13b11a16aSHongbin Zheng //===--- BlockGenerators.cpp - Generate code for statements -----*- C++ -*-===//
23b11a16aSHongbin Zheng //
33b11a16aSHongbin Zheng //                     The LLVM Compiler Infrastructure
43b11a16aSHongbin Zheng //
53b11a16aSHongbin Zheng // This file is distributed under the University of Illinois Open Source
63b11a16aSHongbin Zheng // License. See LICENSE.TXT for details.
73b11a16aSHongbin Zheng //
83b11a16aSHongbin Zheng //===----------------------------------------------------------------------===//
93b11a16aSHongbin Zheng //
103b11a16aSHongbin Zheng // This file implements the BlockGenerator and VectorBlockGenerator classes,
113b11a16aSHongbin Zheng // which generate sequential code and vectorized code for a polyhedral
123b11a16aSHongbin Zheng // statement, respectively.
133b11a16aSHongbin Zheng //
143b11a16aSHongbin Zheng //===----------------------------------------------------------------------===//
153b11a16aSHongbin Zheng 
163b11a16aSHongbin Zheng #include "polly/ScopInfo.h"
1783628182STobias Grosser #include "isl/aff.h"
18a63b2579SJohannes Doerfert #include "isl/ast.h"
19a63b2579SJohannes Doerfert #include "isl/ast_build.h"
2083628182STobias Grosser #include "isl/set.h"
218a846610SHongbin Zheng #include "polly/CodeGen/BlockGenerators.h"
2283628182STobias Grosser #include "polly/CodeGen/CodeGeneration.h"
23a63b2579SJohannes Doerfert #include "polly/CodeGen/IslExprBuilder.h"
24637bd631STobias Grosser #include "polly/Options.h"
253b11a16aSHongbin Zheng #include "polly/Support/GICHelper.h"
2697cb813cSSebastian Pop #include "polly/Support/SCEVValidator.h"
27ecfe21b7STobias Grosser #include "polly/Support/ScopHelper.h"
28e71c6ab5STobias Grosser #include "llvm/Analysis/LoopInfo.h"
29e71c6ab5STobias Grosser #include "llvm/Analysis/ScalarEvolution.h"
30e71c6ab5STobias Grosser #include "llvm/Analysis/ScalarEvolutionExpander.h"
31030237d0STobias Grosser #include "llvm/IR/IntrinsicInst.h"
323b11a16aSHongbin Zheng #include "llvm/Transforms/Utils/BasicBlockUtils.h"
333b11a16aSHongbin Zheng 
343b11a16aSHongbin Zheng using namespace llvm;
353b11a16aSHongbin Zheng using namespace polly;
363b11a16aSHongbin Zheng 
37878aba49STobias Grosser static cl::opt<bool> Aligned("enable-polly-aligned",
38878aba49STobias Grosser                              cl::desc("Assumed aligned memory accesses."),
39878aba49STobias Grosser                              cl::Hidden, cl::init(false), cl::ZeroOrMore,
40878aba49STobias Grosser                              cl::cat(PollyCategory));
413b11a16aSHongbin Zheng 
42e602a076STobias Grosser static cl::opt<bool, true>
43483a90d1STobias Grosser     SCEVCodegenF("polly-codegen-scev",
44483a90d1STobias Grosser                  cl::desc("Use SCEV based code generation."), cl::Hidden,
45483a90d1STobias Grosser                  cl::location(SCEVCodegen), cl::init(false), cl::ZeroOrMore,
46483a90d1STobias Grosser                  cl::cat(PollyCategory));
47ecfe21b7STobias Grosser 
48ecfe21b7STobias Grosser bool polly::SCEVCodegen;
49ecfe21b7STobias Grosser 
50ecfe21b7STobias Grosser bool polly::canSynthesize(const Instruction *I, const llvm::LoopInfo *LI,
51ecfe21b7STobias Grosser                           ScalarEvolution *SE, const Region *R) {
52ecfe21b7STobias Grosser   if (SCEVCodegen) {
53ecfe21b7STobias Grosser     if (!I || !SE->isSCEVable(I->getType()))
54ecfe21b7STobias Grosser       return false;
55ecfe21b7STobias Grosser 
56ecfe21b7STobias Grosser     if (const SCEV *Scev = SE->getSCEV(const_cast<Instruction *>(I)))
57ecfe21b7STobias Grosser       if (!isa<SCEVCouldNotCompute>(Scev))
58ecfe21b7STobias Grosser         if (!hasScalarDepsInsideRegion(Scev, R))
59ecfe21b7STobias Grosser           return true;
60ecfe21b7STobias Grosser 
61ecfe21b7STobias Grosser     return false;
62ecfe21b7STobias Grosser   }
63ecfe21b7STobias Grosser 
64ecfe21b7STobias Grosser   Loop *L = LI->getLoopFor(I->getParent());
65e8df5bd9STobias Grosser   return L && I == L->getCanonicalInductionVariable() && R->contains(L);
66ecfe21b7STobias Grosser }
67ecfe21b7STobias Grosser 
68a63b2579SJohannes Doerfert BlockGenerator::BlockGenerator(PollyIRBuilder &B, ScopStmt &Stmt, Pass *P,
692ef3f4fdSJohannes Doerfert                                LoopInfo &LI, ScalarEvolution &SE,
70a63b2579SJohannes Doerfert                                isl_ast_build *Build,
71a63b2579SJohannes Doerfert                                IslExprBuilder *ExprBuilder)
722ef3f4fdSJohannes Doerfert     : Builder(B), Statement(Stmt), P(P), LI(LI), SE(SE), Build(Build),
732ef3f4fdSJohannes Doerfert       ExprBuilder(ExprBuilder) {}
74e71c6ab5STobias Grosser 
75d213a8b8STobias Grosser Value *BlockGenerator::getNewValue(const Value *Old, ValueMapT &BBMap,
76d213a8b8STobias Grosser                                    ValueMapT &GlobalMap, LoopToScevMapT &LTS,
77d213a8b8STobias Grosser                                    Loop *L) const {
783b11a16aSHongbin Zheng   // We assume constants never change.
793b11a16aSHongbin Zheng   // This avoids map lookups for many calls to this function.
803b11a16aSHongbin Zheng   if (isa<Constant>(Old))
813b11a16aSHongbin Zheng     return const_cast<Value *>(Old);
823b11a16aSHongbin Zheng 
83fe11e287SHongbin Zheng   if (Value *New = GlobalMap.lookup(Old)) {
84c14582f2STobias Grosser     if (Old->getType()->getScalarSizeInBits() <
85c14582f2STobias Grosser         New->getType()->getScalarSizeInBits())
863b11a16aSHongbin Zheng       New = Builder.CreateTruncOrBitCast(New, Old->getType());
873b11a16aSHongbin Zheng 
883b11a16aSHongbin Zheng     return New;
893b11a16aSHongbin Zheng   }
903b11a16aSHongbin Zheng 
91fe11e287SHongbin Zheng   if (Value *New = BBMap.lookup(Old))
92fe11e287SHongbin Zheng     return New;
933b11a16aSHongbin Zheng 
94e71c6ab5STobias Grosser   if (SCEVCodegen && SE.isSCEVable(Old->getType()))
95369430ffSTobias Grosser     if (const SCEV *Scev = SE.getSCEVAtScope(const_cast<Value *>(Old), L)) {
96e71c6ab5STobias Grosser       if (!isa<SCEVCouldNotCompute>(Scev)) {
97637b23dcSSebastian Pop         const SCEV *NewScev = apply(Scev, LTS, SE);
98637b23dcSSebastian Pop         ValueToValueMap VTV;
99637b23dcSSebastian Pop         VTV.insert(BBMap.begin(), BBMap.end());
100637b23dcSSebastian Pop         VTV.insert(GlobalMap.begin(), GlobalMap.end());
10147d4ee3eSSebastian Pop         NewScev = SCEVParameterRewriter::rewrite(NewScev, SE, VTV);
102e71c6ab5STobias Grosser         SCEVExpander Expander(SE, "polly");
103e71c6ab5STobias Grosser         Value *Expanded = Expander.expandCodeFor(NewScev, Old->getType(),
104e71c6ab5STobias Grosser                                                  Builder.GetInsertPoint());
105e71c6ab5STobias Grosser 
106e71c6ab5STobias Grosser         BBMap[Old] = Expanded;
107e71c6ab5STobias Grosser         return Expanded;
108e71c6ab5STobias Grosser       }
109369430ffSTobias Grosser     }
110e71c6ab5STobias Grosser 
111*16371acdSTobias Grosser   // A scop-constant value defined by a global or a function parameter.
112*16371acdSTobias Grosser   if (isa<GlobalValue>(Old) || isa<Argument>(Old))
113*16371acdSTobias Grosser     return const_cast<Value *>(Old);
114*16371acdSTobias Grosser 
115*16371acdSTobias Grosser   // A scop-constant value defined by an instruction executed outside the scop.
116*16371acdSTobias Grosser   if (const Instruction *Inst = dyn_cast<Instruction>(Old))
117*16371acdSTobias Grosser     if (!Statement.getParent()->getRegion().contains(Inst->getParent()))
118*16371acdSTobias Grosser       return const_cast<Value *>(Old);
119*16371acdSTobias Grosser 
120*16371acdSTobias Grosser   // The scalar dependence is neither available nor SCEVCodegenable.
1215b463ceaSHongbin Zheng   llvm_unreachable("Unexpected scalar dependence in region!");
1225a56cbf4STobias Grosser   return nullptr;
1233b11a16aSHongbin Zheng }
1243b11a16aSHongbin Zheng 
1253b11a16aSHongbin Zheng void BlockGenerator::copyInstScalar(const Instruction *Inst, ValueMapT &BBMap,
1269d10fffaSSebastian Pop                                     ValueMapT &GlobalMap, LoopToScevMapT &LTS) {
127030237d0STobias Grosser   // We do not generate debug intrinsics as we did not investigate how to
128030237d0STobias Grosser   // copy them correctly. At the current state, they just crash the code
129030237d0STobias Grosser   // generation as the meta-data operands are not correctly copied.
130030237d0STobias Grosser   if (isa<DbgInfoIntrinsic>(Inst))
131030237d0STobias Grosser     return;
132030237d0STobias Grosser 
1333b11a16aSHongbin Zheng   Instruction *NewInst = Inst->clone();
1343b11a16aSHongbin Zheng 
1353b11a16aSHongbin Zheng   // Replace old operands with the new ones.
13691f5b262STobias Grosser   for (Value *OldOperand : Inst->operands()) {
137369430ffSTobias Grosser     Value *NewOperand =
138369430ffSTobias Grosser         getNewValue(OldOperand, BBMap, GlobalMap, LTS, getLoopForInst(Inst));
1393b11a16aSHongbin Zheng 
1403b11a16aSHongbin Zheng     if (!NewOperand) {
141c14582f2STobias Grosser       assert(!isa<StoreInst>(NewInst) &&
142c14582f2STobias Grosser              "Store instructions are always needed!");
1433b11a16aSHongbin Zheng       delete NewInst;
1443b11a16aSHongbin Zheng       return;
1453b11a16aSHongbin Zheng     }
1463b11a16aSHongbin Zheng 
1473b11a16aSHongbin Zheng     NewInst->replaceUsesOfWith(OldOperand, NewOperand);
1483b11a16aSHongbin Zheng   }
1493b11a16aSHongbin Zheng 
1503b11a16aSHongbin Zheng   Builder.Insert(NewInst);
1513b11a16aSHongbin Zheng   BBMap[Inst] = NewInst;
1523b11a16aSHongbin Zheng 
1533b11a16aSHongbin Zheng   if (!NewInst->getType()->isVoidTy())
1543b11a16aSHongbin Zheng     NewInst->setName("p_" + Inst->getName());
1553b11a16aSHongbin Zheng }
1563b11a16aSHongbin Zheng 
157a63b2579SJohannes Doerfert Value *BlockGenerator::getNewAccessOperand(const MemoryAccess &MA) {
158a99130f0SJohannes Doerfert   isl_pw_multi_aff *PWAccRel;
159a99130f0SJohannes Doerfert   isl_union_map *Schedule;
160a63b2579SJohannes Doerfert   isl_ast_expr *Expr;
1613b11a16aSHongbin Zheng 
162a63b2579SJohannes Doerfert   assert(ExprBuilder && Build &&
163a63b2579SJohannes Doerfert          "Cannot generate new value without IslExprBuilder!");
1643b11a16aSHongbin Zheng 
165a99130f0SJohannes Doerfert   Schedule = isl_ast_build_get_schedule(Build);
166a99130f0SJohannes Doerfert   PWAccRel = MA.applyScheduleToAccessRelation(Schedule);
1673b11a16aSHongbin Zheng 
168a63b2579SJohannes Doerfert   Expr = isl_ast_build_access_from_pw_multi_aff(Build, PWAccRel);
169dcb5f1dcSJohannes Doerfert   Expr = isl_ast_expr_address_of(Expr);
170a63b2579SJohannes Doerfert 
171a63b2579SJohannes Doerfert   return ExprBuilder->create(Expr);
1723b11a16aSHongbin Zheng }
1733b11a16aSHongbin Zheng 
174e602a076STobias Grosser Value *BlockGenerator::generateLocationAccessed(const Instruction *Inst,
175e602a076STobias Grosser                                                 const Value *Pointer,
176e602a076STobias Grosser                                                 ValueMapT &BBMap,
177e602a076STobias Grosser                                                 ValueMapT &GlobalMap,
178e602a076STobias Grosser                                                 LoopToScevMapT &LTS) {
179a63b2579SJohannes Doerfert   const MemoryAccess &MA = Statement.getAccessFor(Inst);
1803b11a16aSHongbin Zheng 
1813b11a16aSHongbin Zheng   Value *NewPointer;
182a99130f0SJohannes Doerfert   if (MA.hasNewAccessRelation())
183a63b2579SJohannes Doerfert     NewPointer = getNewAccessOperand(MA);
184a63b2579SJohannes Doerfert   else
185369430ffSTobias Grosser     NewPointer =
186369430ffSTobias Grosser         getNewValue(Pointer, BBMap, GlobalMap, LTS, getLoopForInst(Inst));
1873b11a16aSHongbin Zheng 
1883b11a16aSHongbin Zheng   return NewPointer;
1893b11a16aSHongbin Zheng }
1903b11a16aSHongbin Zheng 
1914d96c8d7STobias Grosser Loop *BlockGenerator::getLoopForInst(const llvm::Instruction *Inst) {
1922ef3f4fdSJohannes Doerfert   return LI.getLoopFor(Inst->getParent());
193369430ffSTobias Grosser }
194369430ffSTobias Grosser 
195e602a076STobias Grosser Value *BlockGenerator::generateScalarLoad(const LoadInst *Load,
196e602a076STobias Grosser                                           ValueMapT &BBMap,
197e602a076STobias Grosser                                           ValueMapT &GlobalMap,
198e602a076STobias Grosser                                           LoopToScevMapT &LTS) {
1993b11a16aSHongbin Zheng   const Value *Pointer = Load->getPointerOperand();
2007242ad92STobias Grosser   Value *NewPointer =
2011947f863SJohannes Doerfert       generateLocationAccessed(Load, Pointer, BBMap, GlobalMap, LTS);
20287901453SJohannes Doerfert   Value *ScalarLoad = Builder.CreateAlignedLoad(
20387901453SJohannes Doerfert       NewPointer, Load->getAlignment(), Load->getName() + "_p_scalar_");
2043b11a16aSHongbin Zheng   return ScalarLoad;
2053b11a16aSHongbin Zheng }
2063b11a16aSHongbin Zheng 
207e602a076STobias Grosser Value *BlockGenerator::generateScalarStore(const StoreInst *Store,
208e602a076STobias Grosser                                            ValueMapT &BBMap,
209e602a076STobias Grosser                                            ValueMapT &GlobalMap,
210e602a076STobias Grosser                                            LoopToScevMapT &LTS) {
2113b11a16aSHongbin Zheng   const Value *Pointer = Store->getPointerOperand();
212c14582f2STobias Grosser   Value *NewPointer =
2139d10fffaSSebastian Pop       generateLocationAccessed(Store, Pointer, BBMap, GlobalMap, LTS);
214369430ffSTobias Grosser   Value *ValueOperand = getNewValue(Store->getValueOperand(), BBMap, GlobalMap,
215369430ffSTobias Grosser                                     LTS, getLoopForInst(Store));
2163b11a16aSHongbin Zheng 
21787901453SJohannes Doerfert   Value *NewStore = Builder.CreateAlignedStore(ValueOperand, NewPointer,
21887901453SJohannes Doerfert                                                Store->getAlignment());
21987901453SJohannes Doerfert   return NewStore;
2203b11a16aSHongbin Zheng }
2213b11a16aSHongbin Zheng 
222e602a076STobias Grosser void BlockGenerator::copyInstruction(const Instruction *Inst, ValueMapT &BBMap,
223e602a076STobias Grosser                                      ValueMapT &GlobalMap,
224e602a076STobias Grosser                                      LoopToScevMapT &LTS) {
2253b11a16aSHongbin Zheng   // Terminator instructions control the control flow. They are explicitly
2263b11a16aSHongbin Zheng   // expressed in the clast and do not need to be copied.
2273b11a16aSHongbin Zheng   if (Inst->isTerminator())
2283b11a16aSHongbin Zheng     return;
2293b11a16aSHongbin Zheng 
230ecfe21b7STobias Grosser   if (canSynthesize(Inst, &P->getAnalysis<LoopInfo>(), &SE,
231ecfe21b7STobias Grosser                     &Statement.getParent()->getRegion()))
232e71c6ab5STobias Grosser     return;
233e71c6ab5STobias Grosser 
2343b11a16aSHongbin Zheng   if (const LoadInst *Load = dyn_cast<LoadInst>(Inst)) {
235753d43f9SSebastian Pop     Value *NewLoad = generateScalarLoad(Load, BBMap, GlobalMap, LTS);
2363d94fedfSSebastian Pop     // Compute NewLoad before its insertion in BBMap to make the insertion
2373d94fedfSSebastian Pop     // deterministic.
238753d43f9SSebastian Pop     BBMap[Load] = NewLoad;
2393b11a16aSHongbin Zheng     return;
2403b11a16aSHongbin Zheng   }
2413b11a16aSHongbin Zheng 
2423b11a16aSHongbin Zheng   if (const StoreInst *Store = dyn_cast<StoreInst>(Inst)) {
243753d43f9SSebastian Pop     Value *NewStore = generateScalarStore(Store, BBMap, GlobalMap, LTS);
2443d94fedfSSebastian Pop     // Compute NewStore before its insertion in BBMap to make the insertion
2453d94fedfSSebastian Pop     // deterministic.
246753d43f9SSebastian Pop     BBMap[Store] = NewStore;
2473b11a16aSHongbin Zheng     return;
2483b11a16aSHongbin Zheng   }
2493b11a16aSHongbin Zheng 
2509d10fffaSSebastian Pop   copyInstScalar(Inst, BBMap, GlobalMap, LTS);
2513b11a16aSHongbin Zheng }
2523b11a16aSHongbin Zheng 
2539d10fffaSSebastian Pop void BlockGenerator::copyBB(ValueMapT &GlobalMap, LoopToScevMapT &LTS) {
2543b11a16aSHongbin Zheng   BasicBlock *BB = Statement.getBasicBlock();
255c14582f2STobias Grosser   BasicBlock *CopyBB =
256c14582f2STobias Grosser       SplitBlock(Builder.GetInsertBlock(), Builder.GetInsertPoint(), P);
2573b11a16aSHongbin Zheng   CopyBB->setName("polly.stmt." + BB->getName());
2583b11a16aSHongbin Zheng   Builder.SetInsertPoint(CopyBB->begin());
2593b11a16aSHongbin Zheng 
2603b11a16aSHongbin Zheng   ValueMapT BBMap;
2613b11a16aSHongbin Zheng 
26291f5b262STobias Grosser   for (Instruction &Inst : *BB)
26391f5b262STobias Grosser     copyInstruction(&Inst, BBMap, GlobalMap, LTS);
2643b11a16aSHongbin Zheng }
2653b11a16aSHongbin Zheng 
2662ef3f4fdSJohannes Doerfert VectorBlockGenerator::VectorBlockGenerator(
2672ef3f4fdSJohannes Doerfert     PollyIRBuilder &B, VectorValueMapT &GlobalMaps,
2682ef3f4fdSJohannes Doerfert     std::vector<LoopToScevMapT> &VLTS, ScopStmt &Stmt,
269731685e6SJohannes Doerfert     __isl_keep isl_map *Schedule, Pass *P, LoopInfo &LI, ScalarEvolution &SE,
270731685e6SJohannes Doerfert     __isl_keep isl_ast_build *Build, IslExprBuilder *ExprBuilder)
271731685e6SJohannes Doerfert     : BlockGenerator(B, Stmt, P, LI, SE, Build, ExprBuilder),
2722ef3f4fdSJohannes Doerfert       GlobalMaps(GlobalMaps), VLTS(VLTS), Schedule(Schedule) {
2733b11a16aSHongbin Zheng   assert(GlobalMaps.size() > 1 && "Only one vector lane found");
274a00a0291SSebastian Pop   assert(Schedule && "No statement domain provided");
2753b11a16aSHongbin Zheng }
2763b11a16aSHongbin Zheng 
277e602a076STobias Grosser Value *VectorBlockGenerator::getVectorValue(const Value *Old,
278e602a076STobias Grosser                                             ValueMapT &VectorMap,
279e602a076STobias Grosser                                             VectorValueMapT &ScalarMaps,
280e602a076STobias Grosser                                             Loop *L) {
281fe11e287SHongbin Zheng   if (Value *NewValue = VectorMap.lookup(Old))
282fe11e287SHongbin Zheng     return NewValue;
2833b11a16aSHongbin Zheng 
2843b11a16aSHongbin Zheng   int Width = getVectorWidth();
2853b11a16aSHongbin Zheng 
2863b11a16aSHongbin Zheng   Value *Vector = UndefValue::get(VectorType::get(Old->getType(), Width));
2873b11a16aSHongbin Zheng 
2883b11a16aSHongbin Zheng   for (int Lane = 0; Lane < Width; Lane++)
289c14582f2STobias Grosser     Vector = Builder.CreateInsertElement(
2907242ad92STobias Grosser         Vector,
291369430ffSTobias Grosser         getNewValue(Old, ScalarMaps[Lane], GlobalMaps[Lane], VLTS[Lane], L),
2927242ad92STobias Grosser         Builder.getInt32(Lane));
2933b11a16aSHongbin Zheng 
2943b11a16aSHongbin Zheng   VectorMap[Old] = Vector;
2953b11a16aSHongbin Zheng 
2963b11a16aSHongbin Zheng   return Vector;
2973b11a16aSHongbin Zheng }
2983b11a16aSHongbin Zheng 
2993b11a16aSHongbin Zheng Type *VectorBlockGenerator::getVectorPtrTy(const Value *Val, int Width) {
3003b11a16aSHongbin Zheng   PointerType *PointerTy = dyn_cast<PointerType>(Val->getType());
3013b11a16aSHongbin Zheng   assert(PointerTy && "PointerType expected");
3023b11a16aSHongbin Zheng 
3033b11a16aSHongbin Zheng   Type *ScalarType = PointerTy->getElementType();
3043b11a16aSHongbin Zheng   VectorType *VectorType = VectorType::get(ScalarType, Width);
3053b11a16aSHongbin Zheng 
3063b11a16aSHongbin Zheng   return PointerType::getUnqual(VectorType);
3073b11a16aSHongbin Zheng }
3083b11a16aSHongbin Zheng 
3090dd463faSTobias Grosser Value *
3100dd463faSTobias Grosser VectorBlockGenerator::generateStrideOneLoad(const LoadInst *Load,
3110dd463faSTobias Grosser                                             VectorValueMapT &ScalarMaps,
3120dd463faSTobias Grosser                                             bool NegativeStride = false) {
3130dd463faSTobias Grosser   unsigned VectorWidth = getVectorWidth();
3143b11a16aSHongbin Zheng   const Value *Pointer = Load->getPointerOperand();
3150dd463faSTobias Grosser   Type *VectorPtrType = getVectorPtrTy(Pointer, VectorWidth);
3160dd463faSTobias Grosser   unsigned Offset = NegativeStride ? VectorWidth - 1 : 0;
3170dd463faSTobias Grosser 
3185a56cbf4STobias Grosser   Value *NewPointer = nullptr;
319731685e6SJohannes Doerfert   NewPointer = generateLocationAccessed(Load, Pointer, ScalarMaps[Offset],
320731685e6SJohannes Doerfert                                         GlobalMaps[Offset], VLTS[Offset]);
321c14582f2STobias Grosser   Value *VectorPtr =
322c14582f2STobias Grosser       Builder.CreateBitCast(NewPointer, VectorPtrType, "vector_ptr");
323c14582f2STobias Grosser   LoadInst *VecLoad =
324c14582f2STobias Grosser       Builder.CreateLoad(VectorPtr, Load->getName() + "_p_vec_full");
3253b11a16aSHongbin Zheng   if (!Aligned)
3263b11a16aSHongbin Zheng     VecLoad->setAlignment(8);
3273b11a16aSHongbin Zheng 
3280dd463faSTobias Grosser   if (NegativeStride) {
3290dd463faSTobias Grosser     SmallVector<Constant *, 16> Indices;
3300dd463faSTobias Grosser     for (int i = VectorWidth - 1; i >= 0; i--)
3310dd463faSTobias Grosser       Indices.push_back(ConstantInt::get(Builder.getInt32Ty(), i));
3320dd463faSTobias Grosser     Constant *SV = llvm::ConstantVector::get(Indices);
3330dd463faSTobias Grosser     Value *RevVecLoad = Builder.CreateShuffleVector(
3340dd463faSTobias Grosser         VecLoad, VecLoad, SV, Load->getName() + "_reverse");
3350dd463faSTobias Grosser     return RevVecLoad;
3360dd463faSTobias Grosser   }
3370dd463faSTobias Grosser 
3383b11a16aSHongbin Zheng   return VecLoad;
3393b11a16aSHongbin Zheng }
3403b11a16aSHongbin Zheng 
3413b11a16aSHongbin Zheng Value *VectorBlockGenerator::generateStrideZeroLoad(const LoadInst *Load,
3423b11a16aSHongbin Zheng                                                     ValueMapT &BBMap) {
3433b11a16aSHongbin Zheng   const Value *Pointer = Load->getPointerOperand();
3443b11a16aSHongbin Zheng   Type *VectorPtrType = getVectorPtrTy(Pointer, 1);
345369430ffSTobias Grosser   Value *NewPointer =
346731685e6SJohannes Doerfert       generateLocationAccessed(Load, Pointer, BBMap, GlobalMaps[0], VLTS[0]);
3473b11a16aSHongbin Zheng   Value *VectorPtr = Builder.CreateBitCast(NewPointer, VectorPtrType,
3483b11a16aSHongbin Zheng                                            Load->getName() + "_p_vec_p");
349c14582f2STobias Grosser   LoadInst *ScalarLoad =
350c14582f2STobias Grosser       Builder.CreateLoad(VectorPtr, Load->getName() + "_p_splat_one");
3513b11a16aSHongbin Zheng 
3523b11a16aSHongbin Zheng   if (!Aligned)
3533b11a16aSHongbin Zheng     ScalarLoad->setAlignment(8);
3543b11a16aSHongbin Zheng 
355c14582f2STobias Grosser   Constant *SplatVector = Constant::getNullValue(
356c14582f2STobias Grosser       VectorType::get(Builder.getInt32Ty(), getVectorWidth()));
3573b11a16aSHongbin Zheng 
358c14582f2STobias Grosser   Value *VectorLoad = Builder.CreateShuffleVector(
359c14582f2STobias Grosser       ScalarLoad, ScalarLoad, SplatVector, Load->getName() + "_p_splat");
3603b11a16aSHongbin Zheng   return VectorLoad;
3613b11a16aSHongbin Zheng }
3623b11a16aSHongbin Zheng 
363e602a076STobias Grosser Value *
364e602a076STobias Grosser VectorBlockGenerator::generateUnknownStrideLoad(const LoadInst *Load,
365e602a076STobias Grosser                                                 VectorValueMapT &ScalarMaps) {
3663b11a16aSHongbin Zheng   int VectorWidth = getVectorWidth();
3673b11a16aSHongbin Zheng   const Value *Pointer = Load->getPointerOperand();
3683b11a16aSHongbin Zheng   VectorType *VectorType = VectorType::get(
3693b11a16aSHongbin Zheng       dyn_cast<PointerType>(Pointer->getType())->getElementType(), VectorWidth);
3703b11a16aSHongbin Zheng 
3713b11a16aSHongbin Zheng   Value *Vector = UndefValue::get(VectorType);
3723b11a16aSHongbin Zheng 
3733b11a16aSHongbin Zheng   for (int i = 0; i < VectorWidth; i++) {
374731685e6SJohannes Doerfert     Value *NewPointer = generateLocationAccessed(Load, Pointer, ScalarMaps[i],
375731685e6SJohannes Doerfert                                                  GlobalMaps[i], VLTS[i]);
376c14582f2STobias Grosser     Value *ScalarLoad =
377c14582f2STobias Grosser         Builder.CreateLoad(NewPointer, Load->getName() + "_p_scalar_");
378c14582f2STobias Grosser     Vector = Builder.CreateInsertElement(
379c14582f2STobias Grosser         Vector, ScalarLoad, Builder.getInt32(i), Load->getName() + "_p_vec_");
3803b11a16aSHongbin Zheng   }
3813b11a16aSHongbin Zheng 
3823b11a16aSHongbin Zheng   return Vector;
3833b11a16aSHongbin Zheng }
3843b11a16aSHongbin Zheng 
385e602a076STobias Grosser void VectorBlockGenerator::generateLoad(const LoadInst *Load,
386e602a076STobias Grosser                                         ValueMapT &VectorMap,
387e602a076STobias Grosser                                         VectorValueMapT &ScalarMaps) {
38868794217SHongbin Zheng   if (PollyVectorizerChoice >= VECTORIZER_FIRST_NEED_GROUPED_UNROLL ||
38968794217SHongbin Zheng       !VectorType::isValidElementType(Load->getType())) {
3903b11a16aSHongbin Zheng     for (int i = 0; i < getVectorWidth(); i++)
391c14582f2STobias Grosser       ScalarMaps[i][Load] =
3929d10fffaSSebastian Pop           generateScalarLoad(Load, ScalarMaps[i], GlobalMaps[i], VLTS[i]);
3933b11a16aSHongbin Zheng     return;
3943b11a16aSHongbin Zheng   }
3953b11a16aSHongbin Zheng 
396b5f24a66SHongbin Zheng   const MemoryAccess &Access = Statement.getAccessFor(Load);
3973b11a16aSHongbin Zheng 
39895493984STobias Grosser   // Make sure we have scalar values available to access the pointer to
39995493984STobias Grosser   // the data location.
40095493984STobias Grosser   extractScalarValues(Load, VectorMap, ScalarMaps);
40195493984STobias Grosser 
4023b11a16aSHongbin Zheng   Value *NewLoad;
403a00a0291SSebastian Pop   if (Access.isStrideZero(isl_map_copy(Schedule)))
4043b11a16aSHongbin Zheng     NewLoad = generateStrideZeroLoad(Load, ScalarMaps[0]);
405a00a0291SSebastian Pop   else if (Access.isStrideOne(isl_map_copy(Schedule)))
4060dd463faSTobias Grosser     NewLoad = generateStrideOneLoad(Load, ScalarMaps);
4070dd463faSTobias Grosser   else if (Access.isStrideX(isl_map_copy(Schedule), -1))
4080dd463faSTobias Grosser     NewLoad = generateStrideOneLoad(Load, ScalarMaps, true);
4093b11a16aSHongbin Zheng   else
4103b11a16aSHongbin Zheng     NewLoad = generateUnknownStrideLoad(Load, ScalarMaps);
4113b11a16aSHongbin Zheng 
4123b11a16aSHongbin Zheng   VectorMap[Load] = NewLoad;
4133b11a16aSHongbin Zheng }
4143b11a16aSHongbin Zheng 
4153b11a16aSHongbin Zheng void VectorBlockGenerator::copyUnaryInst(const UnaryInstruction *Inst,
4163b11a16aSHongbin Zheng                                          ValueMapT &VectorMap,
4173b11a16aSHongbin Zheng                                          VectorValueMapT &ScalarMaps) {
4183b11a16aSHongbin Zheng   int VectorWidth = getVectorWidth();
419369430ffSTobias Grosser   Value *NewOperand = getVectorValue(Inst->getOperand(0), VectorMap, ScalarMaps,
420369430ffSTobias Grosser                                      getLoopForInst(Inst));
4213b11a16aSHongbin Zheng 
4223b11a16aSHongbin Zheng   assert(isa<CastInst>(Inst) && "Can not generate vector code for instruction");
4233b11a16aSHongbin Zheng 
4243b11a16aSHongbin Zheng   const CastInst *Cast = dyn_cast<CastInst>(Inst);
4253b11a16aSHongbin Zheng   VectorType *DestType = VectorType::get(Inst->getType(), VectorWidth);
4263b11a16aSHongbin Zheng   VectorMap[Inst] = Builder.CreateCast(Cast->getOpcode(), NewOperand, DestType);
4273b11a16aSHongbin Zheng }
4283b11a16aSHongbin Zheng 
4293b11a16aSHongbin Zheng void VectorBlockGenerator::copyBinaryInst(const BinaryOperator *Inst,
4303b11a16aSHongbin Zheng                                           ValueMapT &VectorMap,
4313b11a16aSHongbin Zheng                                           VectorValueMapT &ScalarMaps) {
432369430ffSTobias Grosser   Loop *L = getLoopForInst(Inst);
4333b11a16aSHongbin Zheng   Value *OpZero = Inst->getOperand(0);
4343b11a16aSHongbin Zheng   Value *OpOne = Inst->getOperand(1);
4353b11a16aSHongbin Zheng 
4363b11a16aSHongbin Zheng   Value *NewOpZero, *NewOpOne;
437369430ffSTobias Grosser   NewOpZero = getVectorValue(OpZero, VectorMap, ScalarMaps, L);
438369430ffSTobias Grosser   NewOpOne = getVectorValue(OpOne, VectorMap, ScalarMaps, L);
4393b11a16aSHongbin Zheng 
4401bb59b0dSTobias Grosser   Value *NewInst = Builder.CreateBinOp(Inst->getOpcode(), NewOpZero, NewOpOne,
4413b11a16aSHongbin Zheng                                        Inst->getName() + "p_vec");
4423b11a16aSHongbin Zheng   VectorMap[Inst] = NewInst;
4433b11a16aSHongbin Zheng }
4443b11a16aSHongbin Zheng 
445e602a076STobias Grosser void VectorBlockGenerator::copyStore(const StoreInst *Store,
446e602a076STobias Grosser                                      ValueMapT &VectorMap,
447e602a076STobias Grosser                                      VectorValueMapT &ScalarMaps) {
448b5f24a66SHongbin Zheng   const MemoryAccess &Access = Statement.getAccessFor(Store);
4493b11a16aSHongbin Zheng 
4503b11a16aSHongbin Zheng   const Value *Pointer = Store->getPointerOperand();
451369430ffSTobias Grosser   Value *Vector = getVectorValue(Store->getValueOperand(), VectorMap,
452369430ffSTobias Grosser                                  ScalarMaps, getLoopForInst(Store));
4533b11a16aSHongbin Zheng 
45450fd7010STobias Grosser   // Make sure we have scalar values available to access the pointer to
45550fd7010STobias Grosser   // the data location.
45650fd7010STobias Grosser   extractScalarValues(Store, VectorMap, ScalarMaps);
45750fd7010STobias Grosser 
458a00a0291SSebastian Pop   if (Access.isStrideOne(isl_map_copy(Schedule))) {
4591947f863SJohannes Doerfert     Type *VectorPtrType = getVectorPtrTy(Pointer, getVectorWidth());
460731685e6SJohannes Doerfert     Value *NewPointer = generateLocationAccessed(Store, Pointer, ScalarMaps[0],
461731685e6SJohannes Doerfert                                                  GlobalMaps[0], VLTS[0]);
4623b11a16aSHongbin Zheng 
463c14582f2STobias Grosser     Value *VectorPtr =
464c14582f2STobias Grosser         Builder.CreateBitCast(NewPointer, VectorPtrType, "vector_ptr");
4653b11a16aSHongbin Zheng     StoreInst *Store = Builder.CreateStore(Vector, VectorPtr);
4663b11a16aSHongbin Zheng 
4673b11a16aSHongbin Zheng     if (!Aligned)
4683b11a16aSHongbin Zheng       Store->setAlignment(8);
4693b11a16aSHongbin Zheng   } else {
4703b11a16aSHongbin Zheng     for (unsigned i = 0; i < ScalarMaps.size(); i++) {
4711bb59b0dSTobias Grosser       Value *Scalar = Builder.CreateExtractElement(Vector, Builder.getInt32(i));
472731685e6SJohannes Doerfert       Value *NewPointer = generateLocationAccessed(
473731685e6SJohannes Doerfert           Store, Pointer, ScalarMaps[i], GlobalMaps[i], VLTS[i]);
4743b11a16aSHongbin Zheng       Builder.CreateStore(Scalar, NewPointer);
4753b11a16aSHongbin Zheng     }
4763b11a16aSHongbin Zheng   }
4773b11a16aSHongbin Zheng }
4783b11a16aSHongbin Zheng 
4793b11a16aSHongbin Zheng bool VectorBlockGenerator::hasVectorOperands(const Instruction *Inst,
4803b11a16aSHongbin Zheng                                              ValueMapT &VectorMap) {
48191f5b262STobias Grosser   for (Value *Operand : Inst->operands())
48291f5b262STobias Grosser     if (VectorMap.count(Operand))
4833b11a16aSHongbin Zheng       return true;
4843b11a16aSHongbin Zheng   return false;
4853b11a16aSHongbin Zheng }
4863b11a16aSHongbin Zheng 
4873b11a16aSHongbin Zheng bool VectorBlockGenerator::extractScalarValues(const Instruction *Inst,
4883b11a16aSHongbin Zheng                                                ValueMapT &VectorMap,
4893b11a16aSHongbin Zheng                                                VectorValueMapT &ScalarMaps) {
4903b11a16aSHongbin Zheng   bool HasVectorOperand = false;
4913b11a16aSHongbin Zheng   int VectorWidth = getVectorWidth();
4923b11a16aSHongbin Zheng 
49391f5b262STobias Grosser   for (Value *Operand : Inst->operands()) {
49491f5b262STobias Grosser     ValueMapT::iterator VecOp = VectorMap.find(Operand);
4953b11a16aSHongbin Zheng 
4963b11a16aSHongbin Zheng     if (VecOp == VectorMap.end())
4973b11a16aSHongbin Zheng       continue;
4983b11a16aSHongbin Zheng 
4993b11a16aSHongbin Zheng     HasVectorOperand = true;
5003b11a16aSHongbin Zheng     Value *NewVector = VecOp->second;
5013b11a16aSHongbin Zheng 
5023b11a16aSHongbin Zheng     for (int i = 0; i < VectorWidth; ++i) {
5033b11a16aSHongbin Zheng       ValueMapT &SM = ScalarMaps[i];
5043b11a16aSHongbin Zheng 
5053b11a16aSHongbin Zheng       // If there is one scalar extracted, all scalar elements should have
5063b11a16aSHongbin Zheng       // already been extracted by the code here. So no need to check for the
5073b11a16aSHongbin Zheng       // existance of all of them.
50891f5b262STobias Grosser       if (SM.count(Operand))
5093b11a16aSHongbin Zheng         break;
5103b11a16aSHongbin Zheng 
51191f5b262STobias Grosser       SM[Operand] =
51291f5b262STobias Grosser           Builder.CreateExtractElement(NewVector, Builder.getInt32(i));
5133b11a16aSHongbin Zheng     }
5143b11a16aSHongbin Zheng   }
5153b11a16aSHongbin Zheng 
5163b11a16aSHongbin Zheng   return HasVectorOperand;
5173b11a16aSHongbin Zheng }
5183b11a16aSHongbin Zheng 
5193b11a16aSHongbin Zheng void VectorBlockGenerator::copyInstScalarized(const Instruction *Inst,
5203b11a16aSHongbin Zheng                                               ValueMapT &VectorMap,
5213b11a16aSHongbin Zheng                                               VectorValueMapT &ScalarMaps) {
5223b11a16aSHongbin Zheng   bool HasVectorOperand;
5233b11a16aSHongbin Zheng   int VectorWidth = getVectorWidth();
5243b11a16aSHongbin Zheng 
5253b11a16aSHongbin Zheng   HasVectorOperand = extractScalarValues(Inst, VectorMap, ScalarMaps);
5263b11a16aSHongbin Zheng 
5273b11a16aSHongbin Zheng   for (int VectorLane = 0; VectorLane < getVectorWidth(); VectorLane++)
528731685e6SJohannes Doerfert     BlockGenerator::copyInstruction(Inst, ScalarMaps[VectorLane],
529731685e6SJohannes Doerfert                                     GlobalMaps[VectorLane], VLTS[VectorLane]);
5303b11a16aSHongbin Zheng 
5313b11a16aSHongbin Zheng   if (!VectorType::isValidElementType(Inst->getType()) || !HasVectorOperand)
5323b11a16aSHongbin Zheng     return;
5333b11a16aSHongbin Zheng 
5343b11a16aSHongbin Zheng   // Make the result available as vector value.
5353b11a16aSHongbin Zheng   VectorType *VectorType = VectorType::get(Inst->getType(), VectorWidth);
5363b11a16aSHongbin Zheng   Value *Vector = UndefValue::get(VectorType);
5373b11a16aSHongbin Zheng 
5383b11a16aSHongbin Zheng   for (int i = 0; i < VectorWidth; i++)
5393b11a16aSHongbin Zheng     Vector = Builder.CreateInsertElement(Vector, ScalarMaps[i][Inst],
5403b11a16aSHongbin Zheng                                          Builder.getInt32(i));
5413b11a16aSHongbin Zheng 
5423b11a16aSHongbin Zheng   VectorMap[Inst] = Vector;
5433b11a16aSHongbin Zheng }
5443b11a16aSHongbin Zheng 
545c14582f2STobias Grosser int VectorBlockGenerator::getVectorWidth() { return GlobalMaps.size(); }
5463b11a16aSHongbin Zheng 
5473b11a16aSHongbin Zheng void VectorBlockGenerator::copyInstruction(const Instruction *Inst,
5483b11a16aSHongbin Zheng                                            ValueMapT &VectorMap,
5493b11a16aSHongbin Zheng                                            VectorValueMapT &ScalarMaps) {
5503b11a16aSHongbin Zheng   // Terminator instructions control the control flow. They are explicitly
5513b11a16aSHongbin Zheng   // expressed in the clast and do not need to be copied.
5523b11a16aSHongbin Zheng   if (Inst->isTerminator())
5533b11a16aSHongbin Zheng     return;
5543b11a16aSHongbin Zheng 
555ecfe21b7STobias Grosser   if (canSynthesize(Inst, &P->getAnalysis<LoopInfo>(), &SE,
556ecfe21b7STobias Grosser                     &Statement.getParent()->getRegion()))
557e71c6ab5STobias Grosser     return;
558e71c6ab5STobias Grosser 
5593b11a16aSHongbin Zheng   if (const LoadInst *Load = dyn_cast<LoadInst>(Inst)) {
5603b11a16aSHongbin Zheng     generateLoad(Load, VectorMap, ScalarMaps);
5613b11a16aSHongbin Zheng     return;
5623b11a16aSHongbin Zheng   }
5633b11a16aSHongbin Zheng 
5643b11a16aSHongbin Zheng   if (hasVectorOperands(Inst, VectorMap)) {
5653b11a16aSHongbin Zheng     if (const StoreInst *Store = dyn_cast<StoreInst>(Inst)) {
5663b11a16aSHongbin Zheng       copyStore(Store, VectorMap, ScalarMaps);
5673b11a16aSHongbin Zheng       return;
5683b11a16aSHongbin Zheng     }
5693b11a16aSHongbin Zheng 
5703b11a16aSHongbin Zheng     if (const UnaryInstruction *Unary = dyn_cast<UnaryInstruction>(Inst)) {
5713b11a16aSHongbin Zheng       copyUnaryInst(Unary, VectorMap, ScalarMaps);
5723b11a16aSHongbin Zheng       return;
5733b11a16aSHongbin Zheng     }
5743b11a16aSHongbin Zheng 
5753b11a16aSHongbin Zheng     if (const BinaryOperator *Binary = dyn_cast<BinaryOperator>(Inst)) {
5763b11a16aSHongbin Zheng       copyBinaryInst(Binary, VectorMap, ScalarMaps);
5773b11a16aSHongbin Zheng       return;
5783b11a16aSHongbin Zheng     }
5793b11a16aSHongbin Zheng 
5803b11a16aSHongbin Zheng     // Falltrough: We generate scalar instructions, if we don't know how to
5813b11a16aSHongbin Zheng     // generate vector code.
5823b11a16aSHongbin Zheng   }
5833b11a16aSHongbin Zheng 
5843b11a16aSHongbin Zheng   copyInstScalarized(Inst, VectorMap, ScalarMaps);
5853b11a16aSHongbin Zheng }
5863b11a16aSHongbin Zheng 
5873b11a16aSHongbin Zheng void VectorBlockGenerator::copyBB() {
5883b11a16aSHongbin Zheng   BasicBlock *BB = Statement.getBasicBlock();
589c14582f2STobias Grosser   BasicBlock *CopyBB =
590c14582f2STobias Grosser       SplitBlock(Builder.GetInsertBlock(), Builder.GetInsertPoint(), P);
5913b11a16aSHongbin Zheng   CopyBB->setName("polly.stmt." + BB->getName());
5923b11a16aSHongbin Zheng   Builder.SetInsertPoint(CopyBB->begin());
5933b11a16aSHongbin Zheng 
5943b11a16aSHongbin Zheng   // Create two maps that store the mapping from the original instructions of
5953b11a16aSHongbin Zheng   // the old basic block to their copies in the new basic block. Those maps
5963b11a16aSHongbin Zheng   // are basic block local.
5973b11a16aSHongbin Zheng   //
5983b11a16aSHongbin Zheng   // As vector code generation is supported there is one map for scalar values
5993b11a16aSHongbin Zheng   // and one for vector values.
6003b11a16aSHongbin Zheng   //
6013b11a16aSHongbin Zheng   // In case we just do scalar code generation, the vectorMap is not used and
6023b11a16aSHongbin Zheng   // the scalarMap has just one dimension, which contains the mapping.
6033b11a16aSHongbin Zheng   //
6043b11a16aSHongbin Zheng   // In case vector code generation is done, an instruction may either appear
6053b11a16aSHongbin Zheng   // in the vector map once (as it is calculating >vectorwidth< values at a
6063b11a16aSHongbin Zheng   // time. Or (if the values are calculated using scalar operations), it
6073b11a16aSHongbin Zheng   // appears once in every dimension of the scalarMap.
6083b11a16aSHongbin Zheng   VectorValueMapT ScalarBlockMap(getVectorWidth());
6093b11a16aSHongbin Zheng   ValueMapT VectorBlockMap;
6103b11a16aSHongbin Zheng 
61191f5b262STobias Grosser   for (Instruction &Inst : *BB)
61291f5b262STobias Grosser     copyInstruction(&Inst, VectorBlockMap, ScalarBlockMap);
6133b11a16aSHongbin Zheng }
614