1 //===----------- VectorUtils.cpp - Vectorizer utility functions -----------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines vectorizer utilities.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "llvm/Analysis/VectorUtils.h"
14 #include "llvm/ADT/EquivalenceClasses.h"
15 #include "llvm/Analysis/DemandedBits.h"
16 #include "llvm/Analysis/LoopInfo.h"
17 #include "llvm/Analysis/LoopIterator.h"
18 #include "llvm/Analysis/ScalarEvolution.h"
19 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
20 #include "llvm/Analysis/TargetTransformInfo.h"
21 #include "llvm/Analysis/ValueTracking.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/GetElementPtrTypeIterator.h"
24 #include "llvm/IR/IRBuilder.h"
25 #include "llvm/IR/PatternMatch.h"
26 #include "llvm/IR/Value.h"
27 #include "llvm/Support/CommandLine.h"
28
29 #define DEBUG_TYPE "vectorutils"
30
31 using namespace llvm;
32 using namespace llvm::PatternMatch;
33
34 /// Maximum factor for an interleaved memory access.
35 static cl::opt<unsigned> MaxInterleaveGroupFactor(
36 "max-interleave-group-factor", cl::Hidden,
37 cl::desc("Maximum factor for an interleaved access group (default = 8)"),
38 cl::init(8));
39
40 /// Return true if all of the intrinsic's arguments and return type are scalars
41 /// for the scalar form of the intrinsic, and vectors for the vector form of the
42 /// intrinsic (except operands that are marked as always being scalar by
43 /// hasVectorInstrinsicScalarOpd).
isTriviallyVectorizable(Intrinsic::ID ID)44 bool llvm::isTriviallyVectorizable(Intrinsic::ID ID) {
45 switch (ID) {
46 case Intrinsic::abs: // Begin integer bit-manipulation.
47 case Intrinsic::bswap:
48 case Intrinsic::bitreverse:
49 case Intrinsic::ctpop:
50 case Intrinsic::ctlz:
51 case Intrinsic::cttz:
52 case Intrinsic::fshl:
53 case Intrinsic::fshr:
54 case Intrinsic::smax:
55 case Intrinsic::smin:
56 case Intrinsic::umax:
57 case Intrinsic::umin:
58 case Intrinsic::sadd_sat:
59 case Intrinsic::ssub_sat:
60 case Intrinsic::uadd_sat:
61 case Intrinsic::usub_sat:
62 case Intrinsic::smul_fix:
63 case Intrinsic::smul_fix_sat:
64 case Intrinsic::umul_fix:
65 case Intrinsic::umul_fix_sat:
66 case Intrinsic::sqrt: // Begin floating-point.
67 case Intrinsic::sin:
68 case Intrinsic::cos:
69 case Intrinsic::exp:
70 case Intrinsic::exp2:
71 case Intrinsic::log:
72 case Intrinsic::log10:
73 case Intrinsic::log2:
74 case Intrinsic::fabs:
75 case Intrinsic::minnum:
76 case Intrinsic::maxnum:
77 case Intrinsic::minimum:
78 case Intrinsic::maximum:
79 case Intrinsic::copysign:
80 case Intrinsic::floor:
81 case Intrinsic::ceil:
82 case Intrinsic::trunc:
83 case Intrinsic::rint:
84 case Intrinsic::nearbyint:
85 case Intrinsic::round:
86 case Intrinsic::roundeven:
87 case Intrinsic::pow:
88 case Intrinsic::fma:
89 case Intrinsic::fmuladd:
90 case Intrinsic::powi:
91 case Intrinsic::canonicalize:
92 return true;
93 default:
94 return false;
95 }
96 }
97
98 /// Identifies if the vector form of the intrinsic has a scalar operand.
hasVectorInstrinsicScalarOpd(Intrinsic::ID ID,unsigned ScalarOpdIdx)99 bool llvm::hasVectorInstrinsicScalarOpd(Intrinsic::ID ID,
100 unsigned ScalarOpdIdx) {
101 switch (ID) {
102 case Intrinsic::abs:
103 case Intrinsic::ctlz:
104 case Intrinsic::cttz:
105 case Intrinsic::powi:
106 return (ScalarOpdIdx == 1);
107 case Intrinsic::smul_fix:
108 case Intrinsic::smul_fix_sat:
109 case Intrinsic::umul_fix:
110 case Intrinsic::umul_fix_sat:
111 return (ScalarOpdIdx == 2);
112 default:
113 return false;
114 }
115 }
116
117 /// Returns intrinsic ID for call.
118 /// For the input call instruction it finds mapping intrinsic and returns
119 /// its ID, in case it does not found it return not_intrinsic.
getVectorIntrinsicIDForCall(const CallInst * CI,const TargetLibraryInfo * TLI)120 Intrinsic::ID llvm::getVectorIntrinsicIDForCall(const CallInst *CI,
121 const TargetLibraryInfo *TLI) {
122 Intrinsic::ID ID = getIntrinsicForCallSite(*CI, TLI);
123 if (ID == Intrinsic::not_intrinsic)
124 return Intrinsic::not_intrinsic;
125
126 if (isTriviallyVectorizable(ID) || ID == Intrinsic::lifetime_start ||
127 ID == Intrinsic::lifetime_end || ID == Intrinsic::assume ||
128 ID == Intrinsic::experimental_noalias_scope_decl ||
129 ID == Intrinsic::sideeffect || ID == Intrinsic::pseudoprobe)
130 return ID;
131 return Intrinsic::not_intrinsic;
132 }
133
134 /// Find the operand of the GEP that should be checked for consecutive
135 /// stores. This ignores trailing indices that have no effect on the final
136 /// pointer.
getGEPInductionOperand(const GetElementPtrInst * Gep)137 unsigned llvm::getGEPInductionOperand(const GetElementPtrInst *Gep) {
138 const DataLayout &DL = Gep->getModule()->getDataLayout();
139 unsigned LastOperand = Gep->getNumOperands() - 1;
140 TypeSize GEPAllocSize = DL.getTypeAllocSize(Gep->getResultElementType());
141
142 // Walk backwards and try to peel off zeros.
143 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
144 // Find the type we're currently indexing into.
145 gep_type_iterator GEPTI = gep_type_begin(Gep);
146 std::advance(GEPTI, LastOperand - 2);
147
148 // If it's a type with the same allocation size as the result of the GEP we
149 // can peel off the zero index.
150 if (DL.getTypeAllocSize(GEPTI.getIndexedType()) != GEPAllocSize)
151 break;
152 --LastOperand;
153 }
154
155 return LastOperand;
156 }
157
158 /// If the argument is a GEP, then returns the operand identified by
159 /// getGEPInductionOperand. However, if there is some other non-loop-invariant
160 /// operand, it returns that instead.
stripGetElementPtr(Value * Ptr,ScalarEvolution * SE,Loop * Lp)161 Value *llvm::stripGetElementPtr(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
162 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
163 if (!GEP)
164 return Ptr;
165
166 unsigned InductionOperand = getGEPInductionOperand(GEP);
167
168 // Check that all of the gep indices are uniform except for our induction
169 // operand.
170 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
171 if (i != InductionOperand &&
172 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
173 return Ptr;
174 return GEP->getOperand(InductionOperand);
175 }
176
177 /// If a value has only one user that is a CastInst, return it.
getUniqueCastUse(Value * Ptr,Loop * Lp,Type * Ty)178 Value *llvm::getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
179 Value *UniqueCast = nullptr;
180 for (User *U : Ptr->users()) {
181 CastInst *CI = dyn_cast<CastInst>(U);
182 if (CI && CI->getType() == Ty) {
183 if (!UniqueCast)
184 UniqueCast = CI;
185 else
186 return nullptr;
187 }
188 }
189 return UniqueCast;
190 }
191
192 /// Get the stride of a pointer access in a loop. Looks for symbolic
193 /// strides "a[i*stride]". Returns the symbolic stride, or null otherwise.
getStrideFromPointer(Value * Ptr,ScalarEvolution * SE,Loop * Lp)194 Value *llvm::getStrideFromPointer(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
195 auto *PtrTy = dyn_cast<PointerType>(Ptr->getType());
196 if (!PtrTy || PtrTy->isAggregateType())
197 return nullptr;
198
199 // Try to remove a gep instruction to make the pointer (actually index at this
200 // point) easier analyzable. If OrigPtr is equal to Ptr we are analyzing the
201 // pointer, otherwise, we are analyzing the index.
202 Value *OrigPtr = Ptr;
203
204 // The size of the pointer access.
205 int64_t PtrAccessSize = 1;
206
207 Ptr = stripGetElementPtr(Ptr, SE, Lp);
208 const SCEV *V = SE->getSCEV(Ptr);
209
210 if (Ptr != OrigPtr)
211 // Strip off casts.
212 while (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(V))
213 V = C->getOperand();
214
215 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
216 if (!S)
217 return nullptr;
218
219 V = S->getStepRecurrence(*SE);
220 if (!V)
221 return nullptr;
222
223 // Strip off the size of access multiplication if we are still analyzing the
224 // pointer.
225 if (OrigPtr == Ptr) {
226 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
227 if (M->getOperand(0)->getSCEVType() != scConstant)
228 return nullptr;
229
230 const APInt &APStepVal = cast<SCEVConstant>(M->getOperand(0))->getAPInt();
231
232 // Huge step value - give up.
233 if (APStepVal.getBitWidth() > 64)
234 return nullptr;
235
236 int64_t StepVal = APStepVal.getSExtValue();
237 if (PtrAccessSize != StepVal)
238 return nullptr;
239 V = M->getOperand(1);
240 }
241 }
242
243 // Strip off casts.
244 Type *StripedOffRecurrenceCast = nullptr;
245 if (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(V)) {
246 StripedOffRecurrenceCast = C->getType();
247 V = C->getOperand();
248 }
249
250 // Look for the loop invariant symbolic value.
251 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
252 if (!U)
253 return nullptr;
254
255 Value *Stride = U->getValue();
256 if (!Lp->isLoopInvariant(Stride))
257 return nullptr;
258
259 // If we have stripped off the recurrence cast we have to make sure that we
260 // return the value that is used in this loop so that we can replace it later.
261 if (StripedOffRecurrenceCast)
262 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
263
264 return Stride;
265 }
266
267 /// Given a vector and an element number, see if the scalar value is
268 /// already around as a register, for example if it were inserted then extracted
269 /// from the vector.
findScalarElement(Value * V,unsigned EltNo)270 Value *llvm::findScalarElement(Value *V, unsigned EltNo) {
271 assert(V->getType()->isVectorTy() && "Not looking at a vector?");
272 VectorType *VTy = cast<VectorType>(V->getType());
273 // For fixed-length vector, return undef for out of range access.
274 if (auto *FVTy = dyn_cast<FixedVectorType>(VTy)) {
275 unsigned Width = FVTy->getNumElements();
276 if (EltNo >= Width)
277 return UndefValue::get(FVTy->getElementType());
278 }
279
280 if (Constant *C = dyn_cast<Constant>(V))
281 return C->getAggregateElement(EltNo);
282
283 if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
284 // If this is an insert to a variable element, we don't know what it is.
285 if (!isa<ConstantInt>(III->getOperand(2)))
286 return nullptr;
287 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
288
289 // If this is an insert to the element we are looking for, return the
290 // inserted value.
291 if (EltNo == IIElt)
292 return III->getOperand(1);
293
294 // Guard against infinite loop on malformed, unreachable IR.
295 if (III == III->getOperand(0))
296 return nullptr;
297
298 // Otherwise, the insertelement doesn't modify the value, recurse on its
299 // vector input.
300 return findScalarElement(III->getOperand(0), EltNo);
301 }
302
303 ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V);
304 // Restrict the following transformation to fixed-length vector.
305 if (SVI && isa<FixedVectorType>(SVI->getType())) {
306 unsigned LHSWidth =
307 cast<FixedVectorType>(SVI->getOperand(0)->getType())->getNumElements();
308 int InEl = SVI->getMaskValue(EltNo);
309 if (InEl < 0)
310 return UndefValue::get(VTy->getElementType());
311 if (InEl < (int)LHSWidth)
312 return findScalarElement(SVI->getOperand(0), InEl);
313 return findScalarElement(SVI->getOperand(1), InEl - LHSWidth);
314 }
315
316 // Extract a value from a vector add operation with a constant zero.
317 // TODO: Use getBinOpIdentity() to generalize this.
318 Value *Val; Constant *C;
319 if (match(V, m_Add(m_Value(Val), m_Constant(C))))
320 if (Constant *Elt = C->getAggregateElement(EltNo))
321 if (Elt->isNullValue())
322 return findScalarElement(Val, EltNo);
323
324 // Otherwise, we don't know.
325 return nullptr;
326 }
327
getSplatIndex(ArrayRef<int> Mask)328 int llvm::getSplatIndex(ArrayRef<int> Mask) {
329 int SplatIndex = -1;
330 for (int M : Mask) {
331 // Ignore invalid (undefined) mask elements.
332 if (M < 0)
333 continue;
334
335 // There can be only 1 non-negative mask element value if this is a splat.
336 if (SplatIndex != -1 && SplatIndex != M)
337 return -1;
338
339 // Initialize the splat index to the 1st non-negative mask element.
340 SplatIndex = M;
341 }
342 assert((SplatIndex == -1 || SplatIndex >= 0) && "Negative index?");
343 return SplatIndex;
344 }
345
346 /// Get splat value if the input is a splat vector or return nullptr.
347 /// This function is not fully general. It checks only 2 cases:
348 /// the input value is (1) a splat constant vector or (2) a sequence
349 /// of instructions that broadcasts a scalar at element 0.
getSplatValue(const Value * V)350 Value *llvm::getSplatValue(const Value *V) {
351 if (isa<VectorType>(V->getType()))
352 if (auto *C = dyn_cast<Constant>(V))
353 return C->getSplatValue();
354
355 // shuf (inselt ?, Splat, 0), ?, <0, undef, 0, ...>
356 Value *Splat;
357 if (match(V,
358 m_Shuffle(m_InsertElt(m_Value(), m_Value(Splat), m_ZeroInt()),
359 m_Value(), m_ZeroMask())))
360 return Splat;
361
362 return nullptr;
363 }
364
isSplatValue(const Value * V,int Index,unsigned Depth)365 bool llvm::isSplatValue(const Value *V, int Index, unsigned Depth) {
366 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth");
367
368 if (isa<VectorType>(V->getType())) {
369 if (isa<UndefValue>(V))
370 return true;
371 // FIXME: We can allow undefs, but if Index was specified, we may want to
372 // check that the constant is defined at that index.
373 if (auto *C = dyn_cast<Constant>(V))
374 return C->getSplatValue() != nullptr;
375 }
376
377 if (auto *Shuf = dyn_cast<ShuffleVectorInst>(V)) {
378 // FIXME: We can safely allow undefs here. If Index was specified, we will
379 // check that the mask elt is defined at the required index.
380 if (!is_splat(Shuf->getShuffleMask()))
381 return false;
382
383 // Match any index.
384 if (Index == -1)
385 return true;
386
387 // Match a specific element. The mask should be defined at and match the
388 // specified index.
389 return Shuf->getMaskValue(Index) == Index;
390 }
391
392 // The remaining tests are all recursive, so bail out if we hit the limit.
393 if (Depth++ == MaxAnalysisRecursionDepth)
394 return false;
395
396 // If both operands of a binop are splats, the result is a splat.
397 Value *X, *Y, *Z;
398 if (match(V, m_BinOp(m_Value(X), m_Value(Y))))
399 return isSplatValue(X, Index, Depth) && isSplatValue(Y, Index, Depth);
400
401 // If all operands of a select are splats, the result is a splat.
402 if (match(V, m_Select(m_Value(X), m_Value(Y), m_Value(Z))))
403 return isSplatValue(X, Index, Depth) && isSplatValue(Y, Index, Depth) &&
404 isSplatValue(Z, Index, Depth);
405
406 // TODO: Add support for unary ops (fneg), casts, intrinsics (overflow ops).
407
408 return false;
409 }
410
narrowShuffleMaskElts(int Scale,ArrayRef<int> Mask,SmallVectorImpl<int> & ScaledMask)411 void llvm::narrowShuffleMaskElts(int Scale, ArrayRef<int> Mask,
412 SmallVectorImpl<int> &ScaledMask) {
413 assert(Scale > 0 && "Unexpected scaling factor");
414
415 // Fast-path: if no scaling, then it is just a copy.
416 if (Scale == 1) {
417 ScaledMask.assign(Mask.begin(), Mask.end());
418 return;
419 }
420
421 ScaledMask.clear();
422 for (int MaskElt : Mask) {
423 if (MaskElt >= 0) {
424 assert(((uint64_t)Scale * MaskElt + (Scale - 1)) <= INT32_MAX &&
425 "Overflowed 32-bits");
426 }
427 for (int SliceElt = 0; SliceElt != Scale; ++SliceElt)
428 ScaledMask.push_back(MaskElt < 0 ? MaskElt : Scale * MaskElt + SliceElt);
429 }
430 }
431
widenShuffleMaskElts(int Scale,ArrayRef<int> Mask,SmallVectorImpl<int> & ScaledMask)432 bool llvm::widenShuffleMaskElts(int Scale, ArrayRef<int> Mask,
433 SmallVectorImpl<int> &ScaledMask) {
434 assert(Scale > 0 && "Unexpected scaling factor");
435
436 // Fast-path: if no scaling, then it is just a copy.
437 if (Scale == 1) {
438 ScaledMask.assign(Mask.begin(), Mask.end());
439 return true;
440 }
441
442 // We must map the original elements down evenly to a type with less elements.
443 int NumElts = Mask.size();
444 if (NumElts % Scale != 0)
445 return false;
446
447 ScaledMask.clear();
448 ScaledMask.reserve(NumElts / Scale);
449
450 // Step through the input mask by splitting into Scale-sized slices.
451 do {
452 ArrayRef<int> MaskSlice = Mask.take_front(Scale);
453 assert((int)MaskSlice.size() == Scale && "Expected Scale-sized slice.");
454
455 // The first element of the slice determines how we evaluate this slice.
456 int SliceFront = MaskSlice.front();
457 if (SliceFront < 0) {
458 // Negative values (undef or other "sentinel" values) must be equal across
459 // the entire slice.
460 if (!is_splat(MaskSlice))
461 return false;
462 ScaledMask.push_back(SliceFront);
463 } else {
464 // A positive mask element must be cleanly divisible.
465 if (SliceFront % Scale != 0)
466 return false;
467 // Elements of the slice must be consecutive.
468 for (int i = 1; i < Scale; ++i)
469 if (MaskSlice[i] != SliceFront + i)
470 return false;
471 ScaledMask.push_back(SliceFront / Scale);
472 }
473 Mask = Mask.drop_front(Scale);
474 } while (!Mask.empty());
475
476 assert((int)ScaledMask.size() * Scale == NumElts && "Unexpected scaled mask");
477
478 // All elements of the original mask can be scaled down to map to the elements
479 // of a mask with wider elements.
480 return true;
481 }
482
483 MapVector<Instruction *, uint64_t>
computeMinimumValueSizes(ArrayRef<BasicBlock * > Blocks,DemandedBits & DB,const TargetTransformInfo * TTI)484 llvm::computeMinimumValueSizes(ArrayRef<BasicBlock *> Blocks, DemandedBits &DB,
485 const TargetTransformInfo *TTI) {
486
487 // DemandedBits will give us every value's live-out bits. But we want
488 // to ensure no extra casts would need to be inserted, so every DAG
489 // of connected values must have the same minimum bitwidth.
490 EquivalenceClasses<Value *> ECs;
491 SmallVector<Value *, 16> Worklist;
492 SmallPtrSet<Value *, 4> Roots;
493 SmallPtrSet<Value *, 16> Visited;
494 DenseMap<Value *, uint64_t> DBits;
495 SmallPtrSet<Instruction *, 4> InstructionSet;
496 MapVector<Instruction *, uint64_t> MinBWs;
497
498 // Determine the roots. We work bottom-up, from truncs or icmps.
499 bool SeenExtFromIllegalType = false;
500 for (auto *BB : Blocks)
501 for (auto &I : *BB) {
502 InstructionSet.insert(&I);
503
504 if (TTI && (isa<ZExtInst>(&I) || isa<SExtInst>(&I)) &&
505 !TTI->isTypeLegal(I.getOperand(0)->getType()))
506 SeenExtFromIllegalType = true;
507
508 // Only deal with non-vector integers up to 64-bits wide.
509 if ((isa<TruncInst>(&I) || isa<ICmpInst>(&I)) &&
510 !I.getType()->isVectorTy() &&
511 I.getOperand(0)->getType()->getScalarSizeInBits() <= 64) {
512 // Don't make work for ourselves. If we know the loaded type is legal,
513 // don't add it to the worklist.
514 if (TTI && isa<TruncInst>(&I) && TTI->isTypeLegal(I.getType()))
515 continue;
516
517 Worklist.push_back(&I);
518 Roots.insert(&I);
519 }
520 }
521 // Early exit.
522 if (Worklist.empty() || (TTI && !SeenExtFromIllegalType))
523 return MinBWs;
524
525 // Now proceed breadth-first, unioning values together.
526 while (!Worklist.empty()) {
527 Value *Val = Worklist.pop_back_val();
528 Value *Leader = ECs.getOrInsertLeaderValue(Val);
529
530 if (Visited.count(Val))
531 continue;
532 Visited.insert(Val);
533
534 // Non-instructions terminate a chain successfully.
535 if (!isa<Instruction>(Val))
536 continue;
537 Instruction *I = cast<Instruction>(Val);
538
539 // If we encounter a type that is larger than 64 bits, we can't represent
540 // it so bail out.
541 if (DB.getDemandedBits(I).getBitWidth() > 64)
542 return MapVector<Instruction *, uint64_t>();
543
544 uint64_t V = DB.getDemandedBits(I).getZExtValue();
545 DBits[Leader] |= V;
546 DBits[I] = V;
547
548 // Casts, loads and instructions outside of our range terminate a chain
549 // successfully.
550 if (isa<SExtInst>(I) || isa<ZExtInst>(I) || isa<LoadInst>(I) ||
551 !InstructionSet.count(I))
552 continue;
553
554 // Unsafe casts terminate a chain unsuccessfully. We can't do anything
555 // useful with bitcasts, ptrtoints or inttoptrs and it'd be unsafe to
556 // transform anything that relies on them.
557 if (isa<BitCastInst>(I) || isa<PtrToIntInst>(I) || isa<IntToPtrInst>(I) ||
558 !I->getType()->isIntegerTy()) {
559 DBits[Leader] |= ~0ULL;
560 continue;
561 }
562
563 // We don't modify the types of PHIs. Reductions will already have been
564 // truncated if possible, and inductions' sizes will have been chosen by
565 // indvars.
566 if (isa<PHINode>(I))
567 continue;
568
569 if (DBits[Leader] == ~0ULL)
570 // All bits demanded, no point continuing.
571 continue;
572
573 for (Value *O : cast<User>(I)->operands()) {
574 ECs.unionSets(Leader, O);
575 Worklist.push_back(O);
576 }
577 }
578
579 // Now we've discovered all values, walk them to see if there are
580 // any users we didn't see. If there are, we can't optimize that
581 // chain.
582 for (auto &I : DBits)
583 for (auto *U : I.first->users())
584 if (U->getType()->isIntegerTy() && DBits.count(U) == 0)
585 DBits[ECs.getOrInsertLeaderValue(I.first)] |= ~0ULL;
586
587 for (auto I = ECs.begin(), E = ECs.end(); I != E; ++I) {
588 uint64_t LeaderDemandedBits = 0;
589 for (Value *M : llvm::make_range(ECs.member_begin(I), ECs.member_end()))
590 LeaderDemandedBits |= DBits[M];
591
592 uint64_t MinBW = (sizeof(LeaderDemandedBits) * 8) -
593 llvm::countLeadingZeros(LeaderDemandedBits);
594 // Round up to a power of 2
595 if (!isPowerOf2_64((uint64_t)MinBW))
596 MinBW = NextPowerOf2(MinBW);
597
598 // We don't modify the types of PHIs. Reductions will already have been
599 // truncated if possible, and inductions' sizes will have been chosen by
600 // indvars.
601 // If we are required to shrink a PHI, abandon this entire equivalence class.
602 bool Abort = false;
603 for (Value *M : llvm::make_range(ECs.member_begin(I), ECs.member_end()))
604 if (isa<PHINode>(M) && MinBW < M->getType()->getScalarSizeInBits()) {
605 Abort = true;
606 break;
607 }
608 if (Abort)
609 continue;
610
611 for (Value *M : llvm::make_range(ECs.member_begin(I), ECs.member_end())) {
612 if (!isa<Instruction>(M))
613 continue;
614 Type *Ty = M->getType();
615 if (Roots.count(M))
616 Ty = cast<Instruction>(M)->getOperand(0)->getType();
617 if (MinBW < Ty->getScalarSizeInBits())
618 MinBWs[cast<Instruction>(M)] = MinBW;
619 }
620 }
621
622 return MinBWs;
623 }
624
625 /// Add all access groups in @p AccGroups to @p List.
626 template <typename ListT>
addToAccessGroupList(ListT & List,MDNode * AccGroups)627 static void addToAccessGroupList(ListT &List, MDNode *AccGroups) {
628 // Interpret an access group as a list containing itself.
629 if (AccGroups->getNumOperands() == 0) {
630 assert(isValidAsAccessGroup(AccGroups) && "Node must be an access group");
631 List.insert(AccGroups);
632 return;
633 }
634
635 for (auto &AccGroupListOp : AccGroups->operands()) {
636 auto *Item = cast<MDNode>(AccGroupListOp.get());
637 assert(isValidAsAccessGroup(Item) && "List item must be an access group");
638 List.insert(Item);
639 }
640 }
641
uniteAccessGroups(MDNode * AccGroups1,MDNode * AccGroups2)642 MDNode *llvm::uniteAccessGroups(MDNode *AccGroups1, MDNode *AccGroups2) {
643 if (!AccGroups1)
644 return AccGroups2;
645 if (!AccGroups2)
646 return AccGroups1;
647 if (AccGroups1 == AccGroups2)
648 return AccGroups1;
649
650 SmallSetVector<Metadata *, 4> Union;
651 addToAccessGroupList(Union, AccGroups1);
652 addToAccessGroupList(Union, AccGroups2);
653
654 if (Union.size() == 0)
655 return nullptr;
656 if (Union.size() == 1)
657 return cast<MDNode>(Union.front());
658
659 LLVMContext &Ctx = AccGroups1->getContext();
660 return MDNode::get(Ctx, Union.getArrayRef());
661 }
662
intersectAccessGroups(const Instruction * Inst1,const Instruction * Inst2)663 MDNode *llvm::intersectAccessGroups(const Instruction *Inst1,
664 const Instruction *Inst2) {
665 bool MayAccessMem1 = Inst1->mayReadOrWriteMemory();
666 bool MayAccessMem2 = Inst2->mayReadOrWriteMemory();
667
668 if (!MayAccessMem1 && !MayAccessMem2)
669 return nullptr;
670 if (!MayAccessMem1)
671 return Inst2->getMetadata(LLVMContext::MD_access_group);
672 if (!MayAccessMem2)
673 return Inst1->getMetadata(LLVMContext::MD_access_group);
674
675 MDNode *MD1 = Inst1->getMetadata(LLVMContext::MD_access_group);
676 MDNode *MD2 = Inst2->getMetadata(LLVMContext::MD_access_group);
677 if (!MD1 || !MD2)
678 return nullptr;
679 if (MD1 == MD2)
680 return MD1;
681
682 // Use set for scalable 'contains' check.
683 SmallPtrSet<Metadata *, 4> AccGroupSet2;
684 addToAccessGroupList(AccGroupSet2, MD2);
685
686 SmallVector<Metadata *, 4> Intersection;
687 if (MD1->getNumOperands() == 0) {
688 assert(isValidAsAccessGroup(MD1) && "Node must be an access group");
689 if (AccGroupSet2.count(MD1))
690 Intersection.push_back(MD1);
691 } else {
692 for (const MDOperand &Node : MD1->operands()) {
693 auto *Item = cast<MDNode>(Node.get());
694 assert(isValidAsAccessGroup(Item) && "List item must be an access group");
695 if (AccGroupSet2.count(Item))
696 Intersection.push_back(Item);
697 }
698 }
699
700 if (Intersection.size() == 0)
701 return nullptr;
702 if (Intersection.size() == 1)
703 return cast<MDNode>(Intersection.front());
704
705 LLVMContext &Ctx = Inst1->getContext();
706 return MDNode::get(Ctx, Intersection);
707 }
708
709 /// \returns \p I after propagating metadata from \p VL.
propagateMetadata(Instruction * Inst,ArrayRef<Value * > VL)710 Instruction *llvm::propagateMetadata(Instruction *Inst, ArrayRef<Value *> VL) {
711 if (VL.empty())
712 return Inst;
713 Instruction *I0 = cast<Instruction>(VL[0]);
714 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
715 I0->getAllMetadataOtherThanDebugLoc(Metadata);
716
717 for (auto Kind : {LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
718 LLVMContext::MD_noalias, LLVMContext::MD_fpmath,
719 LLVMContext::MD_nontemporal, LLVMContext::MD_invariant_load,
720 LLVMContext::MD_access_group}) {
721 MDNode *MD = I0->getMetadata(Kind);
722
723 for (int J = 1, E = VL.size(); MD && J != E; ++J) {
724 const Instruction *IJ = cast<Instruction>(VL[J]);
725 MDNode *IMD = IJ->getMetadata(Kind);
726 switch (Kind) {
727 case LLVMContext::MD_tbaa:
728 MD = MDNode::getMostGenericTBAA(MD, IMD);
729 break;
730 case LLVMContext::MD_alias_scope:
731 MD = MDNode::getMostGenericAliasScope(MD, IMD);
732 break;
733 case LLVMContext::MD_fpmath:
734 MD = MDNode::getMostGenericFPMath(MD, IMD);
735 break;
736 case LLVMContext::MD_noalias:
737 case LLVMContext::MD_nontemporal:
738 case LLVMContext::MD_invariant_load:
739 MD = MDNode::intersect(MD, IMD);
740 break;
741 case LLVMContext::MD_access_group:
742 MD = intersectAccessGroups(Inst, IJ);
743 break;
744 default:
745 llvm_unreachable("unhandled metadata");
746 }
747 }
748
749 Inst->setMetadata(Kind, MD);
750 }
751
752 return Inst;
753 }
754
755 Constant *
createBitMaskForGaps(IRBuilderBase & Builder,unsigned VF,const InterleaveGroup<Instruction> & Group)756 llvm::createBitMaskForGaps(IRBuilderBase &Builder, unsigned VF,
757 const InterleaveGroup<Instruction> &Group) {
758 // All 1's means mask is not needed.
759 if (Group.getNumMembers() == Group.getFactor())
760 return nullptr;
761
762 // TODO: support reversed access.
763 assert(!Group.isReverse() && "Reversed group not supported.");
764
765 SmallVector<Constant *, 16> Mask;
766 for (unsigned i = 0; i < VF; i++)
767 for (unsigned j = 0; j < Group.getFactor(); ++j) {
768 unsigned HasMember = Group.getMember(j) ? 1 : 0;
769 Mask.push_back(Builder.getInt1(HasMember));
770 }
771
772 return ConstantVector::get(Mask);
773 }
774
775 llvm::SmallVector<int, 16>
createReplicatedMask(unsigned ReplicationFactor,unsigned VF)776 llvm::createReplicatedMask(unsigned ReplicationFactor, unsigned VF) {
777 SmallVector<int, 16> MaskVec;
778 for (unsigned i = 0; i < VF; i++)
779 for (unsigned j = 0; j < ReplicationFactor; j++)
780 MaskVec.push_back(i);
781
782 return MaskVec;
783 }
784
createInterleaveMask(unsigned VF,unsigned NumVecs)785 llvm::SmallVector<int, 16> llvm::createInterleaveMask(unsigned VF,
786 unsigned NumVecs) {
787 SmallVector<int, 16> Mask;
788 for (unsigned i = 0; i < VF; i++)
789 for (unsigned j = 0; j < NumVecs; j++)
790 Mask.push_back(j * VF + i);
791
792 return Mask;
793 }
794
795 llvm::SmallVector<int, 16>
createStrideMask(unsigned Start,unsigned Stride,unsigned VF)796 llvm::createStrideMask(unsigned Start, unsigned Stride, unsigned VF) {
797 SmallVector<int, 16> Mask;
798 for (unsigned i = 0; i < VF; i++)
799 Mask.push_back(Start + i * Stride);
800
801 return Mask;
802 }
803
createSequentialMask(unsigned Start,unsigned NumInts,unsigned NumUndefs)804 llvm::SmallVector<int, 16> llvm::createSequentialMask(unsigned Start,
805 unsigned NumInts,
806 unsigned NumUndefs) {
807 SmallVector<int, 16> Mask;
808 for (unsigned i = 0; i < NumInts; i++)
809 Mask.push_back(Start + i);
810
811 for (unsigned i = 0; i < NumUndefs; i++)
812 Mask.push_back(-1);
813
814 return Mask;
815 }
816
817 /// A helper function for concatenating vectors. This function concatenates two
818 /// vectors having the same element type. If the second vector has fewer
819 /// elements than the first, it is padded with undefs.
concatenateTwoVectors(IRBuilderBase & Builder,Value * V1,Value * V2)820 static Value *concatenateTwoVectors(IRBuilderBase &Builder, Value *V1,
821 Value *V2) {
822 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
823 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
824 assert(VecTy1 && VecTy2 &&
825 VecTy1->getScalarType() == VecTy2->getScalarType() &&
826 "Expect two vectors with the same element type");
827
828 unsigned NumElts1 = cast<FixedVectorType>(VecTy1)->getNumElements();
829 unsigned NumElts2 = cast<FixedVectorType>(VecTy2)->getNumElements();
830 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
831
832 if (NumElts1 > NumElts2) {
833 // Extend with UNDEFs.
834 V2 = Builder.CreateShuffleVector(
835 V2, createSequentialMask(0, NumElts2, NumElts1 - NumElts2));
836 }
837
838 return Builder.CreateShuffleVector(
839 V1, V2, createSequentialMask(0, NumElts1 + NumElts2, 0));
840 }
841
concatenateVectors(IRBuilderBase & Builder,ArrayRef<Value * > Vecs)842 Value *llvm::concatenateVectors(IRBuilderBase &Builder,
843 ArrayRef<Value *> Vecs) {
844 unsigned NumVecs = Vecs.size();
845 assert(NumVecs > 1 && "Should be at least two vectors");
846
847 SmallVector<Value *, 8> ResList;
848 ResList.append(Vecs.begin(), Vecs.end());
849 do {
850 SmallVector<Value *, 8> TmpList;
851 for (unsigned i = 0; i < NumVecs - 1; i += 2) {
852 Value *V0 = ResList[i], *V1 = ResList[i + 1];
853 assert((V0->getType() == V1->getType() || i == NumVecs - 2) &&
854 "Only the last vector may have a different type");
855
856 TmpList.push_back(concatenateTwoVectors(Builder, V0, V1));
857 }
858
859 // Push the last vector if the total number of vectors is odd.
860 if (NumVecs % 2 != 0)
861 TmpList.push_back(ResList[NumVecs - 1]);
862
863 ResList = TmpList;
864 NumVecs = ResList.size();
865 } while (NumVecs > 1);
866
867 return ResList[0];
868 }
869
maskIsAllZeroOrUndef(Value * Mask)870 bool llvm::maskIsAllZeroOrUndef(Value *Mask) {
871 assert(isa<VectorType>(Mask->getType()) &&
872 isa<IntegerType>(Mask->getType()->getScalarType()) &&
873 cast<IntegerType>(Mask->getType()->getScalarType())->getBitWidth() ==
874 1 &&
875 "Mask must be a vector of i1");
876
877 auto *ConstMask = dyn_cast<Constant>(Mask);
878 if (!ConstMask)
879 return false;
880 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
881 return true;
882 if (isa<ScalableVectorType>(ConstMask->getType()))
883 return false;
884 for (unsigned
885 I = 0,
886 E = cast<FixedVectorType>(ConstMask->getType())->getNumElements();
887 I != E; ++I) {
888 if (auto *MaskElt = ConstMask->getAggregateElement(I))
889 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
890 continue;
891 return false;
892 }
893 return true;
894 }
895
896
maskIsAllOneOrUndef(Value * Mask)897 bool llvm::maskIsAllOneOrUndef(Value *Mask) {
898 assert(isa<VectorType>(Mask->getType()) &&
899 isa<IntegerType>(Mask->getType()->getScalarType()) &&
900 cast<IntegerType>(Mask->getType()->getScalarType())->getBitWidth() ==
901 1 &&
902 "Mask must be a vector of i1");
903
904 auto *ConstMask = dyn_cast<Constant>(Mask);
905 if (!ConstMask)
906 return false;
907 if (ConstMask->isAllOnesValue() || isa<UndefValue>(ConstMask))
908 return true;
909 if (isa<ScalableVectorType>(ConstMask->getType()))
910 return false;
911 for (unsigned
912 I = 0,
913 E = cast<FixedVectorType>(ConstMask->getType())->getNumElements();
914 I != E; ++I) {
915 if (auto *MaskElt = ConstMask->getAggregateElement(I))
916 if (MaskElt->isAllOnesValue() || isa<UndefValue>(MaskElt))
917 continue;
918 return false;
919 }
920 return true;
921 }
922
923 /// TODO: This is a lot like known bits, but for
924 /// vectors. Is there something we can common this with?
possiblyDemandedEltsInMask(Value * Mask)925 APInt llvm::possiblyDemandedEltsInMask(Value *Mask) {
926 assert(isa<FixedVectorType>(Mask->getType()) &&
927 isa<IntegerType>(Mask->getType()->getScalarType()) &&
928 cast<IntegerType>(Mask->getType()->getScalarType())->getBitWidth() ==
929 1 &&
930 "Mask must be a fixed width vector of i1");
931
932 const unsigned VWidth =
933 cast<FixedVectorType>(Mask->getType())->getNumElements();
934 APInt DemandedElts = APInt::getAllOnesValue(VWidth);
935 if (auto *CV = dyn_cast<ConstantVector>(Mask))
936 for (unsigned i = 0; i < VWidth; i++)
937 if (CV->getAggregateElement(i)->isNullValue())
938 DemandedElts.clearBit(i);
939 return DemandedElts;
940 }
941
isStrided(int Stride)942 bool InterleavedAccessInfo::isStrided(int Stride) {
943 unsigned Factor = std::abs(Stride);
944 return Factor >= 2 && Factor <= MaxInterleaveGroupFactor;
945 }
946
collectConstStrideAccesses(MapVector<Instruction *,StrideDescriptor> & AccessStrideInfo,const ValueToValueMap & Strides)947 void InterleavedAccessInfo::collectConstStrideAccesses(
948 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
949 const ValueToValueMap &Strides) {
950 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
951
952 // Since it's desired that the load/store instructions be maintained in
953 // "program order" for the interleaved access analysis, we have to visit the
954 // blocks in the loop in reverse postorder (i.e., in a topological order).
955 // Such an ordering will ensure that any load/store that may be executed
956 // before a second load/store will precede the second load/store in
957 // AccessStrideInfo.
958 LoopBlocksDFS DFS(TheLoop);
959 DFS.perform(LI);
960 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
961 for (auto &I : *BB) {
962 Value *Ptr = getLoadStorePointerOperand(&I);
963 if (!Ptr)
964 continue;
965 Type *ElementTy = getLoadStoreType(&I);
966
967 // We don't check wrapping here because we don't know yet if Ptr will be
968 // part of a full group or a group with gaps. Checking wrapping for all
969 // pointers (even those that end up in groups with no gaps) will be overly
970 // conservative. For full groups, wrapping should be ok since if we would
971 // wrap around the address space we would do a memory access at nullptr
972 // even without the transformation. The wrapping checks are therefore
973 // deferred until after we've formed the interleaved groups.
974 int64_t Stride = getPtrStride(PSE, Ptr, TheLoop, Strides,
975 /*Assume=*/true, /*ShouldCheckWrap=*/false);
976
977 const SCEV *Scev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
978 uint64_t Size = DL.getTypeAllocSize(ElementTy);
979 AccessStrideInfo[&I] = StrideDescriptor(Stride, Scev, Size,
980 getLoadStoreAlignment(&I));
981 }
982 }
983
984 // Analyze interleaved accesses and collect them into interleaved load and
985 // store groups.
986 //
987 // When generating code for an interleaved load group, we effectively hoist all
988 // loads in the group to the location of the first load in program order. When
989 // generating code for an interleaved store group, we sink all stores to the
990 // location of the last store. This code motion can change the order of load
991 // and store instructions and may break dependences.
992 //
993 // The code generation strategy mentioned above ensures that we won't violate
994 // any write-after-read (WAR) dependences.
995 //
996 // E.g., for the WAR dependence: a = A[i]; // (1)
997 // A[i] = b; // (2)
998 //
999 // The store group of (2) is always inserted at or below (2), and the load
1000 // group of (1) is always inserted at or above (1). Thus, the instructions will
1001 // never be reordered. All other dependences are checked to ensure the
1002 // correctness of the instruction reordering.
1003 //
1004 // The algorithm visits all memory accesses in the loop in bottom-up program
1005 // order. Program order is established by traversing the blocks in the loop in
1006 // reverse postorder when collecting the accesses.
1007 //
1008 // We visit the memory accesses in bottom-up order because it can simplify the
1009 // construction of store groups in the presence of write-after-write (WAW)
1010 // dependences.
1011 //
1012 // E.g., for the WAW dependence: A[i] = a; // (1)
1013 // A[i] = b; // (2)
1014 // A[i + 1] = c; // (3)
1015 //
1016 // We will first create a store group with (3) and (2). (1) can't be added to
1017 // this group because it and (2) are dependent. However, (1) can be grouped
1018 // with other accesses that may precede it in program order. Note that a
1019 // bottom-up order does not imply that WAW dependences should not be checked.
analyzeInterleaving(bool EnablePredicatedInterleavedMemAccesses)1020 void InterleavedAccessInfo::analyzeInterleaving(
1021 bool EnablePredicatedInterleavedMemAccesses) {
1022 LLVM_DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
1023 const ValueToValueMap &Strides = LAI->getSymbolicStrides();
1024
1025 // Holds all accesses with a constant stride.
1026 MapVector<Instruction *, StrideDescriptor> AccessStrideInfo;
1027 collectConstStrideAccesses(AccessStrideInfo, Strides);
1028
1029 if (AccessStrideInfo.empty())
1030 return;
1031
1032 // Collect the dependences in the loop.
1033 collectDependences();
1034
1035 // Holds all interleaved store groups temporarily.
1036 SmallSetVector<InterleaveGroup<Instruction> *, 4> StoreGroups;
1037 // Holds all interleaved load groups temporarily.
1038 SmallSetVector<InterleaveGroup<Instruction> *, 4> LoadGroups;
1039
1040 // Search in bottom-up program order for pairs of accesses (A and B) that can
1041 // form interleaved load or store groups. In the algorithm below, access A
1042 // precedes access B in program order. We initialize a group for B in the
1043 // outer loop of the algorithm, and then in the inner loop, we attempt to
1044 // insert each A into B's group if:
1045 //
1046 // 1. A and B have the same stride,
1047 // 2. A and B have the same memory object size, and
1048 // 3. A belongs in B's group according to its distance from B.
1049 //
1050 // Special care is taken to ensure group formation will not break any
1051 // dependences.
1052 for (auto BI = AccessStrideInfo.rbegin(), E = AccessStrideInfo.rend();
1053 BI != E; ++BI) {
1054 Instruction *B = BI->first;
1055 StrideDescriptor DesB = BI->second;
1056
1057 // Initialize a group for B if it has an allowable stride. Even if we don't
1058 // create a group for B, we continue with the bottom-up algorithm to ensure
1059 // we don't break any of B's dependences.
1060 InterleaveGroup<Instruction> *Group = nullptr;
1061 if (isStrided(DesB.Stride) &&
1062 (!isPredicated(B->getParent()) || EnablePredicatedInterleavedMemAccesses)) {
1063 Group = getInterleaveGroup(B);
1064 if (!Group) {
1065 LLVM_DEBUG(dbgs() << "LV: Creating an interleave group with:" << *B
1066 << '\n');
1067 Group = createInterleaveGroup(B, DesB.Stride, DesB.Alignment);
1068 }
1069 if (B->mayWriteToMemory())
1070 StoreGroups.insert(Group);
1071 else
1072 LoadGroups.insert(Group);
1073 }
1074
1075 for (auto AI = std::next(BI); AI != E; ++AI) {
1076 Instruction *A = AI->first;
1077 StrideDescriptor DesA = AI->second;
1078
1079 // Our code motion strategy implies that we can't have dependences
1080 // between accesses in an interleaved group and other accesses located
1081 // between the first and last member of the group. Note that this also
1082 // means that a group can't have more than one member at a given offset.
1083 // The accesses in a group can have dependences with other accesses, but
1084 // we must ensure we don't extend the boundaries of the group such that
1085 // we encompass those dependent accesses.
1086 //
1087 // For example, assume we have the sequence of accesses shown below in a
1088 // stride-2 loop:
1089 //
1090 // (1, 2) is a group | A[i] = a; // (1)
1091 // | A[i-1] = b; // (2) |
1092 // A[i-3] = c; // (3)
1093 // A[i] = d; // (4) | (2, 4) is not a group
1094 //
1095 // Because accesses (2) and (3) are dependent, we can group (2) with (1)
1096 // but not with (4). If we did, the dependent access (3) would be within
1097 // the boundaries of the (2, 4) group.
1098 if (!canReorderMemAccessesForInterleavedGroups(&*AI, &*BI)) {
1099 // If a dependence exists and A is already in a group, we know that A
1100 // must be a store since A precedes B and WAR dependences are allowed.
1101 // Thus, A would be sunk below B. We release A's group to prevent this
1102 // illegal code motion. A will then be free to form another group with
1103 // instructions that precede it.
1104 if (isInterleaved(A)) {
1105 InterleaveGroup<Instruction> *StoreGroup = getInterleaveGroup(A);
1106
1107 LLVM_DEBUG(dbgs() << "LV: Invalidated store group due to "
1108 "dependence between " << *A << " and "<< *B << '\n');
1109
1110 StoreGroups.remove(StoreGroup);
1111 releaseGroup(StoreGroup);
1112 }
1113
1114 // If a dependence exists and A is not already in a group (or it was
1115 // and we just released it), B might be hoisted above A (if B is a
1116 // load) or another store might be sunk below A (if B is a store). In
1117 // either case, we can't add additional instructions to B's group. B
1118 // will only form a group with instructions that it precedes.
1119 break;
1120 }
1121
1122 // At this point, we've checked for illegal code motion. If either A or B
1123 // isn't strided, there's nothing left to do.
1124 if (!isStrided(DesA.Stride) || !isStrided(DesB.Stride))
1125 continue;
1126
1127 // Ignore A if it's already in a group or isn't the same kind of memory
1128 // operation as B.
1129 // Note that mayReadFromMemory() isn't mutually exclusive to
1130 // mayWriteToMemory in the case of atomic loads. We shouldn't see those
1131 // here, canVectorizeMemory() should have returned false - except for the
1132 // case we asked for optimization remarks.
1133 if (isInterleaved(A) ||
1134 (A->mayReadFromMemory() != B->mayReadFromMemory()) ||
1135 (A->mayWriteToMemory() != B->mayWriteToMemory()))
1136 continue;
1137
1138 // Check rules 1 and 2. Ignore A if its stride or size is different from
1139 // that of B.
1140 if (DesA.Stride != DesB.Stride || DesA.Size != DesB.Size)
1141 continue;
1142
1143 // Ignore A if the memory object of A and B don't belong to the same
1144 // address space
1145 if (getLoadStoreAddressSpace(A) != getLoadStoreAddressSpace(B))
1146 continue;
1147
1148 // Calculate the distance from A to B.
1149 const SCEVConstant *DistToB = dyn_cast<SCEVConstant>(
1150 PSE.getSE()->getMinusSCEV(DesA.Scev, DesB.Scev));
1151 if (!DistToB)
1152 continue;
1153 int64_t DistanceToB = DistToB->getAPInt().getSExtValue();
1154
1155 // Check rule 3. Ignore A if its distance to B is not a multiple of the
1156 // size.
1157 if (DistanceToB % static_cast<int64_t>(DesB.Size))
1158 continue;
1159
1160 // All members of a predicated interleave-group must have the same predicate,
1161 // and currently must reside in the same BB.
1162 BasicBlock *BlockA = A->getParent();
1163 BasicBlock *BlockB = B->getParent();
1164 if ((isPredicated(BlockA) || isPredicated(BlockB)) &&
1165 (!EnablePredicatedInterleavedMemAccesses || BlockA != BlockB))
1166 continue;
1167
1168 // The index of A is the index of B plus A's distance to B in multiples
1169 // of the size.
1170 int IndexA =
1171 Group->getIndex(B) + DistanceToB / static_cast<int64_t>(DesB.Size);
1172
1173 // Try to insert A into B's group.
1174 if (Group->insertMember(A, IndexA, DesA.Alignment)) {
1175 LLVM_DEBUG(dbgs() << "LV: Inserted:" << *A << '\n'
1176 << " into the interleave group with" << *B
1177 << '\n');
1178 InterleaveGroupMap[A] = Group;
1179
1180 // Set the first load in program order as the insert position.
1181 if (A->mayReadFromMemory())
1182 Group->setInsertPos(A);
1183 }
1184 } // Iteration over A accesses.
1185 } // Iteration over B accesses.
1186
1187 // Remove interleaved store groups with gaps.
1188 for (auto *Group : StoreGroups)
1189 if (Group->getNumMembers() != Group->getFactor()) {
1190 LLVM_DEBUG(
1191 dbgs() << "LV: Invalidate candidate interleaved store group due "
1192 "to gaps.\n");
1193 releaseGroup(Group);
1194 }
1195 // Remove interleaved groups with gaps (currently only loads) whose memory
1196 // accesses may wrap around. We have to revisit the getPtrStride analysis,
1197 // this time with ShouldCheckWrap=true, since collectConstStrideAccesses does
1198 // not check wrapping (see documentation there).
1199 // FORNOW we use Assume=false;
1200 // TODO: Change to Assume=true but making sure we don't exceed the threshold
1201 // of runtime SCEV assumptions checks (thereby potentially failing to
1202 // vectorize altogether).
1203 // Additional optional optimizations:
1204 // TODO: If we are peeling the loop and we know that the first pointer doesn't
1205 // wrap then we can deduce that all pointers in the group don't wrap.
1206 // This means that we can forcefully peel the loop in order to only have to
1207 // check the first pointer for no-wrap. When we'll change to use Assume=true
1208 // we'll only need at most one runtime check per interleaved group.
1209 for (auto *Group : LoadGroups) {
1210 // Case 1: A full group. Can Skip the checks; For full groups, if the wide
1211 // load would wrap around the address space we would do a memory access at
1212 // nullptr even without the transformation.
1213 if (Group->getNumMembers() == Group->getFactor())
1214 continue;
1215
1216 // Case 2: If first and last members of the group don't wrap this implies
1217 // that all the pointers in the group don't wrap.
1218 // So we check only group member 0 (which is always guaranteed to exist),
1219 // and group member Factor - 1; If the latter doesn't exist we rely on
1220 // peeling (if it is a non-reversed accsess -- see Case 3).
1221 Value *FirstMemberPtr = getLoadStorePointerOperand(Group->getMember(0));
1222 if (!getPtrStride(PSE, FirstMemberPtr, TheLoop, Strides, /*Assume=*/false,
1223 /*ShouldCheckWrap=*/true)) {
1224 LLVM_DEBUG(
1225 dbgs() << "LV: Invalidate candidate interleaved group due to "
1226 "first group member potentially pointer-wrapping.\n");
1227 releaseGroup(Group);
1228 continue;
1229 }
1230 Instruction *LastMember = Group->getMember(Group->getFactor() - 1);
1231 if (LastMember) {
1232 Value *LastMemberPtr = getLoadStorePointerOperand(LastMember);
1233 if (!getPtrStride(PSE, LastMemberPtr, TheLoop, Strides, /*Assume=*/false,
1234 /*ShouldCheckWrap=*/true)) {
1235 LLVM_DEBUG(
1236 dbgs() << "LV: Invalidate candidate interleaved group due to "
1237 "last group member potentially pointer-wrapping.\n");
1238 releaseGroup(Group);
1239 }
1240 } else {
1241 // Case 3: A non-reversed interleaved load group with gaps: We need
1242 // to execute at least one scalar epilogue iteration. This will ensure
1243 // we don't speculatively access memory out-of-bounds. We only need
1244 // to look for a member at index factor - 1, since every group must have
1245 // a member at index zero.
1246 if (Group->isReverse()) {
1247 LLVM_DEBUG(
1248 dbgs() << "LV: Invalidate candidate interleaved group due to "
1249 "a reverse access with gaps.\n");
1250 releaseGroup(Group);
1251 continue;
1252 }
1253 LLVM_DEBUG(
1254 dbgs() << "LV: Interleaved group requires epilogue iteration.\n");
1255 RequiresScalarEpilogue = true;
1256 }
1257 }
1258 }
1259
invalidateGroupsRequiringScalarEpilogue()1260 void InterleavedAccessInfo::invalidateGroupsRequiringScalarEpilogue() {
1261 // If no group had triggered the requirement to create an epilogue loop,
1262 // there is nothing to do.
1263 if (!requiresScalarEpilogue())
1264 return;
1265
1266 bool ReleasedGroup = false;
1267 // Release groups requiring scalar epilogues. Note that this also removes them
1268 // from InterleaveGroups.
1269 for (auto *Group : make_early_inc_range(InterleaveGroups)) {
1270 if (!Group->requiresScalarEpilogue())
1271 continue;
1272 LLVM_DEBUG(
1273 dbgs()
1274 << "LV: Invalidate candidate interleaved group due to gaps that "
1275 "require a scalar epilogue (not allowed under optsize) and cannot "
1276 "be masked (not enabled). \n");
1277 releaseGroup(Group);
1278 ReleasedGroup = true;
1279 }
1280 assert(ReleasedGroup && "At least one group must be invalidated, as a "
1281 "scalar epilogue was required");
1282 (void)ReleasedGroup;
1283 RequiresScalarEpilogue = false;
1284 }
1285
1286 template <typename InstT>
addMetadata(InstT * NewInst) const1287 void InterleaveGroup<InstT>::addMetadata(InstT *NewInst) const {
1288 llvm_unreachable("addMetadata can only be used for Instruction");
1289 }
1290
1291 namespace llvm {
1292 template <>
addMetadata(Instruction * NewInst) const1293 void InterleaveGroup<Instruction>::addMetadata(Instruction *NewInst) const {
1294 SmallVector<Value *, 4> VL;
1295 std::transform(Members.begin(), Members.end(), std::back_inserter(VL),
1296 [](std::pair<int, Instruction *> p) { return p.second; });
1297 propagateMetadata(NewInst, VL);
1298 }
1299 }
1300
mangleTLIVectorName(StringRef VectorName,StringRef ScalarName,unsigned numArgs,ElementCount VF)1301 std::string VFABI::mangleTLIVectorName(StringRef VectorName,
1302 StringRef ScalarName, unsigned numArgs,
1303 ElementCount VF) {
1304 SmallString<256> Buffer;
1305 llvm::raw_svector_ostream Out(Buffer);
1306 Out << "_ZGV" << VFABI::_LLVM_ << "N";
1307 if (VF.isScalable())
1308 Out << 'x';
1309 else
1310 Out << VF.getFixedValue();
1311 for (unsigned I = 0; I < numArgs; ++I)
1312 Out << "v";
1313 Out << "_" << ScalarName << "(" << VectorName << ")";
1314 return std::string(Out.str());
1315 }
1316
getVectorVariantNames(const CallInst & CI,SmallVectorImpl<std::string> & VariantMappings)1317 void VFABI::getVectorVariantNames(
1318 const CallInst &CI, SmallVectorImpl<std::string> &VariantMappings) {
1319 const StringRef S =
1320 CI.getAttribute(AttributeList::FunctionIndex, VFABI::MappingsAttrName)
1321 .getValueAsString();
1322 if (S.empty())
1323 return;
1324
1325 SmallVector<StringRef, 8> ListAttr;
1326 S.split(ListAttr, ",");
1327
1328 for (auto &S : SetVector<StringRef>(ListAttr.begin(), ListAttr.end())) {
1329 #ifndef NDEBUG
1330 LLVM_DEBUG(dbgs() << "VFABI: adding mapping '" << S << "'\n");
1331 Optional<VFInfo> Info = VFABI::tryDemangleForVFABI(S, *(CI.getModule()));
1332 assert(Info.hasValue() && "Invalid name for a VFABI variant.");
1333 assert(CI.getModule()->getFunction(Info.getValue().VectorName) &&
1334 "Vector function is missing.");
1335 #endif
1336 VariantMappings.push_back(std::string(S));
1337 }
1338 }
1339
hasValidParameterList() const1340 bool VFShape::hasValidParameterList() const {
1341 for (unsigned Pos = 0, NumParams = Parameters.size(); Pos < NumParams;
1342 ++Pos) {
1343 assert(Parameters[Pos].ParamPos == Pos && "Broken parameter list.");
1344
1345 switch (Parameters[Pos].ParamKind) {
1346 default: // Nothing to check.
1347 break;
1348 case VFParamKind::OMP_Linear:
1349 case VFParamKind::OMP_LinearRef:
1350 case VFParamKind::OMP_LinearVal:
1351 case VFParamKind::OMP_LinearUVal:
1352 // Compile time linear steps must be non-zero.
1353 if (Parameters[Pos].LinearStepOrPos == 0)
1354 return false;
1355 break;
1356 case VFParamKind::OMP_LinearPos:
1357 case VFParamKind::OMP_LinearRefPos:
1358 case VFParamKind::OMP_LinearValPos:
1359 case VFParamKind::OMP_LinearUValPos:
1360 // The runtime linear step must be referring to some other
1361 // parameters in the signature.
1362 if (Parameters[Pos].LinearStepOrPos >= int(NumParams))
1363 return false;
1364 // The linear step parameter must be marked as uniform.
1365 if (Parameters[Parameters[Pos].LinearStepOrPos].ParamKind !=
1366 VFParamKind::OMP_Uniform)
1367 return false;
1368 // The linear step parameter can't point at itself.
1369 if (Parameters[Pos].LinearStepOrPos == int(Pos))
1370 return false;
1371 break;
1372 case VFParamKind::GlobalPredicate:
1373 // The global predicate must be the unique. Can be placed anywhere in the
1374 // signature.
1375 for (unsigned NextPos = Pos + 1; NextPos < NumParams; ++NextPos)
1376 if (Parameters[NextPos].ParamKind == VFParamKind::GlobalPredicate)
1377 return false;
1378 break;
1379 }
1380 }
1381 return true;
1382 }
1383