Bug Summary

File:lib/Transforms/Vectorize/SLPVectorizer.cpp
Warning:line 2308, column 32
Called C++ object pointer is null

Annotated Source Code

1//===- SLPVectorizer.cpp - A bottom up SLP 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// This pass implements the Bottom Up SLP vectorizer. It detects consecutive
10// stores that can be put together into vector-stores. Next, it attempts to
11// construct vectorizable tree using the use-def chains. If a profitable tree
12// was found, the SLP vectorizer performs vectorization on the tree.
13//
14// The pass is inspired by the work described in the paper:
15// "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks.
16//
17//===----------------------------------------------------------------------===//
18#include "llvm/Transforms/Vectorize/SLPVectorizer.h"
19#include "llvm/ADT/Optional.h"
20#include "llvm/ADT/PostOrderIterator.h"
21#include "llvm/ADT/SetVector.h"
22#include "llvm/ADT/Statistic.h"
23#include "llvm/Analysis/CodeMetrics.h"
24#include "llvm/Analysis/GlobalsModRef.h"
25#include "llvm/Analysis/LoopAccessAnalysis.h"
26#include "llvm/Analysis/ScalarEvolutionExpressions.h"
27#include "llvm/Analysis/ValueTracking.h"
28#include "llvm/Analysis/VectorUtils.h"
29#include "llvm/IR/DataLayout.h"
30#include "llvm/IR/Dominators.h"
31#include "llvm/IR/IRBuilder.h"
32#include "llvm/IR/Instructions.h"
33#include "llvm/IR/IntrinsicInst.h"
34#include "llvm/IR/Module.h"
35#include "llvm/IR/NoFolder.h"
36#include "llvm/IR/Type.h"
37#include "llvm/IR/Value.h"
38#include "llvm/IR/Verifier.h"
39#include "llvm/Pass.h"
40#include "llvm/Support/CommandLine.h"
41#include "llvm/Support/Debug.h"
42#include "llvm/Support/raw_ostream.h"
43#include "llvm/Transforms/Vectorize.h"
44#include <algorithm>
45#include <memory>
46
47using namespace llvm;
48using namespace slpvectorizer;
49
50#define SV_NAME"slp-vectorizer" "slp-vectorizer"
51#define DEBUG_TYPE"SLP" "SLP"
52
53STATISTIC(NumVectorInstructions, "Number of vector instructions generated")static llvm::Statistic NumVectorInstructions = {"SLP", "NumVectorInstructions"
, "Number of vector instructions generated", {0}, false}
;
54
55static cl::opt<int>
56 SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden,
57 cl::desc("Only vectorize if you gain more than this "
58 "number "));
59
60static cl::opt<bool>
61ShouldVectorizeHor("slp-vectorize-hor", cl::init(true), cl::Hidden,
62 cl::desc("Attempt to vectorize horizontal reductions"));
63
64static cl::opt<bool> ShouldStartVectorizeHorAtStore(
65 "slp-vectorize-hor-store", cl::init(false), cl::Hidden,
66 cl::desc(
67 "Attempt to vectorize horizontal reductions feeding into a store"));
68
69static cl::opt<int>
70MaxVectorRegSizeOption("slp-max-reg-size", cl::init(128), cl::Hidden,
71 cl::desc("Attempt to vectorize for this register size in bits"));
72
73/// Limits the size of scheduling regions in a block.
74/// It avoid long compile times for _very_ large blocks where vector
75/// instructions are spread over a wide range.
76/// This limit is way higher than needed by real-world functions.
77static cl::opt<int>
78ScheduleRegionSizeBudget("slp-schedule-budget", cl::init(100000), cl::Hidden,
79 cl::desc("Limit the size of the SLP scheduling region per block"));
80
81static cl::opt<int> MinVectorRegSizeOption(
82 "slp-min-reg-size", cl::init(128), cl::Hidden,
83 cl::desc("Attempt to vectorize for this register size in bits"));
84
85static cl::opt<unsigned> RecursionMaxDepth(
86 "slp-recursion-max-depth", cl::init(12), cl::Hidden,
87 cl::desc("Limit the recursion depth when building a vectorizable tree"));
88
89static cl::opt<unsigned> MinTreeSize(
90 "slp-min-tree-size", cl::init(3), cl::Hidden,
91 cl::desc("Only vectorize small trees if they are fully vectorizable"));
92
93// Limit the number of alias checks. The limit is chosen so that
94// it has no negative effect on the llvm benchmarks.
95static const unsigned AliasedCheckLimit = 10;
96
97// Another limit for the alias checks: The maximum distance between load/store
98// instructions where alias checks are done.
99// This limit is useful for very large basic blocks.
100static const unsigned MaxMemDepDistance = 160;
101
102/// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling
103/// regions to be handled.
104static const int MinScheduleRegionSize = 16;
105
106/// \brief Predicate for the element types that the SLP vectorizer supports.
107///
108/// The most important thing to filter here are types which are invalid in LLVM
109/// vectors. We also filter target specific types which have absolutely no
110/// meaningful vectorization path such as x86_fp80 and ppc_f128. This just
111/// avoids spending time checking the cost model and realizing that they will
112/// be inevitably scalarized.
113static bool isValidElementType(Type *Ty) {
114 return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() &&
115 !Ty->isPPC_FP128Ty();
116}
117
118/// \returns true if all of the instructions in \p VL are in the same block or
119/// false otherwise.
120static bool allSameBlock(ArrayRef<Value *> VL) {
121 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
122 if (!I0)
123 return false;
124 BasicBlock *BB = I0->getParent();
125 for (int i = 1, e = VL.size(); i < e; i++) {
126 Instruction *I = dyn_cast<Instruction>(VL[i]);
127 if (!I)
128 return false;
129
130 if (BB != I->getParent())
131 return false;
132 }
133 return true;
134}
135
136/// \returns True if all of the values in \p VL are constants.
137static bool allConstant(ArrayRef<Value *> VL) {
138 for (Value *i : VL)
139 if (!isa<Constant>(i))
140 return false;
141 return true;
142}
143
144/// \returns True if all of the values in \p VL are identical.
145static bool isSplat(ArrayRef<Value *> VL) {
146 for (unsigned i = 1, e = VL.size(); i < e; ++i)
147 if (VL[i] != VL[0])
148 return false;
149 return true;
150}
151
152///\returns Opcode that can be clubbed with \p Op to create an alternate
153/// sequence which can later be merged as a ShuffleVector instruction.
154static unsigned getAltOpcode(unsigned Op) {
155 switch (Op) {
156 case Instruction::FAdd:
157 return Instruction::FSub;
158 case Instruction::FSub:
159 return Instruction::FAdd;
160 case Instruction::Add:
161 return Instruction::Sub;
162 case Instruction::Sub:
163 return Instruction::Add;
164 default:
165 return 0;
166 }
167}
168
169///\returns bool representing if Opcode \p Op can be part
170/// of an alternate sequence which can later be merged as
171/// a ShuffleVector instruction.
172static bool canCombineAsAltInst(unsigned Op) {
173 return Op == Instruction::FAdd || Op == Instruction::FSub ||
174 Op == Instruction::Sub || Op == Instruction::Add;
175}
176
177/// \returns ShuffleVector instruction if instructions in \p VL have
178/// alternate fadd,fsub / fsub,fadd/add,sub/sub,add sequence.
179/// (i.e. e.g. opcodes of fadd,fsub,fadd,fsub...)
180static unsigned isAltInst(ArrayRef<Value *> VL) {
181 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
182 unsigned Opcode = I0->getOpcode();
183 unsigned AltOpcode = getAltOpcode(Opcode);
184 for (int i = 1, e = VL.size(); i < e; i++) {
185 Instruction *I = dyn_cast<Instruction>(VL[i]);
186 if (!I || I->getOpcode() != ((i & 1) ? AltOpcode : Opcode))
187 return 0;
188 }
189 return Instruction::ShuffleVector;
190}
191
192/// \returns The opcode if all of the Instructions in \p VL have the same
193/// opcode, or zero.
194static unsigned getSameOpcode(ArrayRef<Value *> VL) {
195 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
196 if (!I0)
197 return 0;
198 unsigned Opcode = I0->getOpcode();
199 for (int i = 1, e = VL.size(); i < e; i++) {
200 Instruction *I = dyn_cast<Instruction>(VL[i]);
201 if (!I || Opcode != I->getOpcode()) {
202 if (canCombineAsAltInst(Opcode) && i == 1)
203 return isAltInst(VL);
204 return 0;
205 }
206 }
207 return Opcode;
208}
209
210/// Get the intersection (logical and) of all of the potential IR flags
211/// of each scalar operation (VL) that will be converted into a vector (I).
212/// Flag set: NSW, NUW, exact, and all of fast-math.
213static void propagateIRFlags(Value *I, ArrayRef<Value *> VL) {
214 if (auto *VecOp = dyn_cast<Instruction>(I)) {
215 if (auto *Intersection = dyn_cast<Instruction>(VL[0])) {
216 // Intersection is initialized to the 0th scalar,
217 // so start counting from index '1'.
218 for (int i = 1, e = VL.size(); i < e; ++i) {
219 if (auto *Scalar = dyn_cast<Instruction>(VL[i]))
220 Intersection->andIRFlags(Scalar);
221 }
222 VecOp->copyIRFlags(Intersection);
223 }
224 }
225}
226
227/// \returns true if all of the values in \p VL have the same type or false
228/// otherwise.
229static bool allSameType(ArrayRef<Value *> VL) {
230 Type *Ty = VL[0]->getType();
231 for (int i = 1, e = VL.size(); i < e; i++)
232 if (VL[i]->getType() != Ty)
233 return false;
234
235 return true;
236}
237
238/// \returns True if Extract{Value,Element} instruction extracts element Idx.
239static bool matchExtractIndex(Instruction *E, unsigned Idx, unsigned Opcode) {
240 assert(Opcode == Instruction::ExtractElement ||((Opcode == Instruction::ExtractElement || Opcode == Instruction
::ExtractValue) ? static_cast<void> (0) : __assert_fail
("Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 241, __PRETTY_FUNCTION__))
241 Opcode == Instruction::ExtractValue)((Opcode == Instruction::ExtractElement || Opcode == Instruction
::ExtractValue) ? static_cast<void> (0) : __assert_fail
("Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 241, __PRETTY_FUNCTION__))
;
242 if (Opcode == Instruction::ExtractElement) {
243 ConstantInt *CI = dyn_cast<ConstantInt>(E->getOperand(1));
244 return CI && CI->getZExtValue() == Idx;
245 } else {
246 ExtractValueInst *EI = cast<ExtractValueInst>(E);
247 return EI->getNumIndices() == 1 && *EI->idx_begin() == Idx;
248 }
249}
250
251/// \returns True if in-tree use also needs extract. This refers to
252/// possible scalar operand in vectorized instruction.
253static bool InTreeUserNeedToExtract(Value *Scalar, Instruction *UserInst,
254 TargetLibraryInfo *TLI) {
255
256 unsigned Opcode = UserInst->getOpcode();
257 switch (Opcode) {
258 case Instruction::Load: {
259 LoadInst *LI = cast<LoadInst>(UserInst);
260 return (LI->getPointerOperand() == Scalar);
261 }
262 case Instruction::Store: {
263 StoreInst *SI = cast<StoreInst>(UserInst);
264 return (SI->getPointerOperand() == Scalar);
265 }
266 case Instruction::Call: {
267 CallInst *CI = cast<CallInst>(UserInst);
268 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
269 if (hasVectorInstrinsicScalarOpd(ID, 1)) {
270 return (CI->getArgOperand(1) == Scalar);
271 }
272 }
273 default:
274 return false;
275 }
276}
277
278/// \returns the AA location that is being access by the instruction.
279static MemoryLocation getLocation(Instruction *I, AliasAnalysis *AA) {
280 if (StoreInst *SI = dyn_cast<StoreInst>(I))
281 return MemoryLocation::get(SI);
282 if (LoadInst *LI = dyn_cast<LoadInst>(I))
283 return MemoryLocation::get(LI);
284 return MemoryLocation();
285}
286
287/// \returns True if the instruction is not a volatile or atomic load/store.
288static bool isSimple(Instruction *I) {
289 if (LoadInst *LI = dyn_cast<LoadInst>(I))
290 return LI->isSimple();
291 if (StoreInst *SI = dyn_cast<StoreInst>(I))
292 return SI->isSimple();
293 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
294 return !MI->isVolatile();
295 return true;
296}
297
298namespace llvm {
299namespace slpvectorizer {
300/// Bottom Up SLP Vectorizer.
301class BoUpSLP {
302public:
303 typedef SmallVector<Value *, 8> ValueList;
304 typedef SmallVector<Instruction *, 16> InstrList;
305 typedef SmallPtrSet<Value *, 16> ValueSet;
306 typedef SmallVector<StoreInst *, 8> StoreList;
307
308 BoUpSLP(Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti,
309 TargetLibraryInfo *TLi, AliasAnalysis *Aa, LoopInfo *Li,
310 DominatorTree *Dt, AssumptionCache *AC, DemandedBits *DB,
311 const DataLayout *DL)
312 : NumLoadsWantToKeepOrder(0), NumLoadsWantToChangeOrder(0), F(Func),
313 SE(Se), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt), AC(AC), DB(DB),
314 DL(DL), Builder(Se->getContext()) {
315 CodeMetrics::collectEphemeralValues(F, AC, EphValues);
316 // Use the vector register size specified by the target unless overridden
317 // by a command-line option.
318 // TODO: It would be better to limit the vectorization factor based on
319 // data type rather than just register size. For example, x86 AVX has
320 // 256-bit registers, but it does not support integer operations
321 // at that width (that requires AVX2).
322 if (MaxVectorRegSizeOption.getNumOccurrences())
323 MaxVecRegSize = MaxVectorRegSizeOption;
324 else
325 MaxVecRegSize = TTI->getRegisterBitWidth(true);
326
327 MinVecRegSize = MinVectorRegSizeOption;
328 }
329
330 /// \brief Vectorize the tree that starts with the elements in \p VL.
331 /// Returns the vectorized root.
332 Value *vectorizeTree();
333 /// Vectorize the tree but with the list of externally used values \p
334 /// ExternallyUsedValues. Values in this MapVector can be replaced but the
335 /// generated extractvalue instructions.
336 Value *vectorizeTree(MapVector<Value *, DebugLoc> &ExternallyUsedValues);
337
338 /// \returns the cost incurred by unwanted spills and fills, caused by
339 /// holding live values over call sites.
340 int getSpillCost();
341
342 /// \returns the vectorization cost of the subtree that starts at \p VL.
343 /// A negative number means that this is profitable.
344 int getTreeCost();
345
346 /// Construct a vectorizable tree that starts at \p Roots, ignoring users for
347 /// the purpose of scheduling and extraction in the \p UserIgnoreLst.
348 void buildTree(ArrayRef<Value *> Roots,
349 ArrayRef<Value *> UserIgnoreLst = None);
350 /// Construct a vectorizable tree that starts at \p Roots, ignoring users for
351 /// the purpose of scheduling and extraction in the \p UserIgnoreLst taking
352 /// into account (anf updating it, if required) list of externally used
353 /// values stored in \p ExternallyUsedValues.
354 void buildTree(ArrayRef<Value *> Roots,
355 MapVector<Value *, DebugLoc> &ExternallyUsedValues,
356 ArrayRef<Value *> UserIgnoreLst = None);
357
358 /// Clear the internal data structures that are created by 'buildTree'.
359 void deleteTree() {
360 VectorizableTree.clear();
361 ScalarToTreeEntry.clear();
362 MustGather.clear();
363 ExternalUses.clear();
364 NumLoadsWantToKeepOrder = 0;
365 NumLoadsWantToChangeOrder = 0;
366 for (auto &Iter : BlocksSchedules) {
367 BlockScheduling *BS = Iter.second.get();
368 BS->clear();
369 }
370 MinBWs.clear();
371 }
372
373 /// \brief Perform LICM and CSE on the newly generated gather sequences.
374 void optimizeGatherSequence();
375
376 /// \returns true if it is beneficial to reverse the vector order.
377 bool shouldReorder() const {
378 return NumLoadsWantToChangeOrder > NumLoadsWantToKeepOrder;
379 }
380
381 /// \return The vector element size in bits to use when vectorizing the
382 /// expression tree ending at \p V. If V is a store, the size is the width of
383 /// the stored value. Otherwise, the size is the width of the largest loaded
384 /// value reaching V. This method is used by the vectorizer to calculate
385 /// vectorization factors.
386 unsigned getVectorElementSize(Value *V);
387
388 /// Compute the minimum type sizes required to represent the entries in a
389 /// vectorizable tree.
390 void computeMinimumValueSizes();
391
392 // \returns maximum vector register size as set by TTI or overridden by cl::opt.
393 unsigned getMaxVecRegSize() const {
394 return MaxVecRegSize;
395 }
396
397 // \returns minimum vector register size as set by cl::opt.
398 unsigned getMinVecRegSize() const {
399 return MinVecRegSize;
400 }
401
402 /// \brief Check if ArrayType or StructType is isomorphic to some VectorType.
403 ///
404 /// \returns number of elements in vector if isomorphism exists, 0 otherwise.
405 unsigned canMapToVector(Type *T, const DataLayout &DL) const;
406
407 /// \returns True if the VectorizableTree is both tiny and not fully
408 /// vectorizable. We do not vectorize such trees.
409 bool isTreeTinyAndNotFullyVectorizable();
410
411private:
412 struct TreeEntry;
413
414 /// \returns the cost of the vectorizable entry.
415 int getEntryCost(TreeEntry *E);
416
417 /// This is the recursive part of buildTree.
418 void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth);
419
420 /// \returns True if the ExtractElement/ExtractValue instructions in VL can
421 /// be vectorized to use the original vector (or aggregate "bitcast" to a vector).
422 bool canReuseExtract(ArrayRef<Value *> VL, unsigned Opcode) const;
423
424 /// Vectorize a single entry in the tree. VL icontains all isomorphic scalars
425 /// in order of its usage in a user program, for example ADD1, ADD2 and so on
426 /// or LOAD1 , LOAD2 etc.
427 Value *vectorizeTree(ArrayRef<Value *> VL, TreeEntry *E);
428
429 /// Vectorize a single entry in the tree, starting in \p VL.
430 Value *vectorizeTree(ArrayRef<Value *> VL);
431
432 /// \returns the pointer to the vectorized value if \p VL is already
433 /// vectorized, or NULL. They may happen in cycles.
434 Value *alreadyVectorized(ArrayRef<Value *> VL) const;
435
436 /// \returns the scalarization cost for this type. Scalarization in this
437 /// context means the creation of vectors from a group of scalars.
438 int getGatherCost(Type *Ty);
439
440 /// \returns the scalarization cost for this list of values. Assuming that
441 /// this subtree gets vectorized, we may need to extract the values from the
442 /// roots. This method calculates the cost of extracting the values.
443 int getGatherCost(ArrayRef<Value *> VL);
444
445 /// \brief Set the Builder insert point to one after the last instruction in
446 /// the bundle
447 void setInsertPointAfterBundle(ArrayRef<Value *> VL);
448
449 /// \returns a vector from a collection of scalars in \p VL.
450 Value *Gather(ArrayRef<Value *> VL, VectorType *Ty);
451
452 /// \returns whether the VectorizableTree is fully vectorizable and will
453 /// be beneficial even the tree height is tiny.
454 bool isFullyVectorizableTinyTree();
455
456 /// \reorder commutative operands in alt shuffle if they result in
457 /// vectorized code.
458 void reorderAltShuffleOperands(ArrayRef<Value *> VL,
459 SmallVectorImpl<Value *> &Left,
460 SmallVectorImpl<Value *> &Right);
461 /// \reorder commutative operands to get better probability of
462 /// generating vectorized code.
463 void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL,
464 SmallVectorImpl<Value *> &Left,
465 SmallVectorImpl<Value *> &Right);
466 struct TreeEntry {
467 TreeEntry() : Scalars(), VectorizedValue(nullptr),
468 NeedToGather(0), NeedToShuffle(0) {}
469
470 /// \returns true if the scalars in VL are equal to this entry.
471 bool isSame(ArrayRef<Value *> VL) const {
472 assert(VL.size() == Scalars.size() && "Invalid size")((VL.size() == Scalars.size() && "Invalid size") ? static_cast
<void> (0) : __assert_fail ("VL.size() == Scalars.size() && \"Invalid size\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 472, __PRETTY_FUNCTION__))
;
473 return std::equal(VL.begin(), VL.end(), Scalars.begin());
474 }
475
476 /// \returns true if the scalars in VL are found in this tree entry.
477 bool isFoundJumbled(ArrayRef<Value *> VL, const DataLayout &DL,
478 ScalarEvolution &SE) const {
479 assert(VL.size() == Scalars.size() && "Invalid size")((VL.size() == Scalars.size() && "Invalid size") ? static_cast
<void> (0) : __assert_fail ("VL.size() == Scalars.size() && \"Invalid size\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 479, __PRETTY_FUNCTION__))
;
480 SmallVector<Value *, 8> List;
481 if (!sortMemAccesses(VL, DL, SE, List))
482 return false;
483
484 return std::equal(List.begin(), List.end(), Scalars.begin());
485 }
486
487 /// A vector of scalars.
488 ValueList Scalars;
489
490 /// The Scalars are vectorized into this value. It is initialized to Null.
491 Value *VectorizedValue;
492
493 /// Do we need to gather this sequence ?
494 bool NeedToGather;
495
496 /// Do we need to shuffle the load ?
497 bool NeedToShuffle;
498 };
499
500 /// Create a new VectorizableTree entry.
501 TreeEntry *newTreeEntry(ArrayRef<Value *> VL, bool Vectorized,
502 bool NeedToShuffle) {
503 VectorizableTree.emplace_back();
504 int idx = VectorizableTree.size() - 1;
505 TreeEntry *Last = &VectorizableTree[idx];
506 Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end());
507 Last->NeedToGather = !Vectorized;
508 Last->NeedToShuffle = NeedToShuffle;
509 if (Vectorized) {
510 for (int i = 0, e = VL.size(); i != e; ++i) {
511 assert(!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!")((!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!"
) ? static_cast<void> (0) : __assert_fail ("!ScalarToTreeEntry.count(VL[i]) && \"Scalar already in tree!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 511, __PRETTY_FUNCTION__))
;
512 ScalarToTreeEntry[VL[i]] = idx;
513 }
514 } else {
515 MustGather.insert(VL.begin(), VL.end());
516 }
517 return Last;
518 }
519
520 /// -- Vectorization State --
521 /// Holds all of the tree entries.
522 std::vector<TreeEntry> VectorizableTree;
523
524 /// Maps a specific scalar to its tree entry.
525 SmallDenseMap<Value*, int> ScalarToTreeEntry;
526
527 /// A list of scalars that we found that we need to keep as scalars.
528 ValueSet MustGather;
529
530 /// This POD struct describes one external user in the vectorized tree.
531 struct ExternalUser {
532 ExternalUser (Value *S, llvm::User *U, int L) :
533 Scalar(S), User(U), Lane(L){}
534 // Which scalar in our function.
535 Value *Scalar;
536 // Which user that uses the scalar.
537 llvm::User *User;
538 // Which lane does the scalar belong to.
539 int Lane;
540 };
541 typedef SmallVector<ExternalUser, 16> UserList;
542
543 /// Checks if two instructions may access the same memory.
544 ///
545 /// \p Loc1 is the location of \p Inst1. It is passed explicitly because it
546 /// is invariant in the calling loop.
547 bool isAliased(const MemoryLocation &Loc1, Instruction *Inst1,
548 Instruction *Inst2) {
549
550 // First check if the result is already in the cache.
551 AliasCacheKey key = std::make_pair(Inst1, Inst2);
552 Optional<bool> &result = AliasCache[key];
553 if (result.hasValue()) {
554 return result.getValue();
555 }
556 MemoryLocation Loc2 = getLocation(Inst2, AA);
557 bool aliased = true;
558 if (Loc1.Ptr && Loc2.Ptr && isSimple(Inst1) && isSimple(Inst2)) {
559 // Do the alias check.
560 aliased = AA->alias(Loc1, Loc2);
561 }
562 // Store the result in the cache.
563 result = aliased;
564 return aliased;
565 }
566
567 typedef std::pair<Instruction *, Instruction *> AliasCacheKey;
568
569 /// Cache for alias results.
570 /// TODO: consider moving this to the AliasAnalysis itself.
571 DenseMap<AliasCacheKey, Optional<bool>> AliasCache;
572
573 /// Removes an instruction from its block and eventually deletes it.
574 /// It's like Instruction::eraseFromParent() except that the actual deletion
575 /// is delayed until BoUpSLP is destructed.
576 /// This is required to ensure that there are no incorrect collisions in the
577 /// AliasCache, which can happen if a new instruction is allocated at the
578 /// same address as a previously deleted instruction.
579 void eraseInstruction(Instruction *I) {
580 I->removeFromParent();
581 I->dropAllReferences();
582 DeletedInstructions.push_back(std::unique_ptr<Instruction>(I));
583 }
584
585 /// Temporary store for deleted instructions. Instructions will be deleted
586 /// eventually when the BoUpSLP is destructed.
587 SmallVector<std::unique_ptr<Instruction>, 8> DeletedInstructions;
588
589 /// A list of values that need to extracted out of the tree.
590 /// This list holds pairs of (Internal Scalar : External User). External User
591 /// can be nullptr, it means that this Internal Scalar will be used later,
592 /// after vectorization.
593 UserList ExternalUses;
594
595 /// Values used only by @llvm.assume calls.
596 SmallPtrSet<const Value *, 32> EphValues;
597
598 /// Holds all of the instructions that we gathered.
599 SetVector<Instruction *> GatherSeq;
600 /// A list of blocks that we are going to CSE.
601 SetVector<BasicBlock *> CSEBlocks;
602
603 /// Contains all scheduling relevant data for an instruction.
604 /// A ScheduleData either represents a single instruction or a member of an
605 /// instruction bundle (= a group of instructions which is combined into a
606 /// vector instruction).
607 struct ScheduleData {
608
609 // The initial value for the dependency counters. It means that the
610 // dependencies are not calculated yet.
611 enum { InvalidDeps = -1 };
612
613 ScheduleData()
614 : Inst(nullptr), FirstInBundle(nullptr), NextInBundle(nullptr),
615 NextLoadStore(nullptr), SchedulingRegionID(0), SchedulingPriority(0),
616 Dependencies(InvalidDeps), UnscheduledDeps(InvalidDeps),
617 UnscheduledDepsInBundle(InvalidDeps), IsScheduled(false) {}
618
619 void init(int BlockSchedulingRegionID) {
620 FirstInBundle = this;
621 NextInBundle = nullptr;
622 NextLoadStore = nullptr;
623 IsScheduled = false;
624 SchedulingRegionID = BlockSchedulingRegionID;
625 UnscheduledDepsInBundle = UnscheduledDeps;
626 clearDependencies();
627 }
628
629 /// Returns true if the dependency information has been calculated.
630 bool hasValidDependencies() const { return Dependencies != InvalidDeps; }
631
632 /// Returns true for single instructions and for bundle representatives
633 /// (= the head of a bundle).
634 bool isSchedulingEntity() const { return FirstInBundle == this; }
635
636 /// Returns true if it represents an instruction bundle and not only a
637 /// single instruction.
638 bool isPartOfBundle() const {
639 return NextInBundle != nullptr || FirstInBundle != this;
640 }
641
642 /// Returns true if it is ready for scheduling, i.e. it has no more
643 /// unscheduled depending instructions/bundles.
644 bool isReady() const {
645 assert(isSchedulingEntity() &&((isSchedulingEntity() && "can't consider non-scheduling entity for ready list"
) ? static_cast<void> (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 646, __PRETTY_FUNCTION__))
646 "can't consider non-scheduling entity for ready list")((isSchedulingEntity() && "can't consider non-scheduling entity for ready list"
) ? static_cast<void> (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 646, __PRETTY_FUNCTION__))
;
647 return UnscheduledDepsInBundle == 0 && !IsScheduled;
648 }
649
650 /// Modifies the number of unscheduled dependencies, also updating it for
651 /// the whole bundle.
652 int incrementUnscheduledDeps(int Incr) {
653 UnscheduledDeps += Incr;
654 return FirstInBundle->UnscheduledDepsInBundle += Incr;
655 }
656
657 /// Sets the number of unscheduled dependencies to the number of
658 /// dependencies.
659 void resetUnscheduledDeps() {
660 incrementUnscheduledDeps(Dependencies - UnscheduledDeps);
661 }
662
663 /// Clears all dependency information.
664 void clearDependencies() {
665 Dependencies = InvalidDeps;
666 resetUnscheduledDeps();
667 MemoryDependencies.clear();
668 }
669
670 void dump(raw_ostream &os) const {
671 if (!isSchedulingEntity()) {
672 os << "/ " << *Inst;
673 } else if (NextInBundle) {
674 os << '[' << *Inst;
675 ScheduleData *SD = NextInBundle;
676 while (SD) {
677 os << ';' << *SD->Inst;
678 SD = SD->NextInBundle;
679 }
680 os << ']';
681 } else {
682 os << *Inst;
683 }
684 }
685
686 Instruction *Inst;
687
688 /// Points to the head in an instruction bundle (and always to this for
689 /// single instructions).
690 ScheduleData *FirstInBundle;
691
692 /// Single linked list of all instructions in a bundle. Null if it is a
693 /// single instruction.
694 ScheduleData *NextInBundle;
695
696 /// Single linked list of all memory instructions (e.g. load, store, call)
697 /// in the block - until the end of the scheduling region.
698 ScheduleData *NextLoadStore;
699
700 /// The dependent memory instructions.
701 /// This list is derived on demand in calculateDependencies().
702 SmallVector<ScheduleData *, 4> MemoryDependencies;
703
704 /// This ScheduleData is in the current scheduling region if this matches
705 /// the current SchedulingRegionID of BlockScheduling.
706 int SchedulingRegionID;
707
708 /// Used for getting a "good" final ordering of instructions.
709 int SchedulingPriority;
710
711 /// The number of dependencies. Constitutes of the number of users of the
712 /// instruction plus the number of dependent memory instructions (if any).
713 /// This value is calculated on demand.
714 /// If InvalidDeps, the number of dependencies is not calculated yet.
715 ///
716 int Dependencies;
717
718 /// The number of dependencies minus the number of dependencies of scheduled
719 /// instructions. As soon as this is zero, the instruction/bundle gets ready
720 /// for scheduling.
721 /// Note that this is negative as long as Dependencies is not calculated.
722 int UnscheduledDeps;
723
724 /// The sum of UnscheduledDeps in a bundle. Equals to UnscheduledDeps for
725 /// single instructions.
726 int UnscheduledDepsInBundle;
727
728 /// True if this instruction is scheduled (or considered as scheduled in the
729 /// dry-run).
730 bool IsScheduled;
731 };
732
733#ifndef NDEBUG
734 friend inline raw_ostream &operator<<(raw_ostream &os,
735 const BoUpSLP::ScheduleData &SD) {
736 SD.dump(os);
737 return os;
738 }
739#endif
740
741 /// Contains all scheduling data for a basic block.
742 ///
743 struct BlockScheduling {
744
745 BlockScheduling(BasicBlock *BB)
746 : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize),
747 ScheduleStart(nullptr), ScheduleEnd(nullptr),
748 FirstLoadStoreInRegion(nullptr), LastLoadStoreInRegion(nullptr),
749 ScheduleRegionSize(0),
750 ScheduleRegionSizeLimit(ScheduleRegionSizeBudget),
751 // Make sure that the initial SchedulingRegionID is greater than the
752 // initial SchedulingRegionID in ScheduleData (which is 0).
753 SchedulingRegionID(1) {}
754
755 void clear() {
756 ReadyInsts.clear();
757 ScheduleStart = nullptr;
758 ScheduleEnd = nullptr;
759 FirstLoadStoreInRegion = nullptr;
760 LastLoadStoreInRegion = nullptr;
761
762 // Reduce the maximum schedule region size by the size of the
763 // previous scheduling run.
764 ScheduleRegionSizeLimit -= ScheduleRegionSize;
765 if (ScheduleRegionSizeLimit < MinScheduleRegionSize)
766 ScheduleRegionSizeLimit = MinScheduleRegionSize;
767 ScheduleRegionSize = 0;
768
769 // Make a new scheduling region, i.e. all existing ScheduleData is not
770 // in the new region yet.
771 ++SchedulingRegionID;
772 }
773
774 ScheduleData *getScheduleData(Value *V) {
775 ScheduleData *SD = ScheduleDataMap[V];
776 if (SD && SD->SchedulingRegionID == SchedulingRegionID)
777 return SD;
778 return nullptr;
779 }
780
781 bool isInSchedulingRegion(ScheduleData *SD) {
782 return SD->SchedulingRegionID == SchedulingRegionID;
783 }
784
785 /// Marks an instruction as scheduled and puts all dependent ready
786 /// instructions into the ready-list.
787 template <typename ReadyListType>
788 void schedule(ScheduleData *SD, ReadyListType &ReadyList) {
789 SD->IsScheduled = true;
790 DEBUG(dbgs() << "SLP: schedule " << *SD << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: schedule " << *SD <<
"\n"; } } while (false)
;
791
792 ScheduleData *BundleMember = SD;
793 while (BundleMember) {
794 // Handle the def-use chain dependencies.
795 for (Use &U : BundleMember->Inst->operands()) {
796 ScheduleData *OpDef = getScheduleData(U.get());
797 if (OpDef && OpDef->hasValidDependencies() &&
798 OpDef->incrementUnscheduledDeps(-1) == 0) {
799 // There are no more unscheduled dependencies after decrementing,
800 // so we can put the dependent instruction into the ready list.
801 ScheduleData *DepBundle = OpDef->FirstInBundle;
802 assert(!DepBundle->IsScheduled &&((!DepBundle->IsScheduled && "already scheduled bundle gets ready"
) ? static_cast<void> (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 803, __PRETTY_FUNCTION__))
803 "already scheduled bundle gets ready")((!DepBundle->IsScheduled && "already scheduled bundle gets ready"
) ? static_cast<void> (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 803, __PRETTY_FUNCTION__))
;
804 ReadyList.insert(DepBundle);
805 DEBUG(dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: gets ready (def): " <<
*DepBundle << "\n"; } } while (false)
;
806 }
807 }
808 // Handle the memory dependencies.
809 for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) {
810 if (MemoryDepSD->incrementUnscheduledDeps(-1) == 0) {
811 // There are no more unscheduled dependencies after decrementing,
812 // so we can put the dependent instruction into the ready list.
813 ScheduleData *DepBundle = MemoryDepSD->FirstInBundle;
814 assert(!DepBundle->IsScheduled &&((!DepBundle->IsScheduled && "already scheduled bundle gets ready"
) ? static_cast<void> (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 815, __PRETTY_FUNCTION__))
815 "already scheduled bundle gets ready")((!DepBundle->IsScheduled && "already scheduled bundle gets ready"
) ? static_cast<void> (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 815, __PRETTY_FUNCTION__))
;
816 ReadyList.insert(DepBundle);
817 DEBUG(dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: gets ready (mem): " <<
*DepBundle << "\n"; } } while (false)
;
818 }
819 }
820 BundleMember = BundleMember->NextInBundle;
821 }
822 }
823
824 /// Put all instructions into the ReadyList which are ready for scheduling.
825 template <typename ReadyListType>
826 void initialFillReadyList(ReadyListType &ReadyList) {
827 for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
828 ScheduleData *SD = getScheduleData(I);
829 if (SD->isSchedulingEntity() && SD->isReady()) {
830 ReadyList.insert(SD);
831 DEBUG(dbgs() << "SLP: initially in ready list: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: initially in ready list: "
<< *I << "\n"; } } while (false)
;
832 }
833 }
834 }
835
836 /// Checks if a bundle of instructions can be scheduled, i.e. has no
837 /// cyclic dependencies. This is only a dry-run, no instructions are
838 /// actually moved at this stage.
839 bool tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP);
840
841 /// Un-bundles a group of instructions.
842 void cancelScheduling(ArrayRef<Value *> VL);
843
844 /// Extends the scheduling region so that V is inside the region.
845 /// \returns true if the region size is within the limit.
846 bool extendSchedulingRegion(Value *V);
847
848 /// Initialize the ScheduleData structures for new instructions in the
849 /// scheduling region.
850 void initScheduleData(Instruction *FromI, Instruction *ToI,
851 ScheduleData *PrevLoadStore,
852 ScheduleData *NextLoadStore);
853
854 /// Updates the dependency information of a bundle and of all instructions/
855 /// bundles which depend on the original bundle.
856 void calculateDependencies(ScheduleData *SD, bool InsertInReadyList,
857 BoUpSLP *SLP);
858
859 /// Sets all instruction in the scheduling region to un-scheduled.
860 void resetSchedule();
861
862 BasicBlock *BB;
863
864 /// Simple memory allocation for ScheduleData.
865 std::vector<std::unique_ptr<ScheduleData[]>> ScheduleDataChunks;
866
867 /// The size of a ScheduleData array in ScheduleDataChunks.
868 int ChunkSize;
869
870 /// The allocator position in the current chunk, which is the last entry
871 /// of ScheduleDataChunks.
872 int ChunkPos;
873
874 /// Attaches ScheduleData to Instruction.
875 /// Note that the mapping survives during all vectorization iterations, i.e.
876 /// ScheduleData structures are recycled.
877 DenseMap<Value *, ScheduleData *> ScheduleDataMap;
878
879 struct ReadyList : SmallVector<ScheduleData *, 8> {
880 void insert(ScheduleData *SD) { push_back(SD); }
881 };
882
883 /// The ready-list for scheduling (only used for the dry-run).
884 ReadyList ReadyInsts;
885
886 /// The first instruction of the scheduling region.
887 Instruction *ScheduleStart;
888
889 /// The first instruction _after_ the scheduling region.
890 Instruction *ScheduleEnd;
891
892 /// The first memory accessing instruction in the scheduling region
893 /// (can be null).
894 ScheduleData *FirstLoadStoreInRegion;
895
896 /// The last memory accessing instruction in the scheduling region
897 /// (can be null).
898 ScheduleData *LastLoadStoreInRegion;
899
900 /// The current size of the scheduling region.
901 int ScheduleRegionSize;
902
903 /// The maximum size allowed for the scheduling region.
904 int ScheduleRegionSizeLimit;
905
906 /// The ID of the scheduling region. For a new vectorization iteration this
907 /// is incremented which "removes" all ScheduleData from the region.
908 int SchedulingRegionID;
909 };
910
911 /// Attaches the BlockScheduling structures to basic blocks.
912 MapVector<BasicBlock *, std::unique_ptr<BlockScheduling>> BlocksSchedules;
913
914 /// Performs the "real" scheduling. Done before vectorization is actually
915 /// performed in a basic block.
916 void scheduleBlock(BlockScheduling *BS);
917
918 /// List of users to ignore during scheduling and that don't need extracting.
919 ArrayRef<Value *> UserIgnoreList;
920
921 // Number of load bundles that contain consecutive loads.
922 int NumLoadsWantToKeepOrder;
923
924 // Number of load bundles that contain consecutive loads in reversed order.
925 int NumLoadsWantToChangeOrder;
926
927 // Analysis and block reference.
928 Function *F;
929 ScalarEvolution *SE;
930 TargetTransformInfo *TTI;
931 TargetLibraryInfo *TLI;
932 AliasAnalysis *AA;
933 LoopInfo *LI;
934 DominatorTree *DT;
935 AssumptionCache *AC;
936 DemandedBits *DB;
937 const DataLayout *DL;
938 unsigned MaxVecRegSize; // This is set by TTI or overridden by cl::opt.
939 unsigned MinVecRegSize; // Set by cl::opt (default: 128).
940 /// Instruction builder to construct the vectorized tree.
941 IRBuilder<> Builder;
942
943 /// A map of scalar integer values to the smallest bit width with which they
944 /// can legally be represented. The values map to (width, signed) pairs,
945 /// where "width" indicates the minimum bit width and "signed" is True if the
946 /// value must be signed-extended, rather than zero-extended, back to its
947 /// original width.
948 MapVector<Value *, std::pair<uint64_t, bool>> MinBWs;
949};
950
951} // end namespace llvm
952} // end namespace slpvectorizer
953
954void BoUpSLP::buildTree(ArrayRef<Value *> Roots,
955 ArrayRef<Value *> UserIgnoreLst) {
956 MapVector<Value *, DebugLoc> ExternallyUsedValues;
957 buildTree(Roots, ExternallyUsedValues, UserIgnoreLst);
958}
959void BoUpSLP::buildTree(ArrayRef<Value *> Roots,
960 MapVector<Value *, DebugLoc> &ExternallyUsedValues,
961 ArrayRef<Value *> UserIgnoreLst) {
962 deleteTree();
963 UserIgnoreList = UserIgnoreLst;
964 if (!allSameType(Roots))
965 return;
966 buildTree_rec(Roots, 0);
967
968 // Collect the values that we need to extract from the tree.
969 for (TreeEntry &EIdx : VectorizableTree) {
970 TreeEntry *Entry = &EIdx;
971
972 // For each lane:
973 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
974 Value *Scalar = Entry->Scalars[Lane];
975
976 // No need to handle users of gathered values.
977 if (Entry->NeedToGather)
978 continue;
979
980 // Check if the scalar is externally used as an extra arg.
981 auto ExtI = ExternallyUsedValues.find(Scalar);
982 if (ExtI != ExternallyUsedValues.end()) {
983 DEBUG(dbgs() << "SLP: Need to extract: Extra arg from lane " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Need to extract: Extra arg from lane "
<< Lane << " from " << *Scalar << ".\n"
; } } while (false)
984 Lane << " from " << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Need to extract: Extra arg from lane "
<< Lane << " from " << *Scalar << ".\n"
; } } while (false)
;
985 ExternalUses.emplace_back(Scalar, nullptr, Lane);
986 continue;
987 }
988 for (User *U : Scalar->users()) {
989 DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Checking user:" << *U <<
".\n"; } } while (false)
;
990
991 Instruction *UserInst = dyn_cast<Instruction>(U);
992 if (!UserInst)
993 continue;
994
995 // Skip in-tree scalars that become vectors
996 if (ScalarToTreeEntry.count(U)) {
997 int Idx = ScalarToTreeEntry[U];
998 TreeEntry *UseEntry = &VectorizableTree[Idx];
999 Value *UseScalar = UseEntry->Scalars[0];
1000 // Some in-tree scalars will remain as scalar in vectorized
1001 // instructions. If that is the case, the one in Lane 0 will
1002 // be used.
1003 if (UseScalar != U ||
1004 !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) {
1005 DEBUG(dbgs() << "SLP: \tInternal user will be removed:" << *Udo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: \tInternal user will be removed:"
<< *U << ".\n"; } } while (false)
1006 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: \tInternal user will be removed:"
<< *U << ".\n"; } } while (false)
;
1007 assert(!VectorizableTree[Idx].NeedToGather && "Bad state")((!VectorizableTree[Idx].NeedToGather && "Bad state")
? static_cast<void> (0) : __assert_fail ("!VectorizableTree[Idx].NeedToGather && \"Bad state\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1007, __PRETTY_FUNCTION__))
;
1008 continue;
1009 }
1010 }
1011
1012 // Ignore users in the user ignore list.
1013 if (is_contained(UserIgnoreList, UserInst))
1014 continue;
1015
1016 DEBUG(dbgs() << "SLP: Need to extract:" << *U << " from lane " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Need to extract:" << *
U << " from lane " << Lane << " from " <<
*Scalar << ".\n"; } } while (false)
1017 Lane << " from " << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Need to extract:" << *
U << " from lane " << Lane << " from " <<
*Scalar << ".\n"; } } while (false)
;
1018 ExternalUses.push_back(ExternalUser(Scalar, U, Lane));
1019 }
1020 }
1021 }
1022}
1023
1024
1025void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth) {
1026 bool isAltShuffle = false;
1027 assert((allConstant(VL) || allSameType(VL)) && "Invalid types!")(((allConstant(VL) || allSameType(VL)) && "Invalid types!"
) ? static_cast<void> (0) : __assert_fail ("(allConstant(VL) || allSameType(VL)) && \"Invalid types!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1027, __PRETTY_FUNCTION__))
;
1028
1029 if (Depth == RecursionMaxDepth) {
1030 DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering due to max recursion depth.\n"
; } } while (false)
;
1031 newTreeEntry(VL, false, false);
1032 return;
1033 }
1034
1035 // Don't handle vectors.
1036 if (VL[0]->getType()->isVectorTy()) {
1037 DEBUG(dbgs() << "SLP: Gathering due to vector type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering due to vector type.\n"
; } } while (false)
;
1038 newTreeEntry(VL, false, false);
1039 return;
1040 }
1041
1042 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1043 if (SI->getValueOperand()->getType()->isVectorTy()) {
1044 DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering due to store vector type.\n"
; } } while (false)
;
1045 newTreeEntry(VL, false, false);
1046 return;
1047 }
1048 unsigned Opcode = getSameOpcode(VL);
1049
1050 // Check that this shuffle vector refers to the alternate
1051 // sequence of opcodes.
1052 if (Opcode == Instruction::ShuffleVector) {
1053 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
1054 unsigned Op = I0->getOpcode();
1055 if (Op != Instruction::ShuffleVector)
1056 isAltShuffle = true;
1057 }
1058
1059 // If all of the operands are identical or constant we have a simple solution.
1060 if (allConstant(VL) || isSplat(VL) || !allSameBlock(VL) || !Opcode) {
1061 DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering due to C,S,B,O. \n"
; } } while (false)
;
1062 newTreeEntry(VL, false, false);
1063 return;
1064 }
1065
1066 // We now know that this is a vector of instructions of the same type from
1067 // the same block.
1068
1069 // Don't vectorize ephemeral values.
1070 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
1071 if (EphValues.count(VL[i])) {
1072 DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: The instruction (" << *
VL[i] << ") is ephemeral.\n"; } } while (false)
1073 ") is ephemeral.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: The instruction (" << *
VL[i] << ") is ephemeral.\n"; } } while (false)
;
1074 newTreeEntry(VL, false, false);
1075 return;
1076 }
1077 }
1078
1079 // Check if this is a duplicate of another entry.
1080 if (ScalarToTreeEntry.count(VL[0])) {
1081 int Idx = ScalarToTreeEntry[VL[0]];
1082 TreeEntry *E = &VectorizableTree[Idx];
1083 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
1084 DEBUG(dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: \tChecking bundle: " <<
*VL[i] << ".\n"; } } while (false)
;
1085 if (E->Scalars[i] != VL[i]) {
1086 DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering due to partial overlap.\n"
; } } while (false)
;
1087 newTreeEntry(VL, false, false);
1088 return;
1089 }
1090 }
1091 DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *VL[0] << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Perfect diamond merge at " <<
*VL[0] << ".\n"; } } while (false)
;
1092 return;
1093 }
1094
1095 // Check that none of the instructions in the bundle are already in the tree.
1096 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
1097 if (ScalarToTreeEntry.count(VL[i])) {
1098 DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: The instruction (" << *
VL[i] << ") is already in tree.\n"; } } while (false)
1099 ") is already in tree.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: The instruction (" << *
VL[i] << ") is already in tree.\n"; } } while (false)
;
1100 newTreeEntry(VL, false, false);
1101 return;
1102 }
1103 }
1104
1105 // If any of the scalars is marked as a value that needs to stay scalar then
1106 // we need to gather the scalars.
1107 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
1108 if (MustGather.count(VL[i])) {
1109 DEBUG(dbgs() << "SLP: Gathering due to gathered scalar.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering due to gathered scalar.\n"
; } } while (false)
;
1110 newTreeEntry(VL, false, false);
1111 return;
1112 }
1113 }
1114
1115 // Check that all of the users of the scalars that we want to vectorize are
1116 // schedulable.
1117 Instruction *VL0 = cast<Instruction>(VL[0]);
1118 BasicBlock *BB = cast<Instruction>(VL0)->getParent();
1119
1120 if (!DT->isReachableFromEntry(BB)) {
1121 // Don't go into unreachable blocks. They may contain instructions with
1122 // dependency cycles which confuse the final scheduling.
1123 DEBUG(dbgs() << "SLP: bundle in unreachable block.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: bundle in unreachable block.\n"
; } } while (false)
;
1124 newTreeEntry(VL, false, false);
1125 return;
1126 }
1127
1128 // Check that every instructions appears once in this bundle.
1129 for (unsigned i = 0, e = VL.size(); i < e; ++i)
1130 for (unsigned j = i+1; j < e; ++j)
1131 if (VL[i] == VL[j]) {
1132 DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Scalar used twice in bundle.\n"
; } } while (false)
;
1133 newTreeEntry(VL, false, false);
1134 return;
1135 }
1136
1137 auto &BSRef = BlocksSchedules[BB];
1138 if (!BSRef) {
1139 BSRef = llvm::make_unique<BlockScheduling>(BB);
1140 }
1141 BlockScheduling &BS = *BSRef.get();
1142
1143 if (!BS.tryScheduleBundle(VL, this)) {
1144 DEBUG(dbgs() << "SLP: We are not able to schedule this bundle!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: We are not able to schedule this bundle!\n"
; } } while (false)
;
1145 assert((!BS.getScheduleData(VL[0]) ||(((!BS.getScheduleData(VL[0]) || !BS.getScheduleData(VL[0])->
isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure"
) ? static_cast<void> (0) : __assert_fail ("(!BS.getScheduleData(VL[0]) || !BS.getScheduleData(VL[0])->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1147, __PRETTY_FUNCTION__))
1146 !BS.getScheduleData(VL[0])->isPartOfBundle()) &&(((!BS.getScheduleData(VL[0]) || !BS.getScheduleData(VL[0])->
isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure"
) ? static_cast<void> (0) : __assert_fail ("(!BS.getScheduleData(VL[0]) || !BS.getScheduleData(VL[0])->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1147, __PRETTY_FUNCTION__))
1147 "tryScheduleBundle should cancelScheduling on failure")(((!BS.getScheduleData(VL[0]) || !BS.getScheduleData(VL[0])->
isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure"
) ? static_cast<void> (0) : __assert_fail ("(!BS.getScheduleData(VL[0]) || !BS.getScheduleData(VL[0])->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1147, __PRETTY_FUNCTION__))
;
1148 newTreeEntry(VL, false, false);
1149 return;
1150 }
1151 DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: We are able to schedule this bundle.\n"
; } } while (false)
;
1152
1153 switch (Opcode) {
1154 case Instruction::PHI: {
1155 PHINode *PH = dyn_cast<PHINode>(VL0);
1156
1157 // Check for terminator values (e.g. invoke).
1158 for (unsigned j = 0; j < VL.size(); ++j)
1159 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1160 TerminatorInst *Term = dyn_cast<TerminatorInst>(
1161 cast<PHINode>(VL[j])->getIncomingValueForBlock(PH->getIncomingBlock(i)));
1162 if (Term) {
1163 DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n"
; } } while (false)
;
1164 BS.cancelScheduling(VL);
1165 newTreeEntry(VL, false, false);
1166 return;
1167 }
1168 }
1169
1170 newTreeEntry(VL, true, false);
1171 DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: added a vector of PHINodes.\n"
; } } while (false)
;
1172
1173 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1174 ValueList Operands;
1175 // Prepare the operand vector.
1176 for (Value *j : VL)
1177 Operands.push_back(cast<PHINode>(j)->getIncomingValueForBlock(
1178 PH->getIncomingBlock(i)));
1179
1180 buildTree_rec(Operands, Depth + 1);
1181 }
1182 return;
1183 }
1184 case Instruction::ExtractValue:
1185 case Instruction::ExtractElement: {
1186 bool Reuse = canReuseExtract(VL, Opcode);
1187 if (Reuse) {
1188 DEBUG(dbgs() << "SLP: Reusing extract sequence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Reusing extract sequence.\n"
; } } while (false)
;
1189 } else {
1190 BS.cancelScheduling(VL);
1191 }
1192 newTreeEntry(VL, Reuse, false);
1193 return;
1194 }
1195 case Instruction::Load: {
1196 // Check that a vectorized load would load the same memory as a scalar
1197 // load.
1198 // For example we don't want vectorize loads that are smaller than 8 bit.
1199 // Even though we have a packed struct {<i2, i2, i2, i2>} LLVM treats
1200 // loading/storing it as an i8 struct. If we vectorize loads/stores from
1201 // such a struct we read/write packed bits disagreeing with the
1202 // unvectorized version.
1203 Type *ScalarTy = VL[0]->getType();
1204
1205 if (DL->getTypeSizeInBits(ScalarTy) !=
1206 DL->getTypeAllocSizeInBits(ScalarTy)) {
1207 BS.cancelScheduling(VL);
1208 newTreeEntry(VL, false, false);
1209 DEBUG(dbgs() << "SLP: Gathering loads of non-packed type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering loads of non-packed type.\n"
; } } while (false)
;
1210 return;
1211 }
1212
1213 // Make sure all loads in the bundle are simple - we can't vectorize
1214 // atomic or volatile loads.
1215 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) {
1216 LoadInst *L = cast<LoadInst>(VL[i]);
1217 if (!L->isSimple()) {
1218 BS.cancelScheduling(VL);
1219 newTreeEntry(VL, false, false);
1220 DEBUG(dbgs() << "SLP: Gathering non-simple loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering non-simple loads.\n"
; } } while (false)
;
1221 return;
1222 }
1223 }
1224
1225 // Check if the loads are consecutive, reversed, or neither.
1226 bool Consecutive = true;
1227 bool ReverseConsecutive = true;
1228 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) {
1229 if (!isConsecutiveAccess(VL[i], VL[i + 1], *DL, *SE)) {
1230 Consecutive = false;
1231 break;
1232 } else {
1233 ReverseConsecutive = false;
1234 }
1235 }
1236
1237 if (Consecutive) {
1238 ++NumLoadsWantToKeepOrder;
1239 newTreeEntry(VL, true, false);
1240 DEBUG(dbgs() << "SLP: added a vector of loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: added a vector of loads.\n";
} } while (false)
;
1241 return;
1242 }
1243
1244 // If none of the load pairs were consecutive when checked in order,
1245 // check the reverse order.
1246 if (ReverseConsecutive)
1247 for (unsigned i = VL.size() - 1; i > 0; --i)
1248 if (!isConsecutiveAccess(VL[i], VL[i - 1], *DL, *SE)) {
1249 ReverseConsecutive = false;
1250 break;
1251 }
1252
1253 if (VL.size() > 2 && !ReverseConsecutive) {
1254 bool ShuffledLoads = true;
1255 SmallVector<Value *, 8> Sorted;
1256 if (sortMemAccesses(VL, *DL, *SE, Sorted)) {
1257 auto NewVL = makeArrayRef(Sorted.begin(), Sorted.end());
1258 for (unsigned i = 0, e = NewVL.size() - 1; i < e; ++i) {
1259 if (!isConsecutiveAccess(NewVL[i], NewVL[i + 1], *DL, *SE)) {
1260 ShuffledLoads = false;
1261 break;
1262 }
1263 }
1264 if (ShuffledLoads) {
1265 newTreeEntry(NewVL, true, true);
1266 return;
1267 }
1268 }
1269 }
1270
1271 BS.cancelScheduling(VL);
1272 newTreeEntry(VL, false, false);
1273
1274 if (ReverseConsecutive) {
1275 ++NumLoadsWantToChangeOrder;
1276 DEBUG(dbgs() << "SLP: Gathering reversed loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering reversed loads.\n"
; } } while (false)
;
1277 } else {
1278 DEBUG(dbgs() << "SLP: Gathering non-consecutive loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering non-consecutive loads.\n"
; } } while (false)
;
1279 }
1280 return;
1281 }
1282 case Instruction::ZExt:
1283 case Instruction::SExt:
1284 case Instruction::FPToUI:
1285 case Instruction::FPToSI:
1286 case Instruction::FPExt:
1287 case Instruction::PtrToInt:
1288 case Instruction::IntToPtr:
1289 case Instruction::SIToFP:
1290 case Instruction::UIToFP:
1291 case Instruction::Trunc:
1292 case Instruction::FPTrunc:
1293 case Instruction::BitCast: {
1294 Type *SrcTy = VL0->getOperand(0)->getType();
1295 for (Value *Val : VL) {
1296 Type *Ty = cast<Instruction>(Val)->getOperand(0)->getType();
1297 if (Ty != SrcTy || !isValidElementType(Ty)) {
1298 BS.cancelScheduling(VL);
1299 newTreeEntry(VL, false, false);
1300 DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering casts with different src types.\n"
; } } while (false)
;
1301 return;
1302 }
1303 }
1304 newTreeEntry(VL, true, false);
1305 DEBUG(dbgs() << "SLP: added a vector of casts.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: added a vector of casts.\n";
} } while (false)
;
1306
1307 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
1308 ValueList Operands;
1309 // Prepare the operand vector.
1310 for (Value *j : VL)
1311 Operands.push_back(cast<Instruction>(j)->getOperand(i));
1312
1313 buildTree_rec(Operands, Depth+1);
1314 }
1315 return;
1316 }
1317 case Instruction::ICmp:
1318 case Instruction::FCmp: {
1319 // Check that all of the compares have the same predicate.
1320 CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate();
1321 Type *ComparedTy = cast<Instruction>(VL[0])->getOperand(0)->getType();
1322 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
1323 CmpInst *Cmp = cast<CmpInst>(VL[i]);
1324 if (Cmp->getPredicate() != P0 ||
1325 Cmp->getOperand(0)->getType() != ComparedTy) {
1326 BS.cancelScheduling(VL);
1327 newTreeEntry(VL, false, false);
1328 DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering cmp with different predicate.\n"
; } } while (false)
;
1329 return;
1330 }
1331 }
1332
1333 newTreeEntry(VL, true, false);
1334 DEBUG(dbgs() << "SLP: added a vector of compares.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: added a vector of compares.\n"
; } } while (false)
;
1335
1336 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
1337 ValueList Operands;
1338 // Prepare the operand vector.
1339 for (Value *j : VL)
1340 Operands.push_back(cast<Instruction>(j)->getOperand(i));
1341
1342 buildTree_rec(Operands, Depth+1);
1343 }
1344 return;
1345 }
1346 case Instruction::Select:
1347 case Instruction::Add:
1348 case Instruction::FAdd:
1349 case Instruction::Sub:
1350 case Instruction::FSub:
1351 case Instruction::Mul:
1352 case Instruction::FMul:
1353 case Instruction::UDiv:
1354 case Instruction::SDiv:
1355 case Instruction::FDiv:
1356 case Instruction::URem:
1357 case Instruction::SRem:
1358 case Instruction::FRem:
1359 case Instruction::Shl:
1360 case Instruction::LShr:
1361 case Instruction::AShr:
1362 case Instruction::And:
1363 case Instruction::Or:
1364 case Instruction::Xor: {
1365 newTreeEntry(VL, true, false);
1366 DEBUG(dbgs() << "SLP: added a vector of bin op.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: added a vector of bin op.\n"
; } } while (false)
;
1367
1368 // Sort operands of the instructions so that each side is more likely to
1369 // have the same opcode.
1370 if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) {
1371 ValueList Left, Right;
1372 reorderInputsAccordingToOpcode(VL, Left, Right);
1373 buildTree_rec(Left, Depth + 1);
1374 buildTree_rec(Right, Depth + 1);
1375 return;
1376 }
1377
1378 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
1379 ValueList Operands;
1380 // Prepare the operand vector.
1381 for (Value *j : VL)
1382 Operands.push_back(cast<Instruction>(j)->getOperand(i));
1383
1384 buildTree_rec(Operands, Depth+1);
1385 }
1386 return;
1387 }
1388 case Instruction::GetElementPtr: {
1389 // We don't combine GEPs with complicated (nested) indexing.
1390 for (Value *Val : VL) {
1391 if (cast<Instruction>(Val)->getNumOperands() != 2) {
1392 DEBUG(dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n"
; } } while (false)
;
1393 BS.cancelScheduling(VL);
1394 newTreeEntry(VL, false, false);
1395 return;
1396 }
1397 }
1398
1399 // We can't combine several GEPs into one vector if they operate on
1400 // different types.
1401 Type *Ty0 = cast<Instruction>(VL0)->getOperand(0)->getType();
1402 for (Value *Val : VL) {
1403 Type *CurTy = cast<Instruction>(Val)->getOperand(0)->getType();
1404 if (Ty0 != CurTy) {
1405 DEBUG(dbgs() << "SLP: not-vectorizable GEP (different types).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: not-vectorizable GEP (different types).\n"
; } } while (false)
;
1406 BS.cancelScheduling(VL);
1407 newTreeEntry(VL, false, false);
1408 return;
1409 }
1410 }
1411
1412 // We don't combine GEPs with non-constant indexes.
1413 for (Value *Val : VL) {
1414 auto Op = cast<Instruction>(Val)->getOperand(1);
1415 if (!isa<ConstantInt>(Op)) {
1416 DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n"
; } } while (false)
1417 dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n"
; } } while (false)
;
1418 BS.cancelScheduling(VL);
1419 newTreeEntry(VL, false, false);
1420 return;
1421 }
1422 }
1423
1424 newTreeEntry(VL, true, false);
1425 DEBUG(dbgs() << "SLP: added a vector of GEPs.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: added a vector of GEPs.\n"; }
} while (false)
;
1426 for (unsigned i = 0, e = 2; i < e; ++i) {
1427 ValueList Operands;
1428 // Prepare the operand vector.
1429 for (Value *j : VL)
1430 Operands.push_back(cast<Instruction>(j)->getOperand(i));
1431
1432 buildTree_rec(Operands, Depth + 1);
1433 }
1434 return;
1435 }
1436 case Instruction::Store: {
1437 // Check if the stores are consecutive or of we need to swizzle them.
1438 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
1439 if (!isConsecutiveAccess(VL[i], VL[i + 1], *DL, *SE)) {
1440 BS.cancelScheduling(VL);
1441 newTreeEntry(VL, false, false);
1442 DEBUG(dbgs() << "SLP: Non-consecutive store.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Non-consecutive store.\n"; }
} while (false)
;
1443 return;
1444 }
1445
1446 newTreeEntry(VL, true, false);
1447 DEBUG(dbgs() << "SLP: added a vector of stores.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: added a vector of stores.\n"
; } } while (false)
;
1448
1449 ValueList Operands;
1450 for (Value *j : VL)
1451 Operands.push_back(cast<Instruction>(j)->getOperand(0));
1452
1453 buildTree_rec(Operands, Depth + 1);
1454 return;
1455 }
1456 case Instruction::Call: {
1457 // Check if the calls are all to the same vectorizable intrinsic.
1458 CallInst *CI = cast<CallInst>(VL[0]);
1459 // Check if this is an Intrinsic call or something that can be
1460 // represented by an intrinsic call
1461 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
1462 if (!isTriviallyVectorizable(ID)) {
1463 BS.cancelScheduling(VL);
1464 newTreeEntry(VL, false, false);
1465 DEBUG(dbgs() << "SLP: Non-vectorizable call.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Non-vectorizable call.\n"; }
} while (false)
;
1466 return;
1467 }
1468 Function *Int = CI->getCalledFunction();
1469 Value *A1I = nullptr;
1470 if (hasVectorInstrinsicScalarOpd(ID, 1))
1471 A1I = CI->getArgOperand(1);
1472 for (unsigned i = 1, e = VL.size(); i != e; ++i) {
1473 CallInst *CI2 = dyn_cast<CallInst>(VL[i]);
1474 if (!CI2 || CI2->getCalledFunction() != Int ||
1475 getVectorIntrinsicIDForCall(CI2, TLI) != ID ||
1476 !CI->hasIdenticalOperandBundleSchema(*CI2)) {
1477 BS.cancelScheduling(VL);
1478 newTreeEntry(VL, false, false);
1479 DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *VL[i]do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: mismatched calls:" << *
CI << "!=" << *VL[i] << "\n"; } } while (false
)
1480 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: mismatched calls:" << *
CI << "!=" << *VL[i] << "\n"; } } while (false
)
;
1481 return;
1482 }
1483 // ctlz,cttz and powi are special intrinsics whose second argument
1484 // should be same in order for them to be vectorized.
1485 if (hasVectorInstrinsicScalarOpd(ID, 1)) {
1486 Value *A1J = CI2->getArgOperand(1);
1487 if (A1I != A1J) {
1488 BS.cancelScheduling(VL);
1489 newTreeEntry(VL, false, false);
1490 DEBUG(dbgs() << "SLP: mismatched arguments in call:" << *CIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: mismatched arguments in call:"
<< *CI << " argument "<< A1I<<"!=" <<
A1J << "\n"; } } while (false)
1491 << " argument "<< A1I<<"!=" << A1Jdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: mismatched arguments in call:"
<< *CI << " argument "<< A1I<<"!=" <<
A1J << "\n"; } } while (false)
1492 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: mismatched arguments in call:"
<< *CI << " argument "<< A1I<<"!=" <<
A1J << "\n"; } } while (false)
;
1493 return;
1494 }
1495 }
1496 // Verify that the bundle operands are identical between the two calls.
1497 if (CI->hasOperandBundles() &&
1498 !std::equal(CI->op_begin() + CI->getBundleOperandsStartIndex(),
1499 CI->op_begin() + CI->getBundleOperandsEndIndex(),
1500 CI2->op_begin() + CI2->getBundleOperandsStartIndex())) {
1501 BS.cancelScheduling(VL);
1502 newTreeEntry(VL, false, false);
1503 DEBUG(dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!="do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:"
<< *CI << "!=" << *VL[i] << '\n'; } }
while (false)
1504 << *VL[i] << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:"
<< *CI << "!=" << *VL[i] << '\n'; } }
while (false)
;
1505 return;
1506 }
1507 }
1508
1509 newTreeEntry(VL, true, false);
1510 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) {
1511 ValueList Operands;
1512 // Prepare the operand vector.
1513 for (Value *j : VL) {
1514 CallInst *CI2 = dyn_cast<CallInst>(j);
1515 Operands.push_back(CI2->getArgOperand(i));
1516 }
1517 buildTree_rec(Operands, Depth + 1);
1518 }
1519 return;
1520 }
1521 case Instruction::ShuffleVector: {
1522 // If this is not an alternate sequence of opcode like add-sub
1523 // then do not vectorize this instruction.
1524 if (!isAltShuffle) {
1525 BS.cancelScheduling(VL);
1526 newTreeEntry(VL, false, false);
1527 DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: ShuffleVector are not vectorized.\n"
; } } while (false)
;
1528 return;
1529 }
1530 newTreeEntry(VL, true, false);
1531 DEBUG(dbgs() << "SLP: added a ShuffleVector op.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: added a ShuffleVector op.\n"
; } } while (false)
;
1532
1533 // Reorder operands if reordering would enable vectorization.
1534 if (isa<BinaryOperator>(VL0)) {
1535 ValueList Left, Right;
1536 reorderAltShuffleOperands(VL, Left, Right);
1537 buildTree_rec(Left, Depth + 1);
1538 buildTree_rec(Right, Depth + 1);
1539 return;
1540 }
1541
1542 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
1543 ValueList Operands;
1544 // Prepare the operand vector.
1545 for (Value *j : VL)
1546 Operands.push_back(cast<Instruction>(j)->getOperand(i));
1547
1548 buildTree_rec(Operands, Depth + 1);
1549 }
1550 return;
1551 }
1552 default:
1553 BS.cancelScheduling(VL);
1554 newTreeEntry(VL, false, false);
1555 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering unknown instruction.\n"
; } } while (false)
;
1556 return;
1557 }
1558}
1559
1560unsigned BoUpSLP::canMapToVector(Type *T, const DataLayout &DL) const {
1561 unsigned N;
1562 Type *EltTy;
1563 auto *ST = dyn_cast<StructType>(T);
1564 if (ST) {
1565 N = ST->getNumElements();
1566 EltTy = *ST->element_begin();
1567 } else {
1568 N = cast<ArrayType>(T)->getNumElements();
1569 EltTy = cast<ArrayType>(T)->getElementType();
1570 }
1571 if (!isValidElementType(EltTy))
1572 return 0;
1573 uint64_t VTSize = DL.getTypeStoreSizeInBits(VectorType::get(EltTy, N));
1574 if (VTSize < MinVecRegSize || VTSize > MaxVecRegSize || VTSize != DL.getTypeStoreSizeInBits(T))
1575 return 0;
1576 if (ST) {
1577 // Check that struct is homogeneous.
1578 for (const auto *Ty : ST->elements())
1579 if (Ty != EltTy)
1580 return 0;
1581 }
1582 return N;
1583}
1584
1585bool BoUpSLP::canReuseExtract(ArrayRef<Value *> VL, unsigned Opcode) const {
1586 assert(Opcode == Instruction::ExtractElement ||((Opcode == Instruction::ExtractElement || Opcode == Instruction
::ExtractValue) ? static_cast<void> (0) : __assert_fail
("Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1587, __PRETTY_FUNCTION__))
1587 Opcode == Instruction::ExtractValue)((Opcode == Instruction::ExtractElement || Opcode == Instruction
::ExtractValue) ? static_cast<void> (0) : __assert_fail
("Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1587, __PRETTY_FUNCTION__))
;
1588 assert(Opcode == getSameOpcode(VL) && "Invalid opcode")((Opcode == getSameOpcode(VL) && "Invalid opcode") ? static_cast
<void> (0) : __assert_fail ("Opcode == getSameOpcode(VL) && \"Invalid opcode\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1588, __PRETTY_FUNCTION__))
;
1589 // Check if all of the extracts come from the same vector and from the
1590 // correct offset.
1591 Value *VL0 = VL[0];
1592 Instruction *E0 = cast<Instruction>(VL0);
1593 Value *Vec = E0->getOperand(0);
1594
1595 // We have to extract from a vector/aggregate with the same number of elements.
1596 unsigned NElts;
1597 if (Opcode == Instruction::ExtractValue) {
1598 const DataLayout &DL = E0->getModule()->getDataLayout();
1599 NElts = canMapToVector(Vec->getType(), DL);
1600 if (!NElts)
1601 return false;
1602 // Check if load can be rewritten as load of vector.
1603 LoadInst *LI = dyn_cast<LoadInst>(Vec);
1604 if (!LI || !LI->isSimple() || !LI->hasNUses(VL.size()))
1605 return false;
1606 } else {
1607 NElts = Vec->getType()->getVectorNumElements();
1608 }
1609
1610 if (NElts != VL.size())
1611 return false;
1612
1613 // Check that all of the indices extract from the correct offset.
1614 if (!matchExtractIndex(E0, 0, Opcode))
1615 return false;
1616
1617 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
1618 Instruction *E = cast<Instruction>(VL[i]);
1619 if (!matchExtractIndex(E, i, Opcode))
1620 return false;
1621 if (E->getOperand(0) != Vec)
1622 return false;
1623 }
1624
1625 return true;
1626}
1627
1628int BoUpSLP::getEntryCost(TreeEntry *E) {
1629 ArrayRef<Value*> VL = E->Scalars;
1630
1631 Type *ScalarTy = VL[0]->getType();
1632 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1633 ScalarTy = SI->getValueOperand()->getType();
1634 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1635
1636 // If we have computed a smaller type for the expression, update VecTy so
1637 // that the costs will be accurate.
1638 if (MinBWs.count(VL[0]))
1639 VecTy = VectorType::get(
1640 IntegerType::get(F->getContext(), MinBWs[VL[0]].first), VL.size());
1641
1642 if (E->NeedToGather) {
1643 if (allConstant(VL))
1644 return 0;
1645 if (isSplat(VL)) {
1646 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
1647 }
1648 return getGatherCost(E->Scalars);
1649 }
1650 unsigned Opcode = getSameOpcode(VL);
1651 assert(Opcode && allSameType(VL) && allSameBlock(VL) && "Invalid VL")((Opcode && allSameType(VL) && allSameBlock(VL
) && "Invalid VL") ? static_cast<void> (0) : __assert_fail
("Opcode && allSameType(VL) && allSameBlock(VL) && \"Invalid VL\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1651, __PRETTY_FUNCTION__))
;
1652 Instruction *VL0 = cast<Instruction>(VL[0]);
1653 switch (Opcode) {
1654 case Instruction::PHI: {
1655 return 0;
1656 }
1657 case Instruction::ExtractValue:
1658 case Instruction::ExtractElement: {
1659 if (canReuseExtract(VL, Opcode)) {
1660 int DeadCost = 0;
1661 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
1662 Instruction *E = cast<Instruction>(VL[i]);
1663 // If all users are going to be vectorized, instruction can be
1664 // considered as dead.
1665 // The same, if have only one user, it will be vectorized for sure.
1666 if (E->hasOneUse() ||
1667 std::all_of(E->user_begin(), E->user_end(), [this](User *U) {
1668 return ScalarToTreeEntry.count(U) > 0;
1669 }))
1670 // Take credit for instruction that will become dead.
1671 DeadCost +=
1672 TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, i);
1673 }
1674 return -DeadCost;
1675 }
1676 return getGatherCost(VecTy);
1677 }
1678 case Instruction::ZExt:
1679 case Instruction::SExt:
1680 case Instruction::FPToUI:
1681 case Instruction::FPToSI:
1682 case Instruction::FPExt:
1683 case Instruction::PtrToInt:
1684 case Instruction::IntToPtr:
1685 case Instruction::SIToFP:
1686 case Instruction::UIToFP:
1687 case Instruction::Trunc:
1688 case Instruction::FPTrunc:
1689 case Instruction::BitCast: {
1690 Type *SrcTy = VL0->getOperand(0)->getType();
1691
1692 // Calculate the cost of this instruction.
1693 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
1694 VL0->getType(), SrcTy);
1695
1696 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
1697 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
1698 return VecCost - ScalarCost;
1699 }
1700 case Instruction::FCmp:
1701 case Instruction::ICmp:
1702 case Instruction::Select: {
1703 // Calculate the cost of this instruction.
1704 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
1705 int ScalarCost = VecTy->getNumElements() *
1706 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
1707 int VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
1708 return VecCost - ScalarCost;
1709 }
1710 case Instruction::Add:
1711 case Instruction::FAdd:
1712 case Instruction::Sub:
1713 case Instruction::FSub:
1714 case Instruction::Mul:
1715 case Instruction::FMul:
1716 case Instruction::UDiv:
1717 case Instruction::SDiv:
1718 case Instruction::FDiv:
1719 case Instruction::URem:
1720 case Instruction::SRem:
1721 case Instruction::FRem:
1722 case Instruction::Shl:
1723 case Instruction::LShr:
1724 case Instruction::AShr:
1725 case Instruction::And:
1726 case Instruction::Or:
1727 case Instruction::Xor: {
1728 // Certain instructions can be cheaper to vectorize if they have a
1729 // constant second vector operand.
1730 TargetTransformInfo::OperandValueKind Op1VK =
1731 TargetTransformInfo::OK_AnyValue;
1732 TargetTransformInfo::OperandValueKind Op2VK =
1733 TargetTransformInfo::OK_UniformConstantValue;
1734 TargetTransformInfo::OperandValueProperties Op1VP =
1735 TargetTransformInfo::OP_None;
1736 TargetTransformInfo::OperandValueProperties Op2VP =
1737 TargetTransformInfo::OP_None;
1738
1739 // If all operands are exactly the same ConstantInt then set the
1740 // operand kind to OK_UniformConstantValue.
1741 // If instead not all operands are constants, then set the operand kind
1742 // to OK_AnyValue. If all operands are constants but not the same,
1743 // then set the operand kind to OK_NonUniformConstantValue.
1744 ConstantInt *CInt = nullptr;
1745 for (unsigned i = 0; i < VL.size(); ++i) {
1746 const Instruction *I = cast<Instruction>(VL[i]);
1747 if (!isa<ConstantInt>(I->getOperand(1))) {
1748 Op2VK = TargetTransformInfo::OK_AnyValue;
1749 break;
1750 }
1751 if (i == 0) {
1752 CInt = cast<ConstantInt>(I->getOperand(1));
1753 continue;
1754 }
1755 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue &&
1756 CInt != cast<ConstantInt>(I->getOperand(1)))
1757 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1758 }
1759 // FIXME: Currently cost of model modification for division by power of
1760 // 2 is handled for X86 and AArch64. Add support for other targets.
1761 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue && CInt &&
1762 CInt->getValue().isPowerOf2())
1763 Op2VP = TargetTransformInfo::OP_PowerOf2;
1764
1765 int ScalarCost = VecTy->getNumElements() *
1766 TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK,
1767 Op2VK, Op1VP, Op2VP);
1768 int VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK,
1769 Op1VP, Op2VP);
1770 return VecCost - ScalarCost;
1771 }
1772 case Instruction::GetElementPtr: {
1773 TargetTransformInfo::OperandValueKind Op1VK =
1774 TargetTransformInfo::OK_AnyValue;
1775 TargetTransformInfo::OperandValueKind Op2VK =
1776 TargetTransformInfo::OK_UniformConstantValue;
1777
1778 int ScalarCost =
1779 VecTy->getNumElements() *
1780 TTI->getArithmeticInstrCost(Instruction::Add, ScalarTy, Op1VK, Op2VK);
1781 int VecCost =
1782 TTI->getArithmeticInstrCost(Instruction::Add, VecTy, Op1VK, Op2VK);
1783
1784 return VecCost - ScalarCost;
1785 }
1786 case Instruction::Load: {
1787 // Cost of wide load - cost of scalar loads.
1788 unsigned alignment = dyn_cast<LoadInst>(VL0)->getAlignment();
1789 int ScalarLdCost = VecTy->getNumElements() *
1790 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, alignment, 0);
1791 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load,
1792 VecTy, alignment, 0);
1793 if (E->NeedToShuffle) {
1794 VecLdCost += TTI->getShuffleCost(
1795 TargetTransformInfo::SK_PermuteSingleSrc, VecTy, 0);
1796 }
1797 return VecLdCost - ScalarLdCost;
1798 }
1799 case Instruction::Store: {
1800 // We know that we can merge the stores. Calculate the cost.
1801 unsigned alignment = dyn_cast<StoreInst>(VL0)->getAlignment();
1802 int ScalarStCost = VecTy->getNumElements() *
1803 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, alignment, 0);
1804 int VecStCost = TTI->getMemoryOpCost(Instruction::Store,
1805 VecTy, alignment, 0);
1806 return VecStCost - ScalarStCost;
1807 }
1808 case Instruction::Call: {
1809 CallInst *CI = cast<CallInst>(VL0);
1810 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
1811
1812 // Calculate the cost of the scalar and vector calls.
1813 SmallVector<Type*, 4> ScalarTys, VecTys;
1814 for (unsigned op = 0, opc = CI->getNumArgOperands(); op!= opc; ++op) {
1815 ScalarTys.push_back(CI->getArgOperand(op)->getType());
1816 VecTys.push_back(VectorType::get(CI->getArgOperand(op)->getType(),
1817 VecTy->getNumElements()));
1818 }
1819
1820 FastMathFlags FMF;
1821 if (auto *FPMO = dyn_cast<FPMathOperator>(CI))
1822 FMF = FPMO->getFastMathFlags();
1823
1824 int ScalarCallCost = VecTy->getNumElements() *
1825 TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys, FMF);
1826
1827 int VecCallCost = TTI->getIntrinsicInstrCost(ID, VecTy, VecTys, FMF);
1828
1829 DEBUG(dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Call cost "<< VecCallCost
- ScalarCallCost << " (" << VecCallCost <<
"-" << ScalarCallCost << ")" << " for " <<
*CI << "\n"; } } while (false)
1830 << " (" << VecCallCost << "-" << ScalarCallCost << ")"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Call cost "<< VecCallCost
- ScalarCallCost << " (" << VecCallCost <<
"-" << ScalarCallCost << ")" << " for " <<
*CI << "\n"; } } while (false)
1831 << " for " << *CI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Call cost "<< VecCallCost
- ScalarCallCost << " (" << VecCallCost <<
"-" << ScalarCallCost << ")" << " for " <<
*CI << "\n"; } } while (false)
;
1832
1833 return VecCallCost - ScalarCallCost;
1834 }
1835 case Instruction::ShuffleVector: {
1836 TargetTransformInfo::OperandValueKind Op1VK =
1837 TargetTransformInfo::OK_AnyValue;
1838 TargetTransformInfo::OperandValueKind Op2VK =
1839 TargetTransformInfo::OK_AnyValue;
1840 int ScalarCost = 0;
1841 int VecCost = 0;
1842 for (Value *i : VL) {
1843 Instruction *I = cast<Instruction>(i);
1844 if (!I)
1845 break;
1846 ScalarCost +=
1847 TTI->getArithmeticInstrCost(I->getOpcode(), ScalarTy, Op1VK, Op2VK);
1848 }
1849 // VecCost is equal to sum of the cost of creating 2 vectors
1850 // and the cost of creating shuffle.
1851 Instruction *I0 = cast<Instruction>(VL[0]);
1852 VecCost =
1853 TTI->getArithmeticInstrCost(I0->getOpcode(), VecTy, Op1VK, Op2VK);
1854 Instruction *I1 = cast<Instruction>(VL[1]);
1855 VecCost +=
1856 TTI->getArithmeticInstrCost(I1->getOpcode(), VecTy, Op1VK, Op2VK);
1857 VecCost +=
1858 TTI->getShuffleCost(TargetTransformInfo::SK_Alternate, VecTy, 0);
1859 return VecCost - ScalarCost;
1860 }
1861 default:
1862 llvm_unreachable("Unknown instruction")::llvm::llvm_unreachable_internal("Unknown instruction", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1862)
;
1863 }
1864}
1865
1866bool BoUpSLP::isFullyVectorizableTinyTree() {
1867 DEBUG(dbgs() << "SLP: Check whether the tree with height " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Check whether the tree with height "
<< VectorizableTree.size() << " is fully vectorizable .\n"
; } } while (false)
1868 VectorizableTree.size() << " is fully vectorizable .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Check whether the tree with height "
<< VectorizableTree.size() << " is fully vectorizable .\n"
; } } while (false)
;
1869
1870 // We only handle trees of heights 1 and 2.
1871 if (VectorizableTree.size() == 1 && !VectorizableTree[0].NeedToGather)
1872 return true;
1873
1874 if (VectorizableTree.size() != 2)
1875 return false;
1876
1877 // Handle splat and all-constants stores.
1878 if (!VectorizableTree[0].NeedToGather &&
1879 (allConstant(VectorizableTree[1].Scalars) ||
1880 isSplat(VectorizableTree[1].Scalars)))
1881 return true;
1882
1883 // Gathering cost would be too much for tiny trees.
1884 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
1885 return false;
1886
1887 return true;
1888}
1889
1890bool BoUpSLP::isTreeTinyAndNotFullyVectorizable() {
1891
1892 // We can vectorize the tree if its size is greater than or equal to the
1893 // minimum size specified by the MinTreeSize command line option.
1894 if (VectorizableTree.size() >= MinTreeSize)
1895 return false;
1896
1897 // If we have a tiny tree (a tree whose size is less than MinTreeSize), we
1898 // can vectorize it if we can prove it fully vectorizable.
1899 if (isFullyVectorizableTinyTree())
1900 return false;
1901
1902 assert(VectorizableTree.empty()((VectorizableTree.empty() ? ExternalUses.empty() : true &&
"We shouldn't have any external users") ? static_cast<void
> (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1904, __PRETTY_FUNCTION__))
1903 ? ExternalUses.empty()((VectorizableTree.empty() ? ExternalUses.empty() : true &&
"We shouldn't have any external users") ? static_cast<void
> (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1904, __PRETTY_FUNCTION__))
1904 : true && "We shouldn't have any external users")((VectorizableTree.empty() ? ExternalUses.empty() : true &&
"We shouldn't have any external users") ? static_cast<void
> (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1904, __PRETTY_FUNCTION__))
;
1905
1906 // Otherwise, we can't vectorize the tree. It is both tiny and not fully
1907 // vectorizable.
1908 return true;
1909}
1910
1911int BoUpSLP::getSpillCost() {
1912 // Walk from the bottom of the tree to the top, tracking which values are
1913 // live. When we see a call instruction that is not part of our tree,
1914 // query TTI to see if there is a cost to keeping values live over it
1915 // (for example, if spills and fills are required).
1916 unsigned BundleWidth = VectorizableTree.front().Scalars.size();
1917 int Cost = 0;
1918
1919 SmallPtrSet<Instruction*, 4> LiveValues;
1920 Instruction *PrevInst = nullptr;
1921
1922 for (const auto &N : VectorizableTree) {
1923 Instruction *Inst = dyn_cast<Instruction>(N.Scalars[0]);
1924 if (!Inst)
1925 continue;
1926
1927 if (!PrevInst) {
1928 PrevInst = Inst;
1929 continue;
1930 }
1931
1932 // Update LiveValues.
1933 LiveValues.erase(PrevInst);
1934 for (auto &J : PrevInst->operands()) {
1935 if (isa<Instruction>(&*J) && ScalarToTreeEntry.count(&*J))
1936 LiveValues.insert(cast<Instruction>(&*J));
1937 }
1938
1939 DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size
(); for (auto *X : LiveValues) dbgs() << " " << X
->getName(); dbgs() << ", Looking at "; Inst->dump
();; } } while (false)
1940 dbgs() << "SLP: #LV: " << LiveValues.size();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size
(); for (auto *X : LiveValues) dbgs() << " " << X
->getName(); dbgs() << ", Looking at "; Inst->dump
();; } } while (false)
1941 for (auto *X : LiveValues)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size
(); for (auto *X : LiveValues) dbgs() << " " << X
->getName(); dbgs() << ", Looking at "; Inst->dump
();; } } while (false)
1942 dbgs() << " " << X->getName();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size
(); for (auto *X : LiveValues) dbgs() << " " << X
->getName(); dbgs() << ", Looking at "; Inst->dump
();; } } while (false)
1943 dbgs() << ", Looking at ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size
(); for (auto *X : LiveValues) dbgs() << " " << X
->getName(); dbgs() << ", Looking at "; Inst->dump
();; } } while (false)
1944 Inst->dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size
(); for (auto *X : LiveValues) dbgs() << " " << X
->getName(); dbgs() << ", Looking at "; Inst->dump
();; } } while (false)
1945 )do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: #LV: " << LiveValues.size
(); for (auto *X : LiveValues) dbgs() << " " << X
->getName(); dbgs() << ", Looking at "; Inst->dump
();; } } while (false)
;
1946
1947 // Now find the sequence of instructions between PrevInst and Inst.
1948 BasicBlock::reverse_iterator InstIt = ++Inst->getIterator().getReverse(),
1949 PrevInstIt =
1950 PrevInst->getIterator().getReverse();
1951 while (InstIt != PrevInstIt) {
1952 if (PrevInstIt == PrevInst->getParent()->rend()) {
1953 PrevInstIt = Inst->getParent()->rbegin();
1954 continue;
1955 }
1956
1957 if (isa<CallInst>(&*PrevInstIt) && &*PrevInstIt != PrevInst) {
1958 SmallVector<Type*, 4> V;
1959 for (auto *II : LiveValues)
1960 V.push_back(VectorType::get(II->getType(), BundleWidth));
1961 Cost += TTI->getCostOfKeepingLiveOverCall(V);
1962 }
1963
1964 ++PrevInstIt;
1965 }
1966
1967 PrevInst = Inst;
1968 }
1969
1970 return Cost;
1971}
1972
1973int BoUpSLP::getTreeCost() {
1974 int Cost = 0;
1975 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Calculating cost for tree of size "
<< VectorizableTree.size() << ".\n"; } } while (
false)
1976 VectorizableTree.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Calculating cost for tree of size "
<< VectorizableTree.size() << ".\n"; } } while (
false)
;
1977
1978 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
1979
1980 for (TreeEntry &TE : VectorizableTree) {
1981 int C = getEntryCost(&TE);
1982 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Adding cost " << C <<
" for bundle that starts with " << *TE.Scalars[0] <<
".\n"; } } while (false)
1983 << *TE.Scalars[0] << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Adding cost " << C <<
" for bundle that starts with " << *TE.Scalars[0] <<
".\n"; } } while (false)
;
1984 Cost += C;
1985 }
1986
1987 SmallSet<Value *, 16> ExtractCostCalculated;
1988 int ExtractCost = 0;
1989 for (ExternalUser &EU : ExternalUses) {
1990 // We only add extract cost once for the same scalar.
1991 if (!ExtractCostCalculated.insert(EU.Scalar).second)
1992 continue;
1993
1994 // Uses by ephemeral values are free (because the ephemeral value will be
1995 // removed prior to code generation, and so the extraction will be
1996 // removed as well).
1997 if (EphValues.count(EU.User))
1998 continue;
1999
2000 // If we plan to rewrite the tree in a smaller type, we will need to sign
2001 // extend the extracted value back to the original type. Here, we account
2002 // for the extract and the added cost of the sign extend if needed.
2003 auto *VecTy = VectorType::get(EU.Scalar->getType(), BundleWidth);
2004 auto *ScalarRoot = VectorizableTree[0].Scalars[0];
2005 if (MinBWs.count(ScalarRoot)) {
2006 auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first);
2007 auto Extend =
2008 MinBWs[ScalarRoot].second ? Instruction::SExt : Instruction::ZExt;
2009 VecTy = VectorType::get(MinTy, BundleWidth);
2010 ExtractCost += TTI->getExtractWithExtendCost(Extend, EU.Scalar->getType(),
2011 VecTy, EU.Lane);
2012 } else {
2013 ExtractCost +=
2014 TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, EU.Lane);
2015 }
2016 }
2017
2018 int SpillCost = getSpillCost();
2019 Cost += SpillCost + ExtractCost;
2020
2021 DEBUG(dbgs() << "SLP: Spill Cost = " << SpillCost << ".\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Spill Cost = " << SpillCost
<< ".\n" << "SLP: Extract Cost = " << ExtractCost
<< ".\n" << "SLP: Total Cost = " << Cost <<
".\n"; } } while (false)
2022 << "SLP: Extract Cost = " << ExtractCost << ".\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Spill Cost = " << SpillCost
<< ".\n" << "SLP: Extract Cost = " << ExtractCost
<< ".\n" << "SLP: Total Cost = " << Cost <<
".\n"; } } while (false)
2023 << "SLP: Total Cost = " << Cost << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Spill Cost = " << SpillCost
<< ".\n" << "SLP: Extract Cost = " << ExtractCost
<< ".\n" << "SLP: Total Cost = " << Cost <<
".\n"; } } while (false)
;
2024 return Cost;
2025}
2026
2027int BoUpSLP::getGatherCost(Type *Ty) {
2028 int Cost = 0;
2029 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
2030 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
2031 return Cost;
2032}
2033
2034int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
2035 // Find the type of the operands in VL.
2036 Type *ScalarTy = VL[0]->getType();
2037 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
2038 ScalarTy = SI->getValueOperand()->getType();
2039 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
2040 // Find the cost of inserting/extracting values from the vector.
2041 return getGatherCost(VecTy);
2042}
2043
2044// Reorder commutative operations in alternate shuffle if the resulting vectors
2045// are consecutive loads. This would allow us to vectorize the tree.
2046// If we have something like-
2047// load a[0] - load b[0]
2048// load b[1] + load a[1]
2049// load a[2] - load b[2]
2050// load a[3] + load b[3]
2051// Reordering the second load b[1] load a[1] would allow us to vectorize this
2052// code.
2053void BoUpSLP::reorderAltShuffleOperands(ArrayRef<Value *> VL,
2054 SmallVectorImpl<Value *> &Left,
2055 SmallVectorImpl<Value *> &Right) {
2056 // Push left and right operands of binary operation into Left and Right
2057 for (Value *i : VL) {
2058 Left.push_back(cast<Instruction>(i)->getOperand(0));
2059 Right.push_back(cast<Instruction>(i)->getOperand(1));
2060 }
2061
2062 // Reorder if we have a commutative operation and consecutive access
2063 // are on either side of the alternate instructions.
2064 for (unsigned j = 0; j < VL.size() - 1; ++j) {
2065 if (LoadInst *L = dyn_cast<LoadInst>(Left[j])) {
2066 if (LoadInst *L1 = dyn_cast<LoadInst>(Right[j + 1])) {
2067 Instruction *VL1 = cast<Instruction>(VL[j]);
2068 Instruction *VL2 = cast<Instruction>(VL[j + 1]);
2069 if (VL1->isCommutative() && isConsecutiveAccess(L, L1, *DL, *SE)) {
2070 std::swap(Left[j], Right[j]);
2071 continue;
2072 } else if (VL2->isCommutative() &&
2073 isConsecutiveAccess(L, L1, *DL, *SE)) {
2074 std::swap(Left[j + 1], Right[j + 1]);
2075 continue;
2076 }
2077 // else unchanged
2078 }
2079 }
2080 if (LoadInst *L = dyn_cast<LoadInst>(Right[j])) {
2081 if (LoadInst *L1 = dyn_cast<LoadInst>(Left[j + 1])) {
2082 Instruction *VL1 = cast<Instruction>(VL[j]);
2083 Instruction *VL2 = cast<Instruction>(VL[j + 1]);
2084 if (VL1->isCommutative() && isConsecutiveAccess(L, L1, *DL, *SE)) {
2085 std::swap(Left[j], Right[j]);
2086 continue;
2087 } else if (VL2->isCommutative() &&
2088 isConsecutiveAccess(L, L1, *DL, *SE)) {
2089 std::swap(Left[j + 1], Right[j + 1]);
2090 continue;
2091 }
2092 // else unchanged
2093 }
2094 }
2095 }
2096}
2097
2098// Return true if I should be commuted before adding it's left and right
2099// operands to the arrays Left and Right.
2100//
2101// The vectorizer is trying to either have all elements one side being
2102// instruction with the same opcode to enable further vectorization, or having
2103// a splat to lower the vectorizing cost.
2104static bool shouldReorderOperands(int i, Instruction &I,
2105 SmallVectorImpl<Value *> &Left,
2106 SmallVectorImpl<Value *> &Right,
2107 bool AllSameOpcodeLeft,
2108 bool AllSameOpcodeRight, bool SplatLeft,
2109 bool SplatRight) {
2110 Value *VLeft = I.getOperand(0);
2111 Value *VRight = I.getOperand(1);
2112 // If we have "SplatRight", try to see if commuting is needed to preserve it.
2113 if (SplatRight) {
2114 if (VRight == Right[i - 1])
2115 // Preserve SplatRight
2116 return false;
2117 if (VLeft == Right[i - 1]) {
2118 // Commuting would preserve SplatRight, but we don't want to break
2119 // SplatLeft either, i.e. preserve the original order if possible.
2120 // (FIXME: why do we care?)
2121 if (SplatLeft && VLeft == Left[i - 1])
2122 return false;
2123 return true;
2124 }
2125 }
2126 // Symmetrically handle Right side.
2127 if (SplatLeft) {
2128 if (VLeft == Left[i - 1])
2129 // Preserve SplatLeft
2130 return false;
2131 if (VRight == Left[i - 1])
2132 return true;
2133 }
2134
2135 Instruction *ILeft = dyn_cast<Instruction>(VLeft);
2136 Instruction *IRight = dyn_cast<Instruction>(VRight);
2137
2138 // If we have "AllSameOpcodeRight", try to see if the left operands preserves
2139 // it and not the right, in this case we want to commute.
2140 if (AllSameOpcodeRight) {
2141 unsigned RightPrevOpcode = cast<Instruction>(Right[i - 1])->getOpcode();
2142 if (IRight && RightPrevOpcode == IRight->getOpcode())
2143 // Do not commute, a match on the right preserves AllSameOpcodeRight
2144 return false;
2145 if (ILeft && RightPrevOpcode == ILeft->getOpcode()) {
2146 // We have a match and may want to commute, but first check if there is
2147 // not also a match on the existing operands on the Left to preserve
2148 // AllSameOpcodeLeft, i.e. preserve the original order if possible.
2149 // (FIXME: why do we care?)
2150 if (AllSameOpcodeLeft && ILeft &&
2151 cast<Instruction>(Left[i - 1])->getOpcode() == ILeft->getOpcode())
2152 return false;
2153 return true;
2154 }
2155 }
2156 // Symmetrically handle Left side.
2157 if (AllSameOpcodeLeft) {
2158 unsigned LeftPrevOpcode = cast<Instruction>(Left[i - 1])->getOpcode();
2159 if (ILeft && LeftPrevOpcode == ILeft->getOpcode())
2160 return false;
2161 if (IRight && LeftPrevOpcode == IRight->getOpcode())
2162 return true;
2163 }
2164 return false;
2165}
2166
2167void BoUpSLP::reorderInputsAccordingToOpcode(ArrayRef<Value *> VL,
2168 SmallVectorImpl<Value *> &Left,
2169 SmallVectorImpl<Value *> &Right) {
2170
2171 if (VL.size()) {
2172 // Peel the first iteration out of the loop since there's nothing
2173 // interesting to do anyway and it simplifies the checks in the loop.
2174 auto VLeft = cast<Instruction>(VL[0])->getOperand(0);
2175 auto VRight = cast<Instruction>(VL[0])->getOperand(1);
2176 if (!isa<Instruction>(VRight) && isa<Instruction>(VLeft))
2177 // Favor having instruction to the right. FIXME: why?
2178 std::swap(VLeft, VRight);
2179 Left.push_back(VLeft);
2180 Right.push_back(VRight);
2181 }
2182
2183 // Keep track if we have instructions with all the same opcode on one side.
2184 bool AllSameOpcodeLeft = isa<Instruction>(Left[0]);
2185 bool AllSameOpcodeRight = isa<Instruction>(Right[0]);
2186 // Keep track if we have one side with all the same value (broadcast).
2187 bool SplatLeft = true;
2188 bool SplatRight = true;
2189
2190 for (unsigned i = 1, e = VL.size(); i != e; ++i) {
2191 Instruction *I = cast<Instruction>(VL[i]);
2192 assert(I->isCommutative() && "Can only process commutative instruction")((I->isCommutative() && "Can only process commutative instruction"
) ? static_cast<void> (0) : __assert_fail ("I->isCommutative() && \"Can only process commutative instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2192, __PRETTY_FUNCTION__))
;
2193 // Commute to favor either a splat or maximizing having the same opcodes on
2194 // one side.
2195 if (shouldReorderOperands(i, *I, Left, Right, AllSameOpcodeLeft,
2196 AllSameOpcodeRight, SplatLeft, SplatRight)) {
2197 Left.push_back(I->getOperand(1));
2198 Right.push_back(I->getOperand(0));
2199 } else {
2200 Left.push_back(I->getOperand(0));
2201 Right.push_back(I->getOperand(1));
2202 }
2203 // Update Splat* and AllSameOpcode* after the insertion.
2204 SplatRight = SplatRight && (Right[i - 1] == Right[i]);
2205 SplatLeft = SplatLeft && (Left[i - 1] == Left[i]);
2206 AllSameOpcodeLeft = AllSameOpcodeLeft && isa<Instruction>(Left[i]) &&
2207 (cast<Instruction>(Left[i - 1])->getOpcode() ==
2208 cast<Instruction>(Left[i])->getOpcode());
2209 AllSameOpcodeRight = AllSameOpcodeRight && isa<Instruction>(Right[i]) &&
2210 (cast<Instruction>(Right[i - 1])->getOpcode() ==
2211 cast<Instruction>(Right[i])->getOpcode());
2212 }
2213
2214 // If one operand end up being broadcast, return this operand order.
2215 if (SplatRight || SplatLeft)
2216 return;
2217
2218 // Finally check if we can get longer vectorizable chain by reordering
2219 // without breaking the good operand order detected above.
2220 // E.g. If we have something like-
2221 // load a[0] load b[0]
2222 // load b[1] load a[1]
2223 // load a[2] load b[2]
2224 // load a[3] load b[3]
2225 // Reordering the second load b[1] load a[1] would allow us to vectorize
2226 // this code and we still retain AllSameOpcode property.
2227 // FIXME: This load reordering might break AllSameOpcode in some rare cases
2228 // such as-
2229 // add a[0],c[0] load b[0]
2230 // add a[1],c[2] load b[1]
2231 // b[2] load b[2]
2232 // add a[3],c[3] load b[3]
2233 for (unsigned j = 0; j < VL.size() - 1; ++j) {
2234 if (LoadInst *L = dyn_cast<LoadInst>(Left[j])) {
2235 if (LoadInst *L1 = dyn_cast<LoadInst>(Right[j + 1])) {
2236 if (isConsecutiveAccess(L, L1, *DL, *SE)) {
2237 std::swap(Left[j + 1], Right[j + 1]);
2238 continue;
2239 }
2240 }
2241 }
2242 if (LoadInst *L = dyn_cast<LoadInst>(Right[j])) {
2243 if (LoadInst *L1 = dyn_cast<LoadInst>(Left[j + 1])) {
2244 if (isConsecutiveAccess(L, L1, *DL, *SE)) {
2245 std::swap(Left[j + 1], Right[j + 1]);
2246 continue;
2247 }
2248 }
2249 }
2250 // else unchanged
2251 }
2252}
2253
2254void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
2255
2256 // Get the basic block this bundle is in. All instructions in the bundle
2257 // should be in this block.
2258 auto *Front = cast<Instruction>(VL.front());
2259 auto *BB = Front->getParent();
2260 assert(all_of(make_range(VL.begin(), VL.end()), [&](Value *V) -> bool {((all_of(make_range(VL.begin(), VL.end()), [&](Value *V) ->
bool { return cast<Instruction>(V)->getParent() == BB
; })) ? static_cast<void> (0) : __assert_fail ("all_of(make_range(VL.begin(), VL.end()), [&](Value *V) -> bool { return cast<Instruction>(V)->getParent() == BB; })"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2262, __PRETTY_FUNCTION__))
2261 return cast<Instruction>(V)->getParent() == BB;((all_of(make_range(VL.begin(), VL.end()), [&](Value *V) ->
bool { return cast<Instruction>(V)->getParent() == BB
; })) ? static_cast<void> (0) : __assert_fail ("all_of(make_range(VL.begin(), VL.end()), [&](Value *V) -> bool { return cast<Instruction>(V)->getParent() == BB; })"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2262, __PRETTY_FUNCTION__))
2262 }))((all_of(make_range(VL.begin(), VL.end()), [&](Value *V) ->
bool { return cast<Instruction>(V)->getParent() == BB
; })) ? static_cast<void> (0) : __assert_fail ("all_of(make_range(VL.begin(), VL.end()), [&](Value *V) -> bool { return cast<Instruction>(V)->getParent() == BB; })"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2262, __PRETTY_FUNCTION__))
;
2263
2264 // The last instruction in the bundle in program order.
2265 Instruction *LastInst = nullptr;
7
'LastInst' initialized to a null pointer value
2266
2267 // Find the last instruction. The common case should be that BB has been
2268 // scheduled, and the last instruction is VL.back(). So we start with
2269 // VL.back() and iterate over schedule data until we reach the end of the
2270 // bundle. The end of the bundle is marked by null ScheduleData.
2271 if (BlocksSchedules.count(BB)) {
8
Assuming the condition is false
9
Taking false branch
2272 auto *Bundle = BlocksSchedules[BB]->getScheduleData(VL.back());
2273 if (Bundle && Bundle->isPartOfBundle())
2274 for (; Bundle; Bundle = Bundle->NextInBundle)
2275 LastInst = Bundle->Inst;
2276 }
2277
2278 // LastInst can still be null at this point if there's either not an entry
2279 // for BB in BlocksSchedules or there's no ScheduleData available for
2280 // VL.back(). This can be the case if buildTree_rec aborts for various
2281 // reasons (e.g., the maximum recursion depth is reached, the maximum region
2282 // size is reached, etc.). ScheduleData is initialized in the scheduling
2283 // "dry-run".
2284 //
2285 // If this happens, we can still find the last instruction by brute force. We
2286 // iterate forwards from Front (inclusive) until we either see all
2287 // instructions in the bundle or reach the end of the block. If Front is the
2288 // last instruction in program order, LastInst will be set to Front, and we
2289 // will visit all the remaining instructions in the block.
2290 //
2291 // One of the reasons we exit early from buildTree_rec is to place an upper
2292 // bound on compile-time. Thus, taking an additional compile-time hit here is
2293 // not ideal. However, this should be exceedingly rare since it requires that
2294 // we both exit early from buildTree_rec and that the bundle be out-of-order
2295 // (causing us to iterate all the way to the end of the block).
2296 if (!LastInst) {
10
Taking true branch
2297 SmallPtrSet<Value *, 16> Bundle(VL.begin(), VL.end());
2298 for (auto &I : make_range(BasicBlock::iterator(Front), BB->end())) {
2299 if (Bundle.erase(&I))
2300 LastInst = &I;
2301 if (Bundle.empty())
2302 break;
2303 }
2304 }
2305
2306 // Set the insertion point after the last instruction in the bundle. Set the
2307 // debug location to Front.
2308 Builder.SetInsertPoint(BB, ++LastInst->getIterator());
11
Called C++ object pointer is null
2309 Builder.SetCurrentDebugLocation(Front->getDebugLoc());
2310}
2311
2312Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
2313 Value *Vec = UndefValue::get(Ty);
2314 // Generate the 'InsertElement' instruction.
2315 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
2316 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
2317 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
2318 GatherSeq.insert(Insrt);
2319 CSEBlocks.insert(Insrt->getParent());
2320
2321 // Add to our 'need-to-extract' list.
2322 if (ScalarToTreeEntry.count(VL[i])) {
2323 int Idx = ScalarToTreeEntry[VL[i]];
2324 TreeEntry *E = &VectorizableTree[Idx];
2325 // Find which lane we need to extract.
2326 int FoundLane = -1;
2327 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
2328 // Is this the lane of the scalar that we are looking for ?
2329 if (E->Scalars[Lane] == VL[i]) {
2330 FoundLane = Lane;
2331 break;
2332 }
2333 }
2334 assert(FoundLane >= 0 && "Could not find the correct lane")((FoundLane >= 0 && "Could not find the correct lane"
) ? static_cast<void> (0) : __assert_fail ("FoundLane >= 0 && \"Could not find the correct lane\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2334, __PRETTY_FUNCTION__))
;
2335 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
2336 }
2337 }
2338 }
2339
2340 return Vec;
2341}
2342
2343Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
2344 SmallDenseMap<Value*, int>::const_iterator Entry
2345 = ScalarToTreeEntry.find(VL[0]);
2346 if (Entry != ScalarToTreeEntry.end()) {
2347 int Idx = Entry->second;
2348 const TreeEntry *En = &VectorizableTree[Idx];
2349 if (En->isSame(VL) && En->VectorizedValue)
2350 return En->VectorizedValue;
2351 }
2352 return nullptr;
2353}
2354
2355Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
2356 if (ScalarToTreeEntry.count(VL[0])) {
2357 int Idx = ScalarToTreeEntry[VL[0]];
2358 TreeEntry *E = &VectorizableTree[Idx];
2359 if (E->isSame(VL) || (E->NeedToShuffle && E->isFoundJumbled(VL, *DL, *SE)))
2360 return vectorizeTree(VL, E);
2361 }
2362
2363 Type *ScalarTy = VL[0]->getType();
2364 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
2365 ScalarTy = SI->getValueOperand()->getType();
2366 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
2367
2368 return Gather(VL, VecTy);
2369}
2370
2371Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL, TreeEntry *E) {
2372 IRBuilder<>::InsertPointGuard Guard(Builder);
2373
2374 if (E->VectorizedValue && !E->NeedToShuffle) {
1
Assuming the condition is false
2375 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Diamond merged for " <<
*E->Scalars[0] << ".\n"; } } while (false)
;
2376 return E->VectorizedValue;
2377 }
2378
2379 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
2380 Type *ScalarTy = VL0->getType();
2381 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
2
Taking false branch
2382 ScalarTy = SI->getValueOperand()->getType();
2383 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
2384
2385 if (E->NeedToGather) {
3
Assuming the condition is false
4
Taking false branch
2386 setInsertPointAfterBundle(E->Scalars);
2387 auto *V = Gather(E->Scalars, VecTy);
2388 E->VectorizedValue = V;
2389 return V;
2390 }
2391
2392 unsigned Opcode = getSameOpcode(E->Scalars);
2393
2394 switch (Opcode) {
5
Control jumps to 'case ShuffleVector:' at line 2748
2395 case Instruction::PHI: {
2396 PHINode *PH = dyn_cast<PHINode>(VL0);
2397 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI());
2398 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
2399 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
2400 E->VectorizedValue = NewPhi;
2401
2402 // PHINodes may have multiple entries from the same block. We want to
2403 // visit every block once.
2404 SmallSet<BasicBlock*, 4> VisitedBBs;
2405
2406 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
2407 ValueList Operands;
2408 BasicBlock *IBB = PH->getIncomingBlock(i);
2409
2410 if (!VisitedBBs.insert(IBB).second) {
2411 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
2412 continue;
2413 }
2414
2415 // Prepare the operand vector.
2416 for (Value *V : E->Scalars)
2417 Operands.push_back(cast<PHINode>(V)->getIncomingValueForBlock(IBB));
2418
2419 Builder.SetInsertPoint(IBB->getTerminator());
2420 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
2421 Value *Vec = vectorizeTree(Operands);
2422 NewPhi->addIncoming(Vec, IBB);
2423 }
2424
2425 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&((NewPhi->getNumIncomingValues() == PH->getNumIncomingValues
() && "Invalid number of incoming values") ? static_cast
<void> (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2426, __PRETTY_FUNCTION__))
2426 "Invalid number of incoming values")((NewPhi->getNumIncomingValues() == PH->getNumIncomingValues
() && "Invalid number of incoming values") ? static_cast
<void> (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2426, __PRETTY_FUNCTION__))
;
2427 return NewPhi;
2428 }
2429
2430 case Instruction::ExtractElement: {
2431 if (canReuseExtract(E->Scalars, Instruction::ExtractElement)) {
2432 Value *V = VL0->getOperand(0);
2433 E->VectorizedValue = V;
2434 return V;
2435 }
2436 setInsertPointAfterBundle(E->Scalars);
2437 auto *V = Gather(E->Scalars, VecTy);
2438 E->VectorizedValue = V;
2439 return V;
2440 }
2441 case Instruction::ExtractValue: {
2442 if (canReuseExtract(E->Scalars, Instruction::ExtractValue)) {
2443 LoadInst *LI = cast<LoadInst>(VL0->getOperand(0));
2444 Builder.SetInsertPoint(LI);
2445 PointerType *PtrTy = PointerType::get(VecTy, LI->getPointerAddressSpace());
2446 Value *Ptr = Builder.CreateBitCast(LI->getOperand(0), PtrTy);
2447 LoadInst *V = Builder.CreateAlignedLoad(Ptr, LI->getAlignment());
2448 E->VectorizedValue = V;
2449 return propagateMetadata(V, E->Scalars);
2450 }
2451 setInsertPointAfterBundle(E->Scalars);
2452 auto *V = Gather(E->Scalars, VecTy);
2453 E->VectorizedValue = V;
2454 return V;
2455 }
2456 case Instruction::ZExt:
2457 case Instruction::SExt:
2458 case Instruction::FPToUI:
2459 case Instruction::FPToSI:
2460 case Instruction::FPExt:
2461 case Instruction::PtrToInt:
2462 case Instruction::IntToPtr:
2463 case Instruction::SIToFP:
2464 case Instruction::UIToFP:
2465 case Instruction::Trunc:
2466 case Instruction::FPTrunc:
2467 case Instruction::BitCast: {
2468 ValueList INVL;
2469 for (Value *V : E->Scalars)
2470 INVL.push_back(cast<Instruction>(V)->getOperand(0));
2471
2472 setInsertPointAfterBundle(E->Scalars);
2473
2474 Value *InVec = vectorizeTree(INVL);
2475
2476 if (Value *V = alreadyVectorized(E->Scalars))
2477 return V;
2478
2479 CastInst *CI = dyn_cast<CastInst>(VL0);
2480 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
2481 E->VectorizedValue = V;
2482 ++NumVectorInstructions;
2483 return V;
2484 }
2485 case Instruction::FCmp:
2486 case Instruction::ICmp: {
2487 ValueList LHSV, RHSV;
2488 for (Value *V : E->Scalars) {
2489 LHSV.push_back(cast<Instruction>(V)->getOperand(0));
2490 RHSV.push_back(cast<Instruction>(V)->getOperand(1));
2491 }
2492
2493 setInsertPointAfterBundle(E->Scalars);
2494
2495 Value *L = vectorizeTree(LHSV);
2496 Value *R = vectorizeTree(RHSV);
2497
2498 if (Value *V = alreadyVectorized(E->Scalars))
2499 return V;
2500
2501 CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate();
2502 Value *V;
2503 if (Opcode == Instruction::FCmp)
2504 V = Builder.CreateFCmp(P0, L, R);
2505 else
2506 V = Builder.CreateICmp(P0, L, R);
2507
2508 E->VectorizedValue = V;
2509 propagateIRFlags(E->VectorizedValue, E->Scalars);
2510 ++NumVectorInstructions;
2511 return V;
2512 }
2513 case Instruction::Select: {
2514 ValueList TrueVec, FalseVec, CondVec;
2515 for (Value *V : E->Scalars) {
2516 CondVec.push_back(cast<Instruction>(V)->getOperand(0));
2517 TrueVec.push_back(cast<Instruction>(V)->getOperand(1));
2518 FalseVec.push_back(cast<Instruction>(V)->getOperand(2));
2519 }
2520
2521 setInsertPointAfterBundle(E->Scalars);
2522
2523 Value *Cond = vectorizeTree(CondVec);
2524 Value *True = vectorizeTree(TrueVec);
2525 Value *False = vectorizeTree(FalseVec);
2526
2527 if (Value *V = alreadyVectorized(E->Scalars))
2528 return V;
2529
2530 Value *V = Builder.CreateSelect(Cond, True, False);
2531 E->VectorizedValue = V;
2532 ++NumVectorInstructions;
2533 return V;
2534 }
2535 case Instruction::Add:
2536 case Instruction::FAdd:
2537 case Instruction::Sub:
2538 case Instruction::FSub:
2539 case Instruction::Mul:
2540 case Instruction::FMul:
2541 case Instruction::UDiv:
2542 case Instruction::SDiv:
2543 case Instruction::FDiv:
2544 case Instruction::URem:
2545 case Instruction::SRem:
2546 case Instruction::FRem:
2547 case Instruction::Shl:
2548 case Instruction::LShr:
2549 case Instruction::AShr:
2550 case Instruction::And:
2551 case Instruction::Or:
2552 case Instruction::Xor: {
2553 ValueList LHSVL, RHSVL;
2554 if (isa<BinaryOperator>(VL0) && VL0->isCommutative())
2555 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL);
2556 else
2557 for (Value *V : E->Scalars) {
2558 LHSVL.push_back(cast<Instruction>(V)->getOperand(0));
2559 RHSVL.push_back(cast<Instruction>(V)->getOperand(1));
2560 }
2561
2562 setInsertPointAfterBundle(E->Scalars);
2563
2564 Value *LHS = vectorizeTree(LHSVL);
2565 Value *RHS = vectorizeTree(RHSVL);
2566
2567 if (Value *V = alreadyVectorized(E->Scalars))
2568 return V;
2569
2570 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
2571 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
2572 E->VectorizedValue = V;
2573 propagateIRFlags(E->VectorizedValue, E->Scalars);
2574 ++NumVectorInstructions;
2575
2576 if (Instruction *I = dyn_cast<Instruction>(V))
2577 return propagateMetadata(I, E->Scalars);
2578
2579 return V;
2580 }
2581 case Instruction::Load: {
2582 // Loads are inserted at the head of the tree because we don't want to
2583 // sink them all the way down past store instructions.
2584 setInsertPointAfterBundle(E->Scalars);
2585
2586 LoadInst *LI = cast<LoadInst>(VL0);
2587 Type *ScalarLoadTy = LI->getType();
2588 unsigned AS = LI->getPointerAddressSpace();
2589
2590 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
2591 VecTy->getPointerTo(AS));
2592
2593 // The pointer operand uses an in-tree scalar so we add the new BitCast to
2594 // ExternalUses list to make sure that an extract will be generated in the
2595 // future.
2596 if (ScalarToTreeEntry.count(LI->getPointerOperand()))
2597 ExternalUses.push_back(
2598 ExternalUser(LI->getPointerOperand(), cast<User>(VecPtr), 0));
2599
2600 unsigned Alignment = LI->getAlignment();
2601 LI = Builder.CreateLoad(VecPtr);
2602 if (!Alignment) {
2603 Alignment = DL->getABITypeAlignment(ScalarLoadTy);
2604 }
2605 LI->setAlignment(Alignment);
2606 E->VectorizedValue = LI;
2607 ++NumVectorInstructions;
2608 propagateMetadata(LI, E->Scalars);
2609
2610 // As program order of scalar loads are jumbled, the vectorized 'load'
2611 // must be followed by a 'shuffle' with the required jumbled mask.
2612 if (!VL.empty() && (E->NeedToShuffle)) {
2613 assert(VL.size() == E->Scalars.size() &&((VL.size() == E->Scalars.size() && "Equal number of scalars expected"
) ? static_cast<void> (0) : __assert_fail ("VL.size() == E->Scalars.size() && \"Equal number of scalars expected\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2614, __PRETTY_FUNCTION__))
2614 "Equal number of scalars expected")((VL.size() == E->Scalars.size() && "Equal number of scalars expected"
) ? static_cast<void> (0) : __assert_fail ("VL.size() == E->Scalars.size() && \"Equal number of scalars expected\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2614, __PRETTY_FUNCTION__))
;
2615 SmallVector<Constant *, 8> Mask;
2616 for (Value *Val : VL) {
2617 if (ScalarToTreeEntry.count(Val)) {
2618 int Idx = ScalarToTreeEntry[Val];
2619 TreeEntry *E = &VectorizableTree[Idx];
2620 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
2621 if (E->Scalars[Lane] == Val) {
2622 Mask.push_back(Builder.getInt32(Lane));
2623 break;
2624 }
2625 }
2626 }
2627 }
2628
2629 // Generate shuffle for jumbled memory access
2630 Value *Undef = UndefValue::get(VecTy);
2631 Value *Shuf = Builder.CreateShuffleVector((Value *)LI, Undef,
2632 ConstantVector::get(Mask));
2633 return Shuf;
2634 }
2635
2636 return LI;
2637 }
2638 case Instruction::Store: {
2639 StoreInst *SI = cast<StoreInst>(VL0);
2640 unsigned Alignment = SI->getAlignment();
2641 unsigned AS = SI->getPointerAddressSpace();
2642
2643 ValueList ValueOp;
2644 for (Value *V : E->Scalars)
2645 ValueOp.push_back(cast<StoreInst>(V)->getValueOperand());
2646
2647 setInsertPointAfterBundle(E->Scalars);
2648
2649 Value *VecValue = vectorizeTree(ValueOp);
2650 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
2651 VecTy->getPointerTo(AS));
2652 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
2653
2654 // The pointer operand uses an in-tree scalar so we add the new BitCast to
2655 // ExternalUses list to make sure that an extract will be generated in the
2656 // future.
2657 if (ScalarToTreeEntry.count(SI->getPointerOperand()))
2658 ExternalUses.push_back(
2659 ExternalUser(SI->getPointerOperand(), cast<User>(VecPtr), 0));
2660
2661 if (!Alignment) {
2662 Alignment = DL->getABITypeAlignment(SI->getValueOperand()->getType());
2663 }
2664 S->setAlignment(Alignment);
2665 E->VectorizedValue = S;
2666 ++NumVectorInstructions;
2667 return propagateMetadata(S, E->Scalars);
2668 }
2669 case Instruction::GetElementPtr: {
2670 setInsertPointAfterBundle(E->Scalars);
2671
2672 ValueList Op0VL;
2673 for (Value *V : E->Scalars)
2674 Op0VL.push_back(cast<GetElementPtrInst>(V)->getOperand(0));
2675
2676 Value *Op0 = vectorizeTree(Op0VL);
2677
2678 std::vector<Value *> OpVecs;
2679 for (int j = 1, e = cast<GetElementPtrInst>(VL0)->getNumOperands(); j < e;
2680 ++j) {
2681 ValueList OpVL;
2682 for (Value *V : E->Scalars)
2683 OpVL.push_back(cast<GetElementPtrInst>(V)->getOperand(j));
2684
2685 Value *OpVec = vectorizeTree(OpVL);
2686 OpVecs.push_back(OpVec);
2687 }
2688
2689 Value *V = Builder.CreateGEP(
2690 cast<GetElementPtrInst>(VL0)->getSourceElementType(), Op0, OpVecs);
2691 E->VectorizedValue = V;
2692 ++NumVectorInstructions;
2693
2694 if (Instruction *I = dyn_cast<Instruction>(V))
2695 return propagateMetadata(I, E->Scalars);
2696
2697 return V;
2698 }
2699 case Instruction::Call: {
2700 CallInst *CI = cast<CallInst>(VL0);
2701 setInsertPointAfterBundle(E->Scalars);
2702 Function *FI;
2703 Intrinsic::ID IID = Intrinsic::not_intrinsic;
2704 Value *ScalarArg = nullptr;
2705 if (CI && (FI = CI->getCalledFunction())) {
2706 IID = FI->getIntrinsicID();
2707 }
2708 std::vector<Value *> OpVecs;
2709 for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) {
2710 ValueList OpVL;
2711 // ctlz,cttz and powi are special intrinsics whose second argument is
2712 // a scalar. This argument should not be vectorized.
2713 if (hasVectorInstrinsicScalarOpd(IID, 1) && j == 1) {
2714 CallInst *CEI = cast<CallInst>(E->Scalars[0]);
2715 ScalarArg = CEI->getArgOperand(j);
2716 OpVecs.push_back(CEI->getArgOperand(j));
2717 continue;
2718 }
2719 for (Value *V : E->Scalars) {
2720 CallInst *CEI = cast<CallInst>(V);
2721 OpVL.push_back(CEI->getArgOperand(j));
2722 }
2723
2724 Value *OpVec = vectorizeTree(OpVL);
2725 DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: OpVec[" << j << "]: "
<< *OpVec << "\n"; } } while (false)
;
2726 OpVecs.push_back(OpVec);
2727 }
2728
2729 Module *M = F->getParent();
2730 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
2731 Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) };
2732 Function *CF = Intrinsic::getDeclaration(M, ID, Tys);
2733 SmallVector<OperandBundleDef, 1> OpBundles;
2734 CI->getOperandBundlesAsDefs(OpBundles);
2735 Value *V = Builder.CreateCall(CF, OpVecs, OpBundles);
2736
2737 // The scalar argument uses an in-tree scalar so we add the new vectorized
2738 // call to ExternalUses list to make sure that an extract will be
2739 // generated in the future.
2740 if (ScalarArg && ScalarToTreeEntry.count(ScalarArg))
2741 ExternalUses.push_back(ExternalUser(ScalarArg, cast<User>(V), 0));
2742
2743 E->VectorizedValue = V;
2744 propagateIRFlags(E->VectorizedValue, E->Scalars);
2745 ++NumVectorInstructions;
2746 return V;
2747 }
2748 case Instruction::ShuffleVector: {
2749 ValueList LHSVL, RHSVL;
2750 assert(isa<BinaryOperator>(VL0) && "Invalid Shuffle Vector Operand")((isa<BinaryOperator>(VL0) && "Invalid Shuffle Vector Operand"
) ? static_cast<void> (0) : __assert_fail ("isa<BinaryOperator>(VL0) && \"Invalid Shuffle Vector Operand\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2750, __PRETTY_FUNCTION__))
;
2751 reorderAltShuffleOperands(E->Scalars, LHSVL, RHSVL);
2752 setInsertPointAfterBundle(E->Scalars);
6
Calling 'BoUpSLP::setInsertPointAfterBundle'
2753
2754 Value *LHS = vectorizeTree(LHSVL);
2755 Value *RHS = vectorizeTree(RHSVL);
2756
2757 if (Value *V = alreadyVectorized(E->Scalars))
2758 return V;
2759
2760 // Create a vector of LHS op1 RHS
2761 BinaryOperator *BinOp0 = cast<BinaryOperator>(VL0);
2762 Value *V0 = Builder.CreateBinOp(BinOp0->getOpcode(), LHS, RHS);
2763
2764 // Create a vector of LHS op2 RHS
2765 Instruction *VL1 = cast<Instruction>(E->Scalars[1]);
2766 BinaryOperator *BinOp1 = cast<BinaryOperator>(VL1);
2767 Value *V1 = Builder.CreateBinOp(BinOp1->getOpcode(), LHS, RHS);
2768
2769 // Create shuffle to take alternate operations from the vector.
2770 // Also, gather up odd and even scalar ops to propagate IR flags to
2771 // each vector operation.
2772 ValueList OddScalars, EvenScalars;
2773 unsigned e = E->Scalars.size();
2774 SmallVector<Constant *, 8> Mask(e);
2775 for (unsigned i = 0; i < e; ++i) {
2776 if (i & 1) {
2777 Mask[i] = Builder.getInt32(e + i);
2778 OddScalars.push_back(E->Scalars[i]);
2779 } else {
2780 Mask[i] = Builder.getInt32(i);
2781 EvenScalars.push_back(E->Scalars[i]);
2782 }
2783 }
2784
2785 Value *ShuffleMask = ConstantVector::get(Mask);
2786 propagateIRFlags(V0, EvenScalars);
2787 propagateIRFlags(V1, OddScalars);
2788
2789 Value *V = Builder.CreateShuffleVector(V0, V1, ShuffleMask);
2790 E->VectorizedValue = V;
2791 ++NumVectorInstructions;
2792 if (Instruction *I = dyn_cast<Instruction>(V))
2793 return propagateMetadata(I, E->Scalars);
2794
2795 return V;
2796 }
2797 default:
2798 llvm_unreachable("unknown inst")::llvm::llvm_unreachable_internal("unknown inst", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2798)
;
2799 }
2800 return nullptr;
2801}
2802
2803Value *BoUpSLP::vectorizeTree() {
2804 MapVector<Value *, DebugLoc> ExternallyUsedValues;
2805 return vectorizeTree(ExternallyUsedValues);
2806}
2807
2808Value *
2809BoUpSLP::vectorizeTree(MapVector<Value *, DebugLoc> &ExternallyUsedValues) {
2810
2811 // All blocks must be scheduled before any instructions are inserted.
2812 for (auto &BSIter : BlocksSchedules) {
2813 scheduleBlock(BSIter.second.get());
2814 }
2815
2816 Builder.SetInsertPoint(&F->getEntryBlock().front());
2817 auto *VectorRoot = vectorizeTree(ArrayRef<Value *>(), &VectorizableTree[0]);
2818
2819 // If the vectorized tree can be rewritten in a smaller type, we truncate the
2820 // vectorized root. InstCombine will then rewrite the entire expression. We
2821 // sign extend the extracted values below.
2822 auto *ScalarRoot = VectorizableTree[0].Scalars[0];
2823 if (MinBWs.count(ScalarRoot)) {
2824 if (auto *I = dyn_cast<Instruction>(VectorRoot))
2825 Builder.SetInsertPoint(&*++BasicBlock::iterator(I));
2826 auto BundleWidth = VectorizableTree[0].Scalars.size();
2827 auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first);
2828 auto *VecTy = VectorType::get(MinTy, BundleWidth);
2829 auto *Trunc = Builder.CreateTrunc(VectorRoot, VecTy);
2830 VectorizableTree[0].VectorizedValue = Trunc;
2831 }
2832
2833 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses
.size() << " values .\n"; } } while (false)
;
2834
2835 // If necessary, sign-extend or zero-extend ScalarRoot to the larger type
2836 // specified by ScalarType.
2837 auto extend = [&](Value *ScalarRoot, Value *Ex, Type *ScalarType) {
2838 if (!MinBWs.count(ScalarRoot))
2839 return Ex;
2840 if (MinBWs[ScalarRoot].second)
2841 return Builder.CreateSExt(Ex, ScalarType);
2842 return Builder.CreateZExt(Ex, ScalarType);
2843 };
2844
2845 // Extract all of the elements with the external uses.
2846 for (const auto &ExternalUse : ExternalUses) {
2847 Value *Scalar = ExternalUse.Scalar;
2848 llvm::User *User = ExternalUse.User;
2849
2850 // Skip users that we already RAUW. This happens when one instruction
2851 // has multiple uses of the same value.
2852 if (User && !is_contained(Scalar->users(), User))
2853 continue;
2854 assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar")((ScalarToTreeEntry.count(Scalar) && "Invalid scalar"
) ? static_cast<void> (0) : __assert_fail ("ScalarToTreeEntry.count(Scalar) && \"Invalid scalar\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2854, __PRETTY_FUNCTION__))
;
2855
2856 int Idx = ScalarToTreeEntry[Scalar];
2857 TreeEntry *E = &VectorizableTree[Idx];
2858 assert(!E->NeedToGather && "Extracting from a gather list")((!E->NeedToGather && "Extracting from a gather list"
) ? static_cast<void> (0) : __assert_fail ("!E->NeedToGather && \"Extracting from a gather list\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2858, __PRETTY_FUNCTION__))
;
2859
2860 Value *Vec = E->VectorizedValue;
2861 assert(Vec && "Can't find vectorizable value")((Vec && "Can't find vectorizable value") ? static_cast
<void> (0) : __assert_fail ("Vec && \"Can't find vectorizable value\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2861, __PRETTY_FUNCTION__))
;
2862
2863 Value *Lane = Builder.getInt32(ExternalUse.Lane);
2864 // If User == nullptr, the Scalar is used as extra arg. Generate
2865 // ExtractElement instruction and update the record for this scalar in
2866 // ExternallyUsedValues.
2867 if (!User) {
2868 assert(ExternallyUsedValues.count(Scalar) &&((ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in "
"ExternallyUsedValues map") ? static_cast<void> (0) : __assert_fail
("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2870, __PRETTY_FUNCTION__))
2869 "Scalar with nullptr as an external user must be registered in "((ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in "
"ExternallyUsedValues map") ? static_cast<void> (0) : __assert_fail
("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2870, __PRETTY_FUNCTION__))
2870 "ExternallyUsedValues map")((ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in "
"ExternallyUsedValues map") ? static_cast<void> (0) : __assert_fail
("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2870, __PRETTY_FUNCTION__))
;
2871 DebugLoc DL = ExternallyUsedValues[Scalar];
2872 if (auto *VecI = dyn_cast<Instruction>(Vec)) {
2873 Builder.SetInsertPoint(VecI->getParent(),
2874 std::next(VecI->getIterator()));
2875 } else {
2876 Builder.SetInsertPoint(&F->getEntryBlock().front());
2877 }
2878 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2879 Ex = extend(ScalarRoot, Ex, Scalar->getType());
2880 CSEBlocks.insert(cast<Instruction>(Scalar)->getParent());
2881 ExternallyUsedValues.erase(Scalar);
2882 ExternallyUsedValues[Ex] = DL;
2883 continue;
2884 }
2885
2886 // Generate extracts for out-of-tree users.
2887 // Find the insertion point for the extractelement lane.
2888 if (auto *VecI = dyn_cast<Instruction>(Vec)) {
2889 if (PHINode *PH = dyn_cast<PHINode>(User)) {
2890 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) {
2891 if (PH->getIncomingValue(i) == Scalar) {
2892 TerminatorInst *IncomingTerminator =
2893 PH->getIncomingBlock(i)->getTerminator();
2894 if (isa<CatchSwitchInst>(IncomingTerminator)) {
2895 Builder.SetInsertPoint(VecI->getParent(),
2896 std::next(VecI->getIterator()));
2897 } else {
2898 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator());
2899 }
2900 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2901 Ex = extend(ScalarRoot, Ex, Scalar->getType());
2902 CSEBlocks.insert(PH->getIncomingBlock(i));
2903 PH->setOperand(i, Ex);
2904 }
2905 }
2906 } else {
2907 Builder.SetInsertPoint(cast<Instruction>(User));
2908 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2909 Ex = extend(ScalarRoot, Ex, Scalar->getType());
2910 CSEBlocks.insert(cast<Instruction>(User)->getParent());
2911 User->replaceUsesOfWith(Scalar, Ex);
2912 }
2913 } else {
2914 Builder.SetInsertPoint(&F->getEntryBlock().front());
2915 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2916 Ex = extend(ScalarRoot, Ex, Scalar->getType());
2917 CSEBlocks.insert(&F->getEntryBlock());
2918 User->replaceUsesOfWith(Scalar, Ex);
2919 }
2920
2921 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Replaced:" << *User <<
".\n"; } } while (false)
;
2922 }
2923
2924 // For each vectorized value:
2925 for (TreeEntry &EIdx : VectorizableTree) {
2926 TreeEntry *Entry = &EIdx;
2927
2928 // For each lane:
2929 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
2930 Value *Scalar = Entry->Scalars[Lane];
2931 // No need to handle users of gathered values.
2932 if (Entry->NeedToGather)
2933 continue;
2934
2935 assert(Entry->VectorizedValue && "Can't find vectorizable value")((Entry->VectorizedValue && "Can't find vectorizable value"
) ? static_cast<void> (0) : __assert_fail ("Entry->VectorizedValue && \"Can't find vectorizable value\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2935, __PRETTY_FUNCTION__))
;
2936
2937 Type *Ty = Scalar->getType();
2938 if (!Ty->isVoidTy()) {
2939#ifndef NDEBUG
2940 for (User *U : Scalar->users()) {
2941 DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: \tvalidating user:" <<
*U << ".\n"; } } while (false)
;
2942
2943 assert((ScalarToTreeEntry.count(U) ||(((ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList,
U)) && "Replacing out-of-tree value with undef") ? static_cast
<void> (0) : __assert_fail ("(ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList, U)) && \"Replacing out-of-tree value with undef\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2946, __PRETTY_FUNCTION__))
2944 // It is legal to replace users in the ignorelist by undef.(((ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList,
U)) && "Replacing out-of-tree value with undef") ? static_cast
<void> (0) : __assert_fail ("(ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList, U)) && \"Replacing out-of-tree value with undef\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2946, __PRETTY_FUNCTION__))
2945 is_contained(UserIgnoreList, U)) &&(((ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList,
U)) && "Replacing out-of-tree value with undef") ? static_cast
<void> (0) : __assert_fail ("(ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList, U)) && \"Replacing out-of-tree value with undef\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2946, __PRETTY_FUNCTION__))
2946 "Replacing out-of-tree value with undef")(((ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList,
U)) && "Replacing out-of-tree value with undef") ? static_cast
<void> (0) : __assert_fail ("(ScalarToTreeEntry.count(U) || is_contained(UserIgnoreList, U)) && \"Replacing out-of-tree value with undef\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2946, __PRETTY_FUNCTION__))
;
2947 }
2948#endif
2949 Value *Undef = UndefValue::get(Ty);
2950 Scalar->replaceAllUsesWith(Undef);
2951 }
2952 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: \tErasing scalar:" << *
Scalar << ".\n"; } } while (false)
;
2953 eraseInstruction(cast<Instruction>(Scalar));
2954 }
2955 }
2956
2957 Builder.ClearInsertionPoint();
2958
2959 return VectorizableTree[0].VectorizedValue;
2960}
2961
2962void BoUpSLP::optimizeGatherSequence() {
2963 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq
.size() << " gather sequences instructions.\n"; } } while
(false)
2964 << " gather sequences instructions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq
.size() << " gather sequences instructions.\n"; } } while
(false)
;
2965 // LICM InsertElementInst sequences.
2966 for (Instruction *it : GatherSeq) {
2967 InsertElementInst *Insert = dyn_cast<InsertElementInst>(it);
2968
2969 if (!Insert)
2970 continue;
2971
2972 // Check if this block is inside a loop.
2973 Loop *L = LI->getLoopFor(Insert->getParent());
2974 if (!L)
2975 continue;
2976
2977 // Check if it has a preheader.
2978 BasicBlock *PreHeader = L->getLoopPreheader();
2979 if (!PreHeader)
2980 continue;
2981
2982 // If the vector or the element that we insert into it are
2983 // instructions that are defined in this basic block then we can't
2984 // hoist this instruction.
2985 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
2986 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
2987 if (CurrVec && L->contains(CurrVec))
2988 continue;
2989 if (NewElem && L->contains(NewElem))
2990 continue;
2991
2992 // We can hoist this instruction. Move it to the pre-header.
2993 Insert->moveBefore(PreHeader->getTerminator());
2994 }
2995
2996 // Make a list of all reachable blocks in our CSE queue.
2997 SmallVector<const DomTreeNode *, 8> CSEWorkList;
2998 CSEWorkList.reserve(CSEBlocks.size());
2999 for (BasicBlock *BB : CSEBlocks)
3000 if (DomTreeNode *N = DT->getNode(BB)) {
3001 assert(DT->isReachableFromEntry(N))((DT->isReachableFromEntry(N)) ? static_cast<void> (
0) : __assert_fail ("DT->isReachableFromEntry(N)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3001, __PRETTY_FUNCTION__))
;
3002 CSEWorkList.push_back(N);
3003 }
3004
3005 // Sort blocks by domination. This ensures we visit a block after all blocks
3006 // dominating it are visited.
3007 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(),
3008 [this](const DomTreeNode *A, const DomTreeNode *B) {
3009 return DT->properlyDominates(A, B);
3010 });
3011
3012 // Perform O(N^2) search over the gather sequences and merge identical
3013 // instructions. TODO: We can further optimize this scan if we split the
3014 // instructions into different buckets based on the insert lane.
3015 SmallVector<Instruction *, 16> Visited;
3016 for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) {
3017 assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&(((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev
(I))) && "Worklist not sorted properly!") ? static_cast
<void> (0) : __assert_fail ("(I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3018, __PRETTY_FUNCTION__))
3018 "Worklist not sorted properly!")(((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev
(I))) && "Worklist not sorted properly!") ? static_cast
<void> (0) : __assert_fail ("(I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3018, __PRETTY_FUNCTION__))
;
3019 BasicBlock *BB = (*I)->getBlock();
3020 // For all instructions in blocks containing gather sequences:
3021 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) {
3022 Instruction *In = &*it++;
3023 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In))
3024 continue;
3025
3026 // Check if we can replace this instruction with any of the
3027 // visited instructions.
3028 for (Instruction *v : Visited) {
3029 if (In->isIdenticalTo(v) &&
3030 DT->dominates(v->getParent(), In->getParent())) {
3031 In->replaceAllUsesWith(v);
3032 eraseInstruction(In);
3033 In = nullptr;
3034 break;
3035 }
3036 }
3037 if (In) {
3038 assert(!is_contained(Visited, In))((!is_contained(Visited, In)) ? static_cast<void> (0) :
__assert_fail ("!is_contained(Visited, In)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3038, __PRETTY_FUNCTION__))
;
3039 Visited.push_back(In);
3040 }
3041 }
3042 }
3043 CSEBlocks.clear();
3044 GatherSeq.clear();
3045}
3046
3047// Groups the instructions to a bundle (which is then a single scheduling entity)
3048// and schedules instructions until the bundle gets ready.
3049bool BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef<Value *> VL,
3050 BoUpSLP *SLP) {
3051 if (isa<PHINode>(VL[0]))
3052 return true;
3053
3054 // Initialize the instruction bundle.
3055 Instruction *OldScheduleEnd = ScheduleEnd;
3056 ScheduleData *PrevInBundle = nullptr;
3057 ScheduleData *Bundle = nullptr;
3058 bool ReSchedule = false;
3059 DEBUG(dbgs() << "SLP: bundle: " << *VL[0] << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: bundle: " << *VL[0] <<
"\n"; } } while (false)
;
3060
3061 // Make sure that the scheduling region contains all
3062 // instructions of the bundle.
3063 for (Value *V : VL) {
3064 if (!extendSchedulingRegion(V))
3065 return false;
3066 }
3067
3068 for (Value *V : VL) {
3069 ScheduleData *BundleMember = getScheduleData(V);
3070 assert(BundleMember &&((BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)"
) ? static_cast<void> (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3071, __PRETTY_FUNCTION__))
3071 "no ScheduleData for bundle member (maybe not in same basic block)")((BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)"
) ? static_cast<void> (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3071, __PRETTY_FUNCTION__))
;
3072 if (BundleMember->IsScheduled) {
3073 // A bundle member was scheduled as single instruction before and now
3074 // needs to be scheduled as part of the bundle. We just get rid of the
3075 // existing schedule.
3076 DEBUG(dbgs() << "SLP: reset schedule because " << *BundleMemberdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: reset schedule because " <<
*BundleMember << " was already scheduled\n"; } } while
(false)
3077 << " was already scheduled\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: reset schedule because " <<
*BundleMember << " was already scheduled\n"; } } while
(false)
;
3078 ReSchedule = true;
3079 }
3080 assert(BundleMember->isSchedulingEntity() &&((BundleMember->isSchedulingEntity() && "bundle member already part of other bundle"
) ? static_cast<void> (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3081, __PRETTY_FUNCTION__))
3081 "bundle member already part of other bundle")((BundleMember->isSchedulingEntity() && "bundle member already part of other bundle"
) ? static_cast<void> (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3081, __PRETTY_FUNCTION__))
;
3082 if (PrevInBundle) {
3083 PrevInBundle->NextInBundle = BundleMember;
3084 } else {
3085 Bundle = BundleMember;
3086 }
3087 BundleMember->UnscheduledDepsInBundle = 0;
3088 Bundle->UnscheduledDepsInBundle += BundleMember->UnscheduledDeps;
3089
3090 // Group the instructions to a bundle.
3091 BundleMember->FirstInBundle = Bundle;
3092 PrevInBundle = BundleMember;
3093 }
3094 if (ScheduleEnd != OldScheduleEnd) {
3095 // The scheduling region got new instructions at the lower end (or it is a
3096 // new region for the first bundle). This makes it necessary to
3097 // recalculate all dependencies.
3098 // It is seldom that this needs to be done a second time after adding the
3099 // initial bundle to the region.
3100 for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
3101 ScheduleData *SD = getScheduleData(I);
3102 SD->clearDependencies();
3103 }
3104 ReSchedule = true;
3105 }
3106 if (ReSchedule) {
3107 resetSchedule();
3108 initialFillReadyList(ReadyInsts);
3109 }
3110
3111 DEBUG(dbgs() << "SLP: try schedule bundle " << *Bundle << " in block "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: try schedule bundle " <<
*Bundle << " in block " << BB->getName() <<
"\n"; } } while (false)
3112 << BB->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: try schedule bundle " <<
*Bundle << " in block " << BB->getName() <<
"\n"; } } while (false)
;
3113
3114 calculateDependencies(Bundle, true, SLP);
3115
3116 // Now try to schedule the new bundle. As soon as the bundle is "ready" it
3117 // means that there are no cyclic dependencies and we can schedule it.
3118 // Note that's important that we don't "schedule" the bundle yet (see
3119 // cancelScheduling).
3120 while (!Bundle->isReady() && !ReadyInsts.empty()) {
3121
3122 ScheduleData *pickedSD = ReadyInsts.back();
3123 ReadyInsts.pop_back();
3124
3125 if (pickedSD->isSchedulingEntity() && pickedSD->isReady()) {
3126 schedule(pickedSD, ReadyInsts);
3127 }
3128 }
3129 if (!Bundle->isReady()) {
3130 cancelScheduling(VL);
3131 return false;
3132 }
3133 return true;
3134}
3135
3136void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef<Value *> VL) {
3137 if (isa<PHINode>(VL[0]))
3138 return;
3139
3140 ScheduleData *Bundle = getScheduleData(VL[0]);
3141 DEBUG(dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: cancel scheduling of " <<
*Bundle << "\n"; } } while (false)
;
3142 assert(!Bundle->IsScheduled &&((!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled"
) ? static_cast<void> (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3143, __PRETTY_FUNCTION__))
3143 "Can't cancel bundle which is already scheduled")((!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled"
) ? static_cast<void> (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3143, __PRETTY_FUNCTION__))
;
3144 assert(Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() &&((Bundle->isSchedulingEntity() && Bundle->isPartOfBundle
() && "tried to unbundle something which is not a bundle"
) ? static_cast<void> (0) : __assert_fail ("Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && \"tried to unbundle something which is not a bundle\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3145, __PRETTY_FUNCTION__))
3145 "tried to unbundle something which is not a bundle")((Bundle->isSchedulingEntity() && Bundle->isPartOfBundle
() && "tried to unbundle something which is not a bundle"
) ? static_cast<void> (0) : __assert_fail ("Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && \"tried to unbundle something which is not a bundle\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3145, __PRETTY_FUNCTION__))
;
3146
3147 // Un-bundle: make single instructions out of the bundle.
3148 ScheduleData *BundleMember = Bundle;
3149 while (BundleMember) {
3150 assert(BundleMember->FirstInBundle == Bundle && "corrupt bundle links")((BundleMember->FirstInBundle == Bundle && "corrupt bundle links"
) ? static_cast<void> (0) : __assert_fail ("BundleMember->FirstInBundle == Bundle && \"corrupt bundle links\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3150, __PRETTY_FUNCTION__))
;
3151 BundleMember->FirstInBundle = BundleMember;
3152 ScheduleData *Next = BundleMember->NextInBundle;
3153 BundleMember->NextInBundle = nullptr;
3154 BundleMember->UnscheduledDepsInBundle = BundleMember->UnscheduledDeps;
3155 if (BundleMember->UnscheduledDepsInBundle == 0) {
3156 ReadyInsts.insert(BundleMember);
3157 }
3158 BundleMember = Next;
3159 }
3160}
3161
3162bool BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V) {
3163 if (getScheduleData(V))
3164 return true;
3165 Instruction *I = dyn_cast<Instruction>(V);
3166 assert(I && "bundle member must be an instruction")((I && "bundle member must be an instruction") ? static_cast
<void> (0) : __assert_fail ("I && \"bundle member must be an instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3166, __PRETTY_FUNCTION__))
;
3167 assert(!isa<PHINode>(I) && "phi nodes don't need to be scheduled")((!isa<PHINode>(I) && "phi nodes don't need to be scheduled"
) ? static_cast<void> (0) : __assert_fail ("!isa<PHINode>(I) && \"phi nodes don't need to be scheduled\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3167, __PRETTY_FUNCTION__))
;
3168 if (!ScheduleStart) {
3169 // It's the first instruction in the new region.
3170 initScheduleData(I, I->getNextNode(), nullptr, nullptr);
3171 ScheduleStart = I;
3172 ScheduleEnd = I->getNextNode();
3173 assert(ScheduleEnd && "tried to vectorize a TerminatorInst?")((ScheduleEnd && "tried to vectorize a TerminatorInst?"
) ? static_cast<void> (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a TerminatorInst?\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3173, __PRETTY_FUNCTION__))
;
3174 DEBUG(dbgs() << "SLP: initialize schedule region to " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: initialize schedule region to "
<< *I << "\n"; } } while (false)
;
3175 return true;
3176 }
3177 // Search up and down at the same time, because we don't know if the new
3178 // instruction is above or below the existing scheduling region.
3179 BasicBlock::reverse_iterator UpIter =
3180 ++ScheduleStart->getIterator().getReverse();
3181 BasicBlock::reverse_iterator UpperEnd = BB->rend();
3182 BasicBlock::iterator DownIter = ScheduleEnd->getIterator();
3183 BasicBlock::iterator LowerEnd = BB->end();
3184 for (;;) {
3185 if (++ScheduleRegionSize > ScheduleRegionSizeLimit) {
3186 DEBUG(dbgs() << "SLP: exceeded schedule region size limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: exceeded schedule region size limit\n"
; } } while (false)
;
3187 return false;
3188 }
3189
3190 if (UpIter != UpperEnd) {
3191 if (&*UpIter == I) {
3192 initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion);
3193 ScheduleStart = I;
3194 DEBUG(dbgs() << "SLP: extend schedule region start to " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: extend schedule region start to "
<< *I << "\n"; } } while (false)
;
3195 return true;
3196 }
3197 UpIter++;
3198 }
3199 if (DownIter != LowerEnd) {
3200 if (&*DownIter == I) {
3201 initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion,
3202 nullptr);
3203 ScheduleEnd = I->getNextNode();
3204 assert(ScheduleEnd && "tried to vectorize a TerminatorInst?")((ScheduleEnd && "tried to vectorize a TerminatorInst?"
) ? static_cast<void> (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a TerminatorInst?\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3204, __PRETTY_FUNCTION__))
;
3205 DEBUG(dbgs() << "SLP: extend schedule region end to " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: extend schedule region end to "
<< *I << "\n"; } } while (false)
;
3206 return true;
3207 }
3208 DownIter++;
3209 }
3210 assert((UpIter != UpperEnd || DownIter != LowerEnd) &&(((UpIter != UpperEnd || DownIter != LowerEnd) && "instruction not found in block"
) ? static_cast<void> (0) : __assert_fail ("(UpIter != UpperEnd || DownIter != LowerEnd) && \"instruction not found in block\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3211, __PRETTY_FUNCTION__))
3211 "instruction not found in block")(((UpIter != UpperEnd || DownIter != LowerEnd) && "instruction not found in block"
) ? static_cast<void> (0) : __assert_fail ("(UpIter != UpperEnd || DownIter != LowerEnd) && \"instruction not found in block\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3211, __PRETTY_FUNCTION__))
;
3212 }
3213 return true;
3214}
3215
3216void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI,
3217 Instruction *ToI,
3218 ScheduleData *PrevLoadStore,
3219 ScheduleData *NextLoadStore) {
3220 ScheduleData *CurrentLoadStore = PrevLoadStore;
3221 for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) {
3222 ScheduleData *SD = ScheduleDataMap[I];
3223 if (!SD) {
3224 // Allocate a new ScheduleData for the instruction.
3225 if (ChunkPos >= ChunkSize) {
3226 ScheduleDataChunks.push_back(
3227 llvm::make_unique<ScheduleData[]>(ChunkSize));
3228 ChunkPos = 0;
3229 }
3230 SD = &(ScheduleDataChunks.back()[ChunkPos++]);
3231 ScheduleDataMap[I] = SD;
3232 SD->Inst = I;
3233 }
3234 assert(!isInSchedulingRegion(SD) &&((!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region"
) ? static_cast<void> (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3235, __PRETTY_FUNCTION__))
3235 "new ScheduleData already in scheduling region")((!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region"
) ? static_cast<void> (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3235, __PRETTY_FUNCTION__))
;
3236 SD->init(SchedulingRegionID);
3237
3238 if (I->mayReadOrWriteMemory()) {
3239 // Update the linked list of memory accessing instructions.
3240 if (CurrentLoadStore) {
3241 CurrentLoadStore->NextLoadStore = SD;
3242 } else {
3243 FirstLoadStoreInRegion = SD;
3244 }
3245 CurrentLoadStore = SD;
3246 }
3247 }
3248 if (NextLoadStore) {
3249 if (CurrentLoadStore)
3250 CurrentLoadStore->NextLoadStore = NextLoadStore;
3251 } else {
3252 LastLoadStoreInRegion = CurrentLoadStore;
3253 }
3254}
3255
3256void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD,
3257 bool InsertInReadyList,
3258 BoUpSLP *SLP) {
3259 assert(SD->isSchedulingEntity())((SD->isSchedulingEntity()) ? static_cast<void> (0) :
__assert_fail ("SD->isSchedulingEntity()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3259, __PRETTY_FUNCTION__))
;
3260
3261 SmallVector<ScheduleData *, 10> WorkList;
3262 WorkList.push_back(SD);
3263
3264 while (!WorkList.empty()) {
3265 ScheduleData *SD = WorkList.back();
3266 WorkList.pop_back();
3267
3268 ScheduleData *BundleMember = SD;
3269 while (BundleMember) {
3270 assert(isInSchedulingRegion(BundleMember))((isInSchedulingRegion(BundleMember)) ? static_cast<void>
(0) : __assert_fail ("isInSchedulingRegion(BundleMember)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3270, __PRETTY_FUNCTION__))
;
3271 if (!BundleMember->hasValidDependencies()) {
3272
3273 DEBUG(dbgs() << "SLP: update deps of " << *BundleMember << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: update deps of " <<
*BundleMember << "\n"; } } while (false)
;
3274 BundleMember->Dependencies = 0;
3275 BundleMember->resetUnscheduledDeps();
3276
3277 // Handle def-use chain dependencies.
3278 for (User *U : BundleMember->Inst->users()) {
3279 if (isa<Instruction>(U)) {
3280 ScheduleData *UseSD = getScheduleData(U);
3281 if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) {
3282 BundleMember->Dependencies++;
3283 ScheduleData *DestBundle = UseSD->FirstInBundle;
3284 if (!DestBundle->IsScheduled) {
3285 BundleMember->incrementUnscheduledDeps(1);
3286 }
3287 if (!DestBundle->hasValidDependencies()) {
3288 WorkList.push_back(DestBundle);
3289 }
3290 }
3291 } else {
3292 // I'm not sure if this can ever happen. But we need to be safe.
3293 // This lets the instruction/bundle never be scheduled and
3294 // eventually disable vectorization.
3295 BundleMember->Dependencies++;
3296 BundleMember->incrementUnscheduledDeps(1);
3297 }
3298 }
3299
3300 // Handle the memory dependencies.
3301 ScheduleData *DepDest = BundleMember->NextLoadStore;
3302 if (DepDest) {
3303 Instruction *SrcInst = BundleMember->Inst;
3304 MemoryLocation SrcLoc = getLocation(SrcInst, SLP->AA);
3305 bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory();
3306 unsigned numAliased = 0;
3307 unsigned DistToSrc = 1;
3308
3309 while (DepDest) {
3310 assert(isInSchedulingRegion(DepDest))((isInSchedulingRegion(DepDest)) ? static_cast<void> (0
) : __assert_fail ("isInSchedulingRegion(DepDest)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3310, __PRETTY_FUNCTION__))
;
3311
3312 // We have two limits to reduce the complexity:
3313 // 1) AliasedCheckLimit: It's a small limit to reduce calls to
3314 // SLP->isAliased (which is the expensive part in this loop).
3315 // 2) MaxMemDepDistance: It's for very large blocks and it aborts
3316 // the whole loop (even if the loop is fast, it's quadratic).
3317 // It's important for the loop break condition (see below) to
3318 // check this limit even between two read-only instructions.
3319 if (DistToSrc >= MaxMemDepDistance ||
3320 ((SrcMayWrite || DepDest->Inst->mayWriteToMemory()) &&
3321 (numAliased >= AliasedCheckLimit ||
3322 SLP->isAliased(SrcLoc, SrcInst, DepDest->Inst)))) {
3323
3324 // We increment the counter only if the locations are aliased
3325 // (instead of counting all alias checks). This gives a better
3326 // balance between reduced runtime and accurate dependencies.
3327 numAliased++;
3328
3329 DepDest->MemoryDependencies.push_back(BundleMember);
3330 BundleMember->Dependencies++;
3331 ScheduleData *DestBundle = DepDest->FirstInBundle;
3332 if (!DestBundle->IsScheduled) {
3333 BundleMember->incrementUnscheduledDeps(1);
3334 }
3335 if (!DestBundle->hasValidDependencies()) {
3336 WorkList.push_back(DestBundle);
3337 }
3338 }
3339 DepDest = DepDest->NextLoadStore;
3340
3341 // Example, explaining the loop break condition: Let's assume our
3342 // starting instruction is i0 and MaxMemDepDistance = 3.
3343 //
3344 // +--------v--v--v
3345 // i0,i1,i2,i3,i4,i5,i6,i7,i8
3346 // +--------^--^--^
3347 //
3348 // MaxMemDepDistance let us stop alias-checking at i3 and we add
3349 // dependencies from i0 to i3,i4,.. (even if they are not aliased).
3350 // Previously we already added dependencies from i3 to i6,i7,i8
3351 // (because of MaxMemDepDistance). As we added a dependency from
3352 // i0 to i3, we have transitive dependencies from i0 to i6,i7,i8
3353 // and we can abort this loop at i6.
3354 if (DistToSrc >= 2 * MaxMemDepDistance)
3355 break;
3356 DistToSrc++;
3357 }
3358 }
3359 }
3360 BundleMember = BundleMember->NextInBundle;
3361 }
3362 if (InsertInReadyList && SD->isReady()) {
3363 ReadyInsts.push_back(SD);
3364 DEBUG(dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: gets ready on update: " <<
*SD->Inst << "\n"; } } while (false)
;
3365 }
3366 }
3367}
3368
3369void BoUpSLP::BlockScheduling::resetSchedule() {
3370 assert(ScheduleStart &&((ScheduleStart && "tried to reset schedule on block which has not been scheduled"
) ? static_cast<void> (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3371, __PRETTY_FUNCTION__))
3371 "tried to reset schedule on block which has not been scheduled")((ScheduleStart && "tried to reset schedule on block which has not been scheduled"
) ? static_cast<void> (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3371, __PRETTY_FUNCTION__))
;
3372 for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
3373 ScheduleData *SD = getScheduleData(I);
3374 assert(isInSchedulingRegion(SD))((isInSchedulingRegion(SD)) ? static_cast<void> (0) : __assert_fail
("isInSchedulingRegion(SD)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3374, __PRETTY_FUNCTION__))
;
3375 SD->IsScheduled = false;
3376 SD->resetUnscheduledDeps();
3377 }
3378 ReadyInsts.clear();
3379}
3380
3381void BoUpSLP::scheduleBlock(BlockScheduling *BS) {
3382
3383 if (!BS->ScheduleStart)
3384 return;
3385
3386 DEBUG(dbgs() << "SLP: schedule block " << BS->BB->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: schedule block " << BS
->BB->getName() << "\n"; } } while (false)
;
3387
3388 BS->resetSchedule();
3389
3390 // For the real scheduling we use a more sophisticated ready-list: it is
3391 // sorted by the original instruction location. This lets the final schedule
3392 // be as close as possible to the original instruction order.
3393 struct ScheduleDataCompare {
3394 bool operator()(ScheduleData *SD1, ScheduleData *SD2) const {
3395 return SD2->SchedulingPriority < SD1->SchedulingPriority;
3396 }
3397 };
3398 std::set<ScheduleData *, ScheduleDataCompare> ReadyInsts;
3399
3400 // Ensure that all dependency data is updated and fill the ready-list with
3401 // initial instructions.
3402 int Idx = 0;
3403 int NumToSchedule = 0;
3404 for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd;
3405 I = I->getNextNode()) {
3406 ScheduleData *SD = BS->getScheduleData(I);
3407 assert(((SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->
Inst) != 0) && "scheduler and vectorizer have different opinion on what is a bundle"
) ? static_cast<void> (0) : __assert_fail ("SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->Inst) != 0) && \"scheduler and vectorizer have different opinion on what is a bundle\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3409, __PRETTY_FUNCTION__))
3408 SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->Inst) != 0) &&((SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->
Inst) != 0) && "scheduler and vectorizer have different opinion on what is a bundle"
) ? static_cast<void> (0) : __assert_fail ("SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->Inst) != 0) && \"scheduler and vectorizer have different opinion on what is a bundle\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3409, __PRETTY_FUNCTION__))
3409 "scheduler and vectorizer have different opinion on what is a bundle")((SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->
Inst) != 0) && "scheduler and vectorizer have different opinion on what is a bundle"
) ? static_cast<void> (0) : __assert_fail ("SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->Inst) != 0) && \"scheduler and vectorizer have different opinion on what is a bundle\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3409, __PRETTY_FUNCTION__))
;
3410 SD->FirstInBundle->SchedulingPriority = Idx++;
3411 if (SD->isSchedulingEntity()) {
3412 BS->calculateDependencies(SD, false, this);
3413 NumToSchedule++;
3414 }
3415 }
3416 BS->initialFillReadyList(ReadyInsts);
3417
3418 Instruction *LastScheduledInst = BS->ScheduleEnd;
3419
3420 // Do the "real" scheduling.
3421 while (!ReadyInsts.empty()) {
3422 ScheduleData *picked = *ReadyInsts.begin();
3423 ReadyInsts.erase(ReadyInsts.begin());
3424
3425 // Move the scheduled instruction(s) to their dedicated places, if not
3426 // there yet.
3427 ScheduleData *BundleMember = picked;
3428 while (BundleMember) {
3429 Instruction *pickedInst = BundleMember->Inst;
3430 if (LastScheduledInst->getNextNode() != pickedInst) {
3431 BS->BB->getInstList().remove(pickedInst);
3432 BS->BB->getInstList().insert(LastScheduledInst->getIterator(),
3433 pickedInst);
3434 }
3435 LastScheduledInst = pickedInst;
3436 BundleMember = BundleMember->NextInBundle;
3437 }
3438
3439 BS->schedule(picked, ReadyInsts);
3440 NumToSchedule--;
3441 }
3442 assert(NumToSchedule == 0 && "could not schedule all instructions")((NumToSchedule == 0 && "could not schedule all instructions"
) ? static_cast<void> (0) : __assert_fail ("NumToSchedule == 0 && \"could not schedule all instructions\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3442, __PRETTY_FUNCTION__))
;
3443
3444 // Avoid duplicate scheduling of the block.
3445 BS->ScheduleStart = nullptr;
3446}
3447
3448unsigned BoUpSLP::getVectorElementSize(Value *V) {
3449 // If V is a store, just return the width of the stored value without
3450 // traversing the expression tree. This is the common case.
3451 if (auto *Store = dyn_cast<StoreInst>(V))
3452 return DL->getTypeSizeInBits(Store->getValueOperand()->getType());
3453
3454 // If V is not a store, we can traverse the expression tree to find loads
3455 // that feed it. The type of the loaded value may indicate a more suitable
3456 // width than V's type. We want to base the vector element size on the width
3457 // of memory operations where possible.
3458 SmallVector<Instruction *, 16> Worklist;
3459 SmallPtrSet<Instruction *, 16> Visited;
3460 if (auto *I = dyn_cast<Instruction>(V))
3461 Worklist.push_back(I);
3462
3463 // Traverse the expression tree in bottom-up order looking for loads. If we
3464 // encounter an instruciton we don't yet handle, we give up.
3465 auto MaxWidth = 0u;
3466 auto FoundUnknownInst = false;
3467 while (!Worklist.empty() && !FoundUnknownInst) {
3468 auto *I = Worklist.pop_back_val();
3469 Visited.insert(I);
3470
3471 // We should only be looking at scalar instructions here. If the current
3472 // instruction has a vector type, give up.
3473 auto *Ty = I->getType();
3474 if (isa<VectorType>(Ty))
3475 FoundUnknownInst = true;
3476
3477 // If the current instruction is a load, update MaxWidth to reflect the
3478 // width of the loaded value.
3479 else if (isa<LoadInst>(I))
3480 MaxWidth = std::max<unsigned>(MaxWidth, DL->getTypeSizeInBits(Ty));
3481
3482 // Otherwise, we need to visit the operands of the instruction. We only
3483 // handle the interesting cases from buildTree here. If an operand is an
3484 // instruction we haven't yet visited, we add it to the worklist.
3485 else if (isa<PHINode>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) ||
3486 isa<CmpInst>(I) || isa<SelectInst>(I) || isa<BinaryOperator>(I)) {
3487 for (Use &U : I->operands())
3488 if (auto *J = dyn_cast<Instruction>(U.get()))
3489 if (!Visited.count(J))
3490 Worklist.push_back(J);
3491 }
3492
3493 // If we don't yet handle the instruction, give up.
3494 else
3495 FoundUnknownInst = true;
3496 }
3497
3498 // If we didn't encounter a memory access in the expression tree, or if we
3499 // gave up for some reason, just return the width of V.
3500 if (!MaxWidth || FoundUnknownInst)
3501 return DL->getTypeSizeInBits(V->getType());
3502
3503 // Otherwise, return the maximum width we found.
3504 return MaxWidth;
3505}
3506
3507// Determine if a value V in a vectorizable expression Expr can be demoted to a
3508// smaller type with a truncation. We collect the values that will be demoted
3509// in ToDemote and additional roots that require investigating in Roots.
3510static bool collectValuesToDemote(Value *V, SmallPtrSetImpl<Value *> &Expr,
3511 SmallVectorImpl<Value *> &ToDemote,
3512 SmallVectorImpl<Value *> &Roots) {
3513
3514 // We can always demote constants.
3515 if (isa<Constant>(V)) {
3516 ToDemote.push_back(V);
3517 return true;
3518 }
3519
3520 // If the value is not an instruction in the expression with only one use, it
3521 // cannot be demoted.
3522 auto *I = dyn_cast<Instruction>(V);
3523 if (!I || !I->hasOneUse() || !Expr.count(I))
3524 return false;
3525
3526 switch (I->getOpcode()) {
3527
3528 // We can always demote truncations and extensions. Since truncations can
3529 // seed additional demotion, we save the truncated value.
3530 case Instruction::Trunc:
3531 Roots.push_back(I->getOperand(0));
3532 case Instruction::ZExt:
3533 case Instruction::SExt:
3534 break;
3535
3536 // We can demote certain binary operations if we can demote both of their
3537 // operands.
3538 case Instruction::Add:
3539 case Instruction::Sub:
3540 case Instruction::Mul:
3541 case Instruction::And:
3542 case Instruction::Or:
3543 case Instruction::Xor:
3544 if (!collectValuesToDemote(I->getOperand(0), Expr, ToDemote, Roots) ||
3545 !collectValuesToDemote(I->getOperand(1), Expr, ToDemote, Roots))
3546 return false;
3547 break;
3548
3549 // We can demote selects if we can demote their true and false values.
3550 case Instruction::Select: {
3551 SelectInst *SI = cast<SelectInst>(I);
3552 if (!collectValuesToDemote(SI->getTrueValue(), Expr, ToDemote, Roots) ||
3553 !collectValuesToDemote(SI->getFalseValue(), Expr, ToDemote, Roots))
3554 return false;
3555 break;
3556 }
3557
3558 // We can demote phis if we can demote all their incoming operands. Note that
3559 // we don't need to worry about cycles since we ensure single use above.
3560 case Instruction::PHI: {
3561 PHINode *PN = cast<PHINode>(I);
3562 for (Value *IncValue : PN->incoming_values())
3563 if (!collectValuesToDemote(IncValue, Expr, ToDemote, Roots))
3564 return false;
3565 break;
3566 }
3567
3568 // Otherwise, conservatively give up.
3569 default:
3570 return false;
3571 }
3572
3573 // Record the value that we can demote.
3574 ToDemote.push_back(V);
3575 return true;
3576}
3577
3578void BoUpSLP::computeMinimumValueSizes() {
3579 // If there are no external uses, the expression tree must be rooted by a
3580 // store. We can't demote in-memory values, so there is nothing to do here.
3581 if (ExternalUses.empty())
3582 return;
3583
3584 // We only attempt to truncate integer expressions.
3585 auto &TreeRoot = VectorizableTree[0].Scalars;
3586 auto *TreeRootIT = dyn_cast<IntegerType>(TreeRoot[0]->getType());
3587 if (!TreeRootIT)
3588 return;
3589
3590 // If the expression is not rooted by a store, these roots should have
3591 // external uses. We will rely on InstCombine to rewrite the expression in
3592 // the narrower type. However, InstCombine only rewrites single-use values.
3593 // This means that if a tree entry other than a root is used externally, it
3594 // must have multiple uses and InstCombine will not rewrite it. The code
3595 // below ensures that only the roots are used externally.
3596 SmallPtrSet<Value *, 32> Expr(TreeRoot.begin(), TreeRoot.end());
3597 for (auto &EU : ExternalUses)
3598 if (!Expr.erase(EU.Scalar))
3599 return;
3600 if (!Expr.empty())
3601 return;
3602
3603 // Collect the scalar values of the vectorizable expression. We will use this
3604 // context to determine which values can be demoted. If we see a truncation,
3605 // we mark it as seeding another demotion.
3606 for (auto &Entry : VectorizableTree)
3607 Expr.insert(Entry.Scalars.begin(), Entry.Scalars.end());
3608
3609 // Ensure the roots of the vectorizable tree don't form a cycle. They must
3610 // have a single external user that is not in the vectorizable tree.
3611 for (auto *Root : TreeRoot)
3612 if (!Root->hasOneUse() || Expr.count(*Root->user_begin()))
3613 return;
3614
3615 // Conservatively determine if we can actually truncate the roots of the
3616 // expression. Collect the values that can be demoted in ToDemote and
3617 // additional roots that require investigating in Roots.
3618 SmallVector<Value *, 32> ToDemote;
3619 SmallVector<Value *, 4> Roots;
3620 for (auto *Root : TreeRoot)
3621 if (!collectValuesToDemote(Root, Expr, ToDemote, Roots))
3622 return;
3623
3624 // The maximum bit width required to represent all the values that can be
3625 // demoted without loss of precision. It would be safe to truncate the roots
3626 // of the expression to this width.
3627 auto MaxBitWidth = 8u;
3628
3629 // We first check if all the bits of the roots are demanded. If they're not,
3630 // we can truncate the roots to this narrower type.
3631 for (auto *Root : TreeRoot) {
3632 auto Mask = DB->getDemandedBits(cast<Instruction>(Root));
3633 MaxBitWidth = std::max<unsigned>(
3634 Mask.getBitWidth() - Mask.countLeadingZeros(), MaxBitWidth);
3635 }
3636
3637 // True if the roots can be zero-extended back to their original type, rather
3638 // than sign-extended. We know that if the leading bits are not demanded, we
3639 // can safely zero-extend. So we initialize IsKnownPositive to True.
3640 bool IsKnownPositive = true;
3641
3642 // If all the bits of the roots are demanded, we can try a little harder to
3643 // compute a narrower type. This can happen, for example, if the roots are
3644 // getelementptr indices. InstCombine promotes these indices to the pointer
3645 // width. Thus, all their bits are technically demanded even though the
3646 // address computation might be vectorized in a smaller type.
3647 //
3648 // We start by looking at each entry that can be demoted. We compute the
3649 // maximum bit width required to store the scalar by using ValueTracking to
3650 // compute the number of high-order bits we can truncate.
3651 if (MaxBitWidth == DL->getTypeSizeInBits(TreeRoot[0]->getType())) {
3652 MaxBitWidth = 8u;
3653
3654 // Determine if the sign bit of all the roots is known to be zero. If not,
3655 // IsKnownPositive is set to False.
3656 IsKnownPositive = all_of(TreeRoot, [&](Value *R) {
3657 bool KnownZero = false;
3658 bool KnownOne = false;
3659 ComputeSignBit(R, KnownZero, KnownOne, *DL);
3660 return KnownZero;
3661 });
3662
3663 // Determine the maximum number of bits required to store the scalar
3664 // values.
3665 for (auto *Scalar : ToDemote) {
3666 auto NumSignBits = ComputeNumSignBits(Scalar, *DL, 0, AC, 0, DT);
3667 auto NumTypeBits = DL->getTypeSizeInBits(Scalar->getType());
3668 MaxBitWidth = std::max<unsigned>(NumTypeBits - NumSignBits, MaxBitWidth);
3669 }
3670
3671 // If we can't prove that the sign bit is zero, we must add one to the
3672 // maximum bit width to account for the unknown sign bit. This preserves
3673 // the existing sign bit so we can safely sign-extend the root back to the
3674 // original type. Otherwise, if we know the sign bit is zero, we will
3675 // zero-extend the root instead.
3676 //
3677 // FIXME: This is somewhat suboptimal, as there will be cases where adding
3678 // one to the maximum bit width will yield a larger-than-necessary
3679 // type. In general, we need to add an extra bit only if we can't
3680 // prove that the upper bit of the original type is equal to the
3681 // upper bit of the proposed smaller type. If these two bits are the
3682 // same (either zero or one) we know that sign-extending from the
3683 // smaller type will result in the same value. Here, since we can't
3684 // yet prove this, we are just making the proposed smaller type
3685 // larger to ensure correctness.
3686 if (!IsKnownPositive)
3687 ++MaxBitWidth;
3688 }
3689
3690 // Round MaxBitWidth up to the next power-of-two.
3691 if (!isPowerOf2_64(MaxBitWidth))
3692 MaxBitWidth = NextPowerOf2(MaxBitWidth);
3693
3694 // If the maximum bit width we compute is less than the with of the roots'
3695 // type, we can proceed with the narrowing. Otherwise, do nothing.
3696 if (MaxBitWidth >= TreeRootIT->getBitWidth())
3697 return;
3698
3699 // If we can truncate the root, we must collect additional values that might
3700 // be demoted as a result. That is, those seeded by truncations we will
3701 // modify.
3702 while (!Roots.empty())
3703 collectValuesToDemote(Roots.pop_back_val(), Expr, ToDemote, Roots);
3704
3705 // Finally, map the values we can demote to the maximum bit with we computed.
3706 for (auto *Scalar : ToDemote)
3707 MinBWs[Scalar] = std::make_pair(MaxBitWidth, !IsKnownPositive);
3708}
3709
3710namespace {
3711/// The SLPVectorizer Pass.
3712struct SLPVectorizer : public FunctionPass {
3713 SLPVectorizerPass Impl;
3714
3715 /// Pass identification, replacement for typeid
3716 static char ID;
3717
3718 explicit SLPVectorizer() : FunctionPass(ID) {
3719 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
3720 }
3721
3722
3723 bool doInitialization(Module &M) override {
3724 return false;
3725 }
3726
3727 bool runOnFunction(Function &F) override {
3728 if (skipFunction(F))
3729 return false;
3730
3731 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
3732 auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
3733 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
3734 auto *TLI = TLIP ? &TLIP->getTLI() : nullptr;
3735 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
3736 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
3737 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3738 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3739 auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits();
3740
3741 return Impl.runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB);
3742 }
3743
3744 void getAnalysisUsage(AnalysisUsage &AU) const override {
3745 FunctionPass::getAnalysisUsage(AU);
3746 AU.addRequired<AssumptionCacheTracker>();
3747 AU.addRequired<ScalarEvolutionWrapperPass>();
3748 AU.addRequired<AAResultsWrapperPass>();
3749 AU.addRequired<TargetTransformInfoWrapperPass>();
3750 AU.addRequired<LoopInfoWrapperPass>();
3751 AU.addRequired<DominatorTreeWrapperPass>();
3752 AU.addRequired<DemandedBitsWrapperPass>();
3753 AU.addPreserved<LoopInfoWrapperPass>();
3754 AU.addPreserved<DominatorTreeWrapperPass>();
3755 AU.addPreserved<AAResultsWrapperPass>();
3756 AU.addPreserved<GlobalsAAWrapperPass>();
3757 AU.setPreservesCFG();
3758 }
3759};
3760} // end anonymous namespace
3761
3762PreservedAnalyses SLPVectorizerPass::run(Function &F, FunctionAnalysisManager &AM) {
3763 auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F);
3764 auto *TTI = &AM.getResult<TargetIRAnalysis>(F);
3765 auto *TLI = AM.getCachedResult<TargetLibraryAnalysis>(F);
3766 auto *AA = &AM.getResult<AAManager>(F);
3767 auto *LI = &AM.getResult<LoopAnalysis>(F);
3768 auto *DT = &AM.getResult<DominatorTreeAnalysis>(F);
3769 auto *AC = &AM.getResult<AssumptionAnalysis>(F);
3770 auto *DB = &AM.getResult<DemandedBitsAnalysis>(F);
3771
3772 bool Changed = runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB);
3773 if (!Changed)
3774 return PreservedAnalyses::all();
3775
3776 PreservedAnalyses PA;
3777 PA.preserveSet<CFGAnalyses>();
3778 PA.preserve<AAManager>();
3779 PA.preserve<GlobalsAA>();
3780 return PA;
3781}
3782
3783bool SLPVectorizerPass::runImpl(Function &F, ScalarEvolution *SE_,
3784 TargetTransformInfo *TTI_,
3785 TargetLibraryInfo *TLI_, AliasAnalysis *AA_,
3786 LoopInfo *LI_, DominatorTree *DT_,
3787 AssumptionCache *AC_, DemandedBits *DB_) {
3788 SE = SE_;
3789 TTI = TTI_;
3790 TLI = TLI_;
3791 AA = AA_;
3792 LI = LI_;
3793 DT = DT_;
3794 AC = AC_;
3795 DB = DB_;
3796 DL = &F.getParent()->getDataLayout();
3797
3798 Stores.clear();
3799 GEPs.clear();
3800 bool Changed = false;
3801
3802 // If the target claims to have no vector registers don't attempt
3803 // vectorization.
3804 if (!TTI->getNumberOfRegisters(true))
3805 return false;
3806
3807 // Don't vectorize when the attribute NoImplicitFloat is used.
3808 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
3809 return false;
3810
3811 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing blocks in " <<
F.getName() << ".\n"; } } while (false)
;
3812
3813 // Use the bottom up slp vectorizer to construct chains that start with
3814 // store instructions.
3815 BoUpSLP R(&F, SE, TTI, TLI, AA, LI, DT, AC, DB, DL);
3816
3817 // A general note: the vectorizer must use BoUpSLP::eraseInstruction() to
3818 // delete instructions.
3819
3820 // Scan the blocks in the function in post order.
3821 for (auto BB : post_order(&F.getEntryBlock())) {
3822 collectSeedInstructions(BB);
3823
3824 // Vectorize trees that end at stores.
3825 if (!Stores.empty()) {
3826 DEBUG(dbgs() << "SLP: Found stores for " << Stores.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Found stores for " << Stores
.size() << " underlying objects.\n"; } } while (false)
3827 << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Found stores for " << Stores
.size() << " underlying objects.\n"; } } while (false)
;
3828 Changed |= vectorizeStoreChains(R);
3829 }
3830
3831 // Vectorize trees that end at reductions.
3832 Changed |= vectorizeChainsInBlock(BB, R);
3833
3834 // Vectorize the index computations of getelementptr instructions. This
3835 // is primarily intended to catch gather-like idioms ending at
3836 // non-consecutive loads.
3837 if (!GEPs.empty()) {
3838 DEBUG(dbgs() << "SLP: Found GEPs for " << GEPs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs
.size() << " underlying objects.\n"; } } while (false)
3839 << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs
.size() << " underlying objects.\n"; } } while (false)
;
3840 Changed |= vectorizeGEPIndices(BB, R);
3841 }
3842 }
3843
3844 if (Changed) {
3845 R.optimizeGatherSequence();
3846 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: vectorized \"" << F.getName
() << "\"\n"; } } while (false)
;
3847 DEBUG(verifyFunction(F))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { verifyFunction(F); } } while (false)
;
3848 }
3849 return Changed;
3850}
3851
3852/// \brief Check that the Values in the slice in VL array are still existent in
3853/// the WeakVH array.
3854/// Vectorization of part of the VL array may cause later values in the VL array
3855/// to become invalid. We track when this has happened in the WeakVH array.
3856static bool hasValueBeenRAUWed(ArrayRef<Value *> VL, ArrayRef<WeakVH> VH,
3857 unsigned SliceBegin, unsigned SliceSize) {
3858 VL = VL.slice(SliceBegin, SliceSize);
3859 VH = VH.slice(SliceBegin, SliceSize);
3860 return !std::equal(VL.begin(), VL.end(), VH.begin());
3861}
3862
3863bool SLPVectorizerPass::vectorizeStoreChain(ArrayRef<Value *> Chain, BoUpSLP &R,
3864 unsigned VecRegSize) {
3865 unsigned ChainLen = Chain.size();
3866 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLendo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing a store chain of length "
<< ChainLen << "\n"; } } while (false)
3867 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing a store chain of length "
<< ChainLen << "\n"; } } while (false)
;
3868 unsigned Sz = R.getVectorElementSize(Chain[0]);
3869 unsigned VF = VecRegSize / Sz;
3870
3871 if (!isPowerOf2_32(Sz) || VF < 2)
3872 return false;
3873
3874 // Keep track of values that were deleted by vectorizing in the loop below.
3875 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end());
3876
3877 bool Changed = false;
3878 // Look for profitable vectorizable trees at all offsets, starting at zero.
3879 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
3880 if (i + VF > e)
3881 break;
3882
3883 // Check that a previous iteration of this loop did not delete the Value.
3884 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF))
3885 continue;
3886
3887 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing " << VF <<
" stores at offset " << i << "\n"; } } while (false
)
3888 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing " << VF <<
" stores at offset " << i << "\n"; } } while (false
)
;
3889 ArrayRef<Value *> Operands = Chain.slice(i, VF);
3890
3891 R.buildTree(Operands);
3892 if (R.isTreeTinyAndNotFullyVectorizable())
3893 continue;
3894
3895 R.computeMinimumValueSizes();
3896
3897 int Cost = R.getTreeCost();
3898
3899 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Found cost=" << Cost <<
" for VF=" << VF << "\n"; } } while (false)
;
3900 if (Cost < -SLPCostThreshold) {
3901 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Decided to vectorize cost=" <<
Cost << "\n"; } } while (false)
;
3902 R.vectorizeTree();
3903
3904 // Move to the next bundle.
3905 i += VF - 1;
3906 Changed = true;
3907 }
3908 }
3909
3910 return Changed;
3911}
3912
3913bool SLPVectorizerPass::vectorizeStores(ArrayRef<StoreInst *> Stores,
3914 BoUpSLP &R) {
3915 SetVector<StoreInst *> Heads, Tails;
3916 SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain;
3917
3918 // We may run into multiple chains that merge into a single chain. We mark the
3919 // stores that we vectorized so that we don't visit the same store twice.
3920 BoUpSLP::ValueSet VectorizedStores;
3921 bool Changed = false;
3922
3923 // Do a quadratic search on all of the given stores and find
3924 // all of the pairs of stores that follow each other.
3925 SmallVector<unsigned, 16> IndexQueue;
3926 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
3927 IndexQueue.clear();
3928 // If a store has multiple consecutive store candidates, search Stores
3929 // array according to the sequence: from i+1 to e, then from i-1 to 0.
3930 // This is because usually pairing with immediate succeeding or preceding
3931 // candidate create the best chance to find slp vectorization opportunity.
3932 unsigned j = 0;
3933 for (j = i + 1; j < e; ++j)
3934 IndexQueue.push_back(j);
3935 for (j = i; j > 0; --j)
3936 IndexQueue.push_back(j - 1);
3937
3938 for (auto &k : IndexQueue) {
3939 if (isConsecutiveAccess(Stores[i], Stores[k], *DL, *SE)) {
3940 Tails.insert(Stores[k]);
3941 Heads.insert(Stores[i]);
3942 ConsecutiveChain[Stores[i]] = Stores[k];
3943 break;
3944 }
3945 }
3946 }
3947
3948 // For stores that start but don't end a link in the chain:
3949 for (SetVector<StoreInst *>::iterator it = Heads.begin(), e = Heads.end();
3950 it != e; ++it) {
3951 if (Tails.count(*it))
3952 continue;
3953
3954 // We found a store instr that starts a chain. Now follow the chain and try
3955 // to vectorize it.
3956 BoUpSLP::ValueList Operands;
3957 StoreInst *I = *it;
3958 // Collect the chain into a list.
3959 while (Tails.count(I) || Heads.count(I)) {
3960 if (VectorizedStores.count(I))
3961 break;
3962 Operands.push_back(I);
3963 // Move to the next value in the chain.
3964 I = ConsecutiveChain[I];
3965 }
3966
3967 // FIXME: Is division-by-2 the correct step? Should we assert that the
3968 // register size is a power-of-2?
3969 for (unsigned Size = R.getMaxVecRegSize(); Size >= R.getMinVecRegSize();
3970 Size /= 2) {
3971 if (vectorizeStoreChain(Operands, R, Size)) {
3972 // Mark the vectorized stores so that we don't vectorize them again.
3973 VectorizedStores.insert(Operands.begin(), Operands.end());
3974 Changed = true;
3975 break;
3976 }
3977 }
3978 }
3979
3980 return Changed;
3981}
3982
3983void SLPVectorizerPass::collectSeedInstructions(BasicBlock *BB) {
3984
3985 // Initialize the collections. We will make a single pass over the block.
3986 Stores.clear();
3987 GEPs.clear();
3988
3989 // Visit the store and getelementptr instructions in BB and organize them in
3990 // Stores and GEPs according to the underlying objects of their pointer
3991 // operands.
3992 for (Instruction &I : *BB) {
3993
3994 // Ignore store instructions that are volatile or have a pointer operand
3995 // that doesn't point to a scalar type.
3996 if (auto *SI = dyn_cast<StoreInst>(&I)) {
3997 if (!SI->isSimple())
3998 continue;
3999 if (!isValidElementType(SI->getValueOperand()->getType()))
4000 continue;
4001 Stores[GetUnderlyingObject(SI->getPointerOperand(), *DL)].push_back(SI);
4002 }
4003
4004 // Ignore getelementptr instructions that have more than one index, a
4005 // constant index, or a pointer operand that doesn't point to a scalar
4006 // type.
4007 else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
4008 auto Idx = GEP->idx_begin()->get();
4009 if (GEP->getNumIndices() > 1 || isa<Constant>(Idx))
4010 continue;
4011 if (!isValidElementType(Idx->getType()))
4012 continue;
4013 if (GEP->getType()->isVectorTy())
4014 continue;
4015 GEPs[GetUnderlyingObject(GEP->getPointerOperand(), *DL)].push_back(GEP);
4016 }
4017 }
4018}
4019
4020bool SLPVectorizerPass::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
4021 if (!A || !B)
4022 return false;
4023 Value *VL[] = { A, B };
4024 return tryToVectorizeList(VL, R, None, true);
4025}
4026
4027bool SLPVectorizerPass::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
4028 ArrayRef<Value *> BuildVector,
4029 bool AllowReorder) {
4030 if (VL.size() < 2)
4031 return false;
4032
4033 DEBUG(dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = "
<< VL.size() << ".\n"; } } while (false)
4034 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = "
<< VL.size() << ".\n"; } } while (false)
;
4035
4036 // Check that all of the parts are scalar instructions of the same type.
4037 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
4038 if (!I0)
4039 return false;
4040
4041 unsigned Opcode0 = I0->getOpcode();
4042
4043 unsigned Sz = R.getVectorElementSize(I0);
4044 unsigned MinVF = std::max(2U, R.getMinVecRegSize() / Sz);
4045 unsigned MaxVF = std::max<unsigned>(PowerOf2Floor(VL.size()), MinVF);
4046 if (MaxVF < 2)
4047 return false;
4048
4049 for (Value *V : VL) {
4050 Type *Ty = V->getType();
4051 if (!isValidElementType(Ty))
4052 return false;
4053 Instruction *Inst = dyn_cast<Instruction>(V);
4054 if (!Inst || Inst->getOpcode() != Opcode0)
4055 return false;
4056 }
4057
4058 bool Changed = false;
4059
4060 // Keep track of values that were deleted by vectorizing in the loop below.
4061 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end());
4062
4063 unsigned NextInst = 0, MaxInst = VL.size();
4064 for (unsigned VF = MaxVF; NextInst + 1 < MaxInst && VF >= MinVF;
4065 VF /= 2) {
4066 // No actual vectorization should happen, if number of parts is the same as
4067 // provided vectorization factor (i.e. the scalar type is used for vector
4068 // code during codegen).
4069 auto *VecTy = VectorType::get(VL[0]->getType(), VF);
4070 if (TTI->getNumberOfParts(VecTy) == VF)
4071 continue;
4072 for (unsigned I = NextInst; I < MaxInst; ++I) {
4073 unsigned OpsWidth = 0;
4074
4075 if (I + VF > MaxInst)
4076 OpsWidth = MaxInst - I;
4077 else
4078 OpsWidth = VF;
4079
4080 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
4081 break;
4082
4083 // Check that a previous iteration of this loop did not delete the Value.
4084 if (hasValueBeenRAUWed(VL, TrackValues, I, OpsWidth))
4085 continue;
4086
4087 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth
<< " operations " << "\n"; } } while (false)
4088 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth
<< " operations " << "\n"; } } while (false)
;
4089 ArrayRef<Value *> Ops = VL.slice(I, OpsWidth);
4090
4091 ArrayRef<Value *> BuildVectorSlice;
4092 if (!BuildVector.empty())
4093 BuildVectorSlice = BuildVector.slice(I, OpsWidth);
4094
4095 R.buildTree(Ops, BuildVectorSlice);
4096 // TODO: check if we can allow reordering for more cases.
4097 if (AllowReorder && R.shouldReorder()) {
4098 // Conceptually, there is nothing actually preventing us from trying to
4099 // reorder a larger list. In fact, we do exactly this when vectorizing
4100 // reductions. However, at this point, we only expect to get here from
4101 // tryToVectorizePair().
4102 assert(Ops.size() == 2)((Ops.size() == 2) ? static_cast<void> (0) : __assert_fail
("Ops.size() == 2", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4102, __PRETTY_FUNCTION__))
;
4103 assert(BuildVectorSlice.empty())((BuildVectorSlice.empty()) ? static_cast<void> (0) : __assert_fail
("BuildVectorSlice.empty()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4103, __PRETTY_FUNCTION__))
;
4104 Value *ReorderedOps[] = {Ops[1], Ops[0]};
4105 R.buildTree(ReorderedOps, None);
4106 }
4107 if (R.isTreeTinyAndNotFullyVectorizable())
4108 continue;
4109
4110 R.computeMinimumValueSizes();
4111 int Cost = R.getTreeCost();
4112
4113 if (Cost < -SLPCostThreshold) {
4114 DEBUG(dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Vectorizing list at cost:" <<
Cost << ".\n"; } } while (false)
;
4115 Value *VectorizedRoot = R.vectorizeTree();
4116
4117 // Reconstruct the build vector by extracting the vectorized root. This
4118 // way we handle the case where some elements of the vector are
4119 // undefined.
4120 // (return (inserelt <4 xi32> (insertelt undef (opd0) 0) (opd1) 2))
4121 if (!BuildVectorSlice.empty()) {
4122 // The insert point is the last build vector instruction. The
4123 // vectorized root will precede it. This guarantees that we get an
4124 // instruction. The vectorized tree could have been constant folded.
4125 Instruction *InsertAfter = cast<Instruction>(BuildVectorSlice.back());
4126 unsigned VecIdx = 0;
4127 for (auto &V : BuildVectorSlice) {
4128 IRBuilder<NoFolder> Builder(InsertAfter->getParent(),
4129 ++BasicBlock::iterator(InsertAfter));
4130 Instruction *I = cast<Instruction>(V);
4131 assert(isa<InsertElementInst>(I) || isa<InsertValueInst>(I))((isa<InsertElementInst>(I) || isa<InsertValueInst>
(I)) ? static_cast<void> (0) : __assert_fail ("isa<InsertElementInst>(I) || isa<InsertValueInst>(I)"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4131, __PRETTY_FUNCTION__))
;
4132 Instruction *Extract =
4133 cast<Instruction>(Builder.CreateExtractElement(
4134 VectorizedRoot, Builder.getInt32(VecIdx++)));
4135 I->setOperand(1, Extract);
4136 I->removeFromParent();
4137 I->insertAfter(Extract);
4138 InsertAfter = I;
4139 }
4140 }
4141 // Move to the next bundle.
4142 I += VF - 1;
4143 NextInst = I + 1;
4144 Changed = true;
4145 }
4146 }
4147 }
4148
4149 return Changed;
4150}
4151
4152bool SLPVectorizerPass::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
4153 if (!V)
4154 return false;
4155
4156 Value *P = V->getParent();
4157
4158 // Vectorize in current basic block only.
4159 auto *Op0 = dyn_cast<Instruction>(V->getOperand(0));
4160 auto *Op1 = dyn_cast<Instruction>(V->getOperand(1));
4161 if (!Op0 || !Op1 || Op0->getParent() != P || Op1->getParent() != P)
4162 return false;
4163
4164 // Try to vectorize V.
4165 if (tryToVectorizePair(Op0, Op1, R))
4166 return true;
4167
4168 auto *A = dyn_cast<BinaryOperator>(Op0);
4169 auto *B = dyn_cast<BinaryOperator>(Op1);
4170 // Try to skip B.
4171 if (B && B->hasOneUse()) {
4172 auto *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
4173 auto *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
4174 if (B0 && B0->getParent() == P && tryToVectorizePair(A, B0, R))
4175 return true;
4176 if (B1 && B1->getParent() == P && tryToVectorizePair(A, B1, R))
4177 return true;
4178 }
4179
4180 // Try to skip A.
4181 if (A && A->hasOneUse()) {
4182 auto *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
4183 auto *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
4184 if (A0 && A0->getParent() == P && tryToVectorizePair(A0, B, R))
4185 return true;
4186 if (A1 && A1->getParent() == P && tryToVectorizePair(A1, B, R))
4187 return true;
4188 }
4189 return false;
4190}
4191
4192/// \brief Generate a shuffle mask to be used in a reduction tree.
4193///
4194/// \param VecLen The length of the vector to be reduced.
4195/// \param NumEltsToRdx The number of elements that should be reduced in the
4196/// vector.
4197/// \param IsPairwise Whether the reduction is a pairwise or splitting
4198/// reduction. A pairwise reduction will generate a mask of
4199/// <0,2,...> or <1,3,..> while a splitting reduction will generate
4200/// <2,3, undef,undef> for a vector of 4 and NumElts = 2.
4201/// \param IsLeft True will generate a mask of even elements, odd otherwise.
4202static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx,
4203 bool IsPairwise, bool IsLeft,
4204 IRBuilder<> &Builder) {
4205 assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask")(((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask"
) ? static_cast<void> (0) : __assert_fail ("(IsPairwise || !IsLeft) && \"Don't support a <0,1,undef,...> mask\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4205, __PRETTY_FUNCTION__))
;
4206
4207 SmallVector<Constant *, 32> ShuffleMask(
4208 VecLen, UndefValue::get(Builder.getInt32Ty()));
4209
4210 if (IsPairwise)
4211 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right).
4212 for (unsigned i = 0; i != NumEltsToRdx; ++i)
4213 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft);
4214 else
4215 // Move the upper half of the vector to the lower half.
4216 for (unsigned i = 0; i != NumEltsToRdx; ++i)
4217 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i);
4218
4219 return ConstantVector::get(ShuffleMask);
4220}
4221
4222namespace {
4223/// Model horizontal reductions.
4224///
4225/// A horizontal reduction is a tree of reduction operations (currently add and
4226/// fadd) that has operations that can be put into a vector as its leaf.
4227/// For example, this tree:
4228///
4229/// mul mul mul mul
4230/// \ / \ /
4231/// + +
4232/// \ /
4233/// +
4234/// This tree has "mul" as its reduced values and "+" as its reduction
4235/// operations. A reduction might be feeding into a store or a binary operation
4236/// feeding a phi.
4237/// ...
4238/// \ /
4239/// +
4240/// |
4241/// phi +=
4242///
4243/// Or:
4244/// ...
4245/// \ /
4246/// +
4247/// |
4248/// *p =
4249///
4250class HorizontalReduction {
4251 SmallVector<Value *, 16> ReductionOps;
4252 SmallVector<Value *, 32> ReducedVals;
4253 // Use map vector to make stable output.
4254 MapVector<Instruction *, Value *> ExtraArgs;
4255
4256 BinaryOperator *ReductionRoot = nullptr;
4257 // After successfull horizontal reduction vectorization attempt for PHI node
4258 // vectorizer tries to update root binary op by combining vectorized tree and
4259 // the ReductionPHI node. But during vectorization this ReductionPHI can be
4260 // vectorized itself and replaced by the undef value, while the instruction
4261 // itself is marked for deletion. This 'marked for deletion' PHI node then can
4262 // be used in new binary operation, causing "Use still stuck around after Def
4263 // is destroyed" crash upon PHI node deletion.
4264 WeakVH ReductionPHI;
4265
4266 /// The opcode of the reduction.
4267 Instruction::BinaryOps ReductionOpcode = Instruction::BinaryOpsEnd;
4268 /// The opcode of the values we perform a reduction on.
4269 unsigned ReducedValueOpcode = 0;
4270 /// Should we model this reduction as a pairwise reduction tree or a tree that
4271 /// splits the vector in halves and adds those halves.
4272 bool IsPairwiseReduction = false;
4273
4274 /// Checks if the ParentStackElem.first should be marked as a reduction
4275 /// operation with an extra argument or as extra argument itself.
4276 void markExtraArg(std::pair<Instruction *, unsigned> &ParentStackElem,
4277 Value *ExtraArg) {
4278 if (ExtraArgs.count(ParentStackElem.first)) {
4279 ExtraArgs[ParentStackElem.first] = nullptr;
4280 // We ran into something like:
4281 // ParentStackElem.first = ExtraArgs[ParentStackElem.first] + ExtraArg.
4282 // The whole ParentStackElem.first should be considered as an extra value
4283 // in this case.
4284 // Do not perform analysis of remaining operands of ParentStackElem.first
4285 // instruction, this whole instruction is an extra argument.
4286 ParentStackElem.second = ParentStackElem.first->getNumOperands();
4287 } else {
4288 // We ran into something like:
4289 // ParentStackElem.first += ... + ExtraArg + ...
4290 ExtraArgs[ParentStackElem.first] = ExtraArg;
4291 }
4292 }
4293
4294public:
4295 HorizontalReduction() = default;
4296
4297 /// \brief Try to find a reduction tree.
4298 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B) {
4299 assert((!Phi || is_contained(Phi->operands(), B)) &&(((!Phi || is_contained(Phi->operands(), B)) && "Thi phi needs to use the binary operator"
) ? static_cast<void> (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), B)) && \"Thi phi needs to use the binary operator\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4300, __PRETTY_FUNCTION__))
4300 "Thi phi needs to use the binary operator")(((!Phi || is_contained(Phi->operands(), B)) && "Thi phi needs to use the binary operator"
) ? static_cast<void> (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), B)) && \"Thi phi needs to use the binary operator\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4300, __PRETTY_FUNCTION__))
;
4301
4302 // We could have a initial reductions that is not an add.
4303 // r *= v1 + v2 + v3 + v4
4304 // In such a case start looking for a tree rooted in the first '+'.
4305 if (Phi) {
4306 if (B->getOperand(0) == Phi) {
4307 Phi = nullptr;
4308 B = dyn_cast<BinaryOperator>(B->getOperand(1));
4309 } else if (B->getOperand(1) == Phi) {
4310 Phi = nullptr;
4311 B = dyn_cast<BinaryOperator>(B->getOperand(0));
4312 }
4313 }
4314
4315 if (!B)
4316 return false;
4317
4318 Type *Ty = B->getType();
4319 if (!isValidElementType(Ty))
4320 return false;
4321
4322 ReductionOpcode = B->getOpcode();
4323 ReducedValueOpcode = 0;
4324 ReductionRoot = B;
4325 ReductionPHI = Phi;
4326
4327 // We currently only support adds.
4328 if ((ReductionOpcode != Instruction::Add &&
4329 ReductionOpcode != Instruction::FAdd) ||
4330 !B->isAssociative())
4331 return false;
4332
4333 // Post order traverse the reduction tree starting at B. We only handle true
4334 // trees containing only binary operators or selects.
4335 SmallVector<std::pair<Instruction *, unsigned>, 32> Stack;
4336 Stack.push_back(std::make_pair(B, 0));
4337 while (!Stack.empty()) {
4338 Instruction *TreeN = Stack.back().first;
4339 unsigned EdgeToVist = Stack.back().second++;
4340 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode;
4341
4342 // Postorder vist.
4343 if (EdgeToVist == 2 || IsReducedValue) {
4344 if (IsReducedValue)
4345 ReducedVals.push_back(TreeN);
4346 else {
4347 auto I = ExtraArgs.find(TreeN);
4348 if (I != ExtraArgs.end() && !I->second) {
4349 // Check if TreeN is an extra argument of its parent operation.
4350 if (Stack.size() <= 1) {
4351 // TreeN can't be an extra argument as it is a root reduction
4352 // operation.
4353 return false;
4354 }
4355 // Yes, TreeN is an extra argument, do not add it to a list of
4356 // reduction operations.
4357 // Stack[Stack.size() - 2] always points to the parent operation.
4358 markExtraArg(Stack[Stack.size() - 2], TreeN);
4359 ExtraArgs.erase(TreeN);
4360 } else
4361 ReductionOps.push_back(TreeN);
4362 }
4363 // Retract.
4364 Stack.pop_back();
4365 continue;
4366 }
4367
4368 // Visit left or right.
4369 Value *NextV = TreeN->getOperand(EdgeToVist);
4370 if (NextV != Phi) {
4371 auto *I = dyn_cast<Instruction>(NextV);
4372 // Continue analysis if the next operand is a reduction operation or
4373 // (possibly) a reduced value. If the reduced value opcode is not set,
4374 // the first met operation != reduction operation is considered as the
4375 // reduced value class.
4376 if (I && (!ReducedValueOpcode || I->getOpcode() == ReducedValueOpcode ||
4377 I->getOpcode() == ReductionOpcode)) {
4378 // Only handle trees in the current basic block.
4379 if (I->getParent() != B->getParent()) {
4380 // I is an extra argument for TreeN (its parent operation).
4381 markExtraArg(Stack.back(), I);
4382 continue;
4383 }
4384
4385 // Each tree node needs to have one user except for the ultimate
4386 // reduction.
4387 if (!I->hasOneUse() && I != B) {
4388 // I is an extra argument for TreeN (its parent operation).
4389 markExtraArg(Stack.back(), I);
4390 continue;
4391 }
4392
4393 if (I->getOpcode() == ReductionOpcode) {
4394 // We need to be able to reassociate the reduction operations.
4395 if (!I->isAssociative()) {
4396 // I is an extra argument for TreeN (its parent operation).
4397 markExtraArg(Stack.back(), I);
4398 continue;
4399 }
4400 } else if (ReducedValueOpcode &&
4401 ReducedValueOpcode != I->getOpcode()) {
4402 // Make sure that the opcodes of the operations that we are going to
4403 // reduce match.
4404 // I is an extra argument for TreeN (its parent operation).
4405 markExtraArg(Stack.back(), I);
4406 continue;
4407 } else if (!ReducedValueOpcode)
4408 ReducedValueOpcode = I->getOpcode();
4409
4410 Stack.push_back(std::make_pair(I, 0));
4411 continue;
4412 }
4413 // NextV is an extra argument for TreeN (its parent operation).
4414 markExtraArg(Stack.back(), NextV);
4415 }
4416 }
4417 return true;
4418 }
4419
4420 /// \brief Attempt to vectorize the tree found by
4421 /// matchAssociativeReduction.
4422 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
4423 if (ReducedVals.empty())
4424 return false;
4425
4426 // If there is a sufficient number of reduction values, reduce
4427 // to a nearby power-of-2. Can safely generate oversized
4428 // vectors and rely on the backend to split them to legal sizes.
4429 unsigned NumReducedVals = ReducedVals.size();
4430 if (NumReducedVals < 4)
4431 return false;
4432
4433 unsigned ReduxWidth = PowerOf2Floor(NumReducedVals);
4434
4435 Value *VectorizedTree = nullptr;
4436 IRBuilder<> Builder(ReductionRoot);
4437 FastMathFlags Unsafe;
4438 Unsafe.setUnsafeAlgebra();
4439 Builder.setFastMathFlags(Unsafe);
4440 unsigned i = 0;
4441
4442 MapVector<Value *, DebugLoc> ExternallyUsedValues;
4443 for (auto &Pair : ExtraArgs)
4444 ExternallyUsedValues[Pair.second] = Pair.first->getDebugLoc();
4445 while (i < NumReducedVals - ReduxWidth + 1 && ReduxWidth > 2) {
4446 auto VL = makeArrayRef(&ReducedVals[i], ReduxWidth);
4447 V.buildTree(VL, ExternallyUsedValues, ReductionOps);
4448 if (V.shouldReorder()) {
4449 SmallVector<Value *, 8> Reversed(VL.rbegin(), VL.rend());
4450 V.buildTree(Reversed, ExternallyUsedValues, ReductionOps);
4451 }
4452 if (V.isTreeTinyAndNotFullyVectorizable())
4453 break;
4454
4455 V.computeMinimumValueSizes();
4456
4457 // Estimate cost.
4458 int Cost =
4459 V.getTreeCost() + getReductionCost(TTI, ReducedVals[i], ReduxWidth);
4460 if (Cost >= -SLPCostThreshold)
4461 break;
4462
4463 DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Costdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:"
<< Cost << ". (HorRdx)\n"; } } while (false)
4464 << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:"
<< Cost << ". (HorRdx)\n"; } } while (false)
;
4465
4466 // Vectorize a tree.
4467 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc();
4468 Value *VectorizedRoot = V.vectorizeTree(ExternallyUsedValues);
4469
4470 // Emit a reduction.
4471 Value *ReducedSubTree =
4472 emitReduction(VectorizedRoot, Builder, ReduxWidth);
4473 if (VectorizedTree) {
4474 Builder.SetCurrentDebugLocation(Loc);
4475 VectorizedTree = Builder.CreateBinOp(ReductionOpcode, VectorizedTree,
4476 ReducedSubTree, "bin.rdx");
4477 } else
4478 VectorizedTree = ReducedSubTree;
4479 i += ReduxWidth;
4480 ReduxWidth = PowerOf2Floor(NumReducedVals - i);
4481 }
4482
4483 if (VectorizedTree) {
4484 // Finish the reduction.
4485 for (; i < NumReducedVals; ++i) {
4486 auto *I = cast<Instruction>(ReducedVals[i]);
4487 Builder.SetCurrentDebugLocation(I->getDebugLoc());
4488 VectorizedTree =
4489 Builder.CreateBinOp(ReductionOpcode, VectorizedTree, I);
4490 }
4491 for (auto &Pair : ExternallyUsedValues) {
4492 Builder.SetCurrentDebugLocation(Pair.second);
4493 VectorizedTree = Builder.CreateBinOp(ReductionOpcode, VectorizedTree,
4494 Pair.first, "bin.extra");
4495 }
4496 // Update users.
4497 if (ReductionPHI && !isa<UndefValue>(ReductionPHI)) {
4498 assert(ReductionRoot && "Need a reduction operation")((ReductionRoot && "Need a reduction operation") ? static_cast
<void> (0) : __assert_fail ("ReductionRoot && \"Need a reduction operation\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4498, __PRETTY_FUNCTION__))
;
4499 ReductionRoot->setOperand(0, VectorizedTree);
4500 ReductionRoot->setOperand(1, ReductionPHI);
4501 } else
4502 ReductionRoot->replaceAllUsesWith(VectorizedTree);
4503 }
4504 return VectorizedTree != nullptr;
4505 }
4506
4507 unsigned numReductionValues() const {
4508 return ReducedVals.size();
4509 }
4510
4511private:
4512 /// \brief Calculate the cost of a reduction.
4513 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal,
4514 unsigned ReduxWidth) {
4515 Type *ScalarTy = FirstReducedVal->getType();
4516 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth);
4517
4518 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true);
4519 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false);
4520
4521 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost;
4522 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost;
4523
4524 int ScalarReduxCost =
4525 (ReduxWidth - 1) *
4526 TTI->getArithmeticInstrCost(ReductionOpcode, ScalarTy);
4527
4528 DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost
- ScalarReduxCost << " for reduction that starts with "
<< *FirstReducedVal << " (It is a " << (IsPairwiseReduction
? "pairwise" : "splitting") << " reduction)\n"; } } while
(false)
4529 << " for reduction that starts with " << *FirstReducedValdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost
- ScalarReduxCost << " for reduction that starts with "
<< *FirstReducedVal << " (It is a " << (IsPairwiseReduction
? "pairwise" : "splitting") << " reduction)\n"; } } while
(false)
4530 << " (It is a "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost
- ScalarReduxCost << " for reduction that starts with "
<< *FirstReducedVal << " (It is a " << (IsPairwiseReduction
? "pairwise" : "splitting") << " reduction)\n"; } } while
(false)
4531 << (IsPairwiseReduction ? "pairwise" : "splitting")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost
- ScalarReduxCost << " for reduction that starts with "
<< *FirstReducedVal << " (It is a " << (IsPairwiseReduction
? "pairwise" : "splitting") << " reduction)\n"; } } while
(false)
4532 << " reduction)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Adding cost " << VecReduxCost
- ScalarReduxCost << " for reduction that starts with "
<< *FirstReducedVal << " (It is a " << (IsPairwiseReduction
? "pairwise" : "splitting") << " reduction)\n"; } } while
(false)
;
4533
4534 return VecReduxCost - ScalarReduxCost;
4535 }
4536
4537 /// \brief Emit a horizontal reduction of the vectorized value.
4538 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder,
4539 unsigned ReduxWidth) {
4540 assert(VectorizedValue && "Need to have a vectorized tree node")((VectorizedValue && "Need to have a vectorized tree node"
) ? static_cast<void> (0) : __assert_fail ("VectorizedValue && \"Need to have a vectorized tree node\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4540, __PRETTY_FUNCTION__))
;
4541 assert(isPowerOf2_32(ReduxWidth) &&((isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now"
) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4542, __PRETTY_FUNCTION__))
4542 "We only handle power-of-two reductions for now")((isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now"
) ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4542, __PRETTY_FUNCTION__))
;
4543
4544 Value *TmpVec = VectorizedValue;
4545 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) {
4546 if (IsPairwiseReduction) {
4547 Value *LeftMask =
4548 createRdxShuffleMask(ReduxWidth, i, true, true, Builder);
4549 Value *RightMask =
4550 createRdxShuffleMask(ReduxWidth, i, true, false, Builder);
4551
4552 Value *LeftShuf = Builder.CreateShuffleVector(
4553 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l");
4554 Value *RightShuf = Builder.CreateShuffleVector(
4555 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask),
4556 "rdx.shuf.r");
4557 TmpVec = Builder.CreateBinOp(ReductionOpcode, LeftShuf, RightShuf,
4558 "bin.rdx");
4559 } else {
4560 Value *UpperHalf =
4561 createRdxShuffleMask(ReduxWidth, i, false, false, Builder);
4562 Value *Shuf = Builder.CreateShuffleVector(
4563 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf");
4564 TmpVec = Builder.CreateBinOp(ReductionOpcode, TmpVec, Shuf, "bin.rdx");
4565 }
4566 }
4567
4568 // The result is in the first element of the vector.
4569 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
4570 }
4571};
4572} // end anonymous namespace
4573
4574/// \brief Recognize construction of vectors like
4575/// %ra = insertelement <4 x float> undef, float %s0, i32 0
4576/// %rb = insertelement <4 x float> %ra, float %s1, i32 1
4577/// %rc = insertelement <4 x float> %rb, float %s2, i32 2
4578/// %rd = insertelement <4 x float> %rc, float %s3, i32 3
4579///
4580/// Returns true if it matches
4581///
4582static bool findBuildVector(InsertElementInst *FirstInsertElem,
4583 SmallVectorImpl<Value *> &BuildVector,
4584 SmallVectorImpl<Value *> &BuildVectorOpds) {
4585 if (!isa<UndefValue>(FirstInsertElem->getOperand(0)))
4586 return false;
4587
4588 InsertElementInst *IE = FirstInsertElem;
4589 while (true) {
4590 BuildVector.push_back(IE);
4591 BuildVectorOpds.push_back(IE->getOperand(1));
4592
4593 if (IE->use_empty())
4594 return false;
4595
4596 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->user_back());
4597 if (!NextUse)
4598 return true;
4599
4600 // If this isn't the final use, make sure the next insertelement is the only
4601 // use. It's OK if the final constructed vector is used multiple times
4602 if (!IE->hasOneUse())
4603 return false;
4604
4605 IE = NextUse;
4606 }
4607
4608 return false;
4609}
4610
4611/// \brief Like findBuildVector, but looks backwards for construction of aggregate.
4612///
4613/// \return true if it matches.
4614static bool findBuildAggregate(InsertValueInst *IV,
4615 SmallVectorImpl<Value *> &BuildVector,
4616 SmallVectorImpl<Value *> &BuildVectorOpds) {
4617 Value *V = nullptr;
4618 do {
4619 BuildVector.push_back(IV);
4620 BuildVectorOpds.push_back(IV->getInsertedValueOperand());
4621 V = IV->getAggregateOperand();
4622 if (isa<UndefValue>(V))
4623 break;
4624 IV = dyn_cast<InsertValueInst>(V);
4625 if (!IV || !IV->hasOneUse())
4626 return false;
4627 } while (true);
4628 std::reverse(BuildVector.begin(), BuildVector.end());
4629 std::reverse(BuildVectorOpds.begin(), BuildVectorOpds.end());
4630 return true;
4631}
4632
4633static bool PhiTypeSorterFunc(Value *V, Value *V2) {
4634 return V->getType() < V2->getType();
4635}
4636
4637/// \brief Try and get a reduction value from a phi node.
4638///
4639/// Given a phi node \p P in a block \p ParentBB, consider possible reductions
4640/// if they come from either \p ParentBB or a containing loop latch.
4641///
4642/// \returns A candidate reduction value if possible, or \code nullptr \endcode
4643/// if not possible.
4644static Value *getReductionValue(const DominatorTree *DT, PHINode *P,
4645 BasicBlock *ParentBB, LoopInfo *LI) {
4646 // There are situations where the reduction value is not dominated by the
4647 // reduction phi. Vectorizing such cases has been reported to cause
4648 // miscompiles. See PR25787.
4649 auto DominatedReduxValue = [&](Value *R) {
4650 return (
4651 dyn_cast<Instruction>(R) &&
4652 DT->dominates(P->getParent(), dyn_cast<Instruction>(R)->getParent()));
4653 };
4654
4655 Value *Rdx = nullptr;
4656
4657 // Return the incoming value if it comes from the same BB as the phi node.
4658 if (P->getIncomingBlock(0) == ParentBB) {
4659 Rdx = P->getIncomingValue(0);
4660 } else if (P->getIncomingBlock(1) == ParentBB) {
4661 Rdx = P->getIncomingValue(1);
4662 }
4663
4664 if (Rdx && DominatedReduxValue(Rdx))
4665 return Rdx;
4666
4667 // Otherwise, check whether we have a loop latch to look at.
4668 Loop *BBL = LI->getLoopFor(ParentBB);
4669 if (!BBL)
4670 return nullptr;
4671 BasicBlock *BBLatch = BBL->getLoopLatch();
4672 if (!BBLatch)
4673 return nullptr;
4674
4675 // There is a loop latch, return the incoming value if it comes from
4676 // that. This reduction pattern occasionally turns up.
4677 if (P->getIncomingBlock(0) == BBLatch) {
4678 Rdx = P->getIncomingValue(0);
4679 } else if (P->getIncomingBlock(1) == BBLatch) {
4680 Rdx = P->getIncomingValue(1);
4681 }
4682
4683 if (Rdx && DominatedReduxValue(Rdx))
4684 return Rdx;
4685
4686 return nullptr;
4687}
4688
4689namespace {
4690/// Tracks instructons and its children.
4691class WeakVHWithLevel final : public CallbackVH {
4692 /// Operand index of the instruction currently beeing analized.
4693 unsigned Level = 0;
4694 /// Is this the instruction that should be vectorized, or are we now
4695 /// processing children (i.e. operands of this instruction) for potential
4696 /// vectorization?
4697 bool IsInitial = true;
4698
4699public:
4700 explicit WeakVHWithLevel() = default;
4701 WeakVHWithLevel(Value *V) : CallbackVH(V){};
4702 /// Restart children analysis each time it is repaced by the new instruction.
4703 void allUsesReplacedWith(Value *New) override {
4704 setValPtr(New);
4705 Level = 0;
4706 IsInitial = true;
4707 }
4708 /// Check if the instruction was not deleted during vectorization.
4709 bool isValid() const { return !getValPtr(); }
4710 /// Is the istruction itself must be vectorized?
4711 bool isInitial() const { return IsInitial; }
4712 /// Try to vectorize children.
4713 void clearInitial() { IsInitial = false; }
4714 /// Are all children processed already?
4715 bool isFinal() const {
4716 assert(getValPtr() &&((getValPtr() && (isa<Instruction>(getValPtr())
&& cast<Instruction>(getValPtr())->getNumOperands
() >= Level)) ? static_cast<void> (0) : __assert_fail
("getValPtr() && (isa<Instruction>(getValPtr()) && cast<Instruction>(getValPtr())->getNumOperands() >= Level)"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4718, __PRETTY_FUNCTION__))
4717 (isa<Instruction>(getValPtr()) &&((getValPtr() && (isa<Instruction>(getValPtr())
&& cast<Instruction>(getValPtr())->getNumOperands
() >= Level)) ? static_cast<void> (0) : __assert_fail
("getValPtr() && (isa<Instruction>(getValPtr()) && cast<Instruction>(getValPtr())->getNumOperands() >= Level)"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4718, __PRETTY_FUNCTION__))
4718 cast<Instruction>(getValPtr())->getNumOperands() >= Level))((getValPtr() && (isa<Instruction>(getValPtr())
&& cast<Instruction>(getValPtr())->getNumOperands
() >= Level)) ? static_cast<void> (0) : __assert_fail
("getValPtr() && (isa<Instruction>(getValPtr()) && cast<Instruction>(getValPtr())->getNumOperands() >= Level)"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4718, __PRETTY_FUNCTION__))
;
4719 return getValPtr() &&
4720 cast<Instruction>(getValPtr())->getNumOperands() == Level;
4721 }
4722 /// Get next child operation.
4723 Value *nextOperand() {
4724 assert(getValPtr() && isa<Instruction>(getValPtr()) &&((getValPtr() && isa<Instruction>(getValPtr()) &&
cast<Instruction>(getValPtr())->getNumOperands() >
Level) ? static_cast<void> (0) : __assert_fail ("getValPtr() && isa<Instruction>(getValPtr()) && cast<Instruction>(getValPtr())->getNumOperands() > Level"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4725, __PRETTY_FUNCTION__))
4725 cast<Instruction>(getValPtr())->getNumOperands() > Level)((getValPtr() && isa<Instruction>(getValPtr()) &&
cast<Instruction>(getValPtr())->getNumOperands() >
Level) ? static_cast<void> (0) : __assert_fail ("getValPtr() && isa<Instruction>(getValPtr()) && cast<Instruction>(getValPtr())->getNumOperands() > Level"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4725, __PRETTY_FUNCTION__))
;
4726 return cast<Instruction>(getValPtr())->getOperand(Level++);
4727 }
4728 virtual ~WeakVHWithLevel() = default;
4729};
4730} // namespace
4731
4732/// \brief Attempt to reduce a horizontal reduction.
4733/// If it is legal to match a horizontal reduction feeding
4734/// the phi node P with reduction operators Root in a basic block BB, then check
4735/// if it can be done.
4736/// \returns true if a horizontal reduction was matched and reduced.
4737/// \returns false if a horizontal reduction was not matched.
4738static bool canBeVectorized(
4739 PHINode *P, Instruction *Root, BasicBlock *BB, BoUpSLP &R,
4740 TargetTransformInfo *TTI,
4741 const function_ref<bool(BinaryOperator *, BoUpSLP &)> Vectorize) {
4742 if (!ShouldVectorizeHor)
4743 return false;
4744
4745 if (!Root)
4746 return false;
4747
4748 if (Root->getParent() != BB)
4749 return false;
4750 SmallVector<WeakVHWithLevel, 8> Stack(1, Root);
4751 SmallSet<Value *, 8> VisitedInstrs;
4752 bool Res = false;
4753 while (!Stack.empty()) {
4754 Value *V = Stack.back();
4755 if (!V) {
4756 Stack.pop_back();
4757 continue;
4758 }
4759 auto *Inst = dyn_cast<Instruction>(V);
4760 if (!Inst || isa<PHINode>(Inst)) {
4761 Stack.pop_back();
4762 continue;
4763 }
4764 if (Stack.back().isInitial()) {
4765 Stack.back().clearInitial();
4766 if (auto *BI = dyn_cast<BinaryOperator>(Inst)) {
4767 HorizontalReduction HorRdx;
4768 if (HorRdx.matchAssociativeReduction(P, BI)) {
4769 if (HorRdx.tryToReduce(R, TTI)) {
4770 Res = true;
4771 P = nullptr;
4772 continue;
4773 }
4774 }
4775 if (P) {
4776 Inst = dyn_cast<Instruction>(BI->getOperand(0));
4777 if (Inst == P)
4778 Inst = dyn_cast<Instruction>(BI->getOperand(1));
4779 if (!Inst) {
4780 P = nullptr;
4781 continue;
4782 }
4783 }
4784 }
4785 P = nullptr;
4786 if (Vectorize(dyn_cast<BinaryOperator>(Inst), R)) {
4787 Res = true;
4788 continue;
4789 }
4790 }
4791 if (Stack.back().isFinal()) {
4792 Stack.pop_back();
4793 continue;
4794 }
4795
4796 if (auto *NextV = dyn_cast<Instruction>(Stack.back().nextOperand()))
4797 if (NextV->getParent() == BB && VisitedInstrs.insert(NextV).second &&
4798 Stack.size() < RecursionMaxDepth)
4799 Stack.push_back(NextV);
4800 }
4801 return Res;
4802}
4803
4804bool SLPVectorizerPass::vectorizeRootInstruction(PHINode *P, Value *V,
4805 BasicBlock *BB, BoUpSLP &R,
4806 TargetTransformInfo *TTI) {
4807 if (!V)
4808 return false;
4809 auto *I = dyn_cast<Instruction>(V);
4810 if (!I)
4811 return false;
4812
4813 if (!isa<BinaryOperator>(I))
4814 P = nullptr;
4815 // Try to match and vectorize a horizontal reduction.
4816 return canBeVectorized(P, I, BB, R, TTI,
4817 [this](BinaryOperator *BI, BoUpSLP &R) -> bool {
4818 return tryToVectorize(BI, R);
4819 });
4820}
4821
4822bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
4823 bool Changed = false;
4824 SmallVector<Value *, 4> Incoming;
4825 SmallSet<Value *, 16> VisitedInstrs;
4826
4827 bool HaveVectorizedPhiNodes = true;
4828 while (HaveVectorizedPhiNodes) {
4829 HaveVectorizedPhiNodes = false;
4830
4831 // Collect the incoming values from the PHIs.
4832 Incoming.clear();
4833 for (Instruction &I : *BB) {
4834 PHINode *P = dyn_cast<PHINode>(&I);
4835 if (!P)
4836 break;
4837
4838 if (!VisitedInstrs.count(P))
4839 Incoming.push_back(P);
4840 }
4841
4842 // Sort by type.
4843 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc);
4844
4845 // Try to vectorize elements base on their type.
4846 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(),
4847 E = Incoming.end();
4848 IncIt != E;) {
4849
4850 // Look for the next elements with the same type.
4851 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt;
4852 while (SameTypeIt != E &&
4853 (*SameTypeIt)->getType() == (*IncIt)->getType()) {
4854 VisitedInstrs.insert(*SameTypeIt);
4855 ++SameTypeIt;
4856 }
4857
4858 // Try to vectorize them.
4859 unsigned NumElts = (SameTypeIt - IncIt);
4860 DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { errs() << "SLP: Trying to vectorize starting at PHIs ("
<< NumElts << ")\n"; } } while (false)
;
4861 if (NumElts > 1 && tryToVectorizeList(makeArrayRef(IncIt, NumElts), R)) {
4862 // Success start over because instructions might have been changed.
4863 HaveVectorizedPhiNodes = true;
4864 Changed = true;
4865 break;
4866 }
4867
4868 // Start over at the next instruction of a different type (or the end).
4869 IncIt = SameTypeIt;
4870 }
4871 }
4872
4873 VisitedInstrs.clear();
4874
4875 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
4876 // We may go through BB multiple times so skip the one we have checked.
4877 if (!VisitedInstrs.insert(&*it).second)
4878 continue;
4879
4880 if (isa<DbgInfoIntrinsic>(it))
4881 continue;
4882
4883 // Try to vectorize reductions that use PHINodes.
4884 if (PHINode *P = dyn_cast<PHINode>(it)) {
4885 // Check that the PHI is a reduction PHI.
4886 if (P->getNumIncomingValues() != 2)
4887 return Changed;
4888
4889 // Try to match and vectorize a horizontal reduction.
4890 if (vectorizeRootInstruction(P, getReductionValue(DT, P, BB, LI), BB, R,
4891 TTI)) {
4892 Changed = true;
4893 it = BB->begin();
4894 e = BB->end();
4895 continue;
4896 }
4897 continue;
4898 }
4899
4900 if (ShouldStartVectorizeHorAtStore) {
4901 if (StoreInst *SI = dyn_cast<StoreInst>(it)) {
4902 // Try to match and vectorize a horizontal reduction.
4903 if (vectorizeRootInstruction(nullptr, SI->getValueOperand(), BB, R,
4904 TTI)) {
4905 Changed = true;
4906 it = BB->begin();
4907 e = BB->end();
4908 continue;
4909 }
4910 }
4911 }
4912
4913 // Try to vectorize horizontal reductions feeding into a return.
4914 if (ReturnInst *RI = dyn_cast<ReturnInst>(it)) {
4915 if (RI->getNumOperands() != 0) {
4916 // Try to match and vectorize a horizontal reduction.
4917 if (vectorizeRootInstruction(nullptr, RI->getOperand(0), BB, R, TTI)) {
4918 Changed = true;
4919 it = BB->begin();
4920 e = BB->end();
4921 continue;
4922 }
4923 }
4924 }
4925
4926 // Try to vectorize trees that start at compare instructions.
4927 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
4928 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
4929 Changed = true;
4930 // We would like to start over since some instructions are deleted
4931 // and the iterator may become invalid value.
4932 it = BB->begin();
4933 e = BB->end();
4934 continue;
4935 }
4936
4937 for (int I = 0; I < 2; ++I) {
4938 if (vectorizeRootInstruction(nullptr, CI->getOperand(I), BB, R, TTI)) {
4939 Changed = true;
4940 // We would like to start over since some instructions are deleted
4941 // and the iterator may become invalid value.
4942 it = BB->begin();
4943 e = BB->end();
4944 break;
4945 }
4946 }
4947 continue;
4948 }
4949
4950 // Try to vectorize trees that start at insertelement instructions.
4951 if (InsertElementInst *FirstInsertElem = dyn_cast<InsertElementInst>(it)) {
4952 SmallVector<Value *, 16> BuildVector;
4953 SmallVector<Value *, 16> BuildVectorOpds;
4954 if (!findBuildVector(FirstInsertElem, BuildVector, BuildVectorOpds))
4955 continue;
4956
4957 // Vectorize starting with the build vector operands ignoring the
4958 // BuildVector instructions for the purpose of scheduling and user
4959 // extraction.
4960 if (tryToVectorizeList(BuildVectorOpds, R, BuildVector)) {
4961 Changed = true;
4962 it = BB->begin();
4963 e = BB->end();
4964 }
4965
4966 continue;
4967 }
4968
4969 // Try to vectorize trees that start at insertvalue instructions feeding into
4970 // a store.
4971 if (StoreInst *SI = dyn_cast<StoreInst>(it)) {
4972 if (InsertValueInst *LastInsertValue = dyn_cast<InsertValueInst>(SI->getValueOperand())) {
4973 const DataLayout &DL = BB->getModule()->getDataLayout();
4974 if (R.canMapToVector(SI->getValueOperand()->getType(), DL)) {
4975 SmallVector<Value *, 16> BuildVector;
4976 SmallVector<Value *, 16> BuildVectorOpds;
4977 if (!findBuildAggregate(LastInsertValue, BuildVector, BuildVectorOpds))
4978 continue;
4979
4980 DEBUG(dbgs() << "SLP: store of array mappable to vector: " << *SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: store of array mappable to vector: "
<< *SI << "\n"; } } while (false)
;
4981 if (tryToVectorizeList(BuildVectorOpds, R, BuildVector, false)) {
4982 Changed = true;
4983 it = BB->begin();
4984 e = BB->end();
4985 }
4986 continue;
4987 }
4988 }
4989 }
4990 }
4991
4992 return Changed;
4993}
4994
4995bool SLPVectorizerPass::vectorizeGEPIndices(BasicBlock *BB, BoUpSLP &R) {
4996 auto Changed = false;
4997 for (auto &Entry : GEPs) {
4998
4999 // If the getelementptr list has fewer than two elements, there's nothing
5000 // to do.
5001 if (Entry.second.size() < 2)
5002 continue;
5003
5004 DEBUG(dbgs() << "SLP: Analyzing a getelementptr list of length "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing a getelementptr list of length "
<< Entry.second.size() << ".\n"; } } while (false
)
5005 << Entry.second.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing a getelementptr list of length "
<< Entry.second.size() << ".\n"; } } while (false
)
;
5006
5007 // We process the getelementptr list in chunks of 16 (like we do for
5008 // stores) to minimize compile-time.
5009 for (unsigned BI = 0, BE = Entry.second.size(); BI < BE; BI += 16) {
5010 auto Len = std::min<unsigned>(BE - BI, 16);
5011 auto GEPList = makeArrayRef(&Entry.second[BI], Len);
5012
5013 // Initialize a set a candidate getelementptrs. Note that we use a
5014 // SetVector here to preserve program order. If the index computations
5015 // are vectorizable and begin with loads, we want to minimize the chance
5016 // of having to reorder them later.
5017 SetVector<Value *> Candidates(GEPList.begin(), GEPList.end());
5018
5019 // Some of the candidates may have already been vectorized after we
5020 // initially collected them. If so, the WeakVHs will have nullified the
5021 // values, so remove them from the set of candidates.
5022 Candidates.remove(nullptr);
5023
5024 // Remove from the set of candidates all pairs of getelementptrs with
5025 // constant differences. Such getelementptrs are likely not good
5026 // candidates for vectorization in a bottom-up phase since one can be
5027 // computed from the other. We also ensure all candidate getelementptr
5028 // indices are unique.
5029 for (int I = 0, E = GEPList.size(); I < E && Candidates.size() > 1; ++I) {
5030 auto *GEPI = cast<GetElementPtrInst>(GEPList[I]);
5031 if (!Candidates.count(GEPI))
5032 continue;
5033 auto *SCEVI = SE->getSCEV(GEPList[I]);
5034 for (int J = I + 1; J < E && Candidates.size() > 1; ++J) {
5035 auto *GEPJ = cast<GetElementPtrInst>(GEPList[J]);
5036 auto *SCEVJ = SE->getSCEV(GEPList[J]);
5037 if (isa<SCEVConstant>(SE->getMinusSCEV(SCEVI, SCEVJ))) {
5038 Candidates.remove(GEPList[I]);
5039 Candidates.remove(GEPList[J]);
5040 } else if (GEPI->idx_begin()->get() == GEPJ->idx_begin()->get()) {
5041 Candidates.remove(GEPList[J]);
5042 }
5043 }
5044 }
5045
5046 // We break out of the above computation as soon as we know there are
5047 // fewer than two candidates remaining.
5048 if (Candidates.size() < 2)
5049 continue;
5050
5051 // Add the single, non-constant index of each candidate to the bundle. We
5052 // ensured the indices met these constraints when we originally collected
5053 // the getelementptrs.
5054 SmallVector<Value *, 16> Bundle(Candidates.size());
5055 auto BundleIndex = 0u;
5056 for (auto *V : Candidates) {
5057 auto *GEP = cast<GetElementPtrInst>(V);
5058 auto *GEPIdx = GEP->idx_begin()->get();
5059 assert(GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx))((GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx
)) ? static_cast<void> (0) : __assert_fail ("GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx)"
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn295818/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 5059, __PRETTY_FUNCTION__))
;
5060 Bundle[BundleIndex++] = GEPIdx;
5061 }
5062
5063 // Try and vectorize the indices. We are currently only interested in
5064 // gather-like cases of the form:
5065 //
5066 // ... = g[a[0] - b[0]] + g[a[1] - b[1]] + ...
5067 //
5068 // where the loads of "a", the loads of "b", and the subtractions can be
5069 // performed in parallel. It's likely that detecting this pattern in a
5070 // bottom-up phase will be simpler and less costly than building a
5071 // full-blown top-down phase beginning at the consecutive loads.
5072 Changed |= tryToVectorizeList(Bundle, R);
5073 }
5074 }
5075 return Changed;
5076}
5077
5078bool SLPVectorizerPass::vectorizeStoreChains(BoUpSLP &R) {
5079 bool Changed = false;
5080 // Attempt to sort and vectorize each of the store-groups.
5081 for (StoreListMap::iterator it = Stores.begin(), e = Stores.end(); it != e;
5082 ++it) {
5083 if (it->second.size() < 2)
5084 continue;
5085
5086 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing a store chain of length "
<< it->second.size() << ".\n"; } } while (false
)
5087 << it->second.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing a store chain of length "
<< it->second.size() << ".\n"; } } while (false
)
;
5088
5089 // Process the stores in chunks of 16.
5090 // TODO: The limit of 16 inhibits greater vectorization factors.
5091 // For example, AVX2 supports v32i8. Increasing this limit, however,
5092 // may cause a significant compile-time increase.
5093 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
5094 unsigned Len = std::min<unsigned>(CE - CI, 16);
5095 Changed |= vectorizeStores(makeArrayRef(&it->second[CI], Len), R);
5096 }
5097 }
5098 return Changed;
5099}
5100
5101char SLPVectorizer::ID = 0;
5102static const char lv_name[] = "SLP Vectorizer";
5103INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)static void *initializeSLPVectorizerPassOnce(PassRegistry &
Registry) {
5104INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
5105INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry);
5106INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry);
5107INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry);
5108INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry);
5109INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry);
5110INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)PassInfo *PI = new PassInfo( lv_name, "slp-vectorizer", &
SLPVectorizer::ID, PassInfo::NormalCtor_t(callDefaultCtor<
SLPVectorizer>), false, false); Registry.registerPass(*PI,
true); return PI; } static llvm::once_flag InitializeSLPVectorizerPassFlag
; void llvm::initializeSLPVectorizerPass(PassRegistry &Registry
) { llvm::call_once(InitializeSLPVectorizerPassFlag, initializeSLPVectorizerPassOnce
, std::ref(Registry)); }
5111
5112namespace llvm {
5113Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }
5114}