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
hasVectorInstrinsicOverloadedScalarOpd(Intrinsic::ID ID,unsigned ScalarOpdIdx)117 bool llvm::hasVectorInstrinsicOverloadedScalarOpd(Intrinsic::ID ID,
118 unsigned ScalarOpdIdx) {
119 switch (ID) {
120 case Intrinsic::powi:
121 return (ScalarOpdIdx == 1);
122 default:
123 return false;
124 }
125 }
126
127 /// Returns intrinsic ID for call.
128 /// For the input call instruction it finds mapping intrinsic and returns
129 /// its ID, in case it does not found it return not_intrinsic.
getVectorIntrinsicIDForCall(const CallInst * CI,const TargetLibraryInfo * TLI)130 Intrinsic::ID llvm::getVectorIntrinsicIDForCall(const CallInst *CI,
131 const TargetLibraryInfo *TLI) {
132 Intrinsic::ID ID = getIntrinsicForCallSite(*CI, TLI);
133 if (ID == Intrinsic::not_intrinsic)
134 return Intrinsic::not_intrinsic;
135
136 if (isTriviallyVectorizable(ID) || ID == Intrinsic::lifetime_start ||
137 ID == Intrinsic::lifetime_end || ID == Intrinsic::assume ||
138 ID == Intrinsic::experimental_noalias_scope_decl ||
139 ID == Intrinsic::sideeffect || ID == Intrinsic::pseudoprobe)
140 return ID;
141 return Intrinsic::not_intrinsic;
142 }
143
144 /// Find the operand of the GEP that should be checked for consecutive
145 /// stores. This ignores trailing indices that have no effect on the final
146 /// pointer.
getGEPInductionOperand(const GetElementPtrInst * Gep)147 unsigned llvm::getGEPInductionOperand(const GetElementPtrInst *Gep) {
148 const DataLayout &DL = Gep->getModule()->getDataLayout();
149 unsigned LastOperand = Gep->getNumOperands() - 1;
150 TypeSize GEPAllocSize = DL.getTypeAllocSize(Gep->getResultElementType());
151
152 // Walk backwards and try to peel off zeros.
153 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
154 // Find the type we're currently indexing into.
155 gep_type_iterator GEPTI = gep_type_begin(Gep);
156 std::advance(GEPTI, LastOperand - 2);
157
158 // If it's a type with the same allocation size as the result of the GEP we
159 // can peel off the zero index.
160 if (DL.getTypeAllocSize(GEPTI.getIndexedType()) != GEPAllocSize)
161 break;
162 --LastOperand;
163 }
164
165 return LastOperand;
166 }
167
168 /// If the argument is a GEP, then returns the operand identified by
169 /// getGEPInductionOperand. However, if there is some other non-loop-invariant
170 /// operand, it returns that instead.
stripGetElementPtr(Value * Ptr,ScalarEvolution * SE,Loop * Lp)171 Value *llvm::stripGetElementPtr(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
172 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
173 if (!GEP)
174 return Ptr;
175
176 unsigned InductionOperand = getGEPInductionOperand(GEP);
177
178 // Check that all of the gep indices are uniform except for our induction
179 // operand.
180 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
181 if (i != InductionOperand &&
182 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
183 return Ptr;
184 return GEP->getOperand(InductionOperand);
185 }
186
187 /// If a value has only one user that is a CastInst, return it.
getUniqueCastUse(Value * Ptr,Loop * Lp,Type * Ty)188 Value *llvm::getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
189 Value *UniqueCast = nullptr;
190 for (User *U : Ptr->users()) {
191 CastInst *CI = dyn_cast<CastInst>(U);
192 if (CI && CI->getType() == Ty) {
193 if (!UniqueCast)
194 UniqueCast = CI;
195 else
196 return nullptr;
197 }
198 }
199 return UniqueCast;
200 }
201
202 /// Get the stride of a pointer access in a loop. Looks for symbolic
203 /// strides "a[i*stride]". Returns the symbolic stride, or null otherwise.
getStrideFromPointer(Value * Ptr,ScalarEvolution * SE,Loop * Lp)204 Value *llvm::getStrideFromPointer(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
205 auto *PtrTy = dyn_cast<PointerType>(Ptr->getType());
206 if (!PtrTy || PtrTy->isAggregateType())
207 return nullptr;
208
209 // Try to remove a gep instruction to make the pointer (actually index at this
210 // point) easier analyzable. If OrigPtr is equal to Ptr we are analyzing the
211 // pointer, otherwise, we are analyzing the index.
212 Value *OrigPtr = Ptr;
213
214 // The size of the pointer access.
215 int64_t PtrAccessSize = 1;
216
217 Ptr = stripGetElementPtr(Ptr, SE, Lp);
218 const SCEV *V = SE->getSCEV(Ptr);
219
220 if (Ptr != OrigPtr)
221 // Strip off casts.
222 while (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(V))
223 V = C->getOperand();
224
225 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
226 if (!S)
227 return nullptr;
228
229 V = S->getStepRecurrence(*SE);
230 if (!V)
231 return nullptr;
232
233 // Strip off the size of access multiplication if we are still analyzing the
234 // pointer.
235 if (OrigPtr == Ptr) {
236 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
237 if (M->getOperand(0)->getSCEVType() != scConstant)
238 return nullptr;
239
240 const APInt &APStepVal = cast<SCEVConstant>(M->getOperand(0))->getAPInt();
241
242 // Huge step value - give up.
243 if (APStepVal.getBitWidth() > 64)
244 return nullptr;
245
246 int64_t StepVal = APStepVal.getSExtValue();
247 if (PtrAccessSize != StepVal)
248 return nullptr;
249 V = M->getOperand(1);
250 }
251 }
252
253 // Strip off casts.
254 Type *StripedOffRecurrenceCast = nullptr;
255 if (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(V)) {
256 StripedOffRecurrenceCast = C->getType();
257 V = C->getOperand();
258 }
259
260 // Look for the loop invariant symbolic value.
261 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
262 if (!U)
263 return nullptr;
264
265 Value *Stride = U->getValue();
266 if (!Lp->isLoopInvariant(Stride))
267 return nullptr;
268
269 // If we have stripped off the recurrence cast we have to make sure that we
270 // return the value that is used in this loop so that we can replace it later.
271 if (StripedOffRecurrenceCast)
272 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
273
274 return Stride;
275 }
276
277 /// Given a vector and an element number, see if the scalar value is
278 /// already around as a register, for example if it were inserted then extracted
279 /// from the vector.
findScalarElement(Value * V,unsigned EltNo)280 Value *llvm::findScalarElement(Value *V, unsigned EltNo) {
281 assert(V->getType()->isVectorTy() && "Not looking at a vector?");
282 VectorType *VTy = cast<VectorType>(V->getType());
283 // For fixed-length vector, return undef for out of range access.
284 if (auto *FVTy = dyn_cast<FixedVectorType>(VTy)) {
285 unsigned Width = FVTy->getNumElements();
286 if (EltNo >= Width)
287 return UndefValue::get(FVTy->getElementType());
288 }
289
290 if (Constant *C = dyn_cast<Constant>(V))
291 return C->getAggregateElement(EltNo);
292
293 if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
294 // If this is an insert to a variable element, we don't know what it is.
295 if (!isa<ConstantInt>(III->getOperand(2)))
296 return nullptr;
297 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
298
299 // If this is an insert to the element we are looking for, return the
300 // inserted value.
301 if (EltNo == IIElt)
302 return III->getOperand(1);
303
304 // Guard against infinite loop on malformed, unreachable IR.
305 if (III == III->getOperand(0))
306 return nullptr;
307
308 // Otherwise, the insertelement doesn't modify the value, recurse on its
309 // vector input.
310 return findScalarElement(III->getOperand(0), EltNo);
311 }
312
313 ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V);
314 // Restrict the following transformation to fixed-length vector.
315 if (SVI && isa<FixedVectorType>(SVI->getType())) {
316 unsigned LHSWidth =
317 cast<FixedVectorType>(SVI->getOperand(0)->getType())->getNumElements();
318 int InEl = SVI->getMaskValue(EltNo);
319 if (InEl < 0)
320 return UndefValue::get(VTy->getElementType());
321 if (InEl < (int)LHSWidth)
322 return findScalarElement(SVI->getOperand(0), InEl);
323 return findScalarElement(SVI->getOperand(1), InEl - LHSWidth);
324 }
325
326 // Extract a value from a vector add operation with a constant zero.
327 // TODO: Use getBinOpIdentity() to generalize this.
328 Value *Val; Constant *C;
329 if (match(V, m_Add(m_Value(Val), m_Constant(C))))
330 if (Constant *Elt = C->getAggregateElement(EltNo))
331 if (Elt->isNullValue())
332 return findScalarElement(Val, EltNo);
333
334 // Otherwise, we don't know.
335 return nullptr;
336 }
337
getSplatIndex(ArrayRef<int> Mask)338 int llvm::getSplatIndex(ArrayRef<int> Mask) {
339 int SplatIndex = -1;
340 for (int M : Mask) {
341 // Ignore invalid (undefined) mask elements.
342 if (M < 0)
343 continue;
344
345 // There can be only 1 non-negative mask element value if this is a splat.
346 if (SplatIndex != -1 && SplatIndex != M)
347 return -1;
348
349 // Initialize the splat index to the 1st non-negative mask element.
350 SplatIndex = M;
351 }
352 assert((SplatIndex == -1 || SplatIndex >= 0) && "Negative index?");
353 return SplatIndex;
354 }
355
356 /// Get splat value if the input is a splat vector or return nullptr.
357 /// This function is not fully general. It checks only 2 cases:
358 /// the input value is (1) a splat constant vector or (2) a sequence
359 /// of instructions that broadcasts a scalar at element 0.
getSplatValue(const Value * V)360 Value *llvm::getSplatValue(const Value *V) {
361 if (isa<VectorType>(V->getType()))
362 if (auto *C = dyn_cast<Constant>(V))
363 return C->getSplatValue();
364
365 // shuf (inselt ?, Splat, 0), ?, <0, undef, 0, ...>
366 Value *Splat;
367 if (match(V,
368 m_Shuffle(m_InsertElt(m_Value(), m_Value(Splat), m_ZeroInt()),
369 m_Value(), m_ZeroMask())))
370 return Splat;
371
372 return nullptr;
373 }
374
isSplatValue(const Value * V,int Index,unsigned Depth)375 bool llvm::isSplatValue(const Value *V, int Index, unsigned Depth) {
376 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth");
377
378 if (isa<VectorType>(V->getType())) {
379 if (isa<UndefValue>(V))
380 return true;
381 // FIXME: We can allow undefs, but if Index was specified, we may want to
382 // check that the constant is defined at that index.
383 if (auto *C = dyn_cast<Constant>(V))
384 return C->getSplatValue() != nullptr;
385 }
386
387 if (auto *Shuf = dyn_cast<ShuffleVectorInst>(V)) {
388 // FIXME: We can safely allow undefs here. If Index was specified, we will
389 // check that the mask elt is defined at the required index.
390 if (!is_splat(Shuf->getShuffleMask()))
391 return false;
392
393 // Match any index.
394 if (Index == -1)
395 return true;
396
397 // Match a specific element. The mask should be defined at and match the
398 // specified index.
399 return Shuf->getMaskValue(Index) == Index;
400 }
401
402 // The remaining tests are all recursive, so bail out if we hit the limit.
403 if (Depth++ == MaxAnalysisRecursionDepth)
404 return false;
405
406 // If both operands of a binop are splats, the result is a splat.
407 Value *X, *Y, *Z;
408 if (match(V, m_BinOp(m_Value(X), m_Value(Y))))
409 return isSplatValue(X, Index, Depth) && isSplatValue(Y, Index, Depth);
410
411 // If all operands of a select are splats, the result is a splat.
412 if (match(V, m_Select(m_Value(X), m_Value(Y), m_Value(Z))))
413 return isSplatValue(X, Index, Depth) && isSplatValue(Y, Index, Depth) &&
414 isSplatValue(Z, Index, Depth);
415
416 // TODO: Add support for unary ops (fneg), casts, intrinsics (overflow ops).
417
418 return false;
419 }
420
narrowShuffleMaskElts(int Scale,ArrayRef<int> Mask,SmallVectorImpl<int> & ScaledMask)421 void llvm::narrowShuffleMaskElts(int Scale, ArrayRef<int> Mask,
422 SmallVectorImpl<int> &ScaledMask) {
423 assert(Scale > 0 && "Unexpected scaling factor");
424
425 // Fast-path: if no scaling, then it is just a copy.
426 if (Scale == 1) {
427 ScaledMask.assign(Mask.begin(), Mask.end());
428 return;
429 }
430
431 ScaledMask.clear();
432 for (int MaskElt : Mask) {
433 if (MaskElt >= 0) {
434 assert(((uint64_t)Scale * MaskElt + (Scale - 1)) <= INT32_MAX &&
435 "Overflowed 32-bits");
436 }
437 for (int SliceElt = 0; SliceElt != Scale; ++SliceElt)
438 ScaledMask.push_back(MaskElt < 0 ? MaskElt : Scale * MaskElt + SliceElt);
439 }
440 }
441
widenShuffleMaskElts(int Scale,ArrayRef<int> Mask,SmallVectorImpl<int> & ScaledMask)442 bool llvm::widenShuffleMaskElts(int Scale, ArrayRef<int> Mask,
443 SmallVectorImpl<int> &ScaledMask) {
444 assert(Scale > 0 && "Unexpected scaling factor");
445
446 // Fast-path: if no scaling, then it is just a copy.
447 if (Scale == 1) {
448 ScaledMask.assign(Mask.begin(), Mask.end());
449 return true;
450 }
451
452 // We must map the original elements down evenly to a type with less elements.
453 int NumElts = Mask.size();
454 if (NumElts % Scale != 0)
455 return false;
456
457 ScaledMask.clear();
458 ScaledMask.reserve(NumElts / Scale);
459
460 // Step through the input mask by splitting into Scale-sized slices.
461 do {
462 ArrayRef<int> MaskSlice = Mask.take_front(Scale);
463 assert((int)MaskSlice.size() == Scale && "Expected Scale-sized slice.");
464
465 // The first element of the slice determines how we evaluate this slice.
466 int SliceFront = MaskSlice.front();
467 if (SliceFront < 0) {
468 // Negative values (undef or other "sentinel" values) must be equal across
469 // the entire slice.
470 if (!is_splat(MaskSlice))
471 return false;
472 ScaledMask.push_back(SliceFront);
473 } else {
474 // A positive mask element must be cleanly divisible.
475 if (SliceFront % Scale != 0)
476 return false;
477 // Elements of the slice must be consecutive.
478 for (int i = 1; i < Scale; ++i)
479 if (MaskSlice[i] != SliceFront + i)
480 return false;
481 ScaledMask.push_back(SliceFront / Scale);
482 }
483 Mask = Mask.drop_front(Scale);
484 } while (!Mask.empty());
485
486 assert((int)ScaledMask.size() * Scale == NumElts && "Unexpected scaled mask");
487
488 // All elements of the original mask can be scaled down to map to the elements
489 // of a mask with wider elements.
490 return true;
491 }
492
493 MapVector<Instruction *, uint64_t>
computeMinimumValueSizes(ArrayRef<BasicBlock * > Blocks,DemandedBits & DB,const TargetTransformInfo * TTI)494 llvm::computeMinimumValueSizes(ArrayRef<BasicBlock *> Blocks, DemandedBits &DB,
495 const TargetTransformInfo *TTI) {
496
497 // DemandedBits will give us every value's live-out bits. But we want
498 // to ensure no extra casts would need to be inserted, so every DAG
499 // of connected values must have the same minimum bitwidth.
500 EquivalenceClasses<Value *> ECs;
501 SmallVector<Value *, 16> Worklist;
502 SmallPtrSet<Value *, 4> Roots;
503 SmallPtrSet<Value *, 16> Visited;
504 DenseMap<Value *, uint64_t> DBits;
505 SmallPtrSet<Instruction *, 4> InstructionSet;
506 MapVector<Instruction *, uint64_t> MinBWs;
507
508 // Determine the roots. We work bottom-up, from truncs or icmps.
509 bool SeenExtFromIllegalType = false;
510 for (auto *BB : Blocks)
511 for (auto &I : *BB) {
512 InstructionSet.insert(&I);
513
514 if (TTI && (isa<ZExtInst>(&I) || isa<SExtInst>(&I)) &&
515 !TTI->isTypeLegal(I.getOperand(0)->getType()))
516 SeenExtFromIllegalType = true;
517
518 // Only deal with non-vector integers up to 64-bits wide.
519 if ((isa<TruncInst>(&I) || isa<ICmpInst>(&I)) &&
520 !I.getType()->isVectorTy() &&
521 I.getOperand(0)->getType()->getScalarSizeInBits() <= 64) {
522 // Don't make work for ourselves. If we know the loaded type is legal,
523 // don't add it to the worklist.
524 if (TTI && isa<TruncInst>(&I) && TTI->isTypeLegal(I.getType()))
525 continue;
526
527 Worklist.push_back(&I);
528 Roots.insert(&I);
529 }
530 }
531 // Early exit.
532 if (Worklist.empty() || (TTI && !SeenExtFromIllegalType))
533 return MinBWs;
534
535 // Now proceed breadth-first, unioning values together.
536 while (!Worklist.empty()) {
537 Value *Val = Worklist.pop_back_val();
538 Value *Leader = ECs.getOrInsertLeaderValue(Val);
539
540 if (Visited.count(Val))
541 continue;
542 Visited.insert(Val);
543
544 // Non-instructions terminate a chain successfully.
545 if (!isa<Instruction>(Val))
546 continue;
547 Instruction *I = cast<Instruction>(Val);
548
549 // If we encounter a type that is larger than 64 bits, we can't represent
550 // it so bail out.
551 if (DB.getDemandedBits(I).getBitWidth() > 64)
552 return MapVector<Instruction *, uint64_t>();
553
554 uint64_t V = DB.getDemandedBits(I).getZExtValue();
555 DBits[Leader] |= V;
556 DBits[I] = V;
557
558 // Casts, loads and instructions outside of our range terminate a chain
559 // successfully.
560 if (isa<SExtInst>(I) || isa<ZExtInst>(I) || isa<LoadInst>(I) ||
561 !InstructionSet.count(I))
562 continue;
563
564 // Unsafe casts terminate a chain unsuccessfully. We can't do anything
565 // useful with bitcasts, ptrtoints or inttoptrs and it'd be unsafe to
566 // transform anything that relies on them.
567 if (isa<BitCastInst>(I) || isa<PtrToIntInst>(I) || isa<IntToPtrInst>(I) ||
568 !I->getType()->isIntegerTy()) {
569 DBits[Leader] |= ~0ULL;
570 continue;
571 }
572
573 // We don't modify the types of PHIs. Reductions will already have been
574 // truncated if possible, and inductions' sizes will have been chosen by
575 // indvars.
576 if (isa<PHINode>(I))
577 continue;
578
579 if (DBits[Leader] == ~0ULL)
580 // All bits demanded, no point continuing.
581 continue;
582
583 for (Value *O : cast<User>(I)->operands()) {
584 ECs.unionSets(Leader, O);
585 Worklist.push_back(O);
586 }
587 }
588
589 // Now we've discovered all values, walk them to see if there are
590 // any users we didn't see. If there are, we can't optimize that
591 // chain.
592 for (auto &I : DBits)
593 for (auto *U : I.first->users())
594 if (U->getType()->isIntegerTy() && DBits.count(U) == 0)
595 DBits[ECs.getOrInsertLeaderValue(I.first)] |= ~0ULL;
596
597 for (auto I = ECs.begin(), E = ECs.end(); I != E; ++I) {
598 uint64_t LeaderDemandedBits = 0;
599 for (Value *M : llvm::make_range(ECs.member_begin(I), ECs.member_end()))
600 LeaderDemandedBits |= DBits[M];
601
602 uint64_t MinBW = (sizeof(LeaderDemandedBits) * 8) -
603 llvm::countLeadingZeros(LeaderDemandedBits);
604 // Round up to a power of 2
605 if (!isPowerOf2_64((uint64_t)MinBW))
606 MinBW = NextPowerOf2(MinBW);
607
608 // We don't modify the types of PHIs. Reductions will already have been
609 // truncated if possible, and inductions' sizes will have been chosen by
610 // indvars.
611 // If we are required to shrink a PHI, abandon this entire equivalence class.
612 bool Abort = false;
613 for (Value *M : llvm::make_range(ECs.member_begin(I), ECs.member_end()))
614 if (isa<PHINode>(M) && MinBW < M->getType()->getScalarSizeInBits()) {
615 Abort = true;
616 break;
617 }
618 if (Abort)
619 continue;
620
621 for (Value *M : llvm::make_range(ECs.member_begin(I), ECs.member_end())) {
622 if (!isa<Instruction>(M))
623 continue;
624 Type *Ty = M->getType();
625 if (Roots.count(M))
626 Ty = cast<Instruction>(M)->getOperand(0)->getType();
627 if (MinBW < Ty->getScalarSizeInBits())
628 MinBWs[cast<Instruction>(M)] = MinBW;
629 }
630 }
631
632 return MinBWs;
633 }
634
635 /// Add all access groups in @p AccGroups to @p List.
636 template <typename ListT>
addToAccessGroupList(ListT & List,MDNode * AccGroups)637 static void addToAccessGroupList(ListT &List, MDNode *AccGroups) {
638 // Interpret an access group as a list containing itself.
639 if (AccGroups->getNumOperands() == 0) {
640 assert(isValidAsAccessGroup(AccGroups) && "Node must be an access group");
641 List.insert(AccGroups);
642 return;
643 }
644
645 for (auto &AccGroupListOp : AccGroups->operands()) {
646 auto *Item = cast<MDNode>(AccGroupListOp.get());
647 assert(isValidAsAccessGroup(Item) && "List item must be an access group");
648 List.insert(Item);
649 }
650 }
651
uniteAccessGroups(MDNode * AccGroups1,MDNode * AccGroups2)652 MDNode *llvm::uniteAccessGroups(MDNode *AccGroups1, MDNode *AccGroups2) {
653 if (!AccGroups1)
654 return AccGroups2;
655 if (!AccGroups2)
656 return AccGroups1;
657 if (AccGroups1 == AccGroups2)
658 return AccGroups1;
659
660 SmallSetVector<Metadata *, 4> Union;
661 addToAccessGroupList(Union, AccGroups1);
662 addToAccessGroupList(Union, AccGroups2);
663
664 if (Union.size() == 0)
665 return nullptr;
666 if (Union.size() == 1)
667 return cast<MDNode>(Union.front());
668
669 LLVMContext &Ctx = AccGroups1->getContext();
670 return MDNode::get(Ctx, Union.getArrayRef());
671 }
672
intersectAccessGroups(const Instruction * Inst1,const Instruction * Inst2)673 MDNode *llvm::intersectAccessGroups(const Instruction *Inst1,
674 const Instruction *Inst2) {
675 bool MayAccessMem1 = Inst1->mayReadOrWriteMemory();
676 bool MayAccessMem2 = Inst2->mayReadOrWriteMemory();
677
678 if (!MayAccessMem1 && !MayAccessMem2)
679 return nullptr;
680 if (!MayAccessMem1)
681 return Inst2->getMetadata(LLVMContext::MD_access_group);
682 if (!MayAccessMem2)
683 return Inst1->getMetadata(LLVMContext::MD_access_group);
684
685 MDNode *MD1 = Inst1->getMetadata(LLVMContext::MD_access_group);
686 MDNode *MD2 = Inst2->getMetadata(LLVMContext::MD_access_group);
687 if (!MD1 || !MD2)
688 return nullptr;
689 if (MD1 == MD2)
690 return MD1;
691
692 // Use set for scalable 'contains' check.
693 SmallPtrSet<Metadata *, 4> AccGroupSet2;
694 addToAccessGroupList(AccGroupSet2, MD2);
695
696 SmallVector<Metadata *, 4> Intersection;
697 if (MD1->getNumOperands() == 0) {
698 assert(isValidAsAccessGroup(MD1) && "Node must be an access group");
699 if (AccGroupSet2.count(MD1))
700 Intersection.push_back(MD1);
701 } else {
702 for (const MDOperand &Node : MD1->operands()) {
703 auto *Item = cast<MDNode>(Node.get());
704 assert(isValidAsAccessGroup(Item) && "List item must be an access group");
705 if (AccGroupSet2.count(Item))
706 Intersection.push_back(Item);
707 }
708 }
709
710 if (Intersection.size() == 0)
711 return nullptr;
712 if (Intersection.size() == 1)
713 return cast<MDNode>(Intersection.front());
714
715 LLVMContext &Ctx = Inst1->getContext();
716 return MDNode::get(Ctx, Intersection);
717 }
718
719 /// \returns \p I after propagating metadata from \p VL.
propagateMetadata(Instruction * Inst,ArrayRef<Value * > VL)720 Instruction *llvm::propagateMetadata(Instruction *Inst, ArrayRef<Value *> VL) {
721 if (VL.empty())
722 return Inst;
723 Instruction *I0 = cast<Instruction>(VL[0]);
724 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
725 I0->getAllMetadataOtherThanDebugLoc(Metadata);
726
727 for (auto Kind : {LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
728 LLVMContext::MD_noalias, LLVMContext::MD_fpmath,
729 LLVMContext::MD_nontemporal, LLVMContext::MD_invariant_load,
730 LLVMContext::MD_access_group}) {
731 MDNode *MD = I0->getMetadata(Kind);
732
733 for (int J = 1, E = VL.size(); MD && J != E; ++J) {
734 const Instruction *IJ = cast<Instruction>(VL[J]);
735 MDNode *IMD = IJ->getMetadata(Kind);
736 switch (Kind) {
737 case LLVMContext::MD_tbaa:
738 MD = MDNode::getMostGenericTBAA(MD, IMD);
739 break;
740 case LLVMContext::MD_alias_scope:
741 MD = MDNode::getMostGenericAliasScope(MD, IMD);
742 break;
743 case LLVMContext::MD_fpmath:
744 MD = MDNode::getMostGenericFPMath(MD, IMD);
745 break;
746 case LLVMContext::MD_noalias:
747 case LLVMContext::MD_nontemporal:
748 case LLVMContext::MD_invariant_load:
749 MD = MDNode::intersect(MD, IMD);
750 break;
751 case LLVMContext::MD_access_group:
752 MD = intersectAccessGroups(Inst, IJ);
753 break;
754 default:
755 llvm_unreachable("unhandled metadata");
756 }
757 }
758
759 Inst->setMetadata(Kind, MD);
760 }
761
762 return Inst;
763 }
764
765 Constant *
createBitMaskForGaps(IRBuilderBase & Builder,unsigned VF,const InterleaveGroup<Instruction> & Group)766 llvm::createBitMaskForGaps(IRBuilderBase &Builder, unsigned VF,
767 const InterleaveGroup<Instruction> &Group) {
768 // All 1's means mask is not needed.
769 if (Group.getNumMembers() == Group.getFactor())
770 return nullptr;
771
772 // TODO: support reversed access.
773 assert(!Group.isReverse() && "Reversed group not supported.");
774
775 SmallVector<Constant *, 16> Mask;
776 for (unsigned i = 0; i < VF; i++)
777 for (unsigned j = 0; j < Group.getFactor(); ++j) {
778 unsigned HasMember = Group.getMember(j) ? 1 : 0;
779 Mask.push_back(Builder.getInt1(HasMember));
780 }
781
782 return ConstantVector::get(Mask);
783 }
784
785 llvm::SmallVector<int, 16>
createReplicatedMask(unsigned ReplicationFactor,unsigned VF)786 llvm::createReplicatedMask(unsigned ReplicationFactor, unsigned VF) {
787 SmallVector<int, 16> MaskVec;
788 for (unsigned i = 0; i < VF; i++)
789 for (unsigned j = 0; j < ReplicationFactor; j++)
790 MaskVec.push_back(i);
791
792 return MaskVec;
793 }
794
createInterleaveMask(unsigned VF,unsigned NumVecs)795 llvm::SmallVector<int, 16> llvm::createInterleaveMask(unsigned VF,
796 unsigned NumVecs) {
797 SmallVector<int, 16> Mask;
798 for (unsigned i = 0; i < VF; i++)
799 for (unsigned j = 0; j < NumVecs; j++)
800 Mask.push_back(j * VF + i);
801
802 return Mask;
803 }
804
805 llvm::SmallVector<int, 16>
createStrideMask(unsigned Start,unsigned Stride,unsigned VF)806 llvm::createStrideMask(unsigned Start, unsigned Stride, unsigned VF) {
807 SmallVector<int, 16> Mask;
808 for (unsigned i = 0; i < VF; i++)
809 Mask.push_back(Start + i * Stride);
810
811 return Mask;
812 }
813
createSequentialMask(unsigned Start,unsigned NumInts,unsigned NumUndefs)814 llvm::SmallVector<int, 16> llvm::createSequentialMask(unsigned Start,
815 unsigned NumInts,
816 unsigned NumUndefs) {
817 SmallVector<int, 16> Mask;
818 for (unsigned i = 0; i < NumInts; i++)
819 Mask.push_back(Start + i);
820
821 for (unsigned i = 0; i < NumUndefs; i++)
822 Mask.push_back(-1);
823
824 return Mask;
825 }
826
827 /// A helper function for concatenating vectors. This function concatenates two
828 /// vectors having the same element type. If the second vector has fewer
829 /// elements than the first, it is padded with undefs.
concatenateTwoVectors(IRBuilderBase & Builder,Value * V1,Value * V2)830 static Value *concatenateTwoVectors(IRBuilderBase &Builder, Value *V1,
831 Value *V2) {
832 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
833 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
834 assert(VecTy1 && VecTy2 &&
835 VecTy1->getScalarType() == VecTy2->getScalarType() &&
836 "Expect two vectors with the same element type");
837
838 unsigned NumElts1 = cast<FixedVectorType>(VecTy1)->getNumElements();
839 unsigned NumElts2 = cast<FixedVectorType>(VecTy2)->getNumElements();
840 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements");
841
842 if (NumElts1 > NumElts2) {
843 // Extend with UNDEFs.
844 V2 = Builder.CreateShuffleVector(
845 V2, createSequentialMask(0, NumElts2, NumElts1 - NumElts2));
846 }
847
848 return Builder.CreateShuffleVector(
849 V1, V2, createSequentialMask(0, NumElts1 + NumElts2, 0));
850 }
851
concatenateVectors(IRBuilderBase & Builder,ArrayRef<Value * > Vecs)852 Value *llvm::concatenateVectors(IRBuilderBase &Builder,
853 ArrayRef<Value *> Vecs) {
854 unsigned NumVecs = Vecs.size();
855 assert(NumVecs > 1 && "Should be at least two vectors");
856
857 SmallVector<Value *, 8> ResList;
858 ResList.append(Vecs.begin(), Vecs.end());
859 do {
860 SmallVector<Value *, 8> TmpList;
861 for (unsigned i = 0; i < NumVecs - 1; i += 2) {
862 Value *V0 = ResList[i], *V1 = ResList[i + 1];
863 assert((V0->getType() == V1->getType() || i == NumVecs - 2) &&
864 "Only the last vector may have a different type");
865
866 TmpList.push_back(concatenateTwoVectors(Builder, V0, V1));
867 }
868
869 // Push the last vector if the total number of vectors is odd.
870 if (NumVecs % 2 != 0)
871 TmpList.push_back(ResList[NumVecs - 1]);
872
873 ResList = TmpList;
874 NumVecs = ResList.size();
875 } while (NumVecs > 1);
876
877 return ResList[0];
878 }
879
maskIsAllZeroOrUndef(Value * Mask)880 bool llvm::maskIsAllZeroOrUndef(Value *Mask) {
881 assert(isa<VectorType>(Mask->getType()) &&
882 isa<IntegerType>(Mask->getType()->getScalarType()) &&
883 cast<IntegerType>(Mask->getType()->getScalarType())->getBitWidth() ==
884 1 &&
885 "Mask must be a vector of i1");
886
887 auto *ConstMask = dyn_cast<Constant>(Mask);
888 if (!ConstMask)
889 return false;
890 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
891 return true;
892 if (isa<ScalableVectorType>(ConstMask->getType()))
893 return false;
894 for (unsigned
895 I = 0,
896 E = cast<FixedVectorType>(ConstMask->getType())->getNumElements();
897 I != E; ++I) {
898 if (auto *MaskElt = ConstMask->getAggregateElement(I))
899 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
900 continue;
901 return false;
902 }
903 return true;
904 }
905
maskIsAllOneOrUndef(Value * Mask)906 bool llvm::maskIsAllOneOrUndef(Value *Mask) {
907 assert(isa<VectorType>(Mask->getType()) &&
908 isa<IntegerType>(Mask->getType()->getScalarType()) &&
909 cast<IntegerType>(Mask->getType()->getScalarType())->getBitWidth() ==
910 1 &&
911 "Mask must be a vector of i1");
912
913 auto *ConstMask = dyn_cast<Constant>(Mask);
914 if (!ConstMask)
915 return false;
916 if (ConstMask->isAllOnesValue() || isa<UndefValue>(ConstMask))
917 return true;
918 if (isa<ScalableVectorType>(ConstMask->getType()))
919 return false;
920 for (unsigned
921 I = 0,
922 E = cast<FixedVectorType>(ConstMask->getType())->getNumElements();
923 I != E; ++I) {
924 if (auto *MaskElt = ConstMask->getAggregateElement(I))
925 if (MaskElt->isAllOnesValue() || isa<UndefValue>(MaskElt))
926 continue;
927 return false;
928 }
929 return true;
930 }
931
932 /// TODO: This is a lot like known bits, but for
933 /// vectors. Is there something we can common this with?
possiblyDemandedEltsInMask(Value * Mask)934 APInt llvm::possiblyDemandedEltsInMask(Value *Mask) {
935 assert(isa<FixedVectorType>(Mask->getType()) &&
936 isa<IntegerType>(Mask->getType()->getScalarType()) &&
937 cast<IntegerType>(Mask->getType()->getScalarType())->getBitWidth() ==
938 1 &&
939 "Mask must be a fixed width vector of i1");
940
941 const unsigned VWidth =
942 cast<FixedVectorType>(Mask->getType())->getNumElements();
943 APInt DemandedElts = APInt::getAllOnesValue(VWidth);
944 if (auto *CV = dyn_cast<ConstantVector>(Mask))
945 for (unsigned i = 0; i < VWidth; i++)
946 if (CV->getAggregateElement(i)->isNullValue())
947 DemandedElts.clearBit(i);
948 return DemandedElts;
949 }
950
isStrided(int Stride)951 bool InterleavedAccessInfo::isStrided(int Stride) {
952 unsigned Factor = std::abs(Stride);
953 return Factor >= 2 && Factor <= MaxInterleaveGroupFactor;
954 }
955
collectConstStrideAccesses(MapVector<Instruction *,StrideDescriptor> & AccessStrideInfo,const ValueToValueMap & Strides)956 void InterleavedAccessInfo::collectConstStrideAccesses(
957 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
958 const ValueToValueMap &Strides) {
959 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
960
961 // Since it's desired that the load/store instructions be maintained in
962 // "program order" for the interleaved access analysis, we have to visit the
963 // blocks in the loop in reverse postorder (i.e., in a topological order).
964 // Such an ordering will ensure that any load/store that may be executed
965 // before a second load/store will precede the second load/store in
966 // AccessStrideInfo.
967 LoopBlocksDFS DFS(TheLoop);
968 DFS.perform(LI);
969 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
970 for (auto &I : *BB) {
971 Value *Ptr = getLoadStorePointerOperand(&I);
972 if (!Ptr)
973 continue;
974 Type *ElementTy = getLoadStoreType(&I);
975
976 // We don't check wrapping here because we don't know yet if Ptr will be
977 // part of a full group or a group with gaps. Checking wrapping for all
978 // pointers (even those that end up in groups with no gaps) will be overly
979 // conservative. For full groups, wrapping should be ok since if we would
980 // wrap around the address space we would do a memory access at nullptr
981 // even without the transformation. The wrapping checks are therefore
982 // deferred until after we've formed the interleaved groups.
983 int64_t Stride = getPtrStride(PSE, Ptr, TheLoop, Strides,
984 /*Assume=*/true, /*ShouldCheckWrap=*/false);
985
986 const SCEV *Scev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
987 uint64_t Size = DL.getTypeAllocSize(ElementTy);
988 AccessStrideInfo[&I] = StrideDescriptor(Stride, Scev, Size,
989 getLoadStoreAlignment(&I));
990 }
991 }
992
993 // Analyze interleaved accesses and collect them into interleaved load and
994 // store groups.
995 //
996 // When generating code for an interleaved load group, we effectively hoist all
997 // loads in the group to the location of the first load in program order. When
998 // generating code for an interleaved store group, we sink all stores to the
999 // location of the last store. This code motion can change the order of load
1000 // and store instructions and may break dependences.
1001 //
1002 // The code generation strategy mentioned above ensures that we won't violate
1003 // any write-after-read (WAR) dependences.
1004 //
1005 // E.g., for the WAR dependence: a = A[i]; // (1)
1006 // A[i] = b; // (2)
1007 //
1008 // The store group of (2) is always inserted at or below (2), and the load
1009 // group of (1) is always inserted at or above (1). Thus, the instructions will
1010 // never be reordered. All other dependences are checked to ensure the
1011 // correctness of the instruction reordering.
1012 //
1013 // The algorithm visits all memory accesses in the loop in bottom-up program
1014 // order. Program order is established by traversing the blocks in the loop in
1015 // reverse postorder when collecting the accesses.
1016 //
1017 // We visit the memory accesses in bottom-up order because it can simplify the
1018 // construction of store groups in the presence of write-after-write (WAW)
1019 // dependences.
1020 //
1021 // E.g., for the WAW dependence: A[i] = a; // (1)
1022 // A[i] = b; // (2)
1023 // A[i + 1] = c; // (3)
1024 //
1025 // We will first create a store group with (3) and (2). (1) can't be added to
1026 // this group because it and (2) are dependent. However, (1) can be grouped
1027 // with other accesses that may precede it in program order. Note that a
1028 // bottom-up order does not imply that WAW dependences should not be checked.
analyzeInterleaving(bool EnablePredicatedInterleavedMemAccesses)1029 void InterleavedAccessInfo::analyzeInterleaving(
1030 bool EnablePredicatedInterleavedMemAccesses) {
1031 LLVM_DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n");
1032 const ValueToValueMap &Strides = LAI->getSymbolicStrides();
1033
1034 // Holds all accesses with a constant stride.
1035 MapVector<Instruction *, StrideDescriptor> AccessStrideInfo;
1036 collectConstStrideAccesses(AccessStrideInfo, Strides);
1037
1038 if (AccessStrideInfo.empty())
1039 return;
1040
1041 // Collect the dependences in the loop.
1042 collectDependences();
1043
1044 // Holds all interleaved store groups temporarily.
1045 SmallSetVector<InterleaveGroup<Instruction> *, 4> StoreGroups;
1046 // Holds all interleaved load groups temporarily.
1047 SmallSetVector<InterleaveGroup<Instruction> *, 4> LoadGroups;
1048
1049 // Search in bottom-up program order for pairs of accesses (A and B) that can
1050 // form interleaved load or store groups. In the algorithm below, access A
1051 // precedes access B in program order. We initialize a group for B in the
1052 // outer loop of the algorithm, and then in the inner loop, we attempt to
1053 // insert each A into B's group if:
1054 //
1055 // 1. A and B have the same stride,
1056 // 2. A and B have the same memory object size, and
1057 // 3. A belongs in B's group according to its distance from B.
1058 //
1059 // Special care is taken to ensure group formation will not break any
1060 // dependences.
1061 for (auto BI = AccessStrideInfo.rbegin(), E = AccessStrideInfo.rend();
1062 BI != E; ++BI) {
1063 Instruction *B = BI->first;
1064 StrideDescriptor DesB = BI->second;
1065
1066 // Initialize a group for B if it has an allowable stride. Even if we don't
1067 // create a group for B, we continue with the bottom-up algorithm to ensure
1068 // we don't break any of B's dependences.
1069 InterleaveGroup<Instruction> *Group = nullptr;
1070 if (isStrided(DesB.Stride) &&
1071 (!isPredicated(B->getParent()) || EnablePredicatedInterleavedMemAccesses)) {
1072 Group = getInterleaveGroup(B);
1073 if (!Group) {
1074 LLVM_DEBUG(dbgs() << "LV: Creating an interleave group with:" << *B
1075 << '\n');
1076 Group = createInterleaveGroup(B, DesB.Stride, DesB.Alignment);
1077 }
1078 if (B->mayWriteToMemory())
1079 StoreGroups.insert(Group);
1080 else
1081 LoadGroups.insert(Group);
1082 }
1083
1084 for (auto AI = std::next(BI); AI != E; ++AI) {
1085 Instruction *A = AI->first;
1086 StrideDescriptor DesA = AI->second;
1087
1088 // Our code motion strategy implies that we can't have dependences
1089 // between accesses in an interleaved group and other accesses located
1090 // between the first and last member of the group. Note that this also
1091 // means that a group can't have more than one member at a given offset.
1092 // The accesses in a group can have dependences with other accesses, but
1093 // we must ensure we don't extend the boundaries of the group such that
1094 // we encompass those dependent accesses.
1095 //
1096 // For example, assume we have the sequence of accesses shown below in a
1097 // stride-2 loop:
1098 //
1099 // (1, 2) is a group | A[i] = a; // (1)
1100 // | A[i-1] = b; // (2) |
1101 // A[i-3] = c; // (3)
1102 // A[i] = d; // (4) | (2, 4) is not a group
1103 //
1104 // Because accesses (2) and (3) are dependent, we can group (2) with (1)
1105 // but not with (4). If we did, the dependent access (3) would be within
1106 // the boundaries of the (2, 4) group.
1107 if (!canReorderMemAccessesForInterleavedGroups(&*AI, &*BI)) {
1108 // If a dependence exists and A is already in a group, we know that A
1109 // must be a store since A precedes B and WAR dependences are allowed.
1110 // Thus, A would be sunk below B. We release A's group to prevent this
1111 // illegal code motion. A will then be free to form another group with
1112 // instructions that precede it.
1113 if (isInterleaved(A)) {
1114 InterleaveGroup<Instruction> *StoreGroup = getInterleaveGroup(A);
1115
1116 LLVM_DEBUG(dbgs() << "LV: Invalidated store group due to "
1117 "dependence between " << *A << " and "<< *B << '\n');
1118
1119 StoreGroups.remove(StoreGroup);
1120 releaseGroup(StoreGroup);
1121 }
1122
1123 // If a dependence exists and A is not already in a group (or it was
1124 // and we just released it), B might be hoisted above A (if B is a
1125 // load) or another store might be sunk below A (if B is a store). In
1126 // either case, we can't add additional instructions to B's group. B
1127 // will only form a group with instructions that it precedes.
1128 break;
1129 }
1130
1131 // At this point, we've checked for illegal code motion. If either A or B
1132 // isn't strided, there's nothing left to do.
1133 if (!isStrided(DesA.Stride) || !isStrided(DesB.Stride))
1134 continue;
1135
1136 // Ignore A if it's already in a group or isn't the same kind of memory
1137 // operation as B.
1138 // Note that mayReadFromMemory() isn't mutually exclusive to
1139 // mayWriteToMemory in the case of atomic loads. We shouldn't see those
1140 // here, canVectorizeMemory() should have returned false - except for the
1141 // case we asked for optimization remarks.
1142 if (isInterleaved(A) ||
1143 (A->mayReadFromMemory() != B->mayReadFromMemory()) ||
1144 (A->mayWriteToMemory() != B->mayWriteToMemory()))
1145 continue;
1146
1147 // Check rules 1 and 2. Ignore A if its stride or size is different from
1148 // that of B.
1149 if (DesA.Stride != DesB.Stride || DesA.Size != DesB.Size)
1150 continue;
1151
1152 // Ignore A if the memory object of A and B don't belong to the same
1153 // address space
1154 if (getLoadStoreAddressSpace(A) != getLoadStoreAddressSpace(B))
1155 continue;
1156
1157 // Calculate the distance from A to B.
1158 const SCEVConstant *DistToB = dyn_cast<SCEVConstant>(
1159 PSE.getSE()->getMinusSCEV(DesA.Scev, DesB.Scev));
1160 if (!DistToB)
1161 continue;
1162 int64_t DistanceToB = DistToB->getAPInt().getSExtValue();
1163
1164 // Check rule 3. Ignore A if its distance to B is not a multiple of the
1165 // size.
1166 if (DistanceToB % static_cast<int64_t>(DesB.Size))
1167 continue;
1168
1169 // All members of a predicated interleave-group must have the same predicate,
1170 // and currently must reside in the same BB.
1171 BasicBlock *BlockA = A->getParent();
1172 BasicBlock *BlockB = B->getParent();
1173 if ((isPredicated(BlockA) || isPredicated(BlockB)) &&
1174 (!EnablePredicatedInterleavedMemAccesses || BlockA != BlockB))
1175 continue;
1176
1177 // The index of A is the index of B plus A's distance to B in multiples
1178 // of the size.
1179 int IndexA =
1180 Group->getIndex(B) + DistanceToB / static_cast<int64_t>(DesB.Size);
1181
1182 // Try to insert A into B's group.
1183 if (Group->insertMember(A, IndexA, DesA.Alignment)) {
1184 LLVM_DEBUG(dbgs() << "LV: Inserted:" << *A << '\n'
1185 << " into the interleave group with" << *B
1186 << '\n');
1187 InterleaveGroupMap[A] = Group;
1188
1189 // Set the first load in program order as the insert position.
1190 if (A->mayReadFromMemory())
1191 Group->setInsertPos(A);
1192 }
1193 } // Iteration over A accesses.
1194 } // Iteration over B accesses.
1195
1196 // Remove interleaved store groups with gaps.
1197 for (auto *Group : StoreGroups)
1198 if (Group->getNumMembers() != Group->getFactor()) {
1199 LLVM_DEBUG(
1200 dbgs() << "LV: Invalidate candidate interleaved store group due "
1201 "to gaps.\n");
1202 releaseGroup(Group);
1203 }
1204 // Remove interleaved groups with gaps (currently only loads) whose memory
1205 // accesses may wrap around. We have to revisit the getPtrStride analysis,
1206 // this time with ShouldCheckWrap=true, since collectConstStrideAccesses does
1207 // not check wrapping (see documentation there).
1208 // FORNOW we use Assume=false;
1209 // TODO: Change to Assume=true but making sure we don't exceed the threshold
1210 // of runtime SCEV assumptions checks (thereby potentially failing to
1211 // vectorize altogether).
1212 // Additional optional optimizations:
1213 // TODO: If we are peeling the loop and we know that the first pointer doesn't
1214 // wrap then we can deduce that all pointers in the group don't wrap.
1215 // This means that we can forcefully peel the loop in order to only have to
1216 // check the first pointer for no-wrap. When we'll change to use Assume=true
1217 // we'll only need at most one runtime check per interleaved group.
1218 for (auto *Group : LoadGroups) {
1219 // Case 1: A full group. Can Skip the checks; For full groups, if the wide
1220 // load would wrap around the address space we would do a memory access at
1221 // nullptr even without the transformation.
1222 if (Group->getNumMembers() == Group->getFactor())
1223 continue;
1224
1225 // Case 2: If first and last members of the group don't wrap this implies
1226 // that all the pointers in the group don't wrap.
1227 // So we check only group member 0 (which is always guaranteed to exist),
1228 // and group member Factor - 1; If the latter doesn't exist we rely on
1229 // peeling (if it is a non-reversed accsess -- see Case 3).
1230 Value *FirstMemberPtr = getLoadStorePointerOperand(Group->getMember(0));
1231 if (!getPtrStride(PSE, FirstMemberPtr, TheLoop, Strides, /*Assume=*/false,
1232 /*ShouldCheckWrap=*/true)) {
1233 LLVM_DEBUG(
1234 dbgs() << "LV: Invalidate candidate interleaved group due to "
1235 "first group member potentially pointer-wrapping.\n");
1236 releaseGroup(Group);
1237 continue;
1238 }
1239 Instruction *LastMember = Group->getMember(Group->getFactor() - 1);
1240 if (LastMember) {
1241 Value *LastMemberPtr = getLoadStorePointerOperand(LastMember);
1242 if (!getPtrStride(PSE, LastMemberPtr, TheLoop, Strides, /*Assume=*/false,
1243 /*ShouldCheckWrap=*/true)) {
1244 LLVM_DEBUG(
1245 dbgs() << "LV: Invalidate candidate interleaved group due to "
1246 "last group member potentially pointer-wrapping.\n");
1247 releaseGroup(Group);
1248 }
1249 } else {
1250 // Case 3: A non-reversed interleaved load group with gaps: We need
1251 // to execute at least one scalar epilogue iteration. This will ensure
1252 // we don't speculatively access memory out-of-bounds. We only need
1253 // to look for a member at index factor - 1, since every group must have
1254 // a member at index zero.
1255 if (Group->isReverse()) {
1256 LLVM_DEBUG(
1257 dbgs() << "LV: Invalidate candidate interleaved group due to "
1258 "a reverse access with gaps.\n");
1259 releaseGroup(Group);
1260 continue;
1261 }
1262 LLVM_DEBUG(
1263 dbgs() << "LV: Interleaved group requires epilogue iteration.\n");
1264 RequiresScalarEpilogue = true;
1265 }
1266 }
1267 }
1268
invalidateGroupsRequiringScalarEpilogue()1269 void InterleavedAccessInfo::invalidateGroupsRequiringScalarEpilogue() {
1270 // If no group had triggered the requirement to create an epilogue loop,
1271 // there is nothing to do.
1272 if (!requiresScalarEpilogue())
1273 return;
1274
1275 bool ReleasedGroup = false;
1276 // Release groups requiring scalar epilogues. Note that this also removes them
1277 // from InterleaveGroups.
1278 for (auto *Group : make_early_inc_range(InterleaveGroups)) {
1279 if (!Group->requiresScalarEpilogue())
1280 continue;
1281 LLVM_DEBUG(
1282 dbgs()
1283 << "LV: Invalidate candidate interleaved group due to gaps that "
1284 "require a scalar epilogue (not allowed under optsize) and cannot "
1285 "be masked (not enabled). \n");
1286 releaseGroup(Group);
1287 ReleasedGroup = true;
1288 }
1289 assert(ReleasedGroup && "At least one group must be invalidated, as a "
1290 "scalar epilogue was required");
1291 (void)ReleasedGroup;
1292 RequiresScalarEpilogue = false;
1293 }
1294
1295 template <typename InstT>
addMetadata(InstT * NewInst) const1296 void InterleaveGroup<InstT>::addMetadata(InstT *NewInst) const {
1297 llvm_unreachable("addMetadata can only be used for Instruction");
1298 }
1299
1300 namespace llvm {
1301 template <>
addMetadata(Instruction * NewInst) const1302 void InterleaveGroup<Instruction>::addMetadata(Instruction *NewInst) const {
1303 SmallVector<Value *, 4> VL;
1304 std::transform(Members.begin(), Members.end(), std::back_inserter(VL),
1305 [](std::pair<int, Instruction *> p) { return p.second; });
1306 propagateMetadata(NewInst, VL);
1307 }
1308 }
1309
mangleTLIVectorName(StringRef VectorName,StringRef ScalarName,unsigned numArgs,ElementCount VF)1310 std::string VFABI::mangleTLIVectorName(StringRef VectorName,
1311 StringRef ScalarName, unsigned numArgs,
1312 ElementCount VF) {
1313 SmallString<256> Buffer;
1314 llvm::raw_svector_ostream Out(Buffer);
1315 Out << "_ZGV" << VFABI::_LLVM_ << "N";
1316 if (VF.isScalable())
1317 Out << 'x';
1318 else
1319 Out << VF.getFixedValue();
1320 for (unsigned I = 0; I < numArgs; ++I)
1321 Out << "v";
1322 Out << "_" << ScalarName << "(" << VectorName << ")";
1323 return std::string(Out.str());
1324 }
1325
getVectorVariantNames(const CallInst & CI,SmallVectorImpl<std::string> & VariantMappings)1326 void VFABI::getVectorVariantNames(
1327 const CallInst &CI, SmallVectorImpl<std::string> &VariantMappings) {
1328 const StringRef S =
1329 CI.getAttribute(AttributeList::FunctionIndex, VFABI::MappingsAttrName)
1330 .getValueAsString();
1331 if (S.empty())
1332 return;
1333
1334 SmallVector<StringRef, 8> ListAttr;
1335 S.split(ListAttr, ",");
1336
1337 for (auto &S : SetVector<StringRef>(ListAttr.begin(), ListAttr.end())) {
1338 #ifndef NDEBUG
1339 LLVM_DEBUG(dbgs() << "VFABI: adding mapping '" << S << "'\n");
1340 Optional<VFInfo> Info = VFABI::tryDemangleForVFABI(S, *(CI.getModule()));
1341 assert(Info.hasValue() && "Invalid name for a VFABI variant.");
1342 assert(CI.getModule()->getFunction(Info.getValue().VectorName) &&
1343 "Vector function is missing.");
1344 #endif
1345 VariantMappings.push_back(std::string(S));
1346 }
1347 }
1348
hasValidParameterList() const1349 bool VFShape::hasValidParameterList() const {
1350 for (unsigned Pos = 0, NumParams = Parameters.size(); Pos < NumParams;
1351 ++Pos) {
1352 assert(Parameters[Pos].ParamPos == Pos && "Broken parameter list.");
1353
1354 switch (Parameters[Pos].ParamKind) {
1355 default: // Nothing to check.
1356 break;
1357 case VFParamKind::OMP_Linear:
1358 case VFParamKind::OMP_LinearRef:
1359 case VFParamKind::OMP_LinearVal:
1360 case VFParamKind::OMP_LinearUVal:
1361 // Compile time linear steps must be non-zero.
1362 if (Parameters[Pos].LinearStepOrPos == 0)
1363 return false;
1364 break;
1365 case VFParamKind::OMP_LinearPos:
1366 case VFParamKind::OMP_LinearRefPos:
1367 case VFParamKind::OMP_LinearValPos:
1368 case VFParamKind::OMP_LinearUValPos:
1369 // The runtime linear step must be referring to some other
1370 // parameters in the signature.
1371 if (Parameters[Pos].LinearStepOrPos >= int(NumParams))
1372 return false;
1373 // The linear step parameter must be marked as uniform.
1374 if (Parameters[Parameters[Pos].LinearStepOrPos].ParamKind !=
1375 VFParamKind::OMP_Uniform)
1376 return false;
1377 // The linear step parameter can't point at itself.
1378 if (Parameters[Pos].LinearStepOrPos == int(Pos))
1379 return false;
1380 break;
1381 case VFParamKind::GlobalPredicate:
1382 // The global predicate must be the unique. Can be placed anywhere in the
1383 // signature.
1384 for (unsigned NextPos = Pos + 1; NextPos < NumParams; ++NextPos)
1385 if (Parameters[NextPos].ParamKind == VFParamKind::GlobalPredicate)
1386 return false;
1387 break;
1388 }
1389 }
1390 return true;
1391 }
1392