1 //===- SLPVectorizer.cpp - A bottom up SLP Vectorizer ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
9 // This pass implements the Bottom Up SLP vectorizer. It detects consecutive
10 // stores that can be put together into vector-stores. Next, it attempts to
11 // construct vectorizable tree using the use-def chains. If a profitable tree
12 // was found, the SLP vectorizer performs vectorization on the tree.
14 // The pass is inspired by the work described in the paper:
15 // "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks.
17 //===----------------------------------------------------------------------===//
18 #define SV_NAME "slp-vectorizer"
19 #define DEBUG_TYPE SV_NAME
22 #include "llvm/Transforms/Vectorize.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/ScalarEvolution.h"
25 #include "llvm/Analysis/TargetTransformInfo.h"
26 #include "llvm/Analysis/Verifier.h"
27 #include "llvm/Analysis/LoopInfo.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/IR/Type.h"
33 #include "llvm/IR/Value.h"
34 #include "llvm/Pass.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/raw_ostream.h"
43 SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden,
44 cl::desc("Only vectorize trees if the gain is above this "
45 "number. (gain = -cost of vectorization)"));
48 /// The SLPVectorizer Pass.
49 struct SLPVectorizer : public FunctionPass {
50 typedef std::map<Value*, BoUpSLP::StoreList> StoreListMap;
52 /// Pass identification, replacement for typeid
55 explicit SLPVectorizer() : FunctionPass(ID) {
56 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
61 TargetTransformInfo *TTI;
65 virtual bool runOnFunction(Function &F) {
66 SE = &getAnalysis<ScalarEvolution>();
67 DL = getAnalysisIfAvailable<DataLayout>();
68 TTI = &getAnalysis<TargetTransformInfo>();
69 AA = &getAnalysis<AliasAnalysis>();
70 LI = &getAnalysis<LoopInfo>();
75 // Must have DataLayout. We can't require it because some tests run w/o
80 for (Function::iterator it = F.begin(), e = F.end(); it != e; ++it) {
82 bool BBChanged = false;
84 // Use the bollom up slp vectorizer to construct chains that start with
85 // he store instructions.
86 BoUpSLP R(BB, SE, DL, TTI, AA);
88 // Vectorize trees that end at reductions.
89 BBChanged |= vectorizeReductions(BB, R);
91 // Vectorize trees that end at stores.
92 if (unsigned count = collectStores(BB, R)) {
93 DEBUG(dbgs()<<"SLP: Found " << count << " stores to vectorize.\n");
94 BBChanged |= vectorizeStoreChains(R);
97 // Try to hoist some of the scalarization code to the preheader.
98 if (BBChanged) hoistGatherSequence(LI, BB, R);
100 Changed |= BBChanged;
104 DEBUG(dbgs()<<"SLP: vectorized \""<<F.getName()<<"\"\n");
105 DEBUG(verifyFunction(F));
110 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
111 FunctionPass::getAnalysisUsage(AU);
112 AU.addRequired<ScalarEvolution>();
113 AU.addRequired<AliasAnalysis>();
114 AU.addRequired<TargetTransformInfo>();
115 AU.addRequired<LoopInfo>();
120 /// \brief Collect memory references and sort them according to their base
121 /// object. We sort the stores to their base objects to reduce the cost of the
122 /// quadratic search on the stores. TODO: We can further reduce this cost
123 /// if we flush the chain creation every time we run into a memory barrier.
124 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
126 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
127 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
129 /// \brief Try to vectorize a chain that may start at the operands of \V;
130 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
132 /// \brief Vectorize the stores that were collected in StoreRefs.
133 bool vectorizeStoreChains(BoUpSLP &R);
135 /// \brief Try to hoist gather sequences outside of the loop in cases where
136 /// all of the sources are loop invariant.
137 void hoistGatherSequence(LoopInfo *LI, BasicBlock *BB, BoUpSLP &R);
139 /// \brief Scan the basic block and look for reductions that may start a
140 /// vectorization chain.
141 bool vectorizeReductions(BasicBlock *BB, BoUpSLP &R);
144 StoreListMap StoreRefs;
147 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
150 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
151 StoreInst *SI = dyn_cast<StoreInst>(it);
155 // Check that the pointer points to scalars.
156 if (SI->getValueOperand()->getType()->isAggregateType())
159 // Find the base of the GEP.
160 Value *Ptr = SI->getPointerOperand();
161 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
162 Ptr = GEP->getPointerOperand();
164 // Save the store locations.
165 StoreRefs[Ptr].push_back(SI);
171 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
172 if (!A || !B) return false;
173 BoUpSLP::ValueList VL;
176 int Cost = R.getTreeCost(VL);
177 int ExtrCost = R.getScalarizationCost(VL);
178 DEBUG(dbgs()<<"SLP: Cost of pair:" << Cost <<
179 " Cost of extract:" << ExtrCost << ".\n");
180 if ((Cost+ExtrCost) >= -SLPCostThreshold) return false;
181 DEBUG(dbgs()<<"SLP: Vectorizing pair.\n");
182 R.vectorizeArith(VL);
186 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
187 if (!V) return false;
188 // Try to vectorize V.
189 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
192 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
193 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
195 if (B && B->hasOneUse()) {
196 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
197 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
198 if (tryToVectorizePair(A, B0, R)) {
202 if (tryToVectorizePair(A, B1, R)) {
209 if (A && A->hasOneUse()) {
210 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
211 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
212 if (tryToVectorizePair(A0, B, R)) {
216 if (tryToVectorizePair(A1, B, R)) {
224 bool SLPVectorizer::vectorizeReductions(BasicBlock *BB, BoUpSLP &R) {
225 bool Changed = false;
226 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
227 if (isa<DbgInfoIntrinsic>(it)) continue;
229 // Try to vectorize reductions that use PHINodes.
230 if (PHINode *P = dyn_cast<PHINode>(it)) {
231 // Check that the PHI is a reduction PHI.
232 if (P->getNumIncomingValues() != 2) return Changed;
233 Value *Rdx = (P->getIncomingBlock(0) == BB ? P->getIncomingValue(0) :
234 (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1) :
236 // Check if this is a Binary Operator.
237 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
241 Value *Inst = BI->getOperand(0);
242 if (Inst == P) Inst = BI->getOperand(1);
243 Changed |= tryToVectorize(dyn_cast<BinaryOperator>(Inst), R);
247 // Try to vectorize trees that start at compare instructions.
248 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
249 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
253 for (int i = 0; i < 2; ++i)
254 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i)))
255 Changed |= tryToVectorize(BI, R);
263 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
264 bool Changed = false;
265 // Attempt to sort and vectorize each of the store-groups.
266 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
268 if (it->second.size() < 2)
271 DEBUG(dbgs()<<"SLP: Analyzing a store chain of length " <<
272 it->second.size() << ".\n");
274 Changed |= R.vectorizeStores(it->second, -SLPCostThreshold);
279 void SLPVectorizer::hoistGatherSequence(LoopInfo *LI, BasicBlock *BB,
281 // Check if this block is inside a loop.
282 Loop *L = LI->getLoopFor(BB);
286 // Check if it has a preheader.
287 BasicBlock *PreHeader = L->getLoopPreheader();
291 // Mark the insertion point for the block.
292 Instruction *Location = PreHeader->getTerminator();
294 BoUpSLP::ValueList &Gathers = R.getGatherSeqInstructions();
295 for (BoUpSLP::ValueList::iterator it = Gathers.begin(), e = Gathers.end();
297 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
299 // The InsertElement sequence can be simplified into a constant.
303 // If the vector or the element that we insert into it are
304 // instructions that are defined in this basic block then we can't
305 // hoist this instruction.
306 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
307 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
308 if (CurrVec && L->contains(CurrVec)) continue;
309 if (NewElem && L->contains(NewElem)) continue;
311 // We can hoist this instruction. Move it to the pre-header.
312 Insert->moveBefore(Location);
316 } // end anonymous namespace
318 char SLPVectorizer::ID = 0;
319 static const char lv_name[] = "SLP Vectorizer";
320 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
321 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
322 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
323 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
324 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
325 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
328 Pass *createSLPVectorizerPass() {
329 return new SLPVectorizer();