1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
14 //
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
18 //
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
28 //
29 //===----------------------------------------------------------------------===//
30 //
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
33 //
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
36 //
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
39 //
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
42 //
43 //===----------------------------------------------------------------------===//
44
45 #include "llvm/Transforms/Vectorize.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/EquivalenceClasses.h"
48 #include "llvm/ADT/Hashing.h"
49 #include "llvm/ADT/MapVector.h"
50 #include "llvm/ADT/SetVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/ADT/StringExtras.h"
56 #include "llvm/Analysis/AliasAnalysis.h"
57 #include "llvm/Analysis/AliasSetTracker.h"
58 #include "llvm/Analysis/AssumptionCache.h"
59 #include "llvm/Analysis/BlockFrequencyInfo.h"
60 #include "llvm/Analysis/CodeMetrics.h"
61 #include "llvm/Analysis/LoopInfo.h"
62 #include "llvm/Analysis/LoopIterator.h"
63 #include "llvm/Analysis/LoopPass.h"
64 #include "llvm/Analysis/ScalarEvolution.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
67 #include "llvm/Analysis/TargetTransformInfo.h"
68 #include "llvm/Analysis/ValueTracking.h"
69 #include "llvm/IR/Constants.h"
70 #include "llvm/IR/DataLayout.h"
71 #include "llvm/IR/DebugInfo.h"
72 #include "llvm/IR/DerivedTypes.h"
73 #include "llvm/IR/DiagnosticInfo.h"
74 #include "llvm/IR/Dominators.h"
75 #include "llvm/IR/Function.h"
76 #include "llvm/IR/IRBuilder.h"
77 #include "llvm/IR/Instructions.h"
78 #include "llvm/IR/IntrinsicInst.h"
79 #include "llvm/IR/LLVMContext.h"
80 #include "llvm/IR/Module.h"
81 #include "llvm/IR/PatternMatch.h"
82 #include "llvm/IR/Type.h"
83 #include "llvm/IR/Value.h"
84 #include "llvm/IR/ValueHandle.h"
85 #include "llvm/IR/Verifier.h"
86 #include "llvm/Pass.h"
87 #include "llvm/Support/BranchProbability.h"
88 #include "llvm/Support/CommandLine.h"
89 #include "llvm/Support/Debug.h"
90 #include "llvm/Support/raw_ostream.h"
91 #include "llvm/Transforms/Scalar.h"
92 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
93 #include "llvm/Transforms/Utils/Local.h"
94 #include "llvm/Transforms/Utils/VectorUtils.h"
95 #include <algorithm>
96 #include <map>
97 #include <tuple>
98
99 using namespace llvm;
100 using namespace llvm::PatternMatch;
101
102 #define LV_NAME "loop-vectorize"
103 #define DEBUG_TYPE LV_NAME
104
105 STATISTIC(LoopsVectorized, "Number of loops vectorized");
106 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
107
108 static cl::opt<unsigned>
109 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
110 cl::desc("Sets the SIMD width. Zero is autoselect."));
111
112 static cl::opt<unsigned>
113 VectorizationInterleave("force-vector-interleave", cl::init(0), cl::Hidden,
114 cl::desc("Sets the vectorization interleave count. "
115 "Zero is autoselect."));
116
117 static cl::opt<bool>
118 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
119 cl::desc("Enable if-conversion during vectorization."));
120
121 /// We don't vectorize loops with a known constant trip count below this number.
122 static cl::opt<unsigned>
123 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
124 cl::Hidden,
125 cl::desc("Don't vectorize loops with a constant "
126 "trip count that is smaller than this "
127 "value."));
128
129 /// This enables versioning on the strides of symbolically striding memory
130 /// accesses in code like the following.
131 /// for (i = 0; i < N; ++i)
132 /// A[i * Stride1] += B[i * Stride2] ...
133 ///
134 /// Will be roughly translated to
135 /// if (Stride1 == 1 && Stride2 == 1) {
136 /// for (i = 0; i < N; i+=4)
137 /// A[i:i+3] += ...
138 /// } else
139 /// ...
140 static cl::opt<bool> EnableMemAccessVersioning(
141 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
142 cl::desc("Enable symblic stride memory access versioning"));
143
144 /// We don't unroll loops with a known constant trip count below this number.
145 static const unsigned TinyTripCountUnrollThreshold = 128;
146
147 /// When performing memory disambiguation checks at runtime do not make more
148 /// than this number of comparisons.
149 static const unsigned RuntimeMemoryCheckThreshold = 8;
150
151 /// Maximum simd width.
152 static const unsigned MaxVectorWidth = 64;
153
154 static cl::opt<unsigned> ForceTargetNumScalarRegs(
155 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
156 cl::desc("A flag that overrides the target's number of scalar registers."));
157
158 static cl::opt<unsigned> ForceTargetNumVectorRegs(
159 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
160 cl::desc("A flag that overrides the target's number of vector registers."));
161
162 /// Maximum vectorization interleave count.
163 static const unsigned MaxInterleaveFactor = 16;
164
165 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
166 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
167 cl::desc("A flag that overrides the target's max interleave factor for "
168 "scalar loops."));
169
170 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
171 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
172 cl::desc("A flag that overrides the target's max interleave factor for "
173 "vectorized loops."));
174
175 static cl::opt<unsigned> ForceTargetInstructionCost(
176 "force-target-instruction-cost", cl::init(0), cl::Hidden,
177 cl::desc("A flag that overrides the target's expected cost for "
178 "an instruction to a single constant value. Mostly "
179 "useful for getting consistent testing."));
180
181 static cl::opt<unsigned> SmallLoopCost(
182 "small-loop-cost", cl::init(20), cl::Hidden,
183 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
184
185 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
186 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
187 cl::desc("Enable the use of the block frequency analysis to access PGO "
188 "heuristics minimizing code growth in cold regions and being more "
189 "aggressive in hot regions."));
190
191 // Runtime unroll loops for load/store throughput.
192 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
193 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
194 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
195
196 /// The number of stores in a loop that are allowed to need predication.
197 static cl::opt<unsigned> NumberOfStoresToPredicate(
198 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
199 cl::desc("Max number of stores to be predicated behind an if."));
200
201 static cl::opt<bool> EnableIndVarRegisterHeur(
202 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
203 cl::desc("Count the induction variable only once when unrolling"));
204
205 static cl::opt<bool> EnableCondStoresVectorization(
206 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
207 cl::desc("Enable if predication of stores during vectorization."));
208
209 static cl::opt<unsigned> MaxNestedScalarReductionUF(
210 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
211 cl::desc("The maximum unroll factor to use when unrolling a scalar "
212 "reduction in a nested loop."));
213
214 namespace {
215
216 // Forward declarations.
217 class LoopVectorizationLegality;
218 class LoopVectorizationCostModel;
219 class LoopVectorizeHints;
220
221 /// Optimization analysis message produced during vectorization. Messages inform
222 /// the user why vectorization did not occur.
223 class Report {
224 std::string Message;
225 raw_string_ostream Out;
226 Instruction *Instr;
227
228 public:
Report(Instruction * I=nullptr)229 Report(Instruction *I = nullptr) : Out(Message), Instr(I) {
230 Out << "loop not vectorized: ";
231 }
232
operator <<(const A & Value)233 template <typename A> Report &operator<<(const A &Value) {
234 Out << Value;
235 return *this;
236 }
237
getInstr()238 Instruction *getInstr() { return Instr; }
239
str()240 std::string &str() { return Out.str(); }
operator Twine()241 operator Twine() { return Out.str(); }
242 };
243
244 /// InnerLoopVectorizer vectorizes loops which contain only one basic
245 /// block to a specified vectorization factor (VF).
246 /// This class performs the widening of scalars into vectors, or multiple
247 /// scalars. This class also implements the following features:
248 /// * It inserts an epilogue loop for handling loops that don't have iteration
249 /// counts that are known to be a multiple of the vectorization factor.
250 /// * It handles the code generation for reduction variables.
251 /// * Scalarization (implementation using scalars) of un-vectorizable
252 /// instructions.
253 /// InnerLoopVectorizer does not perform any vectorization-legality
254 /// checks, and relies on the caller to check for the different legality
255 /// aspects. The InnerLoopVectorizer relies on the
256 /// LoopVectorizationLegality class to provide information about the induction
257 /// and reduction variables that were found to a given vectorization factor.
258 class InnerLoopVectorizer {
259 public:
InnerLoopVectorizer(Loop * OrigLoop,ScalarEvolution * SE,LoopInfo * LI,DominatorTree * DT,const DataLayout * DL,const TargetLibraryInfo * TLI,unsigned VecWidth,unsigned UnrollFactor)260 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
261 DominatorTree *DT, const DataLayout *DL,
262 const TargetLibraryInfo *TLI, unsigned VecWidth,
263 unsigned UnrollFactor)
264 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
265 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
266 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
267 Legal(nullptr) {}
268
269 // Perform the actual loop widening (vectorization).
vectorize(LoopVectorizationLegality * L)270 void vectorize(LoopVectorizationLegality *L) {
271 Legal = L;
272 // Create a new empty loop. Unlink the old loop and connect the new one.
273 createEmptyLoop();
274 // Widen each instruction in the old loop to a new one in the new loop.
275 // Use the Legality module to find the induction and reduction variables.
276 vectorizeLoop();
277 // Register the new loop and update the analysis passes.
278 updateAnalysis();
279 }
280
~InnerLoopVectorizer()281 virtual ~InnerLoopVectorizer() {}
282
283 protected:
284 /// A small list of PHINodes.
285 typedef SmallVector<PHINode*, 4> PhiVector;
286 /// When we unroll loops we have multiple vector values for each scalar.
287 /// This data structure holds the unrolled and vectorized values that
288 /// originated from one scalar instruction.
289 typedef SmallVector<Value*, 2> VectorParts;
290
291 // When we if-convert we need create edge masks. We have to cache values so
292 // that we don't end up with exponential recursion/IR.
293 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
294 VectorParts> EdgeMaskCache;
295
296 /// \brief Add code that checks at runtime if the accessed arrays overlap.
297 ///
298 /// Returns a pair of instructions where the first element is the first
299 /// instruction generated in possibly a sequence of instructions and the
300 /// second value is the final comparator value or NULL if no check is needed.
301 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
302
303 /// \brief Add checks for strides that where assumed to be 1.
304 ///
305 /// Returns the last check instruction and the first check instruction in the
306 /// pair as (first, last).
307 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
308
309 /// Create an empty loop, based on the loop ranges of the old loop.
310 void createEmptyLoop();
311 /// Copy and widen the instructions from the old loop.
312 virtual void vectorizeLoop();
313
314 /// \brief The Loop exit block may have single value PHI nodes where the
315 /// incoming value is 'Undef'. While vectorizing we only handled real values
316 /// that were defined inside the loop. Here we fix the 'undef case'.
317 /// See PR14725.
318 void fixLCSSAPHIs();
319
320 /// A helper function that computes the predicate of the block BB, assuming
321 /// that the header block of the loop is set to True. It returns the *entry*
322 /// mask for the block BB.
323 VectorParts createBlockInMask(BasicBlock *BB);
324 /// A helper function that computes the predicate of the edge between SRC
325 /// and DST.
326 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
327
328 /// A helper function to vectorize a single BB within the innermost loop.
329 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
330
331 /// Vectorize a single PHINode in a block. This method handles the induction
332 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
333 /// arbitrary length vectors.
334 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
335 unsigned UF, unsigned VF, PhiVector *PV);
336
337 /// Insert the new loop to the loop hierarchy and pass manager
338 /// and update the analysis passes.
339 void updateAnalysis();
340
341 /// This instruction is un-vectorizable. Implement it as a sequence
342 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
343 /// scalarized instruction behind an if block predicated on the control
344 /// dependence of the instruction.
345 virtual void scalarizeInstruction(Instruction *Instr,
346 bool IfPredicateStore=false);
347
348 /// Vectorize Load and Store instructions,
349 virtual void vectorizeMemoryInstruction(Instruction *Instr);
350
351 /// Create a broadcast instruction. This method generates a broadcast
352 /// instruction (shuffle) for loop invariant values and for the induction
353 /// value. If this is the induction variable then we extend it to N, N+1, ...
354 /// this is needed because each iteration in the loop corresponds to a SIMD
355 /// element.
356 virtual Value *getBroadcastInstrs(Value *V);
357
358 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
359 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
360 /// The sequence starts at StartIndex.
361 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
362
363 /// When we go over instructions in the basic block we rely on previous
364 /// values within the current basic block or on loop invariant values.
365 /// When we widen (vectorize) values we place them in the map. If the values
366 /// are not within the map, they have to be loop invariant, so we simply
367 /// broadcast them into a vector.
368 VectorParts &getVectorValue(Value *V);
369
370 /// Generate a shuffle sequence that will reverse the vector Vec.
371 virtual Value *reverseVector(Value *Vec);
372
373 /// This is a helper class that holds the vectorizer state. It maps scalar
374 /// instructions to vector instructions. When the code is 'unrolled' then
375 /// then a single scalar value is mapped to multiple vector parts. The parts
376 /// are stored in the VectorPart type.
377 struct ValueMap {
378 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
379 /// are mapped.
ValueMap__anon03606d180111::InnerLoopVectorizer::ValueMap380 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
381
382 /// \return True if 'Key' is saved in the Value Map.
has__anon03606d180111::InnerLoopVectorizer::ValueMap383 bool has(Value *Key) const { return MapStorage.count(Key); }
384
385 /// Initializes a new entry in the map. Sets all of the vector parts to the
386 /// save value in 'Val'.
387 /// \return A reference to a vector with splat values.
splat__anon03606d180111::InnerLoopVectorizer::ValueMap388 VectorParts &splat(Value *Key, Value *Val) {
389 VectorParts &Entry = MapStorage[Key];
390 Entry.assign(UF, Val);
391 return Entry;
392 }
393
394 ///\return A reference to the value that is stored at 'Key'.
get__anon03606d180111::InnerLoopVectorizer::ValueMap395 VectorParts &get(Value *Key) {
396 VectorParts &Entry = MapStorage[Key];
397 if (Entry.empty())
398 Entry.resize(UF);
399 assert(Entry.size() == UF);
400 return Entry;
401 }
402
403 private:
404 /// The unroll factor. Each entry in the map stores this number of vector
405 /// elements.
406 unsigned UF;
407
408 /// Map storage. We use std::map and not DenseMap because insertions to a
409 /// dense map invalidates its iterators.
410 std::map<Value *, VectorParts> MapStorage;
411 };
412
413 /// The original loop.
414 Loop *OrigLoop;
415 /// Scev analysis to use.
416 ScalarEvolution *SE;
417 /// Loop Info.
418 LoopInfo *LI;
419 /// Dominator Tree.
420 DominatorTree *DT;
421 /// Alias Analysis.
422 AliasAnalysis *AA;
423 /// Data Layout.
424 const DataLayout *DL;
425 /// Target Library Info.
426 const TargetLibraryInfo *TLI;
427
428 /// The vectorization SIMD factor to use. Each vector will have this many
429 /// vector elements.
430 unsigned VF;
431
432 protected:
433 /// The vectorization unroll factor to use. Each scalar is vectorized to this
434 /// many different vector instructions.
435 unsigned UF;
436
437 /// The builder that we use
438 IRBuilder<> Builder;
439
440 // --- Vectorization state ---
441
442 /// The vector-loop preheader.
443 BasicBlock *LoopVectorPreHeader;
444 /// The scalar-loop preheader.
445 BasicBlock *LoopScalarPreHeader;
446 /// Middle Block between the vector and the scalar.
447 BasicBlock *LoopMiddleBlock;
448 ///The ExitBlock of the scalar loop.
449 BasicBlock *LoopExitBlock;
450 ///The vector loop body.
451 SmallVector<BasicBlock *, 4> LoopVectorBody;
452 ///The scalar loop body.
453 BasicBlock *LoopScalarBody;
454 /// A list of all bypass blocks. The first block is the entry of the loop.
455 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
456
457 /// The new Induction variable which was added to the new block.
458 PHINode *Induction;
459 /// The induction variable of the old basic block.
460 PHINode *OldInduction;
461 /// Holds the extended (to the widest induction type) start index.
462 Value *ExtendedIdx;
463 /// Maps scalars to widened vectors.
464 ValueMap WidenMap;
465 EdgeMaskCache MaskCache;
466
467 LoopVectorizationLegality *Legal;
468 };
469
470 class InnerLoopUnroller : public InnerLoopVectorizer {
471 public:
InnerLoopUnroller(Loop * OrigLoop,ScalarEvolution * SE,LoopInfo * LI,DominatorTree * DT,const DataLayout * DL,const TargetLibraryInfo * TLI,unsigned UnrollFactor)472 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
473 DominatorTree *DT, const DataLayout *DL,
474 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
475 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
476
477 private:
478 void scalarizeInstruction(Instruction *Instr,
479 bool IfPredicateStore = false) override;
480 void vectorizeMemoryInstruction(Instruction *Instr) override;
481 Value *getBroadcastInstrs(Value *V) override;
482 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
483 Value *reverseVector(Value *Vec) override;
484 };
485
486 /// \brief Look for a meaningful debug location on the instruction or it's
487 /// operands.
getDebugLocFromInstOrOperands(Instruction * I)488 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
489 if (!I)
490 return I;
491
492 DebugLoc Empty;
493 if (I->getDebugLoc() != Empty)
494 return I;
495
496 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
497 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
498 if (OpInst->getDebugLoc() != Empty)
499 return OpInst;
500 }
501
502 return I;
503 }
504
505 /// \brief Set the debug location in the builder using the debug location in the
506 /// instruction.
setDebugLocFromInst(IRBuilder<> & B,const Value * Ptr)507 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
508 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
509 B.SetCurrentDebugLocation(Inst->getDebugLoc());
510 else
511 B.SetCurrentDebugLocation(DebugLoc());
512 }
513
514 #ifndef NDEBUG
515 /// \return string containing a file name and a line # for the given loop.
getDebugLocString(const Loop * L)516 static std::string getDebugLocString(const Loop *L) {
517 std::string Result;
518 if (L) {
519 raw_string_ostream OS(Result);
520 const DebugLoc LoopDbgLoc = L->getStartLoc();
521 if (!LoopDbgLoc.isUnknown())
522 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
523 else
524 // Just print the module name.
525 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
526 OS.flush();
527 }
528 return Result;
529 }
530 #endif
531
532 /// \brief Propagate known metadata from one instruction to another.
propagateMetadata(Instruction * To,const Instruction * From)533 static void propagateMetadata(Instruction *To, const Instruction *From) {
534 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
535 From->getAllMetadataOtherThanDebugLoc(Metadata);
536
537 for (auto M : Metadata) {
538 unsigned Kind = M.first;
539
540 // These are safe to transfer (this is safe for TBAA, even when we
541 // if-convert, because should that metadata have had a control dependency
542 // on the condition, and thus actually aliased with some other
543 // non-speculated memory access when the condition was false, this would be
544 // caught by the runtime overlap checks).
545 if (Kind != LLVMContext::MD_tbaa &&
546 Kind != LLVMContext::MD_alias_scope &&
547 Kind != LLVMContext::MD_noalias &&
548 Kind != LLVMContext::MD_fpmath)
549 continue;
550
551 To->setMetadata(Kind, M.second);
552 }
553 }
554
555 /// \brief Propagate known metadata from one instruction to a vector of others.
propagateMetadata(SmallVectorImpl<Value * > & To,const Instruction * From)556 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
557 for (Value *V : To)
558 if (Instruction *I = dyn_cast<Instruction>(V))
559 propagateMetadata(I, From);
560 }
561
562 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
563 /// to what vectorization factor.
564 /// This class does not look at the profitability of vectorization, only the
565 /// legality. This class has two main kinds of checks:
566 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
567 /// will change the order of memory accesses in a way that will change the
568 /// correctness of the program.
569 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
570 /// checks for a number of different conditions, such as the availability of a
571 /// single induction variable, that all types are supported and vectorize-able,
572 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
573 /// This class is also used by InnerLoopVectorizer for identifying
574 /// induction variable and the different reduction variables.
575 class LoopVectorizationLegality {
576 public:
577 unsigned NumLoads;
578 unsigned NumStores;
579 unsigned NumPredStores;
580
LoopVectorizationLegality(Loop * L,ScalarEvolution * SE,const DataLayout * DL,DominatorTree * DT,TargetLibraryInfo * TLI,AliasAnalysis * AA,Function * F,const TargetTransformInfo * TTI)581 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
582 DominatorTree *DT, TargetLibraryInfo *TLI,
583 AliasAnalysis *AA, Function *F,
584 const TargetTransformInfo *TTI)
585 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
586 DT(DT), TLI(TLI), AA(AA), TheFunction(F), TTI(TTI), Induction(nullptr),
587 WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
588 }
589
590 /// This enum represents the kinds of reductions that we support.
591 enum ReductionKind {
592 RK_NoReduction, ///< Not a reduction.
593 RK_IntegerAdd, ///< Sum of integers.
594 RK_IntegerMult, ///< Product of integers.
595 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
596 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
597 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
598 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
599 RK_FloatAdd, ///< Sum of floats.
600 RK_FloatMult, ///< Product of floats.
601 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
602 };
603
604 /// This enum represents the kinds of inductions that we support.
605 enum InductionKind {
606 IK_NoInduction, ///< Not an induction variable.
607 IK_IntInduction, ///< Integer induction variable. Step = 1.
608 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
609 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
610 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
611 };
612
613 // This enum represents the kind of minmax reduction.
614 enum MinMaxReductionKind {
615 MRK_Invalid,
616 MRK_UIntMin,
617 MRK_UIntMax,
618 MRK_SIntMin,
619 MRK_SIntMax,
620 MRK_FloatMin,
621 MRK_FloatMax
622 };
623
624 /// This struct holds information about reduction variables.
625 struct ReductionDescriptor {
ReductionDescriptor__anon03606d180111::LoopVectorizationLegality::ReductionDescriptor626 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
627 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
628
ReductionDescriptor__anon03606d180111::LoopVectorizationLegality::ReductionDescriptor629 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
630 MinMaxReductionKind MK)
631 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
632
633 // The starting value of the reduction.
634 // It does not have to be zero!
635 TrackingVH<Value> StartValue;
636 // The instruction who's value is used outside the loop.
637 Instruction *LoopExitInstr;
638 // The kind of the reduction.
639 ReductionKind Kind;
640 // If this a min/max reduction the kind of reduction.
641 MinMaxReductionKind MinMaxKind;
642 };
643
644 /// This POD struct holds information about a potential reduction operation.
645 struct ReductionInstDesc {
ReductionInstDesc__anon03606d180111::LoopVectorizationLegality::ReductionInstDesc646 ReductionInstDesc(bool IsRedux, Instruction *I) :
647 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
648
ReductionInstDesc__anon03606d180111::LoopVectorizationLegality::ReductionInstDesc649 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
650 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
651
652 // Is this instruction a reduction candidate.
653 bool IsReduction;
654 // The last instruction in a min/max pattern (select of the select(icmp())
655 // pattern), or the current reduction instruction otherwise.
656 Instruction *PatternLastInst;
657 // If this is a min/max pattern the comparison predicate.
658 MinMaxReductionKind MinMaxKind;
659 };
660
661 /// This struct holds information about the memory runtime legality
662 /// check that a group of pointers do not overlap.
663 struct RuntimePointerCheck {
RuntimePointerCheck__anon03606d180111::LoopVectorizationLegality::RuntimePointerCheck664 RuntimePointerCheck() : Need(false) {}
665
666 /// Reset the state of the pointer runtime information.
reset__anon03606d180111::LoopVectorizationLegality::RuntimePointerCheck667 void reset() {
668 Need = false;
669 Pointers.clear();
670 Starts.clear();
671 Ends.clear();
672 IsWritePtr.clear();
673 DependencySetId.clear();
674 AliasSetId.clear();
675 }
676
677 /// Insert a pointer and calculate the start and end SCEVs.
678 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
679 unsigned DepSetId, unsigned ASId, ValueToValueMap &Strides);
680
681 /// This flag indicates if we need to add the runtime check.
682 bool Need;
683 /// Holds the pointers that we need to check.
684 SmallVector<TrackingVH<Value>, 2> Pointers;
685 /// Holds the pointer value at the beginning of the loop.
686 SmallVector<const SCEV*, 2> Starts;
687 /// Holds the pointer value at the end of the loop.
688 SmallVector<const SCEV*, 2> Ends;
689 /// Holds the information if this pointer is used for writing to memory.
690 SmallVector<bool, 2> IsWritePtr;
691 /// Holds the id of the set of pointers that could be dependent because of a
692 /// shared underlying object.
693 SmallVector<unsigned, 2> DependencySetId;
694 /// Holds the id of the disjoint alias set to which this pointer belongs.
695 SmallVector<unsigned, 2> AliasSetId;
696 };
697
698 /// A struct for saving information about induction variables.
699 struct InductionInfo {
InductionInfo__anon03606d180111::LoopVectorizationLegality::InductionInfo700 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
InductionInfo__anon03606d180111::LoopVectorizationLegality::InductionInfo701 InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
702 /// Start value.
703 TrackingVH<Value> StartValue;
704 /// Induction kind.
705 InductionKind IK;
706 };
707
708 /// ReductionList contains the reduction descriptors for all
709 /// of the reductions that were found in the loop.
710 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
711
712 /// InductionList saves induction variables and maps them to the
713 /// induction descriptor.
714 typedef MapVector<PHINode*, InductionInfo> InductionList;
715
716 /// Returns true if it is legal to vectorize this loop.
717 /// This does not mean that it is profitable to vectorize this
718 /// loop, only that it is legal to do so.
719 bool canVectorize();
720
721 /// Returns the Induction variable.
getInduction()722 PHINode *getInduction() { return Induction; }
723
724 /// Returns the reduction variables found in the loop.
getReductionVars()725 ReductionList *getReductionVars() { return &Reductions; }
726
727 /// Returns the induction variables found in the loop.
getInductionVars()728 InductionList *getInductionVars() { return &Inductions; }
729
730 /// Returns the widest induction type.
getWidestInductionType()731 Type *getWidestInductionType() { return WidestIndTy; }
732
733 /// Returns True if V is an induction variable in this loop.
734 bool isInductionVariable(const Value *V);
735
736 /// Return true if the block BB needs to be predicated in order for the loop
737 /// to be vectorized.
738 bool blockNeedsPredication(BasicBlock *BB);
739
740 /// Check if this pointer is consecutive when vectorizing. This happens
741 /// when the last index of the GEP is the induction variable, or that the
742 /// pointer itself is an induction variable.
743 /// This check allows us to vectorize A[idx] into a wide load/store.
744 /// Returns:
745 /// 0 - Stride is unknown or non-consecutive.
746 /// 1 - Address is consecutive.
747 /// -1 - Address is consecutive, and decreasing.
748 int isConsecutivePtr(Value *Ptr);
749
750 /// Returns true if the value V is uniform within the loop.
751 bool isUniform(Value *V);
752
753 /// Returns true if this instruction will remain scalar after vectorization.
isUniformAfterVectorization(Instruction * I)754 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
755
756 /// Returns the information that we collected about runtime memory check.
getRuntimePointerCheck()757 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
758
759 /// This function returns the identity element (or neutral element) for
760 /// the operation K.
761 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
762
getMaxSafeDepDistBytes()763 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
764
hasStride(Value * V)765 bool hasStride(Value *V) { return StrideSet.count(V); }
mustCheckStrides()766 bool mustCheckStrides() { return !StrideSet.empty(); }
strides_begin()767 SmallPtrSet<Value *, 8>::iterator strides_begin() {
768 return StrideSet.begin();
769 }
strides_end()770 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
771
772 /// Returns true if the target machine supports masked store operation
773 /// for the given \p DataType and kind of access to \p Ptr.
isLegalMaskedStore(Type * DataType,Value * Ptr)774 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
775 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
776 }
777 /// Returns true if the target machine supports masked load operation
778 /// for the given \p DataType and kind of access to \p Ptr.
isLegalMaskedLoad(Type * DataType,Value * Ptr)779 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
780 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
781 }
782 /// Returns true if vector representation of the instruction \p I
783 /// requires mask.
isMaskRequired(const Instruction * I)784 bool isMaskRequired(const Instruction* I) {
785 return (MaskedOp.count(I) != 0);
786 }
787 private:
788 /// Check if a single basic block loop is vectorizable.
789 /// At this point we know that this is a loop with a constant trip count
790 /// and we only need to check individual instructions.
791 bool canVectorizeInstrs();
792
793 /// When we vectorize loops we may change the order in which
794 /// we read and write from memory. This method checks if it is
795 /// legal to vectorize the code, considering only memory constrains.
796 /// Returns true if the loop is vectorizable
797 bool canVectorizeMemory();
798
799 /// Return true if we can vectorize this loop using the IF-conversion
800 /// transformation.
801 bool canVectorizeWithIfConvert();
802
803 /// Collect the variables that need to stay uniform after vectorization.
804 void collectLoopUniforms();
805
806 /// Return true if all of the instructions in the block can be speculatively
807 /// executed. \p SafePtrs is a list of addresses that are known to be legal
808 /// and we know that we can read from them without segfault.
809 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
810
811 /// Returns True, if 'Phi' is the kind of reduction variable for type
812 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
813 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
814 /// Returns a struct describing if the instruction 'I' can be a reduction
815 /// variable of type 'Kind'. If the reduction is a min/max pattern of
816 /// select(icmp()) this function advances the instruction pointer 'I' from the
817 /// compare instruction to the select instruction and stores this pointer in
818 /// 'PatternLastInst' member of the returned struct.
819 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
820 ReductionInstDesc &Desc);
821 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
822 /// pattern corresponding to a min(X, Y) or max(X, Y).
823 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
824 ReductionInstDesc &Prev);
825 /// Returns the induction kind of Phi. This function may return NoInduction
826 /// if the PHI is not an induction variable.
827 InductionKind isInductionVariable(PHINode *Phi);
828
829 /// \brief Collect memory access with loop invariant strides.
830 ///
831 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
832 /// invariant.
833 void collectStridedAccess(Value *LoadOrStoreInst);
834
835 /// Report an analysis message to assist the user in diagnosing loops that are
836 /// not vectorized.
emitAnalysis(Report & Message)837 void emitAnalysis(Report &Message) {
838 DebugLoc DL = TheLoop->getStartLoc();
839 if (Instruction *I = Message.getInstr())
840 DL = I->getDebugLoc();
841 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
842 *TheFunction, DL, Message.str());
843 }
844
845 /// The loop that we evaluate.
846 Loop *TheLoop;
847 /// Scev analysis.
848 ScalarEvolution *SE;
849 /// DataLayout analysis.
850 const DataLayout *DL;
851 /// Dominators.
852 DominatorTree *DT;
853 /// Target Library Info.
854 TargetLibraryInfo *TLI;
855 /// Alias analysis.
856 AliasAnalysis *AA;
857 /// Parent function
858 Function *TheFunction;
859 /// Target Transform Info
860 const TargetTransformInfo *TTI;
861
862 // --- vectorization state --- //
863
864 /// Holds the integer induction variable. This is the counter of the
865 /// loop.
866 PHINode *Induction;
867 /// Holds the reduction variables.
868 ReductionList Reductions;
869 /// Holds all of the induction variables that we found in the loop.
870 /// Notice that inductions don't need to start at zero and that induction
871 /// variables can be pointers.
872 InductionList Inductions;
873 /// Holds the widest induction type encountered.
874 Type *WidestIndTy;
875
876 /// Allowed outside users. This holds the reduction
877 /// vars which can be accessed from outside the loop.
878 SmallPtrSet<Value*, 4> AllowedExit;
879 /// This set holds the variables which are known to be uniform after
880 /// vectorization.
881 SmallPtrSet<Instruction*, 4> Uniforms;
882 /// We need to check that all of the pointers in this list are disjoint
883 /// at runtime.
884 RuntimePointerCheck PtrRtCheck;
885 /// Can we assume the absence of NaNs.
886 bool HasFunNoNaNAttr;
887
888 unsigned MaxSafeDepDistBytes;
889
890 ValueToValueMap Strides;
891 SmallPtrSet<Value *, 8> StrideSet;
892
893 /// While vectorizing these instructions we have to generate a
894 /// call to the appropriate masked intrinsic
895 SmallPtrSet<const Instruction*, 8> MaskedOp;
896 };
897
898 /// LoopVectorizationCostModel - estimates the expected speedups due to
899 /// vectorization.
900 /// In many cases vectorization is not profitable. This can happen because of
901 /// a number of reasons. In this class we mainly attempt to predict the
902 /// expected speedup/slowdowns due to the supported instruction set. We use the
903 /// TargetTransformInfo to query the different backends for the cost of
904 /// different operations.
905 class LoopVectorizationCostModel {
906 public:
LoopVectorizationCostModel(Loop * L,ScalarEvolution * SE,LoopInfo * LI,LoopVectorizationLegality * Legal,const TargetTransformInfo & TTI,const DataLayout * DL,const TargetLibraryInfo * TLI,AssumptionCache * AC,const Function * F,const LoopVectorizeHints * Hints)907 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
908 LoopVectorizationLegality *Legal,
909 const TargetTransformInfo &TTI,
910 const DataLayout *DL, const TargetLibraryInfo *TLI,
911 AssumptionCache *AC, const Function *F,
912 const LoopVectorizeHints *Hints)
913 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
914 TheFunction(F), Hints(Hints) {
915 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
916 }
917
918 /// Information about vectorization costs
919 struct VectorizationFactor {
920 unsigned Width; // Vector width with best cost
921 unsigned Cost; // Cost of the loop with that width
922 };
923 /// \return The most profitable vectorization factor and the cost of that VF.
924 /// This method checks every power of two up to VF. If UserVF is not ZERO
925 /// then this vectorization factor will be selected if vectorization is
926 /// possible.
927 VectorizationFactor selectVectorizationFactor(bool OptForSize);
928
929 /// \return The size (in bits) of the widest type in the code that
930 /// needs to be vectorized. We ignore values that remain scalar such as
931 /// 64 bit loop indices.
932 unsigned getWidestType();
933
934 /// \return The most profitable unroll factor.
935 /// If UserUF is non-zero then this method finds the best unroll-factor
936 /// based on register pressure and other parameters.
937 /// VF and LoopCost are the selected vectorization factor and the cost of the
938 /// selected VF.
939 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
940
941 /// \brief A struct that represents some properties of the register usage
942 /// of a loop.
943 struct RegisterUsage {
944 /// Holds the number of loop invariant values that are used in the loop.
945 unsigned LoopInvariantRegs;
946 /// Holds the maximum number of concurrent live intervals in the loop.
947 unsigned MaxLocalUsers;
948 /// Holds the number of instructions in the loop.
949 unsigned NumInstructions;
950 };
951
952 /// \return information about the register usage of the loop.
953 RegisterUsage calculateRegisterUsage();
954
955 private:
956 /// Returns the expected execution cost. The unit of the cost does
957 /// not matter because we use the 'cost' units to compare different
958 /// vector widths. The cost that is returned is *not* normalized by
959 /// the factor width.
960 unsigned expectedCost(unsigned VF);
961
962 /// Returns the execution time cost of an instruction for a given vector
963 /// width. Vector width of one means scalar.
964 unsigned getInstructionCost(Instruction *I, unsigned VF);
965
966 /// A helper function for converting Scalar types to vector types.
967 /// If the incoming type is void, we return void. If the VF is 1, we return
968 /// the scalar type.
969 static Type* ToVectorTy(Type *Scalar, unsigned VF);
970
971 /// Returns whether the instruction is a load or store and will be a emitted
972 /// as a vector operation.
973 bool isConsecutiveLoadOrStore(Instruction *I);
974
975 /// Report an analysis message to assist the user in diagnosing loops that are
976 /// not vectorized.
emitAnalysis(Report & Message)977 void emitAnalysis(Report &Message) {
978 DebugLoc DL = TheLoop->getStartLoc();
979 if (Instruction *I = Message.getInstr())
980 DL = I->getDebugLoc();
981 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
982 *TheFunction, DL, Message.str());
983 }
984
985 /// Values used only by @llvm.assume calls.
986 SmallPtrSet<const Value *, 32> EphValues;
987
988 /// The loop that we evaluate.
989 Loop *TheLoop;
990 /// Scev analysis.
991 ScalarEvolution *SE;
992 /// Loop Info analysis.
993 LoopInfo *LI;
994 /// Vectorization legality.
995 LoopVectorizationLegality *Legal;
996 /// Vector target information.
997 const TargetTransformInfo &TTI;
998 /// Target data layout information.
999 const DataLayout *DL;
1000 /// Target Library Info.
1001 const TargetLibraryInfo *TLI;
1002 const Function *TheFunction;
1003 // Loop Vectorize Hint.
1004 const LoopVectorizeHints *Hints;
1005 };
1006
1007 /// Utility class for getting and setting loop vectorizer hints in the form
1008 /// of loop metadata.
1009 /// This class keeps a number of loop annotations locally (as member variables)
1010 /// and can, upon request, write them back as metadata on the loop. It will
1011 /// initially scan the loop for existing metadata, and will update the local
1012 /// values based on information in the loop.
1013 /// We cannot write all values to metadata, as the mere presence of some info,
1014 /// for example 'force', means a decision has been made. So, we need to be
1015 /// careful NOT to add them if the user hasn't specifically asked so.
1016 class LoopVectorizeHints {
1017 enum HintKind {
1018 HK_WIDTH,
1019 HK_UNROLL,
1020 HK_FORCE
1021 };
1022
1023 /// Hint - associates name and validation with the hint value.
1024 struct Hint {
1025 const char * Name;
1026 unsigned Value; // This may have to change for non-numeric values.
1027 HintKind Kind;
1028
Hint__anon03606d180111::LoopVectorizeHints::Hint1029 Hint(const char * Name, unsigned Value, HintKind Kind)
1030 : Name(Name), Value(Value), Kind(Kind) { }
1031
validate__anon03606d180111::LoopVectorizeHints::Hint1032 bool validate(unsigned Val) {
1033 switch (Kind) {
1034 case HK_WIDTH:
1035 return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
1036 case HK_UNROLL:
1037 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1038 case HK_FORCE:
1039 return (Val <= 1);
1040 }
1041 return false;
1042 }
1043 };
1044
1045 /// Vectorization width.
1046 Hint Width;
1047 /// Vectorization interleave factor.
1048 Hint Interleave;
1049 /// Vectorization forced
1050 Hint Force;
1051
1052 /// Return the loop metadata prefix.
Prefix()1053 static StringRef Prefix() { return "llvm.loop."; }
1054
1055 public:
1056 enum ForceKind {
1057 FK_Undefined = -1, ///< Not selected.
1058 FK_Disabled = 0, ///< Forcing disabled.
1059 FK_Enabled = 1, ///< Forcing enabled.
1060 };
1061
LoopVectorizeHints(const Loop * L,bool DisableInterleaving)1062 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1063 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1064 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1065 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1066 TheLoop(L) {
1067 // Populate values with existing loop metadata.
1068 getHintsFromMetadata();
1069
1070 // force-vector-interleave overrides DisableInterleaving.
1071 if (VectorizationInterleave.getNumOccurrences() > 0)
1072 Interleave.Value = VectorizationInterleave;
1073
1074 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1075 << "LV: Interleaving disabled by the pass manager\n");
1076 }
1077
1078 /// Mark the loop L as already vectorized by setting the width to 1.
setAlreadyVectorized()1079 void setAlreadyVectorized() {
1080 Width.Value = Interleave.Value = 1;
1081 Hint Hints[] = {Width, Interleave};
1082 writeHintsToMetadata(Hints);
1083 }
1084
1085 /// Dumps all the hint information.
emitRemark() const1086 std::string emitRemark() const {
1087 Report R;
1088 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1089 R << "vectorization is explicitly disabled";
1090 else {
1091 R << "use -Rpass-analysis=loop-vectorize for more info";
1092 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1093 R << " (Force=true";
1094 if (Width.Value != 0)
1095 R << ", Vector Width=" << Width.Value;
1096 if (Interleave.Value != 0)
1097 R << ", Interleave Count=" << Interleave.Value;
1098 R << ")";
1099 }
1100 }
1101
1102 return R.str();
1103 }
1104
getWidth() const1105 unsigned getWidth() const { return Width.Value; }
getInterleave() const1106 unsigned getInterleave() const { return Interleave.Value; }
getForce() const1107 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1108
1109 private:
1110 /// Find hints specified in the loop metadata and update local values.
getHintsFromMetadata()1111 void getHintsFromMetadata() {
1112 MDNode *LoopID = TheLoop->getLoopID();
1113 if (!LoopID)
1114 return;
1115
1116 // First operand should refer to the loop id itself.
1117 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1118 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1119
1120 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1121 const MDString *S = nullptr;
1122 SmallVector<Metadata *, 4> Args;
1123
1124 // The expected hint is either a MDString or a MDNode with the first
1125 // operand a MDString.
1126 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1127 if (!MD || MD->getNumOperands() == 0)
1128 continue;
1129 S = dyn_cast<MDString>(MD->getOperand(0));
1130 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1131 Args.push_back(MD->getOperand(i));
1132 } else {
1133 S = dyn_cast<MDString>(LoopID->getOperand(i));
1134 assert(Args.size() == 0 && "too many arguments for MDString");
1135 }
1136
1137 if (!S)
1138 continue;
1139
1140 // Check if the hint starts with the loop metadata prefix.
1141 StringRef Name = S->getString();
1142 if (Args.size() == 1)
1143 setHint(Name, Args[0]);
1144 }
1145 }
1146
1147 /// Checks string hint with one operand and set value if valid.
setHint(StringRef Name,Metadata * Arg)1148 void setHint(StringRef Name, Metadata *Arg) {
1149 if (!Name.startswith(Prefix()))
1150 return;
1151 Name = Name.substr(Prefix().size(), StringRef::npos);
1152
1153 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1154 if (!C) return;
1155 unsigned Val = C->getZExtValue();
1156
1157 Hint *Hints[] = {&Width, &Interleave, &Force};
1158 for (auto H : Hints) {
1159 if (Name == H->Name) {
1160 if (H->validate(Val))
1161 H->Value = Val;
1162 else
1163 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1164 break;
1165 }
1166 }
1167 }
1168
1169 /// Create a new hint from name / value pair.
createHintMetadata(StringRef Name,unsigned V) const1170 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1171 LLVMContext &Context = TheLoop->getHeader()->getContext();
1172 Metadata *MDs[] = {MDString::get(Context, Name),
1173 ConstantAsMetadata::get(
1174 ConstantInt::get(Type::getInt32Ty(Context), V))};
1175 return MDNode::get(Context, MDs);
1176 }
1177
1178 /// Matches metadata with hint name.
matchesHintMetadataName(MDNode * Node,ArrayRef<Hint> HintTypes)1179 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1180 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1181 if (!Name)
1182 return false;
1183
1184 for (auto H : HintTypes)
1185 if (Name->getString().endswith(H.Name))
1186 return true;
1187 return false;
1188 }
1189
1190 /// Sets current hints into loop metadata, keeping other values intact.
writeHintsToMetadata(ArrayRef<Hint> HintTypes)1191 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1192 if (HintTypes.size() == 0)
1193 return;
1194
1195 // Reserve the first element to LoopID (see below).
1196 SmallVector<Metadata *, 4> MDs(1);
1197 // If the loop already has metadata, then ignore the existing operands.
1198 MDNode *LoopID = TheLoop->getLoopID();
1199 if (LoopID) {
1200 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1201 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1202 // If node in update list, ignore old value.
1203 if (!matchesHintMetadataName(Node, HintTypes))
1204 MDs.push_back(Node);
1205 }
1206 }
1207
1208 // Now, add the missing hints.
1209 for (auto H : HintTypes)
1210 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1211
1212 // Replace current metadata node with new one.
1213 LLVMContext &Context = TheLoop->getHeader()->getContext();
1214 MDNode *NewLoopID = MDNode::get(Context, MDs);
1215 // Set operand 0 to refer to the loop id itself.
1216 NewLoopID->replaceOperandWith(0, NewLoopID);
1217
1218 TheLoop->setLoopID(NewLoopID);
1219 }
1220
1221 /// The loop these hints belong to.
1222 const Loop *TheLoop;
1223 };
1224
emitMissedWarning(Function * F,Loop * L,const LoopVectorizeHints & LH)1225 static void emitMissedWarning(Function *F, Loop *L,
1226 const LoopVectorizeHints &LH) {
1227 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1228 L->getStartLoc(), LH.emitRemark());
1229
1230 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1231 if (LH.getWidth() != 1)
1232 emitLoopVectorizeWarning(
1233 F->getContext(), *F, L->getStartLoc(),
1234 "failed explicitly specified loop vectorization");
1235 else if (LH.getInterleave() != 1)
1236 emitLoopInterleaveWarning(
1237 F->getContext(), *F, L->getStartLoc(),
1238 "failed explicitly specified loop interleaving");
1239 }
1240 }
1241
addInnerLoop(Loop & L,SmallVectorImpl<Loop * > & V)1242 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1243 if (L.empty())
1244 return V.push_back(&L);
1245
1246 for (Loop *InnerL : L)
1247 addInnerLoop(*InnerL, V);
1248 }
1249
1250 /// The LoopVectorize Pass.
1251 struct LoopVectorize : public FunctionPass {
1252 /// Pass identification, replacement for typeid
1253 static char ID;
1254
LoopVectorize__anon03606d180111::LoopVectorize1255 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1256 : FunctionPass(ID),
1257 DisableUnrolling(NoUnrolling),
1258 AlwaysVectorize(AlwaysVectorize) {
1259 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1260 }
1261
1262 ScalarEvolution *SE;
1263 const DataLayout *DL;
1264 LoopInfo *LI;
1265 TargetTransformInfo *TTI;
1266 DominatorTree *DT;
1267 BlockFrequencyInfo *BFI;
1268 TargetLibraryInfo *TLI;
1269 AliasAnalysis *AA;
1270 AssumptionCache *AC;
1271 bool DisableUnrolling;
1272 bool AlwaysVectorize;
1273
1274 BlockFrequency ColdEntryFreq;
1275
runOnFunction__anon03606d180111::LoopVectorize1276 bool runOnFunction(Function &F) override {
1277 SE = &getAnalysis<ScalarEvolution>();
1278 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1279 DL = DLP ? &DLP->getDataLayout() : nullptr;
1280 LI = &getAnalysis<LoopInfo>();
1281 TTI = &getAnalysis<TargetTransformInfo>();
1282 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1283 BFI = &getAnalysis<BlockFrequencyInfo>();
1284 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1285 AA = &getAnalysis<AliasAnalysis>();
1286 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1287
1288 // Compute some weights outside of the loop over the loops. Compute this
1289 // using a BranchProbability to re-use its scaling math.
1290 const BranchProbability ColdProb(1, 5); // 20%
1291 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1292
1293 // If the target claims to have no vector registers don't attempt
1294 // vectorization.
1295 if (!TTI->getNumberOfRegisters(true))
1296 return false;
1297
1298 if (!DL) {
1299 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1300 << ": Missing data layout\n");
1301 return false;
1302 }
1303
1304 // Build up a worklist of inner-loops to vectorize. This is necessary as
1305 // the act of vectorizing or partially unrolling a loop creates new loops
1306 // and can invalidate iterators across the loops.
1307 SmallVector<Loop *, 8> Worklist;
1308
1309 for (Loop *L : *LI)
1310 addInnerLoop(*L, Worklist);
1311
1312 LoopsAnalyzed += Worklist.size();
1313
1314 // Now walk the identified inner loops.
1315 bool Changed = false;
1316 while (!Worklist.empty())
1317 Changed |= processLoop(Worklist.pop_back_val());
1318
1319 // Process each loop nest in the function.
1320 return Changed;
1321 }
1322
processLoop__anon03606d180111::LoopVectorize1323 bool processLoop(Loop *L) {
1324 assert(L->empty() && "Only process inner loops.");
1325
1326 #ifndef NDEBUG
1327 const std::string DebugLocStr = getDebugLocString(L);
1328 #endif /* NDEBUG */
1329
1330 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1331 << L->getHeader()->getParent()->getName() << "\" from "
1332 << DebugLocStr << "\n");
1333
1334 LoopVectorizeHints Hints(L, DisableUnrolling);
1335
1336 DEBUG(dbgs() << "LV: Loop hints:"
1337 << " force="
1338 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1339 ? "disabled"
1340 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1341 ? "enabled"
1342 : "?")) << " width=" << Hints.getWidth()
1343 << " unroll=" << Hints.getInterleave() << "\n");
1344
1345 // Function containing loop
1346 Function *F = L->getHeader()->getParent();
1347
1348 // Looking at the diagnostic output is the only way to determine if a loop
1349 // was vectorized (other than looking at the IR or machine code), so it
1350 // is important to generate an optimization remark for each loop. Most of
1351 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1352 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1353 // less verbose reporting vectorized loops and unvectorized loops that may
1354 // benefit from vectorization, respectively.
1355
1356 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1357 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1358 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1359 L->getStartLoc(), Hints.emitRemark());
1360 return false;
1361 }
1362
1363 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1364 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1365 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1366 L->getStartLoc(), Hints.emitRemark());
1367 return false;
1368 }
1369
1370 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1371 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1372 emitOptimizationRemarkAnalysis(
1373 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1374 "loop not vectorized: vector width and interleave count are "
1375 "explicitly set to 1");
1376 return false;
1377 }
1378
1379 // Check the loop for a trip count threshold:
1380 // do not vectorize loops with a tiny trip count.
1381 const unsigned TC = SE->getSmallConstantTripCount(L);
1382 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1383 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1384 << "This loop is not worth vectorizing.");
1385 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1386 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1387 else {
1388 DEBUG(dbgs() << "\n");
1389 emitOptimizationRemarkAnalysis(
1390 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1391 "vectorization is not beneficial and is not explicitly forced");
1392 return false;
1393 }
1394 }
1395
1396 // Check if it is legal to vectorize the loop.
1397 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI);
1398 if (!LVL.canVectorize()) {
1399 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1400 emitMissedWarning(F, L, Hints);
1401 return false;
1402 }
1403
1404 // Use the cost model.
1405 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
1406 &Hints);
1407
1408 // Check the function attributes to find out if this function should be
1409 // optimized for size.
1410 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1411 F->hasFnAttribute(Attribute::OptimizeForSize);
1412
1413 // Compute the weighted frequency of this loop being executed and see if it
1414 // is less than 20% of the function entry baseline frequency. Note that we
1415 // always have a canonical loop here because we think we *can* vectoriez.
1416 // FIXME: This is hidden behind a flag due to pervasive problems with
1417 // exactly what block frequency models.
1418 if (LoopVectorizeWithBlockFrequency) {
1419 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1420 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1421 LoopEntryFreq < ColdEntryFreq)
1422 OptForSize = true;
1423 }
1424
1425 // Check the function attributes to see if implicit floats are allowed.a
1426 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1427 // an integer loop and the vector instructions selected are purely integer
1428 // vector instructions?
1429 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1430 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1431 "attribute is used.\n");
1432 emitOptimizationRemarkAnalysis(
1433 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1434 "loop not vectorized due to NoImplicitFloat attribute");
1435 emitMissedWarning(F, L, Hints);
1436 return false;
1437 }
1438
1439 // Select the optimal vectorization factor.
1440 const LoopVectorizationCostModel::VectorizationFactor VF =
1441 CM.selectVectorizationFactor(OptForSize);
1442
1443 // Select the unroll factor.
1444 const unsigned UF =
1445 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1446
1447 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1448 << DebugLocStr << '\n');
1449 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1450
1451 if (VF.Width == 1) {
1452 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1453
1454 if (UF == 1) {
1455 emitOptimizationRemarkAnalysis(
1456 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1457 "not beneficial to vectorize and user disabled interleaving");
1458 return false;
1459 }
1460 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1461
1462 // Report the unrolling decision.
1463 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1464 Twine("unrolled with interleaving factor " +
1465 Twine(UF) +
1466 " (vectorization not beneficial)"));
1467
1468 // We decided not to vectorize, but we may want to unroll.
1469
1470 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1471 Unroller.vectorize(&LVL);
1472 } else {
1473 // If we decided that it is *legal* to vectorize the loop then do it.
1474 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1475 LB.vectorize(&LVL);
1476 ++LoopsVectorized;
1477
1478 // Report the vectorization decision.
1479 emitOptimizationRemark(
1480 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1481 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1482 ", unrolling interleave factor: " + Twine(UF) + ")");
1483 }
1484
1485 // Mark the loop as already vectorized to avoid vectorizing again.
1486 Hints.setAlreadyVectorized();
1487
1488 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1489 return true;
1490 }
1491
getAnalysisUsage__anon03606d180111::LoopVectorize1492 void getAnalysisUsage(AnalysisUsage &AU) const override {
1493 AU.addRequired<AssumptionCacheTracker>();
1494 AU.addRequiredID(LoopSimplifyID);
1495 AU.addRequiredID(LCSSAID);
1496 AU.addRequired<BlockFrequencyInfo>();
1497 AU.addRequired<DominatorTreeWrapperPass>();
1498 AU.addRequired<LoopInfo>();
1499 AU.addRequired<ScalarEvolution>();
1500 AU.addRequired<TargetTransformInfo>();
1501 AU.addRequired<AliasAnalysis>();
1502 AU.addPreserved<LoopInfo>();
1503 AU.addPreserved<DominatorTreeWrapperPass>();
1504 AU.addPreserved<AliasAnalysis>();
1505 }
1506
1507 };
1508
1509 } // end anonymous namespace
1510
1511 //===----------------------------------------------------------------------===//
1512 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1513 // LoopVectorizationCostModel.
1514 //===----------------------------------------------------------------------===//
1515
stripIntegerCast(Value * V)1516 static Value *stripIntegerCast(Value *V) {
1517 if (CastInst *CI = dyn_cast<CastInst>(V))
1518 if (CI->getOperand(0)->getType()->isIntegerTy())
1519 return CI->getOperand(0);
1520 return V;
1521 }
1522
1523 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1524 ///
1525 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1526 /// \p Ptr.
replaceSymbolicStrideSCEV(ScalarEvolution * SE,ValueToValueMap & PtrToStride,Value * Ptr,Value * OrigPtr=nullptr)1527 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1528 ValueToValueMap &PtrToStride,
1529 Value *Ptr, Value *OrigPtr = nullptr) {
1530
1531 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1532
1533 // If there is an entry in the map return the SCEV of the pointer with the
1534 // symbolic stride replaced by one.
1535 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1536 if (SI != PtrToStride.end()) {
1537 Value *StrideVal = SI->second;
1538
1539 // Strip casts.
1540 StrideVal = stripIntegerCast(StrideVal);
1541
1542 // Replace symbolic stride by one.
1543 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1544 ValueToValueMap RewriteMap;
1545 RewriteMap[StrideVal] = One;
1546
1547 const SCEV *ByOne =
1548 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1549 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1550 << "\n");
1551 return ByOne;
1552 }
1553
1554 // Otherwise, just return the SCEV of the original pointer.
1555 return SE->getSCEV(Ptr);
1556 }
1557
insert(ScalarEvolution * SE,Loop * Lp,Value * Ptr,bool WritePtr,unsigned DepSetId,unsigned ASId,ValueToValueMap & Strides)1558 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1559 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1560 unsigned ASId, ValueToValueMap &Strides) {
1561 // Get the stride replaced scev.
1562 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1563 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1564 assert(AR && "Invalid addrec expression");
1565 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1566 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1567 Pointers.push_back(Ptr);
1568 Starts.push_back(AR->getStart());
1569 Ends.push_back(ScEnd);
1570 IsWritePtr.push_back(WritePtr);
1571 DependencySetId.push_back(DepSetId);
1572 AliasSetId.push_back(ASId);
1573 }
1574
getBroadcastInstrs(Value * V)1575 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1576 // We need to place the broadcast of invariant variables outside the loop.
1577 Instruction *Instr = dyn_cast<Instruction>(V);
1578 bool NewInstr =
1579 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1580 Instr->getParent()) != LoopVectorBody.end());
1581 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1582
1583 // Place the code for broadcasting invariant variables in the new preheader.
1584 IRBuilder<>::InsertPointGuard Guard(Builder);
1585 if (Invariant)
1586 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1587
1588 // Broadcast the scalar into all locations in the vector.
1589 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1590
1591 return Shuf;
1592 }
1593
getConsecutiveVector(Value * Val,int StartIdx,bool Negate)1594 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1595 bool Negate) {
1596 assert(Val->getType()->isVectorTy() && "Must be a vector");
1597 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1598 "Elem must be an integer");
1599 // Create the types.
1600 Type *ITy = Val->getType()->getScalarType();
1601 VectorType *Ty = cast<VectorType>(Val->getType());
1602 int VLen = Ty->getNumElements();
1603 SmallVector<Constant*, 8> Indices;
1604
1605 // Create a vector of consecutive numbers from zero to VF.
1606 for (int i = 0; i < VLen; ++i) {
1607 int64_t Idx = Negate ? (-i) : i;
1608 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1609 }
1610
1611 // Add the consecutive indices to the vector value.
1612 Constant *Cv = ConstantVector::get(Indices);
1613 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1614 return Builder.CreateAdd(Val, Cv, "induction");
1615 }
1616
1617 /// \brief Find the operand of the GEP that should be checked for consecutive
1618 /// stores. This ignores trailing indices that have no effect on the final
1619 /// pointer.
getGEPInductionOperand(const DataLayout * DL,const GetElementPtrInst * Gep)1620 static unsigned getGEPInductionOperand(const DataLayout *DL,
1621 const GetElementPtrInst *Gep) {
1622 unsigned LastOperand = Gep->getNumOperands() - 1;
1623 unsigned GEPAllocSize = DL->getTypeAllocSize(
1624 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1625
1626 // Walk backwards and try to peel off zeros.
1627 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1628 // Find the type we're currently indexing into.
1629 gep_type_iterator GEPTI = gep_type_begin(Gep);
1630 std::advance(GEPTI, LastOperand - 1);
1631
1632 // If it's a type with the same allocation size as the result of the GEP we
1633 // can peel off the zero index.
1634 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1635 break;
1636 --LastOperand;
1637 }
1638
1639 return LastOperand;
1640 }
1641
isConsecutivePtr(Value * Ptr)1642 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1643 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1644 // Make sure that the pointer does not point to structs.
1645 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1646 return 0;
1647
1648 // If this value is a pointer induction variable we know it is consecutive.
1649 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1650 if (Phi && Inductions.count(Phi)) {
1651 InductionInfo II = Inductions[Phi];
1652 if (IK_PtrInduction == II.IK)
1653 return 1;
1654 else if (IK_ReversePtrInduction == II.IK)
1655 return -1;
1656 }
1657
1658 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1659 if (!Gep)
1660 return 0;
1661
1662 unsigned NumOperands = Gep->getNumOperands();
1663 Value *GpPtr = Gep->getPointerOperand();
1664 // If this GEP value is a consecutive pointer induction variable and all of
1665 // the indices are constant then we know it is consecutive. We can
1666 Phi = dyn_cast<PHINode>(GpPtr);
1667 if (Phi && Inductions.count(Phi)) {
1668
1669 // Make sure that the pointer does not point to structs.
1670 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1671 if (GepPtrType->getElementType()->isAggregateType())
1672 return 0;
1673
1674 // Make sure that all of the index operands are loop invariant.
1675 for (unsigned i = 1; i < NumOperands; ++i)
1676 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1677 return 0;
1678
1679 InductionInfo II = Inductions[Phi];
1680 if (IK_PtrInduction == II.IK)
1681 return 1;
1682 else if (IK_ReversePtrInduction == II.IK)
1683 return -1;
1684 }
1685
1686 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1687
1688 // Check that all of the gep indices are uniform except for our induction
1689 // operand.
1690 for (unsigned i = 0; i != NumOperands; ++i)
1691 if (i != InductionOperand &&
1692 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1693 return 0;
1694
1695 // We can emit wide load/stores only if the last non-zero index is the
1696 // induction variable.
1697 const SCEV *Last = nullptr;
1698 if (!Strides.count(Gep))
1699 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1700 else {
1701 // Because of the multiplication by a stride we can have a s/zext cast.
1702 // We are going to replace this stride by 1 so the cast is safe to ignore.
1703 //
1704 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1705 // %0 = trunc i64 %indvars.iv to i32
1706 // %mul = mul i32 %0, %Stride1
1707 // %idxprom = zext i32 %mul to i64 << Safe cast.
1708 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1709 //
1710 Last = replaceSymbolicStrideSCEV(SE, Strides,
1711 Gep->getOperand(InductionOperand), Gep);
1712 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1713 Last =
1714 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1715 ? C->getOperand()
1716 : Last;
1717 }
1718 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1719 const SCEV *Step = AR->getStepRecurrence(*SE);
1720
1721 // The memory is consecutive because the last index is consecutive
1722 // and all other indices are loop invariant.
1723 if (Step->isOne())
1724 return 1;
1725 if (Step->isAllOnesValue())
1726 return -1;
1727 }
1728
1729 return 0;
1730 }
1731
isUniform(Value * V)1732 bool LoopVectorizationLegality::isUniform(Value *V) {
1733 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1734 }
1735
1736 InnerLoopVectorizer::VectorParts&
getVectorValue(Value * V)1737 InnerLoopVectorizer::getVectorValue(Value *V) {
1738 assert(V != Induction && "The new induction variable should not be used.");
1739 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1740
1741 // If we have a stride that is replaced by one, do it here.
1742 if (Legal->hasStride(V))
1743 V = ConstantInt::get(V->getType(), 1);
1744
1745 // If we have this scalar in the map, return it.
1746 if (WidenMap.has(V))
1747 return WidenMap.get(V);
1748
1749 // If this scalar is unknown, assume that it is a constant or that it is
1750 // loop invariant. Broadcast V and save the value for future uses.
1751 Value *B = getBroadcastInstrs(V);
1752 return WidenMap.splat(V, B);
1753 }
1754
reverseVector(Value * Vec)1755 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1756 assert(Vec->getType()->isVectorTy() && "Invalid type");
1757 SmallVector<Constant*, 8> ShuffleMask;
1758 for (unsigned i = 0; i < VF; ++i)
1759 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1760
1761 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1762 ConstantVector::get(ShuffleMask),
1763 "reverse");
1764 }
1765
vectorizeMemoryInstruction(Instruction * Instr)1766 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1767 // Attempt to issue a wide load.
1768 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1769 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1770
1771 assert((LI || SI) && "Invalid Load/Store instruction");
1772
1773 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1774 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1775 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1776 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1777 // An alignment of 0 means target abi alignment. We need to use the scalar's
1778 // target abi alignment in such a case.
1779 if (!Alignment)
1780 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1781 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1782 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1783 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1784
1785 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1786 !Legal->isMaskRequired(SI))
1787 return scalarizeInstruction(Instr, true);
1788
1789 if (ScalarAllocatedSize != VectorElementSize)
1790 return scalarizeInstruction(Instr);
1791
1792 // If the pointer is loop invariant or if it is non-consecutive,
1793 // scalarize the load.
1794 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1795 bool Reverse = ConsecutiveStride < 0;
1796 bool UniformLoad = LI && Legal->isUniform(Ptr);
1797 if (!ConsecutiveStride || UniformLoad)
1798 return scalarizeInstruction(Instr);
1799
1800 Constant *Zero = Builder.getInt32(0);
1801 VectorParts &Entry = WidenMap.get(Instr);
1802
1803 // Handle consecutive loads/stores.
1804 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1805 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1806 setDebugLocFromInst(Builder, Gep);
1807 Value *PtrOperand = Gep->getPointerOperand();
1808 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1809 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1810
1811 // Create the new GEP with the new induction variable.
1812 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1813 Gep2->setOperand(0, FirstBasePtr);
1814 Gep2->setName("gep.indvar.base");
1815 Ptr = Builder.Insert(Gep2);
1816 } else if (Gep) {
1817 setDebugLocFromInst(Builder, Gep);
1818 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1819 OrigLoop) && "Base ptr must be invariant");
1820
1821 // The last index does not have to be the induction. It can be
1822 // consecutive and be a function of the index. For example A[I+1];
1823 unsigned NumOperands = Gep->getNumOperands();
1824 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1825 // Create the new GEP with the new induction variable.
1826 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1827
1828 for (unsigned i = 0; i < NumOperands; ++i) {
1829 Value *GepOperand = Gep->getOperand(i);
1830 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1831
1832 // Update last index or loop invariant instruction anchored in loop.
1833 if (i == InductionOperand ||
1834 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1835 assert((i == InductionOperand ||
1836 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1837 "Must be last index or loop invariant");
1838
1839 VectorParts &GEPParts = getVectorValue(GepOperand);
1840 Value *Index = GEPParts[0];
1841 Index = Builder.CreateExtractElement(Index, Zero);
1842 Gep2->setOperand(i, Index);
1843 Gep2->setName("gep.indvar.idx");
1844 }
1845 }
1846 Ptr = Builder.Insert(Gep2);
1847 } else {
1848 // Use the induction element ptr.
1849 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1850 setDebugLocFromInst(Builder, Ptr);
1851 VectorParts &PtrVal = getVectorValue(Ptr);
1852 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1853 }
1854
1855 VectorParts Mask = createBlockInMask(Instr->getParent());
1856 // Handle Stores:
1857 if (SI) {
1858 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1859 "We do not allow storing to uniform addresses");
1860 setDebugLocFromInst(Builder, SI);
1861 // We don't want to update the value in the map as it might be used in
1862 // another expression. So don't use a reference type for "StoredVal".
1863 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1864
1865 for (unsigned Part = 0; Part < UF; ++Part) {
1866 // Calculate the pointer for the specific unroll-part.
1867 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1868
1869 if (Reverse) {
1870 // If we store to reverse consecutive memory locations then we need
1871 // to reverse the order of elements in the stored value.
1872 StoredVal[Part] = reverseVector(StoredVal[Part]);
1873 // If the address is consecutive but reversed, then the
1874 // wide store needs to start at the last vector element.
1875 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1876 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1877 Mask[Part] = reverseVector(Mask[Part]);
1878 }
1879
1880 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1881 DataTy->getPointerTo(AddressSpace));
1882
1883 Instruction *NewSI;
1884 if (Legal->isMaskRequired(SI))
1885 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1886 Mask[Part]);
1887 else
1888 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1889 propagateMetadata(NewSI, SI);
1890 }
1891 return;
1892 }
1893
1894 // Handle loads.
1895 assert(LI && "Must have a load instruction");
1896 setDebugLocFromInst(Builder, LI);
1897 for (unsigned Part = 0; Part < UF; ++Part) {
1898 // Calculate the pointer for the specific unroll-part.
1899 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1900
1901 if (Reverse) {
1902 // If the address is consecutive but reversed, then the
1903 // wide load needs to start at the last vector element.
1904 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1905 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1906 Mask[Part] = reverseVector(Mask[Part]);
1907 }
1908
1909 Instruction* NewLI;
1910 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1911 DataTy->getPointerTo(AddressSpace));
1912 if (Legal->isMaskRequired(LI))
1913 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1914 UndefValue::get(DataTy),
1915 "wide.masked.load");
1916 else
1917 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1918 propagateMetadata(NewLI, LI);
1919 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1920 }
1921 }
1922
scalarizeInstruction(Instruction * Instr,bool IfPredicateStore)1923 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1924 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1925 // Holds vector parameters or scalars, in case of uniform vals.
1926 SmallVector<VectorParts, 4> Params;
1927
1928 setDebugLocFromInst(Builder, Instr);
1929
1930 // Find all of the vectorized parameters.
1931 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1932 Value *SrcOp = Instr->getOperand(op);
1933
1934 // If we are accessing the old induction variable, use the new one.
1935 if (SrcOp == OldInduction) {
1936 Params.push_back(getVectorValue(SrcOp));
1937 continue;
1938 }
1939
1940 // Try using previously calculated values.
1941 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1942
1943 // If the src is an instruction that appeared earlier in the basic block
1944 // then it should already be vectorized.
1945 if (SrcInst && OrigLoop->contains(SrcInst)) {
1946 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1947 // The parameter is a vector value from earlier.
1948 Params.push_back(WidenMap.get(SrcInst));
1949 } else {
1950 // The parameter is a scalar from outside the loop. Maybe even a constant.
1951 VectorParts Scalars;
1952 Scalars.append(UF, SrcOp);
1953 Params.push_back(Scalars);
1954 }
1955 }
1956
1957 assert(Params.size() == Instr->getNumOperands() &&
1958 "Invalid number of operands");
1959
1960 // Does this instruction return a value ?
1961 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1962
1963 Value *UndefVec = IsVoidRetTy ? nullptr :
1964 UndefValue::get(VectorType::get(Instr->getType(), VF));
1965 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1966 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1967
1968 Instruction *InsertPt = Builder.GetInsertPoint();
1969 BasicBlock *IfBlock = Builder.GetInsertBlock();
1970 BasicBlock *CondBlock = nullptr;
1971
1972 VectorParts Cond;
1973 Loop *VectorLp = nullptr;
1974 if (IfPredicateStore) {
1975 assert(Instr->getParent()->getSinglePredecessor() &&
1976 "Only support single predecessor blocks");
1977 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1978 Instr->getParent());
1979 VectorLp = LI->getLoopFor(IfBlock);
1980 assert(VectorLp && "Must have a loop for this block");
1981 }
1982
1983 // For each vector unroll 'part':
1984 for (unsigned Part = 0; Part < UF; ++Part) {
1985 // For each scalar that we create:
1986 for (unsigned Width = 0; Width < VF; ++Width) {
1987
1988 // Start if-block.
1989 Value *Cmp = nullptr;
1990 if (IfPredicateStore) {
1991 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1992 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1993 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1994 LoopVectorBody.push_back(CondBlock);
1995 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1996 // Update Builder with newly created basic block.
1997 Builder.SetInsertPoint(InsertPt);
1998 }
1999
2000 Instruction *Cloned = Instr->clone();
2001 if (!IsVoidRetTy)
2002 Cloned->setName(Instr->getName() + ".cloned");
2003 // Replace the operands of the cloned instructions with extracted scalars.
2004 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
2005 Value *Op = Params[op][Part];
2006 // Param is a vector. Need to extract the right lane.
2007 if (Op->getType()->isVectorTy())
2008 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
2009 Cloned->setOperand(op, Op);
2010 }
2011
2012 // Place the cloned scalar in the new loop.
2013 Builder.Insert(Cloned);
2014
2015 // If the original scalar returns a value we need to place it in a vector
2016 // so that future users will be able to use it.
2017 if (!IsVoidRetTy)
2018 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
2019 Builder.getInt32(Width));
2020 // End if-block.
2021 if (IfPredicateStore) {
2022 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2023 LoopVectorBody.push_back(NewIfBlock);
2024 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
2025 Builder.SetInsertPoint(InsertPt);
2026 Instruction *OldBr = IfBlock->getTerminator();
2027 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2028 OldBr->eraseFromParent();
2029 IfBlock = NewIfBlock;
2030 }
2031 }
2032 }
2033 }
2034
getFirstInst(Instruction * FirstInst,Value * V,Instruction * Loc)2035 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2036 Instruction *Loc) {
2037 if (FirstInst)
2038 return FirstInst;
2039 if (Instruction *I = dyn_cast<Instruction>(V))
2040 return I->getParent() == Loc->getParent() ? I : nullptr;
2041 return nullptr;
2042 }
2043
2044 std::pair<Instruction *, Instruction *>
addStrideCheck(Instruction * Loc)2045 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2046 Instruction *tnullptr = nullptr;
2047 if (!Legal->mustCheckStrides())
2048 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2049
2050 IRBuilder<> ChkBuilder(Loc);
2051
2052 // Emit checks.
2053 Value *Check = nullptr;
2054 Instruction *FirstInst = nullptr;
2055 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2056 SE = Legal->strides_end();
2057 SI != SE; ++SI) {
2058 Value *Ptr = stripIntegerCast(*SI);
2059 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2060 "stride.chk");
2061 // Store the first instruction we create.
2062 FirstInst = getFirstInst(FirstInst, C, Loc);
2063 if (Check)
2064 Check = ChkBuilder.CreateOr(Check, C);
2065 else
2066 Check = C;
2067 }
2068
2069 // We have to do this trickery because the IRBuilder might fold the check to a
2070 // constant expression in which case there is no Instruction anchored in a
2071 // the block.
2072 LLVMContext &Ctx = Loc->getContext();
2073 Instruction *TheCheck =
2074 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2075 ChkBuilder.Insert(TheCheck, "stride.not.one");
2076 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2077
2078 return std::make_pair(FirstInst, TheCheck);
2079 }
2080
2081 std::pair<Instruction *, Instruction *>
addRuntimeCheck(Instruction * Loc)2082 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
2083 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
2084 Legal->getRuntimePointerCheck();
2085
2086 Instruction *tnullptr = nullptr;
2087 if (!PtrRtCheck->Need)
2088 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2089
2090 unsigned NumPointers = PtrRtCheck->Pointers.size();
2091 SmallVector<TrackingVH<Value> , 2> Starts;
2092 SmallVector<TrackingVH<Value> , 2> Ends;
2093
2094 LLVMContext &Ctx = Loc->getContext();
2095 SCEVExpander Exp(*SE, "induction");
2096 Instruction *FirstInst = nullptr;
2097
2098 for (unsigned i = 0; i < NumPointers; ++i) {
2099 Value *Ptr = PtrRtCheck->Pointers[i];
2100 const SCEV *Sc = SE->getSCEV(Ptr);
2101
2102 if (SE->isLoopInvariant(Sc, OrigLoop)) {
2103 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
2104 *Ptr <<"\n");
2105 Starts.push_back(Ptr);
2106 Ends.push_back(Ptr);
2107 } else {
2108 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
2109 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2110
2111 // Use this type for pointer arithmetic.
2112 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2113
2114 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2115 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2116 Starts.push_back(Start);
2117 Ends.push_back(End);
2118 }
2119 }
2120
2121 IRBuilder<> ChkBuilder(Loc);
2122 // Our instructions might fold to a constant.
2123 Value *MemoryRuntimeCheck = nullptr;
2124 for (unsigned i = 0; i < NumPointers; ++i) {
2125 for (unsigned j = i+1; j < NumPointers; ++j) {
2126 // No need to check if two readonly pointers intersect.
2127 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2128 continue;
2129
2130 // Only need to check pointers between two different dependency sets.
2131 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2132 continue;
2133 // Only need to check pointers in the same alias set.
2134 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2135 continue;
2136
2137 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2138 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2139
2140 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2141 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2142 "Trying to bounds check pointers with different address spaces");
2143
2144 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2145 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2146
2147 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2148 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2149 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2150 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2151
2152 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2153 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2154 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2155 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2156 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2157 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2158 if (MemoryRuntimeCheck) {
2159 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2160 "conflict.rdx");
2161 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2162 }
2163 MemoryRuntimeCheck = IsConflict;
2164 }
2165 }
2166
2167 // We have to do this trickery because the IRBuilder might fold the check to a
2168 // constant expression in which case there is no Instruction anchored in a
2169 // the block.
2170 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2171 ConstantInt::getTrue(Ctx));
2172 ChkBuilder.Insert(Check, "memcheck.conflict");
2173 FirstInst = getFirstInst(FirstInst, Check, Loc);
2174 return std::make_pair(FirstInst, Check);
2175 }
2176
createEmptyLoop()2177 void InnerLoopVectorizer::createEmptyLoop() {
2178 /*
2179 In this function we generate a new loop. The new loop will contain
2180 the vectorized instructions while the old loop will continue to run the
2181 scalar remainder.
2182
2183 [ ] <-- Back-edge taken count overflow check.
2184 / |
2185 / v
2186 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2187 | / |
2188 | / v
2189 || [ ] <-- vector pre header.
2190 || |
2191 || v
2192 || [ ] \
2193 || [ ]_| <-- vector loop.
2194 || |
2195 | \ v
2196 | >[ ] <--- middle-block.
2197 | / |
2198 | / v
2199 -|- >[ ] <--- new preheader.
2200 | |
2201 | v
2202 | [ ] \
2203 | [ ]_| <-- old scalar loop to handle remainder.
2204 \ |
2205 \ v
2206 >[ ] <-- exit block.
2207 ...
2208 */
2209
2210 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2211 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2212 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2213 assert(BypassBlock && "Invalid loop structure");
2214 assert(ExitBlock && "Must have an exit block");
2215
2216 // Some loops have a single integer induction variable, while other loops
2217 // don't. One example is c++ iterators that often have multiple pointer
2218 // induction variables. In the code below we also support a case where we
2219 // don't have a single induction variable.
2220 OldInduction = Legal->getInduction();
2221 Type *IdxTy = Legal->getWidestInductionType();
2222
2223 // Find the loop boundaries.
2224 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2225 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2226
2227 // The exit count might have the type of i64 while the phi is i32. This can
2228 // happen if we have an induction variable that is sign extended before the
2229 // compare. The only way that we get a backedge taken count is that the
2230 // induction variable was signed and as such will not overflow. In such a case
2231 // truncation is legal.
2232 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2233 IdxTy->getPrimitiveSizeInBits())
2234 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2235
2236 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2237 // Get the total trip count from the count by adding 1.
2238 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2239 SE->getConstant(BackedgeTakeCount->getType(), 1));
2240
2241 // Expand the trip count and place the new instructions in the preheader.
2242 // Notice that the pre-header does not change, only the loop body.
2243 SCEVExpander Exp(*SE, "induction");
2244
2245 // We need to test whether the backedge-taken count is uint##_max. Adding one
2246 // to it will cause overflow and an incorrect loop trip count in the vector
2247 // body. In case of overflow we want to directly jump to the scalar remainder
2248 // loop.
2249 Value *BackedgeCount =
2250 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2251 BypassBlock->getTerminator());
2252 if (BackedgeCount->getType()->isPointerTy())
2253 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2254 "backedge.ptrcnt.to.int",
2255 BypassBlock->getTerminator());
2256 Instruction *CheckBCOverflow =
2257 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2258 Constant::getAllOnesValue(BackedgeCount->getType()),
2259 "backedge.overflow", BypassBlock->getTerminator());
2260
2261 // The loop index does not have to start at Zero. Find the original start
2262 // value from the induction PHI node. If we don't have an induction variable
2263 // then we know that it starts at zero.
2264 Builder.SetInsertPoint(BypassBlock->getTerminator());
2265 Value *StartIdx = ExtendedIdx = OldInduction ?
2266 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2267 IdxTy):
2268 ConstantInt::get(IdxTy, 0);
2269
2270 // We need an instruction to anchor the overflow check on. StartIdx needs to
2271 // be defined before the overflow check branch. Because the scalar preheader
2272 // is going to merge the start index and so the overflow branch block needs to
2273 // contain a definition of the start index.
2274 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2275 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2276 BypassBlock->getTerminator());
2277
2278 // Count holds the overall loop count (N).
2279 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2280 BypassBlock->getTerminator());
2281
2282 LoopBypassBlocks.push_back(BypassBlock);
2283
2284 // Split the single block loop into the two loop structure described above.
2285 BasicBlock *VectorPH =
2286 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2287 BasicBlock *VecBody =
2288 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2289 BasicBlock *MiddleBlock =
2290 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2291 BasicBlock *ScalarPH =
2292 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2293
2294 // Create and register the new vector loop.
2295 Loop* Lp = new Loop();
2296 Loop *ParentLoop = OrigLoop->getParentLoop();
2297
2298 // Insert the new loop into the loop nest and register the new basic blocks
2299 // before calling any utilities such as SCEV that require valid LoopInfo.
2300 if (ParentLoop) {
2301 ParentLoop->addChildLoop(Lp);
2302 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2303 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2304 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2305 } else {
2306 LI->addTopLevelLoop(Lp);
2307 }
2308 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2309
2310 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2311 // inside the loop.
2312 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2313
2314 // Generate the induction variable.
2315 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2316 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2317 // The loop step is equal to the vectorization factor (num of SIMD elements)
2318 // times the unroll factor (num of SIMD instructions).
2319 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2320
2321 // This is the IR builder that we use to add all of the logic for bypassing
2322 // the new vector loop.
2323 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2324 setDebugLocFromInst(BypassBuilder,
2325 getDebugLocFromInstOrOperands(OldInduction));
2326
2327 // We may need to extend the index in case there is a type mismatch.
2328 // We know that the count starts at zero and does not overflow.
2329 if (Count->getType() != IdxTy) {
2330 // The exit count can be of pointer type. Convert it to the correct
2331 // integer type.
2332 if (ExitCount->getType()->isPointerTy())
2333 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2334 else
2335 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2336 }
2337
2338 // Add the start index to the loop count to get the new end index.
2339 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2340
2341 // Now we need to generate the expression for N - (N % VF), which is
2342 // the part that the vectorized body will execute.
2343 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2344 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2345 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2346 "end.idx.rnd.down");
2347
2348 // Now, compare the new count to zero. If it is zero skip the vector loop and
2349 // jump to the scalar loop.
2350 Value *Cmp =
2351 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2352
2353 BasicBlock *LastBypassBlock = BypassBlock;
2354
2355 // Generate code to check that the loops trip count that we computed by adding
2356 // one to the backedge-taken count will not overflow.
2357 {
2358 auto PastOverflowCheck =
2359 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2360 BasicBlock *CheckBlock =
2361 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2362 if (ParentLoop)
2363 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2364 LoopBypassBlocks.push_back(CheckBlock);
2365 Instruction *OldTerm = LastBypassBlock->getTerminator();
2366 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2367 OldTerm->eraseFromParent();
2368 LastBypassBlock = CheckBlock;
2369 }
2370
2371 // Generate the code to check that the strides we assumed to be one are really
2372 // one. We want the new basic block to start at the first instruction in a
2373 // sequence of instructions that form a check.
2374 Instruction *StrideCheck;
2375 Instruction *FirstCheckInst;
2376 std::tie(FirstCheckInst, StrideCheck) =
2377 addStrideCheck(LastBypassBlock->getTerminator());
2378 if (StrideCheck) {
2379 // Create a new block containing the stride check.
2380 BasicBlock *CheckBlock =
2381 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2382 if (ParentLoop)
2383 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2384 LoopBypassBlocks.push_back(CheckBlock);
2385
2386 // Replace the branch into the memory check block with a conditional branch
2387 // for the "few elements case".
2388 Instruction *OldTerm = LastBypassBlock->getTerminator();
2389 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2390 OldTerm->eraseFromParent();
2391
2392 Cmp = StrideCheck;
2393 LastBypassBlock = CheckBlock;
2394 }
2395
2396 // Generate the code that checks in runtime if arrays overlap. We put the
2397 // checks into a separate block to make the more common case of few elements
2398 // faster.
2399 Instruction *MemRuntimeCheck;
2400 std::tie(FirstCheckInst, MemRuntimeCheck) =
2401 addRuntimeCheck(LastBypassBlock->getTerminator());
2402 if (MemRuntimeCheck) {
2403 // Create a new block containing the memory check.
2404 BasicBlock *CheckBlock =
2405 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2406 if (ParentLoop)
2407 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2408 LoopBypassBlocks.push_back(CheckBlock);
2409
2410 // Replace the branch into the memory check block with a conditional branch
2411 // for the "few elements case".
2412 Instruction *OldTerm = LastBypassBlock->getTerminator();
2413 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2414 OldTerm->eraseFromParent();
2415
2416 Cmp = MemRuntimeCheck;
2417 LastBypassBlock = CheckBlock;
2418 }
2419
2420 LastBypassBlock->getTerminator()->eraseFromParent();
2421 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2422 LastBypassBlock);
2423
2424 // We are going to resume the execution of the scalar loop.
2425 // Go over all of the induction variables that we found and fix the
2426 // PHIs that are left in the scalar version of the loop.
2427 // The starting values of PHI nodes depend on the counter of the last
2428 // iteration in the vectorized loop.
2429 // If we come from a bypass edge then we need to start from the original
2430 // start value.
2431
2432 // This variable saves the new starting index for the scalar loop.
2433 PHINode *ResumeIndex = nullptr;
2434 LoopVectorizationLegality::InductionList::iterator I, E;
2435 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2436 // Set builder to point to last bypass block.
2437 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2438 for (I = List->begin(), E = List->end(); I != E; ++I) {
2439 PHINode *OrigPhi = I->first;
2440 LoopVectorizationLegality::InductionInfo II = I->second;
2441
2442 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2443 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2444 MiddleBlock->getTerminator());
2445 // We might have extended the type of the induction variable but we need a
2446 // truncated version for the scalar loop.
2447 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2448 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2449 MiddleBlock->getTerminator()) : nullptr;
2450
2451 // Create phi nodes to merge from the backedge-taken check block.
2452 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2453 ScalarPH->getTerminator());
2454 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2455
2456 PHINode *BCTruncResumeVal = nullptr;
2457 if (OrigPhi == OldInduction) {
2458 BCTruncResumeVal =
2459 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2460 ScalarPH->getTerminator());
2461 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2462 }
2463
2464 Value *EndValue = nullptr;
2465 switch (II.IK) {
2466 case LoopVectorizationLegality::IK_NoInduction:
2467 llvm_unreachable("Unknown induction");
2468 case LoopVectorizationLegality::IK_IntInduction: {
2469 // Handle the integer induction counter.
2470 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2471
2472 // We have the canonical induction variable.
2473 if (OrigPhi == OldInduction) {
2474 // Create a truncated version of the resume value for the scalar loop,
2475 // we might have promoted the type to a larger width.
2476 EndValue =
2477 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2478 // The new PHI merges the original incoming value, in case of a bypass,
2479 // or the value at the end of the vectorized loop.
2480 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2481 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2482 TruncResumeVal->addIncoming(EndValue, VecBody);
2483
2484 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2485
2486 // We know what the end value is.
2487 EndValue = IdxEndRoundDown;
2488 // We also know which PHI node holds it.
2489 ResumeIndex = ResumeVal;
2490 break;
2491 }
2492
2493 // Not the canonical induction variable - add the vector loop count to the
2494 // start value.
2495 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2496 II.StartValue->getType(),
2497 "cast.crd");
2498 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2499 break;
2500 }
2501 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2502 // Convert the CountRoundDown variable to the PHI size.
2503 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2504 II.StartValue->getType(),
2505 "cast.crd");
2506 // Handle reverse integer induction counter.
2507 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2508 break;
2509 }
2510 case LoopVectorizationLegality::IK_PtrInduction: {
2511 // For pointer induction variables, calculate the offset using
2512 // the end index.
2513 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2514 "ptr.ind.end");
2515 break;
2516 }
2517 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2518 // The value at the end of the loop for the reverse pointer is calculated
2519 // by creating a GEP with a negative index starting from the start value.
2520 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2521 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2522 "rev.ind.end");
2523 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2524 "rev.ptr.ind.end");
2525 break;
2526 }
2527 }// end of case
2528
2529 // The new PHI merges the original incoming value, in case of a bypass,
2530 // or the value at the end of the vectorized loop.
2531 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2532 if (OrigPhi == OldInduction)
2533 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2534 else
2535 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2536 }
2537 ResumeVal->addIncoming(EndValue, VecBody);
2538
2539 // Fix the scalar body counter (PHI node).
2540 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2541
2542 // The old induction's phi node in the scalar body needs the truncated
2543 // value.
2544 if (OrigPhi == OldInduction) {
2545 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2546 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2547 } else {
2548 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2549 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2550 }
2551 }
2552
2553 // If we are generating a new induction variable then we also need to
2554 // generate the code that calculates the exit value. This value is not
2555 // simply the end of the counter because we may skip the vectorized body
2556 // in case of a runtime check.
2557 if (!OldInduction){
2558 assert(!ResumeIndex && "Unexpected resume value found");
2559 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2560 MiddleBlock->getTerminator());
2561 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2562 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2563 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2564 }
2565
2566 // Make sure that we found the index where scalar loop needs to continue.
2567 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2568 "Invalid resume Index");
2569
2570 // Add a check in the middle block to see if we have completed
2571 // all of the iterations in the first vector loop.
2572 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2573 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2574 ResumeIndex, "cmp.n",
2575 MiddleBlock->getTerminator());
2576
2577 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2578 // Remove the old terminator.
2579 MiddleBlock->getTerminator()->eraseFromParent();
2580
2581 // Create i+1 and fill the PHINode.
2582 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2583 Induction->addIncoming(StartIdx, VectorPH);
2584 Induction->addIncoming(NextIdx, VecBody);
2585 // Create the compare.
2586 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2587 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2588
2589 // Now we have two terminators. Remove the old one from the block.
2590 VecBody->getTerminator()->eraseFromParent();
2591
2592 // Get ready to start creating new instructions into the vectorized body.
2593 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2594
2595 // Save the state.
2596 LoopVectorPreHeader = VectorPH;
2597 LoopScalarPreHeader = ScalarPH;
2598 LoopMiddleBlock = MiddleBlock;
2599 LoopExitBlock = ExitBlock;
2600 LoopVectorBody.push_back(VecBody);
2601 LoopScalarBody = OldBasicBlock;
2602
2603 LoopVectorizeHints Hints(Lp, true);
2604 Hints.setAlreadyVectorized();
2605 }
2606
2607 /// This function returns the identity element (or neutral element) for
2608 /// the operation K.
2609 Constant*
getReductionIdentity(ReductionKind K,Type * Tp)2610 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2611 switch (K) {
2612 case RK_IntegerXor:
2613 case RK_IntegerAdd:
2614 case RK_IntegerOr:
2615 // Adding, Xoring, Oring zero to a number does not change it.
2616 return ConstantInt::get(Tp, 0);
2617 case RK_IntegerMult:
2618 // Multiplying a number by 1 does not change it.
2619 return ConstantInt::get(Tp, 1);
2620 case RK_IntegerAnd:
2621 // AND-ing a number with an all-1 value does not change it.
2622 return ConstantInt::get(Tp, -1, true);
2623 case RK_FloatMult:
2624 // Multiplying a number by 1 does not change it.
2625 return ConstantFP::get(Tp, 1.0L);
2626 case RK_FloatAdd:
2627 // Adding zero to a number does not change it.
2628 return ConstantFP::get(Tp, 0.0L);
2629 default:
2630 llvm_unreachable("Unknown reduction kind");
2631 }
2632 }
2633
2634 /// This function translates the reduction kind to an LLVM binary operator.
2635 static unsigned
getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind)2636 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2637 switch (Kind) {
2638 case LoopVectorizationLegality::RK_IntegerAdd:
2639 return Instruction::Add;
2640 case LoopVectorizationLegality::RK_IntegerMult:
2641 return Instruction::Mul;
2642 case LoopVectorizationLegality::RK_IntegerOr:
2643 return Instruction::Or;
2644 case LoopVectorizationLegality::RK_IntegerAnd:
2645 return Instruction::And;
2646 case LoopVectorizationLegality::RK_IntegerXor:
2647 return Instruction::Xor;
2648 case LoopVectorizationLegality::RK_FloatMult:
2649 return Instruction::FMul;
2650 case LoopVectorizationLegality::RK_FloatAdd:
2651 return Instruction::FAdd;
2652 case LoopVectorizationLegality::RK_IntegerMinMax:
2653 return Instruction::ICmp;
2654 case LoopVectorizationLegality::RK_FloatMinMax:
2655 return Instruction::FCmp;
2656 default:
2657 llvm_unreachable("Unknown reduction operation");
2658 }
2659 }
2660
createMinMaxOp(IRBuilder<> & Builder,LoopVectorizationLegality::MinMaxReductionKind RK,Value * Left,Value * Right)2661 Value *createMinMaxOp(IRBuilder<> &Builder,
2662 LoopVectorizationLegality::MinMaxReductionKind RK,
2663 Value *Left,
2664 Value *Right) {
2665 CmpInst::Predicate P = CmpInst::ICMP_NE;
2666 switch (RK) {
2667 default:
2668 llvm_unreachable("Unknown min/max reduction kind");
2669 case LoopVectorizationLegality::MRK_UIntMin:
2670 P = CmpInst::ICMP_ULT;
2671 break;
2672 case LoopVectorizationLegality::MRK_UIntMax:
2673 P = CmpInst::ICMP_UGT;
2674 break;
2675 case LoopVectorizationLegality::MRK_SIntMin:
2676 P = CmpInst::ICMP_SLT;
2677 break;
2678 case LoopVectorizationLegality::MRK_SIntMax:
2679 P = CmpInst::ICMP_SGT;
2680 break;
2681 case LoopVectorizationLegality::MRK_FloatMin:
2682 P = CmpInst::FCMP_OLT;
2683 break;
2684 case LoopVectorizationLegality::MRK_FloatMax:
2685 P = CmpInst::FCMP_OGT;
2686 break;
2687 }
2688
2689 Value *Cmp;
2690 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2691 RK == LoopVectorizationLegality::MRK_FloatMax)
2692 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2693 else
2694 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2695
2696 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2697 return Select;
2698 }
2699
2700 namespace {
2701 struct CSEDenseMapInfo {
canHandle__anon03606d180211::CSEDenseMapInfo2702 static bool canHandle(Instruction *I) {
2703 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2704 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2705 }
getEmptyKey__anon03606d180211::CSEDenseMapInfo2706 static inline Instruction *getEmptyKey() {
2707 return DenseMapInfo<Instruction *>::getEmptyKey();
2708 }
getTombstoneKey__anon03606d180211::CSEDenseMapInfo2709 static inline Instruction *getTombstoneKey() {
2710 return DenseMapInfo<Instruction *>::getTombstoneKey();
2711 }
getHashValue__anon03606d180211::CSEDenseMapInfo2712 static unsigned getHashValue(Instruction *I) {
2713 assert(canHandle(I) && "Unknown instruction!");
2714 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2715 I->value_op_end()));
2716 }
isEqual__anon03606d180211::CSEDenseMapInfo2717 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2718 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2719 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2720 return LHS == RHS;
2721 return LHS->isIdenticalTo(RHS);
2722 }
2723 };
2724 }
2725
2726 /// \brief Check whether this block is a predicated block.
2727 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2728 /// = ...; " blocks. We start with one vectorized basic block. For every
2729 /// conditional block we split this vectorized block. Therefore, every second
2730 /// block will be a predicated one.
isPredicatedBlock(unsigned BlockNum)2731 static bool isPredicatedBlock(unsigned BlockNum) {
2732 return BlockNum % 2;
2733 }
2734
2735 ///\brief Perform cse of induction variable instructions.
cse(SmallVector<BasicBlock *,4> & BBs)2736 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2737 // Perform simple cse.
2738 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2739 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2740 BasicBlock *BB = BBs[i];
2741 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2742 Instruction *In = I++;
2743
2744 if (!CSEDenseMapInfo::canHandle(In))
2745 continue;
2746
2747 // Check if we can replace this instruction with any of the
2748 // visited instructions.
2749 if (Instruction *V = CSEMap.lookup(In)) {
2750 In->replaceAllUsesWith(V);
2751 In->eraseFromParent();
2752 continue;
2753 }
2754 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2755 // ...;" blocks for predicated stores. Every second block is a predicated
2756 // block.
2757 if (isPredicatedBlock(i))
2758 continue;
2759
2760 CSEMap[In] = In;
2761 }
2762 }
2763 }
2764
2765 /// \brief Adds a 'fast' flag to floating point operations.
addFastMathFlag(Value * V)2766 static Value *addFastMathFlag(Value *V) {
2767 if (isa<FPMathOperator>(V)){
2768 FastMathFlags Flags;
2769 Flags.setUnsafeAlgebra();
2770 cast<Instruction>(V)->setFastMathFlags(Flags);
2771 }
2772 return V;
2773 }
2774
vectorizeLoop()2775 void InnerLoopVectorizer::vectorizeLoop() {
2776 //===------------------------------------------------===//
2777 //
2778 // Notice: any optimization or new instruction that go
2779 // into the code below should be also be implemented in
2780 // the cost-model.
2781 //
2782 //===------------------------------------------------===//
2783 Constant *Zero = Builder.getInt32(0);
2784
2785 // In order to support reduction variables we need to be able to vectorize
2786 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2787 // stages. First, we create a new vector PHI node with no incoming edges.
2788 // We use this value when we vectorize all of the instructions that use the
2789 // PHI. Next, after all of the instructions in the block are complete we
2790 // add the new incoming edges to the PHI. At this point all of the
2791 // instructions in the basic block are vectorized, so we can use them to
2792 // construct the PHI.
2793 PhiVector RdxPHIsToFix;
2794
2795 // Scan the loop in a topological order to ensure that defs are vectorized
2796 // before users.
2797 LoopBlocksDFS DFS(OrigLoop);
2798 DFS.perform(LI);
2799
2800 // Vectorize all of the blocks in the original loop.
2801 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2802 be = DFS.endRPO(); bb != be; ++bb)
2803 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2804
2805 // At this point every instruction in the original loop is widened to
2806 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2807 // that we vectorized. The PHI nodes are currently empty because we did
2808 // not want to introduce cycles. Notice that the remaining PHI nodes
2809 // that we need to fix are reduction variables.
2810
2811 // Create the 'reduced' values for each of the induction vars.
2812 // The reduced values are the vector values that we scalarize and combine
2813 // after the loop is finished.
2814 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2815 it != e; ++it) {
2816 PHINode *RdxPhi = *it;
2817 assert(RdxPhi && "Unable to recover vectorized PHI");
2818
2819 // Find the reduction variable descriptor.
2820 assert(Legal->getReductionVars()->count(RdxPhi) &&
2821 "Unable to find the reduction variable");
2822 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2823 (*Legal->getReductionVars())[RdxPhi];
2824
2825 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2826
2827 // We need to generate a reduction vector from the incoming scalar.
2828 // To do so, we need to generate the 'identity' vector and override
2829 // one of the elements with the incoming scalar reduction. We need
2830 // to do it in the vector-loop preheader.
2831 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2832
2833 // This is the vector-clone of the value that leaves the loop.
2834 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2835 Type *VecTy = VectorExit[0]->getType();
2836
2837 // Find the reduction identity variable. Zero for addition, or, xor,
2838 // one for multiplication, -1 for And.
2839 Value *Identity;
2840 Value *VectorStart;
2841 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2842 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2843 // MinMax reduction have the start value as their identify.
2844 if (VF == 1) {
2845 VectorStart = Identity = RdxDesc.StartValue;
2846 } else {
2847 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2848 RdxDesc.StartValue,
2849 "minmax.ident");
2850 }
2851 } else {
2852 // Handle other reduction kinds:
2853 Constant *Iden =
2854 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2855 VecTy->getScalarType());
2856 if (VF == 1) {
2857 Identity = Iden;
2858 // This vector is the Identity vector where the first element is the
2859 // incoming scalar reduction.
2860 VectorStart = RdxDesc.StartValue;
2861 } else {
2862 Identity = ConstantVector::getSplat(VF, Iden);
2863
2864 // This vector is the Identity vector where the first element is the
2865 // incoming scalar reduction.
2866 VectorStart = Builder.CreateInsertElement(Identity,
2867 RdxDesc.StartValue, Zero);
2868 }
2869 }
2870
2871 // Fix the vector-loop phi.
2872
2873 // Reductions do not have to start at zero. They can start with
2874 // any loop invariant values.
2875 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2876 BasicBlock *Latch = OrigLoop->getLoopLatch();
2877 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2878 VectorParts &Val = getVectorValue(LoopVal);
2879 for (unsigned part = 0; part < UF; ++part) {
2880 // Make sure to add the reduction stat value only to the
2881 // first unroll part.
2882 Value *StartVal = (part == 0) ? VectorStart : Identity;
2883 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2884 LoopVectorPreHeader);
2885 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2886 LoopVectorBody.back());
2887 }
2888
2889 // Before each round, move the insertion point right between
2890 // the PHIs and the values we are going to write.
2891 // This allows us to write both PHINodes and the extractelement
2892 // instructions.
2893 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2894
2895 VectorParts RdxParts;
2896 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2897 for (unsigned part = 0; part < UF; ++part) {
2898 // This PHINode contains the vectorized reduction variable, or
2899 // the initial value vector, if we bypass the vector loop.
2900 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2901 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2902 Value *StartVal = (part == 0) ? VectorStart : Identity;
2903 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2904 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2905 NewPhi->addIncoming(RdxExitVal[part],
2906 LoopVectorBody.back());
2907 RdxParts.push_back(NewPhi);
2908 }
2909
2910 // Reduce all of the unrolled parts into a single vector.
2911 Value *ReducedPartRdx = RdxParts[0];
2912 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2913 setDebugLocFromInst(Builder, ReducedPartRdx);
2914 for (unsigned part = 1; part < UF; ++part) {
2915 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2916 // Floating point operations had to be 'fast' to enable the reduction.
2917 ReducedPartRdx = addFastMathFlag(
2918 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2919 ReducedPartRdx, "bin.rdx"));
2920 else
2921 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2922 ReducedPartRdx, RdxParts[part]);
2923 }
2924
2925 if (VF > 1) {
2926 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2927 // and vector ops, reducing the set of values being computed by half each
2928 // round.
2929 assert(isPowerOf2_32(VF) &&
2930 "Reduction emission only supported for pow2 vectors!");
2931 Value *TmpVec = ReducedPartRdx;
2932 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2933 for (unsigned i = VF; i != 1; i >>= 1) {
2934 // Move the upper half of the vector to the lower half.
2935 for (unsigned j = 0; j != i/2; ++j)
2936 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2937
2938 // Fill the rest of the mask with undef.
2939 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2940 UndefValue::get(Builder.getInt32Ty()));
2941
2942 Value *Shuf =
2943 Builder.CreateShuffleVector(TmpVec,
2944 UndefValue::get(TmpVec->getType()),
2945 ConstantVector::get(ShuffleMask),
2946 "rdx.shuf");
2947
2948 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2949 // Floating point operations had to be 'fast' to enable the reduction.
2950 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2951 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2952 else
2953 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2954 }
2955
2956 // The result is in the first element of the vector.
2957 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2958 Builder.getInt32(0));
2959 }
2960
2961 // Create a phi node that merges control-flow from the backedge-taken check
2962 // block and the middle block.
2963 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2964 LoopScalarPreHeader->getTerminator());
2965 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2966 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2967
2968 // Now, we need to fix the users of the reduction variable
2969 // inside and outside of the scalar remainder loop.
2970 // We know that the loop is in LCSSA form. We need to update the
2971 // PHI nodes in the exit blocks.
2972 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2973 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2974 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2975 if (!LCSSAPhi) break;
2976
2977 // All PHINodes need to have a single entry edge, or two if
2978 // we already fixed them.
2979 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2980
2981 // We found our reduction value exit-PHI. Update it with the
2982 // incoming bypass edge.
2983 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2984 // Add an edge coming from the bypass.
2985 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2986 break;
2987 }
2988 }// end of the LCSSA phi scan.
2989
2990 // Fix the scalar loop reduction variable with the incoming reduction sum
2991 // from the vector body and from the backedge value.
2992 int IncomingEdgeBlockIdx =
2993 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2994 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2995 // Pick the other block.
2996 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2997 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2998 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2999 }// end of for each redux variable.
3000
3001 fixLCSSAPHIs();
3002
3003 // Remove redundant induction instructions.
3004 cse(LoopVectorBody);
3005 }
3006
fixLCSSAPHIs()3007 void InnerLoopVectorizer::fixLCSSAPHIs() {
3008 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
3009 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
3010 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
3011 if (!LCSSAPhi) break;
3012 if (LCSSAPhi->getNumIncomingValues() == 1)
3013 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
3014 LoopMiddleBlock);
3015 }
3016 }
3017
3018 InnerLoopVectorizer::VectorParts
createEdgeMask(BasicBlock * Src,BasicBlock * Dst)3019 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
3020 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
3021 "Invalid edge");
3022
3023 // Look for cached value.
3024 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
3025 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
3026 if (ECEntryIt != MaskCache.end())
3027 return ECEntryIt->second;
3028
3029 VectorParts SrcMask = createBlockInMask(Src);
3030
3031 // The terminator has to be a branch inst!
3032 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3033 assert(BI && "Unexpected terminator found");
3034
3035 if (BI->isConditional()) {
3036 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3037
3038 if (BI->getSuccessor(0) != Dst)
3039 for (unsigned part = 0; part < UF; ++part)
3040 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3041
3042 for (unsigned part = 0; part < UF; ++part)
3043 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3044
3045 MaskCache[Edge] = EdgeMask;
3046 return EdgeMask;
3047 }
3048
3049 MaskCache[Edge] = SrcMask;
3050 return SrcMask;
3051 }
3052
3053 InnerLoopVectorizer::VectorParts
createBlockInMask(BasicBlock * BB)3054 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3055 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3056
3057 // Loop incoming mask is all-one.
3058 if (OrigLoop->getHeader() == BB) {
3059 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3060 return getVectorValue(C);
3061 }
3062
3063 // This is the block mask. We OR all incoming edges, and with zero.
3064 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3065 VectorParts BlockMask = getVectorValue(Zero);
3066
3067 // For each pred:
3068 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3069 VectorParts EM = createEdgeMask(*it, BB);
3070 for (unsigned part = 0; part < UF; ++part)
3071 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3072 }
3073
3074 return BlockMask;
3075 }
3076
widenPHIInstruction(Instruction * PN,InnerLoopVectorizer::VectorParts & Entry,unsigned UF,unsigned VF,PhiVector * PV)3077 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3078 InnerLoopVectorizer::VectorParts &Entry,
3079 unsigned UF, unsigned VF, PhiVector *PV) {
3080 PHINode* P = cast<PHINode>(PN);
3081 // Handle reduction variables:
3082 if (Legal->getReductionVars()->count(P)) {
3083 for (unsigned part = 0; part < UF; ++part) {
3084 // This is phase one of vectorizing PHIs.
3085 Type *VecTy = (VF == 1) ? PN->getType() :
3086 VectorType::get(PN->getType(), VF);
3087 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3088 LoopVectorBody.back()-> getFirstInsertionPt());
3089 }
3090 PV->push_back(P);
3091 return;
3092 }
3093
3094 setDebugLocFromInst(Builder, P);
3095 // Check for PHI nodes that are lowered to vector selects.
3096 if (P->getParent() != OrigLoop->getHeader()) {
3097 // We know that all PHIs in non-header blocks are converted into
3098 // selects, so we don't have to worry about the insertion order and we
3099 // can just use the builder.
3100 // At this point we generate the predication tree. There may be
3101 // duplications since this is a simple recursive scan, but future
3102 // optimizations will clean it up.
3103
3104 unsigned NumIncoming = P->getNumIncomingValues();
3105
3106 // Generate a sequence of selects of the form:
3107 // SELECT(Mask3, In3,
3108 // SELECT(Mask2, In2,
3109 // ( ...)))
3110 for (unsigned In = 0; In < NumIncoming; In++) {
3111 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3112 P->getParent());
3113 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3114
3115 for (unsigned part = 0; part < UF; ++part) {
3116 // We might have single edge PHIs (blocks) - use an identity
3117 // 'select' for the first PHI operand.
3118 if (In == 0)
3119 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3120 In0[part]);
3121 else
3122 // Select between the current value and the previous incoming edge
3123 // based on the incoming mask.
3124 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3125 Entry[part], "predphi");
3126 }
3127 }
3128 return;
3129 }
3130
3131 // This PHINode must be an induction variable.
3132 // Make sure that we know about it.
3133 assert(Legal->getInductionVars()->count(P) &&
3134 "Not an induction variable");
3135
3136 LoopVectorizationLegality::InductionInfo II =
3137 Legal->getInductionVars()->lookup(P);
3138
3139 switch (II.IK) {
3140 case LoopVectorizationLegality::IK_NoInduction:
3141 llvm_unreachable("Unknown induction");
3142 case LoopVectorizationLegality::IK_IntInduction: {
3143 assert(P->getType() == II.StartValue->getType() && "Types must match");
3144 Type *PhiTy = P->getType();
3145 Value *Broadcasted;
3146 if (P == OldInduction) {
3147 // Handle the canonical induction variable. We might have had to
3148 // extend the type.
3149 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3150 } else {
3151 // Handle other induction variables that are now based on the
3152 // canonical one.
3153 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3154 "normalized.idx");
3155 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3156 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
3157 "offset.idx");
3158 }
3159 Broadcasted = getBroadcastInstrs(Broadcasted);
3160 // After broadcasting the induction variable we need to make the vector
3161 // consecutive by adding 0, 1, 2, etc.
3162 for (unsigned part = 0; part < UF; ++part)
3163 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3164 return;
3165 }
3166 case LoopVectorizationLegality::IK_ReverseIntInduction:
3167 case LoopVectorizationLegality::IK_PtrInduction:
3168 case LoopVectorizationLegality::IK_ReversePtrInduction:
3169 // Handle reverse integer and pointer inductions.
3170 Value *StartIdx = ExtendedIdx;
3171 // This is the normalized GEP that starts counting at zero.
3172 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3173 "normalized.idx");
3174
3175 // Handle the reverse integer induction variable case.
3176 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3177 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3178 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3179 "resize.norm.idx");
3180 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3181 "reverse.idx");
3182
3183 // This is a new value so do not hoist it out.
3184 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3185 // After broadcasting the induction variable we need to make the
3186 // vector consecutive by adding ... -3, -2, -1, 0.
3187 for (unsigned part = 0; part < UF; ++part)
3188 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3189 true);
3190 return;
3191 }
3192
3193 // Handle the pointer induction variable case.
3194 assert(P->getType()->isPointerTy() && "Unexpected type.");
3195
3196 // Is this a reverse induction ptr or a consecutive induction ptr.
3197 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3198 II.IK);
3199
3200 // This is the vector of results. Notice that we don't generate
3201 // vector geps because scalar geps result in better code.
3202 for (unsigned part = 0; part < UF; ++part) {
3203 if (VF == 1) {
3204 int EltIndex = (part) * (Reverse ? -1 : 1);
3205 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3206 Value *GlobalIdx;
3207 if (Reverse)
3208 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3209 else
3210 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3211
3212 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3213 "next.gep");
3214 Entry[part] = SclrGep;
3215 continue;
3216 }
3217
3218 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3219 for (unsigned int i = 0; i < VF; ++i) {
3220 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3221 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3222 Value *GlobalIdx;
3223 if (!Reverse)
3224 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3225 else
3226 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3227
3228 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3229 "next.gep");
3230 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3231 Builder.getInt32(i),
3232 "insert.gep");
3233 }
3234 Entry[part] = VecVal;
3235 }
3236 return;
3237 }
3238 }
3239
vectorizeBlockInLoop(BasicBlock * BB,PhiVector * PV)3240 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3241 // For each instruction in the old loop.
3242 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3243 VectorParts &Entry = WidenMap.get(it);
3244 switch (it->getOpcode()) {
3245 case Instruction::Br:
3246 // Nothing to do for PHIs and BR, since we already took care of the
3247 // loop control flow instructions.
3248 continue;
3249 case Instruction::PHI:{
3250 // Vectorize PHINodes.
3251 widenPHIInstruction(it, Entry, UF, VF, PV);
3252 continue;
3253 }// End of PHI.
3254
3255 case Instruction::Add:
3256 case Instruction::FAdd:
3257 case Instruction::Sub:
3258 case Instruction::FSub:
3259 case Instruction::Mul:
3260 case Instruction::FMul:
3261 case Instruction::UDiv:
3262 case Instruction::SDiv:
3263 case Instruction::FDiv:
3264 case Instruction::URem:
3265 case Instruction::SRem:
3266 case Instruction::FRem:
3267 case Instruction::Shl:
3268 case Instruction::LShr:
3269 case Instruction::AShr:
3270 case Instruction::And:
3271 case Instruction::Or:
3272 case Instruction::Xor: {
3273 // Just widen binops.
3274 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3275 setDebugLocFromInst(Builder, BinOp);
3276 VectorParts &A = getVectorValue(it->getOperand(0));
3277 VectorParts &B = getVectorValue(it->getOperand(1));
3278
3279 // Use this vector value for all users of the original instruction.
3280 for (unsigned Part = 0; Part < UF; ++Part) {
3281 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3282
3283 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3284 VecOp->copyIRFlags(BinOp);
3285
3286 Entry[Part] = V;
3287 }
3288
3289 propagateMetadata(Entry, it);
3290 break;
3291 }
3292 case Instruction::Select: {
3293 // Widen selects.
3294 // If the selector is loop invariant we can create a select
3295 // instruction with a scalar condition. Otherwise, use vector-select.
3296 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3297 OrigLoop);
3298 setDebugLocFromInst(Builder, it);
3299
3300 // The condition can be loop invariant but still defined inside the
3301 // loop. This means that we can't just use the original 'cond' value.
3302 // We have to take the 'vectorized' value and pick the first lane.
3303 // Instcombine will make this a no-op.
3304 VectorParts &Cond = getVectorValue(it->getOperand(0));
3305 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3306 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3307
3308 Value *ScalarCond = (VF == 1) ? Cond[0] :
3309 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3310
3311 for (unsigned Part = 0; Part < UF; ++Part) {
3312 Entry[Part] = Builder.CreateSelect(
3313 InvariantCond ? ScalarCond : Cond[Part],
3314 Op0[Part],
3315 Op1[Part]);
3316 }
3317
3318 propagateMetadata(Entry, it);
3319 break;
3320 }
3321
3322 case Instruction::ICmp:
3323 case Instruction::FCmp: {
3324 // Widen compares. Generate vector compares.
3325 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3326 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3327 setDebugLocFromInst(Builder, it);
3328 VectorParts &A = getVectorValue(it->getOperand(0));
3329 VectorParts &B = getVectorValue(it->getOperand(1));
3330 for (unsigned Part = 0; Part < UF; ++Part) {
3331 Value *C = nullptr;
3332 if (FCmp)
3333 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3334 else
3335 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3336 Entry[Part] = C;
3337 }
3338
3339 propagateMetadata(Entry, it);
3340 break;
3341 }
3342
3343 case Instruction::Store:
3344 case Instruction::Load:
3345 vectorizeMemoryInstruction(it);
3346 break;
3347 case Instruction::ZExt:
3348 case Instruction::SExt:
3349 case Instruction::FPToUI:
3350 case Instruction::FPToSI:
3351 case Instruction::FPExt:
3352 case Instruction::PtrToInt:
3353 case Instruction::IntToPtr:
3354 case Instruction::SIToFP:
3355 case Instruction::UIToFP:
3356 case Instruction::Trunc:
3357 case Instruction::FPTrunc:
3358 case Instruction::BitCast: {
3359 CastInst *CI = dyn_cast<CastInst>(it);
3360 setDebugLocFromInst(Builder, it);
3361 /// Optimize the special case where the source is the induction
3362 /// variable. Notice that we can only optimize the 'trunc' case
3363 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3364 /// c. other casts depend on pointer size.
3365 if (CI->getOperand(0) == OldInduction &&
3366 it->getOpcode() == Instruction::Trunc) {
3367 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3368 CI->getType());
3369 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3370 for (unsigned Part = 0; Part < UF; ++Part)
3371 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3372 propagateMetadata(Entry, it);
3373 break;
3374 }
3375 /// Vectorize casts.
3376 Type *DestTy = (VF == 1) ? CI->getType() :
3377 VectorType::get(CI->getType(), VF);
3378
3379 VectorParts &A = getVectorValue(it->getOperand(0));
3380 for (unsigned Part = 0; Part < UF; ++Part)
3381 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3382 propagateMetadata(Entry, it);
3383 break;
3384 }
3385
3386 case Instruction::Call: {
3387 // Ignore dbg intrinsics.
3388 if (isa<DbgInfoIntrinsic>(it))
3389 break;
3390 setDebugLocFromInst(Builder, it);
3391
3392 Module *M = BB->getParent()->getParent();
3393 CallInst *CI = cast<CallInst>(it);
3394 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3395 assert(ID && "Not an intrinsic call!");
3396 switch (ID) {
3397 case Intrinsic::assume:
3398 case Intrinsic::lifetime_end:
3399 case Intrinsic::lifetime_start:
3400 scalarizeInstruction(it);
3401 break;
3402 default:
3403 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3404 for (unsigned Part = 0; Part < UF; ++Part) {
3405 SmallVector<Value *, 4> Args;
3406 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3407 if (HasScalarOpd && i == 1) {
3408 Args.push_back(CI->getArgOperand(i));
3409 continue;
3410 }
3411 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3412 Args.push_back(Arg[Part]);
3413 }
3414 Type *Tys[] = {CI->getType()};
3415 if (VF > 1)
3416 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3417
3418 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3419 Entry[Part] = Builder.CreateCall(F, Args);
3420 }
3421
3422 propagateMetadata(Entry, it);
3423 break;
3424 }
3425 break;
3426 }
3427
3428 default:
3429 // All other instructions are unsupported. Scalarize them.
3430 scalarizeInstruction(it);
3431 break;
3432 }// end of switch.
3433 }// end of for_each instr.
3434 }
3435
updateAnalysis()3436 void InnerLoopVectorizer::updateAnalysis() {
3437 // Forget the original basic block.
3438 SE->forgetLoop(OrigLoop);
3439
3440 // Update the dominator tree information.
3441 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3442 "Entry does not dominate exit.");
3443
3444 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3445 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3446 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3447
3448 // Due to if predication of stores we might create a sequence of "if(pred)
3449 // a[i] = ...; " blocks.
3450 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3451 if (i == 0)
3452 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3453 else if (isPredicatedBlock(i)) {
3454 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3455 } else {
3456 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3457 }
3458 }
3459
3460 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3461 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3462 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3463 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3464
3465 DEBUG(DT->verifyDomTree());
3466 }
3467
3468 /// \brief Check whether it is safe to if-convert this phi node.
3469 ///
3470 /// Phi nodes with constant expressions that can trap are not safe to if
3471 /// convert.
canIfConvertPHINodes(BasicBlock * BB)3472 static bool canIfConvertPHINodes(BasicBlock *BB) {
3473 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3474 PHINode *Phi = dyn_cast<PHINode>(I);
3475 if (!Phi)
3476 return true;
3477 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3478 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3479 if (C->canTrap())
3480 return false;
3481 }
3482 return true;
3483 }
3484
canVectorizeWithIfConvert()3485 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3486 if (!EnableIfConversion) {
3487 emitAnalysis(Report() << "if-conversion is disabled");
3488 return false;
3489 }
3490
3491 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3492
3493 // A list of pointers that we can safely read and write to.
3494 SmallPtrSet<Value *, 8> SafePointes;
3495
3496 // Collect safe addresses.
3497 for (Loop::block_iterator BI = TheLoop->block_begin(),
3498 BE = TheLoop->block_end(); BI != BE; ++BI) {
3499 BasicBlock *BB = *BI;
3500
3501 if (blockNeedsPredication(BB))
3502 continue;
3503
3504 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3505 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3506 SafePointes.insert(LI->getPointerOperand());
3507 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3508 SafePointes.insert(SI->getPointerOperand());
3509 }
3510 }
3511
3512 // Collect the blocks that need predication.
3513 BasicBlock *Header = TheLoop->getHeader();
3514 for (Loop::block_iterator BI = TheLoop->block_begin(),
3515 BE = TheLoop->block_end(); BI != BE; ++BI) {
3516 BasicBlock *BB = *BI;
3517
3518 // We don't support switch statements inside loops.
3519 if (!isa<BranchInst>(BB->getTerminator())) {
3520 emitAnalysis(Report(BB->getTerminator())
3521 << "loop contains a switch statement");
3522 return false;
3523 }
3524
3525 // We must be able to predicate all blocks that need to be predicated.
3526 if (blockNeedsPredication(BB)) {
3527 if (!blockCanBePredicated(BB, SafePointes)) {
3528 emitAnalysis(Report(BB->getTerminator())
3529 << "control flow cannot be substituted for a select");
3530 return false;
3531 }
3532 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3533 emitAnalysis(Report(BB->getTerminator())
3534 << "control flow cannot be substituted for a select");
3535 return false;
3536 }
3537 }
3538
3539 // We can if-convert this loop.
3540 return true;
3541 }
3542
canVectorize()3543 bool LoopVectorizationLegality::canVectorize() {
3544 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3545 // be canonicalized.
3546 if (!TheLoop->getLoopPreheader()) {
3547 emitAnalysis(
3548 Report() << "loop control flow is not understood by vectorizer");
3549 return false;
3550 }
3551
3552 // We can only vectorize innermost loops.
3553 if (TheLoop->getSubLoopsVector().size()) {
3554 emitAnalysis(Report() << "loop is not the innermost loop");
3555 return false;
3556 }
3557
3558 // We must have a single backedge.
3559 if (TheLoop->getNumBackEdges() != 1) {
3560 emitAnalysis(
3561 Report() << "loop control flow is not understood by vectorizer");
3562 return false;
3563 }
3564
3565 // We must have a single exiting block.
3566 if (!TheLoop->getExitingBlock()) {
3567 emitAnalysis(
3568 Report() << "loop control flow is not understood by vectorizer");
3569 return false;
3570 }
3571
3572 // We only handle bottom-tested loops, i.e. loop in which the condition is
3573 // checked at the end of each iteration. With that we can assume that all
3574 // instructions in the loop are executed the same number of times.
3575 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3576 emitAnalysis(
3577 Report() << "loop control flow is not understood by vectorizer");
3578 return false;
3579 }
3580
3581 // We need to have a loop header.
3582 DEBUG(dbgs() << "LV: Found a loop: " <<
3583 TheLoop->getHeader()->getName() << '\n');
3584
3585 // Check if we can if-convert non-single-bb loops.
3586 unsigned NumBlocks = TheLoop->getNumBlocks();
3587 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3588 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3589 return false;
3590 }
3591
3592 // ScalarEvolution needs to be able to find the exit count.
3593 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3594 if (ExitCount == SE->getCouldNotCompute()) {
3595 emitAnalysis(Report() << "could not determine number of loop iterations");
3596 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3597 return false;
3598 }
3599
3600 // Check if we can vectorize the instructions and CFG in this loop.
3601 if (!canVectorizeInstrs()) {
3602 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3603 return false;
3604 }
3605
3606 // Go over each instruction and look at memory deps.
3607 if (!canVectorizeMemory()) {
3608 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3609 return false;
3610 }
3611
3612 // Collect all of the variables that remain uniform after vectorization.
3613 collectLoopUniforms();
3614
3615 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3616 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3617 <<"!\n");
3618
3619 // Okay! We can vectorize. At this point we don't have any other mem analysis
3620 // which may limit our maximum vectorization factor, so just return true with
3621 // no restrictions.
3622 return true;
3623 }
3624
convertPointerToIntegerType(const DataLayout & DL,Type * Ty)3625 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3626 if (Ty->isPointerTy())
3627 return DL.getIntPtrType(Ty);
3628
3629 // It is possible that char's or short's overflow when we ask for the loop's
3630 // trip count, work around this by changing the type size.
3631 if (Ty->getScalarSizeInBits() < 32)
3632 return Type::getInt32Ty(Ty->getContext());
3633
3634 return Ty;
3635 }
3636
getWiderType(const DataLayout & DL,Type * Ty0,Type * Ty1)3637 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3638 Ty0 = convertPointerToIntegerType(DL, Ty0);
3639 Ty1 = convertPointerToIntegerType(DL, Ty1);
3640 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3641 return Ty0;
3642 return Ty1;
3643 }
3644
3645 /// \brief Check that the instruction has outside loop users and is not an
3646 /// identified reduction variable.
hasOutsideLoopUser(const Loop * TheLoop,Instruction * Inst,SmallPtrSetImpl<Value * > & Reductions)3647 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3648 SmallPtrSetImpl<Value *> &Reductions) {
3649 // Reduction instructions are allowed to have exit users. All other
3650 // instructions must not have external users.
3651 if (!Reductions.count(Inst))
3652 //Check that all of the users of the loop are inside the BB.
3653 for (User *U : Inst->users()) {
3654 Instruction *UI = cast<Instruction>(U);
3655 // This user may be a reduction exit value.
3656 if (!TheLoop->contains(UI)) {
3657 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3658 return true;
3659 }
3660 }
3661 return false;
3662 }
3663
canVectorizeInstrs()3664 bool LoopVectorizationLegality::canVectorizeInstrs() {
3665 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3666 BasicBlock *Header = TheLoop->getHeader();
3667
3668 // Look for the attribute signaling the absence of NaNs.
3669 Function &F = *Header->getParent();
3670 if (F.hasFnAttribute("no-nans-fp-math"))
3671 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3672 AttributeSet::FunctionIndex,
3673 "no-nans-fp-math").getValueAsString() == "true";
3674
3675 // For each block in the loop.
3676 for (Loop::block_iterator bb = TheLoop->block_begin(),
3677 be = TheLoop->block_end(); bb != be; ++bb) {
3678
3679 // Scan the instructions in the block and look for hazards.
3680 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3681 ++it) {
3682
3683 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3684 Type *PhiTy = Phi->getType();
3685 // Check that this PHI type is allowed.
3686 if (!PhiTy->isIntegerTy() &&
3687 !PhiTy->isFloatingPointTy() &&
3688 !PhiTy->isPointerTy()) {
3689 emitAnalysis(Report(it)
3690 << "loop control flow is not understood by vectorizer");
3691 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3692 return false;
3693 }
3694
3695 // If this PHINode is not in the header block, then we know that we
3696 // can convert it to select during if-conversion. No need to check if
3697 // the PHIs in this block are induction or reduction variables.
3698 if (*bb != Header) {
3699 // Check that this instruction has no outside users or is an
3700 // identified reduction value with an outside user.
3701 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3702 continue;
3703 emitAnalysis(Report(it) << "value could not be identified as "
3704 "an induction or reduction variable");
3705 return false;
3706 }
3707
3708 // We only allow if-converted PHIs with exactly two incoming values.
3709 if (Phi->getNumIncomingValues() != 2) {
3710 emitAnalysis(Report(it)
3711 << "control flow not understood by vectorizer");
3712 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3713 return false;
3714 }
3715
3716 // This is the value coming from the preheader.
3717 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3718 // Check if this is an induction variable.
3719 InductionKind IK = isInductionVariable(Phi);
3720
3721 if (IK_NoInduction != IK) {
3722 // Get the widest type.
3723 if (!WidestIndTy)
3724 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3725 else
3726 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3727
3728 // Int inductions are special because we only allow one IV.
3729 if (IK == IK_IntInduction) {
3730 // Use the phi node with the widest type as induction. Use the last
3731 // one if there are multiple (no good reason for doing this other
3732 // than it is expedient).
3733 if (!Induction || PhiTy == WidestIndTy)
3734 Induction = Phi;
3735 }
3736
3737 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3738 Inductions[Phi] = InductionInfo(StartValue, IK);
3739
3740 // Until we explicitly handle the case of an induction variable with
3741 // an outside loop user we have to give up vectorizing this loop.
3742 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3743 emitAnalysis(Report(it) << "use of induction value outside of the "
3744 "loop is not handled by vectorizer");
3745 return false;
3746 }
3747
3748 continue;
3749 }
3750
3751 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3752 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3753 continue;
3754 }
3755 if (AddReductionVar(Phi, RK_IntegerMult)) {
3756 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3757 continue;
3758 }
3759 if (AddReductionVar(Phi, RK_IntegerOr)) {
3760 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3761 continue;
3762 }
3763 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3764 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3765 continue;
3766 }
3767 if (AddReductionVar(Phi, RK_IntegerXor)) {
3768 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3769 continue;
3770 }
3771 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3772 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3773 continue;
3774 }
3775 if (AddReductionVar(Phi, RK_FloatMult)) {
3776 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3777 continue;
3778 }
3779 if (AddReductionVar(Phi, RK_FloatAdd)) {
3780 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3781 continue;
3782 }
3783 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3784 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3785 "\n");
3786 continue;
3787 }
3788
3789 emitAnalysis(Report(it) << "value that could not be identified as "
3790 "reduction is used outside the loop");
3791 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3792 return false;
3793 }// end of PHI handling
3794
3795 // We still don't handle functions. However, we can ignore dbg intrinsic
3796 // calls and we do handle certain intrinsic and libm functions.
3797 CallInst *CI = dyn_cast<CallInst>(it);
3798 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3799 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3800 DEBUG(dbgs() << "LV: Found a call site.\n");
3801 return false;
3802 }
3803
3804 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3805 // second argument is the same (i.e. loop invariant)
3806 if (CI &&
3807 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3808 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3809 emitAnalysis(Report(it)
3810 << "intrinsic instruction cannot be vectorized");
3811 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3812 return false;
3813 }
3814 }
3815
3816 // Check that the instruction return type is vectorizable.
3817 // Also, we can't vectorize extractelement instructions.
3818 if ((!VectorType::isValidElementType(it->getType()) &&
3819 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3820 emitAnalysis(Report(it)
3821 << "instruction return type cannot be vectorized");
3822 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3823 return false;
3824 }
3825
3826 // Check that the stored type is vectorizable.
3827 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3828 Type *T = ST->getValueOperand()->getType();
3829 if (!VectorType::isValidElementType(T)) {
3830 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3831 return false;
3832 }
3833 if (EnableMemAccessVersioning)
3834 collectStridedAccess(ST);
3835 }
3836
3837 if (EnableMemAccessVersioning)
3838 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3839 collectStridedAccess(LI);
3840
3841 // Reduction instructions are allowed to have exit users.
3842 // All other instructions must not have external users.
3843 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3844 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3845 return false;
3846 }
3847
3848 } // next instr.
3849
3850 }
3851
3852 if (!Induction) {
3853 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3854 if (Inductions.empty()) {
3855 emitAnalysis(Report()
3856 << "loop induction variable could not be identified");
3857 return false;
3858 }
3859 }
3860
3861 return true;
3862 }
3863
3864 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3865 /// return the induction operand of the gep pointer.
stripGetElementPtr(Value * Ptr,ScalarEvolution * SE,const DataLayout * DL,Loop * Lp)3866 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3867 const DataLayout *DL, Loop *Lp) {
3868 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3869 if (!GEP)
3870 return Ptr;
3871
3872 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3873
3874 // Check that all of the gep indices are uniform except for our induction
3875 // operand.
3876 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3877 if (i != InductionOperand &&
3878 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3879 return Ptr;
3880 return GEP->getOperand(InductionOperand);
3881 }
3882
3883 ///\brief Look for a cast use of the passed value.
getUniqueCastUse(Value * Ptr,Loop * Lp,Type * Ty)3884 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3885 Value *UniqueCast = nullptr;
3886 for (User *U : Ptr->users()) {
3887 CastInst *CI = dyn_cast<CastInst>(U);
3888 if (CI && CI->getType() == Ty) {
3889 if (!UniqueCast)
3890 UniqueCast = CI;
3891 else
3892 return nullptr;
3893 }
3894 }
3895 return UniqueCast;
3896 }
3897
3898 ///\brief Get the stride of a pointer access in a loop.
3899 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3900 /// pointer to the Value, or null otherwise.
getStrideFromPointer(Value * Ptr,ScalarEvolution * SE,const DataLayout * DL,Loop * Lp)3901 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3902 const DataLayout *DL, Loop *Lp) {
3903 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3904 if (!PtrTy || PtrTy->isAggregateType())
3905 return nullptr;
3906
3907 // Try to remove a gep instruction to make the pointer (actually index at this
3908 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3909 // pointer, otherwise, we are analyzing the index.
3910 Value *OrigPtr = Ptr;
3911
3912 // The size of the pointer access.
3913 int64_t PtrAccessSize = 1;
3914
3915 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3916 const SCEV *V = SE->getSCEV(Ptr);
3917
3918 if (Ptr != OrigPtr)
3919 // Strip off casts.
3920 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3921 V = C->getOperand();
3922
3923 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3924 if (!S)
3925 return nullptr;
3926
3927 V = S->getStepRecurrence(*SE);
3928 if (!V)
3929 return nullptr;
3930
3931 // Strip off the size of access multiplication if we are still analyzing the
3932 // pointer.
3933 if (OrigPtr == Ptr) {
3934 DL->getTypeAllocSize(PtrTy->getElementType());
3935 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3936 if (M->getOperand(0)->getSCEVType() != scConstant)
3937 return nullptr;
3938
3939 const APInt &APStepVal =
3940 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3941
3942 // Huge step value - give up.
3943 if (APStepVal.getBitWidth() > 64)
3944 return nullptr;
3945
3946 int64_t StepVal = APStepVal.getSExtValue();
3947 if (PtrAccessSize != StepVal)
3948 return nullptr;
3949 V = M->getOperand(1);
3950 }
3951 }
3952
3953 // Strip off casts.
3954 Type *StripedOffRecurrenceCast = nullptr;
3955 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3956 StripedOffRecurrenceCast = C->getType();
3957 V = C->getOperand();
3958 }
3959
3960 // Look for the loop invariant symbolic value.
3961 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3962 if (!U)
3963 return nullptr;
3964
3965 Value *Stride = U->getValue();
3966 if (!Lp->isLoopInvariant(Stride))
3967 return nullptr;
3968
3969 // If we have stripped off the recurrence cast we have to make sure that we
3970 // return the value that is used in this loop so that we can replace it later.
3971 if (StripedOffRecurrenceCast)
3972 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3973
3974 return Stride;
3975 }
3976
collectStridedAccess(Value * MemAccess)3977 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3978 Value *Ptr = nullptr;
3979 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3980 Ptr = LI->getPointerOperand();
3981 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3982 Ptr = SI->getPointerOperand();
3983 else
3984 return;
3985
3986 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3987 if (!Stride)
3988 return;
3989
3990 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3991 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3992 Strides[Ptr] = Stride;
3993 StrideSet.insert(Stride);
3994 }
3995
collectLoopUniforms()3996 void LoopVectorizationLegality::collectLoopUniforms() {
3997 // We now know that the loop is vectorizable!
3998 // Collect variables that will remain uniform after vectorization.
3999 std::vector<Value*> Worklist;
4000 BasicBlock *Latch = TheLoop->getLoopLatch();
4001
4002 // Start with the conditional branch and walk up the block.
4003 Worklist.push_back(Latch->getTerminator()->getOperand(0));
4004
4005 // Also add all consecutive pointer values; these values will be uniform
4006 // after vectorization (and subsequent cleanup) and, until revectorization is
4007 // supported, all dependencies must also be uniform.
4008 for (Loop::block_iterator B = TheLoop->block_begin(),
4009 BE = TheLoop->block_end(); B != BE; ++B)
4010 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
4011 I != IE; ++I)
4012 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
4013 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4014
4015 while (Worklist.size()) {
4016 Instruction *I = dyn_cast<Instruction>(Worklist.back());
4017 Worklist.pop_back();
4018
4019 // Look at instructions inside this loop.
4020 // Stop when reaching PHI nodes.
4021 // TODO: we need to follow values all over the loop, not only in this block.
4022 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
4023 continue;
4024
4025 // This is a known uniform.
4026 Uniforms.insert(I);
4027
4028 // Insert all operands.
4029 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
4030 }
4031 }
4032
4033 namespace {
4034 /// \brief Analyses memory accesses in a loop.
4035 ///
4036 /// Checks whether run time pointer checks are needed and builds sets for data
4037 /// dependence checking.
4038 class AccessAnalysis {
4039 public:
4040 /// \brief Read or write access location.
4041 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4042 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4043
4044 /// \brief Set of potential dependent memory accesses.
4045 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
4046
AccessAnalysis(const DataLayout * Dl,AliasAnalysis * AA,DepCandidates & DA)4047 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
4048 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
4049
4050 /// \brief Register a load and whether it is only read from.
addLoad(AliasAnalysis::Location & Loc,bool IsReadOnly)4051 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
4052 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4053 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4054 Accesses.insert(MemAccessInfo(Ptr, false));
4055 if (IsReadOnly)
4056 ReadOnlyPtr.insert(Ptr);
4057 }
4058
4059 /// \brief Register a store.
addStore(AliasAnalysis::Location & Loc)4060 void addStore(AliasAnalysis::Location &Loc) {
4061 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4062 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4063 Accesses.insert(MemAccessInfo(Ptr, true));
4064 }
4065
4066 /// \brief Check whether we can check the pointers at runtime for
4067 /// non-intersection.
4068 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4069 unsigned &NumComparisons, ScalarEvolution *SE,
4070 Loop *TheLoop, ValueToValueMap &Strides,
4071 bool ShouldCheckStride = false);
4072
4073 /// \brief Goes over all memory accesses, checks whether a RT check is needed
4074 /// and builds sets of dependent accesses.
buildDependenceSets()4075 void buildDependenceSets() {
4076 processMemAccesses();
4077 }
4078
isRTCheckNeeded()4079 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
4080
isDependencyCheckNeeded()4081 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
resetDepChecks()4082 void resetDepChecks() { CheckDeps.clear(); }
4083
getDependenciesToCheck()4084 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
4085
4086 private:
4087 typedef SetVector<MemAccessInfo> PtrAccessSet;
4088
4089 /// \brief Go over all memory access and check whether runtime pointer checks
4090 /// are needed /// and build sets of dependency check candidates.
4091 void processMemAccesses();
4092
4093 /// Set of all accesses.
4094 PtrAccessSet Accesses;
4095
4096 /// Set of accesses that need a further dependence check.
4097 MemAccessInfoSet CheckDeps;
4098
4099 /// Set of pointers that are read only.
4100 SmallPtrSet<Value*, 16> ReadOnlyPtr;
4101
4102 const DataLayout *DL;
4103
4104 /// An alias set tracker to partition the access set by underlying object and
4105 //intrinsic property (such as TBAA metadata).
4106 AliasSetTracker AST;
4107
4108 /// Sets of potentially dependent accesses - members of one set share an
4109 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
4110 /// dependence check.
4111 DepCandidates &DepCands;
4112
4113 bool IsRTCheckNeeded;
4114 };
4115
4116 } // end anonymous namespace
4117
4118 /// \brief Check whether a pointer can participate in a runtime bounds check.
hasComputableBounds(ScalarEvolution * SE,ValueToValueMap & Strides,Value * Ptr)4119 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4120 Value *Ptr) {
4121 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4122 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4123 if (!AR)
4124 return false;
4125
4126 return AR->isAffine();
4127 }
4128
4129 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4130 /// the address space.
4131 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4132 const Loop *Lp, ValueToValueMap &StridesMap);
4133
canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck & RtCheck,unsigned & NumComparisons,ScalarEvolution * SE,Loop * TheLoop,ValueToValueMap & StridesMap,bool ShouldCheckStride)4134 bool AccessAnalysis::canCheckPtrAtRT(
4135 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4136 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4137 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4138 // Find pointers with computable bounds. We are going to use this information
4139 // to place a runtime bound check.
4140 bool CanDoRT = true;
4141
4142 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4143 NumComparisons = 0;
4144
4145 // We assign a consecutive id to access from different alias sets.
4146 // Accesses between different groups doesn't need to be checked.
4147 unsigned ASId = 1;
4148 for (auto &AS : AST) {
4149 unsigned NumReadPtrChecks = 0;
4150 unsigned NumWritePtrChecks = 0;
4151
4152 // We assign consecutive id to access from different dependence sets.
4153 // Accesses within the same set don't need a runtime check.
4154 unsigned RunningDepId = 1;
4155 DenseMap<Value *, unsigned> DepSetId;
4156
4157 for (auto A : AS) {
4158 Value *Ptr = A.getValue();
4159 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4160 MemAccessInfo Access(Ptr, IsWrite);
4161
4162 if (IsWrite)
4163 ++NumWritePtrChecks;
4164 else
4165 ++NumReadPtrChecks;
4166
4167 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4168 // When we run after a failing dependency check we have to make sure we
4169 // don't have wrapping pointers.
4170 (!ShouldCheckStride ||
4171 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4172 // The id of the dependence set.
4173 unsigned DepId;
4174
4175 if (IsDepCheckNeeded) {
4176 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4177 unsigned &LeaderId = DepSetId[Leader];
4178 if (!LeaderId)
4179 LeaderId = RunningDepId++;
4180 DepId = LeaderId;
4181 } else
4182 // Each access has its own dependence set.
4183 DepId = RunningDepId++;
4184
4185 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4186
4187 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4188 } else {
4189 CanDoRT = false;
4190 }
4191 }
4192
4193 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4194 NumComparisons += 0; // Only one dependence set.
4195 else {
4196 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4197 NumWritePtrChecks - 1));
4198 }
4199
4200 ++ASId;
4201 }
4202
4203 // If the pointers that we would use for the bounds comparison have different
4204 // address spaces, assume the values aren't directly comparable, so we can't
4205 // use them for the runtime check. We also have to assume they could
4206 // overlap. In the future there should be metadata for whether address spaces
4207 // are disjoint.
4208 unsigned NumPointers = RtCheck.Pointers.size();
4209 for (unsigned i = 0; i < NumPointers; ++i) {
4210 for (unsigned j = i + 1; j < NumPointers; ++j) {
4211 // Only need to check pointers between two different dependency sets.
4212 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4213 continue;
4214 // Only need to check pointers in the same alias set.
4215 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4216 continue;
4217
4218 Value *PtrI = RtCheck.Pointers[i];
4219 Value *PtrJ = RtCheck.Pointers[j];
4220
4221 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4222 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4223 if (ASi != ASj) {
4224 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4225 " different address spaces\n");
4226 return false;
4227 }
4228 }
4229 }
4230
4231 return CanDoRT;
4232 }
4233
processMemAccesses()4234 void AccessAnalysis::processMemAccesses() {
4235 // We process the set twice: first we process read-write pointers, last we
4236 // process read-only pointers. This allows us to skip dependence tests for
4237 // read-only pointers.
4238
4239 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4240 DEBUG(dbgs() << " AST: "; AST.dump());
4241 DEBUG(dbgs() << "LV: Accesses:\n");
4242 DEBUG({
4243 for (auto A : Accesses)
4244 dbgs() << "\t" << *A.getPointer() << " (" <<
4245 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4246 "read-only" : "read")) << ")\n";
4247 });
4248
4249 // The AliasSetTracker has nicely partitioned our pointers by metadata
4250 // compatibility and potential for underlying-object overlap. As a result, we
4251 // only need to check for potential pointer dependencies within each alias
4252 // set.
4253 for (auto &AS : AST) {
4254 // Note that both the alias-set tracker and the alias sets themselves used
4255 // linked lists internally and so the iteration order here is deterministic
4256 // (matching the original instruction order within each set).
4257
4258 bool SetHasWrite = false;
4259
4260 // Map of pointers to last access encountered.
4261 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4262 UnderlyingObjToAccessMap ObjToLastAccess;
4263
4264 // Set of access to check after all writes have been processed.
4265 PtrAccessSet DeferredAccesses;
4266
4267 // Iterate over each alias set twice, once to process read/write pointers,
4268 // and then to process read-only pointers.
4269 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4270 bool UseDeferred = SetIteration > 0;
4271 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4272
4273 for (auto AV : AS) {
4274 Value *Ptr = AV.getValue();
4275
4276 // For a single memory access in AliasSetTracker, Accesses may contain
4277 // both read and write, and they both need to be handled for CheckDeps.
4278 for (auto AC : S) {
4279 if (AC.getPointer() != Ptr)
4280 continue;
4281
4282 bool IsWrite = AC.getInt();
4283
4284 // If we're using the deferred access set, then it contains only
4285 // reads.
4286 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4287 if (UseDeferred && !IsReadOnlyPtr)
4288 continue;
4289 // Otherwise, the pointer must be in the PtrAccessSet, either as a
4290 // read or a write.
4291 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4292 S.count(MemAccessInfo(Ptr, false))) &&
4293 "Alias-set pointer not in the access set?");
4294
4295 MemAccessInfo Access(Ptr, IsWrite);
4296 DepCands.insert(Access);
4297
4298 // Memorize read-only pointers for later processing and skip them in
4299 // the first round (they need to be checked after we have seen all
4300 // write pointers). Note: we also mark pointer that are not
4301 // consecutive as "read-only" pointers (so that we check
4302 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
4303 if (!UseDeferred && IsReadOnlyPtr) {
4304 DeferredAccesses.insert(Access);
4305 continue;
4306 }
4307
4308 // If this is a write - check other reads and writes for conflicts. If
4309 // this is a read only check other writes for conflicts (but only if
4310 // there is no other write to the ptr - this is an optimization to
4311 // catch "a[i] = a[i] + " without having to do a dependence check).
4312 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4313 CheckDeps.insert(Access);
4314 IsRTCheckNeeded = true;
4315 }
4316
4317 if (IsWrite)
4318 SetHasWrite = true;
4319
4320 // Create sets of pointers connected by a shared alias set and
4321 // underlying object.
4322 typedef SmallVector<Value *, 16> ValueVector;
4323 ValueVector TempObjects;
4324 GetUnderlyingObjects(Ptr, TempObjects, DL);
4325 for (Value *UnderlyingObj : TempObjects) {
4326 UnderlyingObjToAccessMap::iterator Prev =
4327 ObjToLastAccess.find(UnderlyingObj);
4328 if (Prev != ObjToLastAccess.end())
4329 DepCands.unionSets(Access, Prev->second);
4330
4331 ObjToLastAccess[UnderlyingObj] = Access;
4332 }
4333 }
4334 }
4335 }
4336 }
4337 }
4338
4339 namespace {
4340 /// \brief Checks memory dependences among accesses to the same underlying
4341 /// object to determine whether there vectorization is legal or not (and at
4342 /// which vectorization factor).
4343 ///
4344 /// This class works under the assumption that we already checked that memory
4345 /// locations with different underlying pointers are "must-not alias".
4346 /// We use the ScalarEvolution framework to symbolically evalutate access
4347 /// functions pairs. Since we currently don't restructure the loop we can rely
4348 /// on the program order of memory accesses to determine their safety.
4349 /// At the moment we will only deem accesses as safe for:
4350 /// * A negative constant distance assuming program order.
4351 ///
4352 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4353 /// a[i] = tmp; y = a[i];
4354 ///
4355 /// The latter case is safe because later checks guarantuee that there can't
4356 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4357 /// the same variable: a header phi can only be an induction or a reduction, a
4358 /// reduction can't have a memory sink, an induction can't have a memory
4359 /// source). This is important and must not be violated (or we have to
4360 /// resort to checking for cycles through memory).
4361 ///
4362 /// * A positive constant distance assuming program order that is bigger
4363 /// than the biggest memory access.
4364 ///
4365 /// tmp = a[i] OR b[i] = x
4366 /// a[i+2] = tmp y = b[i+2];
4367 ///
4368 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4369 ///
4370 /// * Zero distances and all accesses have the same size.
4371 ///
4372 class MemoryDepChecker {
4373 public:
4374 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4375 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4376
MemoryDepChecker(ScalarEvolution * Se,const DataLayout * Dl,const Loop * L)4377 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4378 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4379 ShouldRetryWithRuntimeCheck(false) {}
4380
4381 /// \brief Register the location (instructions are given increasing numbers)
4382 /// of a write access.
addAccess(StoreInst * SI)4383 void addAccess(StoreInst *SI) {
4384 Value *Ptr = SI->getPointerOperand();
4385 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4386 InstMap.push_back(SI);
4387 ++AccessIdx;
4388 }
4389
4390 /// \brief Register the location (instructions are given increasing numbers)
4391 /// of a write access.
addAccess(LoadInst * LI)4392 void addAccess(LoadInst *LI) {
4393 Value *Ptr = LI->getPointerOperand();
4394 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4395 InstMap.push_back(LI);
4396 ++AccessIdx;
4397 }
4398
4399 /// \brief Check whether the dependencies between the accesses are safe.
4400 ///
4401 /// Only checks sets with elements in \p CheckDeps.
4402 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4403 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4404
4405 /// \brief The maximum number of bytes of a vector register we can vectorize
4406 /// the accesses safely with.
getMaxSafeDepDistBytes()4407 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4408
4409 /// \brief In same cases when the dependency check fails we can still
4410 /// vectorize the loop with a dynamic array access check.
shouldRetryWithRuntimeCheck()4411 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4412
4413 private:
4414 ScalarEvolution *SE;
4415 const DataLayout *DL;
4416 const Loop *InnermostLoop;
4417
4418 /// \brief Maps access locations (ptr, read/write) to program order.
4419 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4420
4421 /// \brief Memory access instructions in program order.
4422 SmallVector<Instruction *, 16> InstMap;
4423
4424 /// \brief The program order index to be used for the next instruction.
4425 unsigned AccessIdx;
4426
4427 // We can access this many bytes in parallel safely.
4428 unsigned MaxSafeDepDistBytes;
4429
4430 /// \brief If we see a non-constant dependence distance we can still try to
4431 /// vectorize this loop with runtime checks.
4432 bool ShouldRetryWithRuntimeCheck;
4433
4434 /// \brief Check whether there is a plausible dependence between the two
4435 /// accesses.
4436 ///
4437 /// Access \p A must happen before \p B in program order. The two indices
4438 /// identify the index into the program order map.
4439 ///
4440 /// This function checks whether there is a plausible dependence (or the
4441 /// absence of such can't be proved) between the two accesses. If there is a
4442 /// plausible dependence but the dependence distance is bigger than one
4443 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4444 /// distance is smaller than any other distance encountered so far).
4445 /// Otherwise, this function returns true signaling a possible dependence.
4446 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4447 const MemAccessInfo &B, unsigned BIdx,
4448 ValueToValueMap &Strides);
4449
4450 /// \brief Check whether the data dependence could prevent store-load
4451 /// forwarding.
4452 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4453 };
4454
4455 } // end anonymous namespace
4456
isInBoundsGep(Value * Ptr)4457 static bool isInBoundsGep(Value *Ptr) {
4458 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4459 return GEP->isInBounds();
4460 return false;
4461 }
4462
4463 /// \brief Check whether the access through \p Ptr has a constant stride.
isStridedPtr(ScalarEvolution * SE,const DataLayout * DL,Value * Ptr,const Loop * Lp,ValueToValueMap & StridesMap)4464 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4465 const Loop *Lp, ValueToValueMap &StridesMap) {
4466 const Type *Ty = Ptr->getType();
4467 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4468
4469 // Make sure that the pointer does not point to aggregate types.
4470 const PointerType *PtrTy = cast<PointerType>(Ty);
4471 if (PtrTy->getElementType()->isAggregateType()) {
4472 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4473 "\n");
4474 return 0;
4475 }
4476
4477 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4478
4479 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4480 if (!AR) {
4481 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4482 << *Ptr << " SCEV: " << *PtrScev << "\n");
4483 return 0;
4484 }
4485
4486 // The accesss function must stride over the innermost loop.
4487 if (Lp != AR->getLoop()) {
4488 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4489 *Ptr << " SCEV: " << *PtrScev << "\n");
4490 }
4491
4492 // The address calculation must not wrap. Otherwise, a dependence could be
4493 // inverted.
4494 // An inbounds getelementptr that is a AddRec with a unit stride
4495 // cannot wrap per definition. The unit stride requirement is checked later.
4496 // An getelementptr without an inbounds attribute and unit stride would have
4497 // to access the pointer value "0" which is undefined behavior in address
4498 // space 0, therefore we can also vectorize this case.
4499 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4500 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4501 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4502 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4503 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4504 << *Ptr << " SCEV: " << *PtrScev << "\n");
4505 return 0;
4506 }
4507
4508 // Check the step is constant.
4509 const SCEV *Step = AR->getStepRecurrence(*SE);
4510
4511 // Calculate the pointer stride and check if it is consecutive.
4512 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4513 if (!C) {
4514 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4515 " SCEV: " << *PtrScev << "\n");
4516 return 0;
4517 }
4518
4519 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4520 const APInt &APStepVal = C->getValue()->getValue();
4521
4522 // Huge step value - give up.
4523 if (APStepVal.getBitWidth() > 64)
4524 return 0;
4525
4526 int64_t StepVal = APStepVal.getSExtValue();
4527
4528 // Strided access.
4529 int64_t Stride = StepVal / Size;
4530 int64_t Rem = StepVal % Size;
4531 if (Rem)
4532 return 0;
4533
4534 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4535 // know we can't "wrap around the address space". In case of address space
4536 // zero we know that this won't happen without triggering undefined behavior.
4537 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4538 Stride != 1 && Stride != -1)
4539 return 0;
4540
4541 return Stride;
4542 }
4543
couldPreventStoreLoadForward(unsigned Distance,unsigned TypeByteSize)4544 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4545 unsigned TypeByteSize) {
4546 // If loads occur at a distance that is not a multiple of a feasible vector
4547 // factor store-load forwarding does not take place.
4548 // Positive dependences might cause troubles because vectorizing them might
4549 // prevent store-load forwarding making vectorized code run a lot slower.
4550 // a[i] = a[i-3] ^ a[i-8];
4551 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4552 // hence on your typical architecture store-load forwarding does not take
4553 // place. Vectorizing in such cases does not make sense.
4554 // Store-load forwarding distance.
4555 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4556 // Maximum vector factor.
4557 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4558 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4559 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4560
4561 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4562 vf *= 2) {
4563 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4564 MaxVFWithoutSLForwardIssues = (vf >>=1);
4565 break;
4566 }
4567 }
4568
4569 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4570 DEBUG(dbgs() << "LV: Distance " << Distance <<
4571 " that could cause a store-load forwarding conflict\n");
4572 return true;
4573 }
4574
4575 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4576 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4577 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4578 return false;
4579 }
4580
isDependent(const MemAccessInfo & A,unsigned AIdx,const MemAccessInfo & B,unsigned BIdx,ValueToValueMap & Strides)4581 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4582 const MemAccessInfo &B, unsigned BIdx,
4583 ValueToValueMap &Strides) {
4584 assert (AIdx < BIdx && "Must pass arguments in program order");
4585
4586 Value *APtr = A.getPointer();
4587 Value *BPtr = B.getPointer();
4588 bool AIsWrite = A.getInt();
4589 bool BIsWrite = B.getInt();
4590
4591 // Two reads are independent.
4592 if (!AIsWrite && !BIsWrite)
4593 return false;
4594
4595 // We cannot check pointers in different address spaces.
4596 if (APtr->getType()->getPointerAddressSpace() !=
4597 BPtr->getType()->getPointerAddressSpace())
4598 return true;
4599
4600 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4601 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4602
4603 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4604 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4605
4606 const SCEV *Src = AScev;
4607 const SCEV *Sink = BScev;
4608
4609 // If the induction step is negative we have to invert source and sink of the
4610 // dependence.
4611 if (StrideAPtr < 0) {
4612 //Src = BScev;
4613 //Sink = AScev;
4614 std::swap(APtr, BPtr);
4615 std::swap(Src, Sink);
4616 std::swap(AIsWrite, BIsWrite);
4617 std::swap(AIdx, BIdx);
4618 std::swap(StrideAPtr, StrideBPtr);
4619 }
4620
4621 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4622
4623 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4624 << "(Induction step: " << StrideAPtr << ")\n");
4625 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4626 << *InstMap[BIdx] << ": " << *Dist << "\n");
4627
4628 // Need consecutive accesses. We don't want to vectorize
4629 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4630 // the address space.
4631 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4632 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4633 return true;
4634 }
4635
4636 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4637 if (!C) {
4638 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4639 ShouldRetryWithRuntimeCheck = true;
4640 return true;
4641 }
4642
4643 Type *ATy = APtr->getType()->getPointerElementType();
4644 Type *BTy = BPtr->getType()->getPointerElementType();
4645 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4646
4647 // Negative distances are not plausible dependencies.
4648 const APInt &Val = C->getValue()->getValue();
4649 if (Val.isNegative()) {
4650 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4651 if (IsTrueDataDependence &&
4652 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4653 ATy != BTy))
4654 return true;
4655
4656 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4657 return false;
4658 }
4659
4660 // Write to the same location with the same size.
4661 // Could be improved to assert type sizes are the same (i32 == float, etc).
4662 if (Val == 0) {
4663 if (ATy == BTy)
4664 return false;
4665 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4666 return true;
4667 }
4668
4669 assert(Val.isStrictlyPositive() && "Expect a positive value");
4670
4671 // Positive distance bigger than max vectorization factor.
4672 if (ATy != BTy) {
4673 DEBUG(dbgs() <<
4674 "LV: ReadWrite-Write positive dependency with different types\n");
4675 return false;
4676 }
4677
4678 unsigned Distance = (unsigned) Val.getZExtValue();
4679
4680 // Bail out early if passed-in parameters make vectorization not feasible.
4681 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4682 unsigned ForcedUnroll = VectorizationInterleave ? VectorizationInterleave : 1;
4683
4684 // The distance must be bigger than the size needed for a vectorized version
4685 // of the operation and the size of the vectorized operation must not be
4686 // bigger than the currrent maximum size.
4687 if (Distance < 2*TypeByteSize ||
4688 2*TypeByteSize > MaxSafeDepDistBytes ||
4689 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4690 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4691 << Val.getSExtValue() << '\n');
4692 return true;
4693 }
4694
4695 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4696 Distance : MaxSafeDepDistBytes;
4697
4698 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4699 if (IsTrueDataDependence &&
4700 couldPreventStoreLoadForward(Distance, TypeByteSize))
4701 return true;
4702
4703 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4704 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4705
4706 return false;
4707 }
4708
areDepsSafe(AccessAnalysis::DepCandidates & AccessSets,MemAccessInfoSet & CheckDeps,ValueToValueMap & Strides)4709 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4710 MemAccessInfoSet &CheckDeps,
4711 ValueToValueMap &Strides) {
4712
4713 MaxSafeDepDistBytes = -1U;
4714 while (!CheckDeps.empty()) {
4715 MemAccessInfo CurAccess = *CheckDeps.begin();
4716
4717 // Get the relevant memory access set.
4718 EquivalenceClasses<MemAccessInfo>::iterator I =
4719 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4720
4721 // Check accesses within this set.
4722 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4723 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4724
4725 // Check every access pair.
4726 while (AI != AE) {
4727 CheckDeps.erase(*AI);
4728 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4729 while (OI != AE) {
4730 // Check every accessing instruction pair in program order.
4731 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4732 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4733 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4734 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4735 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4736 return false;
4737 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4738 return false;
4739 }
4740 ++OI;
4741 }
4742 AI++;
4743 }
4744 }
4745 return true;
4746 }
4747
canVectorizeMemory()4748 bool LoopVectorizationLegality::canVectorizeMemory() {
4749
4750 typedef SmallVector<Value*, 16> ValueVector;
4751 typedef SmallPtrSet<Value*, 16> ValueSet;
4752
4753 // Holds the Load and Store *instructions*.
4754 ValueVector Loads;
4755 ValueVector Stores;
4756
4757 // Holds all the different accesses in the loop.
4758 unsigned NumReads = 0;
4759 unsigned NumReadWrites = 0;
4760
4761 PtrRtCheck.Pointers.clear();
4762 PtrRtCheck.Need = false;
4763
4764 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4765 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4766
4767 // For each block.
4768 for (Loop::block_iterator bb = TheLoop->block_begin(),
4769 be = TheLoop->block_end(); bb != be; ++bb) {
4770
4771 // Scan the BB and collect legal loads and stores.
4772 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4773 ++it) {
4774
4775 // If this is a load, save it. If this instruction can read from memory
4776 // but is not a load, then we quit. Notice that we don't handle function
4777 // calls that read or write.
4778 if (it->mayReadFromMemory()) {
4779 // Many math library functions read the rounding mode. We will only
4780 // vectorize a loop if it contains known function calls that don't set
4781 // the flag. Therefore, it is safe to ignore this read from memory.
4782 CallInst *Call = dyn_cast<CallInst>(it);
4783 if (Call && getIntrinsicIDForCall(Call, TLI))
4784 continue;
4785
4786 LoadInst *Ld = dyn_cast<LoadInst>(it);
4787 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4788 emitAnalysis(Report(Ld)
4789 << "read with atomic ordering or volatile read");
4790 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4791 return false;
4792 }
4793 NumLoads++;
4794 Loads.push_back(Ld);
4795 DepChecker.addAccess(Ld);
4796 continue;
4797 }
4798
4799 // Save 'store' instructions. Abort if other instructions write to memory.
4800 if (it->mayWriteToMemory()) {
4801 StoreInst *St = dyn_cast<StoreInst>(it);
4802 if (!St) {
4803 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4804 return false;
4805 }
4806 if (!St->isSimple() && !IsAnnotatedParallel) {
4807 emitAnalysis(Report(St)
4808 << "write with atomic ordering or volatile write");
4809 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4810 return false;
4811 }
4812 NumStores++;
4813 Stores.push_back(St);
4814 DepChecker.addAccess(St);
4815 }
4816 } // Next instr.
4817 } // Next block.
4818
4819 // Now we have two lists that hold the loads and the stores.
4820 // Next, we find the pointers that they use.
4821
4822 // Check if we see any stores. If there are no stores, then we don't
4823 // care if the pointers are *restrict*.
4824 if (!Stores.size()) {
4825 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4826 return true;
4827 }
4828
4829 AccessAnalysis::DepCandidates DependentAccesses;
4830 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4831
4832 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4833 // multiple times on the same object. If the ptr is accessed twice, once
4834 // for read and once for write, it will only appear once (on the write
4835 // list). This is okay, since we are going to check for conflicts between
4836 // writes and between reads and writes, but not between reads and reads.
4837 ValueSet Seen;
4838
4839 ValueVector::iterator I, IE;
4840 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4841 StoreInst *ST = cast<StoreInst>(*I);
4842 Value* Ptr = ST->getPointerOperand();
4843
4844 if (isUniform(Ptr)) {
4845 emitAnalysis(
4846 Report(ST)
4847 << "write to a loop invariant address could not be vectorized");
4848 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4849 return false;
4850 }
4851
4852 // If we did *not* see this pointer before, insert it to the read-write
4853 // list. At this phase it is only a 'write' list.
4854 if (Seen.insert(Ptr).second) {
4855 ++NumReadWrites;
4856
4857 AliasAnalysis::Location Loc = AA->getLocation(ST);
4858 // The TBAA metadata could have a control dependency on the predication
4859 // condition, so we cannot rely on it when determining whether or not we
4860 // need runtime pointer checks.
4861 if (blockNeedsPredication(ST->getParent()))
4862 Loc.AATags.TBAA = nullptr;
4863
4864 Accesses.addStore(Loc);
4865 }
4866 }
4867
4868 if (IsAnnotatedParallel) {
4869 DEBUG(dbgs()
4870 << "LV: A loop annotated parallel, ignore memory dependency "
4871 << "checks.\n");
4872 return true;
4873 }
4874
4875 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4876 LoadInst *LD = cast<LoadInst>(*I);
4877 Value* Ptr = LD->getPointerOperand();
4878 // If we did *not* see this pointer before, insert it to the
4879 // read list. If we *did* see it before, then it is already in
4880 // the read-write list. This allows us to vectorize expressions
4881 // such as A[i] += x; Because the address of A[i] is a read-write
4882 // pointer. This only works if the index of A[i] is consecutive.
4883 // If the address of i is unknown (for example A[B[i]]) then we may
4884 // read a few words, modify, and write a few words, and some of the
4885 // words may be written to the same address.
4886 bool IsReadOnlyPtr = false;
4887 if (Seen.insert(Ptr).second ||
4888 !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4889 ++NumReads;
4890 IsReadOnlyPtr = true;
4891 }
4892
4893 AliasAnalysis::Location Loc = AA->getLocation(LD);
4894 // The TBAA metadata could have a control dependency on the predication
4895 // condition, so we cannot rely on it when determining whether or not we
4896 // need runtime pointer checks.
4897 if (blockNeedsPredication(LD->getParent()))
4898 Loc.AATags.TBAA = nullptr;
4899
4900 Accesses.addLoad(Loc, IsReadOnlyPtr);
4901 }
4902
4903 // If we write (or read-write) to a single destination and there are no
4904 // other reads in this loop then is it safe to vectorize.
4905 if (NumReadWrites == 1 && NumReads == 0) {
4906 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4907 return true;
4908 }
4909
4910 // Build dependence sets and check whether we need a runtime pointer bounds
4911 // check.
4912 Accesses.buildDependenceSets();
4913 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4914
4915 // Find pointers with computable bounds. We are going to use this information
4916 // to place a runtime bound check.
4917 unsigned NumComparisons = 0;
4918 bool CanDoRT = false;
4919 if (NeedRTCheck)
4920 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4921 Strides);
4922
4923 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4924 " pointer comparisons.\n");
4925
4926 // If we only have one set of dependences to check pointers among we don't
4927 // need a runtime check.
4928 if (NumComparisons == 0 && NeedRTCheck)
4929 NeedRTCheck = false;
4930
4931 // Check that we did not collect too many pointers or found an unsizeable
4932 // pointer.
4933 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4934 PtrRtCheck.reset();
4935 CanDoRT = false;
4936 }
4937
4938 if (CanDoRT) {
4939 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4940 }
4941
4942 if (NeedRTCheck && !CanDoRT) {
4943 emitAnalysis(Report() << "cannot identify array bounds");
4944 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4945 "the array bounds.\n");
4946 PtrRtCheck.reset();
4947 return false;
4948 }
4949
4950 PtrRtCheck.Need = NeedRTCheck;
4951
4952 bool CanVecMem = true;
4953 if (Accesses.isDependencyCheckNeeded()) {
4954 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4955 CanVecMem = DepChecker.areDepsSafe(
4956 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4957 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4958
4959 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4960 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4961 NeedRTCheck = true;
4962
4963 // Clear the dependency checks. We assume they are not needed.
4964 Accesses.resetDepChecks();
4965
4966 PtrRtCheck.reset();
4967 PtrRtCheck.Need = true;
4968
4969 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4970 TheLoop, Strides, true);
4971 // Check that we did not collect too many pointers or found an unsizeable
4972 // pointer.
4973 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4974 if (!CanDoRT && NumComparisons > 0)
4975 emitAnalysis(Report()
4976 << "cannot check memory dependencies at runtime");
4977 else
4978 emitAnalysis(Report()
4979 << NumComparisons << " exceeds limit of "
4980 << RuntimeMemoryCheckThreshold
4981 << " dependent memory operations checked at runtime");
4982 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4983 PtrRtCheck.reset();
4984 return false;
4985 }
4986
4987 CanVecMem = true;
4988 }
4989 }
4990
4991 if (!CanVecMem)
4992 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4993
4994 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4995 " need a runtime memory check.\n");
4996
4997 return CanVecMem;
4998 }
4999
hasMultipleUsesOf(Instruction * I,SmallPtrSetImpl<Instruction * > & Insts)5000 static bool hasMultipleUsesOf(Instruction *I,
5001 SmallPtrSetImpl<Instruction *> &Insts) {
5002 unsigned NumUses = 0;
5003 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
5004 if (Insts.count(dyn_cast<Instruction>(*Use)))
5005 ++NumUses;
5006 if (NumUses > 1)
5007 return true;
5008 }
5009
5010 return false;
5011 }
5012
areAllUsesIn(Instruction * I,SmallPtrSetImpl<Instruction * > & Set)5013 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
5014 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
5015 if (!Set.count(dyn_cast<Instruction>(*Use)))
5016 return false;
5017 return true;
5018 }
5019
AddReductionVar(PHINode * Phi,ReductionKind Kind)5020 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
5021 ReductionKind Kind) {
5022 if (Phi->getNumIncomingValues() != 2)
5023 return false;
5024
5025 // Reduction variables are only found in the loop header block.
5026 if (Phi->getParent() != TheLoop->getHeader())
5027 return false;
5028
5029 // Obtain the reduction start value from the value that comes from the loop
5030 // preheader.
5031 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
5032
5033 // ExitInstruction is the single value which is used outside the loop.
5034 // We only allow for a single reduction value to be used outside the loop.
5035 // This includes users of the reduction, variables (which form a cycle
5036 // which ends in the phi node).
5037 Instruction *ExitInstruction = nullptr;
5038 // Indicates that we found a reduction operation in our scan.
5039 bool FoundReduxOp = false;
5040
5041 // We start with the PHI node and scan for all of the users of this
5042 // instruction. All users must be instructions that can be used as reduction
5043 // variables (such as ADD). We must have a single out-of-block user. The cycle
5044 // must include the original PHI.
5045 bool FoundStartPHI = false;
5046
5047 // To recognize min/max patterns formed by a icmp select sequence, we store
5048 // the number of instruction we saw from the recognized min/max pattern,
5049 // to make sure we only see exactly the two instructions.
5050 unsigned NumCmpSelectPatternInst = 0;
5051 ReductionInstDesc ReduxDesc(false, nullptr);
5052
5053 SmallPtrSet<Instruction *, 8> VisitedInsts;
5054 SmallVector<Instruction *, 8> Worklist;
5055 Worklist.push_back(Phi);
5056 VisitedInsts.insert(Phi);
5057
5058 // A value in the reduction can be used:
5059 // - By the reduction:
5060 // - Reduction operation:
5061 // - One use of reduction value (safe).
5062 // - Multiple use of reduction value (not safe).
5063 // - PHI:
5064 // - All uses of the PHI must be the reduction (safe).
5065 // - Otherwise, not safe.
5066 // - By one instruction outside of the loop (safe).
5067 // - By further instructions outside of the loop (not safe).
5068 // - By an instruction that is not part of the reduction (not safe).
5069 // This is either:
5070 // * An instruction type other than PHI or the reduction operation.
5071 // * A PHI in the header other than the initial PHI.
5072 while (!Worklist.empty()) {
5073 Instruction *Cur = Worklist.back();
5074 Worklist.pop_back();
5075
5076 // No Users.
5077 // If the instruction has no users then this is a broken chain and can't be
5078 // a reduction variable.
5079 if (Cur->use_empty())
5080 return false;
5081
5082 bool IsAPhi = isa<PHINode>(Cur);
5083
5084 // A header PHI use other than the original PHI.
5085 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
5086 return false;
5087
5088 // Reductions of instructions such as Div, and Sub is only possible if the
5089 // LHS is the reduction variable.
5090 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
5091 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
5092 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
5093 return false;
5094
5095 // Any reduction instruction must be of one of the allowed kinds.
5096 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
5097 if (!ReduxDesc.IsReduction)
5098 return false;
5099
5100 // A reduction operation must only have one use of the reduction value.
5101 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
5102 hasMultipleUsesOf(Cur, VisitedInsts))
5103 return false;
5104
5105 // All inputs to a PHI node must be a reduction value.
5106 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
5107 return false;
5108
5109 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
5110 isa<SelectInst>(Cur)))
5111 ++NumCmpSelectPatternInst;
5112 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
5113 isa<SelectInst>(Cur)))
5114 ++NumCmpSelectPatternInst;
5115
5116 // Check whether we found a reduction operator.
5117 FoundReduxOp |= !IsAPhi;
5118
5119 // Process users of current instruction. Push non-PHI nodes after PHI nodes
5120 // onto the stack. This way we are going to have seen all inputs to PHI
5121 // nodes once we get to them.
5122 SmallVector<Instruction *, 8> NonPHIs;
5123 SmallVector<Instruction *, 8> PHIs;
5124 for (User *U : Cur->users()) {
5125 Instruction *UI = cast<Instruction>(U);
5126
5127 // Check if we found the exit user.
5128 BasicBlock *Parent = UI->getParent();
5129 if (!TheLoop->contains(Parent)) {
5130 // Exit if you find multiple outside users or if the header phi node is
5131 // being used. In this case the user uses the value of the previous
5132 // iteration, in which case we would loose "VF-1" iterations of the
5133 // reduction operation if we vectorize.
5134 if (ExitInstruction != nullptr || Cur == Phi)
5135 return false;
5136
5137 // The instruction used by an outside user must be the last instruction
5138 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5139 // operations on the value.
5140 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5141 return false;
5142
5143 ExitInstruction = Cur;
5144 continue;
5145 }
5146
5147 // Process instructions only once (termination). Each reduction cycle
5148 // value must only be used once, except by phi nodes and min/max
5149 // reductions which are represented as a cmp followed by a select.
5150 ReductionInstDesc IgnoredVal(false, nullptr);
5151 if (VisitedInsts.insert(UI).second) {
5152 if (isa<PHINode>(UI))
5153 PHIs.push_back(UI);
5154 else
5155 NonPHIs.push_back(UI);
5156 } else if (!isa<PHINode>(UI) &&
5157 ((!isa<FCmpInst>(UI) &&
5158 !isa<ICmpInst>(UI) &&
5159 !isa<SelectInst>(UI)) ||
5160 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5161 return false;
5162
5163 // Remember that we completed the cycle.
5164 if (UI == Phi)
5165 FoundStartPHI = true;
5166 }
5167 Worklist.append(PHIs.begin(), PHIs.end());
5168 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5169 }
5170
5171 // This means we have seen one but not the other instruction of the
5172 // pattern or more than just a select and cmp.
5173 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5174 NumCmpSelectPatternInst != 2)
5175 return false;
5176
5177 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5178 return false;
5179
5180 // We found a reduction var if we have reached the original phi node and we
5181 // only have a single instruction with out-of-loop users.
5182
5183 // This instruction is allowed to have out-of-loop users.
5184 AllowedExit.insert(ExitInstruction);
5185
5186 // Save the description of this reduction variable.
5187 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5188 ReduxDesc.MinMaxKind);
5189 Reductions[Phi] = RD;
5190 // We've ended the cycle. This is a reduction variable if we have an
5191 // outside user and it has a binary op.
5192
5193 return true;
5194 }
5195
5196 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5197 /// pattern corresponding to a min(X, Y) or max(X, Y).
5198 LoopVectorizationLegality::ReductionInstDesc
isMinMaxSelectCmpPattern(Instruction * I,ReductionInstDesc & Prev)5199 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5200 ReductionInstDesc &Prev) {
5201
5202 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5203 "Expect a select instruction");
5204 Instruction *Cmp = nullptr;
5205 SelectInst *Select = nullptr;
5206
5207 // We must handle the select(cmp()) as a single instruction. Advance to the
5208 // select.
5209 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5210 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5211 return ReductionInstDesc(false, I);
5212 return ReductionInstDesc(Select, Prev.MinMaxKind);
5213 }
5214
5215 // Only handle single use cases for now.
5216 if (!(Select = dyn_cast<SelectInst>(I)))
5217 return ReductionInstDesc(false, I);
5218 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5219 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5220 return ReductionInstDesc(false, I);
5221 if (!Cmp->hasOneUse())
5222 return ReductionInstDesc(false, I);
5223
5224 Value *CmpLeft;
5225 Value *CmpRight;
5226
5227 // Look for a min/max pattern.
5228 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5229 return ReductionInstDesc(Select, MRK_UIntMin);
5230 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5231 return ReductionInstDesc(Select, MRK_UIntMax);
5232 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5233 return ReductionInstDesc(Select, MRK_SIntMax);
5234 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5235 return ReductionInstDesc(Select, MRK_SIntMin);
5236 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5237 return ReductionInstDesc(Select, MRK_FloatMin);
5238 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5239 return ReductionInstDesc(Select, MRK_FloatMax);
5240 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5241 return ReductionInstDesc(Select, MRK_FloatMin);
5242 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5243 return ReductionInstDesc(Select, MRK_FloatMax);
5244
5245 return ReductionInstDesc(false, I);
5246 }
5247
5248 LoopVectorizationLegality::ReductionInstDesc
isReductionInstr(Instruction * I,ReductionKind Kind,ReductionInstDesc & Prev)5249 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5250 ReductionKind Kind,
5251 ReductionInstDesc &Prev) {
5252 bool FP = I->getType()->isFloatingPointTy();
5253 bool FastMath = FP && I->hasUnsafeAlgebra();
5254 switch (I->getOpcode()) {
5255 default:
5256 return ReductionInstDesc(false, I);
5257 case Instruction::PHI:
5258 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5259 Kind != RK_FloatMinMax))
5260 return ReductionInstDesc(false, I);
5261 return ReductionInstDesc(I, Prev.MinMaxKind);
5262 case Instruction::Sub:
5263 case Instruction::Add:
5264 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5265 case Instruction::Mul:
5266 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5267 case Instruction::And:
5268 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5269 case Instruction::Or:
5270 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5271 case Instruction::Xor:
5272 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5273 case Instruction::FMul:
5274 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5275 case Instruction::FSub:
5276 case Instruction::FAdd:
5277 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5278 case Instruction::FCmp:
5279 case Instruction::ICmp:
5280 case Instruction::Select:
5281 if (Kind != RK_IntegerMinMax &&
5282 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5283 return ReductionInstDesc(false, I);
5284 return isMinMaxSelectCmpPattern(I, Prev);
5285 }
5286 }
5287
5288 LoopVectorizationLegality::InductionKind
isInductionVariable(PHINode * Phi)5289 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5290 Type *PhiTy = Phi->getType();
5291 // We only handle integer and pointer inductions variables.
5292 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5293 return IK_NoInduction;
5294
5295 // Check that the PHI is consecutive.
5296 const SCEV *PhiScev = SE->getSCEV(Phi);
5297 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5298 if (!AR) {
5299 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5300 return IK_NoInduction;
5301 }
5302 const SCEV *Step = AR->getStepRecurrence(*SE);
5303
5304 // Integer inductions need to have a stride of one.
5305 if (PhiTy->isIntegerTy()) {
5306 if (Step->isOne())
5307 return IK_IntInduction;
5308 if (Step->isAllOnesValue())
5309 return IK_ReverseIntInduction;
5310 return IK_NoInduction;
5311 }
5312
5313 // Calculate the pointer stride and check if it is consecutive.
5314 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5315 if (!C)
5316 return IK_NoInduction;
5317
5318 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5319 Type *PointerElementType = PhiTy->getPointerElementType();
5320 // The pointer stride cannot be determined if the pointer element type is not
5321 // sized.
5322 if (!PointerElementType->isSized())
5323 return IK_NoInduction;
5324
5325 uint64_t Size = DL->getTypeAllocSize(PointerElementType);
5326 if (C->getValue()->equalsInt(Size))
5327 return IK_PtrInduction;
5328 else if (C->getValue()->equalsInt(0 - Size))
5329 return IK_ReversePtrInduction;
5330
5331 return IK_NoInduction;
5332 }
5333
isInductionVariable(const Value * V)5334 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5335 Value *In0 = const_cast<Value*>(V);
5336 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5337 if (!PN)
5338 return false;
5339
5340 return Inductions.count(PN);
5341 }
5342
blockNeedsPredication(BasicBlock * BB)5343 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5344 assert(TheLoop->contains(BB) && "Unknown block used");
5345
5346 // Blocks that do not dominate the latch need predication.
5347 BasicBlock* Latch = TheLoop->getLoopLatch();
5348 return !DT->dominates(BB, Latch);
5349 }
5350
blockCanBePredicated(BasicBlock * BB,SmallPtrSetImpl<Value * > & SafePtrs)5351 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5352 SmallPtrSetImpl<Value *> &SafePtrs) {
5353
5354 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5355 // Check that we don't have a constant expression that can trap as operand.
5356 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5357 OI != OE; ++OI) {
5358 if (Constant *C = dyn_cast<Constant>(*OI))
5359 if (C->canTrap())
5360 return false;
5361 }
5362 // We might be able to hoist the load.
5363 if (it->mayReadFromMemory()) {
5364 LoadInst *LI = dyn_cast<LoadInst>(it);
5365 if (!LI)
5366 return false;
5367 if (!SafePtrs.count(LI->getPointerOperand())) {
5368 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
5369 MaskedOp.insert(LI);
5370 continue;
5371 }
5372 return false;
5373 }
5374 }
5375
5376 // We don't predicate stores at the moment.
5377 if (it->mayWriteToMemory()) {
5378 StoreInst *SI = dyn_cast<StoreInst>(it);
5379 // We only support predication of stores in basic blocks with one
5380 // predecessor.
5381 if (!SI)
5382 return false;
5383
5384 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
5385 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
5386
5387 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
5388 !isSinglePredecessor) {
5389 // Build a masked store if it is legal for the target, otherwise scalarize
5390 // the block.
5391 bool isLegalMaskedOp =
5392 isLegalMaskedStore(SI->getValueOperand()->getType(),
5393 SI->getPointerOperand());
5394 if (isLegalMaskedOp) {
5395 --NumPredStores;
5396 MaskedOp.insert(SI);
5397 continue;
5398 }
5399 return false;
5400 }
5401 }
5402 if (it->mayThrow())
5403 return false;
5404
5405 // The instructions below can trap.
5406 switch (it->getOpcode()) {
5407 default: continue;
5408 case Instruction::UDiv:
5409 case Instruction::SDiv:
5410 case Instruction::URem:
5411 case Instruction::SRem:
5412 return false;
5413 }
5414 }
5415
5416 return true;
5417 }
5418
5419 LoopVectorizationCostModel::VectorizationFactor
selectVectorizationFactor(bool OptForSize)5420 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
5421 // Width 1 means no vectorize
5422 VectorizationFactor Factor = { 1U, 0U };
5423 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5424 emitAnalysis(Report() << "runtime pointer checks needed. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
5425 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5426 return Factor;
5427 }
5428
5429 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5430 emitAnalysis(Report() << "store that is conditionally executed prevents vectorization");
5431 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5432 return Factor;
5433 }
5434
5435 // Find the trip count.
5436 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5437 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5438
5439 unsigned WidestType = getWidestType();
5440 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5441 unsigned MaxSafeDepDist = -1U;
5442 if (Legal->getMaxSafeDepDistBytes() != -1U)
5443 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5444 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5445 WidestRegister : MaxSafeDepDist);
5446 unsigned MaxVectorSize = WidestRegister / WidestType;
5447 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5448 DEBUG(dbgs() << "LV: The Widest register is: "
5449 << WidestRegister << " bits.\n");
5450
5451 if (MaxVectorSize == 0) {
5452 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5453 MaxVectorSize = 1;
5454 }
5455
5456 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
5457 " into one vector!");
5458
5459 unsigned VF = MaxVectorSize;
5460
5461 // If we optimize the program for size, avoid creating the tail loop.
5462 if (OptForSize) {
5463 // If we are unable to calculate the trip count then don't try to vectorize.
5464 if (TC < 2) {
5465 emitAnalysis(Report() << "unable to calculate the loop count due to complex control flow");
5466 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5467 return Factor;
5468 }
5469
5470 // Find the maximum SIMD width that can fit within the trip count.
5471 VF = TC % MaxVectorSize;
5472
5473 if (VF == 0)
5474 VF = MaxVectorSize;
5475
5476 // If the trip count that we found modulo the vectorization factor is not
5477 // zero then we require a tail.
5478 if (VF < 2) {
5479 emitAnalysis(Report() << "cannot optimize for size and vectorize at the "
5480 "same time. Enable vectorization of this loop "
5481 "with '#pragma clang loop vectorize(enable)' "
5482 "when compiling with -Os");
5483 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5484 return Factor;
5485 }
5486 }
5487
5488 int UserVF = Hints->getWidth();
5489 if (UserVF != 0) {
5490 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5491 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5492
5493 Factor.Width = UserVF;
5494 return Factor;
5495 }
5496
5497 float Cost = expectedCost(1);
5498 #ifndef NDEBUG
5499 const float ScalarCost = Cost;
5500 #endif /* NDEBUG */
5501 unsigned Width = 1;
5502 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5503
5504 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
5505 // Ignore scalar width, because the user explicitly wants vectorization.
5506 if (ForceVectorization && VF > 1) {
5507 Width = 2;
5508 Cost = expectedCost(Width) / (float)Width;
5509 }
5510
5511 for (unsigned i=2; i <= VF; i*=2) {
5512 // Notice that the vector loop needs to be executed less times, so
5513 // we need to divide the cost of the vector loops by the width of
5514 // the vector elements.
5515 float VectorCost = expectedCost(i) / (float)i;
5516 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5517 (int)VectorCost << ".\n");
5518 if (VectorCost < Cost) {
5519 Cost = VectorCost;
5520 Width = i;
5521 }
5522 }
5523
5524 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5525 << "LV: Vectorization seems to be not beneficial, "
5526 << "but was forced by a user.\n");
5527 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5528 Factor.Width = Width;
5529 Factor.Cost = Width * Cost;
5530 return Factor;
5531 }
5532
getWidestType()5533 unsigned LoopVectorizationCostModel::getWidestType() {
5534 unsigned MaxWidth = 8;
5535
5536 // For each block.
5537 for (Loop::block_iterator bb = TheLoop->block_begin(),
5538 be = TheLoop->block_end(); bb != be; ++bb) {
5539 BasicBlock *BB = *bb;
5540
5541 // For each instruction in the loop.
5542 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5543 Type *T = it->getType();
5544
5545 // Ignore ephemeral values.
5546 if (EphValues.count(it))
5547 continue;
5548
5549 // Only examine Loads, Stores and PHINodes.
5550 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5551 continue;
5552
5553 // Examine PHI nodes that are reduction variables.
5554 if (PHINode *PN = dyn_cast<PHINode>(it))
5555 if (!Legal->getReductionVars()->count(PN))
5556 continue;
5557
5558 // Examine the stored values.
5559 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5560 T = ST->getValueOperand()->getType();
5561
5562 // Ignore loaded pointer types and stored pointer types that are not
5563 // consecutive. However, we do want to take consecutive stores/loads of
5564 // pointer vectors into account.
5565 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5566 continue;
5567
5568 MaxWidth = std::max(MaxWidth,
5569 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5570 }
5571 }
5572
5573 return MaxWidth;
5574 }
5575
5576 unsigned
selectUnrollFactor(bool OptForSize,unsigned VF,unsigned LoopCost)5577 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5578 unsigned VF,
5579 unsigned LoopCost) {
5580
5581 // -- The unroll heuristics --
5582 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5583 // There are many micro-architectural considerations that we can't predict
5584 // at this level. For example, frontend pressure (on decode or fetch) due to
5585 // code size, or the number and capabilities of the execution ports.
5586 //
5587 // We use the following heuristics to select the unroll factor:
5588 // 1. If the code has reductions, then we unroll in order to break the cross
5589 // iteration dependency.
5590 // 2. If the loop is really small, then we unroll in order to reduce the loop
5591 // overhead.
5592 // 3. We don't unroll if we think that we will spill registers to memory due
5593 // to the increased register pressure.
5594
5595 // Use the user preference, unless 'auto' is selected.
5596 int UserUF = Hints->getInterleave();
5597 if (UserUF != 0)
5598 return UserUF;
5599
5600 // When we optimize for size, we don't unroll.
5601 if (OptForSize)
5602 return 1;
5603
5604 // We used the distance for the unroll factor.
5605 if (Legal->getMaxSafeDepDistBytes() != -1U)
5606 return 1;
5607
5608 // Do not unroll loops with a relatively small trip count.
5609 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5610 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5611 return 1;
5612
5613 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5614 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5615 " registers\n");
5616
5617 if (VF == 1) {
5618 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5619 TargetNumRegisters = ForceTargetNumScalarRegs;
5620 } else {
5621 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5622 TargetNumRegisters = ForceTargetNumVectorRegs;
5623 }
5624
5625 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5626 // We divide by these constants so assume that we have at least one
5627 // instruction that uses at least one register.
5628 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5629 R.NumInstructions = std::max(R.NumInstructions, 1U);
5630
5631 // We calculate the unroll factor using the following formula.
5632 // Subtract the number of loop invariants from the number of available
5633 // registers. These registers are used by all of the unrolled instances.
5634 // Next, divide the remaining registers by the number of registers that is
5635 // required by the loop, in order to estimate how many parallel instances
5636 // fit without causing spills. All of this is rounded down if necessary to be
5637 // a power of two. We want power of two unroll factors to simplify any
5638 // addressing operations or alignment considerations.
5639 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5640 R.MaxLocalUsers);
5641
5642 // Don't count the induction variable as unrolled.
5643 if (EnableIndVarRegisterHeur)
5644 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5645 std::max(1U, (R.MaxLocalUsers - 1)));
5646
5647 // Clamp the unroll factor ranges to reasonable factors.
5648 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
5649
5650 // Check if the user has overridden the unroll max.
5651 if (VF == 1) {
5652 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
5653 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
5654 } else {
5655 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
5656 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
5657 }
5658
5659 // If we did not calculate the cost for VF (because the user selected the VF)
5660 // then we calculate the cost of VF here.
5661 if (LoopCost == 0)
5662 LoopCost = expectedCost(VF);
5663
5664 // Clamp the calculated UF to be between the 1 and the max unroll factor
5665 // that the target allows.
5666 if (UF > MaxInterleaveSize)
5667 UF = MaxInterleaveSize;
5668 else if (UF < 1)
5669 UF = 1;
5670
5671 // Unroll if we vectorized this loop and there is a reduction that could
5672 // benefit from unrolling.
5673 if (VF > 1 && Legal->getReductionVars()->size()) {
5674 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5675 return UF;
5676 }
5677
5678 // Note that if we've already vectorized the loop we will have done the
5679 // runtime check and so unrolling won't require further checks.
5680 bool UnrollingRequiresRuntimePointerCheck =
5681 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5682
5683 // We want to unroll small loops in order to reduce the loop overhead and
5684 // potentially expose ILP opportunities.
5685 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5686 if (!UnrollingRequiresRuntimePointerCheck &&
5687 LoopCost < SmallLoopCost) {
5688 // We assume that the cost overhead is 1 and we use the cost model
5689 // to estimate the cost of the loop and unroll until the cost of the
5690 // loop overhead is about 5% of the cost of the loop.
5691 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5692
5693 // Unroll until store/load ports (estimated by max unroll factor) are
5694 // saturated.
5695 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5696 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5697
5698 // If we have a scalar reduction (vector reductions are already dealt with
5699 // by this point), we can increase the critical path length if the loop
5700 // we're unrolling is inside another loop. Limit, by default to 2, so the
5701 // critical path only gets increased by one reduction operation.
5702 if (Legal->getReductionVars()->size() &&
5703 TheLoop->getLoopDepth() > 1) {
5704 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
5705 SmallUF = std::min(SmallUF, F);
5706 StoresUF = std::min(StoresUF, F);
5707 LoadsUF = std::min(LoadsUF, F);
5708 }
5709
5710 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5711 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5712 return std::max(StoresUF, LoadsUF);
5713 }
5714
5715 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5716 return SmallUF;
5717 }
5718
5719 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5720 return 1;
5721 }
5722
5723 LoopVectorizationCostModel::RegisterUsage
calculateRegisterUsage()5724 LoopVectorizationCostModel::calculateRegisterUsage() {
5725 // This function calculates the register usage by measuring the highest number
5726 // of values that are alive at a single location. Obviously, this is a very
5727 // rough estimation. We scan the loop in a topological order in order and
5728 // assign a number to each instruction. We use RPO to ensure that defs are
5729 // met before their users. We assume that each instruction that has in-loop
5730 // users starts an interval. We record every time that an in-loop value is
5731 // used, so we have a list of the first and last occurrences of each
5732 // instruction. Next, we transpose this data structure into a multi map that
5733 // holds the list of intervals that *end* at a specific location. This multi
5734 // map allows us to perform a linear search. We scan the instructions linearly
5735 // and record each time that a new interval starts, by placing it in a set.
5736 // If we find this value in the multi-map then we remove it from the set.
5737 // The max register usage is the maximum size of the set.
5738 // We also search for instructions that are defined outside the loop, but are
5739 // used inside the loop. We need this number separately from the max-interval
5740 // usage number because when we unroll, loop-invariant values do not take
5741 // more register.
5742 LoopBlocksDFS DFS(TheLoop);
5743 DFS.perform(LI);
5744
5745 RegisterUsage R;
5746 R.NumInstructions = 0;
5747
5748 // Each 'key' in the map opens a new interval. The values
5749 // of the map are the index of the 'last seen' usage of the
5750 // instruction that is the key.
5751 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5752 // Maps instruction to its index.
5753 DenseMap<unsigned, Instruction*> IdxToInstr;
5754 // Marks the end of each interval.
5755 IntervalMap EndPoint;
5756 // Saves the list of instruction indices that are used in the loop.
5757 SmallSet<Instruction*, 8> Ends;
5758 // Saves the list of values that are used in the loop but are
5759 // defined outside the loop, such as arguments and constants.
5760 SmallPtrSet<Value*, 8> LoopInvariants;
5761
5762 unsigned Index = 0;
5763 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5764 be = DFS.endRPO(); bb != be; ++bb) {
5765 R.NumInstructions += (*bb)->size();
5766 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5767 ++it) {
5768 Instruction *I = it;
5769 IdxToInstr[Index++] = I;
5770
5771 // Save the end location of each USE.
5772 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5773 Value *U = I->getOperand(i);
5774 Instruction *Instr = dyn_cast<Instruction>(U);
5775
5776 // Ignore non-instruction values such as arguments, constants, etc.
5777 if (!Instr) continue;
5778
5779 // If this instruction is outside the loop then record it and continue.
5780 if (!TheLoop->contains(Instr)) {
5781 LoopInvariants.insert(Instr);
5782 continue;
5783 }
5784
5785 // Overwrite previous end points.
5786 EndPoint[Instr] = Index;
5787 Ends.insert(Instr);
5788 }
5789 }
5790 }
5791
5792 // Saves the list of intervals that end with the index in 'key'.
5793 typedef SmallVector<Instruction*, 2> InstrList;
5794 DenseMap<unsigned, InstrList> TransposeEnds;
5795
5796 // Transpose the EndPoints to a list of values that end at each index.
5797 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5798 it != e; ++it)
5799 TransposeEnds[it->second].push_back(it->first);
5800
5801 SmallSet<Instruction*, 8> OpenIntervals;
5802 unsigned MaxUsage = 0;
5803
5804
5805 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5806 for (unsigned int i = 0; i < Index; ++i) {
5807 Instruction *I = IdxToInstr[i];
5808 // Ignore instructions that are never used within the loop.
5809 if (!Ends.count(I)) continue;
5810
5811 // Ignore ephemeral values.
5812 if (EphValues.count(I))
5813 continue;
5814
5815 // Remove all of the instructions that end at this location.
5816 InstrList &List = TransposeEnds[i];
5817 for (unsigned int j=0, e = List.size(); j < e; ++j)
5818 OpenIntervals.erase(List[j]);
5819
5820 // Count the number of live interals.
5821 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5822
5823 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5824 OpenIntervals.size() << '\n');
5825
5826 // Add the current instruction to the list of open intervals.
5827 OpenIntervals.insert(I);
5828 }
5829
5830 unsigned Invariant = LoopInvariants.size();
5831 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5832 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5833 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5834
5835 R.LoopInvariantRegs = Invariant;
5836 R.MaxLocalUsers = MaxUsage;
5837 return R;
5838 }
5839
expectedCost(unsigned VF)5840 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5841 unsigned Cost = 0;
5842
5843 // For each block.
5844 for (Loop::block_iterator bb = TheLoop->block_begin(),
5845 be = TheLoop->block_end(); bb != be; ++bb) {
5846 unsigned BlockCost = 0;
5847 BasicBlock *BB = *bb;
5848
5849 // For each instruction in the old loop.
5850 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5851 // Skip dbg intrinsics.
5852 if (isa<DbgInfoIntrinsic>(it))
5853 continue;
5854
5855 // Ignore ephemeral values.
5856 if (EphValues.count(it))
5857 continue;
5858
5859 unsigned C = getInstructionCost(it, VF);
5860
5861 // Check if we should override the cost.
5862 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5863 C = ForceTargetInstructionCost;
5864
5865 BlockCost += C;
5866 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5867 VF << " For instruction: " << *it << '\n');
5868 }
5869
5870 // We assume that if-converted blocks have a 50% chance of being executed.
5871 // When the code is scalar then some of the blocks are avoided due to CF.
5872 // When the code is vectorized we execute all code paths.
5873 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5874 BlockCost /= 2;
5875
5876 Cost += BlockCost;
5877 }
5878
5879 return Cost;
5880 }
5881
5882 /// \brief Check whether the address computation for a non-consecutive memory
5883 /// access looks like an unlikely candidate for being merged into the indexing
5884 /// mode.
5885 ///
5886 /// We look for a GEP which has one index that is an induction variable and all
5887 /// other indices are loop invariant. If the stride of this access is also
5888 /// within a small bound we decide that this address computation can likely be
5889 /// merged into the addressing mode.
5890 /// In all other cases, we identify the address computation as complex.
isLikelyComplexAddressComputation(Value * Ptr,LoopVectorizationLegality * Legal,ScalarEvolution * SE,const Loop * TheLoop)5891 static bool isLikelyComplexAddressComputation(Value *Ptr,
5892 LoopVectorizationLegality *Legal,
5893 ScalarEvolution *SE,
5894 const Loop *TheLoop) {
5895 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5896 if (!Gep)
5897 return true;
5898
5899 // We are looking for a gep with all loop invariant indices except for one
5900 // which should be an induction variable.
5901 unsigned NumOperands = Gep->getNumOperands();
5902 for (unsigned i = 1; i < NumOperands; ++i) {
5903 Value *Opd = Gep->getOperand(i);
5904 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5905 !Legal->isInductionVariable(Opd))
5906 return true;
5907 }
5908
5909 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5910 // can likely be merged into the address computation.
5911 unsigned MaxMergeDistance = 64;
5912
5913 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5914 if (!AddRec)
5915 return true;
5916
5917 // Check the step is constant.
5918 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5919 // Calculate the pointer stride and check if it is consecutive.
5920 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5921 if (!C)
5922 return true;
5923
5924 const APInt &APStepVal = C->getValue()->getValue();
5925
5926 // Huge step value - give up.
5927 if (APStepVal.getBitWidth() > 64)
5928 return true;
5929
5930 int64_t StepVal = APStepVal.getSExtValue();
5931
5932 return StepVal > MaxMergeDistance;
5933 }
5934
isStrideMul(Instruction * I,LoopVectorizationLegality * Legal)5935 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5936 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5937 return true;
5938 return false;
5939 }
5940
5941 unsigned
getInstructionCost(Instruction * I,unsigned VF)5942 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5943 // If we know that this instruction will remain uniform, check the cost of
5944 // the scalar version.
5945 if (Legal->isUniformAfterVectorization(I))
5946 VF = 1;
5947
5948 Type *RetTy = I->getType();
5949 Type *VectorTy = ToVectorTy(RetTy, VF);
5950
5951 // TODO: We need to estimate the cost of intrinsic calls.
5952 switch (I->getOpcode()) {
5953 case Instruction::GetElementPtr:
5954 // We mark this instruction as zero-cost because the cost of GEPs in
5955 // vectorized code depends on whether the corresponding memory instruction
5956 // is scalarized or not. Therefore, we handle GEPs with the memory
5957 // instruction cost.
5958 return 0;
5959 case Instruction::Br: {
5960 return TTI.getCFInstrCost(I->getOpcode());
5961 }
5962 case Instruction::PHI:
5963 //TODO: IF-converted IFs become selects.
5964 return 0;
5965 case Instruction::Add:
5966 case Instruction::FAdd:
5967 case Instruction::Sub:
5968 case Instruction::FSub:
5969 case Instruction::Mul:
5970 case Instruction::FMul:
5971 case Instruction::UDiv:
5972 case Instruction::SDiv:
5973 case Instruction::FDiv:
5974 case Instruction::URem:
5975 case Instruction::SRem:
5976 case Instruction::FRem:
5977 case Instruction::Shl:
5978 case Instruction::LShr:
5979 case Instruction::AShr:
5980 case Instruction::And:
5981 case Instruction::Or:
5982 case Instruction::Xor: {
5983 // Since we will replace the stride by 1 the multiplication should go away.
5984 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5985 return 0;
5986 // Certain instructions can be cheaper to vectorize if they have a constant
5987 // second vector operand. One example of this are shifts on x86.
5988 TargetTransformInfo::OperandValueKind Op1VK =
5989 TargetTransformInfo::OK_AnyValue;
5990 TargetTransformInfo::OperandValueKind Op2VK =
5991 TargetTransformInfo::OK_AnyValue;
5992 TargetTransformInfo::OperandValueProperties Op1VP =
5993 TargetTransformInfo::OP_None;
5994 TargetTransformInfo::OperandValueProperties Op2VP =
5995 TargetTransformInfo::OP_None;
5996 Value *Op2 = I->getOperand(1);
5997
5998 // Check for a splat of a constant or for a non uniform vector of constants.
5999 if (isa<ConstantInt>(Op2)) {
6000 ConstantInt *CInt = cast<ConstantInt>(Op2);
6001 if (CInt && CInt->getValue().isPowerOf2())
6002 Op2VP = TargetTransformInfo::OP_PowerOf2;
6003 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6004 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
6005 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
6006 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
6007 if (SplatValue) {
6008 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
6009 if (CInt && CInt->getValue().isPowerOf2())
6010 Op2VP = TargetTransformInfo::OP_PowerOf2;
6011 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
6012 }
6013 }
6014
6015 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
6016 Op1VP, Op2VP);
6017 }
6018 case Instruction::Select: {
6019 SelectInst *SI = cast<SelectInst>(I);
6020 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
6021 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
6022 Type *CondTy = SI->getCondition()->getType();
6023 if (!ScalarCond)
6024 CondTy = VectorType::get(CondTy, VF);
6025
6026 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
6027 }
6028 case Instruction::ICmp:
6029 case Instruction::FCmp: {
6030 Type *ValTy = I->getOperand(0)->getType();
6031 VectorTy = ToVectorTy(ValTy, VF);
6032 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
6033 }
6034 case Instruction::Store:
6035 case Instruction::Load: {
6036 StoreInst *SI = dyn_cast<StoreInst>(I);
6037 LoadInst *LI = dyn_cast<LoadInst>(I);
6038 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
6039 LI->getType());
6040 VectorTy = ToVectorTy(ValTy, VF);
6041
6042 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
6043 unsigned AS = SI ? SI->getPointerAddressSpace() :
6044 LI->getPointerAddressSpace();
6045 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
6046 // We add the cost of address computation here instead of with the gep
6047 // instruction because only here we know whether the operation is
6048 // scalarized.
6049 if (VF == 1)
6050 return TTI.getAddressComputationCost(VectorTy) +
6051 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6052
6053 // Scalarized loads/stores.
6054 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
6055 bool Reverse = ConsecutiveStride < 0;
6056 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
6057 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
6058 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
6059 bool IsComplexComputation =
6060 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
6061 unsigned Cost = 0;
6062 // The cost of extracting from the value vector and pointer vector.
6063 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
6064 for (unsigned i = 0; i < VF; ++i) {
6065 // The cost of extracting the pointer operand.
6066 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
6067 // In case of STORE, the cost of ExtractElement from the vector.
6068 // In case of LOAD, the cost of InsertElement into the returned
6069 // vector.
6070 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
6071 Instruction::InsertElement,
6072 VectorTy, i);
6073 }
6074
6075 // The cost of the scalar loads/stores.
6076 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
6077 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
6078 Alignment, AS);
6079 return Cost;
6080 }
6081
6082 // Wide load/stores.
6083 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
6084 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
6085
6086 if (Reverse)
6087 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
6088 VectorTy, 0);
6089 return Cost;
6090 }
6091 case Instruction::ZExt:
6092 case Instruction::SExt:
6093 case Instruction::FPToUI:
6094 case Instruction::FPToSI:
6095 case Instruction::FPExt:
6096 case Instruction::PtrToInt:
6097 case Instruction::IntToPtr:
6098 case Instruction::SIToFP:
6099 case Instruction::UIToFP:
6100 case Instruction::Trunc:
6101 case Instruction::FPTrunc:
6102 case Instruction::BitCast: {
6103 // We optimize the truncation of induction variable.
6104 // The cost of these is the same as the scalar operation.
6105 if (I->getOpcode() == Instruction::Trunc &&
6106 Legal->isInductionVariable(I->getOperand(0)))
6107 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
6108 I->getOperand(0)->getType());
6109
6110 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
6111 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
6112 }
6113 case Instruction::Call: {
6114 CallInst *CI = cast<CallInst>(I);
6115 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
6116 assert(ID && "Not an intrinsic call!");
6117 Type *RetTy = ToVectorTy(CI->getType(), VF);
6118 SmallVector<Type*, 4> Tys;
6119 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
6120 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
6121 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
6122 }
6123 default: {
6124 // We are scalarizing the instruction. Return the cost of the scalar
6125 // instruction, plus the cost of insert and extract into vector
6126 // elements, times the vector width.
6127 unsigned Cost = 0;
6128
6129 if (!RetTy->isVoidTy() && VF != 1) {
6130 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
6131 VectorTy);
6132 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
6133 VectorTy);
6134
6135 // The cost of inserting the results plus extracting each one of the
6136 // operands.
6137 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
6138 }
6139
6140 // The cost of executing VF copies of the scalar instruction. This opcode
6141 // is unknown. Assume that it is the same as 'mul'.
6142 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
6143 return Cost;
6144 }
6145 }// end of switch.
6146 }
6147
ToVectorTy(Type * Scalar,unsigned VF)6148 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
6149 if (Scalar->isVoidTy() || VF == 1)
6150 return Scalar;
6151 return VectorType::get(Scalar, VF);
6152 }
6153
6154 char LoopVectorize::ID = 0;
6155 static const char lv_name[] = "Loop Vectorization";
6156 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
6157 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
6158 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
6159 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
6160 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
6161 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
6162 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
6163 INITIALIZE_PASS_DEPENDENCY(LCSSA)
6164 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
6165 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
6166 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
6167
6168 namespace llvm {
createLoopVectorizePass(bool NoUnrolling,bool AlwaysVectorize)6169 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
6170 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
6171 }
6172 }
6173
isConsecutiveLoadOrStore(Instruction * Inst)6174 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
6175 // Check for a store.
6176 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
6177 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
6178
6179 // Check for a load.
6180 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
6181 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
6182
6183 return false;
6184 }
6185
6186
scalarizeInstruction(Instruction * Instr,bool IfPredicateStore)6187 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
6188 bool IfPredicateStore) {
6189 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
6190 // Holds vector parameters or scalars, in case of uniform vals.
6191 SmallVector<VectorParts, 4> Params;
6192
6193 setDebugLocFromInst(Builder, Instr);
6194
6195 // Find all of the vectorized parameters.
6196 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6197 Value *SrcOp = Instr->getOperand(op);
6198
6199 // If we are accessing the old induction variable, use the new one.
6200 if (SrcOp == OldInduction) {
6201 Params.push_back(getVectorValue(SrcOp));
6202 continue;
6203 }
6204
6205 // Try using previously calculated values.
6206 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6207
6208 // If the src is an instruction that appeared earlier in the basic block
6209 // then it should already be vectorized.
6210 if (SrcInst && OrigLoop->contains(SrcInst)) {
6211 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6212 // The parameter is a vector value from earlier.
6213 Params.push_back(WidenMap.get(SrcInst));
6214 } else {
6215 // The parameter is a scalar from outside the loop. Maybe even a constant.
6216 VectorParts Scalars;
6217 Scalars.append(UF, SrcOp);
6218 Params.push_back(Scalars);
6219 }
6220 }
6221
6222 assert(Params.size() == Instr->getNumOperands() &&
6223 "Invalid number of operands");
6224
6225 // Does this instruction return a value ?
6226 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6227
6228 Value *UndefVec = IsVoidRetTy ? nullptr :
6229 UndefValue::get(Instr->getType());
6230 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6231 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6232
6233 Instruction *InsertPt = Builder.GetInsertPoint();
6234 BasicBlock *IfBlock = Builder.GetInsertBlock();
6235 BasicBlock *CondBlock = nullptr;
6236
6237 VectorParts Cond;
6238 Loop *VectorLp = nullptr;
6239 if (IfPredicateStore) {
6240 assert(Instr->getParent()->getSinglePredecessor() &&
6241 "Only support single predecessor blocks");
6242 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6243 Instr->getParent());
6244 VectorLp = LI->getLoopFor(IfBlock);
6245 assert(VectorLp && "Must have a loop for this block");
6246 }
6247
6248 // For each vector unroll 'part':
6249 for (unsigned Part = 0; Part < UF; ++Part) {
6250 // For each scalar that we create:
6251
6252 // Start an "if (pred) a[i] = ..." block.
6253 Value *Cmp = nullptr;
6254 if (IfPredicateStore) {
6255 if (Cond[Part]->getType()->isVectorTy())
6256 Cond[Part] =
6257 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6258 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6259 ConstantInt::get(Cond[Part]->getType(), 1));
6260 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6261 LoopVectorBody.push_back(CondBlock);
6262 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
6263 // Update Builder with newly created basic block.
6264 Builder.SetInsertPoint(InsertPt);
6265 }
6266
6267 Instruction *Cloned = Instr->clone();
6268 if (!IsVoidRetTy)
6269 Cloned->setName(Instr->getName() + ".cloned");
6270 // Replace the operands of the cloned instructions with extracted scalars.
6271 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6272 Value *Op = Params[op][Part];
6273 Cloned->setOperand(op, Op);
6274 }
6275
6276 // Place the cloned scalar in the new loop.
6277 Builder.Insert(Cloned);
6278
6279 // If the original scalar returns a value we need to place it in a vector
6280 // so that future users will be able to use it.
6281 if (!IsVoidRetTy)
6282 VecResults[Part] = Cloned;
6283
6284 // End if-block.
6285 if (IfPredicateStore) {
6286 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6287 LoopVectorBody.push_back(NewIfBlock);
6288 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6289 Builder.SetInsertPoint(InsertPt);
6290 Instruction *OldBr = IfBlock->getTerminator();
6291 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6292 OldBr->eraseFromParent();
6293 IfBlock = NewIfBlock;
6294 }
6295 }
6296 }
6297
vectorizeMemoryInstruction(Instruction * Instr)6298 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6299 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6300 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6301
6302 return scalarizeInstruction(Instr, IfPredicateStore);
6303 }
6304
reverseVector(Value * Vec)6305 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6306 return Vec;
6307 }
6308
getBroadcastInstrs(Value * V)6309 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6310 return V;
6311 }
6312
getConsecutiveVector(Value * Val,int StartIdx,bool Negate)6313 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6314 bool Negate) {
6315 // When unrolling and the VF is 1, we only need to add a simple scalar.
6316 Type *ITy = Val->getType();
6317 assert(!ITy->isVectorTy() && "Val must be a scalar");
6318 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6319 return Builder.CreateAdd(Val, C, "induction");
6320 }
6321