xref: /freebsd/contrib/llvm-project/llvm/lib/Target/Hexagon/HexagonLoopIdiomRecognition.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
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:
HexagonLoopIdiomRecognize(AliasAnalysis * AA,DominatorTree * DT,LoopInfo * LF,const TargetLibraryInfo * TLI,ScalarEvolution * SE)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 
HexagonLoopIdiomRecognizeLegacyPass()154   explicit HexagonLoopIdiomRecognizeLegacyPass() : LoopPass(ID) {
155     initializeHexagonLoopIdiomRecognizeLegacyPassPass(
156         *PassRegistry::getPassRegistry());
157   }
158 
getPassName() const159   StringRef getPassName() const override {
160     return "Recognize Hexagon-specific loop idioms";
161   }
162 
getAnalysisUsage(AnalysisUsage & AU) const163   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 &)>;
Rule__anona63979960111::Simplifier::Rule180     Rule(StringRef N, FuncType F) : Name(N), Fn(F) {}
181     StringRef Name; // For debugging.
182     FuncType Fn;
183   };
184 
addRule__anona63979960111::Simplifier185   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 
push_back__anona63979960111::Simplifier::WorkListType193     void push_back(Value *V) {
194       // Do not push back duplicates.
195       if (S.insert(V).second)
196         Q.push_back(V);
197     }
198 
pop_front_val__anona63979960111::Simplifier::WorkListType199     Value *pop_front_val() {
200       Value *V = Q.front();
201       Q.pop_front();
202       S.erase(V);
203       return V;
204     }
205 
empty__anona63979960111::Simplifier::WorkListType206     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 
Context__anona63979960111::Simplifier::Context226     Context(Instruction *Exp)
227         : Ctx(Exp->getParent()->getParent()->getContext()) {
228       initialize(Exp);
229     }
230 
~Context__anona63979960111::Simplifier::Context231     ~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 {
PE__anona63979960111::PE258     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
operator <<(raw_ostream & OS,const PE & P)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)
INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)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 
print(raw_ostream & OS,const Value * V) const302 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 
initialize(Instruction * Exp)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 
record(Value * V)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 
use(Value * V)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 
unuse(Value * V)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 
subst(Value * Tree,Value * OldV,Value * NewV)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 
replace(Value * OldV,Value * NewV)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 
cleanup()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 
equal(const Instruction * I,const Instruction * J) const470 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 
find(Value * Tree,Value * Sub) const494 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 
link(Instruction * I,BasicBlock * B,BasicBlock::iterator At)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 
materialize(BasicBlock * B,BasicBlock::iterator At)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 
simplify(Context & C)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:
PolynomialMultiplyRecognize(Loop * loop,const DataLayout & dl,const DominatorTree & dt,const TargetLibraryInfo & tli,ScalarEvolution & se)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 
getPmpyType() const586     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 
getCountIV(BasicBlock * BB)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 
replaceAllUsesOfWithIn(Value * I,Value * J,BasicBlock * BB)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 
matchLeftShift(SelectInst * SelI,Value * CIV,ParsedValues & PV)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_c_Xor(m_Specific(SelI), m_Value(R))))
774       return false;
775     // Matched: xor (select +++ ? 0 : T), R
776     //          xor (select +++ ? T : 0), R
777   } else
778     return false;
779 
780   // The xor input value T is isolated into its own match so that it could
781   // be checked against an induction variable containing a shifted bit
782   // (todo).
783   // For now, check against (Q << i).
784   if (!match(T, m_Shl(m_Value(Q), m_Specific(CIV))) &&
785       !match(T, m_Shl(m_ZExt(m_Value(Q)), m_ZExt(m_Specific(CIV)))))
786     return false;
787   // Matched: select +++ ? R : R ^ (Q << i)
788   //          select +++ ? R ^ (Q << i) : R
789 
790   PV.X = X;
791   PV.Q = Q;
792   PV.R = R;
793   PV.Left = true;
794   return true;
795 }
796 
matchRightShift(SelectInst * SelI,ParsedValues & PV)797 bool PolynomialMultiplyRecognize::matchRightShift(SelectInst *SelI,
798       ParsedValues &PV) {
799   // Match the following:
800   //   select (X & 1) != 0 ? (R >> 1) ^ Q : (R >> 1)
801   //   select (X & 1) == 0 ? (R >> 1) : (R >> 1) ^ Q
802   // The condition may also check for equality with the masked value, i.e
803   //   select (X & 1) == 1 ? (R >> 1) ^ Q : (R >> 1)
804   //   select (X & 1) != 1 ? (R >> 1) : (R >> 1) ^ Q
805 
806   Value *CondV = SelI->getCondition();
807   Value *TrueV = SelI->getTrueValue();
808   Value *FalseV = SelI->getFalseValue();
809 
810   using namespace PatternMatch;
811 
812   Value *C = nullptr;
813   CmpInst::Predicate P;
814   bool TrueIfZero;
815 
816   if (match(CondV, m_c_ICmp(P, m_Value(C), m_Zero()))) {
817     if (P != CmpInst::ICMP_EQ && P != CmpInst::ICMP_NE)
818       return false;
819     // Matched: select C == 0 ? ... : ...
820     //          select C != 0 ? ... : ...
821     TrueIfZero = (P == CmpInst::ICMP_EQ);
822   } else if (match(CondV, m_c_ICmp(P, m_Value(C), m_One()))) {
823     if (P != CmpInst::ICMP_EQ && P != CmpInst::ICMP_NE)
824       return false;
825     // Matched: select C == 1 ? ... : ...
826     //          select C != 1 ? ... : ...
827     TrueIfZero = (P == CmpInst::ICMP_NE);
828   } else
829     return false;
830 
831   Value *X = nullptr;
832   if (!match(C, m_And(m_Value(X), m_One())))
833     return false;
834   // Matched: select (X & 1) == +++ ? ... : ...
835   //          select (X & 1) != +++ ? ... : ...
836 
837   Value *R = nullptr, *Q = nullptr;
838   if (TrueIfZero) {
839     // The select's condition is true if the tested bit is 0.
840     // TrueV must be the shift, FalseV must be the xor.
841     if (!match(TrueV, m_LShr(m_Value(R), m_One())))
842       return false;
843     // Matched: select +++ ? (R >> 1) : ...
844     if (!match(FalseV, m_c_Xor(m_Specific(TrueV), m_Value(Q))))
845       return false;
846     // Matched: select +++ ? (R >> 1) : (R >> 1) ^ Q
847     // with commuting ^.
848   } else {
849     // The select's condition is true if the tested bit is 1.
850     // TrueV must be the xor, FalseV must be the shift.
851     if (!match(FalseV, m_LShr(m_Value(R), m_One())))
852       return false;
853     // Matched: select +++ ? ... : (R >> 1)
854     if (!match(TrueV, m_c_Xor(m_Specific(FalseV), m_Value(Q))))
855       return false;
856     // Matched: select +++ ? (R >> 1) ^ Q : (R >> 1)
857     // with commuting ^.
858   }
859 
860   PV.X = X;
861   PV.Q = Q;
862   PV.R = R;
863   PV.Left = false;
864   return true;
865 }
866 
scanSelect(SelectInst * SelI,BasicBlock * LoopB,BasicBlock * PrehB,Value * CIV,ParsedValues & PV,bool PreScan)867 bool PolynomialMultiplyRecognize::scanSelect(SelectInst *SelI,
868       BasicBlock *LoopB, BasicBlock *PrehB, Value *CIV, ParsedValues &PV,
869       bool PreScan) {
870   using namespace PatternMatch;
871 
872   // The basic pattern for R = P.Q is:
873   // for i = 0..31
874   //   R = phi (0, R')
875   //   if (P & (1 << i))        ; test-bit(P, i)
876   //     R' = R ^ (Q << i)
877   //
878   // Similarly, the basic pattern for R = (P/Q).Q - P
879   // for i = 0..31
880   //   R = phi(P, R')
881   //   if (R & (1 << i))
882   //     R' = R ^ (Q << i)
883 
884   // There exist idioms, where instead of Q being shifted left, P is shifted
885   // right. This produces a result that is shifted right by 32 bits (the
886   // non-shifted result is 64-bit).
887   //
888   // For R = P.Q, this would be:
889   // for i = 0..31
890   //   R = phi (0, R')
891   //   if ((P >> i) & 1)
892   //     R' = (R >> 1) ^ Q      ; R is cycled through the loop, so it must
893   //   else                     ; be shifted by 1, not i.
894   //     R' = R >> 1
895   //
896   // And for the inverse:
897   // for i = 0..31
898   //   R = phi (P, R')
899   //   if (R & 1)
900   //     R' = (R >> 1) ^ Q
901   //   else
902   //     R' = R >> 1
903 
904   // The left-shifting idioms share the same pattern:
905   //   select (X & (1 << i)) ? R ^ (Q << i) : R
906   // Similarly for right-shifting idioms:
907   //   select (X & 1) ? (R >> 1) ^ Q
908 
909   if (matchLeftShift(SelI, CIV, PV)) {
910     // If this is a pre-scan, getting this far is sufficient.
911     if (PreScan)
912       return true;
913 
914     // Need to make sure that the SelI goes back into R.
915     auto *RPhi = dyn_cast<PHINode>(PV.R);
916     if (!RPhi)
917       return false;
918     if (SelI != RPhi->getIncomingValueForBlock(LoopB))
919       return false;
920     PV.Res = SelI;
921 
922     // If X is loop invariant, it must be the input polynomial, and the
923     // idiom is the basic polynomial multiply.
924     if (CurLoop->isLoopInvariant(PV.X)) {
925       PV.P = PV.X;
926       PV.Inv = false;
927     } else {
928       // X is not loop invariant. If X == R, this is the inverse pmpy.
929       // Otherwise, check for an xor with an invariant value. If the
930       // variable argument to the xor is R, then this is still a valid
931       // inverse pmpy.
932       PV.Inv = true;
933       if (PV.X != PV.R) {
934         Value *Var = nullptr, *Inv = nullptr, *X1 = nullptr, *X2 = nullptr;
935         if (!match(PV.X, m_Xor(m_Value(X1), m_Value(X2))))
936           return false;
937         auto *I1 = dyn_cast<Instruction>(X1);
938         auto *I2 = dyn_cast<Instruction>(X2);
939         if (!I1 || I1->getParent() != LoopB) {
940           Var = X2;
941           Inv = X1;
942         } else if (!I2 || I2->getParent() != LoopB) {
943           Var = X1;
944           Inv = X2;
945         } else
946           return false;
947         if (Var != PV.R)
948           return false;
949         PV.M = Inv;
950       }
951       // The input polynomial P still needs to be determined. It will be
952       // the entry value of R.
953       Value *EntryP = RPhi->getIncomingValueForBlock(PrehB);
954       PV.P = EntryP;
955     }
956 
957     return true;
958   }
959 
960   if (matchRightShift(SelI, PV)) {
961     // If this is an inverse pattern, the Q polynomial must be known at
962     // compile time.
963     if (PV.Inv && !isa<ConstantInt>(PV.Q))
964       return false;
965     if (PreScan)
966       return true;
967     // There is no exact matching of right-shift pmpy.
968     return false;
969   }
970 
971   return false;
972 }
973 
isPromotableTo(Value * Val,IntegerType * DestTy)974 bool PolynomialMultiplyRecognize::isPromotableTo(Value *Val,
975       IntegerType *DestTy) {
976   IntegerType *T = dyn_cast<IntegerType>(Val->getType());
977   if (!T || T->getBitWidth() > DestTy->getBitWidth())
978     return false;
979   if (T->getBitWidth() == DestTy->getBitWidth())
980     return true;
981   // Non-instructions are promotable. The reason why an instruction may not
982   // be promotable is that it may produce a different result if its operands
983   // and the result are promoted, for example, it may produce more non-zero
984   // bits. While it would still be possible to represent the proper result
985   // in a wider type, it may require adding additional instructions (which
986   // we don't want to do).
987   Instruction *In = dyn_cast<Instruction>(Val);
988   if (!In)
989     return true;
990   // The bitwidth of the source type is smaller than the destination.
991   // Check if the individual operation can be promoted.
992   switch (In->getOpcode()) {
993     case Instruction::PHI:
994     case Instruction::ZExt:
995     case Instruction::And:
996     case Instruction::Or:
997     case Instruction::Xor:
998     case Instruction::LShr: // Shift right is ok.
999     case Instruction::Select:
1000     case Instruction::Trunc:
1001       return true;
1002     case Instruction::ICmp:
1003       if (CmpInst *CI = cast<CmpInst>(In))
1004         return CI->isEquality() || CI->isUnsigned();
1005       llvm_unreachable("Cast failed unexpectedly");
1006     case Instruction::Add:
1007       return In->hasNoSignedWrap() && In->hasNoUnsignedWrap();
1008   }
1009   return false;
1010 }
1011 
promoteTo(Instruction * In,IntegerType * DestTy,BasicBlock * LoopB)1012 void PolynomialMultiplyRecognize::promoteTo(Instruction *In,
1013       IntegerType *DestTy, BasicBlock *LoopB) {
1014   Type *OrigTy = In->getType();
1015   assert(!OrigTy->isVoidTy() && "Invalid instruction to promote");
1016 
1017   // Leave boolean values alone.
1018   if (!In->getType()->isIntegerTy(1))
1019     In->mutateType(DestTy);
1020   unsigned DestBW = DestTy->getBitWidth();
1021 
1022   // Handle PHIs.
1023   if (PHINode *P = dyn_cast<PHINode>(In)) {
1024     unsigned N = P->getNumIncomingValues();
1025     for (unsigned i = 0; i != N; ++i) {
1026       BasicBlock *InB = P->getIncomingBlock(i);
1027       if (InB == LoopB)
1028         continue;
1029       Value *InV = P->getIncomingValue(i);
1030       IntegerType *Ty = cast<IntegerType>(InV->getType());
1031       // Do not promote values in PHI nodes of type i1.
1032       if (Ty != P->getType()) {
1033         // If the value type does not match the PHI type, the PHI type
1034         // must have been promoted.
1035         assert(Ty->getBitWidth() < DestBW);
1036         InV = IRBuilder<>(InB->getTerminator()).CreateZExt(InV, DestTy);
1037         P->setIncomingValue(i, InV);
1038       }
1039     }
1040   } else if (ZExtInst *Z = dyn_cast<ZExtInst>(In)) {
1041     Value *Op = Z->getOperand(0);
1042     if (Op->getType() == Z->getType())
1043       Z->replaceAllUsesWith(Op);
1044     Z->eraseFromParent();
1045     return;
1046   }
1047   if (TruncInst *T = dyn_cast<TruncInst>(In)) {
1048     IntegerType *TruncTy = cast<IntegerType>(OrigTy);
1049     Value *Mask = ConstantInt::get(DestTy, (1u << TruncTy->getBitWidth()) - 1);
1050     Value *And = IRBuilder<>(In).CreateAnd(T->getOperand(0), Mask);
1051     T->replaceAllUsesWith(And);
1052     T->eraseFromParent();
1053     return;
1054   }
1055 
1056   // Promote immediates.
1057   for (unsigned i = 0, n = In->getNumOperands(); i != n; ++i) {
1058     if (ConstantInt *CI = dyn_cast<ConstantInt>(In->getOperand(i)))
1059       if (CI->getBitWidth() < DestBW)
1060         In->setOperand(i, ConstantInt::get(DestTy, CI->getZExtValue()));
1061   }
1062 }
1063 
promoteTypes(BasicBlock * LoopB,BasicBlock * ExitB)1064 bool PolynomialMultiplyRecognize::promoteTypes(BasicBlock *LoopB,
1065       BasicBlock *ExitB) {
1066   assert(LoopB);
1067   // Skip loops where the exit block has more than one predecessor. The values
1068   // coming from the loop block will be promoted to another type, and so the
1069   // values coming into the exit block from other predecessors would also have
1070   // to be promoted.
1071   if (!ExitB || (ExitB->getSinglePredecessor() != LoopB))
1072     return false;
1073   IntegerType *DestTy = getPmpyType();
1074   // Check if the exit values have types that are no wider than the type
1075   // that we want to promote to.
1076   unsigned DestBW = DestTy->getBitWidth();
1077   for (PHINode &P : ExitB->phis()) {
1078     if (P.getNumIncomingValues() != 1)
1079       return false;
1080     assert(P.getIncomingBlock(0) == LoopB);
1081     IntegerType *T = dyn_cast<IntegerType>(P.getType());
1082     if (!T || T->getBitWidth() > DestBW)
1083       return false;
1084   }
1085 
1086   // Check all instructions in the loop.
1087   for (Instruction &In : *LoopB)
1088     if (!In.isTerminator() && !isPromotableTo(&In, DestTy))
1089       return false;
1090 
1091   // Perform the promotion.
1092   std::vector<Instruction*> LoopIns;
1093   std::transform(LoopB->begin(), LoopB->end(), std::back_inserter(LoopIns),
1094                  [](Instruction &In) { return &In; });
1095   for (Instruction *In : LoopIns)
1096     if (!In->isTerminator())
1097       promoteTo(In, DestTy, LoopB);
1098 
1099   // Fix up the PHI nodes in the exit block.
1100   Instruction *EndI = ExitB->getFirstNonPHI();
1101   BasicBlock::iterator End = EndI ? EndI->getIterator() : ExitB->end();
1102   for (auto I = ExitB->begin(); I != End; ++I) {
1103     PHINode *P = dyn_cast<PHINode>(I);
1104     if (!P)
1105       break;
1106     Type *Ty0 = P->getIncomingValue(0)->getType();
1107     Type *PTy = P->getType();
1108     if (PTy != Ty0) {
1109       assert(Ty0 == DestTy);
1110       // In order to create the trunc, P must have the promoted type.
1111       P->mutateType(Ty0);
1112       Value *T = IRBuilder<>(ExitB, End).CreateTrunc(P, PTy);
1113       // In order for the RAUW to work, the types of P and T must match.
1114       P->mutateType(PTy);
1115       P->replaceAllUsesWith(T);
1116       // Final update of the P's type.
1117       P->mutateType(Ty0);
1118       cast<Instruction>(T)->setOperand(0, P);
1119     }
1120   }
1121 
1122   return true;
1123 }
1124 
findCycle(Value * Out,Value * In,ValueSeq & Cycle)1125 bool PolynomialMultiplyRecognize::findCycle(Value *Out, Value *In,
1126       ValueSeq &Cycle) {
1127   // Out = ..., In, ...
1128   if (Out == In)
1129     return true;
1130 
1131   auto *BB = cast<Instruction>(Out)->getParent();
1132   bool HadPhi = false;
1133 
1134   for (auto *U : Out->users()) {
1135     auto *I = dyn_cast<Instruction>(&*U);
1136     if (I == nullptr || I->getParent() != BB)
1137       continue;
1138     // Make sure that there are no multi-iteration cycles, e.g.
1139     //   p1 = phi(p2)
1140     //   p2 = phi(p1)
1141     // The cycle p1->p2->p1 would span two loop iterations.
1142     // Check that there is only one phi in the cycle.
1143     bool IsPhi = isa<PHINode>(I);
1144     if (IsPhi && HadPhi)
1145       return false;
1146     HadPhi |= IsPhi;
1147     if (!Cycle.insert(I))
1148       return false;
1149     if (findCycle(I, In, Cycle))
1150       break;
1151     Cycle.remove(I);
1152   }
1153   return !Cycle.empty();
1154 }
1155 
classifyCycle(Instruction * DivI,ValueSeq & Cycle,ValueSeq & Early,ValueSeq & Late)1156 void PolynomialMultiplyRecognize::classifyCycle(Instruction *DivI,
1157       ValueSeq &Cycle, ValueSeq &Early, ValueSeq &Late) {
1158   // All the values in the cycle that are between the phi node and the
1159   // divider instruction will be classified as "early", all other values
1160   // will be "late".
1161 
1162   bool IsE = true;
1163   unsigned I, N = Cycle.size();
1164   for (I = 0; I < N; ++I) {
1165     Value *V = Cycle[I];
1166     if (DivI == V)
1167       IsE = false;
1168     else if (!isa<PHINode>(V))
1169       continue;
1170     // Stop if found either.
1171     break;
1172   }
1173   // "I" is the index of either DivI or the phi node, whichever was first.
1174   // "E" is "false" or "true" respectively.
1175   ValueSeq &First = !IsE ? Early : Late;
1176   for (unsigned J = 0; J < I; ++J)
1177     First.insert(Cycle[J]);
1178 
1179   ValueSeq &Second = IsE ? Early : Late;
1180   Second.insert(Cycle[I]);
1181   for (++I; I < N; ++I) {
1182     Value *V = Cycle[I];
1183     if (DivI == V || isa<PHINode>(V))
1184       break;
1185     Second.insert(V);
1186   }
1187 
1188   for (; I < N; ++I)
1189     First.insert(Cycle[I]);
1190 }
1191 
classifyInst(Instruction * UseI,ValueSeq & Early,ValueSeq & Late)1192 bool PolynomialMultiplyRecognize::classifyInst(Instruction *UseI,
1193       ValueSeq &Early, ValueSeq &Late) {
1194   // Select is an exception, since the condition value does not have to be
1195   // classified in the same way as the true/false values. The true/false
1196   // values do have to be both early or both late.
1197   if (UseI->getOpcode() == Instruction::Select) {
1198     Value *TV = UseI->getOperand(1), *FV = UseI->getOperand(2);
1199     if (Early.count(TV) || Early.count(FV)) {
1200       if (Late.count(TV) || Late.count(FV))
1201         return false;
1202       Early.insert(UseI);
1203     } else if (Late.count(TV) || Late.count(FV)) {
1204       if (Early.count(TV) || Early.count(FV))
1205         return false;
1206       Late.insert(UseI);
1207     }
1208     return true;
1209   }
1210 
1211   // Not sure what would be the example of this, but the code below relies
1212   // on having at least one operand.
1213   if (UseI->getNumOperands() == 0)
1214     return true;
1215 
1216   bool AE = true, AL = true;
1217   for (auto &I : UseI->operands()) {
1218     if (Early.count(&*I))
1219       AL = false;
1220     else if (Late.count(&*I))
1221       AE = false;
1222   }
1223   // If the operands appear "all early" and "all late" at the same time,
1224   // then it means that none of them are actually classified as either.
1225   // This is harmless.
1226   if (AE && AL)
1227     return true;
1228   // Conversely, if they are neither "all early" nor "all late", then
1229   // we have a mixture of early and late operands that is not a known
1230   // exception.
1231   if (!AE && !AL)
1232     return false;
1233 
1234   // Check that we have covered the two special cases.
1235   assert(AE != AL);
1236 
1237   if (AE)
1238     Early.insert(UseI);
1239   else
1240     Late.insert(UseI);
1241   return true;
1242 }
1243 
commutesWithShift(Instruction * I)1244 bool PolynomialMultiplyRecognize::commutesWithShift(Instruction *I) {
1245   switch (I->getOpcode()) {
1246     case Instruction::And:
1247     case Instruction::Or:
1248     case Instruction::Xor:
1249     case Instruction::LShr:
1250     case Instruction::Shl:
1251     case Instruction::Select:
1252     case Instruction::ICmp:
1253     case Instruction::PHI:
1254       break;
1255     default:
1256       return false;
1257   }
1258   return true;
1259 }
1260 
highBitsAreZero(Value * V,unsigned IterCount)1261 bool PolynomialMultiplyRecognize::highBitsAreZero(Value *V,
1262       unsigned IterCount) {
1263   auto *T = dyn_cast<IntegerType>(V->getType());
1264   if (!T)
1265     return false;
1266 
1267   KnownBits Known(T->getBitWidth());
1268   computeKnownBits(V, Known, DL);
1269   return Known.countMinLeadingZeros() >= IterCount;
1270 }
1271 
keepsHighBitsZero(Value * V,unsigned IterCount)1272 bool PolynomialMultiplyRecognize::keepsHighBitsZero(Value *V,
1273       unsigned IterCount) {
1274   // Assume that all inputs to the value have the high bits zero.
1275   // Check if the value itself preserves the zeros in the high bits.
1276   if (auto *C = dyn_cast<ConstantInt>(V))
1277     return C->getValue().countl_zero() >= IterCount;
1278 
1279   if (auto *I = dyn_cast<Instruction>(V)) {
1280     switch (I->getOpcode()) {
1281       case Instruction::And:
1282       case Instruction::Or:
1283       case Instruction::Xor:
1284       case Instruction::LShr:
1285       case Instruction::Select:
1286       case Instruction::ICmp:
1287       case Instruction::PHI:
1288       case Instruction::ZExt:
1289         return true;
1290     }
1291   }
1292 
1293   return false;
1294 }
1295 
isOperandShifted(Instruction * I,Value * Op)1296 bool PolynomialMultiplyRecognize::isOperandShifted(Instruction *I, Value *Op) {
1297   unsigned Opc = I->getOpcode();
1298   if (Opc == Instruction::Shl || Opc == Instruction::LShr)
1299     return Op != I->getOperand(1);
1300   return true;
1301 }
1302 
convertShiftsToLeft(BasicBlock * LoopB,BasicBlock * ExitB,unsigned IterCount)1303 bool PolynomialMultiplyRecognize::convertShiftsToLeft(BasicBlock *LoopB,
1304       BasicBlock *ExitB, unsigned IterCount) {
1305   Value *CIV = getCountIV(LoopB);
1306   if (CIV == nullptr)
1307     return false;
1308   auto *CIVTy = dyn_cast<IntegerType>(CIV->getType());
1309   if (CIVTy == nullptr)
1310     return false;
1311 
1312   ValueSeq RShifts;
1313   ValueSeq Early, Late, Cycled;
1314 
1315   // Find all value cycles that contain logical right shifts by 1.
1316   for (Instruction &I : *LoopB) {
1317     using namespace PatternMatch;
1318 
1319     Value *V = nullptr;
1320     if (!match(&I, m_LShr(m_Value(V), m_One())))
1321       continue;
1322     ValueSeq C;
1323     if (!findCycle(&I, V, C))
1324       continue;
1325 
1326     // Found a cycle.
1327     C.insert(&I);
1328     classifyCycle(&I, C, Early, Late);
1329     Cycled.insert(C.begin(), C.end());
1330     RShifts.insert(&I);
1331   }
1332 
1333   // Find the set of all values affected by the shift cycles, i.e. all
1334   // cycled values, and (recursively) all their users.
1335   ValueSeq Users(Cycled.begin(), Cycled.end());
1336   for (unsigned i = 0; i < Users.size(); ++i) {
1337     Value *V = Users[i];
1338     if (!isa<IntegerType>(V->getType()))
1339       return false;
1340     auto *R = cast<Instruction>(V);
1341     // If the instruction does not commute with shifts, the loop cannot
1342     // be unshifted.
1343     if (!commutesWithShift(R))
1344       return false;
1345     for (User *U : R->users()) {
1346       auto *T = cast<Instruction>(U);
1347       // Skip users from outside of the loop. They will be handled later.
1348       // Also, skip the right-shifts and phi nodes, since they mix early
1349       // and late values.
1350       if (T->getParent() != LoopB || RShifts.count(T) || isa<PHINode>(T))
1351         continue;
1352 
1353       Users.insert(T);
1354       if (!classifyInst(T, Early, Late))
1355         return false;
1356     }
1357   }
1358 
1359   if (Users.empty())
1360     return false;
1361 
1362   // Verify that high bits remain zero.
1363   ValueSeq Internal(Users.begin(), Users.end());
1364   ValueSeq Inputs;
1365   for (unsigned i = 0; i < Internal.size(); ++i) {
1366     auto *R = dyn_cast<Instruction>(Internal[i]);
1367     if (!R)
1368       continue;
1369     for (Value *Op : R->operands()) {
1370       auto *T = dyn_cast<Instruction>(Op);
1371       if (T && T->getParent() != LoopB)
1372         Inputs.insert(Op);
1373       else
1374         Internal.insert(Op);
1375     }
1376   }
1377   for (Value *V : Inputs)
1378     if (!highBitsAreZero(V, IterCount))
1379       return false;
1380   for (Value *V : Internal)
1381     if (!keepsHighBitsZero(V, IterCount))
1382       return false;
1383 
1384   // Finally, the work can be done. Unshift each user.
1385   IRBuilder<> IRB(LoopB);
1386   std::map<Value*,Value*> ShiftMap;
1387 
1388   using CastMapType = std::map<std::pair<Value *, Type *>, Value *>;
1389 
1390   CastMapType CastMap;
1391 
1392   auto upcast = [] (CastMapType &CM, IRBuilder<> &IRB, Value *V,
1393         IntegerType *Ty) -> Value* {
1394     auto H = CM.find(std::make_pair(V, Ty));
1395     if (H != CM.end())
1396       return H->second;
1397     Value *CV = IRB.CreateIntCast(V, Ty, false);
1398     CM.insert(std::make_pair(std::make_pair(V, Ty), CV));
1399     return CV;
1400   };
1401 
1402   for (auto I = LoopB->begin(), E = LoopB->end(); I != E; ++I) {
1403     using namespace PatternMatch;
1404 
1405     if (isa<PHINode>(I) || !Users.count(&*I))
1406       continue;
1407 
1408     // Match lshr x, 1.
1409     Value *V = nullptr;
1410     if (match(&*I, m_LShr(m_Value(V), m_One()))) {
1411       replaceAllUsesOfWithIn(&*I, V, LoopB);
1412       continue;
1413     }
1414     // For each non-cycled operand, replace it with the corresponding
1415     // value shifted left.
1416     for (auto &J : I->operands()) {
1417       Value *Op = J.get();
1418       if (!isOperandShifted(&*I, Op))
1419         continue;
1420       if (Users.count(Op))
1421         continue;
1422       // Skip shifting zeros.
1423       if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
1424         continue;
1425       // Check if we have already generated a shift for this value.
1426       auto F = ShiftMap.find(Op);
1427       Value *W = (F != ShiftMap.end()) ? F->second : nullptr;
1428       if (W == nullptr) {
1429         IRB.SetInsertPoint(&*I);
1430         // First, the shift amount will be CIV or CIV+1, depending on
1431         // whether the value is early or late. Instead of creating CIV+1,
1432         // do a single shift of the value.
1433         Value *ShAmt = CIV, *ShVal = Op;
1434         auto *VTy = cast<IntegerType>(ShVal->getType());
1435         auto *ATy = cast<IntegerType>(ShAmt->getType());
1436         if (Late.count(&*I))
1437           ShVal = IRB.CreateShl(Op, ConstantInt::get(VTy, 1));
1438         // Second, the types of the shifted value and the shift amount
1439         // must match.
1440         if (VTy != ATy) {
1441           if (VTy->getBitWidth() < ATy->getBitWidth())
1442             ShVal = upcast(CastMap, IRB, ShVal, ATy);
1443           else
1444             ShAmt = upcast(CastMap, IRB, ShAmt, VTy);
1445         }
1446         // Ready to generate the shift and memoize it.
1447         W = IRB.CreateShl(ShVal, ShAmt);
1448         ShiftMap.insert(std::make_pair(Op, W));
1449       }
1450       I->replaceUsesOfWith(Op, W);
1451     }
1452   }
1453 
1454   // Update the users outside of the loop to account for having left
1455   // shifts. They would normally be shifted right in the loop, so shift
1456   // them right after the loop exit.
1457   // Take advantage of the loop-closed SSA form, which has all the post-
1458   // loop values in phi nodes.
1459   IRB.SetInsertPoint(ExitB, ExitB->getFirstInsertionPt());
1460   for (auto P = ExitB->begin(), Q = ExitB->end(); P != Q; ++P) {
1461     if (!isa<PHINode>(P))
1462       break;
1463     auto *PN = cast<PHINode>(P);
1464     Value *U = PN->getIncomingValueForBlock(LoopB);
1465     if (!Users.count(U))
1466       continue;
1467     Value *S = IRB.CreateLShr(PN, ConstantInt::get(PN->getType(), IterCount));
1468     PN->replaceAllUsesWith(S);
1469     // The above RAUW will create
1470     //   S = lshr S, IterCount
1471     // so we need to fix it back into
1472     //   S = lshr PN, IterCount
1473     cast<User>(S)->replaceUsesOfWith(S, PN);
1474   }
1475 
1476   return true;
1477 }
1478 
cleanupLoopBody(BasicBlock * LoopB)1479 void PolynomialMultiplyRecognize::cleanupLoopBody(BasicBlock *LoopB) {
1480   for (auto &I : *LoopB)
1481     if (Value *SV = simplifyInstruction(&I, {DL, &TLI, &DT}))
1482       I.replaceAllUsesWith(SV);
1483 
1484   for (Instruction &I : llvm::make_early_inc_range(*LoopB))
1485     RecursivelyDeleteTriviallyDeadInstructions(&I, &TLI);
1486 }
1487 
getInverseMxN(unsigned QP)1488 unsigned PolynomialMultiplyRecognize::getInverseMxN(unsigned QP) {
1489   // Arrays of coefficients of Q and the inverse, C.
1490   // Q[i] = coefficient at x^i.
1491   std::array<char,32> Q, C;
1492 
1493   for (unsigned i = 0; i < 32; ++i) {
1494     Q[i] = QP & 1;
1495     QP >>= 1;
1496   }
1497   assert(Q[0] == 1);
1498 
1499   // Find C, such that
1500   // (Q[n]*x^n + ... + Q[1]*x + Q[0]) * (C[n]*x^n + ... + C[1]*x + C[0]) = 1
1501   //
1502   // For it to have a solution, Q[0] must be 1. Since this is Z2[x], the
1503   // operations * and + are & and ^ respectively.
1504   //
1505   // Find C[i] recursively, by comparing i-th coefficient in the product
1506   // with 0 (or 1 for i=0).
1507   //
1508   // C[0] = 1, since C[0] = Q[0], and Q[0] = 1.
1509   C[0] = 1;
1510   for (unsigned i = 1; i < 32; ++i) {
1511     // Solve for C[i] in:
1512     //   C[0]Q[i] ^ C[1]Q[i-1] ^ ... ^ C[i-1]Q[1] ^ C[i]Q[0] = 0
1513     // This is equivalent to
1514     //   C[0]Q[i] ^ C[1]Q[i-1] ^ ... ^ C[i-1]Q[1] ^ C[i] = 0
1515     // which is
1516     //   C[0]Q[i] ^ C[1]Q[i-1] ^ ... ^ C[i-1]Q[1] = C[i]
1517     unsigned T = 0;
1518     for (unsigned j = 0; j < i; ++j)
1519       T = T ^ (C[j] & Q[i-j]);
1520     C[i] = T;
1521   }
1522 
1523   unsigned QV = 0;
1524   for (unsigned i = 0; i < 32; ++i)
1525     if (C[i])
1526       QV |= (1 << i);
1527 
1528   return QV;
1529 }
1530 
generate(BasicBlock::iterator At,ParsedValues & PV)1531 Value *PolynomialMultiplyRecognize::generate(BasicBlock::iterator At,
1532       ParsedValues &PV) {
1533   IRBuilder<> B(&*At);
1534   Module *M = At->getParent()->getParent()->getParent();
1535   Function *PMF = Intrinsic::getDeclaration(M, Intrinsic::hexagon_M4_pmpyw);
1536 
1537   Value *P = PV.P, *Q = PV.Q, *P0 = P;
1538   unsigned IC = PV.IterCount;
1539 
1540   if (PV.M != nullptr)
1541     P0 = P = B.CreateXor(P, PV.M);
1542 
1543   // Create a bit mask to clear the high bits beyond IterCount.
1544   auto *BMI = ConstantInt::get(P->getType(), APInt::getLowBitsSet(32, IC));
1545 
1546   if (PV.IterCount != 32)
1547     P = B.CreateAnd(P, BMI);
1548 
1549   if (PV.Inv) {
1550     auto *QI = dyn_cast<ConstantInt>(PV.Q);
1551     assert(QI && QI->getBitWidth() <= 32);
1552 
1553     // Again, clearing bits beyond IterCount.
1554     unsigned M = (1 << PV.IterCount) - 1;
1555     unsigned Tmp = (QI->getZExtValue() | 1) & M;
1556     unsigned QV = getInverseMxN(Tmp) & M;
1557     auto *QVI = ConstantInt::get(QI->getType(), QV);
1558     P = B.CreateCall(PMF, {P, QVI});
1559     P = B.CreateTrunc(P, QI->getType());
1560     if (IC != 32)
1561       P = B.CreateAnd(P, BMI);
1562   }
1563 
1564   Value *R = B.CreateCall(PMF, {P, Q});
1565 
1566   if (PV.M != nullptr)
1567     R = B.CreateXor(R, B.CreateIntCast(P0, R->getType(), false));
1568 
1569   return R;
1570 }
1571 
hasZeroSignBit(const Value * V)1572 static bool hasZeroSignBit(const Value *V) {
1573   if (const auto *CI = dyn_cast<const ConstantInt>(V))
1574     return CI->getValue().isNonNegative();
1575   const Instruction *I = dyn_cast<const Instruction>(V);
1576   if (!I)
1577     return false;
1578   switch (I->getOpcode()) {
1579     case Instruction::LShr:
1580       if (const auto SI = dyn_cast<const ConstantInt>(I->getOperand(1)))
1581         return SI->getZExtValue() > 0;
1582       return false;
1583     case Instruction::Or:
1584     case Instruction::Xor:
1585       return hasZeroSignBit(I->getOperand(0)) &&
1586              hasZeroSignBit(I->getOperand(1));
1587     case Instruction::And:
1588       return hasZeroSignBit(I->getOperand(0)) ||
1589              hasZeroSignBit(I->getOperand(1));
1590   }
1591   return false;
1592 }
1593 
setupPreSimplifier(Simplifier & S)1594 void PolynomialMultiplyRecognize::setupPreSimplifier(Simplifier &S) {
1595   S.addRule("sink-zext",
1596     // Sink zext past bitwise operations.
1597     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1598       if (I->getOpcode() != Instruction::ZExt)
1599         return nullptr;
1600       Instruction *T = dyn_cast<Instruction>(I->getOperand(0));
1601       if (!T)
1602         return nullptr;
1603       switch (T->getOpcode()) {
1604         case Instruction::And:
1605         case Instruction::Or:
1606         case Instruction::Xor:
1607           break;
1608         default:
1609           return nullptr;
1610       }
1611       IRBuilder<> B(Ctx);
1612       return B.CreateBinOp(cast<BinaryOperator>(T)->getOpcode(),
1613                            B.CreateZExt(T->getOperand(0), I->getType()),
1614                            B.CreateZExt(T->getOperand(1), I->getType()));
1615     });
1616   S.addRule("xor/and -> and/xor",
1617     // (xor (and x a) (and y a)) -> (and (xor x y) a)
1618     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1619       if (I->getOpcode() != Instruction::Xor)
1620         return nullptr;
1621       Instruction *And0 = dyn_cast<Instruction>(I->getOperand(0));
1622       Instruction *And1 = dyn_cast<Instruction>(I->getOperand(1));
1623       if (!And0 || !And1)
1624         return nullptr;
1625       if (And0->getOpcode() != Instruction::And ||
1626           And1->getOpcode() != Instruction::And)
1627         return nullptr;
1628       if (And0->getOperand(1) != And1->getOperand(1))
1629         return nullptr;
1630       IRBuilder<> B(Ctx);
1631       return B.CreateAnd(B.CreateXor(And0->getOperand(0), And1->getOperand(0)),
1632                          And0->getOperand(1));
1633     });
1634   S.addRule("sink binop into select",
1635     // (Op (select c x y) z) -> (select c (Op x z) (Op y z))
1636     // (Op x (select c y z)) -> (select c (Op x y) (Op x z))
1637     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1638       BinaryOperator *BO = dyn_cast<BinaryOperator>(I);
1639       if (!BO)
1640         return nullptr;
1641       Instruction::BinaryOps Op = BO->getOpcode();
1642       if (SelectInst *Sel = dyn_cast<SelectInst>(BO->getOperand(0))) {
1643         IRBuilder<> B(Ctx);
1644         Value *X = Sel->getTrueValue(), *Y = Sel->getFalseValue();
1645         Value *Z = BO->getOperand(1);
1646         return B.CreateSelect(Sel->getCondition(),
1647                               B.CreateBinOp(Op, X, Z),
1648                               B.CreateBinOp(Op, Y, Z));
1649       }
1650       if (SelectInst *Sel = dyn_cast<SelectInst>(BO->getOperand(1))) {
1651         IRBuilder<> B(Ctx);
1652         Value *X = BO->getOperand(0);
1653         Value *Y = Sel->getTrueValue(), *Z = Sel->getFalseValue();
1654         return B.CreateSelect(Sel->getCondition(),
1655                               B.CreateBinOp(Op, X, Y),
1656                               B.CreateBinOp(Op, X, Z));
1657       }
1658       return nullptr;
1659     });
1660   S.addRule("fold select-select",
1661     // (select c (select c x y) z) -> (select c x z)
1662     // (select c x (select c y z)) -> (select c x z)
1663     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1664       SelectInst *Sel = dyn_cast<SelectInst>(I);
1665       if (!Sel)
1666         return nullptr;
1667       IRBuilder<> B(Ctx);
1668       Value *C = Sel->getCondition();
1669       if (SelectInst *Sel0 = dyn_cast<SelectInst>(Sel->getTrueValue())) {
1670         if (Sel0->getCondition() == C)
1671           return B.CreateSelect(C, Sel0->getTrueValue(), Sel->getFalseValue());
1672       }
1673       if (SelectInst *Sel1 = dyn_cast<SelectInst>(Sel->getFalseValue())) {
1674         if (Sel1->getCondition() == C)
1675           return B.CreateSelect(C, Sel->getTrueValue(), Sel1->getFalseValue());
1676       }
1677       return nullptr;
1678     });
1679   S.addRule("or-signbit -> xor-signbit",
1680     // (or (lshr x 1) 0x800.0) -> (xor (lshr x 1) 0x800.0)
1681     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1682       if (I->getOpcode() != Instruction::Or)
1683         return nullptr;
1684       ConstantInt *Msb = dyn_cast<ConstantInt>(I->getOperand(1));
1685       if (!Msb || !Msb->getValue().isSignMask())
1686         return nullptr;
1687       if (!hasZeroSignBit(I->getOperand(0)))
1688         return nullptr;
1689       return IRBuilder<>(Ctx).CreateXor(I->getOperand(0), Msb);
1690     });
1691   S.addRule("sink lshr into binop",
1692     // (lshr (BitOp x y) c) -> (BitOp (lshr x c) (lshr y c))
1693     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1694       if (I->getOpcode() != Instruction::LShr)
1695         return nullptr;
1696       BinaryOperator *BitOp = dyn_cast<BinaryOperator>(I->getOperand(0));
1697       if (!BitOp)
1698         return nullptr;
1699       switch (BitOp->getOpcode()) {
1700         case Instruction::And:
1701         case Instruction::Or:
1702         case Instruction::Xor:
1703           break;
1704         default:
1705           return nullptr;
1706       }
1707       IRBuilder<> B(Ctx);
1708       Value *S = I->getOperand(1);
1709       return B.CreateBinOp(BitOp->getOpcode(),
1710                 B.CreateLShr(BitOp->getOperand(0), S),
1711                 B.CreateLShr(BitOp->getOperand(1), S));
1712     });
1713   S.addRule("expose bitop-const",
1714     // (BitOp1 (BitOp2 x a) b) -> (BitOp2 x (BitOp1 a b))
1715     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1716       auto IsBitOp = [](unsigned Op) -> bool {
1717         switch (Op) {
1718           case Instruction::And:
1719           case Instruction::Or:
1720           case Instruction::Xor:
1721             return true;
1722         }
1723         return false;
1724       };
1725       BinaryOperator *BitOp1 = dyn_cast<BinaryOperator>(I);
1726       if (!BitOp1 || !IsBitOp(BitOp1->getOpcode()))
1727         return nullptr;
1728       BinaryOperator *BitOp2 = dyn_cast<BinaryOperator>(BitOp1->getOperand(0));
1729       if (!BitOp2 || !IsBitOp(BitOp2->getOpcode()))
1730         return nullptr;
1731       ConstantInt *CA = dyn_cast<ConstantInt>(BitOp2->getOperand(1));
1732       ConstantInt *CB = dyn_cast<ConstantInt>(BitOp1->getOperand(1));
1733       if (!CA || !CB)
1734         return nullptr;
1735       IRBuilder<> B(Ctx);
1736       Value *X = BitOp2->getOperand(0);
1737       return B.CreateBinOp(BitOp2->getOpcode(), X,
1738                 B.CreateBinOp(BitOp1->getOpcode(), CA, CB));
1739     });
1740 }
1741 
setupPostSimplifier(Simplifier & S)1742 void PolynomialMultiplyRecognize::setupPostSimplifier(Simplifier &S) {
1743   S.addRule("(and (xor (and x a) y) b) -> (and (xor x y) b), if b == b&a",
1744     [](Instruction *I, LLVMContext &Ctx) -> Value* {
1745       if (I->getOpcode() != Instruction::And)
1746         return nullptr;
1747       Instruction *Xor = dyn_cast<Instruction>(I->getOperand(0));
1748       ConstantInt *C0 = dyn_cast<ConstantInt>(I->getOperand(1));
1749       if (!Xor || !C0)
1750         return nullptr;
1751       if (Xor->getOpcode() != Instruction::Xor)
1752         return nullptr;
1753       Instruction *And0 = dyn_cast<Instruction>(Xor->getOperand(0));
1754       Instruction *And1 = dyn_cast<Instruction>(Xor->getOperand(1));
1755       // Pick the first non-null and.
1756       if (!And0 || And0->getOpcode() != Instruction::And)
1757         std::swap(And0, And1);
1758       ConstantInt *C1 = dyn_cast<ConstantInt>(And0->getOperand(1));
1759       if (!C1)
1760         return nullptr;
1761       uint32_t V0 = C0->getZExtValue();
1762       uint32_t V1 = C1->getZExtValue();
1763       if (V0 != (V0 & V1))
1764         return nullptr;
1765       IRBuilder<> B(Ctx);
1766       return B.CreateAnd(B.CreateXor(And0->getOperand(0), And1), C0);
1767     });
1768 }
1769 
recognize()1770 bool PolynomialMultiplyRecognize::recognize() {
1771   LLVM_DEBUG(dbgs() << "Starting PolynomialMultiplyRecognize on loop\n"
1772                     << *CurLoop << '\n');
1773   // Restrictions:
1774   // - The loop must consist of a single block.
1775   // - The iteration count must be known at compile-time.
1776   // - The loop must have an induction variable starting from 0, and
1777   //   incremented in each iteration of the loop.
1778   BasicBlock *LoopB = CurLoop->getHeader();
1779   LLVM_DEBUG(dbgs() << "Loop header:\n" << *LoopB);
1780 
1781   if (LoopB != CurLoop->getLoopLatch())
1782     return false;
1783   BasicBlock *ExitB = CurLoop->getExitBlock();
1784   if (ExitB == nullptr)
1785     return false;
1786   BasicBlock *EntryB = CurLoop->getLoopPreheader();
1787   if (EntryB == nullptr)
1788     return false;
1789 
1790   unsigned IterCount = 0;
1791   const SCEV *CT = SE.getBackedgeTakenCount(CurLoop);
1792   if (isa<SCEVCouldNotCompute>(CT))
1793     return false;
1794   if (auto *CV = dyn_cast<SCEVConstant>(CT))
1795     IterCount = CV->getValue()->getZExtValue() + 1;
1796 
1797   Value *CIV = getCountIV(LoopB);
1798   ParsedValues PV;
1799   Simplifier PreSimp;
1800   PV.IterCount = IterCount;
1801   LLVM_DEBUG(dbgs() << "Loop IV: " << *CIV << "\nIterCount: " << IterCount
1802                     << '\n');
1803 
1804   setupPreSimplifier(PreSimp);
1805 
1806   // Perform a preliminary scan of select instructions to see if any of them
1807   // looks like a generator of the polynomial multiply steps. Assume that a
1808   // loop can only contain a single transformable operation, so stop the
1809   // traversal after the first reasonable candidate was found.
1810   // XXX: Currently this approach can modify the loop before being 100% sure
1811   // that the transformation can be carried out.
1812   bool FoundPreScan = false;
1813   auto FeedsPHI = [LoopB](const Value *V) -> bool {
1814     for (const Value *U : V->users()) {
1815       if (const auto *P = dyn_cast<const PHINode>(U))
1816         if (P->getParent() == LoopB)
1817           return true;
1818     }
1819     return false;
1820   };
1821   for (Instruction &In : *LoopB) {
1822     SelectInst *SI = dyn_cast<SelectInst>(&In);
1823     if (!SI || !FeedsPHI(SI))
1824       continue;
1825 
1826     Simplifier::Context C(SI);
1827     Value *T = PreSimp.simplify(C);
1828     SelectInst *SelI = (T && isa<SelectInst>(T)) ? cast<SelectInst>(T) : SI;
1829     LLVM_DEBUG(dbgs() << "scanSelect(pre-scan): " << PE(C, SelI) << '\n');
1830     if (scanSelect(SelI, LoopB, EntryB, CIV, PV, true)) {
1831       FoundPreScan = true;
1832       if (SelI != SI) {
1833         Value *NewSel = C.materialize(LoopB, SI->getIterator());
1834         SI->replaceAllUsesWith(NewSel);
1835         RecursivelyDeleteTriviallyDeadInstructions(SI, &TLI);
1836       }
1837       break;
1838     }
1839   }
1840 
1841   if (!FoundPreScan) {
1842     LLVM_DEBUG(dbgs() << "Have not found candidates for pmpy\n");
1843     return false;
1844   }
1845 
1846   if (!PV.Left) {
1847     // The right shift version actually only returns the higher bits of
1848     // the result (each iteration discards the LSB). If we want to convert it
1849     // to a left-shifting loop, the working data type must be at least as
1850     // wide as the target's pmpy instruction.
1851     if (!promoteTypes(LoopB, ExitB))
1852       return false;
1853     // Run post-promotion simplifications.
1854     Simplifier PostSimp;
1855     setupPostSimplifier(PostSimp);
1856     for (Instruction &In : *LoopB) {
1857       SelectInst *SI = dyn_cast<SelectInst>(&In);
1858       if (!SI || !FeedsPHI(SI))
1859         continue;
1860       Simplifier::Context C(SI);
1861       Value *T = PostSimp.simplify(C);
1862       SelectInst *SelI = dyn_cast_or_null<SelectInst>(T);
1863       if (SelI != SI) {
1864         Value *NewSel = C.materialize(LoopB, SI->getIterator());
1865         SI->replaceAllUsesWith(NewSel);
1866         RecursivelyDeleteTriviallyDeadInstructions(SI, &TLI);
1867       }
1868       break;
1869     }
1870 
1871     if (!convertShiftsToLeft(LoopB, ExitB, IterCount))
1872       return false;
1873     cleanupLoopBody(LoopB);
1874   }
1875 
1876   // Scan the loop again, find the generating select instruction.
1877   bool FoundScan = false;
1878   for (Instruction &In : *LoopB) {
1879     SelectInst *SelI = dyn_cast<SelectInst>(&In);
1880     if (!SelI)
1881       continue;
1882     LLVM_DEBUG(dbgs() << "scanSelect: " << *SelI << '\n');
1883     FoundScan = scanSelect(SelI, LoopB, EntryB, CIV, PV, false);
1884     if (FoundScan)
1885       break;
1886   }
1887   assert(FoundScan);
1888 
1889   LLVM_DEBUG({
1890     StringRef PP = (PV.M ? "(P+M)" : "P");
1891     if (!PV.Inv)
1892       dbgs() << "Found pmpy idiom: R = " << PP << ".Q\n";
1893     else
1894       dbgs() << "Found inverse pmpy idiom: R = (" << PP << "/Q).Q) + "
1895              << PP << "\n";
1896     dbgs() << "  Res:" << *PV.Res << "\n  P:" << *PV.P << "\n";
1897     if (PV.M)
1898       dbgs() << "  M:" << *PV.M << "\n";
1899     dbgs() << "  Q:" << *PV.Q << "\n";
1900     dbgs() << "  Iteration count:" << PV.IterCount << "\n";
1901   });
1902 
1903   BasicBlock::iterator At(EntryB->getTerminator());
1904   Value *PM = generate(At, PV);
1905   if (PM == nullptr)
1906     return false;
1907 
1908   if (PM->getType() != PV.Res->getType())
1909     PM = IRBuilder<>(&*At).CreateIntCast(PM, PV.Res->getType(), false);
1910 
1911   PV.Res->replaceAllUsesWith(PM);
1912   PV.Res->eraseFromParent();
1913   return true;
1914 }
1915 
getSCEVStride(const SCEVAddRecExpr * S)1916 int HexagonLoopIdiomRecognize::getSCEVStride(const SCEVAddRecExpr *S) {
1917   if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(1)))
1918     return SC->getAPInt().getSExtValue();
1919   return 0;
1920 }
1921 
isLegalStore(Loop * CurLoop,StoreInst * SI)1922 bool HexagonLoopIdiomRecognize::isLegalStore(Loop *CurLoop, StoreInst *SI) {
1923   // Allow volatile stores if HexagonVolatileMemcpy is enabled.
1924   if (!(SI->isVolatile() && HexagonVolatileMemcpy) && !SI->isSimple())
1925     return false;
1926 
1927   Value *StoredVal = SI->getValueOperand();
1928   Value *StorePtr = SI->getPointerOperand();
1929 
1930   // Reject stores that are so large that they overflow an unsigned.
1931   uint64_t SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
1932   if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
1933     return false;
1934 
1935   // See if the pointer expression is an AddRec like {base,+,1} on the current
1936   // loop, which indicates a strided store.  If we have something else, it's a
1937   // random store we can't handle.
1938   auto *StoreEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
1939   if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
1940     return false;
1941 
1942   // Check to see if the stride matches the size of the store.  If so, then we
1943   // know that every byte is touched in the loop.
1944   int Stride = getSCEVStride(StoreEv);
1945   if (Stride == 0)
1946     return false;
1947   unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType());
1948   if (StoreSize != unsigned(std::abs(Stride)))
1949     return false;
1950 
1951   // The store must be feeding a non-volatile load.
1952   LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand());
1953   if (!LI || !LI->isSimple())
1954     return false;
1955 
1956   // See if the pointer expression is an AddRec like {base,+,1} on the current
1957   // loop, which indicates a strided load.  If we have something else, it's a
1958   // random load we can't handle.
1959   Value *LoadPtr = LI->getPointerOperand();
1960   auto *LoadEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LoadPtr));
1961   if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
1962     return false;
1963 
1964   // The store and load must share the same stride.
1965   if (StoreEv->getOperand(1) != LoadEv->getOperand(1))
1966     return false;
1967 
1968   // Success.  This store can be converted into a memcpy.
1969   return true;
1970 }
1971 
1972 /// mayLoopAccessLocation - Return true if the specified loop might access the
1973 /// specified pointer location, which is a loop-strided access.  The 'Access'
1974 /// argument specifies what the verboten forms of access are (read or write).
1975 static bool
mayLoopAccessLocation(Value * Ptr,ModRefInfo Access,Loop * L,const SCEV * BECount,unsigned StoreSize,AliasAnalysis & AA,SmallPtrSetImpl<Instruction * > & Ignored)1976 mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
1977                       const SCEV *BECount, unsigned StoreSize,
1978                       AliasAnalysis &AA,
1979                       SmallPtrSetImpl<Instruction *> &Ignored) {
1980   // Get the location that may be stored across the loop.  Since the access
1981   // is strided positively through memory, we say that the modified location
1982   // starts at the pointer and has infinite size.
1983   LocationSize AccessSize = LocationSize::afterPointer();
1984 
1985   // If the loop iterates a fixed number of times, we can refine the access
1986   // size to be exactly the size of the memset, which is (BECount+1)*StoreSize
1987   if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
1988     AccessSize = LocationSize::precise((BECst->getValue()->getZExtValue() + 1) *
1989                                        StoreSize);
1990 
1991   // TODO: For this to be really effective, we have to dive into the pointer
1992   // operand in the store.  Store to &A[i] of 100 will always return may alias
1993   // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
1994   // which will then no-alias a store to &A[100].
1995   MemoryLocation StoreLoc(Ptr, AccessSize);
1996 
1997   for (auto *B : L->blocks())
1998     for (auto &I : *B)
1999       if (Ignored.count(&I) == 0 &&
2000           isModOrRefSet(AA.getModRefInfo(&I, StoreLoc) & Access))
2001         return true;
2002 
2003   return false;
2004 }
2005 
collectStores(Loop * CurLoop,BasicBlock * BB,SmallVectorImpl<StoreInst * > & Stores)2006 void HexagonLoopIdiomRecognize::collectStores(Loop *CurLoop, BasicBlock *BB,
2007       SmallVectorImpl<StoreInst*> &Stores) {
2008   Stores.clear();
2009   for (Instruction &I : *BB)
2010     if (StoreInst *SI = dyn_cast<StoreInst>(&I))
2011       if (isLegalStore(CurLoop, SI))
2012         Stores.push_back(SI);
2013 }
2014 
processCopyingStore(Loop * CurLoop,StoreInst * SI,const SCEV * BECount)2015 bool HexagonLoopIdiomRecognize::processCopyingStore(Loop *CurLoop,
2016       StoreInst *SI, const SCEV *BECount) {
2017   assert((SI->isSimple() || (SI->isVolatile() && HexagonVolatileMemcpy)) &&
2018          "Expected only non-volatile stores, or Hexagon-specific memcpy"
2019          "to volatile destination.");
2020 
2021   Value *StorePtr = SI->getPointerOperand();
2022   auto *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
2023   unsigned Stride = getSCEVStride(StoreEv);
2024   unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType());
2025   if (Stride != StoreSize)
2026     return false;
2027 
2028   // See if the pointer expression is an AddRec like {base,+,1} on the current
2029   // loop, which indicates a strided load.  If we have something else, it's a
2030   // random load we can't handle.
2031   auto *LI = cast<LoadInst>(SI->getValueOperand());
2032   auto *LoadEv = cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
2033 
2034   // The trip count of the loop and the base pointer of the addrec SCEV is
2035   // guaranteed to be loop invariant, which means that it should dominate the
2036   // header.  This allows us to insert code for it in the preheader.
2037   BasicBlock *Preheader = CurLoop->getLoopPreheader();
2038   Instruction *ExpPt = Preheader->getTerminator();
2039   IRBuilder<> Builder(ExpPt);
2040   SCEVExpander Expander(*SE, *DL, "hexagon-loop-idiom");
2041 
2042   Type *IntPtrTy = Builder.getIntPtrTy(*DL, SI->getPointerAddressSpace());
2043 
2044   // Okay, we have a strided store "p[i]" of a loaded value.  We can turn
2045   // this into a memcpy/memmove in the loop preheader now if we want.  However,
2046   // this would be unsafe to do if there is anything else in the loop that may
2047   // read or write the memory region we're storing to.  For memcpy, this
2048   // includes the load that feeds the stores.  Check for an alias by generating
2049   // the base address and checking everything.
2050   Value *StoreBasePtr = Expander.expandCodeFor(StoreEv->getStart(),
2051       Builder.getPtrTy(SI->getPointerAddressSpace()), ExpPt);
2052   Value *LoadBasePtr = nullptr;
2053 
2054   bool Overlap = false;
2055   bool DestVolatile = SI->isVolatile();
2056   Type *BECountTy = BECount->getType();
2057 
2058   if (DestVolatile) {
2059     // The trip count must fit in i32, since it is the type of the "num_words"
2060     // argument to hexagon_memcpy_forward_vp4cp4n2.
2061     if (StoreSize != 4 || DL->getTypeSizeInBits(BECountTy) > 32) {
2062 CleanupAndExit:
2063       // If we generated new code for the base pointer, clean up.
2064       Expander.clear();
2065       if (StoreBasePtr && (LoadBasePtr != StoreBasePtr)) {
2066         RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
2067         StoreBasePtr = nullptr;
2068       }
2069       if (LoadBasePtr) {
2070         RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
2071         LoadBasePtr = nullptr;
2072       }
2073       return false;
2074     }
2075   }
2076 
2077   SmallPtrSet<Instruction*, 2> Ignore1;
2078   Ignore1.insert(SI);
2079   if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount,
2080                             StoreSize, *AA, Ignore1)) {
2081     // Check if the load is the offending instruction.
2082     Ignore1.insert(LI);
2083     if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop,
2084                               BECount, StoreSize, *AA, Ignore1)) {
2085       // Still bad. Nothing we can do.
2086       goto CleanupAndExit;
2087     }
2088     // It worked with the load ignored.
2089     Overlap = true;
2090   }
2091 
2092   if (!Overlap) {
2093     if (DisableMemcpyIdiom || !HasMemcpy)
2094       goto CleanupAndExit;
2095   } else {
2096     // Don't generate memmove if this function will be inlined. This is
2097     // because the caller will undergo this transformation after inlining.
2098     Function *Func = CurLoop->getHeader()->getParent();
2099     if (Func->hasFnAttribute(Attribute::AlwaysInline))
2100       goto CleanupAndExit;
2101 
2102     // In case of a memmove, the call to memmove will be executed instead
2103     // of the loop, so we need to make sure that there is nothing else in
2104     // the loop than the load, store and instructions that these two depend
2105     // on.
2106     SmallVector<Instruction*,2> Insts;
2107     Insts.push_back(SI);
2108     Insts.push_back(LI);
2109     if (!coverLoop(CurLoop, Insts))
2110       goto CleanupAndExit;
2111 
2112     if (DisableMemmoveIdiom || !HasMemmove)
2113       goto CleanupAndExit;
2114     bool IsNested = CurLoop->getParentLoop() != nullptr;
2115     if (IsNested && OnlyNonNestedMemmove)
2116       goto CleanupAndExit;
2117   }
2118 
2119   // For a memcpy, we have to make sure that the input array is not being
2120   // mutated by the loop.
2121   LoadBasePtr = Expander.expandCodeFor(LoadEv->getStart(),
2122       Builder.getPtrTy(LI->getPointerAddressSpace()), ExpPt);
2123 
2124   SmallPtrSet<Instruction*, 2> Ignore2;
2125   Ignore2.insert(SI);
2126   if (mayLoopAccessLocation(LoadBasePtr, ModRefInfo::Mod, CurLoop, BECount,
2127                             StoreSize, *AA, Ignore2))
2128     goto CleanupAndExit;
2129 
2130   // Check the stride.
2131   bool StridePos = getSCEVStride(LoadEv) >= 0;
2132 
2133   // Currently, the volatile memcpy only emulates traversing memory forward.
2134   if (!StridePos && DestVolatile)
2135     goto CleanupAndExit;
2136 
2137   bool RuntimeCheck = (Overlap || DestVolatile);
2138 
2139   BasicBlock *ExitB;
2140   if (RuntimeCheck) {
2141     // The runtime check needs a single exit block.
2142     SmallVector<BasicBlock*, 8> ExitBlocks;
2143     CurLoop->getUniqueExitBlocks(ExitBlocks);
2144     if (ExitBlocks.size() != 1)
2145       goto CleanupAndExit;
2146     ExitB = ExitBlocks[0];
2147   }
2148 
2149   // The # stored bytes is (BECount+1)*Size.  Expand the trip count out to
2150   // pointer size if it isn't already.
2151   LLVMContext &Ctx = SI->getContext();
2152   BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
2153   DebugLoc DLoc = SI->getDebugLoc();
2154 
2155   const SCEV *NumBytesS =
2156       SE->getAddExpr(BECount, SE->getOne(IntPtrTy), SCEV::FlagNUW);
2157   if (StoreSize != 1)
2158     NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
2159                                SCEV::FlagNUW);
2160   Value *NumBytes = Expander.expandCodeFor(NumBytesS, IntPtrTy, ExpPt);
2161   if (Instruction *In = dyn_cast<Instruction>(NumBytes))
2162     if (Value *Simp = simplifyInstruction(In, {*DL, TLI, DT}))
2163       NumBytes = Simp;
2164 
2165   CallInst *NewCall;
2166 
2167   if (RuntimeCheck) {
2168     unsigned Threshold = RuntimeMemSizeThreshold;
2169     if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes)) {
2170       uint64_t C = CI->getZExtValue();
2171       if (Threshold != 0 && C < Threshold)
2172         goto CleanupAndExit;
2173       if (C < CompileTimeMemSizeThreshold)
2174         goto CleanupAndExit;
2175     }
2176 
2177     BasicBlock *Header = CurLoop->getHeader();
2178     Function *Func = Header->getParent();
2179     Loop *ParentL = LF->getLoopFor(Preheader);
2180     StringRef HeaderName = Header->getName();
2181 
2182     // Create a new (empty) preheader, and update the PHI nodes in the
2183     // header to use the new preheader.
2184     BasicBlock *NewPreheader = BasicBlock::Create(Ctx, HeaderName+".rtli.ph",
2185                                                   Func, Header);
2186     if (ParentL)
2187       ParentL->addBasicBlockToLoop(NewPreheader, *LF);
2188     IRBuilder<>(NewPreheader).CreateBr(Header);
2189     for (auto &In : *Header) {
2190       PHINode *PN = dyn_cast<PHINode>(&In);
2191       if (!PN)
2192         break;
2193       int bx = PN->getBasicBlockIndex(Preheader);
2194       if (bx >= 0)
2195         PN->setIncomingBlock(bx, NewPreheader);
2196     }
2197     DT->addNewBlock(NewPreheader, Preheader);
2198     DT->changeImmediateDominator(Header, NewPreheader);
2199 
2200     // Check for safe conditions to execute memmove.
2201     // If stride is positive, copying things from higher to lower addresses
2202     // is equivalent to memmove.  For negative stride, it's the other way
2203     // around.  Copying forward in memory with positive stride may not be
2204     // same as memmove since we may be copying values that we just stored
2205     // in some previous iteration.
2206     Value *LA = Builder.CreatePtrToInt(LoadBasePtr, IntPtrTy);
2207     Value *SA = Builder.CreatePtrToInt(StoreBasePtr, IntPtrTy);
2208     Value *LowA = StridePos ? SA : LA;
2209     Value *HighA = StridePos ? LA : SA;
2210     Value *CmpA = Builder.CreateICmpULT(LowA, HighA);
2211     Value *Cond = CmpA;
2212 
2213     // Check for distance between pointers. Since the case LowA < HighA
2214     // is checked for above, assume LowA >= HighA.
2215     Value *Dist = Builder.CreateSub(LowA, HighA);
2216     Value *CmpD = Builder.CreateICmpSLE(NumBytes, Dist);
2217     Value *CmpEither = Builder.CreateOr(Cond, CmpD);
2218     Cond = CmpEither;
2219 
2220     if (Threshold != 0) {
2221       Type *Ty = NumBytes->getType();
2222       Value *Thr = ConstantInt::get(Ty, Threshold);
2223       Value *CmpB = Builder.CreateICmpULT(Thr, NumBytes);
2224       Value *CmpBoth = Builder.CreateAnd(Cond, CmpB);
2225       Cond = CmpBoth;
2226     }
2227     BasicBlock *MemmoveB = BasicBlock::Create(Ctx, Header->getName()+".rtli",
2228                                               Func, NewPreheader);
2229     if (ParentL)
2230       ParentL->addBasicBlockToLoop(MemmoveB, *LF);
2231     Instruction *OldT = Preheader->getTerminator();
2232     Builder.CreateCondBr(Cond, MemmoveB, NewPreheader);
2233     OldT->eraseFromParent();
2234     Preheader->setName(Preheader->getName()+".old");
2235     DT->addNewBlock(MemmoveB, Preheader);
2236     // Find the new immediate dominator of the exit block.
2237     BasicBlock *ExitD = Preheader;
2238     for (BasicBlock *PB : predecessors(ExitB)) {
2239       ExitD = DT->findNearestCommonDominator(ExitD, PB);
2240       if (!ExitD)
2241         break;
2242     }
2243     // If the prior immediate dominator of ExitB was dominated by the
2244     // old preheader, then the old preheader becomes the new immediate
2245     // dominator.  Otherwise don't change anything (because the newly
2246     // added blocks are dominated by the old preheader).
2247     if (ExitD && DT->dominates(Preheader, ExitD)) {
2248       DomTreeNode *BN = DT->getNode(ExitB);
2249       DomTreeNode *DN = DT->getNode(ExitD);
2250       BN->setIDom(DN);
2251     }
2252 
2253     // Add a call to memmove to the conditional block.
2254     IRBuilder<> CondBuilder(MemmoveB);
2255     CondBuilder.CreateBr(ExitB);
2256     CondBuilder.SetInsertPoint(MemmoveB->getTerminator());
2257 
2258     if (DestVolatile) {
2259       Type *Int32Ty = Type::getInt32Ty(Ctx);
2260       Type *PtrTy = PointerType::get(Ctx, 0);
2261       Type *VoidTy = Type::getVoidTy(Ctx);
2262       Module *M = Func->getParent();
2263       FunctionCallee Fn = M->getOrInsertFunction(
2264           HexagonVolatileMemcpyName, VoidTy, PtrTy, PtrTy, Int32Ty);
2265 
2266       const SCEV *OneS = SE->getConstant(Int32Ty, 1);
2267       const SCEV *BECount32 = SE->getTruncateOrZeroExtend(BECount, Int32Ty);
2268       const SCEV *NumWordsS = SE->getAddExpr(BECount32, OneS, SCEV::FlagNUW);
2269       Value *NumWords = Expander.expandCodeFor(NumWordsS, Int32Ty,
2270                                                MemmoveB->getTerminator());
2271       if (Instruction *In = dyn_cast<Instruction>(NumWords))
2272         if (Value *Simp = simplifyInstruction(In, {*DL, TLI, DT}))
2273           NumWords = Simp;
2274 
2275       NewCall = CondBuilder.CreateCall(Fn,
2276                                        {StoreBasePtr, LoadBasePtr, NumWords});
2277     } else {
2278       NewCall = CondBuilder.CreateMemMove(
2279           StoreBasePtr, SI->getAlign(), LoadBasePtr, LI->getAlign(), NumBytes);
2280     }
2281   } else {
2282     NewCall = Builder.CreateMemCpy(StoreBasePtr, SI->getAlign(), LoadBasePtr,
2283                                    LI->getAlign(), NumBytes);
2284     // Okay, the memcpy has been formed.  Zap the original store and
2285     // anything that feeds into it.
2286     RecursivelyDeleteTriviallyDeadInstructions(SI, TLI);
2287   }
2288 
2289   NewCall->setDebugLoc(DLoc);
2290 
2291   LLVM_DEBUG(dbgs() << "  Formed " << (Overlap ? "memmove: " : "memcpy: ")
2292                     << *NewCall << "\n"
2293                     << "    from load ptr=" << *LoadEv << " at: " << *LI << "\n"
2294                     << "    from store ptr=" << *StoreEv << " at: " << *SI
2295                     << "\n");
2296 
2297   return true;
2298 }
2299 
2300 // Check if the instructions in Insts, together with their dependencies
2301 // cover the loop in the sense that the loop could be safely eliminated once
2302 // the instructions in Insts are removed.
coverLoop(Loop * L,SmallVectorImpl<Instruction * > & Insts) const2303 bool HexagonLoopIdiomRecognize::coverLoop(Loop *L,
2304       SmallVectorImpl<Instruction*> &Insts) const {
2305   SmallSet<BasicBlock*,8> LoopBlocks;
2306   for (auto *B : L->blocks())
2307     LoopBlocks.insert(B);
2308 
2309   SetVector<Instruction*> Worklist(Insts.begin(), Insts.end());
2310 
2311   // Collect all instructions from the loop that the instructions in Insts
2312   // depend on (plus their dependencies, etc.).  These instructions will
2313   // constitute the expression trees that feed those in Insts, but the trees
2314   // will be limited only to instructions contained in the loop.
2315   for (unsigned i = 0; i < Worklist.size(); ++i) {
2316     Instruction *In = Worklist[i];
2317     for (auto I = In->op_begin(), E = In->op_end(); I != E; ++I) {
2318       Instruction *OpI = dyn_cast<Instruction>(I);
2319       if (!OpI)
2320         continue;
2321       BasicBlock *PB = OpI->getParent();
2322       if (!LoopBlocks.count(PB))
2323         continue;
2324       Worklist.insert(OpI);
2325     }
2326   }
2327 
2328   // Scan all instructions in the loop, if any of them have a user outside
2329   // of the loop, or outside of the expressions collected above, then either
2330   // the loop has a side-effect visible outside of it, or there are
2331   // instructions in it that are not involved in the original set Insts.
2332   for (auto *B : L->blocks()) {
2333     for (auto &In : *B) {
2334       if (isa<BranchInst>(In) || isa<DbgInfoIntrinsic>(In))
2335         continue;
2336       if (!Worklist.count(&In) && In.mayHaveSideEffects())
2337         return false;
2338       for (auto *K : In.users()) {
2339         Instruction *UseI = dyn_cast<Instruction>(K);
2340         if (!UseI)
2341           continue;
2342         BasicBlock *UseB = UseI->getParent();
2343         if (LF->getLoopFor(UseB) != L)
2344           return false;
2345       }
2346     }
2347   }
2348 
2349   return true;
2350 }
2351 
2352 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
2353 /// with the specified backedge count.  This block is known to be in the current
2354 /// loop and not in any subloops.
runOnLoopBlock(Loop * CurLoop,BasicBlock * BB,const SCEV * BECount,SmallVectorImpl<BasicBlock * > & ExitBlocks)2355 bool HexagonLoopIdiomRecognize::runOnLoopBlock(Loop *CurLoop, BasicBlock *BB,
2356       const SCEV *BECount, SmallVectorImpl<BasicBlock*> &ExitBlocks) {
2357   // We can only promote stores in this block if they are unconditionally
2358   // executed in the loop.  For a block to be unconditionally executed, it has
2359   // to dominate all the exit blocks of the loop.  Verify this now.
2360   auto DominatedByBB = [this,BB] (BasicBlock *EB) -> bool {
2361     return DT->dominates(BB, EB);
2362   };
2363   if (!all_of(ExitBlocks, DominatedByBB))
2364     return false;
2365 
2366   bool MadeChange = false;
2367   // Look for store instructions, which may be optimized to memset/memcpy.
2368   SmallVector<StoreInst*,8> Stores;
2369   collectStores(CurLoop, BB, Stores);
2370 
2371   // Optimize the store into a memcpy, if it feeds an similarly strided load.
2372   for (auto &SI : Stores)
2373     MadeChange |= processCopyingStore(CurLoop, SI, BECount);
2374 
2375   return MadeChange;
2376 }
2377 
runOnCountableLoop(Loop * L)2378 bool HexagonLoopIdiomRecognize::runOnCountableLoop(Loop *L) {
2379   PolynomialMultiplyRecognize PMR(L, *DL, *DT, *TLI, *SE);
2380   if (PMR.recognize())
2381     return true;
2382 
2383   if (!HasMemcpy && !HasMemmove)
2384     return false;
2385 
2386   const SCEV *BECount = SE->getBackedgeTakenCount(L);
2387   assert(!isa<SCEVCouldNotCompute>(BECount) &&
2388          "runOnCountableLoop() called on a loop without a predictable"
2389          "backedge-taken count");
2390 
2391   SmallVector<BasicBlock *, 8> ExitBlocks;
2392   L->getUniqueExitBlocks(ExitBlocks);
2393 
2394   bool Changed = false;
2395 
2396   // Scan all the blocks in the loop that are not in subloops.
2397   for (auto *BB : L->getBlocks()) {
2398     // Ignore blocks in subloops.
2399     if (LF->getLoopFor(BB) != L)
2400       continue;
2401     Changed |= runOnLoopBlock(L, BB, BECount, ExitBlocks);
2402   }
2403 
2404   return Changed;
2405 }
2406 
run(Loop * L)2407 bool HexagonLoopIdiomRecognize::run(Loop *L) {
2408   const Module &M = *L->getHeader()->getParent()->getParent();
2409   if (Triple(M.getTargetTriple()).getArch() != Triple::hexagon)
2410     return false;
2411 
2412   // If the loop could not be converted to canonical form, it must have an
2413   // indirectbr in it, just give up.
2414   if (!L->getLoopPreheader())
2415     return false;
2416 
2417   // Disable loop idiom recognition if the function's name is a common idiom.
2418   StringRef Name = L->getHeader()->getParent()->getName();
2419   if (Name == "memset" || Name == "memcpy" || Name == "memmove")
2420     return false;
2421 
2422   DL = &L->getHeader()->getDataLayout();
2423 
2424   HasMemcpy = TLI->has(LibFunc_memcpy);
2425   HasMemmove = TLI->has(LibFunc_memmove);
2426 
2427   if (SE->hasLoopInvariantBackedgeTakenCount(L))
2428     return runOnCountableLoop(L);
2429   return false;
2430 }
2431 
runOnLoop(Loop * L,LPPassManager & LPM)2432 bool HexagonLoopIdiomRecognizeLegacyPass::runOnLoop(Loop *L,
2433                                                     LPPassManager &LPM) {
2434   if (skipLoop(L))
2435     return false;
2436 
2437   auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2438   auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2439   auto *LF = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2440   auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
2441       *L->getHeader()->getParent());
2442   auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2443   return HexagonLoopIdiomRecognize(AA, DT, LF, TLI, SE).run(L);
2444 }
2445 
createHexagonLoopIdiomPass()2446 Pass *llvm::createHexagonLoopIdiomPass() {
2447   return new HexagonLoopIdiomRecognizeLegacyPass();
2448 }
2449 
2450 PreservedAnalyses
run(Loop & L,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater & U)2451 HexagonLoopIdiomRecognitionPass::run(Loop &L, LoopAnalysisManager &AM,
2452                                      LoopStandardAnalysisResults &AR,
2453                                      LPMUpdater &U) {
2454   return HexagonLoopIdiomRecognize(&AR.AA, &AR.DT, &AR.LI, &AR.TLI, &AR.SE)
2455                  .run(&L)
2456              ? getLoopPassPreservedAnalyses()
2457              : PreservedAnalyses::all();
2458 }
2459