xref: /freebsd/contrib/llvm-project/llvm/lib/Target/Hexagon/HexagonLoopIdiomRecognition.cpp (revision 0ad011ececb978e22a9bff2acf76633b094f1ff6)
1 //===- HexagonLoopIdiomRecognition.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 #include "HexagonLoopIdiomRecognition.h"
10 #include "llvm/ADT/APInt.h"
11 #include "llvm/ADT/DenseMap.h"
12 #include "llvm/ADT/SetVector.h"
13 #include "llvm/ADT/SmallPtrSet.h"
14 #include "llvm/ADT/SmallSet.h"
15 #include "llvm/ADT/SmallVector.h"
16 #include "llvm/ADT/StringRef.h"
17 #include "llvm/Analysis/AliasAnalysis.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/LoopAnalysisManager.h"
20 #include "llvm/Analysis/LoopInfo.h"
21 #include "llvm/Analysis/LoopPass.h"
22 #include "llvm/Analysis/MemoryLocation.h"
23 #include "llvm/Analysis/ScalarEvolution.h"
24 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
25 #include "llvm/Analysis/TargetLibraryInfo.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/IR/Attributes.h"
28 #include "llvm/IR/BasicBlock.h"
29 #include "llvm/IR/Constant.h"
30 #include "llvm/IR/Constants.h"
31 #include "llvm/IR/DataLayout.h"
32 #include "llvm/IR/DebugLoc.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/IRBuilder.h"
37 #include "llvm/IR/InstrTypes.h"
38 #include "llvm/IR/Instruction.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/IntrinsicInst.h"
41 #include "llvm/IR/Intrinsics.h"
42 #include "llvm/IR/IntrinsicsHexagon.h"
43 #include "llvm/IR/Module.h"
44 #include "llvm/IR/PassManager.h"
45 #include "llvm/IR/PatternMatch.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/IR/User.h"
48 #include "llvm/IR/Value.h"
49 #include "llvm/InitializePasses.h"
50 #include "llvm/Pass.h"
51 #include "llvm/Support/Casting.h"
52 #include "llvm/Support/CommandLine.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/Support/Debug.h"
55 #include "llvm/Support/ErrorHandling.h"
56 #include "llvm/Support/KnownBits.h"
57 #include "llvm/Support/raw_ostream.h"
58 #include "llvm/TargetParser/Triple.h"
59 #include "llvm/Transforms/Scalar.h"
60 #include "llvm/Transforms/Utils.h"
61 #include "llvm/Transforms/Utils/Local.h"
62 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
63 #include <algorithm>
64 #include <array>
65 #include <cassert>
66 #include <cstdint>
67 #include <cstdlib>
68 #include <deque>
69 #include <functional>
70 #include <iterator>
71 #include <map>
72 #include <set>
73 #include <utility>
74 #include <vector>
75 
76 #define DEBUG_TYPE "hexagon-lir"
77 
78 using namespace llvm;
79 
80 static cl::opt<bool> DisableMemcpyIdiom("disable-memcpy-idiom",
81   cl::Hidden, cl::init(false),
82   cl::desc("Disable generation of memcpy in loop idiom recognition"));
83 
84 static cl::opt<bool> DisableMemmoveIdiom("disable-memmove-idiom",
85   cl::Hidden, cl::init(false),
86   cl::desc("Disable generation of memmove in loop idiom recognition"));
87 
88 static cl::opt<unsigned> RuntimeMemSizeThreshold("runtime-mem-idiom-threshold",
89   cl::Hidden, cl::init(0), cl::desc("Threshold (in bytes) for the runtime "
90   "check guarding the memmove."));
91 
92 static cl::opt<unsigned> CompileTimeMemSizeThreshold(
93   "compile-time-mem-idiom-threshold", cl::Hidden, cl::init(64),
94   cl::desc("Threshold (in bytes) to perform the transformation, if the "
95     "runtime loop count (mem transfer size) is known at compile-time."));
96 
97 static cl::opt<bool> OnlyNonNestedMemmove("only-nonnested-memmove-idiom",
98   cl::Hidden, cl::init(true),
99   cl::desc("Only enable generating memmove in non-nested loops"));
100 
101 static cl::opt<bool> HexagonVolatileMemcpy(
102     "disable-hexagon-volatile-memcpy", cl::Hidden, cl::init(false),
103     cl::desc("Enable Hexagon-specific memcpy for volatile destination."));
104 
105 static cl::opt<unsigned> SimplifyLimit("hlir-simplify-limit", cl::init(10000),
106   cl::Hidden, cl::desc("Maximum number of simplification steps in HLIR"));
107 
108 static const char *HexagonVolatileMemcpyName
109   = "hexagon_memcpy_forward_vp4cp4n2";
110 
111 
112 namespace llvm {
113 
114 void initializeHexagonLoopIdiomRecognizeLegacyPassPass(PassRegistry &);
115 Pass *createHexagonLoopIdiomPass();
116 
117 } // end namespace llvm
118 
119 namespace {
120 
121 class HexagonLoopIdiomRecognize {
122 public:
123   explicit HexagonLoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT,
124                                      LoopInfo *LF, const TargetLibraryInfo *TLI,
125                                      ScalarEvolution *SE)
126       : AA(AA), DT(DT), LF(LF), TLI(TLI), SE(SE) {}
127 
128   bool run(Loop *L);
129 
130 private:
131   int getSCEVStride(const SCEVAddRecExpr *StoreEv);
132   bool isLegalStore(Loop *CurLoop, StoreInst *SI);
133   void collectStores(Loop *CurLoop, BasicBlock *BB,
134                      SmallVectorImpl<StoreInst *> &Stores);
135   bool processCopyingStore(Loop *CurLoop, StoreInst *SI, const SCEV *BECount);
136   bool coverLoop(Loop *L, SmallVectorImpl<Instruction *> &Insts) const;
137   bool runOnLoopBlock(Loop *CurLoop, BasicBlock *BB, const SCEV *BECount,
138                       SmallVectorImpl<BasicBlock *> &ExitBlocks);
139   bool runOnCountableLoop(Loop *L);
140 
141   AliasAnalysis *AA;
142   const DataLayout *DL;
143   DominatorTree *DT;
144   LoopInfo *LF;
145   const TargetLibraryInfo *TLI;
146   ScalarEvolution *SE;
147   bool HasMemcpy, HasMemmove;
148 };
149 
150 class HexagonLoopIdiomRecognizeLegacyPass : public LoopPass {
151 public:
152   static char ID;
153 
154   explicit HexagonLoopIdiomRecognizeLegacyPass() : LoopPass(ID) {
155     initializeHexagonLoopIdiomRecognizeLegacyPassPass(
156         *PassRegistry::getPassRegistry());
157   }
158 
159   StringRef getPassName() const override {
160     return "Recognize Hexagon-specific loop idioms";
161   }
162 
163   void getAnalysisUsage(AnalysisUsage &AU) const override {
164     AU.addRequired<LoopInfoWrapperPass>();
165     AU.addRequiredID(LoopSimplifyID);
166     AU.addRequiredID(LCSSAID);
167     AU.addRequired<AAResultsWrapperPass>();
168     AU.addRequired<ScalarEvolutionWrapperPass>();
169     AU.addRequired<DominatorTreeWrapperPass>();
170     AU.addRequired<TargetLibraryInfoWrapperPass>();
171     AU.addPreserved<TargetLibraryInfoWrapperPass>();
172   }
173 
174   bool runOnLoop(Loop *L, LPPassManager &LPM) override;
175 };
176 
177 struct Simplifier {
178   struct Rule {
179     using FuncType = std::function<Value *(Instruction *, LLVMContext &)>;
180     Rule(StringRef N, FuncType F) : Name(N), Fn(F) {}
181     StringRef Name; // For debugging.
182     FuncType Fn;
183   };
184 
185   void addRule(StringRef N, const Rule::FuncType &F) {
186     Rules.push_back(Rule(N, F));
187   }
188 
189 private:
190   struct WorkListType {
191     WorkListType() = default;
192 
193     void push_back(Value *V) {
194       // Do not push back duplicates.
195       if (S.insert(V).second)
196         Q.push_back(V);
197     }
198 
199     Value *pop_front_val() {
200       Value *V = Q.front();
201       Q.pop_front();
202       S.erase(V);
203       return V;
204     }
205 
206     bool empty() const { return Q.empty(); }
207 
208   private:
209     std::deque<Value *> Q;
210     std::set<Value *> S;
211   };
212 
213   using ValueSetType = std::set<Value *>;
214 
215   std::vector<Rule> Rules;
216 
217 public:
218   struct Context {
219     using ValueMapType = DenseMap<Value *, Value *>;
220 
221     Value *Root;
222     ValueSetType Used;   // The set of all cloned values used by Root.
223     ValueSetType Clones; // The set of all cloned values.
224     LLVMContext &Ctx;
225 
226     Context(Instruction *Exp)
227         : Ctx(Exp->getParent()->getParent()->getContext()) {
228       initialize(Exp);
229     }
230 
231     ~Context() { cleanup(); }
232 
233     void print(raw_ostream &OS, const Value *V) const;
234     Value *materialize(BasicBlock *B, BasicBlock::iterator At);
235 
236   private:
237     friend struct Simplifier;
238 
239     void initialize(Instruction *Exp);
240     void cleanup();
241 
242     template <typename FuncT> void traverse(Value *V, FuncT F);
243     void record(Value *V);
244     void use(Value *V);
245     void unuse(Value *V);
246 
247     bool equal(const Instruction *I, const Instruction *J) const;
248     Value *find(Value *Tree, Value *Sub) const;
249     Value *subst(Value *Tree, Value *OldV, Value *NewV);
250     void replace(Value *OldV, Value *NewV);
251     void link(Instruction *I, BasicBlock *B, BasicBlock::iterator At);
252   };
253 
254   Value *simplify(Context &C);
255 };
256 
257   struct PE {
258     PE(const Simplifier::Context &c, Value *v = nullptr) : C(c), V(v) {}
259 
260     const Simplifier::Context &C;
261     const Value *V;
262   };
263 
264   LLVM_ATTRIBUTE_USED
265   raw_ostream &operator<<(raw_ostream &OS, const PE &P) {
266     P.C.print(OS, P.V ? P.V : P.C.Root);
267     return OS;
268   }
269 
270 } // end anonymous namespace
271 
272 char HexagonLoopIdiomRecognizeLegacyPass::ID = 0;
273 
274 INITIALIZE_PASS_BEGIN(HexagonLoopIdiomRecognizeLegacyPass, "hexagon-loop-idiom",
275                       "Recognize Hexagon-specific loop idioms", false, false)
276 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
277 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
278 INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
279 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
280 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
281 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
282 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
283 INITIALIZE_PASS_END(HexagonLoopIdiomRecognizeLegacyPass, "hexagon-loop-idiom",
284                     "Recognize Hexagon-specific loop idioms", false, false)
285 
286 template <typename FuncT>
287 void Simplifier::Context::traverse(Value *V, FuncT F) {
288   WorkListType Q;
289   Q.push_back(V);
290 
291   while (!Q.empty()) {
292     Instruction *U = dyn_cast<Instruction>(Q.pop_front_val());
293     if (!U || U->getParent())
294       continue;
295     if (!F(U))
296       continue;
297     for (Value *Op : U->operands())
298       Q.push_back(Op);
299   }
300 }
301 
302 void Simplifier::Context::print(raw_ostream &OS, const Value *V) const {
303   const auto *U = dyn_cast<const Instruction>(V);
304   if (!U) {
305     OS << V << '(' << *V << ')';
306     return;
307   }
308 
309   if (U->getParent()) {
310     OS << U << '(';
311     U->printAsOperand(OS, true);
312     OS << ')';
313     return;
314   }
315 
316   unsigned N = U->getNumOperands();
317   if (N != 0)
318     OS << U << '(';
319   OS << U->getOpcodeName();
320   for (const Value *Op : U->operands()) {
321     OS << ' ';
322     print(OS, Op);
323   }
324   if (N != 0)
325     OS << ')';
326 }
327 
328 void Simplifier::Context::initialize(Instruction *Exp) {
329   // Perform a deep clone of the expression, set Root to the root
330   // of the clone, and build a map from the cloned values to the
331   // original ones.
332   ValueMapType M;
333   BasicBlock *Block = Exp->getParent();
334   WorkListType Q;
335   Q.push_back(Exp);
336 
337   while (!Q.empty()) {
338     Value *V = Q.pop_front_val();
339     if (M.contains(V))
340       continue;
341     if (Instruction *U = dyn_cast<Instruction>(V)) {
342       if (isa<PHINode>(U) || U->getParent() != Block)
343         continue;
344       for (Value *Op : U->operands())
345         Q.push_back(Op);
346       M.insert({U, U->clone()});
347     }
348   }
349 
350   for (std::pair<Value*,Value*> P : M) {
351     Instruction *U = cast<Instruction>(P.second);
352     for (unsigned i = 0, n = U->getNumOperands(); i != n; ++i) {
353       auto F = M.find(U->getOperand(i));
354       if (F != M.end())
355         U->setOperand(i, F->second);
356     }
357   }
358 
359   auto R = M.find(Exp);
360   assert(R != M.end());
361   Root = R->second;
362 
363   record(Root);
364   use(Root);
365 }
366 
367 void Simplifier::Context::record(Value *V) {
368   auto Record = [this](Instruction *U) -> bool {
369     Clones.insert(U);
370     return true;
371   };
372   traverse(V, Record);
373 }
374 
375 void Simplifier::Context::use(Value *V) {
376   auto Use = [this](Instruction *U) -> bool {
377     Used.insert(U);
378     return true;
379   };
380   traverse(V, Use);
381 }
382 
383 void Simplifier::Context::unuse(Value *V) {
384   if (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != nullptr)
385     return;
386 
387   auto Unuse = [this](Instruction *U) -> bool {
388     if (!U->use_empty())
389       return false;
390     Used.erase(U);
391     return true;
392   };
393   traverse(V, Unuse);
394 }
395 
396 Value *Simplifier::Context::subst(Value *Tree, Value *OldV, Value *NewV) {
397   if (Tree == OldV)
398     return NewV;
399   if (OldV == NewV)
400     return Tree;
401 
402   WorkListType Q;
403   Q.push_back(Tree);
404   while (!Q.empty()) {
405     Instruction *U = dyn_cast<Instruction>(Q.pop_front_val());
406     // If U is not an instruction, or it's not a clone, skip it.
407     if (!U || U->getParent())
408       continue;
409     for (unsigned i = 0, n = U->getNumOperands(); i != n; ++i) {
410       Value *Op = U->getOperand(i);
411       if (Op == OldV) {
412         U->setOperand(i, NewV);
413         unuse(OldV);
414       } else {
415         Q.push_back(Op);
416       }
417     }
418   }
419   return Tree;
420 }
421 
422 void Simplifier::Context::replace(Value *OldV, Value *NewV) {
423   if (Root == OldV) {
424     Root = NewV;
425     use(Root);
426     return;
427   }
428 
429   // NewV may be a complex tree that has just been created by one of the
430   // transformation rules. We need to make sure that it is commoned with
431   // the existing Root to the maximum extent possible.
432   // Identify all subtrees of NewV (including NewV itself) that have
433   // equivalent counterparts in Root, and replace those subtrees with
434   // these counterparts.
435   WorkListType Q;
436   Q.push_back(NewV);
437   while (!Q.empty()) {
438     Value *V = Q.pop_front_val();
439     Instruction *U = dyn_cast<Instruction>(V);
440     if (!U || U->getParent())
441       continue;
442     if (Value *DupV = find(Root, V)) {
443       if (DupV != V)
444         NewV = subst(NewV, V, DupV);
445     } else {
446       for (Value *Op : U->operands())
447         Q.push_back(Op);
448     }
449   }
450 
451   // Now, simply replace OldV with NewV in Root.
452   Root = subst(Root, OldV, NewV);
453   use(Root);
454 }
455 
456 void Simplifier::Context::cleanup() {
457   for (Value *V : Clones) {
458     Instruction *U = cast<Instruction>(V);
459     if (!U->getParent())
460       U->dropAllReferences();
461   }
462 
463   for (Value *V : Clones) {
464     Instruction *U = cast<Instruction>(V);
465     if (!U->getParent())
466       U->deleteValue();
467   }
468 }
469 
470 bool Simplifier::Context::equal(const Instruction *I,
471                                 const Instruction *J) const {
472   if (I == J)
473     return true;
474   if (!I->isSameOperationAs(J))
475     return false;
476   if (isa<PHINode>(I))
477     return I->isIdenticalTo(J);
478 
479   for (unsigned i = 0, n = I->getNumOperands(); i != n; ++i) {
480     Value *OpI = I->getOperand(i), *OpJ = J->getOperand(i);
481     if (OpI == OpJ)
482       continue;
483     auto *InI = dyn_cast<const Instruction>(OpI);
484     auto *InJ = dyn_cast<const Instruction>(OpJ);
485     if (InI && InJ) {
486       if (!equal(InI, InJ))
487         return false;
488     } else if (InI != InJ || !InI)
489       return false;
490   }
491   return true;
492 }
493 
494 Value *Simplifier::Context::find(Value *Tree, Value *Sub) const {
495   Instruction *SubI = dyn_cast<Instruction>(Sub);
496   WorkListType Q;
497   Q.push_back(Tree);
498 
499   while (!Q.empty()) {
500     Value *V = Q.pop_front_val();
501     if (V == Sub)
502       return V;
503     Instruction *U = dyn_cast<Instruction>(V);
504     if (!U || U->getParent())
505       continue;
506     if (SubI && equal(SubI, U))
507       return U;
508     assert(!isa<PHINode>(U));
509     for (Value *Op : U->operands())
510       Q.push_back(Op);
511   }
512   return nullptr;
513 }
514 
515 void Simplifier::Context::link(Instruction *I, BasicBlock *B,
516       BasicBlock::iterator At) {
517   if (I->getParent())
518     return;
519 
520   for (Value *Op : I->operands()) {
521     if (Instruction *OpI = dyn_cast<Instruction>(Op))
522       link(OpI, B, At);
523   }
524 
525   I->insertInto(B, At);
526 }
527 
528 Value *Simplifier::Context::materialize(BasicBlock *B,
529       BasicBlock::iterator At) {
530   if (Instruction *RootI = dyn_cast<Instruction>(Root))
531     link(RootI, B, At);
532   return Root;
533 }
534 
535 Value *Simplifier::simplify(Context &C) {
536   WorkListType Q;
537   Q.push_back(C.Root);
538   unsigned Count = 0;
539   const unsigned Limit = SimplifyLimit;
540 
541   while (!Q.empty()) {
542     if (Count++ >= Limit)
543       break;
544     Instruction *U = dyn_cast<Instruction>(Q.pop_front_val());
545     if (!U || U->getParent() || !C.Used.count(U))
546       continue;
547     bool Changed = false;
548     for (Rule &R : Rules) {
549       Value *W = R.Fn(U, C.Ctx);
550       if (!W)
551         continue;
552       Changed = true;
553       C.record(W);
554       C.replace(U, W);
555       Q.push_back(C.Root);
556       break;
557     }
558     if (!Changed) {
559       for (Value *Op : U->operands())
560         Q.push_back(Op);
561     }
562   }
563   return Count < Limit ? C.Root : nullptr;
564 }
565 
566 //===----------------------------------------------------------------------===//
567 //
568 //          Implementation of PolynomialMultiplyRecognize
569 //
570 //===----------------------------------------------------------------------===//
571 
572 namespace {
573 
574   class PolynomialMultiplyRecognize {
575   public:
576     explicit PolynomialMultiplyRecognize(Loop *loop, const DataLayout &dl,
577         const DominatorTree &dt, const TargetLibraryInfo &tli,
578         ScalarEvolution &se)
579       : CurLoop(loop), DL(dl), DT(dt), TLI(tli), SE(se) {}
580 
581     bool recognize();
582 
583   private:
584     using ValueSeq = SetVector<Value *>;
585 
586     IntegerType *getPmpyType() const {
587       LLVMContext &Ctx = CurLoop->getHeader()->getParent()->getContext();
588       return IntegerType::get(Ctx, 32);
589     }
590 
591     bool isPromotableTo(Value *V, IntegerType *Ty);
592     void promoteTo(Instruction *In, IntegerType *DestTy, BasicBlock *LoopB);
593     bool promoteTypes(BasicBlock *LoopB, BasicBlock *ExitB);
594 
595     Value *getCountIV(BasicBlock *BB);
596     bool findCycle(Value *Out, Value *In, ValueSeq &Cycle);
597     void classifyCycle(Instruction *DivI, ValueSeq &Cycle, ValueSeq &Early,
598           ValueSeq &Late);
599     bool classifyInst(Instruction *UseI, ValueSeq &Early, ValueSeq &Late);
600     bool commutesWithShift(Instruction *I);
601     bool highBitsAreZero(Value *V, unsigned IterCount);
602     bool keepsHighBitsZero(Value *V, unsigned IterCount);
603     bool isOperandShifted(Instruction *I, Value *Op);
604     bool convertShiftsToLeft(BasicBlock *LoopB, BasicBlock *ExitB,
605           unsigned IterCount);
606     void cleanupLoopBody(BasicBlock *LoopB);
607 
608     struct ParsedValues {
609       ParsedValues() = default;
610 
611       Value *M = nullptr;
612       Value *P = nullptr;
613       Value *Q = nullptr;
614       Value *R = nullptr;
615       Value *X = nullptr;
616       Instruction *Res = nullptr;
617       unsigned IterCount = 0;
618       bool Left = false;
619       bool Inv = false;
620     };
621 
622     bool matchLeftShift(SelectInst *SelI, Value *CIV, ParsedValues &PV);
623     bool matchRightShift(SelectInst *SelI, ParsedValues &PV);
624     bool scanSelect(SelectInst *SI, BasicBlock *LoopB, BasicBlock *PrehB,
625           Value *CIV, ParsedValues &PV, bool PreScan);
626     unsigned getInverseMxN(unsigned QP);
627     Value *generate(BasicBlock::iterator At, ParsedValues &PV);
628 
629     void setupPreSimplifier(Simplifier &S);
630     void setupPostSimplifier(Simplifier &S);
631 
632     Loop *CurLoop;
633     const DataLayout &DL;
634     const DominatorTree &DT;
635     const TargetLibraryInfo &TLI;
636     ScalarEvolution &SE;
637   };
638 
639 } // end anonymous namespace
640 
641 Value *PolynomialMultiplyRecognize::getCountIV(BasicBlock *BB) {
642   pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
643   if (std::distance(PI, PE) != 2)
644     return nullptr;
645   BasicBlock *PB = (*PI == BB) ? *std::next(PI) : *PI;
646 
647   for (auto I = BB->begin(), E = BB->end(); I != E && isa<PHINode>(I); ++I) {
648     auto *PN = cast<PHINode>(I);
649     Value *InitV = PN->getIncomingValueForBlock(PB);
650     if (!isa<ConstantInt>(InitV) || !cast<ConstantInt>(InitV)->isZero())
651       continue;
652     Value *IterV = PN->getIncomingValueForBlock(BB);
653     auto *BO = dyn_cast<BinaryOperator>(IterV);
654     if (!BO)
655       continue;
656     if (BO->getOpcode() != Instruction::Add)
657       continue;
658     Value *IncV = nullptr;
659     if (BO->getOperand(0) == PN)
660       IncV = BO->getOperand(1);
661     else if (BO->getOperand(1) == PN)
662       IncV = BO->getOperand(0);
663     if (IncV == nullptr)
664       continue;
665 
666     if (auto *T = dyn_cast<ConstantInt>(IncV))
667       if (T->isOne())
668         return PN;
669   }
670   return nullptr;
671 }
672 
673 static void replaceAllUsesOfWithIn(Value *I, Value *J, BasicBlock *BB) {
674   for (auto UI = I->user_begin(), UE = I->user_end(); UI != UE;) {
675     Use &TheUse = UI.getUse();
676     ++UI;
677     if (auto *II = dyn_cast<Instruction>(TheUse.getUser()))
678       if (BB == II->getParent())
679         II->replaceUsesOfWith(I, J);
680   }
681 }
682 
683 bool PolynomialMultiplyRecognize::matchLeftShift(SelectInst *SelI,
684       Value *CIV, ParsedValues &PV) {
685   // Match the following:
686   //   select (X & (1 << i)) != 0 ? R ^ (Q << i) : R
687   //   select (X & (1 << i)) == 0 ? R : R ^ (Q << i)
688   // The condition may also check for equality with the masked value, i.e
689   //   select (X & (1 << i)) == (1 << i) ? R ^ (Q << i) : R
690   //   select (X & (1 << i)) != (1 << i) ? R : R ^ (Q << i);
691 
692   Value *CondV = SelI->getCondition();
693   Value *TrueV = SelI->getTrueValue();
694   Value *FalseV = SelI->getFalseValue();
695 
696   using namespace PatternMatch;
697 
698   CmpInst::Predicate P;
699   Value *A = nullptr, *B = nullptr, *C = nullptr;
700 
701   if (!match(CondV, m_ICmp(P, m_And(m_Value(A), m_Value(B)), m_Value(C))) &&
702       !match(CondV, m_ICmp(P, m_Value(C), m_And(m_Value(A), m_Value(B)))))
703     return false;
704   if (P != CmpInst::ICMP_EQ && P != CmpInst::ICMP_NE)
705     return false;
706   // Matched: select (A & B) == C ? ... : ...
707   //          select (A & B) != C ? ... : ...
708 
709   Value *X = nullptr, *Sh1 = nullptr;
710   // Check (A & B) for (X & (1 << i)):
711   if (match(A, m_Shl(m_One(), m_Specific(CIV)))) {
712     Sh1 = A;
713     X = B;
714   } else if (match(B, m_Shl(m_One(), m_Specific(CIV)))) {
715     Sh1 = B;
716     X = A;
717   } else {
718     // TODO: Could also check for an induction variable containing single
719     // bit shifted left by 1 in each iteration.
720     return false;
721   }
722 
723   bool TrueIfZero;
724 
725   // Check C against the possible values for comparison: 0 and (1 << i):
726   if (match(C, m_Zero()))
727     TrueIfZero = (P == CmpInst::ICMP_EQ);
728   else if (C == Sh1)
729     TrueIfZero = (P == CmpInst::ICMP_NE);
730   else
731     return false;
732 
733   // So far, matched:
734   //   select (X & (1 << i)) ? ... : ...
735   // including variations of the check against zero/non-zero value.
736 
737   Value *ShouldSameV = nullptr, *ShouldXoredV = nullptr;
738   if (TrueIfZero) {
739     ShouldSameV = TrueV;
740     ShouldXoredV = FalseV;
741   } else {
742     ShouldSameV = FalseV;
743     ShouldXoredV = TrueV;
744   }
745 
746   Value *Q = nullptr, *R = nullptr, *Y = nullptr, *Z = nullptr;
747   Value *T = nullptr;
748   if (match(ShouldXoredV, m_Xor(m_Value(Y), m_Value(Z)))) {
749     // Matched: select +++ ? ... : Y ^ Z
750     //          select +++ ? Y ^ Z : ...
751     // where +++ denotes previously checked matches.
752     if (ShouldSameV == Y)
753       T = Z;
754     else if (ShouldSameV == Z)
755       T = Y;
756     else
757       return false;
758     R = ShouldSameV;
759     // Matched: select +++ ? R : R ^ T
760     //          select +++ ? R ^ T : R
761     // depending on TrueIfZero.
762 
763   } else if (match(ShouldSameV, m_Zero())) {
764     // Matched: select +++ ? 0 : ...
765     //          select +++ ? ... : 0
766     if (!SelI->hasOneUse())
767       return false;
768     T = ShouldXoredV;
769     // Matched: select +++ ? 0 : T
770     //          select +++ ? T : 0
771 
772     Value *U = *SelI->user_begin();
773     if (!match(U, m_Xor(m_Specific(SelI), m_Value(R))) &&
774         !match(U, m_Xor(m_Value(R), m_Specific(SelI))))
775       return false;
776     // Matched: xor (select +++ ? 0 : T), R
777     //          xor (select +++ ? T : 0), R
778   } else
779     return false;
780 
781   // The xor input value T is isolated into its own match so that it could
782   // be checked against an induction variable containing a shifted bit
783   // (todo).
784   // For now, check against (Q << i).
785   if (!match(T, m_Shl(m_Value(Q), m_Specific(CIV))) &&
786       !match(T, m_Shl(m_ZExt(m_Value(Q)), m_ZExt(m_Specific(CIV)))))
787     return false;
788   // Matched: select +++ ? R : R ^ (Q << i)
789   //          select +++ ? R ^ (Q << i) : R
790 
791   PV.X = X;
792   PV.Q = Q;
793   PV.R = R;
794   PV.Left = true;
795   return true;
796 }
797 
798 bool PolynomialMultiplyRecognize::matchRightShift(SelectInst *SelI,
799       ParsedValues &PV) {
800   // Match the following:
801   //   select (X & 1) != 0 ? (R >> 1) ^ Q : (R >> 1)
802   //   select (X & 1) == 0 ? (R >> 1) : (R >> 1) ^ Q
803   // The condition may also check for equality with the masked value, i.e
804   //   select (X & 1) == 1 ? (R >> 1) ^ Q : (R >> 1)
805   //   select (X & 1) != 1 ? (R >> 1) : (R >> 1) ^ Q
806 
807   Value *CondV = SelI->getCondition();
808   Value *TrueV = SelI->getTrueValue();
809   Value *FalseV = SelI->getFalseValue();
810 
811   using namespace PatternMatch;
812 
813   Value *C = nullptr;
814   CmpInst::Predicate P;
815   bool TrueIfZero;
816 
817   if (match(CondV, m_ICmp(P, m_Value(C), m_Zero())) ||
818       match(CondV, m_ICmp(P, m_Zero(), m_Value(C)))) {
819     if (P != CmpInst::ICMP_EQ && P != CmpInst::ICMP_NE)
820       return false;
821     // Matched: select C == 0 ? ... : ...
822     //          select C != 0 ? ... : ...
823     TrueIfZero = (P == CmpInst::ICMP_EQ);
824   } else if (match(CondV, m_ICmp(P, m_Value(C), m_One())) ||
825              match(CondV, m_ICmp(P, m_One(), m_Value(C)))) {
826     if (P != CmpInst::ICMP_EQ && P != CmpInst::ICMP_NE)
827       return false;
828     // Matched: select C == 1 ? ... : ...
829     //          select C != 1 ? ... : ...
830     TrueIfZero = (P == CmpInst::ICMP_NE);
831   } else
832     return false;
833 
834   Value *X = nullptr;
835   if (!match(C, m_And(m_Value(X), m_One())) &&
836       !match(C, m_And(m_One(), m_Value(X))))
837     return false;
838   // Matched: select (X & 1) == +++ ? ... : ...
839   //          select (X & 1) != +++ ? ... : ...
840 
841   Value *R = nullptr, *Q = nullptr;
842   if (TrueIfZero) {
843     // The select's condition is true if the tested bit is 0.
844     // TrueV must be the shift, FalseV must be the xor.
845     if (!match(TrueV, m_LShr(m_Value(R), m_One())))
846       return false;
847     // Matched: select +++ ? (R >> 1) : ...
848     if (!match(FalseV, m_Xor(m_Specific(TrueV), m_Value(Q))) &&
849         !match(FalseV, m_Xor(m_Value(Q), m_Specific(TrueV))))
850       return false;
851     // Matched: select +++ ? (R >> 1) : (R >> 1) ^ Q
852     // with commuting ^.
853   } else {
854     // The select's condition is true if the tested bit is 1.
855     // TrueV must be the xor, FalseV must be the shift.
856     if (!match(FalseV, m_LShr(m_Value(R), m_One())))
857       return false;
858     // Matched: select +++ ? ... : (R >> 1)
859     if (!match(TrueV, m_Xor(m_Specific(FalseV), m_Value(Q))) &&
860         !match(TrueV, m_Xor(m_Value(Q), m_Specific(FalseV))))
861       return false;
862     // Matched: select +++ ? (R >> 1) ^ Q : (R >> 1)
863     // with commuting ^.
864   }
865 
866   PV.X = X;
867   PV.Q = Q;
868   PV.R = R;
869   PV.Left = false;
870   return true;
871 }
872 
873 bool PolynomialMultiplyRecognize::scanSelect(SelectInst *SelI,
874       BasicBlock *LoopB, BasicBlock *PrehB, Value *CIV, ParsedValues &PV,
875       bool PreScan) {
876   using namespace PatternMatch;
877 
878   // The basic pattern for R = P.Q is:
879   // for i = 0..31
880   //   R = phi (0, R')
881   //   if (P & (1 << i))        ; test-bit(P, i)
882   //     R' = R ^ (Q << i)
883   //
884   // Similarly, the basic pattern for R = (P/Q).Q - P
885   // for i = 0..31
886   //   R = phi(P, R')
887   //   if (R & (1 << i))
888   //     R' = R ^ (Q << i)
889 
890   // There exist idioms, where instead of Q being shifted left, P is shifted
891   // right. This produces a result that is shifted right by 32 bits (the
892   // non-shifted result is 64-bit).
893   //
894   // For R = P.Q, this would be:
895   // for i = 0..31
896   //   R = phi (0, R')
897   //   if ((P >> i) & 1)
898   //     R' = (R >> 1) ^ Q      ; R is cycled through the loop, so it must
899   //   else                     ; be shifted by 1, not i.
900   //     R' = R >> 1
901   //
902   // And for the inverse:
903   // for i = 0..31
904   //   R = phi (P, R')
905   //   if (R & 1)
906   //     R' = (R >> 1) ^ Q
907   //   else
908   //     R' = R >> 1
909 
910   // The left-shifting idioms share the same pattern:
911   //   select (X & (1 << i)) ? R ^ (Q << i) : R
912   // Similarly for right-shifting idioms:
913   //   select (X & 1) ? (R >> 1) ^ Q
914 
915   if (matchLeftShift(SelI, CIV, PV)) {
916     // If this is a pre-scan, getting this far is sufficient.
917     if (PreScan)
918       return true;
919 
920     // Need to make sure that the SelI goes back into R.
921     auto *RPhi = dyn_cast<PHINode>(PV.R);
922     if (!RPhi)
923       return false;
924     if (SelI != RPhi->getIncomingValueForBlock(LoopB))
925       return false;
926     PV.Res = SelI;
927 
928     // If X is loop invariant, it must be the input polynomial, and the
929     // idiom is the basic polynomial multiply.
930     if (CurLoop->isLoopInvariant(PV.X)) {
931       PV.P = PV.X;
932       PV.Inv = false;
933     } else {
934       // X is not loop invariant. If X == R, this is the inverse pmpy.
935       // Otherwise, check for an xor with an invariant value. If the
936       // variable argument to the xor is R, then this is still a valid
937       // inverse pmpy.
938       PV.Inv = true;
939       if (PV.X != PV.R) {
940         Value *Var = nullptr, *Inv = nullptr, *X1 = nullptr, *X2 = nullptr;
941         if (!match(PV.X, m_Xor(m_Value(X1), m_Value(X2))))
942           return false;
943         auto *I1 = dyn_cast<Instruction>(X1);
944         auto *I2 = dyn_cast<Instruction>(X2);
945         if (!I1 || I1->getParent() != LoopB) {
946           Var = X2;
947           Inv = X1;
948         } else if (!I2 || I2->getParent() != LoopB) {
949           Var = X1;
950           Inv = X2;
951         } else
952           return false;
953         if (Var != PV.R)
954           return false;
955         PV.M = Inv;
956       }
957       // The input polynomial P still needs to be determined. It will be
958       // the entry value of R.
959       Value *EntryP = RPhi->getIncomingValueForBlock(PrehB);
960       PV.P = EntryP;
961     }
962 
963     return true;
964   }
965 
966   if (matchRightShift(SelI, PV)) {
967     // If this is an inverse pattern, the Q polynomial must be known at
968     // compile time.
969     if (PV.Inv && !isa<ConstantInt>(PV.Q))
970       return false;
971     if (PreScan)
972       return true;
973     // There is no exact matching of right-shift pmpy.
974     return false;
975   }
976 
977   return false;
978 }
979 
980 bool PolynomialMultiplyRecognize::isPromotableTo(Value *Val,
981       IntegerType *DestTy) {
982   IntegerType *T = dyn_cast<IntegerType>(Val->getType());
983   if (!T || T->getBitWidth() > DestTy->getBitWidth())
984     return false;
985   if (T->getBitWidth() == DestTy->getBitWidth())
986     return true;
987   // Non-instructions are promotable. The reason why an instruction may not
988   // be promotable is that it may produce a different result if its operands
989   // and the result are promoted, for example, it may produce more non-zero
990   // bits. While it would still be possible to represent the proper result
991   // in a wider type, it may require adding additional instructions (which
992   // we don't want to do).
993   Instruction *In = dyn_cast<Instruction>(Val);
994   if (!In)
995     return true;
996   // The bitwidth of the source type is smaller than the destination.
997   // Check if the individual operation can be promoted.
998   switch (In->getOpcode()) {
999     case Instruction::PHI:
1000     case Instruction::ZExt:
1001     case Instruction::And:
1002     case Instruction::Or:
1003     case Instruction::Xor:
1004     case Instruction::LShr: // Shift right is ok.
1005     case Instruction::Select:
1006     case Instruction::Trunc:
1007       return true;
1008     case Instruction::ICmp:
1009       if (CmpInst *CI = cast<CmpInst>(In))
1010         return CI->isEquality() || CI->isUnsigned();
1011       llvm_unreachable("Cast failed unexpectedly");
1012     case Instruction::Add:
1013       return In->hasNoSignedWrap() && In->hasNoUnsignedWrap();
1014   }
1015   return false;
1016 }
1017 
1018 void PolynomialMultiplyRecognize::promoteTo(Instruction *In,
1019       IntegerType *DestTy, BasicBlock *LoopB) {
1020   Type *OrigTy = In->getType();
1021   assert(!OrigTy->isVoidTy() && "Invalid instruction to promote");
1022 
1023   // Leave boolean values alone.
1024   if (!In->getType()->isIntegerTy(1))
1025     In->mutateType(DestTy);
1026   unsigned DestBW = DestTy->getBitWidth();
1027 
1028   // Handle PHIs.
1029   if (PHINode *P = dyn_cast<PHINode>(In)) {
1030     unsigned N = P->getNumIncomingValues();
1031     for (unsigned i = 0; i != N; ++i) {
1032       BasicBlock *InB = P->getIncomingBlock(i);
1033       if (InB == LoopB)
1034         continue;
1035       Value *InV = P->getIncomingValue(i);
1036       IntegerType *Ty = cast<IntegerType>(InV->getType());
1037       // Do not promote values in PHI nodes of type i1.
1038       if (Ty != P->getType()) {
1039         // If the value type does not match the PHI type, the PHI type
1040         // must have been promoted.
1041         assert(Ty->getBitWidth() < DestBW);
1042         InV = IRBuilder<>(InB->getTerminator()).CreateZExt(InV, DestTy);
1043         P->setIncomingValue(i, InV);
1044       }
1045     }
1046   } else if (ZExtInst *Z = dyn_cast<ZExtInst>(In)) {
1047     Value *Op = Z->getOperand(0);
1048     if (Op->getType() == Z->getType())
1049       Z->replaceAllUsesWith(Op);
1050     Z->eraseFromParent();
1051     return;
1052   }
1053   if (TruncInst *T = dyn_cast<TruncInst>(In)) {
1054     IntegerType *TruncTy = cast<IntegerType>(OrigTy);
1055     Value *Mask = ConstantInt::get(DestTy, (1u << TruncTy->getBitWidth()) - 1);
1056     Value *And = IRBuilder<>(In).CreateAnd(T->getOperand(0), Mask);
1057     T->replaceAllUsesWith(And);
1058     T->eraseFromParent();
1059     return;
1060   }
1061 
1062   // Promote immediates.
1063   for (unsigned i = 0, n = In->getNumOperands(); i != n; ++i) {
1064     if (ConstantInt *CI = dyn_cast<ConstantInt>(In->getOperand(i)))
1065       if (CI->getType()->getBitWidth() < DestBW)
1066         In->setOperand(i, ConstantInt::get(DestTy, CI->getZExtValue()));
1067   }
1068 }
1069 
1070 bool PolynomialMultiplyRecognize::promoteTypes(BasicBlock *LoopB,
1071       BasicBlock *ExitB) {
1072   assert(LoopB);
1073   // Skip loops where the exit block has more than one predecessor. The values
1074   // coming from the loop block will be promoted to another type, and so the
1075   // values coming into the exit block from other predecessors would also have
1076   // to be promoted.
1077   if (!ExitB || (ExitB->getSinglePredecessor() != LoopB))
1078     return false;
1079   IntegerType *DestTy = getPmpyType();
1080   // Check if the exit values have types that are no wider than the type
1081   // that we want to promote to.
1082   unsigned DestBW = DestTy->getBitWidth();
1083   for (PHINode &P : ExitB->phis()) {
1084     if (P.getNumIncomingValues() != 1)
1085       return false;
1086     assert(P.getIncomingBlock(0) == LoopB);
1087     IntegerType *T = dyn_cast<IntegerType>(P.getType());
1088     if (!T || T->getBitWidth() > DestBW)
1089       return false;
1090   }
1091 
1092   // Check all instructions in the loop.
1093   for (Instruction &In : *LoopB)
1094     if (!In.isTerminator() && !isPromotableTo(&In, DestTy))
1095       return false;
1096 
1097   // Perform the promotion.
1098   std::vector<Instruction*> LoopIns;
1099   std::transform(LoopB->begin(), LoopB->end(), std::back_inserter(LoopIns),
1100                  [](Instruction &In) { return &In; });
1101   for (Instruction *In : LoopIns)
1102     if (!In->isTerminator())
1103       promoteTo(In, DestTy, LoopB);
1104 
1105   // Fix up the PHI nodes in the exit block.
1106   Instruction *EndI = ExitB->getFirstNonPHI();
1107   BasicBlock::iterator End = EndI ? EndI->getIterator() : ExitB->end();
1108   for (auto I = ExitB->begin(); I != End; ++I) {
1109     PHINode *P = dyn_cast<PHINode>(I);
1110     if (!P)
1111       break;
1112     Type *Ty0 = P->getIncomingValue(0)->getType();
1113     Type *PTy = P->getType();
1114     if (PTy != Ty0) {
1115       assert(Ty0 == DestTy);
1116       // In order to create the trunc, P must have the promoted type.
1117       P->mutateType(Ty0);
1118       Value *T = IRBuilder<>(ExitB, End).CreateTrunc(P, PTy);
1119       // In order for the RAUW to work, the types of P and T must match.
1120       P->mutateType(PTy);
1121       P->replaceAllUsesWith(T);
1122       // Final update of the P's type.
1123       P->mutateType(Ty0);
1124       cast<Instruction>(T)->setOperand(0, P);
1125     }
1126   }
1127 
1128   return true;
1129 }
1130 
1131 bool PolynomialMultiplyRecognize::findCycle(Value *Out, Value *In,
1132       ValueSeq &Cycle) {
1133   // Out = ..., In, ...
1134   if (Out == In)
1135     return true;
1136 
1137   auto *BB = cast<Instruction>(Out)->getParent();
1138   bool HadPhi = false;
1139 
1140   for (auto *U : Out->users()) {
1141     auto *I = dyn_cast<Instruction>(&*U);
1142     if (I == nullptr || I->getParent() != BB)
1143       continue;
1144     // Make sure that there are no multi-iteration cycles, e.g.
1145     //   p1 = phi(p2)
1146     //   p2 = phi(p1)
1147     // The cycle p1->p2->p1 would span two loop iterations.
1148     // Check that there is only one phi in the cycle.
1149     bool IsPhi = isa<PHINode>(I);
1150     if (IsPhi && HadPhi)
1151       return false;
1152     HadPhi |= IsPhi;
1153     if (!Cycle.insert(I))
1154       return false;
1155     if (findCycle(I, In, Cycle))
1156       break;
1157     Cycle.remove(I);
1158   }
1159   return !Cycle.empty();
1160 }
1161 
1162 void PolynomialMultiplyRecognize::classifyCycle(Instruction *DivI,
1163       ValueSeq &Cycle, ValueSeq &Early, ValueSeq &Late) {
1164   // All the values in the cycle that are between the phi node and the
1165   // divider instruction will be classified as "early", all other values
1166   // will be "late".
1167 
1168   bool IsE = true;
1169   unsigned I, N = Cycle.size();
1170   for (I = 0; I < N; ++I) {
1171     Value *V = Cycle[I];
1172     if (DivI == V)
1173       IsE = false;
1174     else if (!isa<PHINode>(V))
1175       continue;
1176     // Stop if found either.
1177     break;
1178   }
1179   // "I" is the index of either DivI or the phi node, whichever was first.
1180   // "E" is "false" or "true" respectively.
1181   ValueSeq &First = !IsE ? Early : Late;
1182   for (unsigned J = 0; J < I; ++J)
1183     First.insert(Cycle[J]);
1184 
1185   ValueSeq &Second = IsE ? Early : Late;
1186   Second.insert(Cycle[I]);
1187   for (++I; I < N; ++I) {
1188     Value *V = Cycle[I];
1189     if (DivI == V || isa<PHINode>(V))
1190       break;
1191     Second.insert(V);
1192   }
1193 
1194   for (; I < N; ++I)
1195     First.insert(Cycle[I]);
1196 }
1197 
1198 bool PolynomialMultiplyRecognize::classifyInst(Instruction *UseI,
1199       ValueSeq &Early, ValueSeq &Late) {
1200   // Select is an exception, since the condition value does not have to be
1201   // classified in the same way as the true/false values. The true/false
1202   // values do have to be both early or both late.
1203   if (UseI->getOpcode() == Instruction::Select) {
1204     Value *TV = UseI->getOperand(1), *FV = UseI->getOperand(2);
1205     if (Early.count(TV) || Early.count(FV)) {
1206       if (Late.count(TV) || Late.count(FV))
1207         return false;
1208       Early.insert(UseI);
1209     } else if (Late.count(TV) || Late.count(FV)) {
1210       if (Early.count(TV) || Early.count(FV))
1211         return false;
1212       Late.insert(UseI);
1213     }
1214     return true;
1215   }
1216 
1217   // Not sure what would be the example of this, but the code below relies
1218   // on having at least one operand.
1219   if (UseI->getNumOperands() == 0)
1220     return true;
1221 
1222   bool AE = true, AL = true;
1223   for (auto &I : UseI->operands()) {
1224     if (Early.count(&*I))
1225       AL = false;
1226     else if (Late.count(&*I))
1227       AE = false;
1228   }
1229   // If the operands appear "all early" and "all late" at the same time,
1230   // then it means that none of them are actually classified as either.
1231   // This is harmless.
1232   if (AE && AL)
1233     return true;
1234   // Conversely, if they are neither "all early" nor "all late", then
1235   // we have a mixture of early and late operands that is not a known
1236   // exception.
1237   if (!AE && !AL)
1238     return false;
1239 
1240   // Check that we have covered the two special cases.
1241   assert(AE != AL);
1242 
1243   if (AE)
1244     Early.insert(UseI);
1245   else
1246     Late.insert(UseI);
1247   return true;
1248 }
1249 
1250 bool PolynomialMultiplyRecognize::commutesWithShift(Instruction *I) {
1251   switch (I->getOpcode()) {
1252     case Instruction::And:
1253     case Instruction::Or:
1254     case Instruction::Xor:
1255     case Instruction::LShr:
1256     case Instruction::Shl:
1257     case Instruction::Select:
1258     case Instruction::ICmp:
1259     case Instruction::PHI:
1260       break;
1261     default:
1262       return false;
1263   }
1264   return true;
1265 }
1266 
1267 bool PolynomialMultiplyRecognize::highBitsAreZero(Value *V,
1268       unsigned IterCount) {
1269   auto *T = dyn_cast<IntegerType>(V->getType());
1270   if (!T)
1271     return false;
1272 
1273   KnownBits Known(T->getBitWidth());
1274   computeKnownBits(V, Known, DL);
1275   return Known.countMinLeadingZeros() >= IterCount;
1276 }
1277 
1278 bool PolynomialMultiplyRecognize::keepsHighBitsZero(Value *V,
1279       unsigned IterCount) {
1280   // Assume that all inputs to the value have the high bits zero.
1281   // Check if the value itself preserves the zeros in the high bits.
1282   if (auto *C = dyn_cast<ConstantInt>(V))
1283     return C->getValue().countl_zero() >= IterCount;
1284 
1285   if (auto *I = dyn_cast<Instruction>(V)) {
1286     switch (I->getOpcode()) {
1287       case Instruction::And:
1288       case Instruction::Or:
1289       case Instruction::Xor:
1290       case Instruction::LShr:
1291       case Instruction::Select:
1292       case Instruction::ICmp:
1293       case Instruction::PHI:
1294       case Instruction::ZExt:
1295         return true;
1296     }
1297   }
1298 
1299   return false;
1300 }
1301 
1302 bool PolynomialMultiplyRecognize::isOperandShifted(Instruction *I, Value *Op) {
1303   unsigned Opc = I->getOpcode();
1304   if (Opc == Instruction::Shl || Opc == Instruction::LShr)
1305     return Op != I->getOperand(1);
1306   return true;
1307 }
1308 
1309 bool PolynomialMultiplyRecognize::convertShiftsToLeft(BasicBlock *LoopB,
1310       BasicBlock *ExitB, unsigned IterCount) {
1311   Value *CIV = getCountIV(LoopB);
1312   if (CIV == nullptr)
1313     return false;
1314   auto *CIVTy = dyn_cast<IntegerType>(CIV->getType());
1315   if (CIVTy == nullptr)
1316     return false;
1317 
1318   ValueSeq RShifts;
1319   ValueSeq Early, Late, Cycled;
1320 
1321   // Find all value cycles that contain logical right shifts by 1.
1322   for (Instruction &I : *LoopB) {
1323     using namespace PatternMatch;
1324 
1325     Value *V = nullptr;
1326     if (!match(&I, m_LShr(m_Value(V), m_One())))
1327       continue;
1328     ValueSeq C;
1329     if (!findCycle(&I, V, C))
1330       continue;
1331 
1332     // Found a cycle.
1333     C.insert(&I);
1334     classifyCycle(&I, C, Early, Late);
1335     Cycled.insert(C.begin(), C.end());
1336     RShifts.insert(&I);
1337   }
1338 
1339   // Find the set of all values affected by the shift cycles, i.e. all
1340   // cycled values, and (recursively) all their users.
1341   ValueSeq Users(Cycled.begin(), Cycled.end());
1342   for (unsigned i = 0; i < Users.size(); ++i) {
1343     Value *V = Users[i];
1344     if (!isa<IntegerType>(V->getType()))
1345       return false;
1346     auto *R = cast<Instruction>(V);
1347     // If the instruction does not commute with shifts, the loop cannot
1348     // be unshifted.
1349     if (!commutesWithShift(R))
1350       return false;
1351     for (User *U : R->users()) {
1352       auto *T = cast<Instruction>(U);
1353       // Skip users from outside of the loop. They will be handled later.
1354       // Also, skip the right-shifts and phi nodes, since they mix early
1355       // and late values.
1356       if (T->getParent() != LoopB || RShifts.count(T) || isa<PHINode>(T))
1357         continue;
1358 
1359       Users.insert(T);
1360       if (!classifyInst(T, Early, Late))
1361         return false;
1362     }
1363   }
1364 
1365   if (Users.empty())
1366     return false;
1367 
1368   // Verify that high bits remain zero.
1369   ValueSeq Internal(Users.begin(), Users.end());
1370   ValueSeq Inputs;
1371   for (unsigned i = 0; i < Internal.size(); ++i) {
1372     auto *R = dyn_cast<Instruction>(Internal[i]);
1373     if (!R)
1374       continue;
1375     for (Value *Op : R->operands()) {
1376       auto *T = dyn_cast<Instruction>(Op);
1377       if (T && T->getParent() != LoopB)
1378         Inputs.insert(Op);
1379       else
1380         Internal.insert(Op);
1381     }
1382   }
1383   for (Value *V : Inputs)
1384     if (!highBitsAreZero(V, IterCount))
1385       return false;
1386   for (Value *V : Internal)
1387     if (!keepsHighBitsZero(V, IterCount))
1388       return false;
1389 
1390   // Finally, the work can be done. Unshift each user.
1391   IRBuilder<> IRB(LoopB);
1392   std::map<Value*,Value*> ShiftMap;
1393 
1394   using CastMapType = std::map<std::pair<Value *, Type *>, Value *>;
1395 
1396   CastMapType CastMap;
1397 
1398   auto upcast = [] (CastMapType &CM, IRBuilder<> &IRB, Value *V,
1399         IntegerType *Ty) -> Value* {
1400     auto H = CM.find(std::make_pair(V, Ty));
1401     if (H != CM.end())
1402       return H->second;
1403     Value *CV = IRB.CreateIntCast(V, Ty, false);
1404     CM.insert(std::make_pair(std::make_pair(V, Ty), CV));
1405     return CV;
1406   };
1407 
1408   for (auto I = LoopB->begin(), E = LoopB->end(); I != E; ++I) {
1409     using namespace PatternMatch;
1410 
1411     if (isa<PHINode>(I) || !Users.count(&*I))
1412       continue;
1413 
1414     // Match lshr x, 1.
1415     Value *V = nullptr;
1416     if (match(&*I, m_LShr(m_Value(V), m_One()))) {
1417       replaceAllUsesOfWithIn(&*I, V, LoopB);
1418       continue;
1419     }
1420     // For each non-cycled operand, replace it with the corresponding
1421     // value shifted left.
1422     for (auto &J : I->operands()) {
1423       Value *Op = J.get();
1424       if (!isOperandShifted(&*I, Op))
1425         continue;
1426       if (Users.count(Op))
1427         continue;
1428       // Skip shifting zeros.
1429       if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
1430         continue;
1431       // Check if we have already generated a shift for this value.
1432       auto F = ShiftMap.find(Op);
1433       Value *W = (F != ShiftMap.end()) ? F->second : nullptr;
1434       if (W == nullptr) {
1435         IRB.SetInsertPoint(&*I);
1436         // First, the shift amount will be CIV or CIV+1, depending on
1437         // whether the value is early or late. Instead of creating CIV+1,
1438         // do a single shift of the value.
1439         Value *ShAmt = CIV, *ShVal = Op;
1440         auto *VTy = cast<IntegerType>(ShVal->getType());
1441         auto *ATy = cast<IntegerType>(ShAmt->getType());
1442         if (Late.count(&*I))
1443           ShVal = IRB.CreateShl(Op, ConstantInt::get(VTy, 1));
1444         // Second, the types of the shifted value and the shift amount
1445         // must match.
1446         if (VTy != ATy) {
1447           if (VTy->getBitWidth() < ATy->getBitWidth())
1448             ShVal = upcast(CastMap, IRB, ShVal, ATy);
1449           else
1450             ShAmt = upcast(CastMap, IRB, ShAmt, VTy);
1451         }
1452         // Ready to generate the shift and memoize it.
1453         W = IRB.CreateShl(ShVal, ShAmt);
1454         ShiftMap.insert(std::make_pair(Op, W));
1455       }
1456       I->replaceUsesOfWith(Op, W);
1457     }
1458   }
1459 
1460   // Update the users outside of the loop to account for having left
1461   // shifts. They would normally be shifted right in the loop, so shift
1462   // them right after the loop exit.
1463   // Take advantage of the loop-closed SSA form, which has all the post-
1464   // loop values in phi nodes.
1465   IRB.SetInsertPoint(ExitB, ExitB->getFirstInsertionPt());
1466   for (auto P = ExitB->begin(), Q = ExitB->end(); P != Q; ++P) {
1467     if (!isa<PHINode>(P))
1468       break;
1469     auto *PN = cast<PHINode>(P);
1470     Value *U = PN->getIncomingValueForBlock(LoopB);
1471     if (!Users.count(U))
1472       continue;
1473     Value *S = IRB.CreateLShr(PN, ConstantInt::get(PN->getType(), IterCount));
1474     PN->replaceAllUsesWith(S);
1475     // The above RAUW will create
1476     //   S = lshr S, IterCount
1477     // so we need to fix it back into
1478     //   S = lshr PN, IterCount
1479     cast<User>(S)->replaceUsesOfWith(S, PN);
1480   }
1481 
1482   return true;
1483 }
1484 
1485 void PolynomialMultiplyRecognize::cleanupLoopBody(BasicBlock *LoopB) {
1486   for (auto &I : *LoopB)
1487     if (Value *SV = simplifyInstruction(&I, {DL, &TLI, &DT}))
1488       I.replaceAllUsesWith(SV);
1489 
1490   for (Instruction &I : llvm::make_early_inc_range(*LoopB))
1491     RecursivelyDeleteTriviallyDeadInstructions(&I, &TLI);
1492 }
1493 
1494 unsigned PolynomialMultiplyRecognize::getInverseMxN(unsigned QP) {
1495   // Arrays of coefficients of Q and the inverse, C.
1496   // Q[i] = coefficient at x^i.
1497   std::array<char,32> Q, C;
1498 
1499   for (unsigned i = 0; i < 32; ++i) {
1500     Q[i] = QP & 1;
1501     QP >>= 1;
1502   }
1503   assert(Q[0] == 1);
1504 
1505   // Find C, such that
1506   // (Q[n]*x^n + ... + Q[1]*x + Q[0]) * (C[n]*x^n + ... + C[1]*x + C[0]) = 1
1507   //
1508   // For it to have a solution, Q[0] must be 1. Since this is Z2[x], the
1509   // operations * and + are & and ^ respectively.
1510   //
1511   // Find C[i] recursively, by comparing i-th coefficient in the product
1512   // with 0 (or 1 for i=0).
1513   //
1514   // C[0] = 1, since C[0] = Q[0], and Q[0] = 1.
1515   C[0] = 1;
1516   for (unsigned i = 1; i < 32; ++i) {
1517     // Solve for C[i] in:
1518     //   C[0]Q[i] ^ C[1]Q[i-1] ^ ... ^ C[i-1]Q[1] ^ C[i]Q[0] = 0
1519     // This is equivalent to
1520     //   C[0]Q[i] ^ C[1]Q[i-1] ^ ... ^ C[i-1]Q[1] ^ C[i] = 0
1521     // which is
1522     //   C[0]Q[i] ^ C[1]Q[i-1] ^ ... ^ C[i-1]Q[1] = C[i]
1523     unsigned T = 0;
1524     for (unsigned j = 0; j < i; ++j)
1525       T = T ^ (C[j] & Q[i-j]);
1526     C[i] = T;
1527   }
1528 
1529   unsigned QV = 0;
1530   for (unsigned i = 0; i < 32; ++i)
1531     if (C[i])
1532       QV |= (1 << i);
1533 
1534   return QV;
1535 }
1536 
1537 Value *PolynomialMultiplyRecognize::generate(BasicBlock::iterator At,
1538       ParsedValues &PV) {
1539   IRBuilder<> B(&*At);
1540   Module *M = At->getParent()->getParent()->getParent();
1541   Function *PMF = Intrinsic::getDeclaration(M, Intrinsic::hexagon_M4_pmpyw);
1542 
1543   Value *P = PV.P, *Q = PV.Q, *P0 = P;
1544   unsigned IC = PV.IterCount;
1545 
1546   if (PV.M != nullptr)
1547     P0 = P = B.CreateXor(P, PV.M);
1548 
1549   // Create a bit mask to clear the high bits beyond IterCount.
1550   auto *BMI = ConstantInt::get(P->getType(), APInt::getLowBitsSet(32, IC));
1551 
1552   if (PV.IterCount != 32)
1553     P = B.CreateAnd(P, BMI);
1554 
1555   if (PV.Inv) {
1556     auto *QI = dyn_cast<ConstantInt>(PV.Q);
1557     assert(QI && QI->getBitWidth() <= 32);
1558 
1559     // Again, clearing bits beyond IterCount.
1560     unsigned M = (1 << PV.IterCount) - 1;
1561     unsigned Tmp = (QI->getZExtValue() | 1) & M;
1562     unsigned QV = getInverseMxN(Tmp) & M;
1563     auto *QVI = ConstantInt::get(QI->getType(), QV);
1564     P = B.CreateCall(PMF, {P, QVI});
1565     P = B.CreateTrunc(P, QI->getType());
1566     if (IC != 32)
1567       P = B.CreateAnd(P, BMI);
1568   }
1569 
1570   Value *R = B.CreateCall(PMF, {P, Q});
1571 
1572   if (PV.M != nullptr)
1573     R = B.CreateXor(R, B.CreateIntCast(P0, R->getType(), false));
1574 
1575   return R;
1576 }
1577 
1578 static bool hasZeroSignBit(const Value *V) {
1579   if (const auto *CI = dyn_cast<const ConstantInt>(V))
1580     return (CI->getType()->getSignBit() & CI->getSExtValue()) == 0;
1581   const Instruction *I = dyn_cast<const Instruction>(V);
1582   if (!I)
1583     return false;
1584   switch (I->getOpcode()) {
1585     case Instruction::LShr:
1586       if (const auto SI = dyn_cast<const ConstantInt>(I->getOperand(1)))
1587         return SI->getZExtValue() > 0;
1588       return false;
1589     case Instruction::Or:
1590     case Instruction::Xor:
1591       return hasZeroSignBit(I->getOperand(0)) &&
1592              hasZeroSignBit(I->getOperand(1));
1593     case Instruction::And:
1594       return hasZeroSignBit(I->getOperand(0)) ||
1595              hasZeroSignBit(I->getOperand(1));
1596   }
1597   return false;
1598 }
1599 
1600 void PolynomialMultiplyRecognize::setupPreSimplifier(Simplifier &S) {
1601   S.addRule("sink-zext",
1602     // Sink zext past bitwise operations.
1603     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1604       if (I->getOpcode() != Instruction::ZExt)
1605         return nullptr;
1606       Instruction *T = dyn_cast<Instruction>(I->getOperand(0));
1607       if (!T)
1608         return nullptr;
1609       switch (T->getOpcode()) {
1610         case Instruction::And:
1611         case Instruction::Or:
1612         case Instruction::Xor:
1613           break;
1614         default:
1615           return nullptr;
1616       }
1617       IRBuilder<> B(Ctx);
1618       return B.CreateBinOp(cast<BinaryOperator>(T)->getOpcode(),
1619                            B.CreateZExt(T->getOperand(0), I->getType()),
1620                            B.CreateZExt(T->getOperand(1), I->getType()));
1621     });
1622   S.addRule("xor/and -> and/xor",
1623     // (xor (and x a) (and y a)) -> (and (xor x y) a)
1624     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1625       if (I->getOpcode() != Instruction::Xor)
1626         return nullptr;
1627       Instruction *And0 = dyn_cast<Instruction>(I->getOperand(0));
1628       Instruction *And1 = dyn_cast<Instruction>(I->getOperand(1));
1629       if (!And0 || !And1)
1630         return nullptr;
1631       if (And0->getOpcode() != Instruction::And ||
1632           And1->getOpcode() != Instruction::And)
1633         return nullptr;
1634       if (And0->getOperand(1) != And1->getOperand(1))
1635         return nullptr;
1636       IRBuilder<> B(Ctx);
1637       return B.CreateAnd(B.CreateXor(And0->getOperand(0), And1->getOperand(0)),
1638                          And0->getOperand(1));
1639     });
1640   S.addRule("sink binop into select",
1641     // (Op (select c x y) z) -> (select c (Op x z) (Op y z))
1642     // (Op x (select c y z)) -> (select c (Op x y) (Op x z))
1643     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1644       BinaryOperator *BO = dyn_cast<BinaryOperator>(I);
1645       if (!BO)
1646         return nullptr;
1647       Instruction::BinaryOps Op = BO->getOpcode();
1648       if (SelectInst *Sel = dyn_cast<SelectInst>(BO->getOperand(0))) {
1649         IRBuilder<> B(Ctx);
1650         Value *X = Sel->getTrueValue(), *Y = Sel->getFalseValue();
1651         Value *Z = BO->getOperand(1);
1652         return B.CreateSelect(Sel->getCondition(),
1653                               B.CreateBinOp(Op, X, Z),
1654                               B.CreateBinOp(Op, Y, Z));
1655       }
1656       if (SelectInst *Sel = dyn_cast<SelectInst>(BO->getOperand(1))) {
1657         IRBuilder<> B(Ctx);
1658         Value *X = BO->getOperand(0);
1659         Value *Y = Sel->getTrueValue(), *Z = Sel->getFalseValue();
1660         return B.CreateSelect(Sel->getCondition(),
1661                               B.CreateBinOp(Op, X, Y),
1662                               B.CreateBinOp(Op, X, Z));
1663       }
1664       return nullptr;
1665     });
1666   S.addRule("fold select-select",
1667     // (select c (select c x y) z) -> (select c x z)
1668     // (select c x (select c y z)) -> (select c x z)
1669     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1670       SelectInst *Sel = dyn_cast<SelectInst>(I);
1671       if (!Sel)
1672         return nullptr;
1673       IRBuilder<> B(Ctx);
1674       Value *C = Sel->getCondition();
1675       if (SelectInst *Sel0 = dyn_cast<SelectInst>(Sel->getTrueValue())) {
1676         if (Sel0->getCondition() == C)
1677           return B.CreateSelect(C, Sel0->getTrueValue(), Sel->getFalseValue());
1678       }
1679       if (SelectInst *Sel1 = dyn_cast<SelectInst>(Sel->getFalseValue())) {
1680         if (Sel1->getCondition() == C)
1681           return B.CreateSelect(C, Sel->getTrueValue(), Sel1->getFalseValue());
1682       }
1683       return nullptr;
1684     });
1685   S.addRule("or-signbit -> xor-signbit",
1686     // (or (lshr x 1) 0x800.0) -> (xor (lshr x 1) 0x800.0)
1687     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1688       if (I->getOpcode() != Instruction::Or)
1689         return nullptr;
1690       ConstantInt *Msb = dyn_cast<ConstantInt>(I->getOperand(1));
1691       if (!Msb || Msb->getZExtValue() != Msb->getType()->getSignBit())
1692         return nullptr;
1693       if (!hasZeroSignBit(I->getOperand(0)))
1694         return nullptr;
1695       return IRBuilder<>(Ctx).CreateXor(I->getOperand(0), Msb);
1696     });
1697   S.addRule("sink lshr into binop",
1698     // (lshr (BitOp x y) c) -> (BitOp (lshr x c) (lshr y c))
1699     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1700       if (I->getOpcode() != Instruction::LShr)
1701         return nullptr;
1702       BinaryOperator *BitOp = dyn_cast<BinaryOperator>(I->getOperand(0));
1703       if (!BitOp)
1704         return nullptr;
1705       switch (BitOp->getOpcode()) {
1706         case Instruction::And:
1707         case Instruction::Or:
1708         case Instruction::Xor:
1709           break;
1710         default:
1711           return nullptr;
1712       }
1713       IRBuilder<> B(Ctx);
1714       Value *S = I->getOperand(1);
1715       return B.CreateBinOp(BitOp->getOpcode(),
1716                 B.CreateLShr(BitOp->getOperand(0), S),
1717                 B.CreateLShr(BitOp->getOperand(1), S));
1718     });
1719   S.addRule("expose bitop-const",
1720     // (BitOp1 (BitOp2 x a) b) -> (BitOp2 x (BitOp1 a b))
1721     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1722       auto IsBitOp = [](unsigned Op) -> bool {
1723         switch (Op) {
1724           case Instruction::And:
1725           case Instruction::Or:
1726           case Instruction::Xor:
1727             return true;
1728         }
1729         return false;
1730       };
1731       BinaryOperator *BitOp1 = dyn_cast<BinaryOperator>(I);
1732       if (!BitOp1 || !IsBitOp(BitOp1->getOpcode()))
1733         return nullptr;
1734       BinaryOperator *BitOp2 = dyn_cast<BinaryOperator>(BitOp1->getOperand(0));
1735       if (!BitOp2 || !IsBitOp(BitOp2->getOpcode()))
1736         return nullptr;
1737       ConstantInt *CA = dyn_cast<ConstantInt>(BitOp2->getOperand(1));
1738       ConstantInt *CB = dyn_cast<ConstantInt>(BitOp1->getOperand(1));
1739       if (!CA || !CB)
1740         return nullptr;
1741       IRBuilder<> B(Ctx);
1742       Value *X = BitOp2->getOperand(0);
1743       return B.CreateBinOp(BitOp2->getOpcode(), X,
1744                 B.CreateBinOp(BitOp1->getOpcode(), CA, CB));
1745     });
1746 }
1747 
1748 void PolynomialMultiplyRecognize::setupPostSimplifier(Simplifier &S) {
1749   S.addRule("(and (xor (and x a) y) b) -> (and (xor x y) b), if b == b&a",
1750     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1751       if (I->getOpcode() != Instruction::And)
1752         return nullptr;
1753       Instruction *Xor = dyn_cast<Instruction>(I->getOperand(0));
1754       ConstantInt *C0 = dyn_cast<ConstantInt>(I->getOperand(1));
1755       if (!Xor || !C0)
1756         return nullptr;
1757       if (Xor->getOpcode() != Instruction::Xor)
1758         return nullptr;
1759       Instruction *And0 = dyn_cast<Instruction>(Xor->getOperand(0));
1760       Instruction *And1 = dyn_cast<Instruction>(Xor->getOperand(1));
1761       // Pick the first non-null and.
1762       if (!And0 || And0->getOpcode() != Instruction::And)
1763         std::swap(And0, And1);
1764       ConstantInt *C1 = dyn_cast<ConstantInt>(And0->getOperand(1));
1765       if (!C1)
1766         return nullptr;
1767       uint32_t V0 = C0->getZExtValue();
1768       uint32_t V1 = C1->getZExtValue();
1769       if (V0 != (V0 & V1))
1770         return nullptr;
1771       IRBuilder<> B(Ctx);
1772       return B.CreateAnd(B.CreateXor(And0->getOperand(0), And1), C0);
1773     });
1774 }
1775 
1776 bool PolynomialMultiplyRecognize::recognize() {
1777   LLVM_DEBUG(dbgs() << "Starting PolynomialMultiplyRecognize on loop\n"
1778                     << *CurLoop << '\n');
1779   // Restrictions:
1780   // - The loop must consist of a single block.
1781   // - The iteration count must be known at compile-time.
1782   // - The loop must have an induction variable starting from 0, and
1783   //   incremented in each iteration of the loop.
1784   BasicBlock *LoopB = CurLoop->getHeader();
1785   LLVM_DEBUG(dbgs() << "Loop header:\n" << *LoopB);
1786 
1787   if (LoopB != CurLoop->getLoopLatch())
1788     return false;
1789   BasicBlock *ExitB = CurLoop->getExitBlock();
1790   if (ExitB == nullptr)
1791     return false;
1792   BasicBlock *EntryB = CurLoop->getLoopPreheader();
1793   if (EntryB == nullptr)
1794     return false;
1795 
1796   unsigned IterCount = 0;
1797   const SCEV *CT = SE.getBackedgeTakenCount(CurLoop);
1798   if (isa<SCEVCouldNotCompute>(CT))
1799     return false;
1800   if (auto *CV = dyn_cast<SCEVConstant>(CT))
1801     IterCount = CV->getValue()->getZExtValue() + 1;
1802 
1803   Value *CIV = getCountIV(LoopB);
1804   ParsedValues PV;
1805   Simplifier PreSimp;
1806   PV.IterCount = IterCount;
1807   LLVM_DEBUG(dbgs() << "Loop IV: " << *CIV << "\nIterCount: " << IterCount
1808                     << '\n');
1809 
1810   setupPreSimplifier(PreSimp);
1811 
1812   // Perform a preliminary scan of select instructions to see if any of them
1813   // looks like a generator of the polynomial multiply steps. Assume that a
1814   // loop can only contain a single transformable operation, so stop the
1815   // traversal after the first reasonable candidate was found.
1816   // XXX: Currently this approach can modify the loop before being 100% sure
1817   // that the transformation can be carried out.
1818   bool FoundPreScan = false;
1819   auto FeedsPHI = [LoopB](const Value *V) -> bool {
1820     for (const Value *U : V->users()) {
1821       if (const auto *P = dyn_cast<const PHINode>(U))
1822         if (P->getParent() == LoopB)
1823           return true;
1824     }
1825     return false;
1826   };
1827   for (Instruction &In : *LoopB) {
1828     SelectInst *SI = dyn_cast<SelectInst>(&In);
1829     if (!SI || !FeedsPHI(SI))
1830       continue;
1831 
1832     Simplifier::Context C(SI);
1833     Value *T = PreSimp.simplify(C);
1834     SelectInst *SelI = (T && isa<SelectInst>(T)) ? cast<SelectInst>(T) : SI;
1835     LLVM_DEBUG(dbgs() << "scanSelect(pre-scan): " << PE(C, SelI) << '\n');
1836     if (scanSelect(SelI, LoopB, EntryB, CIV, PV, true)) {
1837       FoundPreScan = true;
1838       if (SelI != SI) {
1839         Value *NewSel = C.materialize(LoopB, SI->getIterator());
1840         SI->replaceAllUsesWith(NewSel);
1841         RecursivelyDeleteTriviallyDeadInstructions(SI, &TLI);
1842       }
1843       break;
1844     }
1845   }
1846 
1847   if (!FoundPreScan) {
1848     LLVM_DEBUG(dbgs() << "Have not found candidates for pmpy\n");
1849     return false;
1850   }
1851 
1852   if (!PV.Left) {
1853     // The right shift version actually only returns the higher bits of
1854     // the result (each iteration discards the LSB). If we want to convert it
1855     // to a left-shifting loop, the working data type must be at least as
1856     // wide as the target's pmpy instruction.
1857     if (!promoteTypes(LoopB, ExitB))
1858       return false;
1859     // Run post-promotion simplifications.
1860     Simplifier PostSimp;
1861     setupPostSimplifier(PostSimp);
1862     for (Instruction &In : *LoopB) {
1863       SelectInst *SI = dyn_cast<SelectInst>(&In);
1864       if (!SI || !FeedsPHI(SI))
1865         continue;
1866       Simplifier::Context C(SI);
1867       Value *T = PostSimp.simplify(C);
1868       SelectInst *SelI = dyn_cast_or_null<SelectInst>(T);
1869       if (SelI != SI) {
1870         Value *NewSel = C.materialize(LoopB, SI->getIterator());
1871         SI->replaceAllUsesWith(NewSel);
1872         RecursivelyDeleteTriviallyDeadInstructions(SI, &TLI);
1873       }
1874       break;
1875     }
1876 
1877     if (!convertShiftsToLeft(LoopB, ExitB, IterCount))
1878       return false;
1879     cleanupLoopBody(LoopB);
1880   }
1881 
1882   // Scan the loop again, find the generating select instruction.
1883   bool FoundScan = false;
1884   for (Instruction &In : *LoopB) {
1885     SelectInst *SelI = dyn_cast<SelectInst>(&In);
1886     if (!SelI)
1887       continue;
1888     LLVM_DEBUG(dbgs() << "scanSelect: " << *SelI << '\n');
1889     FoundScan = scanSelect(SelI, LoopB, EntryB, CIV, PV, false);
1890     if (FoundScan)
1891       break;
1892   }
1893   assert(FoundScan);
1894 
1895   LLVM_DEBUG({
1896     StringRef PP = (PV.M ? "(P+M)" : "P");
1897     if (!PV.Inv)
1898       dbgs() << "Found pmpy idiom: R = " << PP << ".Q\n";
1899     else
1900       dbgs() << "Found inverse pmpy idiom: R = (" << PP << "/Q).Q) + "
1901              << PP << "\n";
1902     dbgs() << "  Res:" << *PV.Res << "\n  P:" << *PV.P << "\n";
1903     if (PV.M)
1904       dbgs() << "  M:" << *PV.M << "\n";
1905     dbgs() << "  Q:" << *PV.Q << "\n";
1906     dbgs() << "  Iteration count:" << PV.IterCount << "\n";
1907   });
1908 
1909   BasicBlock::iterator At(EntryB->getTerminator());
1910   Value *PM = generate(At, PV);
1911   if (PM == nullptr)
1912     return false;
1913 
1914   if (PM->getType() != PV.Res->getType())
1915     PM = IRBuilder<>(&*At).CreateIntCast(PM, PV.Res->getType(), false);
1916 
1917   PV.Res->replaceAllUsesWith(PM);
1918   PV.Res->eraseFromParent();
1919   return true;
1920 }
1921 
1922 int HexagonLoopIdiomRecognize::getSCEVStride(const SCEVAddRecExpr *S) {
1923   if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(1)))
1924     return SC->getAPInt().getSExtValue();
1925   return 0;
1926 }
1927 
1928 bool HexagonLoopIdiomRecognize::isLegalStore(Loop *CurLoop, StoreInst *SI) {
1929   // Allow volatile stores if HexagonVolatileMemcpy is enabled.
1930   if (!(SI->isVolatile() && HexagonVolatileMemcpy) && !SI->isSimple())
1931     return false;
1932 
1933   Value *StoredVal = SI->getValueOperand();
1934   Value *StorePtr = SI->getPointerOperand();
1935 
1936   // Reject stores that are so large that they overflow an unsigned.
1937   uint64_t SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
1938   if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
1939     return false;
1940 
1941   // See if the pointer expression is an AddRec like {base,+,1} on the current
1942   // loop, which indicates a strided store.  If we have something else, it's a
1943   // random store we can't handle.
1944   auto *StoreEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
1945   if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
1946     return false;
1947 
1948   // Check to see if the stride matches the size of the store.  If so, then we
1949   // know that every byte is touched in the loop.
1950   int Stride = getSCEVStride(StoreEv);
1951   if (Stride == 0)
1952     return false;
1953   unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType());
1954   if (StoreSize != unsigned(std::abs(Stride)))
1955     return false;
1956 
1957   // The store must be feeding a non-volatile load.
1958   LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand());
1959   if (!LI || !LI->isSimple())
1960     return false;
1961 
1962   // See if the pointer expression is an AddRec like {base,+,1} on the current
1963   // loop, which indicates a strided load.  If we have something else, it's a
1964   // random load we can't handle.
1965   Value *LoadPtr = LI->getPointerOperand();
1966   auto *LoadEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LoadPtr));
1967   if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
1968     return false;
1969 
1970   // The store and load must share the same stride.
1971   if (StoreEv->getOperand(1) != LoadEv->getOperand(1))
1972     return false;
1973 
1974   // Success.  This store can be converted into a memcpy.
1975   return true;
1976 }
1977 
1978 /// mayLoopAccessLocation - Return true if the specified loop might access the
1979 /// specified pointer location, which is a loop-strided access.  The 'Access'
1980 /// argument specifies what the verboten forms of access are (read or write).
1981 static bool
1982 mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
1983                       const SCEV *BECount, unsigned StoreSize,
1984                       AliasAnalysis &AA,
1985                       SmallPtrSetImpl<Instruction *> &Ignored) {
1986   // Get the location that may be stored across the loop.  Since the access
1987   // is strided positively through memory, we say that the modified location
1988   // starts at the pointer and has infinite size.
1989   LocationSize AccessSize = LocationSize::afterPointer();
1990 
1991   // If the loop iterates a fixed number of times, we can refine the access
1992   // size to be exactly the size of the memset, which is (BECount+1)*StoreSize
1993   if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
1994     AccessSize = LocationSize::precise((BECst->getValue()->getZExtValue() + 1) *
1995                                        StoreSize);
1996 
1997   // TODO: For this to be really effective, we have to dive into the pointer
1998   // operand in the store.  Store to &A[i] of 100 will always return may alias
1999   // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
2000   // which will then no-alias a store to &A[100].
2001   MemoryLocation StoreLoc(Ptr, AccessSize);
2002 
2003   for (auto *B : L->blocks())
2004     for (auto &I : *B)
2005       if (Ignored.count(&I) == 0 &&
2006           isModOrRefSet(AA.getModRefInfo(&I, StoreLoc) & Access))
2007         return true;
2008 
2009   return false;
2010 }
2011 
2012 void HexagonLoopIdiomRecognize::collectStores(Loop *CurLoop, BasicBlock *BB,
2013       SmallVectorImpl<StoreInst*> &Stores) {
2014   Stores.clear();
2015   for (Instruction &I : *BB)
2016     if (StoreInst *SI = dyn_cast<StoreInst>(&I))
2017       if (isLegalStore(CurLoop, SI))
2018         Stores.push_back(SI);
2019 }
2020 
2021 bool HexagonLoopIdiomRecognize::processCopyingStore(Loop *CurLoop,
2022       StoreInst *SI, const SCEV *BECount) {
2023   assert((SI->isSimple() || (SI->isVolatile() && HexagonVolatileMemcpy)) &&
2024          "Expected only non-volatile stores, or Hexagon-specific memcpy"
2025          "to volatile destination.");
2026 
2027   Value *StorePtr = SI->getPointerOperand();
2028   auto *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
2029   unsigned Stride = getSCEVStride(StoreEv);
2030   unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType());
2031   if (Stride != StoreSize)
2032     return false;
2033 
2034   // See if the pointer expression is an AddRec like {base,+,1} on the current
2035   // loop, which indicates a strided load.  If we have something else, it's a
2036   // random load we can't handle.
2037   auto *LI = cast<LoadInst>(SI->getValueOperand());
2038   auto *LoadEv = cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
2039 
2040   // The trip count of the loop and the base pointer of the addrec SCEV is
2041   // guaranteed to be loop invariant, which means that it should dominate the
2042   // header.  This allows us to insert code for it in the preheader.
2043   BasicBlock *Preheader = CurLoop->getLoopPreheader();
2044   Instruction *ExpPt = Preheader->getTerminator();
2045   IRBuilder<> Builder(ExpPt);
2046   SCEVExpander Expander(*SE, *DL, "hexagon-loop-idiom");
2047 
2048   Type *IntPtrTy = Builder.getIntPtrTy(*DL, SI->getPointerAddressSpace());
2049 
2050   // Okay, we have a strided store "p[i]" of a loaded value.  We can turn
2051   // this into a memcpy/memmove in the loop preheader now if we want.  However,
2052   // this would be unsafe to do if there is anything else in the loop that may
2053   // read or write the memory region we're storing to.  For memcpy, this
2054   // includes the load that feeds the stores.  Check for an alias by generating
2055   // the base address and checking everything.
2056   Value *StoreBasePtr = Expander.expandCodeFor(StoreEv->getStart(),
2057       Builder.getInt8PtrTy(SI->getPointerAddressSpace()), ExpPt);
2058   Value *LoadBasePtr = nullptr;
2059 
2060   bool Overlap = false;
2061   bool DestVolatile = SI->isVolatile();
2062   Type *BECountTy = BECount->getType();
2063 
2064   if (DestVolatile) {
2065     // The trip count must fit in i32, since it is the type of the "num_words"
2066     // argument to hexagon_memcpy_forward_vp4cp4n2.
2067     if (StoreSize != 4 || DL->getTypeSizeInBits(BECountTy) > 32) {
2068 CleanupAndExit:
2069       // If we generated new code for the base pointer, clean up.
2070       Expander.clear();
2071       if (StoreBasePtr && (LoadBasePtr != StoreBasePtr)) {
2072         RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
2073         StoreBasePtr = nullptr;
2074       }
2075       if (LoadBasePtr) {
2076         RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
2077         LoadBasePtr = nullptr;
2078       }
2079       return false;
2080     }
2081   }
2082 
2083   SmallPtrSet<Instruction*, 2> Ignore1;
2084   Ignore1.insert(SI);
2085   if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount,
2086                             StoreSize, *AA, Ignore1)) {
2087     // Check if the load is the offending instruction.
2088     Ignore1.insert(LI);
2089     if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop,
2090                               BECount, StoreSize, *AA, Ignore1)) {
2091       // Still bad. Nothing we can do.
2092       goto CleanupAndExit;
2093     }
2094     // It worked with the load ignored.
2095     Overlap = true;
2096   }
2097 
2098   if (!Overlap) {
2099     if (DisableMemcpyIdiom || !HasMemcpy)
2100       goto CleanupAndExit;
2101   } else {
2102     // Don't generate memmove if this function will be inlined. This is
2103     // because the caller will undergo this transformation after inlining.
2104     Function *Func = CurLoop->getHeader()->getParent();
2105     if (Func->hasFnAttribute(Attribute::AlwaysInline))
2106       goto CleanupAndExit;
2107 
2108     // In case of a memmove, the call to memmove will be executed instead
2109     // of the loop, so we need to make sure that there is nothing else in
2110     // the loop than the load, store and instructions that these two depend
2111     // on.
2112     SmallVector<Instruction*,2> Insts;
2113     Insts.push_back(SI);
2114     Insts.push_back(LI);
2115     if (!coverLoop(CurLoop, Insts))
2116       goto CleanupAndExit;
2117 
2118     if (DisableMemmoveIdiom || !HasMemmove)
2119       goto CleanupAndExit;
2120     bool IsNested = CurLoop->getParentLoop() != nullptr;
2121     if (IsNested && OnlyNonNestedMemmove)
2122       goto CleanupAndExit;
2123   }
2124 
2125   // For a memcpy, we have to make sure that the input array is not being
2126   // mutated by the loop.
2127   LoadBasePtr = Expander.expandCodeFor(LoadEv->getStart(),
2128       Builder.getInt8PtrTy(LI->getPointerAddressSpace()), ExpPt);
2129 
2130   SmallPtrSet<Instruction*, 2> Ignore2;
2131   Ignore2.insert(SI);
2132   if (mayLoopAccessLocation(LoadBasePtr, ModRefInfo::Mod, CurLoop, BECount,
2133                             StoreSize, *AA, Ignore2))
2134     goto CleanupAndExit;
2135 
2136   // Check the stride.
2137   bool StridePos = getSCEVStride(LoadEv) >= 0;
2138 
2139   // Currently, the volatile memcpy only emulates traversing memory forward.
2140   if (!StridePos && DestVolatile)
2141     goto CleanupAndExit;
2142 
2143   bool RuntimeCheck = (Overlap || DestVolatile);
2144 
2145   BasicBlock *ExitB;
2146   if (RuntimeCheck) {
2147     // The runtime check needs a single exit block.
2148     SmallVector<BasicBlock*, 8> ExitBlocks;
2149     CurLoop->getUniqueExitBlocks(ExitBlocks);
2150     if (ExitBlocks.size() != 1)
2151       goto CleanupAndExit;
2152     ExitB = ExitBlocks[0];
2153   }
2154 
2155   // The # stored bytes is (BECount+1)*Size.  Expand the trip count out to
2156   // pointer size if it isn't already.
2157   LLVMContext &Ctx = SI->getContext();
2158   BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
2159   DebugLoc DLoc = SI->getDebugLoc();
2160 
2161   const SCEV *NumBytesS =
2162       SE->getAddExpr(BECount, SE->getOne(IntPtrTy), SCEV::FlagNUW);
2163   if (StoreSize != 1)
2164     NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
2165                                SCEV::FlagNUW);
2166   Value *NumBytes = Expander.expandCodeFor(NumBytesS, IntPtrTy, ExpPt);
2167   if (Instruction *In = dyn_cast<Instruction>(NumBytes))
2168     if (Value *Simp = simplifyInstruction(In, {*DL, TLI, DT}))
2169       NumBytes = Simp;
2170 
2171   CallInst *NewCall;
2172 
2173   if (RuntimeCheck) {
2174     unsigned Threshold = RuntimeMemSizeThreshold;
2175     if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes)) {
2176       uint64_t C = CI->getZExtValue();
2177       if (Threshold != 0 && C < Threshold)
2178         goto CleanupAndExit;
2179       if (C < CompileTimeMemSizeThreshold)
2180         goto CleanupAndExit;
2181     }
2182 
2183     BasicBlock *Header = CurLoop->getHeader();
2184     Function *Func = Header->getParent();
2185     Loop *ParentL = LF->getLoopFor(Preheader);
2186     StringRef HeaderName = Header->getName();
2187 
2188     // Create a new (empty) preheader, and update the PHI nodes in the
2189     // header to use the new preheader.
2190     BasicBlock *NewPreheader = BasicBlock::Create(Ctx, HeaderName+".rtli.ph",
2191                                                   Func, Header);
2192     if (ParentL)
2193       ParentL->addBasicBlockToLoop(NewPreheader, *LF);
2194     IRBuilder<>(NewPreheader).CreateBr(Header);
2195     for (auto &In : *Header) {
2196       PHINode *PN = dyn_cast<PHINode>(&In);
2197       if (!PN)
2198         break;
2199       int bx = PN->getBasicBlockIndex(Preheader);
2200       if (bx >= 0)
2201         PN->setIncomingBlock(bx, NewPreheader);
2202     }
2203     DT->addNewBlock(NewPreheader, Preheader);
2204     DT->changeImmediateDominator(Header, NewPreheader);
2205 
2206     // Check for safe conditions to execute memmove.
2207     // If stride is positive, copying things from higher to lower addresses
2208     // is equivalent to memmove.  For negative stride, it's the other way
2209     // around.  Copying forward in memory with positive stride may not be
2210     // same as memmove since we may be copying values that we just stored
2211     // in some previous iteration.
2212     Value *LA = Builder.CreatePtrToInt(LoadBasePtr, IntPtrTy);
2213     Value *SA = Builder.CreatePtrToInt(StoreBasePtr, IntPtrTy);
2214     Value *LowA = StridePos ? SA : LA;
2215     Value *HighA = StridePos ? LA : SA;
2216     Value *CmpA = Builder.CreateICmpULT(LowA, HighA);
2217     Value *Cond = CmpA;
2218 
2219     // Check for distance between pointers. Since the case LowA < HighA
2220     // is checked for above, assume LowA >= HighA.
2221     Value *Dist = Builder.CreateSub(LowA, HighA);
2222     Value *CmpD = Builder.CreateICmpSLE(NumBytes, Dist);
2223     Value *CmpEither = Builder.CreateOr(Cond, CmpD);
2224     Cond = CmpEither;
2225 
2226     if (Threshold != 0) {
2227       Type *Ty = NumBytes->getType();
2228       Value *Thr = ConstantInt::get(Ty, Threshold);
2229       Value *CmpB = Builder.CreateICmpULT(Thr, NumBytes);
2230       Value *CmpBoth = Builder.CreateAnd(Cond, CmpB);
2231       Cond = CmpBoth;
2232     }
2233     BasicBlock *MemmoveB = BasicBlock::Create(Ctx, Header->getName()+".rtli",
2234                                               Func, NewPreheader);
2235     if (ParentL)
2236       ParentL->addBasicBlockToLoop(MemmoveB, *LF);
2237     Instruction *OldT = Preheader->getTerminator();
2238     Builder.CreateCondBr(Cond, MemmoveB, NewPreheader);
2239     OldT->eraseFromParent();
2240     Preheader->setName(Preheader->getName()+".old");
2241     DT->addNewBlock(MemmoveB, Preheader);
2242     // Find the new immediate dominator of the exit block.
2243     BasicBlock *ExitD = Preheader;
2244     for (BasicBlock *PB : predecessors(ExitB)) {
2245       ExitD = DT->findNearestCommonDominator(ExitD, PB);
2246       if (!ExitD)
2247         break;
2248     }
2249     // If the prior immediate dominator of ExitB was dominated by the
2250     // old preheader, then the old preheader becomes the new immediate
2251     // dominator.  Otherwise don't change anything (because the newly
2252     // added blocks are dominated by the old preheader).
2253     if (ExitD && DT->dominates(Preheader, ExitD)) {
2254       DomTreeNode *BN = DT->getNode(ExitB);
2255       DomTreeNode *DN = DT->getNode(ExitD);
2256       BN->setIDom(DN);
2257     }
2258 
2259     // Add a call to memmove to the conditional block.
2260     IRBuilder<> CondBuilder(MemmoveB);
2261     CondBuilder.CreateBr(ExitB);
2262     CondBuilder.SetInsertPoint(MemmoveB->getTerminator());
2263 
2264     if (DestVolatile) {
2265       Type *Int32Ty = Type::getInt32Ty(Ctx);
2266       Type *Int32PtrTy = Type::getInt32PtrTy(Ctx);
2267       Type *VoidTy = Type::getVoidTy(Ctx);
2268       Module *M = Func->getParent();
2269       FunctionCallee Fn = M->getOrInsertFunction(
2270           HexagonVolatileMemcpyName, VoidTy, Int32PtrTy, Int32PtrTy, Int32Ty);
2271 
2272       const SCEV *OneS = SE->getConstant(Int32Ty, 1);
2273       const SCEV *BECount32 = SE->getTruncateOrZeroExtend(BECount, Int32Ty);
2274       const SCEV *NumWordsS = SE->getAddExpr(BECount32, OneS, SCEV::FlagNUW);
2275       Value *NumWords = Expander.expandCodeFor(NumWordsS, Int32Ty,
2276                                                MemmoveB->getTerminator());
2277       if (Instruction *In = dyn_cast<Instruction>(NumWords))
2278         if (Value *Simp = simplifyInstruction(In, {*DL, TLI, DT}))
2279           NumWords = Simp;
2280 
2281       Value *Op0 = (StoreBasePtr->getType() == Int32PtrTy)
2282                       ? StoreBasePtr
2283                       : CondBuilder.CreateBitCast(StoreBasePtr, Int32PtrTy);
2284       Value *Op1 = (LoadBasePtr->getType() == Int32PtrTy)
2285                       ? LoadBasePtr
2286                       : CondBuilder.CreateBitCast(LoadBasePtr, Int32PtrTy);
2287       NewCall = CondBuilder.CreateCall(Fn, {Op0, Op1, NumWords});
2288     } else {
2289       NewCall = CondBuilder.CreateMemMove(
2290           StoreBasePtr, SI->getAlign(), LoadBasePtr, LI->getAlign(), NumBytes);
2291     }
2292   } else {
2293     NewCall = Builder.CreateMemCpy(StoreBasePtr, SI->getAlign(), LoadBasePtr,
2294                                    LI->getAlign(), NumBytes);
2295     // Okay, the memcpy has been formed.  Zap the original store and
2296     // anything that feeds into it.
2297     RecursivelyDeleteTriviallyDeadInstructions(SI, TLI);
2298   }
2299 
2300   NewCall->setDebugLoc(DLoc);
2301 
2302   LLVM_DEBUG(dbgs() << "  Formed " << (Overlap ? "memmove: " : "memcpy: ")
2303                     << *NewCall << "\n"
2304                     << "    from load ptr=" << *LoadEv << " at: " << *LI << "\n"
2305                     << "    from store ptr=" << *StoreEv << " at: " << *SI
2306                     << "\n");
2307 
2308   return true;
2309 }
2310 
2311 // Check if the instructions in Insts, together with their dependencies
2312 // cover the loop in the sense that the loop could be safely eliminated once
2313 // the instructions in Insts are removed.
2314 bool HexagonLoopIdiomRecognize::coverLoop(Loop *L,
2315       SmallVectorImpl<Instruction*> &Insts) const {
2316   SmallSet<BasicBlock*,8> LoopBlocks;
2317   for (auto *B : L->blocks())
2318     LoopBlocks.insert(B);
2319 
2320   SetVector<Instruction*> Worklist(Insts.begin(), Insts.end());
2321 
2322   // Collect all instructions from the loop that the instructions in Insts
2323   // depend on (plus their dependencies, etc.).  These instructions will
2324   // constitute the expression trees that feed those in Insts, but the trees
2325   // will be limited only to instructions contained in the loop.
2326   for (unsigned i = 0; i < Worklist.size(); ++i) {
2327     Instruction *In = Worklist[i];
2328     for (auto I = In->op_begin(), E = In->op_end(); I != E; ++I) {
2329       Instruction *OpI = dyn_cast<Instruction>(I);
2330       if (!OpI)
2331         continue;
2332       BasicBlock *PB = OpI->getParent();
2333       if (!LoopBlocks.count(PB))
2334         continue;
2335       Worklist.insert(OpI);
2336     }
2337   }
2338 
2339   // Scan all instructions in the loop, if any of them have a user outside
2340   // of the loop, or outside of the expressions collected above, then either
2341   // the loop has a side-effect visible outside of it, or there are
2342   // instructions in it that are not involved in the original set Insts.
2343   for (auto *B : L->blocks()) {
2344     for (auto &In : *B) {
2345       if (isa<BranchInst>(In) || isa<DbgInfoIntrinsic>(In))
2346         continue;
2347       if (!Worklist.count(&In) && In.mayHaveSideEffects())
2348         return false;
2349       for (auto *K : In.users()) {
2350         Instruction *UseI = dyn_cast<Instruction>(K);
2351         if (!UseI)
2352           continue;
2353         BasicBlock *UseB = UseI->getParent();
2354         if (LF->getLoopFor(UseB) != L)
2355           return false;
2356       }
2357     }
2358   }
2359 
2360   return true;
2361 }
2362 
2363 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
2364 /// with the specified backedge count.  This block is known to be in the current
2365 /// loop and not in any subloops.
2366 bool HexagonLoopIdiomRecognize::runOnLoopBlock(Loop *CurLoop, BasicBlock *BB,
2367       const SCEV *BECount, SmallVectorImpl<BasicBlock*> &ExitBlocks) {
2368   // We can only promote stores in this block if they are unconditionally
2369   // executed in the loop.  For a block to be unconditionally executed, it has
2370   // to dominate all the exit blocks of the loop.  Verify this now.
2371   auto DominatedByBB = [this,BB] (BasicBlock *EB) -> bool {
2372     return DT->dominates(BB, EB);
2373   };
2374   if (!all_of(ExitBlocks, DominatedByBB))
2375     return false;
2376 
2377   bool MadeChange = false;
2378   // Look for store instructions, which may be optimized to memset/memcpy.
2379   SmallVector<StoreInst*,8> Stores;
2380   collectStores(CurLoop, BB, Stores);
2381 
2382   // Optimize the store into a memcpy, if it feeds an similarly strided load.
2383   for (auto &SI : Stores)
2384     MadeChange |= processCopyingStore(CurLoop, SI, BECount);
2385 
2386   return MadeChange;
2387 }
2388 
2389 bool HexagonLoopIdiomRecognize::runOnCountableLoop(Loop *L) {
2390   PolynomialMultiplyRecognize PMR(L, *DL, *DT, *TLI, *SE);
2391   if (PMR.recognize())
2392     return true;
2393 
2394   if (!HasMemcpy && !HasMemmove)
2395     return false;
2396 
2397   const SCEV *BECount = SE->getBackedgeTakenCount(L);
2398   assert(!isa<SCEVCouldNotCompute>(BECount) &&
2399          "runOnCountableLoop() called on a loop without a predictable"
2400          "backedge-taken count");
2401 
2402   SmallVector<BasicBlock *, 8> ExitBlocks;
2403   L->getUniqueExitBlocks(ExitBlocks);
2404 
2405   bool Changed = false;
2406 
2407   // Scan all the blocks in the loop that are not in subloops.
2408   for (auto *BB : L->getBlocks()) {
2409     // Ignore blocks in subloops.
2410     if (LF->getLoopFor(BB) != L)
2411       continue;
2412     Changed |= runOnLoopBlock(L, BB, BECount, ExitBlocks);
2413   }
2414 
2415   return Changed;
2416 }
2417 
2418 bool HexagonLoopIdiomRecognize::run(Loop *L) {
2419   const Module &M = *L->getHeader()->getParent()->getParent();
2420   if (Triple(M.getTargetTriple()).getArch() != Triple::hexagon)
2421     return false;
2422 
2423   // If the loop could not be converted to canonical form, it must have an
2424   // indirectbr in it, just give up.
2425   if (!L->getLoopPreheader())
2426     return false;
2427 
2428   // Disable loop idiom recognition if the function's name is a common idiom.
2429   StringRef Name = L->getHeader()->getParent()->getName();
2430   if (Name == "memset" || Name == "memcpy" || Name == "memmove")
2431     return false;
2432 
2433   DL = &L->getHeader()->getModule()->getDataLayout();
2434 
2435   HasMemcpy = TLI->has(LibFunc_memcpy);
2436   HasMemmove = TLI->has(LibFunc_memmove);
2437 
2438   if (SE->hasLoopInvariantBackedgeTakenCount(L))
2439     return runOnCountableLoop(L);
2440   return false;
2441 }
2442 
2443 bool HexagonLoopIdiomRecognizeLegacyPass::runOnLoop(Loop *L,
2444                                                     LPPassManager &LPM) {
2445   if (skipLoop(L))
2446     return false;
2447 
2448   auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2449   auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2450   auto *LF = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2451   auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
2452       *L->getHeader()->getParent());
2453   auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2454   return HexagonLoopIdiomRecognize(AA, DT, LF, TLI, SE).run(L);
2455 }
2456 
2457 Pass *llvm::createHexagonLoopIdiomPass() {
2458   return new HexagonLoopIdiomRecognizeLegacyPass();
2459 }
2460 
2461 PreservedAnalyses
2462 HexagonLoopIdiomRecognitionPass::run(Loop &L, LoopAnalysisManager &AM,
2463                                      LoopStandardAnalysisResults &AR,
2464                                      LPMUpdater &U) {
2465   return HexagonLoopIdiomRecognize(&AR.AA, &AR.DT, &AR.LI, &AR.TLI, &AR.SE)
2466                  .run(&L)
2467              ? getLoopPassPreservedAnalyses()
2468              : PreservedAnalyses::all();
2469 }
2470