than we're
554	// eliminating, which increases register pressure on entry to the PHI's
555	// block.
556	if (NeededPhi)
557	return nullptr;
558
559	FixedOperands[Op] = nullptr; // Needs a PHI.
560	NeededPhi = true;
561	}
562	}
563
564	// If all of the base pointers of the PHI'd GEPs are from allocas, don't
565	// bother doing this transformation. At best, this will just save a bit of
566	// offset calculation, but all the predecessors will have to materialize the
567	// stack address into a register anyway. We'd actually rather *clone* the
568	// load up into the predecessors so that we have a load of a gep of an alloca,
569	// which can usually all be folded into the load.
570	if (AllBasePointersAreAllocas)
571	return nullptr;
572
573	// Otherwise, this is safe to transform. Insert PHI nodes for each operand
574	// that is variable.
575	SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
576
577	bool HasAnyPHIs = false;
578	for (unsigned I = 0, E = FixedOperands.size(); I != E; ++I) {
579	if (FixedOperands[I])
580	continue; // operand doesn't need a phi.
581	Value *FirstOp = FirstInst->getOperand(i_nocapture: I);
582	PHINode *NewPN =
583	PHINode::Create(Ty: FirstOp->getType(), NumReservedValues: E, NameStr: FirstOp->getName() + ".pn");
584	InsertNewInstBefore(New: NewPN, Old: PN.getIterator());
585
586	NewPN->addIncoming(V: FirstOp, BB: PN.getIncomingBlock(i: 0));
587	OperandPhis[I] = NewPN;
588	FixedOperands[I] = NewPN;
589	HasAnyPHIs = true;
590	}
591
592	// Add all operands to the new PHIs.
593	if (HasAnyPHIs) {
594	for (auto Incoming : drop_begin(RangeOrContainer: zip(t: PN.blocks(), u: PN.incoming_values()))) {
595	BasicBlock *InBB = std::get<0>(t&: Incoming);
596	Value *InVal = std::get<1>(t&: Incoming);
597	GetElementPtrInst *InGEP = cast<GetElementPtrInst>(Val: InVal);
598
599	for (unsigned Op = 0, E = OperandPhis.size(); Op != E; ++Op)
600	if (PHINode *OpPhi = OperandPhis[Op])
601	OpPhi->addIncoming(V: InGEP->getOperand(i_nocapture: Op), BB: InBB);
602	}
603	}
604
605	Value *Base = FixedOperands[0];
606	GetElementPtrInst *NewGEP =
607	GetElementPtrInst::Create(PointeeType: FirstInst->getSourceElementType(), Ptr: Base,
608	IdxList: ArrayRef(FixedOperands).slice(N: 1));
609	if (AllInBounds) NewGEP->setIsInBounds();
610	PHIArgMergedDebugLoc(Inst: NewGEP, PN);
611	return NewGEP;
612	}
613
614	/// Return true if we know that it is safe to sink the load out of the block
615	/// that defines it. This means that it must be obvious the value of the load is
616	/// not changed from the point of the load to the end of the block it is in.
617	///
618	/// Finally, it is safe, but not profitable, to sink a load targeting a
619	/// non-address-taken alloca. Doing so will cause us to not promote the alloca
620	/// to a register.
621	static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
622	BasicBlock::iterator BBI = L->getIterator(), E = L->getParent()->end();
623
624	for (++BBI; BBI != E; ++BBI)
625	if (BBI->mayWriteToMemory()) {
626	// Calls that only access inaccessible memory do not block sinking the
627	// load.
628	if (auto *CB = dyn_cast<CallBase>(Val&: BBI))
629	if (CB->onlyAccessesInaccessibleMemory())
630	continue;
631	return false;
632	}
633
634	// Check for non-address taken alloca. If not address-taken already, it isn't
635	// profitable to do this xform.
636	if (AllocaInst *AI = dyn_cast<AllocaInst>(Val: L->getOperand(i_nocapture: 0))) {
637	bool IsAddressTaken = false;
638	for (User *U : AI->users()) {
639	if (isa<LoadInst>(Val: U)) continue;
640	if (StoreInst *SI = dyn_cast<StoreInst>(Val: U)) {
641	// If storing TO the alloca, then the address isn't taken.
642	if (SI->getOperand(i_nocapture: 1) == AI) continue;
643	}
644	IsAddressTaken = true;
645	break;
646	}
647
648	if (!IsAddressTaken && AI->isStaticAlloca())
649	return false;
650	}
651
652	// If this load is a load from a GEP with a constant offset from an alloca,
653	// then we don't want to sink it. In its present form, it will be
654	// load [constant stack offset]. Sinking it will cause us to have to
655	// materialize the stack addresses in each predecessor in a register only to
656	// do a shared load from register in the successor.
657	if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Val: L->getOperand(i_nocapture: 0)))
658	if (AllocaInst *AI = dyn_cast<AllocaInst>(Val: GEP->getOperand(i_nocapture: 0)))
659	if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
660	return false;
661
662	return true;
663	}
664
665	Instruction *InstCombinerImpl::foldPHIArgLoadIntoPHI(PHINode &PN) {
666	LoadInst *FirstLI = cast<LoadInst>(Val: PN.getIncomingValue(i: 0));
667
668	// Can't forward swifterror through a phi.
669	if (FirstLI->getOperand(i_nocapture: 0)->isSwiftError())
670	return nullptr;
671
672	// FIXME: This is overconservative; this transform is allowed in some cases
673	// for atomic operations.
674	if (FirstLI->isAtomic())
675	return nullptr;
676
677	// When processing loads, we need to propagate two bits of information to the
678	// sunk load: whether it is volatile, and what its alignment is.
679	bool IsVolatile = FirstLI->isVolatile();
680	Align LoadAlignment = FirstLI->getAlign();
681	const unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace();
682
683	// We can't sink the load if the loaded value could be modified between the
684	// load and the PHI.
685	if (FirstLI->getParent() != PN.getIncomingBlock(i: 0) ||
686	!isSafeAndProfitableToSinkLoad(L: FirstLI))
687	return nullptr;
688
689	// If the PHI is of volatile loads and the load block has multiple
690	// successors, sinking it would remove a load of the volatile value from
691	// the path through the other successor.
692	if (IsVolatile &&
693	FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1)
694	return nullptr;
695
696	for (auto Incoming : drop_begin(RangeOrContainer: zip(t: PN.blocks(), u: PN.incoming_values()))) {
697	BasicBlock *InBB = std::get<0>(t&: Incoming);
698	Value *InVal = std::get<1>(t&: Incoming);
699	LoadInst *LI = dyn_cast<LoadInst>(Val: InVal);
700	if (!LI || !LI->hasOneUser() || LI->isAtomic())
701	return nullptr;
702
703	// Make sure all arguments are the same type of operation.
704	if (LI->isVolatile() != IsVolatile ||
705	LI->getPointerAddressSpace() != LoadAddrSpace)
706	return nullptr;
707
708	// Can't forward swifterror through a phi.
709	if (LI->getOperand(i_nocapture: 0)->isSwiftError())
710	return nullptr;
711
712	// We can't sink the load if the loaded value could be modified between
713	// the load and the PHI.
714	if (LI->getParent() != InBB || !isSafeAndProfitableToSinkLoad(L: LI))
715	return nullptr;
716
717	LoadAlignment = std::min(a: LoadAlignment, b: LI->getAlign());
718
719	// If the PHI is of volatile loads and the load block has multiple
720	// successors, sinking it would remove a load of the volatile value from
721	// the path through the other successor.
722	if (IsVolatile && LI->getParent()->getTerminator()->getNumSuccessors() != 1)
723	return nullptr;
724	}
725
726	// Okay, they are all the same operation. Create a new PHI node of the
727	// correct type, and PHI together all of the LHS's of the instructions.
728	PHINode *NewPN = PHINode::Create(Ty: FirstLI->getOperand(i_nocapture: 0)->getType(),
729	NumReservedValues: PN.getNumIncomingValues(),
730	NameStr: PN.getName()+".in");
731
732	Value *InVal = FirstLI->getOperand(i_nocapture: 0);
733	NewPN->addIncoming(V: InVal, BB: PN.getIncomingBlock(i: 0));
734	LoadInst *NewLI =
735	new LoadInst(FirstLI->getType(), NewPN, "", IsVolatile, LoadAlignment);
736
737	unsigned KnownIDs[] = {
738	LLVMContext::MD_tbaa,
739	LLVMContext::MD_range,
740	LLVMContext::MD_invariant_load,
741	LLVMContext::MD_alias_scope,
742	LLVMContext::MD_noalias,
743	LLVMContext::MD_nonnull,
744	LLVMContext::MD_align,
745	LLVMContext::MD_dereferenceable,
746	LLVMContext::MD_dereferenceable_or_null,
747	LLVMContext::MD_access_group,
748	LLVMContext::MD_noundef,
749	};
750
751	for (unsigned ID : KnownIDs)
752	NewLI->setMetadata(KindID: ID, Node: FirstLI->getMetadata(KindID: ID));
753
754	// Add all operands to the new PHI and combine TBAA metadata.
755	for (auto Incoming : drop_begin(RangeOrContainer: zip(t: PN.blocks(), u: PN.incoming_values()))) {
756	BasicBlock *BB = std::get<0>(t&: Incoming);
757	Value *V = std::get<1>(t&: Incoming);
758	LoadInst *LI = cast<LoadInst>(Val: V);
759	combineMetadata(K: NewLI, J: LI, KnownIDs, DoesKMove: true);
760	Value *NewInVal = LI->getOperand(i_nocapture: 0);
761	if (NewInVal != InVal)
762	InVal = nullptr;
763	NewPN->addIncoming(V: NewInVal, BB);
764	}
765
766	if (InVal) {
767	// The new PHI unions all of the same values together. This is really
768	// common, so we handle it intelligently here for compile-time speed.
769	NewLI->setOperand(i_nocapture: 0, Val_nocapture: InVal);
770	delete NewPN;
771	} else {
772	InsertNewInstBefore(New: NewPN, Old: PN.getIterator());
773	}
774
775	// If this was a volatile load that we are merging, make sure to loop through
776	// and mark all the input loads as non-volatile. If we don't do this, we will
777	// insert a new volatile load and the old ones will not be deletable.
778	if (IsVolatile)
779	for (Value *IncValue : PN.incoming_values())
780	cast<LoadInst>(Val: IncValue)->setVolatile(false);
781
782	PHIArgMergedDebugLoc(Inst: NewLI, PN);
783	return NewLI;
784	}
785
786	/// TODO: This function could handle other cast types, but then it might
787	/// require special-casing a cast from the 'i1' type. See the comment in
788	/// FoldPHIArgOpIntoPHI() about pessimizing illegal integer types.
789	Instruction *InstCombinerImpl::foldPHIArgZextsIntoPHI(PHINode &Phi) {
790	// We cannot create a new instruction after the PHI if the terminator is an
791	// EHPad because there is no valid insertion point.
792	if (Instruction *TI = Phi.getParent()->getTerminator())
793	if (TI->isEHPad())
794	return nullptr;
795
796	// Early exit for the common case of a phi with two operands. These are
797	// handled elsewhere. See the comment below where we check the count of zexts
798	// and constants for more details.
799	unsigned NumIncomingValues = Phi.getNumIncomingValues();
800	if (NumIncomingValues < 3)
801	return nullptr;
802
803	// Find the narrower type specified by the first zext.
804	Type *NarrowType = nullptr;
805	for (Value *V : Phi.incoming_values()) {
806	if (auto *Zext = dyn_cast<ZExtInst>(Val: V)) {
807	NarrowType = Zext->getSrcTy();
808	break;
809	}
810	}
811	if (!NarrowType)
812	return nullptr;
813
814	// Walk the phi operands checking that we only have zexts or constants that
815	// we can shrink for free. Store the new operands for the new phi.
816	SmallVector<Value *, 4> NewIncoming;
817	unsigned NumZexts = 0;
818	unsigned NumConsts = 0;
819	for (Value *V : Phi.incoming_values()) {
820	if (auto *Zext = dyn_cast<ZExtInst>(Val: V)) {
821	// All zexts must be identical and have one user.
822	if (Zext->getSrcTy() != NarrowType || !Zext->hasOneUser())
823	return nullptr;
824	NewIncoming.push_back(Elt: Zext->getOperand(i_nocapture: 0));
825	NumZexts++;
826	} else if (auto *C = dyn_cast<Constant>(Val: V)) {
827	// Make sure that constants can fit in the new type.
828	Constant *Trunc = getLosslessUnsignedTrunc(C, TruncTy: NarrowType);
829	if (!Trunc)
830	return nullptr;
831	NewIncoming.push_back(Elt: Trunc);
832	NumConsts++;
833	} else {
834	// If it's not a cast or a constant, bail out.
835	return nullptr;
836	}
837	}
838
839	// The more common cases of a phi with no constant operands or just one
840	// variable operand are handled by FoldPHIArgOpIntoPHI() and foldOpIntoPhi()
841	// respectively. foldOpIntoPhi() wants to do the opposite transform that is
842	// performed here. It tries to replicate a cast in the phi operand's basic
843	// block to expose other folding opportunities. Thus, InstCombine will
844	// infinite loop without this check.
845	if (NumConsts == 0 || NumZexts < 2)
846	return nullptr;
847
848	// All incoming values are zexts or constants that are safe to truncate.
849	// Create a new phi node of the narrow type, phi together all of the new
850	// operands, and zext the result back to the original type.
851	PHINode *NewPhi = PHINode::Create(Ty: NarrowType, NumReservedValues: NumIncomingValues,
852	NameStr: Phi.getName() + ".shrunk");
853	for (unsigned I = 0; I != NumIncomingValues; ++I)
854	NewPhi->addIncoming(V: NewIncoming[I], BB: Phi.getIncomingBlock(i: I));
855
856	InsertNewInstBefore(New: NewPhi, Old: Phi.getIterator());
857	return CastInst::CreateZExtOrBitCast(S: NewPhi, Ty: Phi.getType());
858	}
859
860	/// If all operands to a PHI node are the same "unary" operator and they all are
861	/// only used by the PHI, PHI together their inputs, and do the operation once,
862	/// to the result of the PHI.
863	Instruction *InstCombinerImpl::foldPHIArgOpIntoPHI(PHINode &PN) {
864	// We cannot create a new instruction after the PHI if the terminator is an
865	// EHPad because there is no valid insertion point.
866	if (Instruction *TI = PN.getParent()->getTerminator())
867	if (TI->isEHPad())
868	return nullptr;
869
870	Instruction *FirstInst = cast<Instruction>(Val: PN.getIncomingValue(i: 0));
871
872	if (isa<GetElementPtrInst>(Val: FirstInst))
873	return foldPHIArgGEPIntoPHI(PN);
874	if (isa<LoadInst>(Val: FirstInst))
875	return foldPHIArgLoadIntoPHI(PN);
876	if (isa<InsertValueInst>(Val: FirstInst))
877	return foldPHIArgInsertValueInstructionIntoPHI(PN);
878	if (isa<ExtractValueInst>(Val: FirstInst))
879	return foldPHIArgExtractValueInstructionIntoPHI(PN);
880
881	// Scan the instruction, looking for input operations that can be folded away.
882	// If all input operands to the phi are the same instruction (e.g. a cast from
883	// the same type or "+42") we can pull the operation through the PHI, reducing
884	// code size and simplifying code.
885	Constant *ConstantOp = nullptr;
886	Type *CastSrcTy = nullptr;
887
888	if (isa<CastInst>(Val: FirstInst)) {
889	CastSrcTy = FirstInst->getOperand(i: 0)->getType();
890
891	// Be careful about transforming integer PHIs. We don't want to pessimize
892	// the code by turning an i32 into an i1293.
893	if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) {
894	if (!shouldChangeType(From: PN.getType(), To: CastSrcTy))
895	return nullptr;
896	}
897	} else if (isa<BinaryOperator>(Val: FirstInst) || isa<CmpInst>(Val: FirstInst)) {
898	// Can fold binop, compare or shift here if the RHS is a constant,
899	// otherwise call FoldPHIArgBinOpIntoPHI.
900	ConstantOp = dyn_cast<Constant>(Val: FirstInst->getOperand(i: 1));
901	if (!ConstantOp)
902	return foldPHIArgBinOpIntoPHI(PN);
903	} else {
904	return nullptr; // Cannot fold this operation.
905	}
906
907	// Check to see if all arguments are the same operation.
908	for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values())) {
909	Instruction *I = dyn_cast<Instruction>(Val: V);
910	if (!I || !I->hasOneUser() || !I->isSameOperationAs(I: FirstInst))
911	return nullptr;
912	if (CastSrcTy) {
913	if (I->getOperand(i: 0)->getType() != CastSrcTy)
914	return nullptr; // Cast operation must match.
915	} else if (I->getOperand(i: 1) != ConstantOp) {
916	return nullptr;
917	}
918	}
919
920	// Okay, they are all the same operation. Create a new PHI node of the
921	// correct type, and PHI together all of the LHS's of the instructions.
922	PHINode *NewPN = PHINode::Create(Ty: FirstInst->getOperand(i: 0)->getType(),
923	NumReservedValues: PN.getNumIncomingValues(),
924	NameStr: PN.getName()+".in");
925
926	Value *InVal = FirstInst->getOperand(i: 0);
927	NewPN->addIncoming(V: InVal, BB: PN.getIncomingBlock(i: 0));
928
929	// Add all operands to the new PHI.
930	for (auto Incoming : drop_begin(RangeOrContainer: zip(t: PN.blocks(), u: PN.incoming_values()))) {
931	BasicBlock *BB = std::get<0>(t&: Incoming);
932	Value *V = std::get<1>(t&: Incoming);
933	Value *NewInVal = cast<Instruction>(Val: V)->getOperand(i: 0);
934	if (NewInVal != InVal)
935	InVal = nullptr;
936	NewPN->addIncoming(V: NewInVal, BB);
937	}
938
939	Value *PhiVal;
940	if (InVal) {
941	// The new PHI unions all of the same values together. This is really
942	// common, so we handle it intelligently here for compile-time speed.
943	PhiVal = InVal;
944	delete NewPN;
945	} else {
946	InsertNewInstBefore(New: NewPN, Old: PN.getIterator());
947	PhiVal = NewPN;
948	}
949
950	// Insert and return the new operation.
951	if (CastInst *FirstCI = dyn_cast<CastInst>(Val: FirstInst)) {
952	CastInst *NewCI = CastInst::Create(FirstCI->getOpcode(), S: PhiVal,
953	Ty: PN.getType());
954	PHIArgMergedDebugLoc(Inst: NewCI, PN);
955	return NewCI;
956	}
957
958	if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Val: FirstInst)) {
959	BinOp = BinaryOperator::Create(Op: BinOp->getOpcode(), S1: PhiVal, S2: ConstantOp);
960	BinOp->copyIRFlags(V: PN.getIncomingValue(i: 0));
961
962	for (Value *V : drop_begin(RangeOrContainer: PN.incoming_values()))
963	BinOp->andIRFlags(V);
964
965	PHIArgMergedDebugLoc(Inst: BinOp, PN);
966	return BinOp;
967	}
968
969	CmpInst *CIOp = cast<CmpInst>(Val: FirstInst);
970	CmpInst *NewCI = CmpInst::Create(Op: CIOp->getOpcode(), Pred: CIOp->getPredicate(),
971	S1: PhiVal, S2: ConstantOp);
972	PHIArgMergedDebugLoc(Inst: NewCI, PN);
973	return NewCI;
974	}
975
976	/// Return true if this PHI node is only used by a PHI node cycle that is dead.
977	static bool isDeadPHICycle(PHINode *PN,
978	SmallPtrSetImpl<PHINode *> &PotentiallyDeadPHIs) {
979	if (PN->use_empty()) return true;
980	if (!PN->hasOneUse()) return false;
981
982	// Remember this node, and if we find the cycle, return.
983	if (!PotentiallyDeadPHIs.insert(Ptr: PN).second)
984	return true;
985
986	// Don't scan crazily complex things.
987	if (PotentiallyDeadPHIs.size() == 16)
988	return false;
989
990	if (PHINode *PU = dyn_cast<PHINode>(Val: PN->user_back()))
991	return isDeadPHICycle(PN: PU, PotentiallyDeadPHIs);
992
993	return false;
994	}
995
996	/// Return true if this phi node is always equal to NonPhiInVal.
997	/// This happens with mutually cyclic phi nodes like:
998	/// z = some value; x = phi (y, z); y = phi (x, z)
999	static bool PHIsEqualValue(PHINode *PN, Value *&NonPhiInVal,
1000	SmallPtrSetImpl<PHINode *> &ValueEqualPHIs) {
1001	// See if we already saw this PHI node.
1002	if (!ValueEqualPHIs.insert(Ptr: PN).second)
1003	return true;
1004
1005	// Don't scan crazily complex things.
1006	if (ValueEqualPHIs.size() == 16)
1007	return false;
1008
1009	// Scan the operands to see if they are either phi nodes or are equal to
1010	// the value.
1011	for (Value *Op : PN->incoming_values()) {
1012	if (PHINode *OpPN = dyn_cast<PHINode>(Val: Op)) {
1013	if (!PHIsEqualValue(PN: OpPN, NonPhiInVal, ValueEqualPHIs)) {
1014	if (NonPhiInVal)
1015	return false;
1016	NonPhiInVal = OpPN;
1017	}
1018	} else if (Op != NonPhiInVal)
1019	return false;
1020	}
1021
1022	return true;
1023	}
1024
1025	/// Return an existing non-zero constant if this phi node has one, otherwise
1026	/// return constant 1.
1027	static ConstantInt *getAnyNonZeroConstInt(PHINode &PN) {
1028	assert(isa<IntegerType>(PN.getType()) && "Expect only integer type phi");
1029	for (Value *V : PN.operands())
1030	if (auto *ConstVA = dyn_cast<ConstantInt>(Val: V))
1031	if (!ConstVA->isZero())
1032	return ConstVA;
1033	return ConstantInt::get(Ty: cast<IntegerType>(Val: PN.getType()), V: 1);
1034	}
1035
1036	namespace {
1037	struct PHIUsageRecord {
1038	unsigned PHIId; // The ID # of the PHI (something determinstic to sort on)
1039	unsigned Shift; // The amount shifted.
1040	Instruction *Inst; // The trunc instruction.
1041
1042	PHIUsageRecord(unsigned Pn, unsigned Sh, Instruction *User)
1043	: PHIId(Pn), Shift(Sh), Inst(User) {}
1044
1045	bool operator<(const PHIUsageRecord &RHS) const {
1046	if (PHIId < RHS.PHIId) return true;
1047	if (PHIId > RHS.PHIId) return false;
1048	if (Shift < RHS.Shift) return true;
1049	if (Shift > RHS.Shift) return false;
1050	return Inst->getType()->getPrimitiveSizeInBits() <
1051	RHS.Inst->getType()->getPrimitiveSizeInBits();
1052	}
1053	};
1054
1055	struct LoweredPHIRecord {
1056	PHINode *PN; // The PHI that was lowered.
1057	unsigned Shift; // The amount shifted.
1058	unsigned Width; // The width extracted.
1059
1060	LoweredPHIRecord(PHINode *Phi, unsigned Sh, Type *Ty)
1061	: PN(Phi), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {}
1062
1063	// Ctor form used by DenseMap.
1064	LoweredPHIRecord(PHINode *Phi, unsigned Sh) : PN(Phi), Shift(Sh), Width(0) {}
1065	};
1066	} // namespace
1067
1068	namespace llvm {
1069	template<>
1070	struct DenseMapInfo<LoweredPHIRecord> {
1071	static inline LoweredPHIRecord getEmptyKey() {
1072	return LoweredPHIRecord(nullptr, 0);
1073	}
1074	static inline LoweredPHIRecord getTombstoneKey() {
1075	return LoweredPHIRecord(nullptr, 1);
1076	}
1077	static unsigned getHashValue(const LoweredPHIRecord &Val) {
1078	return DenseMapInfo<PHINode*>::getHashValue(PtrVal: Val.PN) ^ (Val.Shift>>3) ^
1079	(Val.Width>>3);
1080	}
1081	static bool isEqual(const LoweredPHIRecord &LHS,
1082	const LoweredPHIRecord &RHS) {
1083	return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift &&
1084	LHS.Width == RHS.Width;
1085	}
1086	};
1087	} // namespace llvm
1088
1089
1090	/// This is an integer PHI and we know that it has an illegal type: see if it is
1091	/// only used by trunc or trunc(lshr) operations. If so, we split the PHI into
1092	/// the various pieces being extracted. This sort of thing is introduced when
1093	/// SROA promotes an aggregate to large integer values.
1094	///
1095	/// TODO: The user of the trunc may be an bitcast to float/double/vector or an
1096	/// inttoptr. We should produce new PHIs in the right type.
1097	///
1098	Instruction *InstCombinerImpl::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) {
1099	// PHIUsers - Keep track of all of the truncated values extracted from a set
1100	// of PHIs, along with their offset. These are the things we want to rewrite.
1101	SmallVector<PHIUsageRecord, 16> PHIUsers;
1102
1103	// PHIs are often mutually cyclic, so we keep track of a whole set of PHI
1104	// nodes which are extracted from. PHIsToSlice is a set we use to avoid
1105	// revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to
1106	// check the uses of (to ensure they are all extracts).
1107	SmallVector<PHINode*, 8> PHIsToSlice;
1108	SmallPtrSet<PHINode*, 8> PHIsInspected;
1109
1110	PHIsToSlice.push_back(Elt: &FirstPhi);
1111	PHIsInspected.insert(Ptr: &FirstPhi);
1112
1113	for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) {
1114	PHINode *PN = PHIsToSlice[PHIId];
1115
1116	// Scan the input list of the PHI. If any input is an invoke, and if the
1117	// input is defined in the predecessor, then we won't be split the critical
1118	// edge which is required to insert a truncate. Because of this, we have to
1119	// bail out.
1120	for (auto Incoming : zip(t: PN->blocks(), u: PN->incoming_values())) {
1121	BasicBlock *BB = std::get<0>(t&: Incoming);
1122	Value *V = std::get<1>(t&: Incoming);
1123	InvokeInst *II = dyn_cast<InvokeInst>(Val: V);
1124	if (!II)
1125	continue;
1126	if (II->getParent() != BB)
1127	continue;
1128
1129	// If we have a phi, and if it's directly in the predecessor, then we have
1130	// a critical edge where we need to put the truncate. Since we can't
1131	// split the edge in instcombine, we have to bail out.
1132	return nullptr;
1133	}
1134
1135	// If the incoming value is a PHI node before a catchswitch, we cannot
1136	// extract the value within that BB because we cannot insert any non-PHI
1137	// instructions in the BB.
1138	for (auto *Pred : PN->blocks())
1139	if (Pred->getFirstInsertionPt() == Pred->end())
1140	return nullptr;
1141
1142	for (User *U : PN->users()) {
1143	Instruction *UserI = cast<Instruction>(Val: U);
1144
1145	// If the user is a PHI, inspect its uses recursively.
1146	if (PHINode *UserPN = dyn_cast<PHINode>(Val: UserI)) {
1147	if (PHIsInspected.insert(Ptr: UserPN).second)
1148	PHIsToSlice.push_back(Elt: UserPN);
1149	continue;
1150	}
1151
1152	// Truncates are always ok.
1153	if (isa<TruncInst>(Val: UserI)) {
1154	PHIUsers.push_back(Elt: PHIUsageRecord(PHIId, 0, UserI));
1155	continue;
1156	}
1157
1158	// Otherwise it must be a lshr which can only be used by one trunc.
1159	if (UserI->getOpcode() != Instruction::LShr ||
1160	!UserI->hasOneUse() || !isa<TruncInst>(Val: UserI->user_back()) ||
1161	!isa<ConstantInt>(Val: UserI->getOperand(i: 1)))
1162	return nullptr;
1163
1164	// Bail on out of range shifts.
1165	unsigned SizeInBits = UserI->getType()->getScalarSizeInBits();
1166	if (cast<ConstantInt>(Val: UserI->getOperand(i: 1))->getValue().uge(RHS: SizeInBits))
1167	return nullptr;
1168
1169	unsigned Shift = cast<ConstantInt>(Val: UserI->getOperand(i: 1))->getZExtValue();
1170	PHIUsers.push_back(Elt: PHIUsageRecord(PHIId, Shift, UserI->user_back()));
1171	}
1172	}
1173
1174	// If we have no users, they must be all self uses, just nuke the PHI.
1175	if (PHIUsers.empty())
1176	return replaceInstUsesWith(I&: FirstPhi, V: PoisonValue::get(T: FirstPhi.getType()));
1177
1178	// If this phi node is transformable, create new PHIs for all the pieces
1179	// extracted out of it. First, sort the users by their offset and size.
1180	array_pod_sort(Start: PHIUsers.begin(), End: PHIUsers.end());
1181
1182	LLVM_DEBUG(dbgs() << "SLICING UP PHI: " << FirstPhi << '\n';
1183	for (unsigned I = 1; I != PHIsToSlice.size(); ++I) dbgs()
1184	<< "AND USER PHI #" << I << ": " << *PHIsToSlice[I] << '\n');
1185
1186	// PredValues - This is a temporary used when rewriting PHI nodes. It is
1187	// hoisted out here to avoid construction/destruction thrashing.
1188	DenseMap<BasicBlock*, Value*> PredValues;
1189
1190	// ExtractedVals - Each new PHI we introduce is saved here so we don't
1191	// introduce redundant PHIs.
1192	DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals;
1193
1194	for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) {
1195	unsigned PHIId = PHIUsers[UserI].PHIId;
1196	PHINode *PN = PHIsToSlice[PHIId];
1197	unsigned Offset = PHIUsers[UserI].Shift;
1198	Type *Ty = PHIUsers[UserI].Inst->getType();
1199
1200	PHINode *EltPHI;
1201
1202	// If we've already lowered a user like this, reuse the previously lowered
1203	// value.
1204	if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == nullptr) {
1205
1206	// Otherwise, Create the new PHI node for this user.
1207	EltPHI = PHINode::Create(Ty, NumReservedValues: PN->getNumIncomingValues(),
1208	NameStr: PN->getName() + ".off" + Twine(Offset),
1209	InsertBefore: PN->getIterator());
1210	assert(EltPHI->getType() != PN->getType() &&
1211	"Truncate didn't shrink phi?");
1212
1213	for (auto Incoming : zip(t: PN->blocks(), u: PN->incoming_values())) {
1214	BasicBlock *Pred = std::get<0>(t&: Incoming);
1215	Value *InVal = std::get<1>(t&: Incoming);
1216	Value *&PredVal = PredValues[Pred];
1217
1218	// If we already have a value for this predecessor, reuse it.
1219	if (PredVal) {
1220	EltPHI->addIncoming(V: PredVal, BB: Pred);
1221	continue;
1222	}
1223
1224	// Handle the PHI self-reuse case.
1225	if (InVal == PN) {
1226	PredVal = EltPHI;
1227	EltPHI->addIncoming(V: PredVal, BB: Pred);
1228	continue;
1229	}
1230
1231	if (PHINode *InPHI = dyn_cast<PHINode>(Val: PN)) {
1232	// If the incoming value was a PHI, and if it was one of the PHIs we
1233	// already rewrote it, just use the lowered value.
1234	if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) {
1235	PredVal = Res;
1236	EltPHI->addIncoming(V: PredVal, BB: Pred);
1237	continue;
1238	}
1239	}
1240
1241	// Otherwise, do an extract in the predecessor.
1242	Builder.SetInsertPoint(Pred->getTerminator());
1243	Value *Res = InVal;
1244	if (Offset)
1245	Res = Builder.CreateLShr(
1246	LHS: Res, RHS: ConstantInt::get(Ty: InVal->getType(), V: Offset), Name: "extract");
1247	Res = Builder.CreateTrunc(V: Res, DestTy: Ty, Name: "extract.t");
1248	PredVal = Res;
1249	EltPHI->addIncoming(V: Res, BB: Pred);
1250
1251	// If the incoming value was a PHI, and if it was one of the PHIs we are
1252	// rewriting, we will ultimately delete the code we inserted. This
1253	// means we need to revisit that PHI to make sure we extract out the
1254	// needed piece.
1255	if (PHINode *OldInVal = dyn_cast<PHINode>(Val: InVal))
1256	if (PHIsInspected.count(Ptr: OldInVal)) {
1257	unsigned RefPHIId =
1258	find(Range&: PHIsToSlice, Val: OldInVal) - PHIsToSlice.begin();
1259	PHIUsers.push_back(
1260	Elt: PHIUsageRecord(RefPHIId, Offset, cast<Instruction>(Val: Res)));
1261	++UserE;
1262	}
1263	}
1264	PredValues.clear();
1265
1266	LLVM_DEBUG(dbgs() << " Made element PHI for offset " << Offset << ": "
1267	<< *EltPHI << '\n');
1268	ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI;
1269	}
1270
1271	// Replace the use of this piece with the PHI node.
1272	replaceInstUsesWith(I&: *PHIUsers[UserI].Inst, V: EltPHI);
1273	}
1274
1275	// Replace all the remaining uses of the PHI nodes (self uses and the lshrs)
1276	// with poison.
1277	Value *Poison = PoisonValue::get(T: FirstPhi.getType());
1278	for (PHINode *PHI : drop_begin(RangeOrContainer&: PHIsToSlice))
1279	replaceInstUsesWith(I&: *PHI, V: Poison);
1280	return replaceInstUsesWith(I&: FirstPhi, V: Poison);
1281	}
1282
1283	static Value *simplifyUsingControlFlow(InstCombiner &Self, PHINode &PN,
1284	const DominatorTree &DT) {
1285	// Simplify the following patterns:
1286	// if (cond)
1287	// / \
1288	// ... ...
1289	// \ /
1290	// phi [true] [false]
1291	// and
1292	// switch (cond)
1293	// case v1: / \ case v2:
1294	// ... ...
1295	// \ /
1296	// phi [v1] [v2]
1297	// Make sure all inputs are constants.
1298	if (!all_of(Range: PN.operands(), P: [](Value *V) { return isa<ConstantInt>(Val: V); }))
1299	return nullptr;
1300
1301	BasicBlock *BB = PN.getParent();
1302	// Do not bother with unreachable instructions.
1303	if (!DT.isReachableFromEntry(A: BB))
1304	return nullptr;
1305
1306	// Determine which value the condition of the idom has for which successor.
1307	LLVMContext &Context = PN.getContext();
1308	auto *IDom = DT.getNode(BB)->getIDom()->getBlock();
1309	Value *Cond;
1310	SmallDenseMap<ConstantInt *, BasicBlock *, 8> SuccForValue;
1311	SmallDenseMap<BasicBlock *, unsigned, 8> SuccCount;
1312	auto AddSucc = [&](ConstantInt *C, BasicBlock *Succ) {
1313	SuccForValue[C] = Succ;
1314	++SuccCount[Succ];
1315	};
1316	if (auto *BI = dyn_cast<BranchInst>(Val: IDom->getTerminator())) {
1317	if (BI->isUnconditional())
1318	return nullptr;
1319
1320	Cond = BI->getCondition();
1321	AddSucc(ConstantInt::getTrue(Context), BI->getSuccessor(i: 0));
1322	AddSucc(ConstantInt::getFalse(Context), BI->getSuccessor(i: 1));
1323	} else if (auto *SI = dyn_cast<SwitchInst>(Val: IDom->getTerminator())) {
1324	Cond = SI->getCondition();
1325	++SuccCount[SI->getDefaultDest()];
1326	for (auto Case : SI->cases())
1327	AddSucc(Case.getCaseValue(), Case.getCaseSuccessor());
1328	} else {
1329	return nullptr;
1330	}
1331
1332	if (Cond->getType() != PN.getType())
1333	return nullptr;
1334
1335	// Check that edges outgoing from the idom's terminators dominate respective
1336	// inputs of the Phi.
1337	std::optional<bool> Invert;
1338	for (auto Pair : zip(t: PN.incoming_values(), u: PN.blocks())) {
1339	auto *Input = cast<ConstantInt>(Val&: std::get<0>(t&: Pair));
1340	BasicBlock *Pred = std::get<1>(t&: Pair);
1341	auto IsCorrectInput = [&](ConstantInt *Input) {
1342	// The input needs to be dominated by the corresponding edge of the idom.
1343	// This edge cannot be a multi-edge, as that would imply that multiple
1344	// different condition values follow the same edge.
1345	auto It = SuccForValue.find(Val: Input);
1346	return It != SuccForValue.end() && SuccCount[It->second] == 1 &&
1347	DT.dominates(BBE1: BasicBlockEdge(IDom, It->second),
1348	BBE2: BasicBlockEdge(Pred, BB));
1349	};
1350
1351	// Depending on the constant, the condition may need to be inverted.
1352	bool NeedsInvert;
1353	if (IsCorrectInput(Input))
1354	NeedsInvert = false;
1355	else if (IsCorrectInput(cast<ConstantInt>(Val: ConstantExpr::getNot(C: Input))))
1356	NeedsInvert = true;
1357	else
1358	return nullptr;
1359
1360	// Make sure the inversion requirement is always the same.
1361	if (Invert && *Invert != NeedsInvert)
1362	return nullptr;
1363
1364	Invert = NeedsInvert;
1365	}
1366
1367	if (!*Invert)
1368	return Cond;
1369
1370	// This Phi is actually opposite to branching condition of IDom. We invert
1371	// the condition that will potentially open up some opportunities for
1372	// sinking.
1373	auto InsertPt = BB->getFirstInsertionPt();
1374	if (InsertPt != BB->end()) {
1375	Self.Builder.SetInsertPoint(TheBB: &*BB, IP: InsertPt);
1376	return Self.Builder.CreateNot(V: Cond);
1377	}
1378
1379	return nullptr;
1380	}
1381
1382	// Fold iv = phi(start, iv.next = iv2.next op start)
1383	// where iv2 = phi(iv2.start, iv2.next = iv2 + iv2.step)
1384	// and iv2.start op start = start
1385	// to iv = iv2 op start
1386	static Value *foldDependentIVs(PHINode &PN, IRBuilderBase &Builder) {
1387	BasicBlock *BB = PN.getParent();
1388	if (PN.getNumIncomingValues() != 2)
1389	return nullptr;
1390
1391	Value *Start;
1392	Instruction *IvNext;
1393	BinaryOperator *Iv2Next;
1394	auto MatchOuterIV = [&](Value *V1, Value *V2) {
1395	if (match(V: V2, P: m_c_BinOp(L: m_Specific(V: V1), R: m_BinOp(I&: Iv2Next))) ||
1396	match(V: V2, P: m_GEP(Ops: m_Specific(V: V1), Ops: m_BinOp(I&: Iv2Next)))) {
1397	Start = V1;
1398	IvNext = cast<Instruction>(Val: V2);
1399	return true;
1400	}
1401	return false;
1402	};
1403
1404	if (!MatchOuterIV(PN.getIncomingValue(i: 0), PN.getIncomingValue(i: 1)) &&
1405	!MatchOuterIV(PN.getIncomingValue(i: 1), PN.getIncomingValue(i: 0)))
1406	return nullptr;
1407
1408	PHINode *Iv2;
1409	Value *Iv2Start, *Iv2Step;
1410	if (!matchSimpleRecurrence(I: Iv2Next, P&: Iv2, Start&: Iv2Start, Step&: Iv2Step) ||
1411	Iv2->getParent() != BB)
1412	return nullptr;
1413
1414	auto *BO = dyn_cast<BinaryOperator>(Val: IvNext);
1415	Constant *Identity =
1416	BO ? ConstantExpr::getBinOpIdentity(Opcode: BO->getOpcode(), Ty: Iv2Start->getType())
1417	: Constant::getNullValue(Ty: Iv2Start->getType());
1418	if (Iv2Start != Identity)
1419	return nullptr;
1420
1421	Builder.SetInsertPoint(TheBB: &*BB, IP: BB->getFirstInsertionPt());
1422	if (!BO) {
1423	auto *GEP = cast<GEPOperator>(Val: IvNext);
1424	return Builder.CreateGEP(Ty: GEP->getSourceElementType(), Ptr: Start, IdxList: Iv2, Name: "",
1425	IsInBounds: cast<GEPOperator>(Val: IvNext)->isInBounds());
1426	}
1427
1428	assert(BO->isCommutative() && "Must be commutative");
1429	Value *Res = Builder.CreateBinOp(Opc: BO->getOpcode(), LHS: Iv2, RHS: Start);
1430	cast<Instruction>(Val: Res)->copyIRFlags(V: BO);
1431	return Res;
1432	}
1433
1434	// PHINode simplification
1435	//
1436	Instruction *InstCombinerImpl::visitPHINode(PHINode &PN) {
1437	if (Value *V = simplifyInstruction(I: &PN, Q: SQ.getWithInstruction(I: &PN)))
1438	return replaceInstUsesWith(I&: PN, V);
1439
1440	if (Instruction *Result = foldPHIArgZextsIntoPHI(Phi&: PN))
1441	return Result;
1442
1443	if (Instruction *Result = foldPHIArgIntToPtrToPHI(PN))
1444	return Result;
1445
1446	// If all PHI operands are the same operation, pull them through the PHI,
1447	// reducing code size.
1448	auto *Inst0 = dyn_cast<Instruction>(Val: PN.getIncomingValue(i: 0));
1449	auto *Inst1 = dyn_cast<Instruction>(Val: PN.getIncomingValue(i: 1));
1450	if (Inst0 && Inst1 && Inst0->getOpcode() == Inst1->getOpcode() &&
1451	Inst0->hasOneUser())
1452	if (Instruction *Result = foldPHIArgOpIntoPHI(PN))
1453	return Result;
1454
1455	// If the incoming values are pointer casts of the same original value,
1456	// replace the phi with a single cast iff we can insert a non-PHI instruction.
1457	if (PN.getType()->isPointerTy() &&
1458	PN.getParent()->getFirstInsertionPt() != PN.getParent()->end()) {
1459	Value *IV0 = PN.getIncomingValue(i: 0);
1460	Value *IV0Stripped = IV0->stripPointerCasts();
1461	// Set to keep track of values known to be equal to IV0Stripped after
1462	// stripping pointer casts.
1463	SmallPtrSet<Value *, 4> CheckedIVs;
1464	CheckedIVs.insert(Ptr: IV0);
1465	if (IV0 != IV0Stripped &&
1466	all_of(Range: PN.incoming_values(), P: [&CheckedIVs, IV0Stripped](Value *IV) {
1467	return !CheckedIVs.insert(Ptr: IV).second ||
1468	IV0Stripped == IV->stripPointerCasts();
1469	})) {
1470	return CastInst::CreatePointerCast(S: IV0Stripped, Ty: PN.getType());
1471	}
1472	}
1473
1474	// If this is a trivial cycle in the PHI node graph, remove it. Basically, if
1475	// this PHI only has a single use (a PHI), and if that PHI only has one use (a
1476	// PHI)... break the cycle.
1477	if (PN.hasOneUse()) {
1478	if (foldIntegerTypedPHI(PN))
1479	return nullptr;
1480
1481	Instruction *PHIUser = cast<Instruction>(Val: PN.user_back());
1482	if (PHINode *PU = dyn_cast<PHINode>(Val: PHIUser)) {
1483	SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
1484	PotentiallyDeadPHIs.insert(Ptr: &PN);
1485	if (isDeadPHICycle(PN: PU, PotentiallyDeadPHIs))
1486	return replaceInstUsesWith(I&: PN, V: PoisonValue::get(T: PN.getType()));
1487	}
1488
1489	// If this phi has a single use, and if that use just computes a value for
1490	// the next iteration of a loop, delete the phi. This occurs with unused
1491	// induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
1492	// common case here is good because the only other things that catch this
1493	// are induction variable analysis (sometimes) and ADCE, which is only run
1494	// late.
1495	if (PHIUser->hasOneUse() &&
1496	(isa<BinaryOperator>(Val: PHIUser) || isa<UnaryOperator>(Val: PHIUser) ||
1497	isa<GetElementPtrInst>(Val: PHIUser)) &&
1498	PHIUser->user_back() == &PN) {
1499	return replaceInstUsesWith(I&: PN, V: PoisonValue::get(T: PN.getType()));
1500	}
1501	}
1502
1503	// When a PHI is used only to be compared with zero, it is safe to replace
1504	// an incoming value proved as known nonzero with any non-zero constant.
1505	// For example, in the code below, the incoming value %v can be replaced
1506	// with any non-zero constant based on the fact that the PHI is only used to
1507	// be compared with zero and %v is a known non-zero value:
1508	// %v = select %cond, 1, 2
1509	// %p = phi [%v, BB] ...
1510	// icmp eq, %p, 0
1511	// FIXME: To be simple, handle only integer type for now.
1512	// This handles a small number of uses to keep the complexity down, and an
1513	// icmp(or(phi)) can equally be replaced with any non-zero constant as the
1514	// "or" will only add bits.
1515	if (!PN.hasNUsesOrMore(N: 3)) {
1516	SmallVector<Instruction *> DropPoisonFlags;
1517	bool AllUsesOfPhiEndsInCmp = all_of(Range: PN.users(), P: [&](User *U) {
1518	auto *CmpInst = dyn_cast<ICmpInst>(Val: U);
1519	if (!CmpInst) {
1520	// This is always correct as OR only add bits and we are checking
1521	// against 0.
1522	if (U->hasOneUse() && match(V: U, P: m_c_Or(L: m_Specific(V: &PN), R: m_Value()))) {
1523	DropPoisonFlags.push_back(Elt: cast<Instruction>(Val: U));
1524	CmpInst = dyn_cast<ICmpInst>(Val: U->user_back());
1525	}
1526	}
1527	if (!CmpInst || !isa<IntegerType>(Val: PN.getType()) ||
1528	!CmpInst->isEquality() || !match(V: CmpInst->getOperand(i_nocapture: 1), P: m_Zero())) {
1529	return false;
1530	}
1531	return true;
1532	});
1533	// All uses of PHI results in a compare with zero.
1534	if (AllUsesOfPhiEndsInCmp) {
1535	ConstantInt *NonZeroConst = nullptr;
1536	bool MadeChange = false;
1537	for (unsigned I = 0, E = PN.getNumIncomingValues(); I != E; ++I) {
1538	Instruction *CtxI = PN.getIncomingBlock(i: I)->getTerminator();
1539	Value *VA = PN.getIncomingValue(i: I);
1540	if (isKnownNonZero(V: VA, Q: getSimplifyQuery().getWithInstruction(I: CtxI))) {
1541	if (!NonZeroConst)
1542	NonZeroConst = getAnyNonZeroConstInt(PN);
1543	if (NonZeroConst != VA) {
1544	replaceOperand(I&: PN, OpNum: I, V: NonZeroConst);
1545	// The "disjoint" flag may no longer hold after the transform.
1546	for (Instruction *I : DropPoisonFlags)
1547	I->dropPoisonGeneratingFlags();
1548	MadeChange = true;
1549	}
1550	}
1551	}
1552	if (MadeChange)
1553	return &PN;
1554	}
1555	}
1556
1557	// We sometimes end up with phi cycles that non-obviously end up being the
1558	// same value, for example:
1559	// z = some value; x = phi (y, z); y = phi (x, z)
1560	// where the phi nodes don't necessarily need to be in the same block. Do a
1561	// quick check to see if the PHI node only contains a single non-phi value, if
1562	// so, scan to see if the phi cycle is actually equal to that value. If the
1563	// phi has no non-phi values then allow the "NonPhiInVal" to be set later if
1564	// one of the phis itself does not have a single input.
1565	{
1566	unsigned InValNo = 0, NumIncomingVals = PN.getNumIncomingValues();
1567	// Scan for the first non-phi operand.
1568	while (InValNo != NumIncomingVals &&
1569	isa<PHINode>(Val: PN.getIncomingValue(i: InValNo)))
1570	++InValNo;
1571
1572	Value *NonPhiInVal =
1573	InValNo != NumIncomingVals ? PN.getIncomingValue(i: InValNo) : nullptr;
1574
1575	// Scan the rest of the operands to see if there are any conflicts, if so
1576	// there is no need to recursively scan other phis.
1577	if (NonPhiInVal)
1578	for (++InValNo; InValNo != NumIncomingVals; ++InValNo) {
1579	Value *OpVal = PN.getIncomingValue(i: InValNo);
1580	if (OpVal != NonPhiInVal && !isa<PHINode>(Val: OpVal))
1581	break;
1582	}
1583
1584	// If we scanned over all operands, then we have one unique value plus
1585	// phi values. Scan PHI nodes to see if they all merge in each other or
1586	// the value.
1587	if (InValNo == NumIncomingVals) {
1588	SmallPtrSet<PHINode *, 16> ValueEqualPHIs;
1589	if (PHIsEqualValue(PN: &PN, NonPhiInVal, ValueEqualPHIs))
1590	return replaceInstUsesWith(I&: PN, V: NonPhiInVal);
1591	}
1592	}
1593
1594	// If there are multiple PHIs, sort their operands so that they all list
1595	// the blocks in the same order. This will help identical PHIs be eliminated
1596	// by other passes. Other passes shouldn't depend on this for correctness
1597	// however.
1598	auto Res = PredOrder.try_emplace(Key: PN.getParent());
1599	if (!Res.second) {
1600	const auto &Preds = Res.first->second;
1601	for (unsigned I = 0, E = PN.getNumIncomingValues(); I != E; ++I) {
1602	BasicBlock *BBA = PN.getIncomingBlock(i: I);
1603	BasicBlock *BBB = Preds[I];
1604	if (BBA != BBB) {
1605	Value *VA = PN.getIncomingValue(i: I);
1606	unsigned J = PN.getBasicBlockIndex(BB: BBB);
1607	Value *VB = PN.getIncomingValue(i: J);
1608	PN.setIncomingBlock(i: I, BB: BBB);
1609	PN.setIncomingValue(i: I, V: VB);
1610	PN.setIncomingBlock(i: J, BB: BBA);
1611	PN.setIncomingValue(i: J, V: VA);
1612	// NOTE: Instcombine normally would want us to "return &PN" if we
1613	// modified any of the operands of an instruction. However, since we
1614	// aren't adding or removing uses (just rearranging them) we don't do
1615	// this in this case.
1616	}
1617	}
1618	} else {
1619	// Remember the block order of the first encountered phi node.
1620	append_range(C&: Res.first->second, R: PN.blocks());
1621	}
1622
1623	// Is there an identical PHI node in this basic block?
1624	for (PHINode &IdenticalPN : PN.getParent()->phis()) {
1625	// Ignore the PHI node itself.
1626	if (&IdenticalPN == &PN)
1627	continue;
1628	// Note that even though we've just canonicalized this PHI, due to the
1629	// worklist visitation order, there are no guarantess that *every* PHI
1630	// has been canonicalized, so we can't just compare operands ranges.
1631	if (!PN.isIdenticalToWhenDefined(I: &IdenticalPN))
1632	continue;
1633	// Just use that PHI instead then.
1634	++NumPHICSEs;
1635	return replaceInstUsesWith(I&: PN, V: &IdenticalPN);
1636	}
1637
1638	// If this is an integer PHI and we know that it has an illegal type, see if
1639	// it is only used by trunc or trunc(lshr) operations. If so, we split the
1640	// PHI into the various pieces being extracted. This sort of thing is
1641	// introduced when SROA promotes an aggregate to a single large integer type.
1642	if (PN.getType()->isIntegerTy() &&
1643	!DL.isLegalInteger(Width: PN.getType()->getPrimitiveSizeInBits()))
1644	if (Instruction *Res = SliceUpIllegalIntegerPHI(FirstPhi&: PN))
1645	return Res;
1646
1647	// Ultimately, try to replace this Phi with a dominating condition.
1648	if (auto *V = simplifyUsingControlFlow(Self&: *this, PN, DT))
1649	return replaceInstUsesWith(I&: PN, V);
1650
1651	if (Value *Res = foldDependentIVs(PN, Builder))
1652	return replaceInstUsesWith(I&: PN, V: Res);
1653
1654	return nullptr;
1655	}
1656