Bug Summary

File:lib/Transforms/Vectorize/SLPVectorizer.cpp
Warning:line 2383, 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/GraphWriter.h"
43#include "llvm/Support/raw_ostream.h"
44#include "llvm/Transforms/Vectorize.h"
45#include <algorithm>
46#include <memory>
47
48using namespace llvm;
49using namespace slpvectorizer;
50
51#define SV_NAME"slp-vectorizer" "slp-vectorizer"
52#define DEBUG_TYPE"SLP" "SLP"
53
54STATISTIC(NumVectorInstructions, "Number of vector instructions generated")static llvm::Statistic NumVectorInstructions = {"SLP", "NumVectorInstructions"
, "Number of vector instructions generated", {0}, false}
;
55
56static cl::opt<int>
57 SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden,
58 cl::desc("Only vectorize if you gain more than this "
59 "number "));
60
61static cl::opt<bool>
62ShouldVectorizeHor("slp-vectorize-hor", cl::init(true), cl::Hidden,
63 cl::desc("Attempt to vectorize horizontal reductions"));
64
65static cl::opt<bool> ShouldStartVectorizeHorAtStore(
66 "slp-vectorize-hor-store", cl::init(false), cl::Hidden,
67 cl::desc(
68 "Attempt to vectorize horizontal reductions feeding into a store"));
69
70static cl::opt<int>
71MaxVectorRegSizeOption("slp-max-reg-size", cl::init(128), cl::Hidden,
72 cl::desc("Attempt to vectorize for this register size in bits"));
73
74/// Limits the size of scheduling regions in a block.
75/// It avoid long compile times for _very_ large blocks where vector
76/// instructions are spread over a wide range.
77/// This limit is way higher than needed by real-world functions.
78static cl::opt<int>
79ScheduleRegionSizeBudget("slp-schedule-budget", cl::init(100000), cl::Hidden,
80 cl::desc("Limit the size of the SLP scheduling region per block"));
81
82static cl::opt<int> MinVectorRegSizeOption(
83 "slp-min-reg-size", cl::init(128), cl::Hidden,
84 cl::desc("Attempt to vectorize for this register size in bits"));
85
86static cl::opt<unsigned> RecursionMaxDepth(
87 "slp-recursion-max-depth", cl::init(12), cl::Hidden,
88 cl::desc("Limit the recursion depth when building a vectorizable tree"));
89
90static cl::opt<unsigned> MinTreeSize(
91 "slp-min-tree-size", cl::init(3), cl::Hidden,
92 cl::desc("Only vectorize small trees if they are fully vectorizable"));
93
94static cl::opt<bool>
95 ViewSLPTree("view-slp-tree", cl::Hidden,
96 cl::desc("Display the SLP trees with Graphviz"));
97
98// Limit the number of alias checks. The limit is chosen so that
99// it has no negative effect on the llvm benchmarks.
100static const unsigned AliasedCheckLimit = 10;
101
102// Another limit for the alias checks: The maximum distance between load/store
103// instructions where alias checks are done.
104// This limit is useful for very large basic blocks.
105static const unsigned MaxMemDepDistance = 160;
106
107/// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling
108/// regions to be handled.
109static const int MinScheduleRegionSize = 16;
110
111/// \brief Predicate for the element types that the SLP vectorizer supports.
112///
113/// The most important thing to filter here are types which are invalid in LLVM
114/// vectors. We also filter target specific types which have absolutely no
115/// meaningful vectorization path such as x86_fp80 and ppc_f128. This just
116/// avoids spending time checking the cost model and realizing that they will
117/// be inevitably scalarized.
118static bool isValidElementType(Type *Ty) {
119 return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() &&
120 !Ty->isPPC_FP128Ty();
121}
122
123/// \returns true if all of the instructions in \p VL are in the same block or
124/// false otherwise.
125static bool allSameBlock(ArrayRef<Value *> VL) {
126 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
127 if (!I0)
128 return false;
129 BasicBlock *BB = I0->getParent();
130 for (int i = 1, e = VL.size(); i < e; i++) {
131 Instruction *I = dyn_cast<Instruction>(VL[i]);
132 if (!I)
133 return false;
134
135 if (BB != I->getParent())
136 return false;
137 }
138 return true;
139}
140
141/// \returns True if all of the values in \p VL are constants.
142static bool allConstant(ArrayRef<Value *> VL) {
143 for (Value *i : VL)
144 if (!isa<Constant>(i))
145 return false;
146 return true;
147}
148
149/// \returns True if all of the values in \p VL are identical.
150static bool isSplat(ArrayRef<Value *> VL) {
151 for (unsigned i = 1, e = VL.size(); i < e; ++i)
152 if (VL[i] != VL[0])
153 return false;
154 return true;
155}
156
157///\returns Opcode that can be clubbed with \p Op to create an alternate
158/// sequence which can later be merged as a ShuffleVector instruction.
159static unsigned getAltOpcode(unsigned Op) {
160 switch (Op) {
161 case Instruction::FAdd:
162 return Instruction::FSub;
163 case Instruction::FSub:
164 return Instruction::FAdd;
165 case Instruction::Add:
166 return Instruction::Sub;
167 case Instruction::Sub:
168 return Instruction::Add;
169 default:
170 return 0;
171 }
172}
173
174///\returns bool representing if Opcode \p Op can be part
175/// of an alternate sequence which can later be merged as
176/// a ShuffleVector instruction.
177static bool canCombineAsAltInst(unsigned Op) {
178 return Op == Instruction::FAdd || Op == Instruction::FSub ||
179 Op == Instruction::Sub || Op == Instruction::Add;
180}
181
182/// \returns ShuffleVector instruction if instructions in \p VL have
183/// alternate fadd,fsub / fsub,fadd/add,sub/sub,add sequence.
184/// (i.e. e.g. opcodes of fadd,fsub,fadd,fsub...)
185static unsigned isAltInst(ArrayRef<Value *> VL) {
186 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
187 unsigned Opcode = I0->getOpcode();
188 unsigned AltOpcode = getAltOpcode(Opcode);
189 for (int i = 1, e = VL.size(); i < e; i++) {
190 Instruction *I = dyn_cast<Instruction>(VL[i]);
191 if (!I || I->getOpcode() != ((i & 1) ? AltOpcode : Opcode))
192 return 0;
193 }
194 return Instruction::ShuffleVector;
195}
196
197/// \returns The opcode if all of the Instructions in \p VL have the same
198/// opcode, or zero.
199static unsigned getSameOpcode(ArrayRef<Value *> VL) {
200 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
201 if (!I0)
202 return 0;
203 unsigned Opcode = I0->getOpcode();
204 for (int i = 1, e = VL.size(); i < e; i++) {
205 Instruction *I = dyn_cast<Instruction>(VL[i]);
206 if (!I || Opcode != I->getOpcode()) {
207 if (canCombineAsAltInst(Opcode) && i == 1)
208 return isAltInst(VL);
209 return 0;
210 }
211 }
212 return Opcode;
213}
214
215/// Get the intersection (logical and) of all of the potential IR flags
216/// of each scalar operation (VL) that will be converted into a vector (I).
217/// Flag set: NSW, NUW, exact, and all of fast-math.
218static void propagateIRFlags(Value *I, ArrayRef<Value *> VL) {
219 if (auto *VecOp = dyn_cast<Instruction>(I)) {
220 if (auto *I0 = dyn_cast<Instruction>(VL[0])) {
221 // VecOVp is initialized to the 0th scalar, so start counting from index
222 // '1'.
223 VecOp->copyIRFlags(I0);
224 for (int i = 1, e = VL.size(); i < e; ++i) {
225 if (auto *Scalar = dyn_cast<Instruction>(VL[i]))
226 VecOp->andIRFlags(Scalar);
227 }
228 }
229 }
230}
231
232/// \returns true if all of the values in \p VL have the same type or false
233/// otherwise.
234static bool allSameType(ArrayRef<Value *> VL) {
235 Type *Ty = VL[0]->getType();
236 for (int i = 1, e = VL.size(); i < e; i++)
237 if (VL[i]->getType() != Ty)
238 return false;
239
240 return true;
241}
242
243/// \returns True if Extract{Value,Element} instruction extracts element Idx.
244static bool matchExtractIndex(Instruction *E, unsigned Idx, unsigned Opcode) {
245 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 246, __PRETTY_FUNCTION__))
246 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 246, __PRETTY_FUNCTION__))
;
247 if (Opcode == Instruction::ExtractElement) {
248 ConstantInt *CI = dyn_cast<ConstantInt>(E->getOperand(1));
249 return CI && CI->getZExtValue() == Idx;
250 } else {
251 ExtractValueInst *EI = cast<ExtractValueInst>(E);
252 return EI->getNumIndices() == 1 && *EI->idx_begin() == Idx;
253 }
254}
255
256/// \returns True if in-tree use also needs extract. This refers to
257/// possible scalar operand in vectorized instruction.
258static bool InTreeUserNeedToExtract(Value *Scalar, Instruction *UserInst,
259 TargetLibraryInfo *TLI) {
260
261 unsigned Opcode = UserInst->getOpcode();
262 switch (Opcode) {
263 case Instruction::Load: {
264 LoadInst *LI = cast<LoadInst>(UserInst);
265 return (LI->getPointerOperand() == Scalar);
266 }
267 case Instruction::Store: {
268 StoreInst *SI = cast<StoreInst>(UserInst);
269 return (SI->getPointerOperand() == Scalar);
270 }
271 case Instruction::Call: {
272 CallInst *CI = cast<CallInst>(UserInst);
273 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
274 if (hasVectorInstrinsicScalarOpd(ID, 1)) {
275 return (CI->getArgOperand(1) == Scalar);
276 }
277 }
278 default:
279 return false;
280 }
281}
282
283/// \returns the AA location that is being access by the instruction.
284static MemoryLocation getLocation(Instruction *I, AliasAnalysis *AA) {
285 if (StoreInst *SI = dyn_cast<StoreInst>(I))
286 return MemoryLocation::get(SI);
287 if (LoadInst *LI = dyn_cast<LoadInst>(I))
288 return MemoryLocation::get(LI);
289 return MemoryLocation();
290}
291
292/// \returns True if the instruction is not a volatile or atomic load/store.
293static bool isSimple(Instruction *I) {
294 if (LoadInst *LI = dyn_cast<LoadInst>(I))
295 return LI->isSimple();
296 if (StoreInst *SI = dyn_cast<StoreInst>(I))
297 return SI->isSimple();
298 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
299 return !MI->isVolatile();
300 return true;
301}
302
303namespace llvm {
304namespace slpvectorizer {
305/// Bottom Up SLP Vectorizer.
306class BoUpSLP {
307public:
308 typedef SmallVector<Value *, 8> ValueList;
309 typedef SmallVector<Instruction *, 16> InstrList;
310 typedef SmallPtrSet<Value *, 16> ValueSet;
311 typedef SmallVector<StoreInst *, 8> StoreList;
312 typedef MapVector<Value *, SmallVector<Instruction *, 2>>
313 ExtraValueToDebugLocsMap;
314
315 BoUpSLP(Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti,
316 TargetLibraryInfo *TLi, AliasAnalysis *Aa, LoopInfo *Li,
317 DominatorTree *Dt, AssumptionCache *AC, DemandedBits *DB,
318 const DataLayout *DL)
319 : NumLoadsWantToKeepOrder(0), NumLoadsWantToChangeOrder(0), F(Func),
320 SE(Se), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt), AC(AC), DB(DB),
321 DL(DL), Builder(Se->getContext()) {
322 CodeMetrics::collectEphemeralValues(F, AC, EphValues);
323 // Use the vector register size specified by the target unless overridden
324 // by a command-line option.
325 // TODO: It would be better to limit the vectorization factor based on
326 // data type rather than just register size. For example, x86 AVX has
327 // 256-bit registers, but it does not support integer operations
328 // at that width (that requires AVX2).
329 if (MaxVectorRegSizeOption.getNumOccurrences())
330 MaxVecRegSize = MaxVectorRegSizeOption;
331 else
332 MaxVecRegSize = TTI->getRegisterBitWidth(true);
333
334 MinVecRegSize = MinVectorRegSizeOption;
335 }
336
337 /// \brief Vectorize the tree that starts with the elements in \p VL.
338 /// Returns the vectorized root.
339 Value *vectorizeTree();
340 /// Vectorize the tree but with the list of externally used values \p
341 /// ExternallyUsedValues. Values in this MapVector can be replaced but the
342 /// generated extractvalue instructions.
343 Value *vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues);
344
345 /// \returns the cost incurred by unwanted spills and fills, caused by
346 /// holding live values over call sites.
347 int getSpillCost();
348
349 /// \returns the vectorization cost of the subtree that starts at \p VL.
350 /// A negative number means that this is profitable.
351 int getTreeCost();
352
353 /// Construct a vectorizable tree that starts at \p Roots, ignoring users for
354 /// the purpose of scheduling and extraction in the \p UserIgnoreLst.
355 void buildTree(ArrayRef<Value *> Roots,
356 ArrayRef<Value *> UserIgnoreLst = None);
357 /// Construct a vectorizable tree that starts at \p Roots, ignoring users for
358 /// the purpose of scheduling and extraction in the \p UserIgnoreLst taking
359 /// into account (anf updating it, if required) list of externally used
360 /// values stored in \p ExternallyUsedValues.
361 void buildTree(ArrayRef<Value *> Roots,
362 ExtraValueToDebugLocsMap &ExternallyUsedValues,
363 ArrayRef<Value *> UserIgnoreLst = None);
364
365 /// Clear the internal data structures that are created by 'buildTree'.
366 void deleteTree() {
367 VectorizableTree.clear();
368 ScalarToTreeEntry.clear();
369 MustGather.clear();
370 ExternalUses.clear();
371 NumLoadsWantToKeepOrder = 0;
372 NumLoadsWantToChangeOrder = 0;
373 for (auto &Iter : BlocksSchedules) {
374 BlockScheduling *BS = Iter.second.get();
375 BS->clear();
376 }
377 MinBWs.clear();
378 }
379
380 /// \brief Perform LICM and CSE on the newly generated gather sequences.
381 void optimizeGatherSequence();
382
383 /// \returns true if it is beneficial to reverse the vector order.
384 bool shouldReorder() const {
385 return NumLoadsWantToChangeOrder > NumLoadsWantToKeepOrder;
386 }
387
388 /// \return The vector element size in bits to use when vectorizing the
389 /// expression tree ending at \p V. If V is a store, the size is the width of
390 /// the stored value. Otherwise, the size is the width of the largest loaded
391 /// value reaching V. This method is used by the vectorizer to calculate
392 /// vectorization factors.
393 unsigned getVectorElementSize(Value *V);
394
395 /// Compute the minimum type sizes required to represent the entries in a
396 /// vectorizable tree.
397 void computeMinimumValueSizes();
398
399 // \returns maximum vector register size as set by TTI or overridden by cl::opt.
400 unsigned getMaxVecRegSize() const {
401 return MaxVecRegSize;
402 }
403
404 // \returns minimum vector register size as set by cl::opt.
405 unsigned getMinVecRegSize() const {
406 return MinVecRegSize;
407 }
408
409 /// \brief Check if ArrayType or StructType is isomorphic to some VectorType.
410 ///
411 /// \returns number of elements in vector if isomorphism exists, 0 otherwise.
412 unsigned canMapToVector(Type *T, const DataLayout &DL) const;
413
414 /// \returns True if the VectorizableTree is both tiny and not fully
415 /// vectorizable. We do not vectorize such trees.
416 bool isTreeTinyAndNotFullyVectorizable();
417
418private:
419 struct TreeEntry;
420
421 /// \returns the cost of the vectorizable entry.
422 int getEntryCost(TreeEntry *E);
423
424 /// This is the recursive part of buildTree.
425 void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth, int);
426
427 /// \returns True if the ExtractElement/ExtractValue instructions in VL can
428 /// be vectorized to use the original vector (or aggregate "bitcast" to a vector).
429 bool canReuseExtract(ArrayRef<Value *> VL, unsigned Opcode) const;
430
431 /// Vectorize a single entry in the tree.
432 Value *vectorizeTree(TreeEntry *E);
433
434 /// Vectorize a single entry in the tree, starting in \p VL.
435 Value *vectorizeTree(ArrayRef<Value *> VL);
436
437 /// \returns the pointer to the vectorized value if \p VL is already
438 /// vectorized, or NULL. They may happen in cycles.
439 Value *alreadyVectorized(ArrayRef<Value *> VL) const;
440
441 /// \returns the scalarization cost for this type. Scalarization in this
442 /// context means the creation of vectors from a group of scalars.
443 int getGatherCost(Type *Ty);
444
445 /// \returns the scalarization cost for this list of values. Assuming that
446 /// this subtree gets vectorized, we may need to extract the values from the
447 /// roots. This method calculates the cost of extracting the values.
448 int getGatherCost(ArrayRef<Value *> VL);
449
450 /// \brief Set the Builder insert point to one after the last instruction in
451 /// the bundle
452 void setInsertPointAfterBundle(ArrayRef<Value *> VL);
453
454 /// \returns a vector from a collection of scalars in \p VL.
455 Value *Gather(ArrayRef<Value *> VL, VectorType *Ty);
456
457 /// \returns whether the VectorizableTree is fully vectorizable and will
458 /// be beneficial even the tree height is tiny.
459 bool isFullyVectorizableTinyTree();
460
461 /// \reorder commutative operands in alt shuffle if they result in
462 /// vectorized code.
463 void reorderAltShuffleOperands(ArrayRef<Value *> VL,
464 SmallVectorImpl<Value *> &Left,
465 SmallVectorImpl<Value *> &Right);
466 /// \reorder commutative operands to get better probability of
467 /// generating vectorized code.
468 void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL,
469 SmallVectorImpl<Value *> &Left,
470 SmallVectorImpl<Value *> &Right);
471 struct TreeEntry {
472 TreeEntry(std::vector<TreeEntry> &Container)
473 : Scalars(), VectorizedValue(nullptr), NeedToGather(0),
474 Container(Container) {}
475
476 /// \returns true if the scalars in VL are equal to this entry.
477 bool isSame(ArrayRef<Value *> VL) const {
478 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 478, __PRETTY_FUNCTION__))
;
479 return std::equal(VL.begin(), VL.end(), Scalars.begin());
480 }
481
482 /// A vector of scalars.
483 ValueList Scalars;
484
485 /// The Scalars are vectorized into this value. It is initialized to Null.
486 Value *VectorizedValue;
487
488 /// Do we need to gather this sequence ?
489 bool NeedToGather;
490
491 /// Points back to the VectorizableTree.
492 ///
493 /// Only used for Graphviz right now. Unfortunately GraphTrait::NodeRef has
494 /// to be a pointer and needs to be able to initialize the child iterator.
495 /// Thus we need a reference back to the container to translate the indices
496 /// to entries.
497 std::vector<TreeEntry> &Container;
498
499 /// The TreeEntry index containing the user of this entry. We can actually
500 /// have multiple users so the data structure is not truly a tree.
501 SmallVector<int, 1> UserTreeIndices;
502 };
503
504 /// Create a new VectorizableTree entry.
505 TreeEntry *newTreeEntry(ArrayRef<Value *> VL, bool Vectorized,
506 int &UserTreeIdx) {
507 VectorizableTree.emplace_back(VectorizableTree);
508 int idx = VectorizableTree.size() - 1;
509 TreeEntry *Last = &VectorizableTree[idx];
510 Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end());
511 Last->NeedToGather = !Vectorized;
512 if (Vectorized) {
513 for (int i = 0, e = VL.size(); i != e; ++i) {
514 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 514, __PRETTY_FUNCTION__))
;
515 ScalarToTreeEntry[VL[i]] = idx;
516 }
517 } else {
518 MustGather.insert(VL.begin(), VL.end());
519 }
520
521 if (UserTreeIdx >= 0)
522 Last->UserTreeIndices.push_back(UserTreeIdx);
523 UserTreeIdx = idx;
524 return Last;
525 }
526
527 /// -- Vectorization State --
528 /// Holds all of the tree entries.
529 std::vector<TreeEntry> VectorizableTree;
530
531 /// Maps a specific scalar to its tree entry.
532 SmallDenseMap<Value*, int> ScalarToTreeEntry;
533
534 /// A list of scalars that we found that we need to keep as scalars.
535 ValueSet MustGather;
536
537 /// This POD struct describes one external user in the vectorized tree.
538 struct ExternalUser {
539 ExternalUser (Value *S, llvm::User *U, int L) :
540 Scalar(S), User(U), Lane(L){}
541 // Which scalar in our function.
542 Value *Scalar;
543 // Which user that uses the scalar.
544 llvm::User *User;
545 // Which lane does the scalar belong to.
546 int Lane;
547 };
548 typedef SmallVector<ExternalUser, 16> UserList;
549
550 /// Checks if two instructions may access the same memory.
551 ///
552 /// \p Loc1 is the location of \p Inst1. It is passed explicitly because it
553 /// is invariant in the calling loop.
554 bool isAliased(const MemoryLocation &Loc1, Instruction *Inst1,
555 Instruction *Inst2) {
556
557 // First check if the result is already in the cache.
558 AliasCacheKey key = std::make_pair(Inst1, Inst2);
559 Optional<bool> &result = AliasCache[key];
560 if (result.hasValue()) {
561 return result.getValue();
562 }
563 MemoryLocation Loc2 = getLocation(Inst2, AA);
564 bool aliased = true;
565 if (Loc1.Ptr && Loc2.Ptr && isSimple(Inst1) && isSimple(Inst2)) {
566 // Do the alias check.
567 aliased = AA->alias(Loc1, Loc2);
568 }
569 // Store the result in the cache.
570 result = aliased;
571 return aliased;
572 }
573
574 typedef std::pair<Instruction *, Instruction *> AliasCacheKey;
575
576 /// Cache for alias results.
577 /// TODO: consider moving this to the AliasAnalysis itself.
578 DenseMap<AliasCacheKey, Optional<bool>> AliasCache;
579
580 /// Removes an instruction from its block and eventually deletes it.
581 /// It's like Instruction::eraseFromParent() except that the actual deletion
582 /// is delayed until BoUpSLP is destructed.
583 /// This is required to ensure that there are no incorrect collisions in the
584 /// AliasCache, which can happen if a new instruction is allocated at the
585 /// same address as a previously deleted instruction.
586 void eraseInstruction(Instruction *I) {
587 I->removeFromParent();
588 I->dropAllReferences();
589 DeletedInstructions.push_back(std::unique_ptr<Instruction>(I));
590 }
591
592 /// Temporary store for deleted instructions. Instructions will be deleted
593 /// eventually when the BoUpSLP is destructed.
594 SmallVector<std::unique_ptr<Instruction>, 8> DeletedInstructions;
595
596 /// A list of values that need to extracted out of the tree.
597 /// This list holds pairs of (Internal Scalar : External User). External User
598 /// can be nullptr, it means that this Internal Scalar will be used later,
599 /// after vectorization.
600 UserList ExternalUses;
601
602 /// Values used only by @llvm.assume calls.
603 SmallPtrSet<const Value *, 32> EphValues;
604
605 /// Holds all of the instructions that we gathered.
606 SetVector<Instruction *> GatherSeq;
607 /// A list of blocks that we are going to CSE.
608 SetVector<BasicBlock *> CSEBlocks;
609
610 /// Contains all scheduling relevant data for an instruction.
611 /// A ScheduleData either represents a single instruction or a member of an
612 /// instruction bundle (= a group of instructions which is combined into a
613 /// vector instruction).
614 struct ScheduleData {
615
616 // The initial value for the dependency counters. It means that the
617 // dependencies are not calculated yet.
618 enum { InvalidDeps = -1 };
619
620 ScheduleData()
621 : Inst(nullptr), FirstInBundle(nullptr), NextInBundle(nullptr),
622 NextLoadStore(nullptr), SchedulingRegionID(0), SchedulingPriority(0),
623 Dependencies(InvalidDeps), UnscheduledDeps(InvalidDeps),
624 UnscheduledDepsInBundle(InvalidDeps), IsScheduled(false) {}
625
626 void init(int BlockSchedulingRegionID) {
627 FirstInBundle = this;
628 NextInBundle = nullptr;
629 NextLoadStore = nullptr;
630 IsScheduled = false;
631 SchedulingRegionID = BlockSchedulingRegionID;
632 UnscheduledDepsInBundle = UnscheduledDeps;
633 clearDependencies();
634 }
635
636 /// Returns true if the dependency information has been calculated.
637 bool hasValidDependencies() const { return Dependencies != InvalidDeps; }
638
639 /// Returns true for single instructions and for bundle representatives
640 /// (= the head of a bundle).
641 bool isSchedulingEntity() const { return FirstInBundle == this; }
642
643 /// Returns true if it represents an instruction bundle and not only a
644 /// single instruction.
645 bool isPartOfBundle() const {
646 return NextInBundle != nullptr || FirstInBundle != this;
647 }
648
649 /// Returns true if it is ready for scheduling, i.e. it has no more
650 /// unscheduled depending instructions/bundles.
651 bool isReady() const {
652 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 653, __PRETTY_FUNCTION__))
653 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 653, __PRETTY_FUNCTION__))
;
654 return UnscheduledDepsInBundle == 0 && !IsScheduled;
655 }
656
657 /// Modifies the number of unscheduled dependencies, also updating it for
658 /// the whole bundle.
659 int incrementUnscheduledDeps(int Incr) {
660 UnscheduledDeps += Incr;
661 return FirstInBundle->UnscheduledDepsInBundle += Incr;
662 }
663
664 /// Sets the number of unscheduled dependencies to the number of
665 /// dependencies.
666 void resetUnscheduledDeps() {
667 incrementUnscheduledDeps(Dependencies - UnscheduledDeps);
668 }
669
670 /// Clears all dependency information.
671 void clearDependencies() {
672 Dependencies = InvalidDeps;
673 resetUnscheduledDeps();
674 MemoryDependencies.clear();
675 }
676
677 void dump(raw_ostream &os) const {
678 if (!isSchedulingEntity()) {
679 os << "/ " << *Inst;
680 } else if (NextInBundle) {
681 os << '[' << *Inst;
682 ScheduleData *SD = NextInBundle;
683 while (SD) {
684 os << ';' << *SD->Inst;
685 SD = SD->NextInBundle;
686 }
687 os << ']';
688 } else {
689 os << *Inst;
690 }
691 }
692
693 Instruction *Inst;
694
695 /// Points to the head in an instruction bundle (and always to this for
696 /// single instructions).
697 ScheduleData *FirstInBundle;
698
699 /// Single linked list of all instructions in a bundle. Null if it is a
700 /// single instruction.
701 ScheduleData *NextInBundle;
702
703 /// Single linked list of all memory instructions (e.g. load, store, call)
704 /// in the block - until the end of the scheduling region.
705 ScheduleData *NextLoadStore;
706
707 /// The dependent memory instructions.
708 /// This list is derived on demand in calculateDependencies().
709 SmallVector<ScheduleData *, 4> MemoryDependencies;
710
711 /// This ScheduleData is in the current scheduling region if this matches
712 /// the current SchedulingRegionID of BlockScheduling.
713 int SchedulingRegionID;
714
715 /// Used for getting a "good" final ordering of instructions.
716 int SchedulingPriority;
717
718 /// The number of dependencies. Constitutes of the number of users of the
719 /// instruction plus the number of dependent memory instructions (if any).
720 /// This value is calculated on demand.
721 /// If InvalidDeps, the number of dependencies is not calculated yet.
722 ///
723 int Dependencies;
724
725 /// The number of dependencies minus the number of dependencies of scheduled
726 /// instructions. As soon as this is zero, the instruction/bundle gets ready
727 /// for scheduling.
728 /// Note that this is negative as long as Dependencies is not calculated.
729 int UnscheduledDeps;
730
731 /// The sum of UnscheduledDeps in a bundle. Equals to UnscheduledDeps for
732 /// single instructions.
733 int UnscheduledDepsInBundle;
734
735 /// True if this instruction is scheduled (or considered as scheduled in the
736 /// dry-run).
737 bool IsScheduled;
738 };
739
740#ifndef NDEBUG
741 friend inline raw_ostream &operator<<(raw_ostream &os,
742 const BoUpSLP::ScheduleData &SD) {
743 SD.dump(os);
744 return os;
745 }
746#endif
747 friend struct GraphTraits<BoUpSLP *>;
748 friend struct DOTGraphTraits<BoUpSLP *>;
749
750 /// Contains all scheduling data for a basic block.
751 ///
752 struct BlockScheduling {
753
754 BlockScheduling(BasicBlock *BB)
755 : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize),
756 ScheduleStart(nullptr), ScheduleEnd(nullptr),
757 FirstLoadStoreInRegion(nullptr), LastLoadStoreInRegion(nullptr),
758 ScheduleRegionSize(0),
759 ScheduleRegionSizeLimit(ScheduleRegionSizeBudget),
760 // Make sure that the initial SchedulingRegionID is greater than the
761 // initial SchedulingRegionID in ScheduleData (which is 0).
762 SchedulingRegionID(1) {}
763
764 void clear() {
765 ReadyInsts.clear();
766 ScheduleStart = nullptr;
767 ScheduleEnd = nullptr;
768 FirstLoadStoreInRegion = nullptr;
769 LastLoadStoreInRegion = nullptr;
770
771 // Reduce the maximum schedule region size by the size of the
772 // previous scheduling run.
773 ScheduleRegionSizeLimit -= ScheduleRegionSize;
774 if (ScheduleRegionSizeLimit < MinScheduleRegionSize)
775 ScheduleRegionSizeLimit = MinScheduleRegionSize;
776 ScheduleRegionSize = 0;
777
778 // Make a new scheduling region, i.e. all existing ScheduleData is not
779 // in the new region yet.
780 ++SchedulingRegionID;
781 }
782
783 ScheduleData *getScheduleData(Value *V) {
784 ScheduleData *SD = ScheduleDataMap[V];
785 if (SD && SD->SchedulingRegionID == SchedulingRegionID)
786 return SD;
787 return nullptr;
788 }
789
790 bool isInSchedulingRegion(ScheduleData *SD) {
791 return SD->SchedulingRegionID == SchedulingRegionID;
792 }
793
794 /// Marks an instruction as scheduled and puts all dependent ready
795 /// instructions into the ready-list.
796 template <typename ReadyListType>
797 void schedule(ScheduleData *SD, ReadyListType &ReadyList) {
798 SD->IsScheduled = true;
799 DEBUG(dbgs() << "SLP: schedule " << *SD << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: schedule " << *SD <<
"\n"; } } while (false)
;
800
801 ScheduleData *BundleMember = SD;
802 while (BundleMember) {
803 // Handle the def-use chain dependencies.
804 for (Use &U : BundleMember->Inst->operands()) {
805 ScheduleData *OpDef = getScheduleData(U.get());
806 if (OpDef && OpDef->hasValidDependencies() &&
807 OpDef->incrementUnscheduledDeps(-1) == 0) {
808 // There are no more unscheduled dependencies after decrementing,
809 // so we can put the dependent instruction into the ready list.
810 ScheduleData *DepBundle = OpDef->FirstInBundle;
811 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 812, __PRETTY_FUNCTION__))
812 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 812, __PRETTY_FUNCTION__))
;
813 ReadyList.insert(DepBundle);
814 DEBUG(dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: gets ready (def): " <<
*DepBundle << "\n"; } } while (false)
;
815 }
816 }
817 // Handle the memory dependencies.
818 for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) {
819 if (MemoryDepSD->incrementUnscheduledDeps(-1) == 0) {
820 // There are no more unscheduled dependencies after decrementing,
821 // so we can put the dependent instruction into the ready list.
822 ScheduleData *DepBundle = MemoryDepSD->FirstInBundle;
823 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 824, __PRETTY_FUNCTION__))
824 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 824, __PRETTY_FUNCTION__))
;
825 ReadyList.insert(DepBundle);
826 DEBUG(dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: gets ready (mem): " <<
*DepBundle << "\n"; } } while (false)
;
827 }
828 }
829 BundleMember = BundleMember->NextInBundle;
830 }
831 }
832
833 /// Put all instructions into the ReadyList which are ready for scheduling.
834 template <typename ReadyListType>
835 void initialFillReadyList(ReadyListType &ReadyList) {
836 for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
837 ScheduleData *SD = getScheduleData(I);
838 if (SD->isSchedulingEntity() && SD->isReady()) {
839 ReadyList.insert(SD);
840 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)
;
841 }
842 }
843 }
844
845 /// Checks if a bundle of instructions can be scheduled, i.e. has no
846 /// cyclic dependencies. This is only a dry-run, no instructions are
847 /// actually moved at this stage.
848 bool tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP);
849
850 /// Un-bundles a group of instructions.
851 void cancelScheduling(ArrayRef<Value *> VL);
852
853 /// Extends the scheduling region so that V is inside the region.
854 /// \returns true if the region size is within the limit.
855 bool extendSchedulingRegion(Value *V);
856
857 /// Initialize the ScheduleData structures for new instructions in the
858 /// scheduling region.
859 void initScheduleData(Instruction *FromI, Instruction *ToI,
860 ScheduleData *PrevLoadStore,
861 ScheduleData *NextLoadStore);
862
863 /// Updates the dependency information of a bundle and of all instructions/
864 /// bundles which depend on the original bundle.
865 void calculateDependencies(ScheduleData *SD, bool InsertInReadyList,
866 BoUpSLP *SLP);
867
868 /// Sets all instruction in the scheduling region to un-scheduled.
869 void resetSchedule();
870
871 BasicBlock *BB;
872
873 /// Simple memory allocation for ScheduleData.
874 std::vector<std::unique_ptr<ScheduleData[]>> ScheduleDataChunks;
875
876 /// The size of a ScheduleData array in ScheduleDataChunks.
877 int ChunkSize;
878
879 /// The allocator position in the current chunk, which is the last entry
880 /// of ScheduleDataChunks.
881 int ChunkPos;
882
883 /// Attaches ScheduleData to Instruction.
884 /// Note that the mapping survives during all vectorization iterations, i.e.
885 /// ScheduleData structures are recycled.
886 DenseMap<Value *, ScheduleData *> ScheduleDataMap;
887
888 struct ReadyList : SmallVector<ScheduleData *, 8> {
889 void insert(ScheduleData *SD) { push_back(SD); }
890 };
891
892 /// The ready-list for scheduling (only used for the dry-run).
893 ReadyList ReadyInsts;
894
895 /// The first instruction of the scheduling region.
896 Instruction *ScheduleStart;
897
898 /// The first instruction _after_ the scheduling region.
899 Instruction *ScheduleEnd;
900
901 /// The first memory accessing instruction in the scheduling region
902 /// (can be null).
903 ScheduleData *FirstLoadStoreInRegion;
904
905 /// The last memory accessing instruction in the scheduling region
906 /// (can be null).
907 ScheduleData *LastLoadStoreInRegion;
908
909 /// The current size of the scheduling region.
910 int ScheduleRegionSize;
911
912 /// The maximum size allowed for the scheduling region.
913 int ScheduleRegionSizeLimit;
914
915 /// The ID of the scheduling region. For a new vectorization iteration this
916 /// is incremented which "removes" all ScheduleData from the region.
917 int SchedulingRegionID;
918 };
919
920 /// Attaches the BlockScheduling structures to basic blocks.
921 MapVector<BasicBlock *, std::unique_ptr<BlockScheduling>> BlocksSchedules;
922
923 /// Performs the "real" scheduling. Done before vectorization is actually
924 /// performed in a basic block.
925 void scheduleBlock(BlockScheduling *BS);
926
927 /// List of users to ignore during scheduling and that don't need extracting.
928 ArrayRef<Value *> UserIgnoreList;
929
930 // Number of load bundles that contain consecutive loads.
931 int NumLoadsWantToKeepOrder;
932
933 // Number of load bundles that contain consecutive loads in reversed order.
934 int NumLoadsWantToChangeOrder;
935
936 // Analysis and block reference.
937 Function *F;
938 ScalarEvolution *SE;
939 TargetTransformInfo *TTI;
940 TargetLibraryInfo *TLI;
941 AliasAnalysis *AA;
942 LoopInfo *LI;
943 DominatorTree *DT;
944 AssumptionCache *AC;
945 DemandedBits *DB;
946 const DataLayout *DL;
947 unsigned MaxVecRegSize; // This is set by TTI or overridden by cl::opt.
948 unsigned MinVecRegSize; // Set by cl::opt (default: 128).
949 /// Instruction builder to construct the vectorized tree.
950 IRBuilder<> Builder;
951
952 /// A map of scalar integer values to the smallest bit width with which they
953 /// can legally be represented. The values map to (width, signed) pairs,
954 /// where "width" indicates the minimum bit width and "signed" is True if the
955 /// value must be signed-extended, rather than zero-extended, back to its
956 /// original width.
957 MapVector<Value *, std::pair<uint64_t, bool>> MinBWs;
958};
959} // end namespace slpvectorizer
960
961template <> struct GraphTraits<BoUpSLP *> {
962 typedef BoUpSLP::TreeEntry TreeEntry;
963
964 /// NodeRef has to be a pointer per the GraphWriter.
965 typedef TreeEntry *NodeRef;
966
967 /// \brief Add the VectorizableTree to the index iterator to be able to return
968 /// TreeEntry pointers.
969 struct ChildIteratorType
970 : public iterator_adaptor_base<ChildIteratorType,
971 SmallVector<int, 1>::iterator> {
972
973 std::vector<TreeEntry> &VectorizableTree;
974
975 ChildIteratorType(SmallVector<int, 1>::iterator W,
976 std::vector<TreeEntry> &VT)
977 : ChildIteratorType::iterator_adaptor_base(W), VectorizableTree(VT) {}
978
979 NodeRef operator*() { return &VectorizableTree[*I]; }
980 };
981
982 static NodeRef getEntryNode(BoUpSLP &R) { return &R.VectorizableTree[0]; }
983
984 static ChildIteratorType child_begin(NodeRef N) {
985 return {N->UserTreeIndices.begin(), N->Container};
986 }
987 static ChildIteratorType child_end(NodeRef N) {
988 return {N->UserTreeIndices.end(), N->Container};
989 }
990
991 /// For the node iterator we just need to turn the TreeEntry iterator into a
992 /// TreeEntry* iterator so that it dereferences to NodeRef.
993 typedef pointer_iterator<std::vector<TreeEntry>::iterator> nodes_iterator;
994
995 static nodes_iterator nodes_begin(BoUpSLP *R) {
996 return nodes_iterator(R->VectorizableTree.begin());
997 }
998 static nodes_iterator nodes_end(BoUpSLP *R) {
999 return nodes_iterator(R->VectorizableTree.end());
1000 }
1001
1002 static unsigned size(BoUpSLP *R) { return R->VectorizableTree.size(); }
1003};
1004
1005template <> struct DOTGraphTraits<BoUpSLP *> : public DefaultDOTGraphTraits {
1006 typedef BoUpSLP::TreeEntry TreeEntry;
1007
1008 DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {}
1009
1010 std::string getNodeLabel(const TreeEntry *Entry, const BoUpSLP *R) {
1011 std::string Str;
1012 raw_string_ostream OS(Str);
1013 if (isSplat(Entry->Scalars)) {
1014 OS << "<splat> " << *Entry->Scalars[0];
1015 return Str;
1016 }
1017 for (auto V : Entry->Scalars) {
1018 OS << *V;
1019 if (std::any_of(
1020 R->ExternalUses.begin(), R->ExternalUses.end(),
1021 [&](const BoUpSLP::ExternalUser &EU) { return EU.Scalar == V; }))
1022 OS << " <extract>";
1023 OS << "\n";
1024 }
1025 return Str;
1026 }
1027
1028 static std::string getNodeAttributes(const TreeEntry *Entry,
1029 const BoUpSLP *) {
1030 if (Entry->NeedToGather)
1031 return "color=red";
1032 return "";
1033 }
1034};
1035
1036} // end namespace llvm
1037
1038void BoUpSLP::buildTree(ArrayRef<Value *> Roots,
1039 ArrayRef<Value *> UserIgnoreLst) {
1040 ExtraValueToDebugLocsMap ExternallyUsedValues;
1041 buildTree(Roots, ExternallyUsedValues, UserIgnoreLst);
1042}
1043void BoUpSLP::buildTree(ArrayRef<Value *> Roots,
1044 ExtraValueToDebugLocsMap &ExternallyUsedValues,
1045 ArrayRef<Value *> UserIgnoreLst) {
1046 deleteTree();
1047 UserIgnoreList = UserIgnoreLst;
1048 if (!allSameType(Roots))
1049 return;
1050 buildTree_rec(Roots, 0, -1);
1051
1052 // Collect the values that we need to extract from the tree.
1053 for (TreeEntry &EIdx : VectorizableTree) {
1054 TreeEntry *Entry = &EIdx;
1055
1056 // For each lane:
1057 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
1058 Value *Scalar = Entry->Scalars[Lane];
1059
1060 // No need to handle users of gathered values.
1061 if (Entry->NeedToGather)
1062 continue;
1063
1064 // Check if the scalar is externally used as an extra arg.
1065 auto ExtI = ExternallyUsedValues.find(Scalar);
1066 if (ExtI != ExternallyUsedValues.end()) {
1067 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)
1068 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)
;
1069 ExternalUses.emplace_back(Scalar, nullptr, Lane);
1070 continue;
1071 }
1072 for (User *U : Scalar->users()) {
1073 DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Checking user:" << *U <<
".\n"; } } while (false)
;
1074
1075 Instruction *UserInst = dyn_cast<Instruction>(U);
1076 if (!UserInst)
1077 continue;
1078
1079 // Skip in-tree scalars that become vectors
1080 if (ScalarToTreeEntry.count(U)) {
1081 int Idx = ScalarToTreeEntry[U];
1082 TreeEntry *UseEntry = &VectorizableTree[Idx];
1083 Value *UseScalar = UseEntry->Scalars[0];
1084 // Some in-tree scalars will remain as scalar in vectorized
1085 // instructions. If that is the case, the one in Lane 0 will
1086 // be used.
1087 if (UseScalar != U ||
1088 !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) {
1089 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)
1090 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: \tInternal user will be removed:"
<< *U << ".\n"; } } while (false)
;
1091 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1091, __PRETTY_FUNCTION__))
;
1092 continue;
1093 }
1094 }
1095
1096 // Ignore users in the user ignore list.
1097 if (is_contained(UserIgnoreList, UserInst))
1098 continue;
1099
1100 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)
1101 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)
;
1102 ExternalUses.push_back(ExternalUser(Scalar, U, Lane));
1103 }
1104 }
1105 }
1106}
1107
1108void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
1109 int UserTreeIdx) {
1110 bool isAltShuffle = false;
1111 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1111, __PRETTY_FUNCTION__))
;
1112
1113 if (Depth == RecursionMaxDepth) {
1114 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)
;
1115 newTreeEntry(VL, false, UserTreeIdx);
1116 return;
1117 }
1118
1119 // Don't handle vectors.
1120 if (VL[0]->getType()->isVectorTy()) {
1121 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)
;
1122 newTreeEntry(VL, false, UserTreeIdx);
1123 return;
1124 }
1125
1126 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1127 if (SI->getValueOperand()->getType()->isVectorTy()) {
1128 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)
;
1129 newTreeEntry(VL, false, UserTreeIdx);
1130 return;
1131 }
1132 unsigned Opcode = getSameOpcode(VL);
1133
1134 // Check that this shuffle vector refers to the alternate
1135 // sequence of opcodes.
1136 if (Opcode == Instruction::ShuffleVector) {
1137 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
1138 unsigned Op = I0->getOpcode();
1139 if (Op != Instruction::ShuffleVector)
1140 isAltShuffle = true;
1141 }
1142
1143 // If all of the operands are identical or constant we have a simple solution.
1144 if (allConstant(VL) || isSplat(VL) || !allSameBlock(VL) || !Opcode) {
1145 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)
;
1146 newTreeEntry(VL, false, UserTreeIdx);
1147 return;
1148 }
1149
1150 // We now know that this is a vector of instructions of the same type from
1151 // the same block.
1152
1153 // Don't vectorize ephemeral values.
1154 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
1155 if (EphValues.count(VL[i])) {
1156 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)
1157 ") is ephemeral.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: The instruction (" << *
VL[i] << ") is ephemeral.\n"; } } while (false)
;
1158 newTreeEntry(VL, false, UserTreeIdx);
1159 return;
1160 }
1161 }
1162
1163 // Check if this is a duplicate of another entry.
1164 if (ScalarToTreeEntry.count(VL[0])) {
1165 int Idx = ScalarToTreeEntry[VL[0]];
1166 TreeEntry *E = &VectorizableTree[Idx];
1167 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
1168 DEBUG(dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: \tChecking bundle: " <<
*VL[i] << ".\n"; } } while (false)
;
1169 if (E->Scalars[i] != VL[i]) {
1170 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)
;
1171 newTreeEntry(VL, false, UserTreeIdx);
1172 return;
1173 }
1174 }
1175 // Record the reuse of the tree node. FIXME, currently this is only used to
1176 // properly draw the graph rather than for the actual vectorization.
1177 E->UserTreeIndices.push_back(UserTreeIdx);
1178 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)
;
1179 return;
1180 }
1181
1182 // Check that none of the instructions in the bundle are already in the tree.
1183 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
1184 if (ScalarToTreeEntry.count(VL[i])) {
1185 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)
1186 ") 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)
;
1187 newTreeEntry(VL, false, UserTreeIdx);
1188 return;
1189 }
1190 }
1191
1192 // If any of the scalars is marked as a value that needs to stay scalar then
1193 // we need to gather the scalars.
1194 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
1195 if (MustGather.count(VL[i])) {
1196 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)
;
1197 newTreeEntry(VL, false, UserTreeIdx);
1198 return;
1199 }
1200 }
1201
1202 // Check that all of the users of the scalars that we want to vectorize are
1203 // schedulable.
1204 Instruction *VL0 = cast<Instruction>(VL[0]);
1205 BasicBlock *BB = cast<Instruction>(VL0)->getParent();
1206
1207 if (!DT->isReachableFromEntry(BB)) {
1208 // Don't go into unreachable blocks. They may contain instructions with
1209 // dependency cycles which confuse the final scheduling.
1210 DEBUG(dbgs() << "SLP: bundle in unreachable block.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: bundle in unreachable block.\n"
; } } while (false)
;
1211 newTreeEntry(VL, false, UserTreeIdx);
1212 return;
1213 }
1214
1215 // Check that every instructions appears once in this bundle.
1216 for (unsigned i = 0, e = VL.size(); i < e; ++i)
1217 for (unsigned j = i+1; j < e; ++j)
1218 if (VL[i] == VL[j]) {
1219 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)
;
1220 newTreeEntry(VL, false, UserTreeIdx);
1221 return;
1222 }
1223
1224 auto &BSRef = BlocksSchedules[BB];
1225 if (!BSRef) {
1226 BSRef = llvm::make_unique<BlockScheduling>(BB);
1227 }
1228 BlockScheduling &BS = *BSRef.get();
1229
1230 if (!BS.tryScheduleBundle(VL, this)) {
1231 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)
;
1232 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1234, __PRETTY_FUNCTION__))
1233 !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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1234, __PRETTY_FUNCTION__))
1234 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1234, __PRETTY_FUNCTION__))
;
1235 newTreeEntry(VL, false, UserTreeIdx);
1236 return;
1237 }
1238 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)
;
1239
1240 switch (Opcode) {
1241 case Instruction::PHI: {
1242 PHINode *PH = dyn_cast<PHINode>(VL0);
1243
1244 // Check for terminator values (e.g. invoke).
1245 for (unsigned j = 0; j < VL.size(); ++j)
1246 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1247 TerminatorInst *Term = dyn_cast<TerminatorInst>(
1248 cast<PHINode>(VL[j])->getIncomingValueForBlock(PH->getIncomingBlock(i)));
1249 if (Term) {
1250 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)
;
1251 BS.cancelScheduling(VL);
1252 newTreeEntry(VL, false, UserTreeIdx);
1253 return;
1254 }
1255 }
1256
1257 newTreeEntry(VL, true, UserTreeIdx);
1258 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)
;
1259
1260 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1261 ValueList Operands;
1262 // Prepare the operand vector.
1263 for (Value *j : VL)
1264 Operands.push_back(cast<PHINode>(j)->getIncomingValueForBlock(
1265 PH->getIncomingBlock(i)));
1266
1267 buildTree_rec(Operands, Depth + 1, UserTreeIdx);
1268 }
1269 return;
1270 }
1271 case Instruction::ExtractValue:
1272 case Instruction::ExtractElement: {
1273 bool Reuse = canReuseExtract(VL, Opcode);
1274 if (Reuse) {
1275 DEBUG(dbgs() << "SLP: Reusing extract sequence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Reusing extract sequence.\n"
; } } while (false)
;
1276 } else {
1277 BS.cancelScheduling(VL);
1278 }
1279 newTreeEntry(VL, Reuse, UserTreeIdx);
1280 return;
1281 }
1282 case Instruction::Load: {
1283 // Check that a vectorized load would load the same memory as a scalar
1284 // load.
1285 // For example we don't want vectorize loads that are smaller than 8 bit.
1286 // Even though we have a packed struct {<i2, i2, i2, i2>} LLVM treats
1287 // loading/storing it as an i8 struct. If we vectorize loads/stores from
1288 // such a struct we read/write packed bits disagreeing with the
1289 // unvectorized version.
1290 Type *ScalarTy = VL[0]->getType();
1291
1292 if (DL->getTypeSizeInBits(ScalarTy) !=
1293 DL->getTypeAllocSizeInBits(ScalarTy)) {
1294 BS.cancelScheduling(VL);
1295 newTreeEntry(VL, false, UserTreeIdx);
1296 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)
;
1297 return;
1298 }
1299
1300 // Make sure all loads in the bundle are simple - we can't vectorize
1301 // atomic or volatile loads.
1302 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) {
1303 LoadInst *L = cast<LoadInst>(VL[i]);
1304 if (!L->isSimple()) {
1305 BS.cancelScheduling(VL);
1306 newTreeEntry(VL, false, UserTreeIdx);
1307 DEBUG(dbgs() << "SLP: Gathering non-simple loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering non-simple loads.\n"
; } } while (false)
;
1308 return;
1309 }
1310 }
1311
1312 // Check if the loads are consecutive, reversed, or neither.
1313 // TODO: What we really want is to sort the loads, but for now, check
1314 // the two likely directions.
1315 bool Consecutive = true;
1316 bool ReverseConsecutive = true;
1317 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) {
1318 if (!isConsecutiveAccess(VL[i], VL[i + 1], *DL, *SE)) {
1319 Consecutive = false;
1320 break;
1321 } else {
1322 ReverseConsecutive = false;
1323 }
1324 }
1325
1326 if (Consecutive) {
1327 ++NumLoadsWantToKeepOrder;
1328 newTreeEntry(VL, true, UserTreeIdx);
1329 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)
;
1330 return;
1331 }
1332
1333 // If none of the load pairs were consecutive when checked in order,
1334 // check the reverse order.
1335 if (ReverseConsecutive)
1336 for (unsigned i = VL.size() - 1; i > 0; --i)
1337 if (!isConsecutiveAccess(VL[i], VL[i - 1], *DL, *SE)) {
1338 ReverseConsecutive = false;
1339 break;
1340 }
1341
1342 BS.cancelScheduling(VL);
1343 newTreeEntry(VL, false, UserTreeIdx);
1344
1345 if (ReverseConsecutive) {
1346 ++NumLoadsWantToChangeOrder;
1347 DEBUG(dbgs() << "SLP: Gathering reversed loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering reversed loads.\n"
; } } while (false)
;
1348 } else {
1349 DEBUG(dbgs() << "SLP: Gathering non-consecutive loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering non-consecutive loads.\n"
; } } while (false)
;
1350 }
1351 return;
1352 }
1353 case Instruction::ZExt:
1354 case Instruction::SExt:
1355 case Instruction::FPToUI:
1356 case Instruction::FPToSI:
1357 case Instruction::FPExt:
1358 case Instruction::PtrToInt:
1359 case Instruction::IntToPtr:
1360 case Instruction::SIToFP:
1361 case Instruction::UIToFP:
1362 case Instruction::Trunc:
1363 case Instruction::FPTrunc:
1364 case Instruction::BitCast: {
1365 Type *SrcTy = VL0->getOperand(0)->getType();
1366 for (unsigned i = 0; i < VL.size(); ++i) {
1367 Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType();
1368 if (Ty != SrcTy || !isValidElementType(Ty)) {
1369 BS.cancelScheduling(VL);
1370 newTreeEntry(VL, false, UserTreeIdx);
1371 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)
;
1372 return;
1373 }
1374 }
1375 newTreeEntry(VL, true, UserTreeIdx);
1376 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)
;
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, UserTreeIdx);
1385 }
1386 return;
1387 }
1388 case Instruction::ICmp:
1389 case Instruction::FCmp: {
1390 // Check that all of the compares have the same predicate.
1391 CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate();
1392 Type *ComparedTy = cast<Instruction>(VL[0])->getOperand(0)->getType();
1393 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
1394 CmpInst *Cmp = cast<CmpInst>(VL[i]);
1395 if (Cmp->getPredicate() != P0 ||
1396 Cmp->getOperand(0)->getType() != ComparedTy) {
1397 BS.cancelScheduling(VL);
1398 newTreeEntry(VL, false, UserTreeIdx);
1399 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)
;
1400 return;
1401 }
1402 }
1403
1404 newTreeEntry(VL, true, UserTreeIdx);
1405 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)
;
1406
1407 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
1408 ValueList Operands;
1409 // Prepare the operand vector.
1410 for (Value *j : VL)
1411 Operands.push_back(cast<Instruction>(j)->getOperand(i));
1412
1413 buildTree_rec(Operands, Depth + 1, UserTreeIdx);
1414 }
1415 return;
1416 }
1417 case Instruction::Select:
1418 case Instruction::Add:
1419 case Instruction::FAdd:
1420 case Instruction::Sub:
1421 case Instruction::FSub:
1422 case Instruction::Mul:
1423 case Instruction::FMul:
1424 case Instruction::UDiv:
1425 case Instruction::SDiv:
1426 case Instruction::FDiv:
1427 case Instruction::URem:
1428 case Instruction::SRem:
1429 case Instruction::FRem:
1430 case Instruction::Shl:
1431 case Instruction::LShr:
1432 case Instruction::AShr:
1433 case Instruction::And:
1434 case Instruction::Or:
1435 case Instruction::Xor: {
1436 newTreeEntry(VL, true, UserTreeIdx);
1437 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)
;
1438
1439 // Sort operands of the instructions so that each side is more likely to
1440 // have the same opcode.
1441 if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) {
1442 ValueList Left, Right;
1443 reorderInputsAccordingToOpcode(VL, Left, Right);
1444 buildTree_rec(Left, Depth + 1, UserTreeIdx);
1445 buildTree_rec(Right, Depth + 1, UserTreeIdx);
1446 return;
1447 }
1448
1449 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
1450 ValueList Operands;
1451 // Prepare the operand vector.
1452 for (Value *j : VL)
1453 Operands.push_back(cast<Instruction>(j)->getOperand(i));
1454
1455 buildTree_rec(Operands, Depth + 1, UserTreeIdx);
1456 }
1457 return;
1458 }
1459 case Instruction::GetElementPtr: {
1460 // We don't combine GEPs with complicated (nested) indexing.
1461 for (unsigned j = 0; j < VL.size(); ++j) {
1462 if (cast<Instruction>(VL[j])->getNumOperands() != 2) {
1463 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)
;
1464 BS.cancelScheduling(VL);
1465 newTreeEntry(VL, false, UserTreeIdx);
1466 return;
1467 }
1468 }
1469
1470 // We can't combine several GEPs into one vector if they operate on
1471 // different types.
1472 Type *Ty0 = cast<Instruction>(VL0)->getOperand(0)->getType();
1473 for (unsigned j = 0; j < VL.size(); ++j) {
1474 Type *CurTy = cast<Instruction>(VL[j])->getOperand(0)->getType();
1475 if (Ty0 != CurTy) {
1476 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)
;
1477 BS.cancelScheduling(VL);
1478 newTreeEntry(VL, false, UserTreeIdx);
1479 return;
1480 }
1481 }
1482
1483 // We don't combine GEPs with non-constant indexes.
1484 for (unsigned j = 0; j < VL.size(); ++j) {
1485 auto Op = cast<Instruction>(VL[j])->getOperand(1);
1486 if (!isa<ConstantInt>(Op)) {
1487 DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n"
; } } while (false)
1488 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)
;
1489 BS.cancelScheduling(VL);
1490 newTreeEntry(VL, false, UserTreeIdx);
1491 return;
1492 }
1493 }
1494
1495 newTreeEntry(VL, true, UserTreeIdx);
1496 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)
;
1497 for (unsigned i = 0, e = 2; i < e; ++i) {
1498 ValueList Operands;
1499 // Prepare the operand vector.
1500 for (Value *j : VL)
1501 Operands.push_back(cast<Instruction>(j)->getOperand(i));
1502
1503 buildTree_rec(Operands, Depth + 1, UserTreeIdx);
1504 }
1505 return;
1506 }
1507 case Instruction::Store: {
1508 // Check if the stores are consecutive or of we need to swizzle them.
1509 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
1510 if (!isConsecutiveAccess(VL[i], VL[i + 1], *DL, *SE)) {
1511 BS.cancelScheduling(VL);
1512 newTreeEntry(VL, false, UserTreeIdx);
1513 DEBUG(dbgs() << "SLP: Non-consecutive store.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Non-consecutive store.\n"; }
} while (false)
;
1514 return;
1515 }
1516
1517 newTreeEntry(VL, true, UserTreeIdx);
1518 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)
;
1519
1520 ValueList Operands;
1521 for (Value *j : VL)
1522 Operands.push_back(cast<Instruction>(j)->getOperand(0));
1523
1524 buildTree_rec(Operands, Depth + 1, UserTreeIdx);
1525 return;
1526 }
1527 case Instruction::Call: {
1528 // Check if the calls are all to the same vectorizable intrinsic.
1529 CallInst *CI = cast<CallInst>(VL[0]);
1530 // Check if this is an Intrinsic call or something that can be
1531 // represented by an intrinsic call
1532 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
1533 if (!isTriviallyVectorizable(ID)) {
1534 BS.cancelScheduling(VL);
1535 newTreeEntry(VL, false, UserTreeIdx);
1536 DEBUG(dbgs() << "SLP: Non-vectorizable call.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Non-vectorizable call.\n"; }
} while (false)
;
1537 return;
1538 }
1539 Function *Int = CI->getCalledFunction();
1540 Value *A1I = nullptr;
1541 if (hasVectorInstrinsicScalarOpd(ID, 1))
1542 A1I = CI->getArgOperand(1);
1543 for (unsigned i = 1, e = VL.size(); i != e; ++i) {
1544 CallInst *CI2 = dyn_cast<CallInst>(VL[i]);
1545 if (!CI2 || CI2->getCalledFunction() != Int ||
1546 getVectorIntrinsicIDForCall(CI2, TLI) != ID ||
1547 !CI->hasIdenticalOperandBundleSchema(*CI2)) {
1548 BS.cancelScheduling(VL);
1549 newTreeEntry(VL, false, UserTreeIdx);
1550 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
)
1551 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: mismatched calls:" << *
CI << "!=" << *VL[i] << "\n"; } } while (false
)
;
1552 return;
1553 }
1554 // ctlz,cttz and powi are special intrinsics whose second argument
1555 // should be same in order for them to be vectorized.
1556 if (hasVectorInstrinsicScalarOpd(ID, 1)) {
1557 Value *A1J = CI2->getArgOperand(1);
1558 if (A1I != A1J) {
1559 BS.cancelScheduling(VL);
1560 newTreeEntry(VL, false, UserTreeIdx);
1561 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)
1562 << " argument "<< A1I<<"!=" << A1Jdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: mismatched arguments in call:"
<< *CI << " argument "<< A1I<<"!=" <<
A1J << "\n"; } } while (false)
1563 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: mismatched arguments in call:"
<< *CI << " argument "<< A1I<<"!=" <<
A1J << "\n"; } } while (false)
;
1564 return;
1565 }
1566 }
1567 // Verify that the bundle operands are identical between the two calls.
1568 if (CI->hasOperandBundles() &&
1569 !std::equal(CI->op_begin() + CI->getBundleOperandsStartIndex(),
1570 CI->op_begin() + CI->getBundleOperandsEndIndex(),
1571 CI2->op_begin() + CI2->getBundleOperandsStartIndex())) {
1572 BS.cancelScheduling(VL);
1573 newTreeEntry(VL, false, UserTreeIdx);
1574 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)
1575 << *VL[i] << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:"
<< *CI << "!=" << *VL[i] << '\n'; } }
while (false)
;
1576 return;
1577 }
1578 }
1579
1580 newTreeEntry(VL, true, UserTreeIdx);
1581 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) {
1582 ValueList Operands;
1583 // Prepare the operand vector.
1584 for (Value *j : VL) {
1585 CallInst *CI2 = dyn_cast<CallInst>(j);
1586 Operands.push_back(CI2->getArgOperand(i));
1587 }
1588 buildTree_rec(Operands, Depth + 1, UserTreeIdx);
1589 }
1590 return;
1591 }
1592 case Instruction::ShuffleVector: {
1593 // If this is not an alternate sequence of opcode like add-sub
1594 // then do not vectorize this instruction.
1595 if (!isAltShuffle) {
1596 BS.cancelScheduling(VL);
1597 newTreeEntry(VL, false, UserTreeIdx);
1598 DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: ShuffleVector are not vectorized.\n"
; } } while (false)
;
1599 return;
1600 }
1601 newTreeEntry(VL, true, UserTreeIdx);
1602 DEBUG(dbgs() << "SLP: added a ShuffleVector op.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: added a ShuffleVector op.\n"
; } } while (false)
;
1603
1604 // Reorder operands if reordering would enable vectorization.
1605 if (isa<BinaryOperator>(VL0)) {
1606 ValueList Left, Right;
1607 reorderAltShuffleOperands(VL, Left, Right);
1608 buildTree_rec(Left, Depth + 1, UserTreeIdx);
1609 buildTree_rec(Right, Depth + 1, UserTreeIdx);
1610 return;
1611 }
1612
1613 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
1614 ValueList Operands;
1615 // Prepare the operand vector.
1616 for (Value *j : VL)
1617 Operands.push_back(cast<Instruction>(j)->getOperand(i));
1618
1619 buildTree_rec(Operands, Depth + 1, UserTreeIdx);
1620 }
1621 return;
1622 }
1623 default:
1624 BS.cancelScheduling(VL);
1625 newTreeEntry(VL, false, UserTreeIdx);
1626 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Gathering unknown instruction.\n"
; } } while (false)
;
1627 return;
1628 }
1629}
1630
1631unsigned BoUpSLP::canMapToVector(Type *T, const DataLayout &DL) const {
1632 unsigned N;
1633 Type *EltTy;
1634 auto *ST = dyn_cast<StructType>(T);
1635 if (ST) {
1636 N = ST->getNumElements();
1637 EltTy = *ST->element_begin();
1638 } else {
1639 N = cast<ArrayType>(T)->getNumElements();
1640 EltTy = cast<ArrayType>(T)->getElementType();
1641 }
1642 if (!isValidElementType(EltTy))
1643 return 0;
1644 uint64_t VTSize = DL.getTypeStoreSizeInBits(VectorType::get(EltTy, N));
1645 if (VTSize < MinVecRegSize || VTSize > MaxVecRegSize || VTSize != DL.getTypeStoreSizeInBits(T))
1646 return 0;
1647 if (ST) {
1648 // Check that struct is homogeneous.
1649 for (const auto *Ty : ST->elements())
1650 if (Ty != EltTy)
1651 return 0;
1652 }
1653 return N;
1654}
1655
1656bool BoUpSLP::canReuseExtract(ArrayRef<Value *> VL, unsigned Opcode) const {
1657 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1658, __PRETTY_FUNCTION__))
1658 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1658, __PRETTY_FUNCTION__))
;
1659 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1659, __PRETTY_FUNCTION__))
;
1660 // Check if all of the extracts come from the same vector and from the
1661 // correct offset.
1662 Value *VL0 = VL[0];
1663 Instruction *E0 = cast<Instruction>(VL0);
1664 Value *Vec = E0->getOperand(0);
1665
1666 // We have to extract from a vector/aggregate with the same number of elements.
1667 unsigned NElts;
1668 if (Opcode == Instruction::ExtractValue) {
1669 const DataLayout &DL = E0->getModule()->getDataLayout();
1670 NElts = canMapToVector(Vec->getType(), DL);
1671 if (!NElts)
1672 return false;
1673 // Check if load can be rewritten as load of vector.
1674 LoadInst *LI = dyn_cast<LoadInst>(Vec);
1675 if (!LI || !LI->isSimple() || !LI->hasNUses(VL.size()))
1676 return false;
1677 } else {
1678 NElts = Vec->getType()->getVectorNumElements();
1679 }
1680
1681 if (NElts != VL.size())
1682 return false;
1683
1684 // Check that all of the indices extract from the correct offset.
1685 if (!matchExtractIndex(E0, 0, Opcode))
1686 return false;
1687
1688 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
1689 Instruction *E = cast<Instruction>(VL[i]);
1690 if (!matchExtractIndex(E, i, Opcode))
1691 return false;
1692 if (E->getOperand(0) != Vec)
1693 return false;
1694 }
1695
1696 return true;
1697}
1698
1699int BoUpSLP::getEntryCost(TreeEntry *E) {
1700 ArrayRef<Value*> VL = E->Scalars;
1701
1702 Type *ScalarTy = VL[0]->getType();
1703 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1704 ScalarTy = SI->getValueOperand()->getType();
1705 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1706
1707 // If we have computed a smaller type for the expression, update VecTy so
1708 // that the costs will be accurate.
1709 if (MinBWs.count(VL[0]))
1710 VecTy = VectorType::get(
1711 IntegerType::get(F->getContext(), MinBWs[VL[0]].first), VL.size());
1712
1713 if (E->NeedToGather) {
1714 if (allConstant(VL))
1715 return 0;
1716 if (isSplat(VL)) {
1717 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
1718 }
1719 return getGatherCost(E->Scalars);
1720 }
1721 unsigned Opcode = getSameOpcode(VL);
1722 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1722, __PRETTY_FUNCTION__))
;
1723 Instruction *VL0 = cast<Instruction>(VL[0]);
1724 switch (Opcode) {
1725 case Instruction::PHI: {
1726 return 0;
1727 }
1728 case Instruction::ExtractValue:
1729 case Instruction::ExtractElement: {
1730 if (canReuseExtract(VL, Opcode)) {
1731 int DeadCost = 0;
1732 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
1733 Instruction *E = cast<Instruction>(VL[i]);
1734 // If all users are going to be vectorized, instruction can be
1735 // considered as dead.
1736 // The same, if have only one user, it will be vectorized for sure.
1737 if (E->hasOneUse() ||
1738 std::all_of(E->user_begin(), E->user_end(), [this](User *U) {
1739 return ScalarToTreeEntry.count(U) > 0;
1740 }))
1741 // Take credit for instruction that will become dead.
1742 DeadCost +=
1743 TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, i);
1744 }
1745 return -DeadCost;
1746 }
1747 return getGatherCost(VecTy);
1748 }
1749 case Instruction::ZExt:
1750 case Instruction::SExt:
1751 case Instruction::FPToUI:
1752 case Instruction::FPToSI:
1753 case Instruction::FPExt:
1754 case Instruction::PtrToInt:
1755 case Instruction::IntToPtr:
1756 case Instruction::SIToFP:
1757 case Instruction::UIToFP:
1758 case Instruction::Trunc:
1759 case Instruction::FPTrunc:
1760 case Instruction::BitCast: {
1761 Type *SrcTy = VL0->getOperand(0)->getType();
1762
1763 // Calculate the cost of this instruction.
1764 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
1765 VL0->getType(), SrcTy);
1766
1767 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
1768 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
1769 return VecCost - ScalarCost;
1770 }
1771 case Instruction::FCmp:
1772 case Instruction::ICmp:
1773 case Instruction::Select: {
1774 // Calculate the cost of this instruction.
1775 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
1776 int ScalarCost = VecTy->getNumElements() *
1777 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
1778 int VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
1779 return VecCost - ScalarCost;
1780 }
1781 case Instruction::Add:
1782 case Instruction::FAdd:
1783 case Instruction::Sub:
1784 case Instruction::FSub:
1785 case Instruction::Mul:
1786 case Instruction::FMul:
1787 case Instruction::UDiv:
1788 case Instruction::SDiv:
1789 case Instruction::FDiv:
1790 case Instruction::URem:
1791 case Instruction::SRem:
1792 case Instruction::FRem:
1793 case Instruction::Shl:
1794 case Instruction::LShr:
1795 case Instruction::AShr:
1796 case Instruction::And:
1797 case Instruction::Or:
1798 case Instruction::Xor: {
1799 // Certain instructions can be cheaper to vectorize if they have a
1800 // constant second vector operand.
1801 TargetTransformInfo::OperandValueKind Op1VK =
1802 TargetTransformInfo::OK_AnyValue;
1803 TargetTransformInfo::OperandValueKind Op2VK =
1804 TargetTransformInfo::OK_UniformConstantValue;
1805 TargetTransformInfo::OperandValueProperties Op1VP =
1806 TargetTransformInfo::OP_None;
1807 TargetTransformInfo::OperandValueProperties Op2VP =
1808 TargetTransformInfo::OP_None;
1809
1810 // If all operands are exactly the same ConstantInt then set the
1811 // operand kind to OK_UniformConstantValue.
1812 // If instead not all operands are constants, then set the operand kind
1813 // to OK_AnyValue. If all operands are constants but not the same,
1814 // then set the operand kind to OK_NonUniformConstantValue.
1815 ConstantInt *CInt = nullptr;
1816 for (unsigned i = 0; i < VL.size(); ++i) {
1817 const Instruction *I = cast<Instruction>(VL[i]);
1818 if (!isa<ConstantInt>(I->getOperand(1))) {
1819 Op2VK = TargetTransformInfo::OK_AnyValue;
1820 break;
1821 }
1822 if (i == 0) {
1823 CInt = cast<ConstantInt>(I->getOperand(1));
1824 continue;
1825 }
1826 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue &&
1827 CInt != cast<ConstantInt>(I->getOperand(1)))
1828 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1829 }
1830 // FIXME: Currently cost of model modification for division by power of
1831 // 2 is handled for X86 and AArch64. Add support for other targets.
1832 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue && CInt &&
1833 CInt->getValue().isPowerOf2())
1834 Op2VP = TargetTransformInfo::OP_PowerOf2;
1835
1836 int ScalarCost = VecTy->getNumElements() *
1837 TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK,
1838 Op2VK, Op1VP, Op2VP);
1839 int VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK,
1840 Op1VP, Op2VP);
1841 return VecCost - ScalarCost;
1842 }
1843 case Instruction::GetElementPtr: {
1844 TargetTransformInfo::OperandValueKind Op1VK =
1845 TargetTransformInfo::OK_AnyValue;
1846 TargetTransformInfo::OperandValueKind Op2VK =
1847 TargetTransformInfo::OK_UniformConstantValue;
1848
1849 int ScalarCost =
1850 VecTy->getNumElements() *
1851 TTI->getArithmeticInstrCost(Instruction::Add, ScalarTy, Op1VK, Op2VK);
1852 int VecCost =
1853 TTI->getArithmeticInstrCost(Instruction::Add, VecTy, Op1VK, Op2VK);
1854
1855 return VecCost - ScalarCost;
1856 }
1857 case Instruction::Load: {
1858 // Cost of wide load - cost of scalar loads.
1859 unsigned alignment = dyn_cast<LoadInst>(VL0)->getAlignment();
1860 int ScalarLdCost = VecTy->getNumElements() *
1861 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, alignment, 0);
1862 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load,
1863 VecTy, alignment, 0);
1864 return VecLdCost - ScalarLdCost;
1865 }
1866 case Instruction::Store: {
1867 // We know that we can merge the stores. Calculate the cost.
1868 unsigned alignment = dyn_cast<StoreInst>(VL0)->getAlignment();
1869 int ScalarStCost = VecTy->getNumElements() *
1870 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, alignment, 0);
1871 int VecStCost = TTI->getMemoryOpCost(Instruction::Store,
1872 VecTy, alignment, 0);
1873 return VecStCost - ScalarStCost;
1874 }
1875 case Instruction::Call: {
1876 CallInst *CI = cast<CallInst>(VL0);
1877 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
1878
1879 // Calculate the cost of the scalar and vector calls.
1880 SmallVector<Type*, 4> ScalarTys;
1881 for (unsigned op = 0, opc = CI->getNumArgOperands(); op!= opc; ++op)
1882 ScalarTys.push_back(CI->getArgOperand(op)->getType());
1883
1884 FastMathFlags FMF;
1885 if (auto *FPMO = dyn_cast<FPMathOperator>(CI))
1886 FMF = FPMO->getFastMathFlags();
1887
1888 int ScalarCallCost = VecTy->getNumElements() *
1889 TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys, FMF);
1890
1891 SmallVector<Value *, 4> Args(CI->arg_operands());
1892 int VecCallCost = TTI->getIntrinsicInstrCost(ID, CI->getType(), Args, FMF,
1893 VecTy->getNumElements());
1894
1895 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)
1896 << " (" << VecCallCost << "-" << ScalarCallCost << ")"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Call cost "<< VecCallCost
- ScalarCallCost << " (" << VecCallCost <<
"-" << ScalarCallCost << ")" << " for " <<
*CI << "\n"; } } while (false)
1897 << " for " << *CI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Call cost "<< VecCallCost
- ScalarCallCost << " (" << VecCallCost <<
"-" << ScalarCallCost << ")" << " for " <<
*CI << "\n"; } } while (false)
;
1898
1899 return VecCallCost - ScalarCallCost;
1900 }
1901 case Instruction::ShuffleVector: {
1902 TargetTransformInfo::OperandValueKind Op1VK =
1903 TargetTransformInfo::OK_AnyValue;
1904 TargetTransformInfo::OperandValueKind Op2VK =
1905 TargetTransformInfo::OK_AnyValue;
1906 int ScalarCost = 0;
1907 int VecCost = 0;
1908 for (Value *i : VL) {
1909 Instruction *I = cast<Instruction>(i);
1910 if (!I)
1911 break;
1912 ScalarCost +=
1913 TTI->getArithmeticInstrCost(I->getOpcode(), ScalarTy, Op1VK, Op2VK);
1914 }
1915 // VecCost is equal to sum of the cost of creating 2 vectors
1916 // and the cost of creating shuffle.
1917 Instruction *I0 = cast<Instruction>(VL[0]);
1918 VecCost =
1919 TTI->getArithmeticInstrCost(I0->getOpcode(), VecTy, Op1VK, Op2VK);
1920 Instruction *I1 = cast<Instruction>(VL[1]);
1921 VecCost +=
1922 TTI->getArithmeticInstrCost(I1->getOpcode(), VecTy, Op1VK, Op2VK);
1923 VecCost +=
1924 TTI->getShuffleCost(TargetTransformInfo::SK_Alternate, VecTy, 0);
1925 return VecCost - ScalarCost;
1926 }
1927 default:
1928 llvm_unreachable("Unknown instruction")::llvm::llvm_unreachable_internal("Unknown instruction", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1928)
;
1929 }
1930}
1931
1932bool BoUpSLP::isFullyVectorizableTinyTree() {
1933 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)
1934 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)
;
1935
1936 // We only handle trees of heights 1 and 2.
1937 if (VectorizableTree.size() == 1 && !VectorizableTree[0].NeedToGather)
1938 return true;
1939
1940 if (VectorizableTree.size() != 2)
1941 return false;
1942
1943 // Handle splat and all-constants stores.
1944 if (!VectorizableTree[0].NeedToGather &&
1945 (allConstant(VectorizableTree[1].Scalars) ||
1946 isSplat(VectorizableTree[1].Scalars)))
1947 return true;
1948
1949 // Gathering cost would be too much for tiny trees.
1950 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
1951 return false;
1952
1953 return true;
1954}
1955
1956bool BoUpSLP::isTreeTinyAndNotFullyVectorizable() {
1957
1958 // We can vectorize the tree if its size is greater than or equal to the
1959 // minimum size specified by the MinTreeSize command line option.
1960 if (VectorizableTree.size() >= MinTreeSize)
1961 return false;
1962
1963 // If we have a tiny tree (a tree whose size is less than MinTreeSize), we
1964 // can vectorize it if we can prove it fully vectorizable.
1965 if (isFullyVectorizableTinyTree())
1966 return false;
1967
1968 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1970, __PRETTY_FUNCTION__))
1969 ? 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1970, __PRETTY_FUNCTION__))
1970 : 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 1970, __PRETTY_FUNCTION__))
;
1971
1972 // Otherwise, we can't vectorize the tree. It is both tiny and not fully
1973 // vectorizable.
1974 return true;
1975}
1976
1977int BoUpSLP::getSpillCost() {
1978 // Walk from the bottom of the tree to the top, tracking which values are
1979 // live. When we see a call instruction that is not part of our tree,
1980 // query TTI to see if there is a cost to keeping values live over it
1981 // (for example, if spills and fills are required).
1982 unsigned BundleWidth = VectorizableTree.front().Scalars.size();
1983 int Cost = 0;
1984
1985 SmallPtrSet<Instruction*, 4> LiveValues;
1986 Instruction *PrevInst = nullptr;
1987
1988 for (const auto &N : VectorizableTree) {
1989 Instruction *Inst = dyn_cast<Instruction>(N.Scalars[0]);
1990 if (!Inst)
1991 continue;
1992
1993 if (!PrevInst) {
1994 PrevInst = Inst;
1995 continue;
1996 }
1997
1998 // Update LiveValues.
1999 LiveValues.erase(PrevInst);
2000 for (auto &J : PrevInst->operands()) {
2001 if (isa<Instruction>(&*J) && ScalarToTreeEntry.count(&*J))
2002 LiveValues.insert(cast<Instruction>(&*J));
2003 }
2004
2005 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)
2006 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)
2007 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)
2008 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)
2009 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)
2010 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)
2011 )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)
;
2012
2013 // Now find the sequence of instructions between PrevInst and Inst.
2014 BasicBlock::reverse_iterator InstIt = ++Inst->getIterator().getReverse(),
2015 PrevInstIt =
2016 PrevInst->getIterator().getReverse();
2017 while (InstIt != PrevInstIt) {
2018 if (PrevInstIt == PrevInst->getParent()->rend()) {
2019 PrevInstIt = Inst->getParent()->rbegin();
2020 continue;
2021 }
2022
2023 if (isa<CallInst>(&*PrevInstIt) && &*PrevInstIt != PrevInst) {
2024 SmallVector<Type*, 4> V;
2025 for (auto *II : LiveValues)
2026 V.push_back(VectorType::get(II->getType(), BundleWidth));
2027 Cost += TTI->getCostOfKeepingLiveOverCall(V);
2028 }
2029
2030 ++PrevInstIt;
2031 }
2032
2033 PrevInst = Inst;
2034 }
2035
2036 return Cost;
2037}
2038
2039int BoUpSLP::getTreeCost() {
2040 int Cost = 0;
2041 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)
2042 VectorizableTree.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Calculating cost for tree of size "
<< VectorizableTree.size() << ".\n"; } } while (
false)
;
2043
2044 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
2045
2046 for (TreeEntry &TE : VectorizableTree) {
2047 int C = getEntryCost(&TE);
2048 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)
2049 << *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)
;
2050 Cost += C;
2051 }
2052
2053 SmallSet<Value *, 16> ExtractCostCalculated;
2054 int ExtractCost = 0;
2055 for (ExternalUser &EU : ExternalUses) {
2056 // We only add extract cost once for the same scalar.
2057 if (!ExtractCostCalculated.insert(EU.Scalar).second)
2058 continue;
2059
2060 // Uses by ephemeral values are free (because the ephemeral value will be
2061 // removed prior to code generation, and so the extraction will be
2062 // removed as well).
2063 if (EphValues.count(EU.User))
2064 continue;
2065
2066 // If we plan to rewrite the tree in a smaller type, we will need to sign
2067 // extend the extracted value back to the original type. Here, we account
2068 // for the extract and the added cost of the sign extend if needed.
2069 auto *VecTy = VectorType::get(EU.Scalar->getType(), BundleWidth);
2070 auto *ScalarRoot = VectorizableTree[0].Scalars[0];
2071 if (MinBWs.count(ScalarRoot)) {
2072 auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first);
2073 auto Extend =
2074 MinBWs[ScalarRoot].second ? Instruction::SExt : Instruction::ZExt;
2075 VecTy = VectorType::get(MinTy, BundleWidth);
2076 ExtractCost += TTI->getExtractWithExtendCost(Extend, EU.Scalar->getType(),
2077 VecTy, EU.Lane);
2078 } else {
2079 ExtractCost +=
2080 TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, EU.Lane);
2081 }
2082 }
2083
2084 int SpillCost = getSpillCost();
2085 Cost += SpillCost + ExtractCost;
2086
2087 std::string Str;
2088 {
2089 raw_string_ostream OS(Str);
2090 OS << "SLP: Spill Cost = " << SpillCost << ".\n"
2091 << "SLP: Extract Cost = " << ExtractCost << ".\n"
2092 << "SLP: Total Cost = " << Cost << ".\n";
2093 }
2094 DEBUG(dbgs() << Str)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << Str; } } while (false)
;
2095
2096 if (ViewSLPTree)
2097 ViewGraph(this, "SLP" + F->getName(), false, Str);
2098
2099 return Cost;
2100}
2101
2102int BoUpSLP::getGatherCost(Type *Ty) {
2103 int Cost = 0;
2104 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
2105 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
2106 return Cost;
2107}
2108
2109int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
2110 // Find the type of the operands in VL.
2111 Type *ScalarTy = VL[0]->getType();
2112 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
2113 ScalarTy = SI->getValueOperand()->getType();
2114 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
2115 // Find the cost of inserting/extracting values from the vector.
2116 return getGatherCost(VecTy);
2117}
2118
2119// Reorder commutative operations in alternate shuffle if the resulting vectors
2120// are consecutive loads. This would allow us to vectorize the tree.
2121// If we have something like-
2122// load a[0] - load b[0]
2123// load b[1] + load a[1]
2124// load a[2] - load b[2]
2125// load a[3] + load b[3]
2126// Reordering the second load b[1] load a[1] would allow us to vectorize this
2127// code.
2128void BoUpSLP::reorderAltShuffleOperands(ArrayRef<Value *> VL,
2129 SmallVectorImpl<Value *> &Left,
2130 SmallVectorImpl<Value *> &Right) {
2131 // Push left and right operands of binary operation into Left and Right
2132 for (Value *i : VL) {
2133 Left.push_back(cast<Instruction>(i)->getOperand(0));
2134 Right.push_back(cast<Instruction>(i)->getOperand(1));
2135 }
2136
2137 // Reorder if we have a commutative operation and consecutive access
2138 // are on either side of the alternate instructions.
2139 for (unsigned j = 0; j < VL.size() - 1; ++j) {
2140 if (LoadInst *L = dyn_cast<LoadInst>(Left[j])) {
2141 if (LoadInst *L1 = dyn_cast<LoadInst>(Right[j + 1])) {
2142 Instruction *VL1 = cast<Instruction>(VL[j]);
2143 Instruction *VL2 = cast<Instruction>(VL[j + 1]);
2144 if (VL1->isCommutative() && isConsecutiveAccess(L, L1, *DL, *SE)) {
2145 std::swap(Left[j], Right[j]);
2146 continue;
2147 } else if (VL2->isCommutative() &&
2148 isConsecutiveAccess(L, L1, *DL, *SE)) {
2149 std::swap(Left[j + 1], Right[j + 1]);
2150 continue;
2151 }
2152 // else unchanged
2153 }
2154 }
2155 if (LoadInst *L = dyn_cast<LoadInst>(Right[j])) {
2156 if (LoadInst *L1 = dyn_cast<LoadInst>(Left[j + 1])) {
2157 Instruction *VL1 = cast<Instruction>(VL[j]);
2158 Instruction *VL2 = cast<Instruction>(VL[j + 1]);
2159 if (VL1->isCommutative() && isConsecutiveAccess(L, L1, *DL, *SE)) {
2160 std::swap(Left[j], Right[j]);
2161 continue;
2162 } else if (VL2->isCommutative() &&
2163 isConsecutiveAccess(L, L1, *DL, *SE)) {
2164 std::swap(Left[j + 1], Right[j + 1]);
2165 continue;
2166 }
2167 // else unchanged
2168 }
2169 }
2170 }
2171}
2172
2173// Return true if I should be commuted before adding it's left and right
2174// operands to the arrays Left and Right.
2175//
2176// The vectorizer is trying to either have all elements one side being
2177// instruction with the same opcode to enable further vectorization, or having
2178// a splat to lower the vectorizing cost.
2179static bool shouldReorderOperands(int i, Instruction &I,
2180 SmallVectorImpl<Value *> &Left,
2181 SmallVectorImpl<Value *> &Right,
2182 bool AllSameOpcodeLeft,
2183 bool AllSameOpcodeRight, bool SplatLeft,
2184 bool SplatRight) {
2185 Value *VLeft = I.getOperand(0);
2186 Value *VRight = I.getOperand(1);
2187 // If we have "SplatRight", try to see if commuting is needed to preserve it.
2188 if (SplatRight) {
2189 if (VRight == Right[i - 1])
2190 // Preserve SplatRight
2191 return false;
2192 if (VLeft == Right[i - 1]) {
2193 // Commuting would preserve SplatRight, but we don't want to break
2194 // SplatLeft either, i.e. preserve the original order if possible.
2195 // (FIXME: why do we care?)
2196 if (SplatLeft && VLeft == Left[i - 1])
2197 return false;
2198 return true;
2199 }
2200 }
2201 // Symmetrically handle Right side.
2202 if (SplatLeft) {
2203 if (VLeft == Left[i - 1])
2204 // Preserve SplatLeft
2205 return false;
2206 if (VRight == Left[i - 1])
2207 return true;
2208 }
2209
2210 Instruction *ILeft = dyn_cast<Instruction>(VLeft);
2211 Instruction *IRight = dyn_cast<Instruction>(VRight);
2212
2213 // If we have "AllSameOpcodeRight", try to see if the left operands preserves
2214 // it and not the right, in this case we want to commute.
2215 if (AllSameOpcodeRight) {
2216 unsigned RightPrevOpcode = cast<Instruction>(Right[i - 1])->getOpcode();
2217 if (IRight && RightPrevOpcode == IRight->getOpcode())
2218 // Do not commute, a match on the right preserves AllSameOpcodeRight
2219 return false;
2220 if (ILeft && RightPrevOpcode == ILeft->getOpcode()) {
2221 // We have a match and may want to commute, but first check if there is
2222 // not also a match on the existing operands on the Left to preserve
2223 // AllSameOpcodeLeft, i.e. preserve the original order if possible.
2224 // (FIXME: why do we care?)
2225 if (AllSameOpcodeLeft && ILeft &&
2226 cast<Instruction>(Left[i - 1])->getOpcode() == ILeft->getOpcode())
2227 return false;
2228 return true;
2229 }
2230 }
2231 // Symmetrically handle Left side.
2232 if (AllSameOpcodeLeft) {
2233 unsigned LeftPrevOpcode = cast<Instruction>(Left[i - 1])->getOpcode();
2234 if (ILeft && LeftPrevOpcode == ILeft->getOpcode())
2235 return false;
2236 if (IRight && LeftPrevOpcode == IRight->getOpcode())
2237 return true;
2238 }
2239 return false;
2240}
2241
2242void BoUpSLP::reorderInputsAccordingToOpcode(ArrayRef<Value *> VL,
2243 SmallVectorImpl<Value *> &Left,
2244 SmallVectorImpl<Value *> &Right) {
2245
2246 if (VL.size()) {
2247 // Peel the first iteration out of the loop since there's nothing
2248 // interesting to do anyway and it simplifies the checks in the loop.
2249 auto VLeft = cast<Instruction>(VL[0])->getOperand(0);
2250 auto VRight = cast<Instruction>(VL[0])->getOperand(1);
2251 if (!isa<Instruction>(VRight) && isa<Instruction>(VLeft))
2252 // Favor having instruction to the right. FIXME: why?
2253 std::swap(VLeft, VRight);
2254 Left.push_back(VLeft);
2255 Right.push_back(VRight);
2256 }
2257
2258 // Keep track if we have instructions with all the same opcode on one side.
2259 bool AllSameOpcodeLeft = isa<Instruction>(Left[0]);
2260 bool AllSameOpcodeRight = isa<Instruction>(Right[0]);
2261 // Keep track if we have one side with all the same value (broadcast).
2262 bool SplatLeft = true;
2263 bool SplatRight = true;
2264
2265 for (unsigned i = 1, e = VL.size(); i != e; ++i) {
2266 Instruction *I = cast<Instruction>(VL[i]);
2267 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2267, __PRETTY_FUNCTION__))
;
2268 // Commute to favor either a splat or maximizing having the same opcodes on
2269 // one side.
2270 if (shouldReorderOperands(i, *I, Left, Right, AllSameOpcodeLeft,
2271 AllSameOpcodeRight, SplatLeft, SplatRight)) {
2272 Left.push_back(I->getOperand(1));
2273 Right.push_back(I->getOperand(0));
2274 } else {
2275 Left.push_back(I->getOperand(0));
2276 Right.push_back(I->getOperand(1));
2277 }
2278 // Update Splat* and AllSameOpcode* after the insertion.
2279 SplatRight = SplatRight && (Right[i - 1] == Right[i]);
2280 SplatLeft = SplatLeft && (Left[i - 1] == Left[i]);
2281 AllSameOpcodeLeft = AllSameOpcodeLeft && isa<Instruction>(Left[i]) &&
2282 (cast<Instruction>(Left[i - 1])->getOpcode() ==
2283 cast<Instruction>(Left[i])->getOpcode());
2284 AllSameOpcodeRight = AllSameOpcodeRight && isa<Instruction>(Right[i]) &&
2285 (cast<Instruction>(Right[i - 1])->getOpcode() ==
2286 cast<Instruction>(Right[i])->getOpcode());
2287 }
2288
2289 // If one operand end up being broadcast, return this operand order.
2290 if (SplatRight || SplatLeft)
2291 return;
2292
2293 // Finally check if we can get longer vectorizable chain by reordering
2294 // without breaking the good operand order detected above.
2295 // E.g. If we have something like-
2296 // load a[0] load b[0]
2297 // load b[1] load a[1]
2298 // load a[2] load b[2]
2299 // load a[3] load b[3]
2300 // Reordering the second load b[1] load a[1] would allow us to vectorize
2301 // this code and we still retain AllSameOpcode property.
2302 // FIXME: This load reordering might break AllSameOpcode in some rare cases
2303 // such as-
2304 // add a[0],c[0] load b[0]
2305 // add a[1],c[2] load b[1]
2306 // b[2] load b[2]
2307 // add a[3],c[3] load b[3]
2308 for (unsigned j = 0; j < VL.size() - 1; ++j) {
2309 if (LoadInst *L = dyn_cast<LoadInst>(Left[j])) {
2310 if (LoadInst *L1 = dyn_cast<LoadInst>(Right[j + 1])) {
2311 if (isConsecutiveAccess(L, L1, *DL, *SE)) {
2312 std::swap(Left[j + 1], Right[j + 1]);
2313 continue;
2314 }
2315 }
2316 }
2317 if (LoadInst *L = dyn_cast<LoadInst>(Right[j])) {
2318 if (LoadInst *L1 = dyn_cast<LoadInst>(Left[j + 1])) {
2319 if (isConsecutiveAccess(L, L1, *DL, *SE)) {
2320 std::swap(Left[j + 1], Right[j + 1]);
2321 continue;
2322 }
2323 }
2324 }
2325 // else unchanged
2326 }
2327}
2328
2329void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
2330
2331 // Get the basic block this bundle is in. All instructions in the bundle
2332 // should be in this block.
2333 auto *Front = cast<Instruction>(VL.front());
2334 auto *BB = Front->getParent();
2335 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2337, __PRETTY_FUNCTION__))
2336 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2337, __PRETTY_FUNCTION__))
2337 }))((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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2337, __PRETTY_FUNCTION__))
;
2338
2339 // The last instruction in the bundle in program order.
2340 Instruction *LastInst = nullptr;
8
'LastInst' initialized to a null pointer value
2341
2342 // Find the last instruction. The common case should be that BB has been
2343 // scheduled, and the last instruction is VL.back(). So we start with
2344 // VL.back() and iterate over schedule data until we reach the end of the
2345 // bundle. The end of the bundle is marked by null ScheduleData.
2346 if (BlocksSchedules.count(BB)) {
9
Assuming the condition is false
10
Taking false branch
2347 auto *Bundle = BlocksSchedules[BB]->getScheduleData(VL.back());
2348 if (Bundle && Bundle->isPartOfBundle())
2349 for (; Bundle; Bundle = Bundle->NextInBundle)
2350 LastInst = Bundle->Inst;
2351 }
2352
2353 // LastInst can still be null at this point if there's either not an entry
2354 // for BB in BlocksSchedules or there's no ScheduleData available for
2355 // VL.back(). This can be the case if buildTree_rec aborts for various
2356 // reasons (e.g., the maximum recursion depth is reached, the maximum region
2357 // size is reached, etc.). ScheduleData is initialized in the scheduling
2358 // "dry-run".
2359 //
2360 // If this happens, we can still find the last instruction by brute force. We
2361 // iterate forwards from Front (inclusive) until we either see all
2362 // instructions in the bundle or reach the end of the block. If Front is the
2363 // last instruction in program order, LastInst will be set to Front, and we
2364 // will visit all the remaining instructions in the block.
2365 //
2366 // One of the reasons we exit early from buildTree_rec is to place an upper
2367 // bound on compile-time. Thus, taking an additional compile-time hit here is
2368 // not ideal. However, this should be exceedingly rare since it requires that
2369 // we both exit early from buildTree_rec and that the bundle be out-of-order
2370 // (causing us to iterate all the way to the end of the block).
2371 if (!LastInst) {
11
Taking true branch
2372 SmallPtrSet<Value *, 16> Bundle(VL.begin(), VL.end());
2373 for (auto &I : make_range(BasicBlock::iterator(Front), BB->end())) {
2374 if (Bundle.erase(&I))
2375 LastInst = &I;
2376 if (Bundle.empty())
2377 break;
2378 }
2379 }
2380
2381 // Set the insertion point after the last instruction in the bundle. Set the
2382 // debug location to Front.
2383 Builder.SetInsertPoint(BB, ++LastInst->getIterator());
12
Called C++ object pointer is null
2384 Builder.SetCurrentDebugLocation(Front->getDebugLoc());
2385}
2386
2387Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
2388 Value *Vec = UndefValue::get(Ty);
2389 // Generate the 'InsertElement' instruction.
2390 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
2391 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
2392 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
2393 GatherSeq.insert(Insrt);
2394 CSEBlocks.insert(Insrt->getParent());
2395
2396 // Add to our 'need-to-extract' list.
2397 if (ScalarToTreeEntry.count(VL[i])) {
2398 int Idx = ScalarToTreeEntry[VL[i]];
2399 TreeEntry *E = &VectorizableTree[Idx];
2400 // Find which lane we need to extract.
2401 int FoundLane = -1;
2402 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
2403 // Is this the lane of the scalar that we are looking for ?
2404 if (E->Scalars[Lane] == VL[i]) {
2405 FoundLane = Lane;
2406 break;
2407 }
2408 }
2409 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2409, __PRETTY_FUNCTION__))
;
2410 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
2411 }
2412 }
2413 }
2414
2415 return Vec;
2416}
2417
2418Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
2419 SmallDenseMap<Value*, int>::const_iterator Entry
2420 = ScalarToTreeEntry.find(VL[0]);
2421 if (Entry != ScalarToTreeEntry.end()) {
2422 int Idx = Entry->second;
2423 const TreeEntry *En = &VectorizableTree[Idx];
2424 if (En->isSame(VL) && En->VectorizedValue)
2425 return En->VectorizedValue;
2426 }
2427 return nullptr;
2428}
2429
2430Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
2431 if (ScalarToTreeEntry.count(VL[0])) {
2432 int Idx = ScalarToTreeEntry[VL[0]];
2433 TreeEntry *E = &VectorizableTree[Idx];
2434 if (E->isSame(VL))
2435 return vectorizeTree(E);
2436 }
2437
2438 Type *ScalarTy = VL[0]->getType();
2439 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
2440 ScalarTy = SI->getValueOperand()->getType();
2441 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
2442
2443 return Gather(VL, VecTy);
2444}
2445
2446Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
2447 IRBuilder<>::InsertPointGuard Guard(Builder);
2448
2449 if (E->VectorizedValue) {
1
Assuming the condition is false
2
Taking false branch
2450 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)
;
2451 return E->VectorizedValue;
2452 }
2453
2454 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
2455 Type *ScalarTy = VL0->getType();
2456 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
3
Taking false branch
2457 ScalarTy = SI->getValueOperand()->getType();
2458 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
2459
2460 if (E->NeedToGather) {
4
Assuming the condition is false
5
Taking false branch
2461 setInsertPointAfterBundle(E->Scalars);
2462 auto *V = Gather(E->Scalars, VecTy);
2463 E->VectorizedValue = V;
2464 return V;
2465 }
2466
2467 unsigned Opcode = getSameOpcode(E->Scalars);
2468
2469 switch (Opcode) {
6
Control jumps to 'case ShuffleVector:' at line 2795
2470 case Instruction::PHI: {
2471 PHINode *PH = dyn_cast<PHINode>(VL0);
2472 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI());
2473 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
2474 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
2475 E->VectorizedValue = NewPhi;
2476
2477 // PHINodes may have multiple entries from the same block. We want to
2478 // visit every block once.
2479 SmallSet<BasicBlock*, 4> VisitedBBs;
2480
2481 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
2482 ValueList Operands;
2483 BasicBlock *IBB = PH->getIncomingBlock(i);
2484
2485 if (!VisitedBBs.insert(IBB).second) {
2486 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
2487 continue;
2488 }
2489
2490 // Prepare the operand vector.
2491 for (Value *V : E->Scalars)
2492 Operands.push_back(cast<PHINode>(V)->getIncomingValueForBlock(IBB));
2493
2494 Builder.SetInsertPoint(IBB->getTerminator());
2495 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
2496 Value *Vec = vectorizeTree(Operands);
2497 NewPhi->addIncoming(Vec, IBB);
2498 }
2499
2500 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2501, __PRETTY_FUNCTION__))
2501 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2501, __PRETTY_FUNCTION__))
;
2502 return NewPhi;
2503 }
2504
2505 case Instruction::ExtractElement: {
2506 if (canReuseExtract(E->Scalars, Instruction::ExtractElement)) {
2507 Value *V = VL0->getOperand(0);
2508 E->VectorizedValue = V;
2509 return V;
2510 }
2511 setInsertPointAfterBundle(E->Scalars);
2512 auto *V = Gather(E->Scalars, VecTy);
2513 E->VectorizedValue = V;
2514 return V;
2515 }
2516 case Instruction::ExtractValue: {
2517 if (canReuseExtract(E->Scalars, Instruction::ExtractValue)) {
2518 LoadInst *LI = cast<LoadInst>(VL0->getOperand(0));
2519 Builder.SetInsertPoint(LI);
2520 PointerType *PtrTy = PointerType::get(VecTy, LI->getPointerAddressSpace());
2521 Value *Ptr = Builder.CreateBitCast(LI->getOperand(0), PtrTy);
2522 LoadInst *V = Builder.CreateAlignedLoad(Ptr, LI->getAlignment());
2523 E->VectorizedValue = V;
2524 return propagateMetadata(V, E->Scalars);
2525 }
2526 setInsertPointAfterBundle(E->Scalars);
2527 auto *V = Gather(E->Scalars, VecTy);
2528 E->VectorizedValue = V;
2529 return V;
2530 }
2531 case Instruction::ZExt:
2532 case Instruction::SExt:
2533 case Instruction::FPToUI:
2534 case Instruction::FPToSI:
2535 case Instruction::FPExt:
2536 case Instruction::PtrToInt:
2537 case Instruction::IntToPtr:
2538 case Instruction::SIToFP:
2539 case Instruction::UIToFP:
2540 case Instruction::Trunc:
2541 case Instruction::FPTrunc:
2542 case Instruction::BitCast: {
2543 ValueList INVL;
2544 for (Value *V : E->Scalars)
2545 INVL.push_back(cast<Instruction>(V)->getOperand(0));
2546
2547 setInsertPointAfterBundle(E->Scalars);
2548
2549 Value *InVec = vectorizeTree(INVL);
2550
2551 if (Value *V = alreadyVectorized(E->Scalars))
2552 return V;
2553
2554 CastInst *CI = dyn_cast<CastInst>(VL0);
2555 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
2556 E->VectorizedValue = V;
2557 ++NumVectorInstructions;
2558 return V;
2559 }
2560 case Instruction::FCmp:
2561 case Instruction::ICmp: {
2562 ValueList LHSV, RHSV;
2563 for (Value *V : E->Scalars) {
2564 LHSV.push_back(cast<Instruction>(V)->getOperand(0));
2565 RHSV.push_back(cast<Instruction>(V)->getOperand(1));
2566 }
2567
2568 setInsertPointAfterBundle(E->Scalars);
2569
2570 Value *L = vectorizeTree(LHSV);
2571 Value *R = vectorizeTree(RHSV);
2572
2573 if (Value *V = alreadyVectorized(E->Scalars))
2574 return V;
2575
2576 CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate();
2577 Value *V;
2578 if (Opcode == Instruction::FCmp)
2579 V = Builder.CreateFCmp(P0, L, R);
2580 else
2581 V = Builder.CreateICmp(P0, L, R);
2582
2583 E->VectorizedValue = V;
2584 propagateIRFlags(E->VectorizedValue, E->Scalars);
2585 ++NumVectorInstructions;
2586 return V;
2587 }
2588 case Instruction::Select: {
2589 ValueList TrueVec, FalseVec, CondVec;
2590 for (Value *V : E->Scalars) {
2591 CondVec.push_back(cast<Instruction>(V)->getOperand(0));
2592 TrueVec.push_back(cast<Instruction>(V)->getOperand(1));
2593 FalseVec.push_back(cast<Instruction>(V)->getOperand(2));
2594 }
2595
2596 setInsertPointAfterBundle(E->Scalars);
2597
2598 Value *Cond = vectorizeTree(CondVec);
2599 Value *True = vectorizeTree(TrueVec);
2600 Value *False = vectorizeTree(FalseVec);
2601
2602 if (Value *V = alreadyVectorized(E->Scalars))
2603 return V;
2604
2605 Value *V = Builder.CreateSelect(Cond, True, False);
2606 E->VectorizedValue = V;
2607 ++NumVectorInstructions;
2608 return V;
2609 }
2610 case Instruction::Add:
2611 case Instruction::FAdd:
2612 case Instruction::Sub:
2613 case Instruction::FSub:
2614 case Instruction::Mul:
2615 case Instruction::FMul:
2616 case Instruction::UDiv:
2617 case Instruction::SDiv:
2618 case Instruction::FDiv:
2619 case Instruction::URem:
2620 case Instruction::SRem:
2621 case Instruction::FRem:
2622 case Instruction::Shl:
2623 case Instruction::LShr:
2624 case Instruction::AShr:
2625 case Instruction::And:
2626 case Instruction::Or:
2627 case Instruction::Xor: {
2628 ValueList LHSVL, RHSVL;
2629 if (isa<BinaryOperator>(VL0) && VL0->isCommutative())
2630 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL);
2631 else
2632 for (Value *V : E->Scalars) {
2633 LHSVL.push_back(cast<Instruction>(V)->getOperand(0));
2634 RHSVL.push_back(cast<Instruction>(V)->getOperand(1));
2635 }
2636
2637 setInsertPointAfterBundle(E->Scalars);
2638
2639 Value *LHS = vectorizeTree(LHSVL);
2640 Value *RHS = vectorizeTree(RHSVL);
2641
2642 if (Value *V = alreadyVectorized(E->Scalars))
2643 return V;
2644
2645 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
2646 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
2647 E->VectorizedValue = V;
2648 propagateIRFlags(E->VectorizedValue, E->Scalars);
2649 ++NumVectorInstructions;
2650
2651 if (Instruction *I = dyn_cast<Instruction>(V))
2652 return propagateMetadata(I, E->Scalars);
2653
2654 return V;
2655 }
2656 case Instruction::Load: {
2657 // Loads are inserted at the head of the tree because we don't want to
2658 // sink them all the way down past store instructions.
2659 setInsertPointAfterBundle(E->Scalars);
2660
2661 LoadInst *LI = cast<LoadInst>(VL0);
2662 Type *ScalarLoadTy = LI->getType();
2663 unsigned AS = LI->getPointerAddressSpace();
2664
2665 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
2666 VecTy->getPointerTo(AS));
2667
2668 // The pointer operand uses an in-tree scalar so we add the new BitCast to
2669 // ExternalUses list to make sure that an extract will be generated in the
2670 // future.
2671 if (ScalarToTreeEntry.count(LI->getPointerOperand()))
2672 ExternalUses.push_back(
2673 ExternalUser(LI->getPointerOperand(), cast<User>(VecPtr), 0));
2674
2675 unsigned Alignment = LI->getAlignment();
2676 LI = Builder.CreateLoad(VecPtr);
2677 if (!Alignment) {
2678 Alignment = DL->getABITypeAlignment(ScalarLoadTy);
2679 }
2680 LI->setAlignment(Alignment);
2681 E->VectorizedValue = LI;
2682 ++NumVectorInstructions;
2683 return propagateMetadata(LI, E->Scalars);
2684 }
2685 case Instruction::Store: {
2686 StoreInst *SI = cast<StoreInst>(VL0);
2687 unsigned Alignment = SI->getAlignment();
2688 unsigned AS = SI->getPointerAddressSpace();
2689
2690 ValueList ValueOp;
2691 for (Value *V : E->Scalars)
2692 ValueOp.push_back(cast<StoreInst>(V)->getValueOperand());
2693
2694 setInsertPointAfterBundle(E->Scalars);
2695
2696 Value *VecValue = vectorizeTree(ValueOp);
2697 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
2698 VecTy->getPointerTo(AS));
2699 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
2700
2701 // The pointer operand uses an in-tree scalar so we add the new BitCast to
2702 // ExternalUses list to make sure that an extract will be generated in the
2703 // future.
2704 if (ScalarToTreeEntry.count(SI->getPointerOperand()))
2705 ExternalUses.push_back(
2706 ExternalUser(SI->getPointerOperand(), cast<User>(VecPtr), 0));
2707
2708 if (!Alignment) {
2709 Alignment = DL->getABITypeAlignment(SI->getValueOperand()->getType());
2710 }
2711 S->setAlignment(Alignment);
2712 E->VectorizedValue = S;
2713 ++NumVectorInstructions;
2714 return propagateMetadata(S, E->Scalars);
2715 }
2716 case Instruction::GetElementPtr: {
2717 setInsertPointAfterBundle(E->Scalars);
2718
2719 ValueList Op0VL;
2720 for (Value *V : E->Scalars)
2721 Op0VL.push_back(cast<GetElementPtrInst>(V)->getOperand(0));
2722
2723 Value *Op0 = vectorizeTree(Op0VL);
2724
2725 std::vector<Value *> OpVecs;
2726 for (int j = 1, e = cast<GetElementPtrInst>(VL0)->getNumOperands(); j < e;
2727 ++j) {
2728 ValueList OpVL;
2729 for (Value *V : E->Scalars)
2730 OpVL.push_back(cast<GetElementPtrInst>(V)->getOperand(j));
2731
2732 Value *OpVec = vectorizeTree(OpVL);
2733 OpVecs.push_back(OpVec);
2734 }
2735
2736 Value *V = Builder.CreateGEP(
2737 cast<GetElementPtrInst>(VL0)->getSourceElementType(), Op0, OpVecs);
2738 E->VectorizedValue = V;
2739 ++NumVectorInstructions;
2740
2741 if (Instruction *I = dyn_cast<Instruction>(V))
2742 return propagateMetadata(I, E->Scalars);
2743
2744 return V;
2745 }
2746 case Instruction::Call: {
2747 CallInst *CI = cast<CallInst>(VL0);
2748 setInsertPointAfterBundle(E->Scalars);
2749 Function *FI;
2750 Intrinsic::ID IID = Intrinsic::not_intrinsic;
2751 Value *ScalarArg = nullptr;
2752 if (CI && (FI = CI->getCalledFunction())) {
2753 IID = FI->getIntrinsicID();
2754 }
2755 std::vector<Value *> OpVecs;
2756 for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) {
2757 ValueList OpVL;
2758 // ctlz,cttz and powi are special intrinsics whose second argument is
2759 // a scalar. This argument should not be vectorized.
2760 if (hasVectorInstrinsicScalarOpd(IID, 1) && j == 1) {
2761 CallInst *CEI = cast<CallInst>(E->Scalars[0]);
2762 ScalarArg = CEI->getArgOperand(j);
2763 OpVecs.push_back(CEI->getArgOperand(j));
2764 continue;
2765 }
2766 for (Value *V : E->Scalars) {
2767 CallInst *CEI = cast<CallInst>(V);
2768 OpVL.push_back(CEI->getArgOperand(j));
2769 }
2770
2771 Value *OpVec = vectorizeTree(OpVL);
2772 DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: OpVec[" << j << "]: "
<< *OpVec << "\n"; } } while (false)
;
2773 OpVecs.push_back(OpVec);
2774 }
2775
2776 Module *M = F->getParent();
2777 Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
2778 Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) };
2779 Function *CF = Intrinsic::getDeclaration(M, ID, Tys);
2780 SmallVector<OperandBundleDef, 1> OpBundles;
2781 CI->getOperandBundlesAsDefs(OpBundles);
2782 Value *V = Builder.CreateCall(CF, OpVecs, OpBundles);
2783
2784 // The scalar argument uses an in-tree scalar so we add the new vectorized
2785 // call to ExternalUses list to make sure that an extract will be
2786 // generated in the future.
2787 if (ScalarArg && ScalarToTreeEntry.count(ScalarArg))
2788 ExternalUses.push_back(ExternalUser(ScalarArg, cast<User>(V), 0));
2789
2790 E->VectorizedValue = V;
2791 propagateIRFlags(E->VectorizedValue, E->Scalars);
2792 ++NumVectorInstructions;
2793 return V;
2794 }
2795 case Instruction::ShuffleVector: {
2796 ValueList LHSVL, RHSVL;
2797 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2797, __PRETTY_FUNCTION__))
;
2798 reorderAltShuffleOperands(E->Scalars, LHSVL, RHSVL);
2799 setInsertPointAfterBundle(E->Scalars);
7
Calling 'BoUpSLP::setInsertPointAfterBundle'
2800
2801 Value *LHS = vectorizeTree(LHSVL);
2802 Value *RHS = vectorizeTree(RHSVL);
2803
2804 if (Value *V = alreadyVectorized(E->Scalars))
2805 return V;
2806
2807 // Create a vector of LHS op1 RHS
2808 BinaryOperator *BinOp0 = cast<BinaryOperator>(VL0);
2809 Value *V0 = Builder.CreateBinOp(BinOp0->getOpcode(), LHS, RHS);
2810
2811 // Create a vector of LHS op2 RHS
2812 Instruction *VL1 = cast<Instruction>(E->Scalars[1]);
2813 BinaryOperator *BinOp1 = cast<BinaryOperator>(VL1);
2814 Value *V1 = Builder.CreateBinOp(BinOp1->getOpcode(), LHS, RHS);
2815
2816 // Create shuffle to take alternate operations from the vector.
2817 // Also, gather up odd and even scalar ops to propagate IR flags to
2818 // each vector operation.
2819 ValueList OddScalars, EvenScalars;
2820 unsigned e = E->Scalars.size();
2821 SmallVector<Constant *, 8> Mask(e);
2822 for (unsigned i = 0; i < e; ++i) {
2823 if (i & 1) {
2824 Mask[i] = Builder.getInt32(e + i);
2825 OddScalars.push_back(E->Scalars[i]);
2826 } else {
2827 Mask[i] = Builder.getInt32(i);
2828 EvenScalars.push_back(E->Scalars[i]);
2829 }
2830 }
2831
2832 Value *ShuffleMask = ConstantVector::get(Mask);
2833 propagateIRFlags(V0, EvenScalars);
2834 propagateIRFlags(V1, OddScalars);
2835
2836 Value *V = Builder.CreateShuffleVector(V0, V1, ShuffleMask);
2837 E->VectorizedValue = V;
2838 ++NumVectorInstructions;
2839 if (Instruction *I = dyn_cast<Instruction>(V))
2840 return propagateMetadata(I, E->Scalars);
2841
2842 return V;
2843 }
2844 default:
2845 llvm_unreachable("unknown inst")::llvm::llvm_unreachable_internal("unknown inst", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2845)
;
2846 }
2847 return nullptr;
2848}
2849
2850Value *BoUpSLP::vectorizeTree() {
2851 ExtraValueToDebugLocsMap ExternallyUsedValues;
2852 return vectorizeTree(ExternallyUsedValues);
2853}
2854
2855Value *
2856BoUpSLP::vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues) {
2857
2858 // All blocks must be scheduled before any instructions are inserted.
2859 for (auto &BSIter : BlocksSchedules) {
2860 scheduleBlock(BSIter.second.get());
2861 }
2862
2863 Builder.SetInsertPoint(&F->getEntryBlock().front());
2864 auto *VectorRoot = vectorizeTree(&VectorizableTree[0]);
2865
2866 // If the vectorized tree can be rewritten in a smaller type, we truncate the
2867 // vectorized root. InstCombine will then rewrite the entire expression. We
2868 // sign extend the extracted values below.
2869 auto *ScalarRoot = VectorizableTree[0].Scalars[0];
2870 if (MinBWs.count(ScalarRoot)) {
2871 if (auto *I = dyn_cast<Instruction>(VectorRoot))
2872 Builder.SetInsertPoint(&*++BasicBlock::iterator(I));
2873 auto BundleWidth = VectorizableTree[0].Scalars.size();
2874 auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first);
2875 auto *VecTy = VectorType::get(MinTy, BundleWidth);
2876 auto *Trunc = Builder.CreateTrunc(VectorRoot, VecTy);
2877 VectorizableTree[0].VectorizedValue = Trunc;
2878 }
2879
2880 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses
.size() << " values .\n"; } } while (false)
;
2881
2882 // If necessary, sign-extend or zero-extend ScalarRoot to the larger type
2883 // specified by ScalarType.
2884 auto extend = [&](Value *ScalarRoot, Value *Ex, Type *ScalarType) {
2885 if (!MinBWs.count(ScalarRoot))
2886 return Ex;
2887 if (MinBWs[ScalarRoot].second)
2888 return Builder.CreateSExt(Ex, ScalarType);
2889 return Builder.CreateZExt(Ex, ScalarType);
2890 };
2891
2892 // Extract all of the elements with the external uses.
2893 for (const auto &ExternalUse : ExternalUses) {
2894 Value *Scalar = ExternalUse.Scalar;
2895 llvm::User *User = ExternalUse.User;
2896
2897 // Skip users that we already RAUW. This happens when one instruction
2898 // has multiple uses of the same value.
2899 if (User && !is_contained(Scalar->users(), User))
2900 continue;
2901 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2901, __PRETTY_FUNCTION__))
;
2902
2903 int Idx = ScalarToTreeEntry[Scalar];
2904 TreeEntry *E = &VectorizableTree[Idx];
2905 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2905, __PRETTY_FUNCTION__))
;
2906
2907 Value *Vec = E->VectorizedValue;
2908 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2908, __PRETTY_FUNCTION__))
;
2909
2910 Value *Lane = Builder.getInt32(ExternalUse.Lane);
2911 // If User == nullptr, the Scalar is used as extra arg. Generate
2912 // ExtractElement instruction and update the record for this scalar in
2913 // ExternallyUsedValues.
2914 if (!User) {
2915 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2917, __PRETTY_FUNCTION__))
2916 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2917, __PRETTY_FUNCTION__))
2917 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2917, __PRETTY_FUNCTION__))
;
2918 if (auto *VecI = dyn_cast<Instruction>(Vec)) {
2919 Builder.SetInsertPoint(VecI->getParent(),
2920 std::next(VecI->getIterator()));
2921 } else {
2922 Builder.SetInsertPoint(&F->getEntryBlock().front());
2923 }
2924 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2925 Ex = extend(ScalarRoot, Ex, Scalar->getType());
2926 CSEBlocks.insert(cast<Instruction>(Scalar)->getParent());
2927 auto &Locs = ExternallyUsedValues[Scalar];
2928 ExternallyUsedValues.insert({Ex, Locs});
2929 ExternallyUsedValues.erase(Scalar);
2930 continue;
2931 }
2932
2933 // Generate extracts for out-of-tree users.
2934 // Find the insertion point for the extractelement lane.
2935 if (auto *VecI = dyn_cast<Instruction>(Vec)) {
2936 if (PHINode *PH = dyn_cast<PHINode>(User)) {
2937 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) {
2938 if (PH->getIncomingValue(i) == Scalar) {
2939 TerminatorInst *IncomingTerminator =
2940 PH->getIncomingBlock(i)->getTerminator();
2941 if (isa<CatchSwitchInst>(IncomingTerminator)) {
2942 Builder.SetInsertPoint(VecI->getParent(),
2943 std::next(VecI->getIterator()));
2944 } else {
2945 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator());
2946 }
2947 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2948 Ex = extend(ScalarRoot, Ex, Scalar->getType());
2949 CSEBlocks.insert(PH->getIncomingBlock(i));
2950 PH->setOperand(i, Ex);
2951 }
2952 }
2953 } else {
2954 Builder.SetInsertPoint(cast<Instruction>(User));
2955 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2956 Ex = extend(ScalarRoot, Ex, Scalar->getType());
2957 CSEBlocks.insert(cast<Instruction>(User)->getParent());
2958 User->replaceUsesOfWith(Scalar, Ex);
2959 }
2960 } else {
2961 Builder.SetInsertPoint(&F->getEntryBlock().front());
2962 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2963 Ex = extend(ScalarRoot, Ex, Scalar->getType());
2964 CSEBlocks.insert(&F->getEntryBlock());
2965 User->replaceUsesOfWith(Scalar, Ex);
2966 }
2967
2968 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Replaced:" << *User <<
".\n"; } } while (false)
;
2969 }
2970
2971 // For each vectorized value:
2972 for (TreeEntry &EIdx : VectorizableTree) {
2973 TreeEntry *Entry = &EIdx;
2974
2975 // For each lane:
2976 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
2977 Value *Scalar = Entry->Scalars[Lane];
2978 // No need to handle users of gathered values.
2979 if (Entry->NeedToGather)
2980 continue;
2981
2982 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2982, __PRETTY_FUNCTION__))
;
2983
2984 Type *Ty = Scalar->getType();
2985 if (!Ty->isVoidTy()) {
2986#ifndef NDEBUG
2987 for (User *U : Scalar->users()) {
2988 DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: \tvalidating user:" <<
*U << ".\n"; } } while (false)
;
2989
2990 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2993, __PRETTY_FUNCTION__))
2991 // 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2993, __PRETTY_FUNCTION__))
2992 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2993, __PRETTY_FUNCTION__))
2993 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 2993, __PRETTY_FUNCTION__))
;
2994 }
2995#endif
2996 Value *Undef = UndefValue::get(Ty);
2997 Scalar->replaceAllUsesWith(Undef);
2998 }
2999 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: \tErasing scalar:" << *
Scalar << ".\n"; } } while (false)
;
3000 eraseInstruction(cast<Instruction>(Scalar));
3001 }
3002 }
3003
3004 Builder.ClearInsertionPoint();
3005
3006 return VectorizableTree[0].VectorizedValue;
3007}
3008
3009void BoUpSLP::optimizeGatherSequence() {
3010 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq
.size() << " gather sequences instructions.\n"; } } while
(false)
3011 << " gather sequences instructions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq
.size() << " gather sequences instructions.\n"; } } while
(false)
;
3012 // LICM InsertElementInst sequences.
3013 for (Instruction *it : GatherSeq) {
3014 InsertElementInst *Insert = dyn_cast<InsertElementInst>(it);
3015
3016 if (!Insert)
3017 continue;
3018
3019 // Check if this block is inside a loop.
3020 Loop *L = LI->getLoopFor(Insert->getParent());
3021 if (!L)
3022 continue;
3023
3024 // Check if it has a preheader.
3025 BasicBlock *PreHeader = L->getLoopPreheader();
3026 if (!PreHeader)
3027 continue;
3028
3029 // If the vector or the element that we insert into it are
3030 // instructions that are defined in this basic block then we can't
3031 // hoist this instruction.
3032 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
3033 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
3034 if (CurrVec && L->contains(CurrVec))
3035 continue;
3036 if (NewElem && L->contains(NewElem))
3037 continue;
3038
3039 // We can hoist this instruction. Move it to the pre-header.
3040 Insert->moveBefore(PreHeader->getTerminator());
3041 }
3042
3043 // Make a list of all reachable blocks in our CSE queue.
3044 SmallVector<const DomTreeNode *, 8> CSEWorkList;
3045 CSEWorkList.reserve(CSEBlocks.size());
3046 for (BasicBlock *BB : CSEBlocks)
3047 if (DomTreeNode *N = DT->getNode(BB)) {
3048 assert(DT->isReachableFromEntry(N))((DT->isReachableFromEntry(N)) ? static_cast<void> (
0) : __assert_fail ("DT->isReachableFromEntry(N)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3048, __PRETTY_FUNCTION__))
;
3049 CSEWorkList.push_back(N);
3050 }
3051
3052 // Sort blocks by domination. This ensures we visit a block after all blocks
3053 // dominating it are visited.
3054 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(),
3055 [this](const DomTreeNode *A, const DomTreeNode *B) {
3056 return DT->properlyDominates(A, B);
3057 });
3058
3059 // Perform O(N^2) search over the gather sequences and merge identical
3060 // instructions. TODO: We can further optimize this scan if we split the
3061 // instructions into different buckets based on the insert lane.
3062 SmallVector<Instruction *, 16> Visited;
3063 for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) {
3064 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3065, __PRETTY_FUNCTION__))
3065 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3065, __PRETTY_FUNCTION__))
;
3066 BasicBlock *BB = (*I)->getBlock();
3067 // For all instructions in blocks containing gather sequences:
3068 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) {
3069 Instruction *In = &*it++;
3070 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In))
3071 continue;
3072
3073 // Check if we can replace this instruction with any of the
3074 // visited instructions.
3075 for (Instruction *v : Visited) {
3076 if (In->isIdenticalTo(v) &&
3077 DT->dominates(v->getParent(), In->getParent())) {
3078 In->replaceAllUsesWith(v);
3079 eraseInstruction(In);
3080 In = nullptr;
3081 break;
3082 }
3083 }
3084 if (In) {
3085 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3085, __PRETTY_FUNCTION__))
;
3086 Visited.push_back(In);
3087 }
3088 }
3089 }
3090 CSEBlocks.clear();
3091 GatherSeq.clear();
3092}
3093
3094// Groups the instructions to a bundle (which is then a single scheduling entity)
3095// and schedules instructions until the bundle gets ready.
3096bool BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef<Value *> VL,
3097 BoUpSLP *SLP) {
3098 if (isa<PHINode>(VL[0]))
3099 return true;
3100
3101 // Initialize the instruction bundle.
3102 Instruction *OldScheduleEnd = ScheduleEnd;
3103 ScheduleData *PrevInBundle = nullptr;
3104 ScheduleData *Bundle = nullptr;
3105 bool ReSchedule = false;
3106 DEBUG(dbgs() << "SLP: bundle: " << *VL[0] << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: bundle: " << *VL[0] <<
"\n"; } } while (false)
;
3107
3108 // Make sure that the scheduling region contains all
3109 // instructions of the bundle.
3110 for (Value *V : VL) {
3111 if (!extendSchedulingRegion(V))
3112 return false;
3113 }
3114
3115 for (Value *V : VL) {
3116 ScheduleData *BundleMember = getScheduleData(V);
3117 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3118, __PRETTY_FUNCTION__))
3118 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3118, __PRETTY_FUNCTION__))
;
3119 if (BundleMember->IsScheduled) {
3120 // A bundle member was scheduled as single instruction before and now
3121 // needs to be scheduled as part of the bundle. We just get rid of the
3122 // existing schedule.
3123 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)
3124 << " was already scheduled\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: reset schedule because " <<
*BundleMember << " was already scheduled\n"; } } while
(false)
;
3125 ReSchedule = true;
3126 }
3127 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3128, __PRETTY_FUNCTION__))
3128 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3128, __PRETTY_FUNCTION__))
;
3129 if (PrevInBundle) {
3130 PrevInBundle->NextInBundle = BundleMember;
3131 } else {
3132 Bundle = BundleMember;
3133 }
3134 BundleMember->UnscheduledDepsInBundle = 0;
3135 Bundle->UnscheduledDepsInBundle += BundleMember->UnscheduledDeps;
3136
3137 // Group the instructions to a bundle.
3138 BundleMember->FirstInBundle = Bundle;
3139 PrevInBundle = BundleMember;
3140 }
3141 if (ScheduleEnd != OldScheduleEnd) {
3142 // The scheduling region got new instructions at the lower end (or it is a
3143 // new region for the first bundle). This makes it necessary to
3144 // recalculate all dependencies.
3145 // It is seldom that this needs to be done a second time after adding the
3146 // initial bundle to the region.
3147 for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
3148 ScheduleData *SD = getScheduleData(I);
3149 SD->clearDependencies();
3150 }
3151 ReSchedule = true;
3152 }
3153 if (ReSchedule) {
3154 resetSchedule();
3155 initialFillReadyList(ReadyInsts);
3156 }
3157
3158 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)
3159 << BB->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: try schedule bundle " <<
*Bundle << " in block " << BB->getName() <<
"\n"; } } while (false)
;
3160
3161 calculateDependencies(Bundle, true, SLP);
3162
3163 // Now try to schedule the new bundle. As soon as the bundle is "ready" it
3164 // means that there are no cyclic dependencies and we can schedule it.
3165 // Note that's important that we don't "schedule" the bundle yet (see
3166 // cancelScheduling).
3167 while (!Bundle->isReady() && !ReadyInsts.empty()) {
3168
3169 ScheduleData *pickedSD = ReadyInsts.back();
3170 ReadyInsts.pop_back();
3171
3172 if (pickedSD->isSchedulingEntity() && pickedSD->isReady()) {
3173 schedule(pickedSD, ReadyInsts);
3174 }
3175 }
3176 if (!Bundle->isReady()) {
3177 cancelScheduling(VL);
3178 return false;
3179 }
3180 return true;
3181}
3182
3183void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef<Value *> VL) {
3184 if (isa<PHINode>(VL[0]))
3185 return;
3186
3187 ScheduleData *Bundle = getScheduleData(VL[0]);
3188 DEBUG(dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: cancel scheduling of " <<
*Bundle << "\n"; } } while (false)
;
3189 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3190, __PRETTY_FUNCTION__))
3190 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3190, __PRETTY_FUNCTION__))
;
3191 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3192, __PRETTY_FUNCTION__))
3192 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3192, __PRETTY_FUNCTION__))
;
3193
3194 // Un-bundle: make single instructions out of the bundle.
3195 ScheduleData *BundleMember = Bundle;
3196 while (BundleMember) {
3197 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3197, __PRETTY_FUNCTION__))
;
3198 BundleMember->FirstInBundle = BundleMember;
3199 ScheduleData *Next = BundleMember->NextInBundle;
3200 BundleMember->NextInBundle = nullptr;
3201 BundleMember->UnscheduledDepsInBundle = BundleMember->UnscheduledDeps;
3202 if (BundleMember->UnscheduledDepsInBundle == 0) {
3203 ReadyInsts.insert(BundleMember);
3204 }
3205 BundleMember = Next;
3206 }
3207}
3208
3209bool BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V) {
3210 if (getScheduleData(V))
3211 return true;
3212 Instruction *I = dyn_cast<Instruction>(V);
3213 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3213, __PRETTY_FUNCTION__))
;
3214 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3214, __PRETTY_FUNCTION__))
;
3215 if (!ScheduleStart) {
3216 // It's the first instruction in the new region.
3217 initScheduleData(I, I->getNextNode(), nullptr, nullptr);
3218 ScheduleStart = I;
3219 ScheduleEnd = I->getNextNode();
3220 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3220, __PRETTY_FUNCTION__))
;
3221 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)
;
3222 return true;
3223 }
3224 // Search up and down at the same time, because we don't know if the new
3225 // instruction is above or below the existing scheduling region.
3226 BasicBlock::reverse_iterator UpIter =
3227 ++ScheduleStart->getIterator().getReverse();
3228 BasicBlock::reverse_iterator UpperEnd = BB->rend();
3229 BasicBlock::iterator DownIter = ScheduleEnd->getIterator();
3230 BasicBlock::iterator LowerEnd = BB->end();
3231 for (;;) {
3232 if (++ScheduleRegionSize > ScheduleRegionSizeLimit) {
3233 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)
;
3234 return false;
3235 }
3236
3237 if (UpIter != UpperEnd) {
3238 if (&*UpIter == I) {
3239 initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion);
3240 ScheduleStart = I;
3241 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)
;
3242 return true;
3243 }
3244 UpIter++;
3245 }
3246 if (DownIter != LowerEnd) {
3247 if (&*DownIter == I) {
3248 initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion,
3249 nullptr);
3250 ScheduleEnd = I->getNextNode();
3251 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3251, __PRETTY_FUNCTION__))
;
3252 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)
;
3253 return true;
3254 }
3255 DownIter++;
3256 }
3257 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3258, __PRETTY_FUNCTION__))
3258 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3258, __PRETTY_FUNCTION__))
;
3259 }
3260 return true;
3261}
3262
3263void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI,
3264 Instruction *ToI,
3265 ScheduleData *PrevLoadStore,
3266 ScheduleData *NextLoadStore) {
3267 ScheduleData *CurrentLoadStore = PrevLoadStore;
3268 for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) {
3269 ScheduleData *SD = ScheduleDataMap[I];
3270 if (!SD) {
3271 // Allocate a new ScheduleData for the instruction.
3272 if (ChunkPos >= ChunkSize) {
3273 ScheduleDataChunks.push_back(
3274 llvm::make_unique<ScheduleData[]>(ChunkSize));
3275 ChunkPos = 0;
3276 }
3277 SD = &(ScheduleDataChunks.back()[ChunkPos++]);
3278 ScheduleDataMap[I] = SD;
3279 SD->Inst = I;
3280 }
3281 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3282, __PRETTY_FUNCTION__))
3282 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3282, __PRETTY_FUNCTION__))
;
3283 SD->init(SchedulingRegionID);
3284
3285 if (I->mayReadOrWriteMemory()) {
3286 // Update the linked list of memory accessing instructions.
3287 if (CurrentLoadStore) {
3288 CurrentLoadStore->NextLoadStore = SD;
3289 } else {
3290 FirstLoadStoreInRegion = SD;
3291 }
3292 CurrentLoadStore = SD;
3293 }
3294 }
3295 if (NextLoadStore) {
3296 if (CurrentLoadStore)
3297 CurrentLoadStore->NextLoadStore = NextLoadStore;
3298 } else {
3299 LastLoadStoreInRegion = CurrentLoadStore;
3300 }
3301}
3302
3303void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD,
3304 bool InsertInReadyList,
3305 BoUpSLP *SLP) {
3306 assert(SD->isSchedulingEntity())((SD->isSchedulingEntity()) ? static_cast<void> (0) :
__assert_fail ("SD->isSchedulingEntity()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3306, __PRETTY_FUNCTION__))
;
3307
3308 SmallVector<ScheduleData *, 10> WorkList;
3309 WorkList.push_back(SD);
3310
3311 while (!WorkList.empty()) {
3312 ScheduleData *SD = WorkList.back();
3313 WorkList.pop_back();
3314
3315 ScheduleData *BundleMember = SD;
3316 while (BundleMember) {
3317 assert(isInSchedulingRegion(BundleMember))((isInSchedulingRegion(BundleMember)) ? static_cast<void>
(0) : __assert_fail ("isInSchedulingRegion(BundleMember)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3317, __PRETTY_FUNCTION__))
;
3318 if (!BundleMember->hasValidDependencies()) {
3319
3320 DEBUG(dbgs() << "SLP: update deps of " << *BundleMember << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: update deps of " <<
*BundleMember << "\n"; } } while (false)
;
3321 BundleMember->Dependencies = 0;
3322 BundleMember->resetUnscheduledDeps();
3323
3324 // Handle def-use chain dependencies.
3325 for (User *U : BundleMember->Inst->users()) {
3326 if (isa<Instruction>(U)) {
3327 ScheduleData *UseSD = getScheduleData(U);
3328 if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) {
3329 BundleMember->Dependencies++;
3330 ScheduleData *DestBundle = UseSD->FirstInBundle;
3331 if (!DestBundle->IsScheduled) {
3332 BundleMember->incrementUnscheduledDeps(1);
3333 }
3334 if (!DestBundle->hasValidDependencies()) {
3335 WorkList.push_back(DestBundle);
3336 }
3337 }
3338 } else {
3339 // I'm not sure if this can ever happen. But we need to be safe.
3340 // This lets the instruction/bundle never be scheduled and
3341 // eventually disable vectorization.
3342 BundleMember->Dependencies++;
3343 BundleMember->incrementUnscheduledDeps(1);
3344 }
3345 }
3346
3347 // Handle the memory dependencies.
3348 ScheduleData *DepDest = BundleMember->NextLoadStore;
3349 if (DepDest) {
3350 Instruction *SrcInst = BundleMember->Inst;
3351 MemoryLocation SrcLoc = getLocation(SrcInst, SLP->AA);
3352 bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory();
3353 unsigned numAliased = 0;
3354 unsigned DistToSrc = 1;
3355
3356 while (DepDest) {
3357 assert(isInSchedulingRegion(DepDest))((isInSchedulingRegion(DepDest)) ? static_cast<void> (0
) : __assert_fail ("isInSchedulingRegion(DepDest)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3357, __PRETTY_FUNCTION__))
;
3358
3359 // We have two limits to reduce the complexity:
3360 // 1) AliasedCheckLimit: It's a small limit to reduce calls to
3361 // SLP->isAliased (which is the expensive part in this loop).
3362 // 2) MaxMemDepDistance: It's for very large blocks and it aborts
3363 // the whole loop (even if the loop is fast, it's quadratic).
3364 // It's important for the loop break condition (see below) to
3365 // check this limit even between two read-only instructions.
3366 if (DistToSrc >= MaxMemDepDistance ||
3367 ((SrcMayWrite || DepDest->Inst->mayWriteToMemory()) &&
3368 (numAliased >= AliasedCheckLimit ||
3369 SLP->isAliased(SrcLoc, SrcInst, DepDest->Inst)))) {
3370
3371 // We increment the counter only if the locations are aliased
3372 // (instead of counting all alias checks). This gives a better
3373 // balance between reduced runtime and accurate dependencies.
3374 numAliased++;
3375
3376 DepDest->MemoryDependencies.push_back(BundleMember);
3377 BundleMember->Dependencies++;
3378 ScheduleData *DestBundle = DepDest->FirstInBundle;
3379 if (!DestBundle->IsScheduled) {
3380 BundleMember->incrementUnscheduledDeps(1);
3381 }
3382 if (!DestBundle->hasValidDependencies()) {
3383 WorkList.push_back(DestBundle);
3384 }
3385 }
3386 DepDest = DepDest->NextLoadStore;
3387
3388 // Example, explaining the loop break condition: Let's assume our
3389 // starting instruction is i0 and MaxMemDepDistance = 3.
3390 //
3391 // +--------v--v--v
3392 // i0,i1,i2,i3,i4,i5,i6,i7,i8
3393 // +--------^--^--^
3394 //
3395 // MaxMemDepDistance let us stop alias-checking at i3 and we add
3396 // dependencies from i0 to i3,i4,.. (even if they are not aliased).
3397 // Previously we already added dependencies from i3 to i6,i7,i8
3398 // (because of MaxMemDepDistance). As we added a dependency from
3399 // i0 to i3, we have transitive dependencies from i0 to i6,i7,i8
3400 // and we can abort this loop at i6.
3401 if (DistToSrc >= 2 * MaxMemDepDistance)
3402 break;
3403 DistToSrc++;
3404 }
3405 }
3406 }
3407 BundleMember = BundleMember->NextInBundle;
3408 }
3409 if (InsertInReadyList && SD->isReady()) {
3410 ReadyInsts.push_back(SD);
3411 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)
;
3412 }
3413 }
3414}
3415
3416void BoUpSLP::BlockScheduling::resetSchedule() {
3417 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3418, __PRETTY_FUNCTION__))
3418 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3418, __PRETTY_FUNCTION__))
;
3419 for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
3420 ScheduleData *SD = getScheduleData(I);
3421 assert(isInSchedulingRegion(SD))((isInSchedulingRegion(SD)) ? static_cast<void> (0) : __assert_fail
("isInSchedulingRegion(SD)", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3421, __PRETTY_FUNCTION__))
;
3422 SD->IsScheduled = false;
3423 SD->resetUnscheduledDeps();
3424 }
3425 ReadyInsts.clear();
3426}
3427
3428void BoUpSLP::scheduleBlock(BlockScheduling *BS) {
3429
3430 if (!BS->ScheduleStart)
3431 return;
3432
3433 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)
;
3434
3435 BS->resetSchedule();
3436
3437 // For the real scheduling we use a more sophisticated ready-list: it is
3438 // sorted by the original instruction location. This lets the final schedule
3439 // be as close as possible to the original instruction order.
3440 struct ScheduleDataCompare {
3441 bool operator()(ScheduleData *SD1, ScheduleData *SD2) const {
3442 return SD2->SchedulingPriority < SD1->SchedulingPriority;
3443 }
3444 };
3445 std::set<ScheduleData *, ScheduleDataCompare> ReadyInsts;
3446
3447 // Ensure that all dependency data is updated and fill the ready-list with
3448 // initial instructions.
3449 int Idx = 0;
3450 int NumToSchedule = 0;
3451 for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd;
3452 I = I->getNextNode()) {
3453 ScheduleData *SD = BS->getScheduleData(I);
3454 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3456, __PRETTY_FUNCTION__))
3455 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3456, __PRETTY_FUNCTION__))
3456 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3456, __PRETTY_FUNCTION__))
;
3457 SD->FirstInBundle->SchedulingPriority = Idx++;
3458 if (SD->isSchedulingEntity()) {
3459 BS->calculateDependencies(SD, false, this);
3460 NumToSchedule++;
3461 }
3462 }
3463 BS->initialFillReadyList(ReadyInsts);
3464
3465 Instruction *LastScheduledInst = BS->ScheduleEnd;
3466
3467 // Do the "real" scheduling.
3468 while (!ReadyInsts.empty()) {
3469 ScheduleData *picked = *ReadyInsts.begin();
3470 ReadyInsts.erase(ReadyInsts.begin());
3471
3472 // Move the scheduled instruction(s) to their dedicated places, if not
3473 // there yet.
3474 ScheduleData *BundleMember = picked;
3475 while (BundleMember) {
3476 Instruction *pickedInst = BundleMember->Inst;
3477 if (LastScheduledInst->getNextNode() != pickedInst) {
3478 BS->BB->getInstList().remove(pickedInst);
3479 BS->BB->getInstList().insert(LastScheduledInst->getIterator(),
3480 pickedInst);
3481 }
3482 LastScheduledInst = pickedInst;
3483 BundleMember = BundleMember->NextInBundle;
3484 }
3485
3486 BS->schedule(picked, ReadyInsts);
3487 NumToSchedule--;
3488 }
3489 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 3489, __PRETTY_FUNCTION__))
;
3490
3491 // Avoid duplicate scheduling of the block.
3492 BS->ScheduleStart = nullptr;
3493}
3494
3495unsigned BoUpSLP::getVectorElementSize(Value *V) {
3496 // If V is a store, just return the width of the stored value without
3497 // traversing the expression tree. This is the common case.
3498 if (auto *Store = dyn_cast<StoreInst>(V))
3499 return DL->getTypeSizeInBits(Store->getValueOperand()->getType());
3500
3501 // If V is not a store, we can traverse the expression tree to find loads
3502 // that feed it. The type of the loaded value may indicate a more suitable
3503 // width than V's type. We want to base the vector element size on the width
3504 // of memory operations where possible.
3505 SmallVector<Instruction *, 16> Worklist;
3506 SmallPtrSet<Instruction *, 16> Visited;
3507 if (auto *I = dyn_cast<Instruction>(V))
3508 Worklist.push_back(I);
3509
3510 // Traverse the expression tree in bottom-up order looking for loads. If we
3511 // encounter an instruciton we don't yet handle, we give up.
3512 auto MaxWidth = 0u;
3513 auto FoundUnknownInst = false;
3514 while (!Worklist.empty() && !FoundUnknownInst) {
3515 auto *I = Worklist.pop_back_val();
3516 Visited.insert(I);
3517
3518 // We should only be looking at scalar instructions here. If the current
3519 // instruction has a vector type, give up.
3520 auto *Ty = I->getType();
3521 if (isa<VectorType>(Ty))
3522 FoundUnknownInst = true;
3523
3524 // If the current instruction is a load, update MaxWidth to reflect the
3525 // width of the loaded value.
3526 else if (isa<LoadInst>(I))
3527 MaxWidth = std::max<unsigned>(MaxWidth, DL->getTypeSizeInBits(Ty));
3528
3529 // Otherwise, we need to visit the operands of the instruction. We only
3530 // handle the interesting cases from buildTree here. If an operand is an
3531 // instruction we haven't yet visited, we add it to the worklist.
3532 else if (isa<PHINode>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) ||
3533 isa<CmpInst>(I) || isa<SelectInst>(I) || isa<BinaryOperator>(I)) {
3534 for (Use &U : I->operands())
3535 if (auto *J = dyn_cast<Instruction>(U.get()))
3536 if (!Visited.count(J))
3537 Worklist.push_back(J);
3538 }
3539
3540 // If we don't yet handle the instruction, give up.
3541 else
3542 FoundUnknownInst = true;
3543 }
3544
3545 // If we didn't encounter a memory access in the expression tree, or if we
3546 // gave up for some reason, just return the width of V.
3547 if (!MaxWidth || FoundUnknownInst)
3548 return DL->getTypeSizeInBits(V->getType());
3549
3550 // Otherwise, return the maximum width we found.
3551 return MaxWidth;
3552}
3553
3554// Determine if a value V in a vectorizable expression Expr can be demoted to a
3555// smaller type with a truncation. We collect the values that will be demoted
3556// in ToDemote and additional roots that require investigating in Roots.
3557static bool collectValuesToDemote(Value *V, SmallPtrSetImpl<Value *> &Expr,
3558 SmallVectorImpl<Value *> &ToDemote,
3559 SmallVectorImpl<Value *> &Roots) {
3560
3561 // We can always demote constants.
3562 if (isa<Constant>(V)) {
3563 ToDemote.push_back(V);
3564 return true;
3565 }
3566
3567 // If the value is not an instruction in the expression with only one use, it
3568 // cannot be demoted.
3569 auto *I = dyn_cast<Instruction>(V);
3570 if (!I || !I->hasOneUse() || !Expr.count(I))
3571 return false;
3572
3573 switch (I->getOpcode()) {
3574
3575 // We can always demote truncations and extensions. Since truncations can
3576 // seed additional demotion, we save the truncated value.
3577 case Instruction::Trunc:
3578 Roots.push_back(I->getOperand(0));
3579 case Instruction::ZExt:
3580 case Instruction::SExt:
3581 break;
3582
3583 // We can demote certain binary operations if we can demote both of their
3584 // operands.
3585 case Instruction::Add:
3586 case Instruction::Sub:
3587 case Instruction::Mul:
3588 case Instruction::And:
3589 case Instruction::Or:
3590 case Instruction::Xor:
3591 if (!collectValuesToDemote(I->getOperand(0), Expr, ToDemote, Roots) ||
3592 !collectValuesToDemote(I->getOperand(1), Expr, ToDemote, Roots))
3593 return false;
3594 break;
3595
3596 // We can demote selects if we can demote their true and false values.
3597 case Instruction::Select: {
3598 SelectInst *SI = cast<SelectInst>(I);
3599 if (!collectValuesToDemote(SI->getTrueValue(), Expr, ToDemote, Roots) ||
3600 !collectValuesToDemote(SI->getFalseValue(), Expr, ToDemote, Roots))
3601 return false;
3602 break;
3603 }
3604
3605 // We can demote phis if we can demote all their incoming operands. Note that
3606 // we don't need to worry about cycles since we ensure single use above.
3607 case Instruction::PHI: {
3608 PHINode *PN = cast<PHINode>(I);
3609 for (Value *IncValue : PN->incoming_values())
3610 if (!collectValuesToDemote(IncValue, Expr, ToDemote, Roots))
3611 return false;
3612 break;
3613 }
3614
3615 // Otherwise, conservatively give up.
3616 default:
3617 return false;
3618 }
3619
3620 // Record the value that we can demote.
3621 ToDemote.push_back(V);
3622 return true;
3623}
3624
3625void BoUpSLP::computeMinimumValueSizes() {
3626 // If there are no external uses, the expression tree must be rooted by a
3627 // store. We can't demote in-memory values, so there is nothing to do here.
3628 if (ExternalUses.empty())
3629 return;
3630
3631 // We only attempt to truncate integer expressions.
3632 auto &TreeRoot = VectorizableTree[0].Scalars;
3633 auto *TreeRootIT = dyn_cast<IntegerType>(TreeRoot[0]->getType());
3634 if (!TreeRootIT)
3635 return;
3636
3637 // If the expression is not rooted by a store, these roots should have
3638 // external uses. We will rely on InstCombine to rewrite the expression in
3639 // the narrower type. However, InstCombine only rewrites single-use values.
3640 // This means that if a tree entry other than a root is used externally, it
3641 // must have multiple uses and InstCombine will not rewrite it. The code
3642 // below ensures that only the roots are used externally.
3643 SmallPtrSet<Value *, 32> Expr(TreeRoot.begin(), TreeRoot.end());
3644 for (auto &EU : ExternalUses)
3645 if (!Expr.erase(EU.Scalar))
3646 return;
3647 if (!Expr.empty())
3648 return;
3649
3650 // Collect the scalar values of the vectorizable expression. We will use this
3651 // context to determine which values can be demoted. If we see a truncation,
3652 // we mark it as seeding another demotion.
3653 for (auto &Entry : VectorizableTree)
3654 Expr.insert(Entry.Scalars.begin(), Entry.Scalars.end());
3655
3656 // Ensure the roots of the vectorizable tree don't form a cycle. They must
3657 // have a single external user that is not in the vectorizable tree.
3658 for (auto *Root : TreeRoot)
3659 if (!Root->hasOneUse() || Expr.count(*Root->user_begin()))
3660 return;
3661
3662 // Conservatively determine if we can actually truncate the roots of the
3663 // expression. Collect the values that can be demoted in ToDemote and
3664 // additional roots that require investigating in Roots.
3665 SmallVector<Value *, 32> ToDemote;
3666 SmallVector<Value *, 4> Roots;
3667 for (auto *Root : TreeRoot)
3668 if (!collectValuesToDemote(Root, Expr, ToDemote, Roots))
3669 return;
3670
3671 // The maximum bit width required to represent all the values that can be
3672 // demoted without loss of precision. It would be safe to truncate the roots
3673 // of the expression to this width.
3674 auto MaxBitWidth = 8u;
3675
3676 // We first check if all the bits of the roots are demanded. If they're not,
3677 // we can truncate the roots to this narrower type.
3678 for (auto *Root : TreeRoot) {
3679 auto Mask = DB->getDemandedBits(cast<Instruction>(Root));
3680 MaxBitWidth = std::max<unsigned>(
3681 Mask.getBitWidth() - Mask.countLeadingZeros(), MaxBitWidth);
3682 }
3683
3684 // True if the roots can be zero-extended back to their original type, rather
3685 // than sign-extended. We know that if the leading bits are not demanded, we
3686 // can safely zero-extend. So we initialize IsKnownPositive to True.
3687 bool IsKnownPositive = true;
3688
3689 // If all the bits of the roots are demanded, we can try a little harder to
3690 // compute a narrower type. This can happen, for example, if the roots are
3691 // getelementptr indices. InstCombine promotes these indices to the pointer
3692 // width. Thus, all their bits are technically demanded even though the
3693 // address computation might be vectorized in a smaller type.
3694 //
3695 // We start by looking at each entry that can be demoted. We compute the
3696 // maximum bit width required to store the scalar by using ValueTracking to
3697 // compute the number of high-order bits we can truncate.
3698 if (MaxBitWidth == DL->getTypeSizeInBits(TreeRoot[0]->getType())) {
3699 MaxBitWidth = 8u;
3700
3701 // Determine if the sign bit of all the roots is known to be zero. If not,
3702 // IsKnownPositive is set to False.
3703 IsKnownPositive = all_of(TreeRoot, [&](Value *R) {
3704 bool KnownZero = false;
3705 bool KnownOne = false;
3706 ComputeSignBit(R, KnownZero, KnownOne, *DL);
3707 return KnownZero;
3708 });
3709
3710 // Determine the maximum number of bits required to store the scalar
3711 // values.
3712 for (auto *Scalar : ToDemote) {
3713 auto NumSignBits = ComputeNumSignBits(Scalar, *DL, 0, AC, 0, DT);
3714 auto NumTypeBits = DL->getTypeSizeInBits(Scalar->getType());
3715 MaxBitWidth = std::max<unsigned>(NumTypeBits - NumSignBits, MaxBitWidth);
3716 }
3717
3718 // If we can't prove that the sign bit is zero, we must add one to the
3719 // maximum bit width to account for the unknown sign bit. This preserves
3720 // the existing sign bit so we can safely sign-extend the root back to the
3721 // original type. Otherwise, if we know the sign bit is zero, we will
3722 // zero-extend the root instead.
3723 //
3724 // FIXME: This is somewhat suboptimal, as there will be cases where adding
3725 // one to the maximum bit width will yield a larger-than-necessary
3726 // type. In general, we need to add an extra bit only if we can't
3727 // prove that the upper bit of the original type is equal to the
3728 // upper bit of the proposed smaller type. If these two bits are the
3729 // same (either zero or one) we know that sign-extending from the
3730 // smaller type will result in the same value. Here, since we can't
3731 // yet prove this, we are just making the proposed smaller type
3732 // larger to ensure correctness.
3733 if (!IsKnownPositive)
3734 ++MaxBitWidth;
3735 }
3736
3737 // Round MaxBitWidth up to the next power-of-two.
3738 if (!isPowerOf2_64(MaxBitWidth))
3739 MaxBitWidth = NextPowerOf2(MaxBitWidth);
3740
3741 // If the maximum bit width we compute is less than the with of the roots'
3742 // type, we can proceed with the narrowing. Otherwise, do nothing.
3743 if (MaxBitWidth >= TreeRootIT->getBitWidth())
3744 return;
3745
3746 // If we can truncate the root, we must collect additional values that might
3747 // be demoted as a result. That is, those seeded by truncations we will
3748 // modify.
3749 while (!Roots.empty())
3750 collectValuesToDemote(Roots.pop_back_val(), Expr, ToDemote, Roots);
3751
3752 // Finally, map the values we can demote to the maximum bit with we computed.
3753 for (auto *Scalar : ToDemote)
3754 MinBWs[Scalar] = std::make_pair(MaxBitWidth, !IsKnownPositive);
3755}
3756
3757namespace {
3758/// The SLPVectorizer Pass.
3759struct SLPVectorizer : public FunctionPass {
3760 SLPVectorizerPass Impl;
3761
3762 /// Pass identification, replacement for typeid
3763 static char ID;
3764
3765 explicit SLPVectorizer() : FunctionPass(ID) {
3766 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
3767 }
3768
3769
3770 bool doInitialization(Module &M) override {
3771 return false;
3772 }
3773
3774 bool runOnFunction(Function &F) override {
3775 if (skipFunction(F))
3776 return false;
3777
3778 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
3779 auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
3780 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
3781 auto *TLI = TLIP ? &TLIP->getTLI() : nullptr;
3782 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
3783 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
3784 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
3785 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
3786 auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits();
3787
3788 return Impl.runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB);
3789 }
3790
3791 void getAnalysisUsage(AnalysisUsage &AU) const override {
3792 FunctionPass::getAnalysisUsage(AU);
3793 AU.addRequired<AssumptionCacheTracker>();
3794 AU.addRequired<ScalarEvolutionWrapperPass>();
3795 AU.addRequired<AAResultsWrapperPass>();
3796 AU.addRequired<TargetTransformInfoWrapperPass>();
3797 AU.addRequired<LoopInfoWrapperPass>();
3798 AU.addRequired<DominatorTreeWrapperPass>();
3799 AU.addRequired<DemandedBitsWrapperPass>();
3800 AU.addPreserved<LoopInfoWrapperPass>();
3801 AU.addPreserved<DominatorTreeWrapperPass>();
3802 AU.addPreserved<AAResultsWrapperPass>();
3803 AU.addPreserved<GlobalsAAWrapperPass>();
3804 AU.setPreservesCFG();
3805 }
3806};
3807} // end anonymous namespace
3808
3809PreservedAnalyses SLPVectorizerPass::run(Function &F, FunctionAnalysisManager &AM) {
3810 auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F);
3811 auto *TTI = &AM.getResult<TargetIRAnalysis>(F);
3812 auto *TLI = AM.getCachedResult<TargetLibraryAnalysis>(F);
3813 auto *AA = &AM.getResult<AAManager>(F);
3814 auto *LI = &AM.getResult<LoopAnalysis>(F);
3815 auto *DT = &AM.getResult<DominatorTreeAnalysis>(F);
3816 auto *AC = &AM.getResult<AssumptionAnalysis>(F);
3817 auto *DB = &AM.getResult<DemandedBitsAnalysis>(F);
3818
3819 bool Changed = runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB);
3820 if (!Changed)
3821 return PreservedAnalyses::all();
3822
3823 PreservedAnalyses PA;
3824 PA.preserveSet<CFGAnalyses>();
3825 PA.preserve<AAManager>();
3826 PA.preserve<GlobalsAA>();
3827 return PA;
3828}
3829
3830bool SLPVectorizerPass::runImpl(Function &F, ScalarEvolution *SE_,
3831 TargetTransformInfo *TTI_,
3832 TargetLibraryInfo *TLI_, AliasAnalysis *AA_,
3833 LoopInfo *LI_, DominatorTree *DT_,
3834 AssumptionCache *AC_, DemandedBits *DB_) {
3835 SE = SE_;
3836 TTI = TTI_;
3837 TLI = TLI_;
3838 AA = AA_;
3839 LI = LI_;
3840 DT = DT_;
3841 AC = AC_;
3842 DB = DB_;
3843 DL = &F.getParent()->getDataLayout();
3844
3845 Stores.clear();
3846 GEPs.clear();
3847 bool Changed = false;
3848
3849 // If the target claims to have no vector registers don't attempt
3850 // vectorization.
3851 if (!TTI->getNumberOfRegisters(true))
3852 return false;
3853
3854 // Don't vectorize when the attribute NoImplicitFloat is used.
3855 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
3856 return false;
3857
3858 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)
;
3859
3860 // Use the bottom up slp vectorizer to construct chains that start with
3861 // store instructions.
3862 BoUpSLP R(&F, SE, TTI, TLI, AA, LI, DT, AC, DB, DL);
3863
3864 // A general note: the vectorizer must use BoUpSLP::eraseInstruction() to
3865 // delete instructions.
3866
3867 // Scan the blocks in the function in post order.
3868 for (auto BB : post_order(&F.getEntryBlock())) {
3869 collectSeedInstructions(BB);
3870
3871 // Vectorize trees that end at stores.
3872 if (!Stores.empty()) {
3873 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)
3874 << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Found stores for " << Stores
.size() << " underlying objects.\n"; } } while (false)
;
3875 Changed |= vectorizeStoreChains(R);
3876 }
3877
3878 // Vectorize trees that end at reductions.
3879 Changed |= vectorizeChainsInBlock(BB, R);
3880
3881 // Vectorize the index computations of getelementptr instructions. This
3882 // is primarily intended to catch gather-like idioms ending at
3883 // non-consecutive loads.
3884 if (!GEPs.empty()) {
3885 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)
3886 << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs
.size() << " underlying objects.\n"; } } while (false)
;
3887 Changed |= vectorizeGEPIndices(BB, R);
3888 }
3889 }
3890
3891 if (Changed) {
3892 R.optimizeGatherSequence();
3893 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: vectorized \"" << F.getName
() << "\"\n"; } } while (false)
;
3894 DEBUG(verifyFunction(F))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { verifyFunction(F); } } while (false)
;
3895 }
3896 return Changed;
3897}
3898
3899/// \brief Check that the Values in the slice in VL array are still existent in
3900/// the WeakVH array.
3901/// Vectorization of part of the VL array may cause later values in the VL array
3902/// to become invalid. We track when this has happened in the WeakVH array.
3903static bool hasValueBeenRAUWed(ArrayRef<Value *> VL, ArrayRef<WeakVH> VH,
3904 unsigned SliceBegin, unsigned SliceSize) {
3905 VL = VL.slice(SliceBegin, SliceSize);
3906 VH = VH.slice(SliceBegin, SliceSize);
3907 return !std::equal(VL.begin(), VL.end(), VH.begin());
3908}
3909
3910bool SLPVectorizerPass::vectorizeStoreChain(ArrayRef<Value *> Chain, BoUpSLP &R,
3911 unsigned VecRegSize) {
3912 unsigned ChainLen = Chain.size();
3913 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)
3914 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing a store chain of length "
<< ChainLen << "\n"; } } while (false)
;
3915 unsigned Sz = R.getVectorElementSize(Chain[0]);
3916 unsigned VF = VecRegSize / Sz;
3917
3918 if (!isPowerOf2_32(Sz) || VF < 2)
3919 return false;
3920
3921 // Keep track of values that were deleted by vectorizing in the loop below.
3922 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end());
3923
3924 bool Changed = false;
3925 // Look for profitable vectorizable trees at all offsets, starting at zero.
3926 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
3927 if (i + VF > e)
3928 break;
3929
3930 // Check that a previous iteration of this loop did not delete the Value.
3931 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF))
3932 continue;
3933
3934 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
)
3935 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing " << VF <<
" stores at offset " << i << "\n"; } } while (false
)
;
3936 ArrayRef<Value *> Operands = Chain.slice(i, VF);
3937
3938 R.buildTree(Operands);
3939 if (R.isTreeTinyAndNotFullyVectorizable())
3940 continue;
3941
3942 R.computeMinimumValueSizes();
3943
3944 int Cost = R.getTreeCost();
3945
3946 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)
;
3947 if (Cost < -SLPCostThreshold) {
3948 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)
;
3949 R.vectorizeTree();
3950
3951 // Move to the next bundle.
3952 i += VF - 1;
3953 Changed = true;
3954 }
3955 }
3956
3957 return Changed;
3958}
3959
3960bool SLPVectorizerPass::vectorizeStores(ArrayRef<StoreInst *> Stores,
3961 BoUpSLP &R) {
3962 SetVector<StoreInst *> Heads, Tails;
3963 SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain;
3964
3965 // We may run into multiple chains that merge into a single chain. We mark the
3966 // stores that we vectorized so that we don't visit the same store twice.
3967 BoUpSLP::ValueSet VectorizedStores;
3968 bool Changed = false;
3969
3970 // Do a quadratic search on all of the given stores and find
3971 // all of the pairs of stores that follow each other.
3972 SmallVector<unsigned, 16> IndexQueue;
3973 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
3974 IndexQueue.clear();
3975 // If a store has multiple consecutive store candidates, search Stores
3976 // array according to the sequence: from i+1 to e, then from i-1 to 0.
3977 // This is because usually pairing with immediate succeeding or preceding
3978 // candidate create the best chance to find slp vectorization opportunity.
3979 unsigned j = 0;
3980 for (j = i + 1; j < e; ++j)
3981 IndexQueue.push_back(j);
3982 for (j = i; j > 0; --j)
3983 IndexQueue.push_back(j - 1);
3984
3985 for (auto &k : IndexQueue) {
3986 if (isConsecutiveAccess(Stores[i], Stores[k], *DL, *SE)) {
3987 Tails.insert(Stores[k]);
3988 Heads.insert(Stores[i]);
3989 ConsecutiveChain[Stores[i]] = Stores[k];
3990 break;
3991 }
3992 }
3993 }
3994
3995 // For stores that start but don't end a link in the chain:
3996 for (SetVector<StoreInst *>::iterator it = Heads.begin(), e = Heads.end();
3997 it != e; ++it) {
3998 if (Tails.count(*it))
3999 continue;
4000
4001 // We found a store instr that starts a chain. Now follow the chain and try
4002 // to vectorize it.
4003 BoUpSLP::ValueList Operands;
4004 StoreInst *I = *it;
4005 // Collect the chain into a list.
4006 while (Tails.count(I) || Heads.count(I)) {
4007 if (VectorizedStores.count(I))
4008 break;
4009 Operands.push_back(I);
4010 // Move to the next value in the chain.
4011 I = ConsecutiveChain[I];
4012 }
4013
4014 // FIXME: Is division-by-2 the correct step? Should we assert that the
4015 // register size is a power-of-2?
4016 for (unsigned Size = R.getMaxVecRegSize(); Size >= R.getMinVecRegSize();
4017 Size /= 2) {
4018 if (vectorizeStoreChain(Operands, R, Size)) {
4019 // Mark the vectorized stores so that we don't vectorize them again.
4020 VectorizedStores.insert(Operands.begin(), Operands.end());
4021 Changed = true;
4022 break;
4023 }
4024 }
4025 }
4026
4027 return Changed;
4028}
4029
4030void SLPVectorizerPass::collectSeedInstructions(BasicBlock *BB) {
4031
4032 // Initialize the collections. We will make a single pass over the block.
4033 Stores.clear();
4034 GEPs.clear();
4035
4036 // Visit the store and getelementptr instructions in BB and organize them in
4037 // Stores and GEPs according to the underlying objects of their pointer
4038 // operands.
4039 for (Instruction &I : *BB) {
4040
4041 // Ignore store instructions that are volatile or have a pointer operand
4042 // that doesn't point to a scalar type.
4043 if (auto *SI = dyn_cast<StoreInst>(&I)) {
4044 if (!SI->isSimple())
4045 continue;
4046 if (!isValidElementType(SI->getValueOperand()->getType()))
4047 continue;
4048 Stores[GetUnderlyingObject(SI->getPointerOperand(), *DL)].push_back(SI);
4049 }
4050
4051 // Ignore getelementptr instructions that have more than one index, a
4052 // constant index, or a pointer operand that doesn't point to a scalar
4053 // type.
4054 else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
4055 auto Idx = GEP->idx_begin()->get();
4056 if (GEP->getNumIndices() > 1 || isa<Constant>(Idx))
4057 continue;
4058 if (!isValidElementType(Idx->getType()))
4059 continue;
4060 if (GEP->getType()->isVectorTy())
4061 continue;
4062 GEPs[GetUnderlyingObject(GEP->getPointerOperand(), *DL)].push_back(GEP);
4063 }
4064 }
4065}
4066
4067bool SLPVectorizerPass::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
4068 if (!A || !B)
4069 return false;
4070 Value *VL[] = { A, B };
4071 return tryToVectorizeList(VL, R, None, true);
4072}
4073
4074bool SLPVectorizerPass::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
4075 ArrayRef<Value *> BuildVector,
4076 bool AllowReorder) {
4077 if (VL.size() < 2)
4078 return false;
4079
4080 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)
4081 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = "
<< VL.size() << ".\n"; } } while (false)
;
4082
4083 // Check that all of the parts are scalar instructions of the same type.
4084 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
4085 if (!I0)
4086 return false;
4087
4088 unsigned Opcode0 = I0->getOpcode();
4089
4090 unsigned Sz = R.getVectorElementSize(I0);
4091 unsigned MinVF = std::max(2U, R.getMinVecRegSize() / Sz);
4092 unsigned MaxVF = std::max<unsigned>(PowerOf2Floor(VL.size()), MinVF);
4093 if (MaxVF < 2)
4094 return false;
4095
4096 for (Value *V : VL) {
4097 Type *Ty = V->getType();
4098 if (!isValidElementType(Ty))
4099 return false;
4100 Instruction *Inst = dyn_cast<Instruction>(V);
4101 if (!Inst || Inst->getOpcode() != Opcode0)
4102 return false;
4103 }
4104
4105 bool Changed = false;
4106
4107 // Keep track of values that were deleted by vectorizing in the loop below.
4108 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end());
4109
4110 unsigned NextInst = 0, MaxInst = VL.size();
4111 for (unsigned VF = MaxVF; NextInst + 1 < MaxInst && VF >= MinVF;
4112 VF /= 2) {
4113 // No actual vectorization should happen, if number of parts is the same as
4114 // provided vectorization factor (i.e. the scalar type is used for vector
4115 // code during codegen).
4116 auto *VecTy = VectorType::get(VL[0]->getType(), VF);
4117 if (TTI->getNumberOfParts(VecTy) == VF)
4118 continue;
4119 for (unsigned I = NextInst; I < MaxInst; ++I) {
4120 unsigned OpsWidth = 0;
4121
4122 if (I + VF > MaxInst)
4123 OpsWidth = MaxInst - I;
4124 else
4125 OpsWidth = VF;
4126
4127 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
4128 break;
4129
4130 // Check that a previous iteration of this loop did not delete the Value.
4131 if (hasValueBeenRAUWed(VL, TrackValues, I, OpsWidth))
4132 continue;
4133
4134 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth
<< " operations " << "\n"; } } while (false)
4135 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth
<< " operations " << "\n"; } } while (false)
;
4136 ArrayRef<Value *> Ops = VL.slice(I, OpsWidth);
4137
4138 ArrayRef<Value *> BuildVectorSlice;
4139 if (!BuildVector.empty())
4140 BuildVectorSlice = BuildVector.slice(I, OpsWidth);
4141
4142 R.buildTree(Ops, BuildVectorSlice);
4143 // TODO: check if we can allow reordering for more cases.
4144 if (AllowReorder && R.shouldReorder()) {
4145 // Conceptually, there is nothing actually preventing us from trying to
4146 // reorder a larger list. In fact, we do exactly this when vectorizing
4147 // reductions. However, at this point, we only expect to get here from
4148 // tryToVectorizePair().
4149 assert(Ops.size() == 2)((Ops.size() == 2) ? static_cast<void> (0) : __assert_fail
("Ops.size() == 2", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4149, __PRETTY_FUNCTION__))
;
4150 assert(BuildVectorSlice.empty())((BuildVectorSlice.empty()) ? static_cast<void> (0) : __assert_fail
("BuildVectorSlice.empty()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4150, __PRETTY_FUNCTION__))
;
4151 Value *ReorderedOps[] = {Ops[1], Ops[0]};
4152 R.buildTree(ReorderedOps, None);
4153 }
4154 if (R.isTreeTinyAndNotFullyVectorizable())
4155 continue;
4156
4157 R.computeMinimumValueSizes();
4158 int Cost = R.getTreeCost();
4159
4160 if (Cost < -SLPCostThreshold) {
4161 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)
;
4162 Value *VectorizedRoot = R.vectorizeTree();
4163
4164 // Reconstruct the build vector by extracting the vectorized root. This
4165 // way we handle the case where some elements of the vector are
4166 // undefined.
4167 // (return (inserelt <4 xi32> (insertelt undef (opd0) 0) (opd1) 2))
4168 if (!BuildVectorSlice.empty()) {
4169 // The insert point is the last build vector instruction. The
4170 // vectorized root will precede it. This guarantees that we get an
4171 // instruction. The vectorized tree could have been constant folded.
4172 Instruction *InsertAfter = cast<Instruction>(BuildVectorSlice.back());
4173 unsigned VecIdx = 0;
4174 for (auto &V : BuildVectorSlice) {
4175 IRBuilder<NoFolder> Builder(InsertAfter->getParent(),
4176 ++BasicBlock::iterator(InsertAfter));
4177 Instruction *I = cast<Instruction>(V);
4178 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4178, __PRETTY_FUNCTION__))
;
4179 Instruction *Extract =
4180 cast<Instruction>(Builder.CreateExtractElement(
4181 VectorizedRoot, Builder.getInt32(VecIdx++)));
4182 I->setOperand(1, Extract);
4183 I->removeFromParent();
4184 I->insertAfter(Extract);
4185 InsertAfter = I;
4186 }
4187 }
4188 // Move to the next bundle.
4189 I += VF - 1;
4190 NextInst = I + 1;
4191 Changed = true;
4192 }
4193 }
4194 }
4195
4196 return Changed;
4197}
4198
4199bool SLPVectorizerPass::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
4200 if (!V)
4201 return false;
4202
4203 Value *P = V->getParent();
4204
4205 // Vectorize in current basic block only.
4206 auto *Op0 = dyn_cast<Instruction>(V->getOperand(0));
4207 auto *Op1 = dyn_cast<Instruction>(V->getOperand(1));
4208 if (!Op0 || !Op1 || Op0->getParent() != P || Op1->getParent() != P)
4209 return false;
4210
4211 // Try to vectorize V.
4212 if (tryToVectorizePair(Op0, Op1, R))
4213 return true;
4214
4215 auto *A = dyn_cast<BinaryOperator>(Op0);
4216 auto *B = dyn_cast<BinaryOperator>(Op1);
4217 // Try to skip B.
4218 if (B && B->hasOneUse()) {
4219 auto *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
4220 auto *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
4221 if (B0 && B0->getParent() == P && tryToVectorizePair(A, B0, R))
4222 return true;
4223 if (B1 && B1->getParent() == P && tryToVectorizePair(A, B1, R))
4224 return true;
4225 }
4226
4227 // Try to skip A.
4228 if (A && A->hasOneUse()) {
4229 auto *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
4230 auto *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
4231 if (A0 && A0->getParent() == P && tryToVectorizePair(A0, B, R))
4232 return true;
4233 if (A1 && A1->getParent() == P && tryToVectorizePair(A1, B, R))
4234 return true;
4235 }
4236 return false;
4237}
4238
4239/// \brief Generate a shuffle mask to be used in a reduction tree.
4240///
4241/// \param VecLen The length of the vector to be reduced.
4242/// \param NumEltsToRdx The number of elements that should be reduced in the
4243/// vector.
4244/// \param IsPairwise Whether the reduction is a pairwise or splitting
4245/// reduction. A pairwise reduction will generate a mask of
4246/// <0,2,...> or <1,3,..> while a splitting reduction will generate
4247/// <2,3, undef,undef> for a vector of 4 and NumElts = 2.
4248/// \param IsLeft True will generate a mask of even elements, odd otherwise.
4249static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx,
4250 bool IsPairwise, bool IsLeft,
4251 IRBuilder<> &Builder) {
4252 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4252, __PRETTY_FUNCTION__))
;
4253
4254 SmallVector<Constant *, 32> ShuffleMask(
4255 VecLen, UndefValue::get(Builder.getInt32Ty()));
4256
4257 if (IsPairwise)
4258 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right).
4259 for (unsigned i = 0; i != NumEltsToRdx; ++i)
4260 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft);
4261 else
4262 // Move the upper half of the vector to the lower half.
4263 for (unsigned i = 0; i != NumEltsToRdx; ++i)
4264 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i);
4265
4266 return ConstantVector::get(ShuffleMask);
4267}
4268
4269namespace {
4270/// Model horizontal reductions.
4271///
4272/// A horizontal reduction is a tree of reduction operations (currently add and
4273/// fadd) that has operations that can be put into a vector as its leaf.
4274/// For example, this tree:
4275///
4276/// mul mul mul mul
4277/// \ / \ /
4278/// + +
4279/// \ /
4280/// +
4281/// This tree has "mul" as its reduced values and "+" as its reduction
4282/// operations. A reduction might be feeding into a store or a binary operation
4283/// feeding a phi.
4284/// ...
4285/// \ /
4286/// +
4287/// |
4288/// phi +=
4289///
4290/// Or:
4291/// ...
4292/// \ /
4293/// +
4294/// |
4295/// *p =
4296///
4297class HorizontalReduction {
4298 SmallVector<Value *, 16> ReductionOps;
4299 SmallVector<Value *, 32> ReducedVals;
4300 // Use map vector to make stable output.
4301 MapVector<Instruction *, Value *> ExtraArgs;
4302
4303 BinaryOperator *ReductionRoot = nullptr;
4304
4305 /// The opcode of the reduction.
4306 Instruction::BinaryOps ReductionOpcode = Instruction::BinaryOpsEnd;
4307 /// The opcode of the values we perform a reduction on.
4308 unsigned ReducedValueOpcode = 0;
4309 /// Should we model this reduction as a pairwise reduction tree or a tree that
4310 /// splits the vector in halves and adds those halves.
4311 bool IsPairwiseReduction = false;
4312
4313 /// Checks if the ParentStackElem.first should be marked as a reduction
4314 /// operation with an extra argument or as extra argument itself.
4315 void markExtraArg(std::pair<Instruction *, unsigned> &ParentStackElem,
4316 Value *ExtraArg) {
4317 if (ExtraArgs.count(ParentStackElem.first)) {
4318 ExtraArgs[ParentStackElem.first] = nullptr;
4319 // We ran into something like:
4320 // ParentStackElem.first = ExtraArgs[ParentStackElem.first] + ExtraArg.
4321 // The whole ParentStackElem.first should be considered as an extra value
4322 // in this case.
4323 // Do not perform analysis of remaining operands of ParentStackElem.first
4324 // instruction, this whole instruction is an extra argument.
4325 ParentStackElem.second = ParentStackElem.first->getNumOperands();
4326 } else {
4327 // We ran into something like:
4328 // ParentStackElem.first += ... + ExtraArg + ...
4329 ExtraArgs[ParentStackElem.first] = ExtraArg;
4330 }
4331 }
4332
4333public:
4334 HorizontalReduction() = default;
4335
4336 /// \brief Try to find a reduction tree.
4337 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B) {
4338 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4339, __PRETTY_FUNCTION__))
4339 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4339, __PRETTY_FUNCTION__))
;
4340
4341 // We could have a initial reductions that is not an add.
4342 // r *= v1 + v2 + v3 + v4
4343 // In such a case start looking for a tree rooted in the first '+'.
4344 if (Phi) {
4345 if (B->getOperand(0) == Phi) {
4346 Phi = nullptr;
4347 B = dyn_cast<BinaryOperator>(B->getOperand(1));
4348 } else if (B->getOperand(1) == Phi) {
4349 Phi = nullptr;
4350 B = dyn_cast<BinaryOperator>(B->getOperand(0));
4351 }
4352 }
4353
4354 if (!B)
4355 return false;
4356
4357 Type *Ty = B->getType();
4358 if (!isValidElementType(Ty))
4359 return false;
4360
4361 ReductionOpcode = B->getOpcode();
4362 ReducedValueOpcode = 0;
4363 ReductionRoot = B;
4364
4365 // We currently only support adds.
4366 if ((ReductionOpcode != Instruction::Add &&
4367 ReductionOpcode != Instruction::FAdd) ||
4368 !B->isAssociative())
4369 return false;
4370
4371 // Post order traverse the reduction tree starting at B. We only handle true
4372 // trees containing only binary operators or selects.
4373 SmallVector<std::pair<Instruction *, unsigned>, 32> Stack;
4374 Stack.push_back(std::make_pair(B, 0));
4375 while (!Stack.empty()) {
4376 Instruction *TreeN = Stack.back().first;
4377 unsigned EdgeToVist = Stack.back().second++;
4378 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode;
4379
4380 // Postorder vist.
4381 if (EdgeToVist == 2 || IsReducedValue) {
4382 if (IsReducedValue)
4383 ReducedVals.push_back(TreeN);
4384 else {
4385 auto I = ExtraArgs.find(TreeN);
4386 if (I != ExtraArgs.end() && !I->second) {
4387 // Check if TreeN is an extra argument of its parent operation.
4388 if (Stack.size() <= 1) {
4389 // TreeN can't be an extra argument as it is a root reduction
4390 // operation.
4391 return false;
4392 }
4393 // Yes, TreeN is an extra argument, do not add it to a list of
4394 // reduction operations.
4395 // Stack[Stack.size() - 2] always points to the parent operation.
4396 markExtraArg(Stack[Stack.size() - 2], TreeN);
4397 ExtraArgs.erase(TreeN);
4398 } else
4399 ReductionOps.push_back(TreeN);
4400 }
4401 // Retract.
4402 Stack.pop_back();
4403 continue;
4404 }
4405
4406 // Visit left or right.
4407 Value *NextV = TreeN->getOperand(EdgeToVist);
4408 if (NextV != Phi) {
4409 auto *I = dyn_cast<Instruction>(NextV);
4410 // Continue analysis if the next operand is a reduction operation or
4411 // (possibly) a reduced value. If the reduced value opcode is not set,
4412 // the first met operation != reduction operation is considered as the
4413 // reduced value class.
4414 if (I && (!ReducedValueOpcode || I->getOpcode() == ReducedValueOpcode ||
4415 I->getOpcode() == ReductionOpcode)) {
4416 // Only handle trees in the current basic block.
4417 if (I->getParent() != B->getParent()) {
4418 // I is an extra argument for TreeN (its parent operation).
4419 markExtraArg(Stack.back(), I);
4420 continue;
4421 }
4422
4423 // Each tree node needs to have one user except for the ultimate
4424 // reduction.
4425 if (!I->hasOneUse() && I != B) {
4426 // I is an extra argument for TreeN (its parent operation).
4427 markExtraArg(Stack.back(), I);
4428 continue;
4429 }
4430
4431 if (I->getOpcode() == ReductionOpcode) {
4432 // We need to be able to reassociate the reduction operations.
4433 if (!I->isAssociative()) {
4434 // I is an extra argument for TreeN (its parent operation).
4435 markExtraArg(Stack.back(), I);
4436 continue;
4437 }
4438 } else if (ReducedValueOpcode &&
4439 ReducedValueOpcode != I->getOpcode()) {
4440 // Make sure that the opcodes of the operations that we are going to
4441 // reduce match.
4442 // I is an extra argument for TreeN (its parent operation).
4443 markExtraArg(Stack.back(), I);
4444 continue;
4445 } else if (!ReducedValueOpcode)
4446 ReducedValueOpcode = I->getOpcode();
4447
4448 Stack.push_back(std::make_pair(I, 0));
4449 continue;
4450 }
4451 }
4452 // NextV is an extra argument for TreeN (its parent operation).
4453 markExtraArg(Stack.back(), NextV);
4454 }
4455 return true;
4456 }
4457
4458 /// \brief Attempt to vectorize the tree found by
4459 /// matchAssociativeReduction.
4460 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
4461 if (ReducedVals.empty())
4462 return false;
4463
4464 // If there is a sufficient number of reduction values, reduce
4465 // to a nearby power-of-2. Can safely generate oversized
4466 // vectors and rely on the backend to split them to legal sizes.
4467 unsigned NumReducedVals = ReducedVals.size();
4468 if (NumReducedVals < 4)
4469 return false;
4470
4471 unsigned ReduxWidth = PowerOf2Floor(NumReducedVals);
4472
4473 Value *VectorizedTree = nullptr;
4474 IRBuilder<> Builder(ReductionRoot);
4475 FastMathFlags Unsafe;
4476 Unsafe.setUnsafeAlgebra();
4477 Builder.setFastMathFlags(Unsafe);
4478 unsigned i = 0;
4479
4480 BoUpSLP::ExtraValueToDebugLocsMap ExternallyUsedValues;
4481 // The same extra argument may be used several time, so log each attempt
4482 // to use it.
4483 for (auto &Pair : ExtraArgs)
4484 ExternallyUsedValues[Pair.second].push_back(Pair.first);
4485 while (i < NumReducedVals - ReduxWidth + 1 && ReduxWidth > 2) {
4486 auto VL = makeArrayRef(&ReducedVals[i], ReduxWidth);
4487 V.buildTree(VL, ExternallyUsedValues, ReductionOps);
4488 if (V.shouldReorder()) {
4489 SmallVector<Value *, 8> Reversed(VL.rbegin(), VL.rend());
4490 V.buildTree(Reversed, ExternallyUsedValues, ReductionOps);
4491 }
4492 if (V.isTreeTinyAndNotFullyVectorizable())
4493 break;
4494
4495 V.computeMinimumValueSizes();
4496
4497 // Estimate cost.
4498 int Cost =
4499 V.getTreeCost() + getReductionCost(TTI, ReducedVals[i], ReduxWidth);
4500 if (Cost >= -SLPCostThreshold)
4501 break;
4502
4503 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)
4504 << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:"
<< Cost << ". (HorRdx)\n"; } } while (false)
;
4505
4506 // Vectorize a tree.
4507 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc();
4508 Value *VectorizedRoot = V.vectorizeTree(ExternallyUsedValues);
4509
4510 // Emit a reduction.
4511 Value *ReducedSubTree =
4512 emitReduction(VectorizedRoot, Builder, ReduxWidth, ReductionOps);
4513 if (VectorizedTree) {
4514 Builder.SetCurrentDebugLocation(Loc);
4515 VectorizedTree = Builder.CreateBinOp(ReductionOpcode, VectorizedTree,
4516 ReducedSubTree, "bin.rdx");
4517 propagateIRFlags(VectorizedTree, ReductionOps);
4518 } else
4519 VectorizedTree = ReducedSubTree;
4520 i += ReduxWidth;
4521 ReduxWidth = PowerOf2Floor(NumReducedVals - i);
4522 }
4523
4524 if (VectorizedTree) {
4525 // Finish the reduction.
4526 for (; i < NumReducedVals; ++i) {
4527 auto *I = cast<Instruction>(ReducedVals[i]);
4528 Builder.SetCurrentDebugLocation(I->getDebugLoc());
4529 VectorizedTree =
4530 Builder.CreateBinOp(ReductionOpcode, VectorizedTree, I);
4531 propagateIRFlags(VectorizedTree, ReductionOps);
4532 }
4533 for (auto &Pair : ExternallyUsedValues) {
4534 assert(!Pair.second.empty() &&((!Pair.second.empty() && "At least one DebugLoc must be inserted"
) ? static_cast<void> (0) : __assert_fail ("!Pair.second.empty() && \"At least one DebugLoc must be inserted\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4535, __PRETTY_FUNCTION__))
4535 "At least one DebugLoc must be inserted")((!Pair.second.empty() && "At least one DebugLoc must be inserted"
) ? static_cast<void> (0) : __assert_fail ("!Pair.second.empty() && \"At least one DebugLoc must be inserted\""
, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4535, __PRETTY_FUNCTION__))
;
4536 // Add each externally used value to the final reduction.
4537 for (auto *I : Pair.second) {
4538 Builder.SetCurrentDebugLocation(I->getDebugLoc());
4539 VectorizedTree = Builder.CreateBinOp(ReductionOpcode, VectorizedTree,
4540 Pair.first, "bin.extra");
4541 propagateIRFlags(VectorizedTree, I);
4542 }
4543 }
4544 // Update users.
4545 ReductionRoot->replaceAllUsesWith(VectorizedTree);
4546 }
4547 return VectorizedTree != nullptr;
4548 }
4549
4550 unsigned numReductionValues() const {
4551 return ReducedVals.size();
4552 }
4553
4554private:
4555 /// \brief Calculate the cost of a reduction.
4556 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal,
4557 unsigned ReduxWidth) {
4558 Type *ScalarTy = FirstReducedVal->getType();
4559 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth);
4560
4561 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true);
4562 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false);
4563
4564 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost;
4565 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost;
4566
4567 int ScalarReduxCost =
4568 (ReduxWidth - 1) *
4569 TTI->getArithmeticInstrCost(ReductionOpcode, ScalarTy);
4570
4571 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)
4572 << " 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)
4573 << " (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)
4574 << (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)
4575 << " 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)
;
4576
4577 return VecReduxCost - ScalarReduxCost;
4578 }
4579
4580 /// \brief Emit a horizontal reduction of the vectorized value.
4581 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder,
4582 unsigned ReduxWidth, ArrayRef<Value *> RedOps) {
4583 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4583, __PRETTY_FUNCTION__))
;
4584 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4585, __PRETTY_FUNCTION__))
4585 "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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4585, __PRETTY_FUNCTION__))
;
4586
4587 Value *TmpVec = VectorizedValue;
4588 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) {
4589 if (IsPairwiseReduction) {
4590 Value *LeftMask =
4591 createRdxShuffleMask(ReduxWidth, i, true, true, Builder);
4592 Value *RightMask =
4593 createRdxShuffleMask(ReduxWidth, i, true, false, Builder);
4594
4595 Value *LeftShuf = Builder.CreateShuffleVector(
4596 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l");
4597 Value *RightShuf = Builder.CreateShuffleVector(
4598 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask),
4599 "rdx.shuf.r");
4600 TmpVec = Builder.CreateBinOp(ReductionOpcode, LeftShuf, RightShuf,
4601 "bin.rdx");
4602 } else {
4603 Value *UpperHalf =
4604 createRdxShuffleMask(ReduxWidth, i, false, false, Builder);
4605 Value *Shuf = Builder.CreateShuffleVector(
4606 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf");
4607 TmpVec = Builder.CreateBinOp(ReductionOpcode, TmpVec, Shuf, "bin.rdx");
4608 }
4609 propagateIRFlags(TmpVec, RedOps);
4610 }
4611
4612 // The result is in the first element of the vector.
4613 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
4614 }
4615};
4616} // end anonymous namespace
4617
4618/// \brief Recognize construction of vectors like
4619/// %ra = insertelement <4 x float> undef, float %s0, i32 0
4620/// %rb = insertelement <4 x float> %ra, float %s1, i32 1
4621/// %rc = insertelement <4 x float> %rb, float %s2, i32 2
4622/// %rd = insertelement <4 x float> %rc, float %s3, i32 3
4623///
4624/// Returns true if it matches
4625///
4626static bool findBuildVector(InsertElementInst *FirstInsertElem,
4627 SmallVectorImpl<Value *> &BuildVector,
4628 SmallVectorImpl<Value *> &BuildVectorOpds) {
4629 if (!isa<UndefValue>(FirstInsertElem->getOperand(0)))
4630 return false;
4631
4632 InsertElementInst *IE = FirstInsertElem;
4633 while (true) {
4634 BuildVector.push_back(IE);
4635 BuildVectorOpds.push_back(IE->getOperand(1));
4636
4637 if (IE->use_empty())
4638 return false;
4639
4640 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->user_back());
4641 if (!NextUse)
4642 return true;
4643
4644 // If this isn't the final use, make sure the next insertelement is the only
4645 // use. It's OK if the final constructed vector is used multiple times
4646 if (!IE->hasOneUse())
4647 return false;
4648
4649 IE = NextUse;
4650 }
4651
4652 return false;
4653}
4654
4655/// \brief Like findBuildVector, but looks backwards for construction of aggregate.
4656///
4657/// \return true if it matches.
4658static bool findBuildAggregate(InsertValueInst *IV,
4659 SmallVectorImpl<Value *> &BuildVector,
4660 SmallVectorImpl<Value *> &BuildVectorOpds) {
4661 Value *V;
4662 do {
4663 BuildVector.push_back(IV);
4664 BuildVectorOpds.push_back(IV->getInsertedValueOperand());
4665 V = IV->getAggregateOperand();
4666 if (isa<UndefValue>(V))
4667 break;
4668 IV = dyn_cast<InsertValueInst>(V);
4669 if (!IV || !IV->hasOneUse())
4670 return false;
4671 } while (true);
4672 std::reverse(BuildVector.begin(), BuildVector.end());
4673 std::reverse(BuildVectorOpds.begin(), BuildVectorOpds.end());
4674 return true;
4675}
4676
4677static bool PhiTypeSorterFunc(Value *V, Value *V2) {
4678 return V->getType() < V2->getType();
4679}
4680
4681/// \brief Try and get a reduction value from a phi node.
4682///
4683/// Given a phi node \p P in a block \p ParentBB, consider possible reductions
4684/// if they come from either \p ParentBB or a containing loop latch.
4685///
4686/// \returns A candidate reduction value if possible, or \code nullptr \endcode
4687/// if not possible.
4688static Value *getReductionValue(const DominatorTree *DT, PHINode *P,
4689 BasicBlock *ParentBB, LoopInfo *LI) {
4690 // There are situations where the reduction value is not dominated by the
4691 // reduction phi. Vectorizing such cases has been reported to cause
4692 // miscompiles. See PR25787.
4693 auto DominatedReduxValue = [&](Value *R) {
4694 return (
4695 dyn_cast<Instruction>(R) &&
4696 DT->dominates(P->getParent(), dyn_cast<Instruction>(R)->getParent()));
4697 };
4698
4699 Value *Rdx = nullptr;
4700
4701 // Return the incoming value if it comes from the same BB as the phi node.
4702 if (P->getIncomingBlock(0) == ParentBB) {
4703 Rdx = P->getIncomingValue(0);
4704 } else if (P->getIncomingBlock(1) == ParentBB) {
4705 Rdx = P->getIncomingValue(1);
4706 }
4707
4708 if (Rdx && DominatedReduxValue(Rdx))
4709 return Rdx;
4710
4711 // Otherwise, check whether we have a loop latch to look at.
4712 Loop *BBL = LI->getLoopFor(ParentBB);
4713 if (!BBL)
4714 return nullptr;
4715 BasicBlock *BBLatch = BBL->getLoopLatch();
4716 if (!BBLatch)
4717 return nullptr;
4718
4719 // There is a loop latch, return the incoming value if it comes from
4720 // that. This reduction pattern occasionally turns up.
4721 if (P->getIncomingBlock(0) == BBLatch) {
4722 Rdx = P->getIncomingValue(0);
4723 } else if (P->getIncomingBlock(1) == BBLatch) {
4724 Rdx = P->getIncomingValue(1);
4725 }
4726
4727 if (Rdx && DominatedReduxValue(Rdx))
4728 return Rdx;
4729
4730 return nullptr;
4731}
4732
4733namespace {
4734/// Tracks instructons and its children.
4735class WeakVHWithLevel final : public CallbackVH {
4736 /// Operand index of the instruction currently beeing analized.
4737 unsigned Level = 0;
4738 /// Is this the instruction that should be vectorized, or are we now
4739 /// processing children (i.e. operands of this instruction) for potential
4740 /// vectorization?
4741 bool IsInitial = true;
4742
4743public:
4744 explicit WeakVHWithLevel() = default;
4745 WeakVHWithLevel(Value *V) : CallbackVH(V){};
4746 /// Restart children analysis each time it is repaced by the new instruction.
4747 void allUsesReplacedWith(Value *New) override {
4748 setValPtr(New);
4749 Level = 0;
4750 IsInitial = true;
4751 }
4752 /// Check if the instruction was not deleted during vectorization.
4753 bool isValid() const { return !getValPtr(); }
4754 /// Is the istruction itself must be vectorized?
4755 bool isInitial() const { return IsInitial; }
4756 /// Try to vectorize children.
4757 void clearInitial() { IsInitial = false; }
4758 /// Are all children processed already?
4759 bool isFinal() const {
4760 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4762, __PRETTY_FUNCTION__))
4761 (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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4762, __PRETTY_FUNCTION__))
4762 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4762, __PRETTY_FUNCTION__))
;
4763 return getValPtr() &&
4764 cast<Instruction>(getValPtr())->getNumOperands() == Level;
4765 }
4766 /// Get next child operation.
4767 Value *nextOperand() {
4768 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4769, __PRETTY_FUNCTION__))
4769 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 4769, __PRETTY_FUNCTION__))
;
4770 return cast<Instruction>(getValPtr())->getOperand(Level++);
4771 }
4772 virtual ~WeakVHWithLevel() = default;
4773};
4774} // namespace
4775
4776/// \brief Attempt to reduce a horizontal reduction.
4777/// If it is legal to match a horizontal reduction feeding
4778/// the phi node P with reduction operators Root in a basic block BB, then check
4779/// if it can be done.
4780/// \returns true if a horizontal reduction was matched and reduced.
4781/// \returns false if a horizontal reduction was not matched.
4782static bool canBeVectorized(
4783 PHINode *P, Instruction *Root, BasicBlock *BB, BoUpSLP &R,
4784 TargetTransformInfo *TTI,
4785 const function_ref<bool(BinaryOperator *, BoUpSLP &)> Vectorize) {
4786 if (!ShouldVectorizeHor)
4787 return false;
4788
4789 if (!Root)
4790 return false;
4791
4792 if (Root->getParent() != BB)
4793 return false;
4794 SmallVector<WeakVHWithLevel, 8> Stack(1, Root);
4795 SmallSet<Value *, 8> VisitedInstrs;
4796 bool Res = false;
4797 while (!Stack.empty()) {
4798 Value *V = Stack.back();
4799 if (!V) {
4800 Stack.pop_back();
4801 continue;
4802 }
4803 auto *Inst = dyn_cast<Instruction>(V);
4804 if (!Inst || isa<PHINode>(Inst)) {
4805 Stack.pop_back();
4806 continue;
4807 }
4808 if (Stack.back().isInitial()) {
4809 Stack.back().clearInitial();
4810 if (auto *BI = dyn_cast<BinaryOperator>(Inst)) {
4811 HorizontalReduction HorRdx;
4812 if (HorRdx.matchAssociativeReduction(P, BI)) {
4813 if (HorRdx.tryToReduce(R, TTI)) {
4814 Res = true;
4815 P = nullptr;
4816 continue;
4817 }
4818 }
4819 if (P) {
4820 Inst = dyn_cast<Instruction>(BI->getOperand(0));
4821 if (Inst == P)
4822 Inst = dyn_cast<Instruction>(BI->getOperand(1));
4823 if (!Inst) {
4824 P = nullptr;
4825 continue;
4826 }
4827 }
4828 }
4829 P = nullptr;
4830 if (Vectorize(dyn_cast<BinaryOperator>(Inst), R)) {
4831 Res = true;
4832 continue;
4833 }
4834 }
4835 if (Stack.back().isFinal()) {
4836 Stack.pop_back();
4837 continue;
4838 }
4839
4840 if (auto *NextV = dyn_cast<Instruction>(Stack.back().nextOperand()))
4841 if (NextV->getParent() == BB && VisitedInstrs.insert(NextV).second &&
4842 Stack.size() < RecursionMaxDepth)
4843 Stack.push_back(NextV);
4844 }
4845 return Res;
4846}
4847
4848bool SLPVectorizerPass::vectorizeRootInstruction(PHINode *P, Value *V,
4849 BasicBlock *BB, BoUpSLP &R,
4850 TargetTransformInfo *TTI) {
4851 if (!V)
4852 return false;
4853 auto *I = dyn_cast<Instruction>(V);
4854 if (!I)
4855 return false;
4856
4857 if (!isa<BinaryOperator>(I))
4858 P = nullptr;
4859 // Try to match and vectorize a horizontal reduction.
4860 return canBeVectorized(P, I, BB, R, TTI,
4861 [this](BinaryOperator *BI, BoUpSLP &R) -> bool {
4862 return tryToVectorize(BI, R);
4863 });
4864}
4865
4866bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
4867 bool Changed = false;
4868 SmallVector<Value *, 4> Incoming;
4869 SmallSet<Value *, 16> VisitedInstrs;
4870
4871 bool HaveVectorizedPhiNodes = true;
4872 while (HaveVectorizedPhiNodes) {
4873 HaveVectorizedPhiNodes = false;
4874
4875 // Collect the incoming values from the PHIs.
4876 Incoming.clear();
4877 for (Instruction &I : *BB) {
4878 PHINode *P = dyn_cast<PHINode>(&I);
4879 if (!P)
4880 break;
4881
4882 if (!VisitedInstrs.count(P))
4883 Incoming.push_back(P);
4884 }
4885
4886 // Sort by type.
4887 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc);
4888
4889 // Try to vectorize elements base on their type.
4890 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(),
4891 E = Incoming.end();
4892 IncIt != E;) {
4893
4894 // Look for the next elements with the same type.
4895 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt;
4896 while (SameTypeIt != E &&
4897 (*SameTypeIt)->getType() == (*IncIt)->getType()) {
4898 VisitedInstrs.insert(*SameTypeIt);
4899 ++SameTypeIt;
4900 }
4901
4902 // Try to vectorize them.
4903 unsigned NumElts = (SameTypeIt - IncIt);
4904 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)
;
4905 if (NumElts > 1 && tryToVectorizeList(makeArrayRef(IncIt, NumElts), R)) {
4906 // Success start over because instructions might have been changed.
4907 HaveVectorizedPhiNodes = true;
4908 Changed = true;
4909 break;
4910 }
4911
4912 // Start over at the next instruction of a different type (or the end).
4913 IncIt = SameTypeIt;
4914 }
4915 }
4916
4917 VisitedInstrs.clear();
4918
4919 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
4920 // We may go through BB multiple times so skip the one we have checked.
4921 if (!VisitedInstrs.insert(&*it).second)
4922 continue;
4923
4924 if (isa<DbgInfoIntrinsic>(it))
4925 continue;
4926
4927 // Try to vectorize reductions that use PHINodes.
4928 if (PHINode *P = dyn_cast<PHINode>(it)) {
4929 // Check that the PHI is a reduction PHI.
4930 if (P->getNumIncomingValues() != 2)
4931 return Changed;
4932
4933 // Try to match and vectorize a horizontal reduction.
4934 if (vectorizeRootInstruction(P, getReductionValue(DT, P, BB, LI), BB, R,
4935 TTI)) {
4936 Changed = true;
4937 it = BB->begin();
4938 e = BB->end();
4939 continue;
4940 }
4941 continue;
4942 }
4943
4944 if (ShouldStartVectorizeHorAtStore) {
4945 if (StoreInst *SI = dyn_cast<StoreInst>(it)) {
4946 // Try to match and vectorize a horizontal reduction.
4947 if (vectorizeRootInstruction(nullptr, SI->getValueOperand(), BB, R,
4948 TTI)) {
4949 Changed = true;
4950 it = BB->begin();
4951 e = BB->end();
4952 continue;
4953 }
4954 }
4955 }
4956
4957 // Try to vectorize horizontal reductions feeding into a return.
4958 if (ReturnInst *RI = dyn_cast<ReturnInst>(it)) {
4959 if (RI->getNumOperands() != 0) {
4960 // Try to match and vectorize a horizontal reduction.
4961 if (vectorizeRootInstruction(nullptr, RI->getOperand(0), BB, R, TTI)) {
4962 Changed = true;
4963 it = BB->begin();
4964 e = BB->end();
4965 continue;
4966 }
4967 }
4968 }
4969
4970 // Try to vectorize trees that start at compare instructions.
4971 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
4972 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
4973 Changed = true;
4974 // We would like to start over since some instructions are deleted
4975 // and the iterator may become invalid value.
4976 it = BB->begin();
4977 e = BB->end();
4978 continue;
4979 }
4980
4981 for (int I = 0; I < 2; ++I) {
4982 if (vectorizeRootInstruction(nullptr, CI->getOperand(I), BB, R, TTI)) {
4983 Changed = true;
4984 // We would like to start over since some instructions are deleted
4985 // and the iterator may become invalid value.
4986 it = BB->begin();
4987 e = BB->end();
4988 break;
4989 }
4990 }
4991 continue;
4992 }
4993
4994 // Try to vectorize trees that start at insertelement instructions.
4995 if (InsertElementInst *FirstInsertElem = dyn_cast<InsertElementInst>(it)) {
4996 SmallVector<Value *, 16> BuildVector;
4997 SmallVector<Value *, 16> BuildVectorOpds;
4998 if (!findBuildVector(FirstInsertElem, BuildVector, BuildVectorOpds))
4999 continue;
5000
5001 // Vectorize starting with the build vector operands ignoring the
5002 // BuildVector instructions for the purpose of scheduling and user
5003 // extraction.
5004 if (tryToVectorizeList(BuildVectorOpds, R, BuildVector)) {
5005 Changed = true;
5006 it = BB->begin();
5007 e = BB->end();
5008 }
5009
5010 continue;
5011 }
5012
5013 // Try to vectorize trees that start at insertvalue instructions feeding into
5014 // a store.
5015 if (StoreInst *SI = dyn_cast<StoreInst>(it)) {
5016 if (InsertValueInst *LastInsertValue = dyn_cast<InsertValueInst>(SI->getValueOperand())) {
5017 const DataLayout &DL = BB->getModule()->getDataLayout();
5018 if (R.canMapToVector(SI->getValueOperand()->getType(), DL)) {
5019 SmallVector<Value *, 16> BuildVector;
5020 SmallVector<Value *, 16> BuildVectorOpds;
5021 if (!findBuildAggregate(LastInsertValue, BuildVector, BuildVectorOpds))
5022 continue;
5023
5024 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)
;
5025 if (tryToVectorizeList(BuildVectorOpds, R, BuildVector, false)) {
5026 Changed = true;
5027 it = BB->begin();
5028 e = BB->end();
5029 }
5030 continue;
5031 }
5032 }
5033 }
5034 }
5035
5036 return Changed;
5037}
5038
5039bool SLPVectorizerPass::vectorizeGEPIndices(BasicBlock *BB, BoUpSLP &R) {
5040 auto Changed = false;
5041 for (auto &Entry : GEPs) {
5042
5043 // If the getelementptr list has fewer than two elements, there's nothing
5044 // to do.
5045 if (Entry.second.size() < 2)
5046 continue;
5047
5048 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
)
5049 << 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
)
;
5050
5051 // We process the getelementptr list in chunks of 16 (like we do for
5052 // stores) to minimize compile-time.
5053 for (unsigned BI = 0, BE = Entry.second.size(); BI < BE; BI += 16) {
5054 auto Len = std::min<unsigned>(BE - BI, 16);
5055 auto GEPList = makeArrayRef(&Entry.second[BI], Len);
5056
5057 // Initialize a set a candidate getelementptrs. Note that we use a
5058 // SetVector here to preserve program order. If the index computations
5059 // are vectorizable and begin with loads, we want to minimize the chance
5060 // of having to reorder them later.
5061 SetVector<Value *> Candidates(GEPList.begin(), GEPList.end());
5062
5063 // Some of the candidates may have already been vectorized after we
5064 // initially collected them. If so, the WeakVHs will have nullified the
5065 // values, so remove them from the set of candidates.
5066 Candidates.remove(nullptr);
5067
5068 // Remove from the set of candidates all pairs of getelementptrs with
5069 // constant differences. Such getelementptrs are likely not good
5070 // candidates for vectorization in a bottom-up phase since one can be
5071 // computed from the other. We also ensure all candidate getelementptr
5072 // indices are unique.
5073 for (int I = 0, E = GEPList.size(); I < E && Candidates.size() > 1; ++I) {
5074 auto *GEPI = cast<GetElementPtrInst>(GEPList[I]);
5075 if (!Candidates.count(GEPI))
5076 continue;
5077 auto *SCEVI = SE->getSCEV(GEPList[I]);
5078 for (int J = I + 1; J < E && Candidates.size() > 1; ++J) {
5079 auto *GEPJ = cast<GetElementPtrInst>(GEPList[J]);
5080 auto *SCEVJ = SE->getSCEV(GEPList[J]);
5081 if (isa<SCEVConstant>(SE->getMinusSCEV(SCEVI, SCEVJ))) {
5082 Candidates.remove(GEPList[I]);
5083 Candidates.remove(GEPList[J]);
5084 } else if (GEPI->idx_begin()->get() == GEPJ->idx_begin()->get()) {
5085 Candidates.remove(GEPList[J]);
5086 }
5087 }
5088 }
5089
5090 // We break out of the above computation as soon as we know there are
5091 // fewer than two candidates remaining.
5092 if (Candidates.size() < 2)
5093 continue;
5094
5095 // Add the single, non-constant index of each candidate to the bundle. We
5096 // ensured the indices met these constraints when we originally collected
5097 // the getelementptrs.
5098 SmallVector<Value *, 16> Bundle(Candidates.size());
5099 auto BundleIndex = 0u;
5100 for (auto *V : Candidates) {
5101 auto *GEP = cast<GetElementPtrInst>(V);
5102 auto *GEPIdx = GEP->idx_begin()->get();
5103 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~svn298304/lib/Transforms/Vectorize/SLPVectorizer.cpp"
, 5103, __PRETTY_FUNCTION__))
;
5104 Bundle[BundleIndex++] = GEPIdx;
5105 }
5106
5107 // Try and vectorize the indices. We are currently only interested in
5108 // gather-like cases of the form:
5109 //
5110 // ... = g[a[0] - b[0]] + g[a[1] - b[1]] + ...
5111 //
5112 // where the loads of "a", the loads of "b", and the subtractions can be
5113 // performed in parallel. It's likely that detecting this pattern in a
5114 // bottom-up phase will be simpler and less costly than building a
5115 // full-blown top-down phase beginning at the consecutive loads.
5116 Changed |= tryToVectorizeList(Bundle, R);
5117 }
5118 }
5119 return Changed;
5120}
5121
5122bool SLPVectorizerPass::vectorizeStoreChains(BoUpSLP &R) {
5123 bool Changed = false;
5124 // Attempt to sort and vectorize each of the store-groups.
5125 for (StoreListMap::iterator it = Stores.begin(), e = Stores.end(); it != e;
5126 ++it) {
5127 if (it->second.size() < 2)
5128 continue;
5129
5130 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
)
5131 << 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
)
;
5132
5133 // Process the stores in chunks of 16.
5134 // TODO: The limit of 16 inhibits greater vectorization factors.
5135 // For example, AVX2 supports v32i8. Increasing this limit, however,
5136 // may cause a significant compile-time increase.
5137 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
5138 unsigned Len = std::min<unsigned>(CE - CI, 16);
5139 Changed |= vectorizeStores(makeArrayRef(&it->second[CI], Len), R);
5140 }
5141 }
5142 return Changed;
5143}
5144
5145char SLPVectorizer::ID = 0;
5146static const char lv_name[] = "SLP Vectorizer";
5147INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)static void *initializeSLPVectorizerPassOnce(PassRegistry &
Registry) {
5148INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
5149INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry);
5150INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry);
5151INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry);
5152INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry);
5153INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry);
5154INITIALIZE_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)); }
5155
5156namespace llvm {
5157Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }
5158}