Bug Summary

File:lib/Analysis/VectorUtils.cpp
Warning:line 942, column 11
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name VectorUtils.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mthread-model posix -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-8/lib/clang/8.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/lib/Analysis -I /build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis -I /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/include -I /build/llvm-toolchain-snapshot-8~svn350071/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/include/clang/8.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-8/lib/clang/8.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/lib/Analysis -fdebug-prefix-map=/build/llvm-toolchain-snapshot-8~svn350071=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2018-12-27-042839-1215-1 -x c++ /build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp -faddrsig
1//===----------- VectorUtils.cpp - Vectorizer utility functions -----------===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file defines vectorizer utilities.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/Analysis/VectorUtils.h"
15#include "llvm/ADT/EquivalenceClasses.h"
16#include "llvm/Analysis/DemandedBits.h"
17#include "llvm/Analysis/LoopInfo.h"
18#include "llvm/Analysis/LoopIterator.h"
19#include "llvm/Analysis/ScalarEvolution.h"
20#include "llvm/Analysis/ScalarEvolutionExpressions.h"
21#include "llvm/Analysis/TargetTransformInfo.h"
22#include "llvm/Analysis/ValueTracking.h"
23#include "llvm/IR/Constants.h"
24#include "llvm/IR/GetElementPtrTypeIterator.h"
25#include "llvm/IR/IRBuilder.h"
26#include "llvm/IR/PatternMatch.h"
27#include "llvm/IR/Value.h"
28
29#define DEBUG_TYPE"vectorutils" "vectorutils"
30
31using namespace llvm;
32using namespace llvm::PatternMatch;
33
34/// Maximum factor for an interleaved memory access.
35static 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.
43bool llvm::isTriviallyVectorizable(Intrinsic::ID ID) {
44 switch (ID) {
45 case Intrinsic::bswap: // Begin integer bit-manipulation.
46 case Intrinsic::bitreverse:
47 case Intrinsic::ctpop:
48 case Intrinsic::ctlz:
49 case Intrinsic::cttz:
50 case Intrinsic::fshl:
51 case Intrinsic::fshr:
52 case Intrinsic::sqrt: // Begin floating-point.
53 case Intrinsic::sin:
54 case Intrinsic::cos:
55 case Intrinsic::exp:
56 case Intrinsic::exp2:
57 case Intrinsic::log:
58 case Intrinsic::log10:
59 case Intrinsic::log2:
60 case Intrinsic::fabs:
61 case Intrinsic::minnum:
62 case Intrinsic::maxnum:
63 case Intrinsic::minimum:
64 case Intrinsic::maximum:
65 case Intrinsic::copysign:
66 case Intrinsic::floor:
67 case Intrinsic::ceil:
68 case Intrinsic::trunc:
69 case Intrinsic::rint:
70 case Intrinsic::nearbyint:
71 case Intrinsic::round:
72 case Intrinsic::pow:
73 case Intrinsic::fma:
74 case Intrinsic::fmuladd:
75 case Intrinsic::powi:
76 case Intrinsic::canonicalize:
77 return true;
78 default:
79 return false;
80 }
81}
82
83/// Identifies if the intrinsic has a scalar operand. It check for
84/// ctlz,cttz and powi special intrinsics whose argument is scalar.
85bool llvm::hasVectorInstrinsicScalarOpd(Intrinsic::ID ID,
86 unsigned ScalarOpdIdx) {
87 switch (ID) {
88 case Intrinsic::ctlz:
89 case Intrinsic::cttz:
90 case Intrinsic::powi:
91 return (ScalarOpdIdx == 1);
92 default:
93 return false;
94 }
95}
96
97/// Returns intrinsic ID for call.
98/// For the input call instruction it finds mapping intrinsic and returns
99/// its ID, in case it does not found it return not_intrinsic.
100Intrinsic::ID llvm::getVectorIntrinsicIDForCall(const CallInst *CI,
101 const TargetLibraryInfo *TLI) {
102 Intrinsic::ID ID = getIntrinsicForCallSite(CI, TLI);
103 if (ID == Intrinsic::not_intrinsic)
104 return Intrinsic::not_intrinsic;
105
106 if (isTriviallyVectorizable(ID) || ID == Intrinsic::lifetime_start ||
107 ID == Intrinsic::lifetime_end || ID == Intrinsic::assume ||
108 ID == Intrinsic::sideeffect)
109 return ID;
110 return Intrinsic::not_intrinsic;
111}
112
113/// Find the operand of the GEP that should be checked for consecutive
114/// stores. This ignores trailing indices that have no effect on the final
115/// pointer.
116unsigned llvm::getGEPInductionOperand(const GetElementPtrInst *Gep) {
117 const DataLayout &DL = Gep->getModule()->getDataLayout();
118 unsigned LastOperand = Gep->getNumOperands() - 1;
119 unsigned GEPAllocSize = DL.getTypeAllocSize(Gep->getResultElementType());
120
121 // Walk backwards and try to peel off zeros.
122 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
123 // Find the type we're currently indexing into.
124 gep_type_iterator GEPTI = gep_type_begin(Gep);
125 std::advance(GEPTI, LastOperand - 2);
126
127 // If it's a type with the same allocation size as the result of the GEP we
128 // can peel off the zero index.
129 if (DL.getTypeAllocSize(GEPTI.getIndexedType()) != GEPAllocSize)
130 break;
131 --LastOperand;
132 }
133
134 return LastOperand;
135}
136
137/// If the argument is a GEP, then returns the operand identified by
138/// getGEPInductionOperand. However, if there is some other non-loop-invariant
139/// operand, it returns that instead.
140Value *llvm::stripGetElementPtr(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
141 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
142 if (!GEP)
143 return Ptr;
144
145 unsigned InductionOperand = getGEPInductionOperand(GEP);
146
147 // Check that all of the gep indices are uniform except for our induction
148 // operand.
149 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
150 if (i != InductionOperand &&
151 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
152 return Ptr;
153 return GEP->getOperand(InductionOperand);
154}
155
156/// If a value has only one user that is a CastInst, return it.
157Value *llvm::getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
158 Value *UniqueCast = nullptr;
159 for (User *U : Ptr->users()) {
160 CastInst *CI = dyn_cast<CastInst>(U);
161 if (CI && CI->getType() == Ty) {
162 if (!UniqueCast)
163 UniqueCast = CI;
164 else
165 return nullptr;
166 }
167 }
168 return UniqueCast;
169}
170
171/// Get the stride of a pointer access in a loop. Looks for symbolic
172/// strides "a[i*stride]". Returns the symbolic stride, or null otherwise.
173Value *llvm::getStrideFromPointer(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
174 auto *PtrTy = dyn_cast<PointerType>(Ptr->getType());
175 if (!PtrTy || PtrTy->isAggregateType())
176 return nullptr;
177
178 // Try to remove a gep instruction to make the pointer (actually index at this
179 // point) easier analyzable. If OrigPtr is equal to Ptr we are analyzing the
180 // pointer, otherwise, we are analyzing the index.
181 Value *OrigPtr = Ptr;
182
183 // The size of the pointer access.
184 int64_t PtrAccessSize = 1;
185
186 Ptr = stripGetElementPtr(Ptr, SE, Lp);
187 const SCEV *V = SE->getSCEV(Ptr);
188
189 if (Ptr != OrigPtr)
190 // Strip off casts.
191 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
192 V = C->getOperand();
193
194 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
195 if (!S)
196 return nullptr;
197
198 V = S->getStepRecurrence(*SE);
199 if (!V)
200 return nullptr;
201
202 // Strip off the size of access multiplication if we are still analyzing the
203 // pointer.
204 if (OrigPtr == Ptr) {
205 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
206 if (M->getOperand(0)->getSCEVType() != scConstant)
207 return nullptr;
208
209 const APInt &APStepVal = cast<SCEVConstant>(M->getOperand(0))->getAPInt();
210
211 // Huge step value - give up.
212 if (APStepVal.getBitWidth() > 64)
213 return nullptr;
214
215 int64_t StepVal = APStepVal.getSExtValue();
216 if (PtrAccessSize != StepVal)
217 return nullptr;
218 V = M->getOperand(1);
219 }
220 }
221
222 // Strip off casts.
223 Type *StripedOffRecurrenceCast = nullptr;
224 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
225 StripedOffRecurrenceCast = C->getType();
226 V = C->getOperand();
227 }
228
229 // Look for the loop invariant symbolic value.
230 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
231 if (!U)
232 return nullptr;
233
234 Value *Stride = U->getValue();
235 if (!Lp->isLoopInvariant(Stride))
236 return nullptr;
237
238 // If we have stripped off the recurrence cast we have to make sure that we
239 // return the value that is used in this loop so that we can replace it later.
240 if (StripedOffRecurrenceCast)
241 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
242
243 return Stride;
244}
245
246/// Given a vector and an element number, see if the scalar value is
247/// already around as a register, for example if it were inserted then extracted
248/// from the vector.
249Value *llvm::findScalarElement(Value *V, unsigned EltNo) {
250 assert(V->getType()->isVectorTy() && "Not looking at a vector?")((V->getType()->isVectorTy() && "Not looking at a vector?"
) ? static_cast<void> (0) : __assert_fail ("V->getType()->isVectorTy() && \"Not looking at a vector?\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 250, __PRETTY_FUNCTION__))
;
251 VectorType *VTy = cast<VectorType>(V->getType());
252 unsigned Width = VTy->getNumElements();
253 if (EltNo >= Width) // Out of range access.
254 return UndefValue::get(VTy->getElementType());
255
256 if (Constant *C = dyn_cast<Constant>(V))
257 return C->getAggregateElement(EltNo);
258
259 if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
260 // If this is an insert to a variable element, we don't know what it is.
261 if (!isa<ConstantInt>(III->getOperand(2)))
262 return nullptr;
263 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
264
265 // If this is an insert to the element we are looking for, return the
266 // inserted value.
267 if (EltNo == IIElt)
268 return III->getOperand(1);
269
270 // Otherwise, the insertelement doesn't modify the value, recurse on its
271 // vector input.
272 return findScalarElement(III->getOperand(0), EltNo);
273 }
274
275 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
276 unsigned LHSWidth = SVI->getOperand(0)->getType()->getVectorNumElements();
277 int InEl = SVI->getMaskValue(EltNo);
278 if (InEl < 0)
279 return UndefValue::get(VTy->getElementType());
280 if (InEl < (int)LHSWidth)
281 return findScalarElement(SVI->getOperand(0), InEl);
282 return findScalarElement(SVI->getOperand(1), InEl - LHSWidth);
283 }
284
285 // Extract a value from a vector add operation with a constant zero.
286 // TODO: Use getBinOpIdentity() to generalize this.
287 Value *Val; Constant *C;
288 if (match(V, m_Add(m_Value(Val), m_Constant(C))))
289 if (Constant *Elt = C->getAggregateElement(EltNo))
290 if (Elt->isNullValue())
291 return findScalarElement(Val, EltNo);
292
293 // Otherwise, we don't know.
294 return nullptr;
295}
296
297/// Get splat value if the input is a splat vector or return nullptr.
298/// This function is not fully general. It checks only 2 cases:
299/// the input value is (1) a splat constants vector or (2) a sequence
300/// of instructions that broadcast a single value into a vector.
301///
302const llvm::Value *llvm::getSplatValue(const Value *V) {
303
304 if (auto *C = dyn_cast<Constant>(V))
305 if (isa<VectorType>(V->getType()))
306 return C->getSplatValue();
307
308 auto *ShuffleInst = dyn_cast<ShuffleVectorInst>(V);
309 if (!ShuffleInst)
310 return nullptr;
311 // All-zero (or undef) shuffle mask elements.
312 for (int MaskElt : ShuffleInst->getShuffleMask())
313 if (MaskElt != 0 && MaskElt != -1)
314 return nullptr;
315 // The first shuffle source is 'insertelement' with index 0.
316 auto *InsertEltInst =
317 dyn_cast<InsertElementInst>(ShuffleInst->getOperand(0));
318 if (!InsertEltInst || !isa<ConstantInt>(InsertEltInst->getOperand(2)) ||
319 !cast<ConstantInt>(InsertEltInst->getOperand(2))->isZero())
320 return nullptr;
321
322 return InsertEltInst->getOperand(1);
323}
324
325MapVector<Instruction *, uint64_t>
326llvm::computeMinimumValueSizes(ArrayRef<BasicBlock *> Blocks, DemandedBits &DB,
327 const TargetTransformInfo *TTI) {
328
329 // DemandedBits will give us every value's live-out bits. But we want
330 // to ensure no extra casts would need to be inserted, so every DAG
331 // of connected values must have the same minimum bitwidth.
332 EquivalenceClasses<Value *> ECs;
333 SmallVector<Value *, 16> Worklist;
334 SmallPtrSet<Value *, 4> Roots;
335 SmallPtrSet<Value *, 16> Visited;
336 DenseMap<Value *, uint64_t> DBits;
337 SmallPtrSet<Instruction *, 4> InstructionSet;
338 MapVector<Instruction *, uint64_t> MinBWs;
339
340 // Determine the roots. We work bottom-up, from truncs or icmps.
341 bool SeenExtFromIllegalType = false;
342 for (auto *BB : Blocks)
343 for (auto &I : *BB) {
344 InstructionSet.insert(&I);
345
346 if (TTI && (isa<ZExtInst>(&I) || isa<SExtInst>(&I)) &&
347 !TTI->isTypeLegal(I.getOperand(0)->getType()))
348 SeenExtFromIllegalType = true;
349
350 // Only deal with non-vector integers up to 64-bits wide.
351 if ((isa<TruncInst>(&I) || isa<ICmpInst>(&I)) &&
352 !I.getType()->isVectorTy() &&
353 I.getOperand(0)->getType()->getScalarSizeInBits() <= 64) {
354 // Don't make work for ourselves. If we know the loaded type is legal,
355 // don't add it to the worklist.
356 if (TTI && isa<TruncInst>(&I) && TTI->isTypeLegal(I.getType()))
357 continue;
358
359 Worklist.push_back(&I);
360 Roots.insert(&I);
361 }
362 }
363 // Early exit.
364 if (Worklist.empty() || (TTI && !SeenExtFromIllegalType))
365 return MinBWs;
366
367 // Now proceed breadth-first, unioning values together.
368 while (!Worklist.empty()) {
369 Value *Val = Worklist.pop_back_val();
370 Value *Leader = ECs.getOrInsertLeaderValue(Val);
371
372 if (Visited.count(Val))
373 continue;
374 Visited.insert(Val);
375
376 // Non-instructions terminate a chain successfully.
377 if (!isa<Instruction>(Val))
378 continue;
379 Instruction *I = cast<Instruction>(Val);
380
381 // If we encounter a type that is larger than 64 bits, we can't represent
382 // it so bail out.
383 if (DB.getDemandedBits(I).getBitWidth() > 64)
384 return MapVector<Instruction *, uint64_t>();
385
386 uint64_t V = DB.getDemandedBits(I).getZExtValue();
387 DBits[Leader] |= V;
388 DBits[I] = V;
389
390 // Casts, loads and instructions outside of our range terminate a chain
391 // successfully.
392 if (isa<SExtInst>(I) || isa<ZExtInst>(I) || isa<LoadInst>(I) ||
393 !InstructionSet.count(I))
394 continue;
395
396 // Unsafe casts terminate a chain unsuccessfully. We can't do anything
397 // useful with bitcasts, ptrtoints or inttoptrs and it'd be unsafe to
398 // transform anything that relies on them.
399 if (isa<BitCastInst>(I) || isa<PtrToIntInst>(I) || isa<IntToPtrInst>(I) ||
400 !I->getType()->isIntegerTy()) {
401 DBits[Leader] |= ~0ULL;
402 continue;
403 }
404
405 // We don't modify the types of PHIs. Reductions will already have been
406 // truncated if possible, and inductions' sizes will have been chosen by
407 // indvars.
408 if (isa<PHINode>(I))
409 continue;
410
411 if (DBits[Leader] == ~0ULL)
412 // All bits demanded, no point continuing.
413 continue;
414
415 for (Value *O : cast<User>(I)->operands()) {
416 ECs.unionSets(Leader, O);
417 Worklist.push_back(O);
418 }
419 }
420
421 // Now we've discovered all values, walk them to see if there are
422 // any users we didn't see. If there are, we can't optimize that
423 // chain.
424 for (auto &I : DBits)
425 for (auto *U : I.first->users())
426 if (U->getType()->isIntegerTy() && DBits.count(U) == 0)
427 DBits[ECs.getOrInsertLeaderValue(I.first)] |= ~0ULL;
428
429 for (auto I = ECs.begin(), E = ECs.end(); I != E; ++I) {
430 uint64_t LeaderDemandedBits = 0;
431 for (auto MI = ECs.member_begin(I), ME = ECs.member_end(); MI != ME; ++MI)
432 LeaderDemandedBits |= DBits[*MI];
433
434 uint64_t MinBW = (sizeof(LeaderDemandedBits) * 8) -
435 llvm::countLeadingZeros(LeaderDemandedBits);
436 // Round up to a power of 2
437 if (!isPowerOf2_64((uint64_t)MinBW))
438 MinBW = NextPowerOf2(MinBW);
439
440 // We don't modify the types of PHIs. Reductions will already have been
441 // truncated if possible, and inductions' sizes will have been chosen by
442 // indvars.
443 // If we are required to shrink a PHI, abandon this entire equivalence class.
444 bool Abort = false;
445 for (auto MI = ECs.member_begin(I), ME = ECs.member_end(); MI != ME; ++MI)
446 if (isa<PHINode>(*MI) && MinBW < (*MI)->getType()->getScalarSizeInBits()) {
447 Abort = true;
448 break;
449 }
450 if (Abort)
451 continue;
452
453 for (auto MI = ECs.member_begin(I), ME = ECs.member_end(); MI != ME; ++MI) {
454 if (!isa<Instruction>(*MI))
455 continue;
456 Type *Ty = (*MI)->getType();
457 if (Roots.count(*MI))
458 Ty = cast<Instruction>(*MI)->getOperand(0)->getType();
459 if (MinBW < Ty->getScalarSizeInBits())
460 MinBWs[cast<Instruction>(*MI)] = MinBW;
461 }
462 }
463
464 return MinBWs;
465}
466
467/// Add all access groups in @p AccGroups to @p List.
468template <typename ListT>
469static void addToAccessGroupList(ListT &List, MDNode *AccGroups) {
470 // Interpret an access group as a list containing itself.
471 if (AccGroups->getNumOperands() == 0) {
472 assert(isValidAsAccessGroup(AccGroups) && "Node must be an access group")((isValidAsAccessGroup(AccGroups) && "Node must be an access group"
) ? static_cast<void> (0) : __assert_fail ("isValidAsAccessGroup(AccGroups) && \"Node must be an access group\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 472, __PRETTY_FUNCTION__))
;
473 List.insert(AccGroups);
474 return;
475 }
476
477 for (auto &AccGroupListOp : AccGroups->operands()) {
478 auto *Item = cast<MDNode>(AccGroupListOp.get());
479 assert(isValidAsAccessGroup(Item) && "List item must be an access group")((isValidAsAccessGroup(Item) && "List item must be an access group"
) ? static_cast<void> (0) : __assert_fail ("isValidAsAccessGroup(Item) && \"List item must be an access group\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 479, __PRETTY_FUNCTION__))
;
480 List.insert(Item);
481 }
482}
483
484MDNode *llvm::uniteAccessGroups(MDNode *AccGroups1, MDNode *AccGroups2) {
485 if (!AccGroups1)
486 return AccGroups2;
487 if (!AccGroups2)
488 return AccGroups1;
489 if (AccGroups1 == AccGroups2)
490 return AccGroups1;
491
492 SmallSetVector<Metadata *, 4> Union;
493 addToAccessGroupList(Union, AccGroups1);
494 addToAccessGroupList(Union, AccGroups2);
495
496 if (Union.size() == 0)
497 return nullptr;
498 if (Union.size() == 1)
499 return cast<MDNode>(Union.front());
500
501 LLVMContext &Ctx = AccGroups1->getContext();
502 return MDNode::get(Ctx, Union.getArrayRef());
503}
504
505MDNode *llvm::intersectAccessGroups(const Instruction *Inst1,
506 const Instruction *Inst2) {
507 bool MayAccessMem1 = Inst1->mayReadOrWriteMemory();
508 bool MayAccessMem2 = Inst2->mayReadOrWriteMemory();
509
510 if (!MayAccessMem1 && !MayAccessMem2)
511 return nullptr;
512 if (!MayAccessMem1)
513 return Inst2->getMetadata(LLVMContext::MD_access_group);
514 if (!MayAccessMem2)
515 return Inst1->getMetadata(LLVMContext::MD_access_group);
516
517 MDNode *MD1 = Inst1->getMetadata(LLVMContext::MD_access_group);
518 MDNode *MD2 = Inst2->getMetadata(LLVMContext::MD_access_group);
519 if (!MD1 || !MD2)
520 return nullptr;
521 if (MD1 == MD2)
522 return MD1;
523
524 // Use set for scalable 'contains' check.
525 SmallPtrSet<Metadata *, 4> AccGroupSet2;
526 addToAccessGroupList(AccGroupSet2, MD2);
527
528 SmallVector<Metadata *, 4> Intersection;
529 if (MD1->getNumOperands() == 0) {
530 assert(isValidAsAccessGroup(MD1) && "Node must be an access group")((isValidAsAccessGroup(MD1) && "Node must be an access group"
) ? static_cast<void> (0) : __assert_fail ("isValidAsAccessGroup(MD1) && \"Node must be an access group\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 530, __PRETTY_FUNCTION__))
;
531 if (AccGroupSet2.count(MD1))
532 Intersection.push_back(MD1);
533 } else {
534 for (const MDOperand &Node : MD1->operands()) {
535 auto *Item = cast<MDNode>(Node.get());
536 assert(isValidAsAccessGroup(Item) && "List item must be an access group")((isValidAsAccessGroup(Item) && "List item must be an access group"
) ? static_cast<void> (0) : __assert_fail ("isValidAsAccessGroup(Item) && \"List item must be an access group\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 536, __PRETTY_FUNCTION__))
;
537 if (AccGroupSet2.count(Item))
538 Intersection.push_back(Item);
539 }
540 }
541
542 if (Intersection.size() == 0)
543 return nullptr;
544 if (Intersection.size() == 1)
545 return cast<MDNode>(Intersection.front());
546
547 LLVMContext &Ctx = Inst1->getContext();
548 return MDNode::get(Ctx, Intersection);
549}
550
551/// \returns \p I after propagating metadata from \p VL.
552Instruction *llvm::propagateMetadata(Instruction *Inst, ArrayRef<Value *> VL) {
553 Instruction *I0 = cast<Instruction>(VL[0]);
554 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
555 I0->getAllMetadataOtherThanDebugLoc(Metadata);
556
557 for (auto Kind : {LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
558 LLVMContext::MD_noalias, LLVMContext::MD_fpmath,
559 LLVMContext::MD_nontemporal, LLVMContext::MD_invariant_load,
560 LLVMContext::MD_access_group}) {
561 MDNode *MD = I0->getMetadata(Kind);
562
563 for (int J = 1, E = VL.size(); MD && J != E; ++J) {
564 const Instruction *IJ = cast<Instruction>(VL[J]);
565 MDNode *IMD = IJ->getMetadata(Kind);
566 switch (Kind) {
567 case LLVMContext::MD_tbaa:
568 MD = MDNode::getMostGenericTBAA(MD, IMD);
569 break;
570 case LLVMContext::MD_alias_scope:
571 MD = MDNode::getMostGenericAliasScope(MD, IMD);
572 break;
573 case LLVMContext::MD_fpmath:
574 MD = MDNode::getMostGenericFPMath(MD, IMD);
575 break;
576 case LLVMContext::MD_noalias:
577 case LLVMContext::MD_nontemporal:
578 case LLVMContext::MD_invariant_load:
579 MD = MDNode::intersect(MD, IMD);
580 break;
581 case LLVMContext::MD_access_group:
582 MD = intersectAccessGroups(Inst, IJ);
583 break;
584 default:
585 llvm_unreachable("unhandled metadata")::llvm::llvm_unreachable_internal("unhandled metadata", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 585)
;
586 }
587 }
588
589 Inst->setMetadata(Kind, MD);
590 }
591
592 return Inst;
593}
594
595Constant *
596llvm::createBitMaskForGaps(IRBuilder<> &Builder, unsigned VF,
597 const InterleaveGroup<Instruction> &Group) {
598 // All 1's means mask is not needed.
599 if (Group.getNumMembers() == Group.getFactor())
600 return nullptr;
601
602 // TODO: support reversed access.
603 assert(!Group.isReverse() && "Reversed group not supported.")((!Group.isReverse() && "Reversed group not supported."
) ? static_cast<void> (0) : __assert_fail ("!Group.isReverse() && \"Reversed group not supported.\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 603, __PRETTY_FUNCTION__))
;
604
605 SmallVector<Constant *, 16> Mask;
606 for (unsigned i = 0; i < VF; i++)
607 for (unsigned j = 0; j < Group.getFactor(); ++j) {
608 unsigned HasMember = Group.getMember(j) ? 1 : 0;
609 Mask.push_back(Builder.getInt1(HasMember));
610 }
611
612 return ConstantVector::get(Mask);
613}
614
615Constant *llvm::createReplicatedMask(IRBuilder<> &Builder,
616 unsigned ReplicationFactor, unsigned VF) {
617 SmallVector<Constant *, 16> MaskVec;
618 for (unsigned i = 0; i < VF; i++)
619 for (unsigned j = 0; j < ReplicationFactor; j++)
620 MaskVec.push_back(Builder.getInt32(i));
621
622 return ConstantVector::get(MaskVec);
623}
624
625Constant *llvm::createInterleaveMask(IRBuilder<> &Builder, unsigned VF,
626 unsigned NumVecs) {
627 SmallVector<Constant *, 16> Mask;
628 for (unsigned i = 0; i < VF; i++)
629 for (unsigned j = 0; j < NumVecs; j++)
630 Mask.push_back(Builder.getInt32(j * VF + i));
631
632 return ConstantVector::get(Mask);
633}
634
635Constant *llvm::createStrideMask(IRBuilder<> &Builder, unsigned Start,
636 unsigned Stride, unsigned VF) {
637 SmallVector<Constant *, 16> Mask;
638 for (unsigned i = 0; i < VF; i++)
639 Mask.push_back(Builder.getInt32(Start + i * Stride));
640
641 return ConstantVector::get(Mask);
642}
643
644Constant *llvm::createSequentialMask(IRBuilder<> &Builder, unsigned Start,
645 unsigned NumInts, unsigned NumUndefs) {
646 SmallVector<Constant *, 16> Mask;
647 for (unsigned i = 0; i < NumInts; i++)
648 Mask.push_back(Builder.getInt32(Start + i));
649
650 Constant *Undef = UndefValue::get(Builder.getInt32Ty());
651 for (unsigned i = 0; i < NumUndefs; i++)
652 Mask.push_back(Undef);
653
654 return ConstantVector::get(Mask);
655}
656
657/// A helper function for concatenating vectors. This function concatenates two
658/// vectors having the same element type. If the second vector has fewer
659/// elements than the first, it is padded with undefs.
660static Value *concatenateTwoVectors(IRBuilder<> &Builder, Value *V1,
661 Value *V2) {
662 VectorType *VecTy1 = dyn_cast<VectorType>(V1->getType());
663 VectorType *VecTy2 = dyn_cast<VectorType>(V2->getType());
664 assert(VecTy1 && VecTy2 &&((VecTy1 && VecTy2 && VecTy1->getScalarType
() == VecTy2->getScalarType() && "Expect two vectors with the same element type"
) ? static_cast<void> (0) : __assert_fail ("VecTy1 && VecTy2 && VecTy1->getScalarType() == VecTy2->getScalarType() && \"Expect two vectors with the same element type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 666, __PRETTY_FUNCTION__))
665 VecTy1->getScalarType() == VecTy2->getScalarType() &&((VecTy1 && VecTy2 && VecTy1->getScalarType
() == VecTy2->getScalarType() && "Expect two vectors with the same element type"
) ? static_cast<void> (0) : __assert_fail ("VecTy1 && VecTy2 && VecTy1->getScalarType() == VecTy2->getScalarType() && \"Expect two vectors with the same element type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 666, __PRETTY_FUNCTION__))
666 "Expect two vectors with the same element type")((VecTy1 && VecTy2 && VecTy1->getScalarType
() == VecTy2->getScalarType() && "Expect two vectors with the same element type"
) ? static_cast<void> (0) : __assert_fail ("VecTy1 && VecTy2 && VecTy1->getScalarType() == VecTy2->getScalarType() && \"Expect two vectors with the same element type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 666, __PRETTY_FUNCTION__))
;
667
668 unsigned NumElts1 = VecTy1->getNumElements();
669 unsigned NumElts2 = VecTy2->getNumElements();
670 assert(NumElts1 >= NumElts2 && "Unexpect the first vector has less elements")((NumElts1 >= NumElts2 && "Unexpect the first vector has less elements"
) ? static_cast<void> (0) : __assert_fail ("NumElts1 >= NumElts2 && \"Unexpect the first vector has less elements\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 670, __PRETTY_FUNCTION__))
;
671
672 if (NumElts1 > NumElts2) {
673 // Extend with UNDEFs.
674 Constant *ExtMask =
675 createSequentialMask(Builder, 0, NumElts2, NumElts1 - NumElts2);
676 V2 = Builder.CreateShuffleVector(V2, UndefValue::get(VecTy2), ExtMask);
677 }
678
679 Constant *Mask = createSequentialMask(Builder, 0, NumElts1 + NumElts2, 0);
680 return Builder.CreateShuffleVector(V1, V2, Mask);
681}
682
683Value *llvm::concatenateVectors(IRBuilder<> &Builder, ArrayRef<Value *> Vecs) {
684 unsigned NumVecs = Vecs.size();
685 assert(NumVecs > 1 && "Should be at least two vectors")((NumVecs > 1 && "Should be at least two vectors")
? static_cast<void> (0) : __assert_fail ("NumVecs > 1 && \"Should be at least two vectors\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 685, __PRETTY_FUNCTION__))
;
686
687 SmallVector<Value *, 8> ResList;
688 ResList.append(Vecs.begin(), Vecs.end());
689 do {
690 SmallVector<Value *, 8> TmpList;
691 for (unsigned i = 0; i < NumVecs - 1; i += 2) {
692 Value *V0 = ResList[i], *V1 = ResList[i + 1];
693 assert((V0->getType() == V1->getType() || i == NumVecs - 2) &&(((V0->getType() == V1->getType() || i == NumVecs - 2) &&
"Only the last vector may have a different type") ? static_cast
<void> (0) : __assert_fail ("(V0->getType() == V1->getType() || i == NumVecs - 2) && \"Only the last vector may have a different type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 694, __PRETTY_FUNCTION__))
694 "Only the last vector may have a different type")(((V0->getType() == V1->getType() || i == NumVecs - 2) &&
"Only the last vector may have a different type") ? static_cast
<void> (0) : __assert_fail ("(V0->getType() == V1->getType() || i == NumVecs - 2) && \"Only the last vector may have a different type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 694, __PRETTY_FUNCTION__))
;
695
696 TmpList.push_back(concatenateTwoVectors(Builder, V0, V1));
697 }
698
699 // Push the last vector if the total number of vectors is odd.
700 if (NumVecs % 2 != 0)
701 TmpList.push_back(ResList[NumVecs - 1]);
702
703 ResList = TmpList;
704 NumVecs = ResList.size();
705 } while (NumVecs > 1);
706
707 return ResList[0];
708}
709
710bool InterleavedAccessInfo::isStrided(int Stride) {
711 unsigned Factor = std::abs(Stride);
712 return Factor >= 2 && Factor <= MaxInterleaveGroupFactor;
713}
714
715void InterleavedAccessInfo::collectConstStrideAccesses(
716 MapVector<Instruction *, StrideDescriptor> &AccessStrideInfo,
717 const ValueToValueMap &Strides) {
718 auto &DL = TheLoop->getHeader()->getModule()->getDataLayout();
719
720 // Since it's desired that the load/store instructions be maintained in
721 // "program order" for the interleaved access analysis, we have to visit the
722 // blocks in the loop in reverse postorder (i.e., in a topological order).
723 // Such an ordering will ensure that any load/store that may be executed
724 // before a second load/store will precede the second load/store in
725 // AccessStrideInfo.
726 LoopBlocksDFS DFS(TheLoop);
727 DFS.perform(LI);
728 for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO()))
729 for (auto &I : *BB) {
730 auto *LI = dyn_cast<LoadInst>(&I);
731 auto *SI = dyn_cast<StoreInst>(&I);
732 if (!LI && !SI)
733 continue;
734
735 Value *Ptr = getLoadStorePointerOperand(&I);
736 // We don't check wrapping here because we don't know yet if Ptr will be
737 // part of a full group or a group with gaps. Checking wrapping for all
738 // pointers (even those that end up in groups with no gaps) will be overly
739 // conservative. For full groups, wrapping should be ok since if we would
740 // wrap around the address space we would do a memory access at nullptr
741 // even without the transformation. The wrapping checks are therefore
742 // deferred until after we've formed the interleaved groups.
743 int64_t Stride = getPtrStride(PSE, Ptr, TheLoop, Strides,
744 /*Assume=*/true, /*ShouldCheckWrap=*/false);
745
746 const SCEV *Scev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
747 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
748 uint64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
749
750 // An alignment of 0 means target ABI alignment.
751 unsigned Align = getLoadStoreAlignment(&I);
752 if (!Align)
753 Align = DL.getABITypeAlignment(PtrTy->getElementType());
754
755 AccessStrideInfo[&I] = StrideDescriptor(Stride, Scev, Size, Align);
756 }
757}
758
759// Analyze interleaved accesses and collect them into interleaved load and
760// store groups.
761//
762// When generating code for an interleaved load group, we effectively hoist all
763// loads in the group to the location of the first load in program order. When
764// generating code for an interleaved store group, we sink all stores to the
765// location of the last store. This code motion can change the order of load
766// and store instructions and may break dependences.
767//
768// The code generation strategy mentioned above ensures that we won't violate
769// any write-after-read (WAR) dependences.
770//
771// E.g., for the WAR dependence: a = A[i]; // (1)
772// A[i] = b; // (2)
773//
774// The store group of (2) is always inserted at or below (2), and the load
775// group of (1) is always inserted at or above (1). Thus, the instructions will
776// never be reordered. All other dependences are checked to ensure the
777// correctness of the instruction reordering.
778//
779// The algorithm visits all memory accesses in the loop in bottom-up program
780// order. Program order is established by traversing the blocks in the loop in
781// reverse postorder when collecting the accesses.
782//
783// We visit the memory accesses in bottom-up order because it can simplify the
784// construction of store groups in the presence of write-after-write (WAW)
785// dependences.
786//
787// E.g., for the WAW dependence: A[i] = a; // (1)
788// A[i] = b; // (2)
789// A[i + 1] = c; // (3)
790//
791// We will first create a store group with (3) and (2). (1) can't be added to
792// this group because it and (2) are dependent. However, (1) can be grouped
793// with other accesses that may precede it in program order. Note that a
794// bottom-up order does not imply that WAW dependences should not be checked.
795void InterleavedAccessInfo::analyzeInterleaving(
796 bool EnablePredicatedInterleavedMemAccesses) {
797 LLVM_DEBUG(dbgs() << "LV: Analyzing interleaved accesses...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Analyzing interleaved accesses...\n"
; } } while (false)
;
1
Assuming 'DebugFlag' is 0
2
Loop condition is false. Exiting loop
798 const ValueToValueMap &Strides = LAI->getSymbolicStrides();
799
800 // Holds all accesses with a constant stride.
801 MapVector<Instruction *, StrideDescriptor> AccessStrideInfo;
802 collectConstStrideAccesses(AccessStrideInfo, Strides);
803
804 if (AccessStrideInfo.empty())
3
Assuming the condition is false
4
Taking false branch
805 return;
806
807 // Collect the dependences in the loop.
808 collectDependences();
809
810 // Holds all interleaved store groups temporarily.
811 SmallSetVector<InterleaveGroup<Instruction> *, 4> StoreGroups;
812 // Holds all interleaved load groups temporarily.
813 SmallSetVector<InterleaveGroup<Instruction> *, 4> LoadGroups;
814
815 // Search in bottom-up program order for pairs of accesses (A and B) that can
816 // form interleaved load or store groups. In the algorithm below, access A
817 // precedes access B in program order. We initialize a group for B in the
818 // outer loop of the algorithm, and then in the inner loop, we attempt to
819 // insert each A into B's group if:
820 //
821 // 1. A and B have the same stride,
822 // 2. A and B have the same memory object size, and
823 // 3. A belongs in B's group according to its distance from B.
824 //
825 // Special care is taken to ensure group formation will not break any
826 // dependences.
827 for (auto BI = AccessStrideInfo.rbegin(), E = AccessStrideInfo.rend();
5
Loop condition is true. Entering loop body
828 BI != E; ++BI) {
829 Instruction *B = BI->first;
830 StrideDescriptor DesB = BI->second;
831
832 // Initialize a group for B if it has an allowable stride. Even if we don't
833 // create a group for B, we continue with the bottom-up algorithm to ensure
834 // we don't break any of B's dependences.
835 InterleaveGroup<Instruction> *Group = nullptr;
6
'Group' initialized to a null pointer value
836 if (isStrided(DesB.Stride) &&
837 (!isPredicated(B->getParent()) || EnablePredicatedInterleavedMemAccesses)) {
838 Group = getInterleaveGroup(B);
839 if (!Group) {
840 LLVM_DEBUG(dbgs() << "LV: Creating an interleave group with:" << *Bdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Creating an interleave group with:"
<< *B << '\n'; } } while (false)
841 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Creating an interleave group with:"
<< *B << '\n'; } } while (false)
;
842 Group = createInterleaveGroup(B, DesB.Stride, DesB.Align);
843 }
844 if (B->mayWriteToMemory())
845 StoreGroups.insert(Group);
846 else
847 LoadGroups.insert(Group);
848 }
849
850 for (auto AI = std::next(BI); AI != E; ++AI) {
7
Loop condition is true. Entering loop body
851 Instruction *A = AI->first;
852 StrideDescriptor DesA = AI->second;
853
854 // Our code motion strategy implies that we can't have dependences
855 // between accesses in an interleaved group and other accesses located
856 // between the first and last member of the group. Note that this also
857 // means that a group can't have more than one member at a given offset.
858 // The accesses in a group can have dependences with other accesses, but
859 // we must ensure we don't extend the boundaries of the group such that
860 // we encompass those dependent accesses.
861 //
862 // For example, assume we have the sequence of accesses shown below in a
863 // stride-2 loop:
864 //
865 // (1, 2) is a group | A[i] = a; // (1)
866 // | A[i-1] = b; // (2) |
867 // A[i-3] = c; // (3)
868 // A[i] = d; // (4) | (2, 4) is not a group
869 //
870 // Because accesses (2) and (3) are dependent, we can group (2) with (1)
871 // but not with (4). If we did, the dependent access (3) would be within
872 // the boundaries of the (2, 4) group.
873 if (!canReorderMemAccessesForInterleavedGroups(&*AI, &*BI)) {
8
Taking false branch
874 // If a dependence exists and A is already in a group, we know that A
875 // must be a store since A precedes B and WAR dependences are allowed.
876 // Thus, A would be sunk below B. We release A's group to prevent this
877 // illegal code motion. A will then be free to form another group with
878 // instructions that precede it.
879 if (isInterleaved(A)) {
880 InterleaveGroup<Instruction> *StoreGroup = getInterleaveGroup(A);
881 StoreGroups.remove(StoreGroup);
882 releaseGroup(StoreGroup);
883 }
884
885 // If a dependence exists and A is not already in a group (or it was
886 // and we just released it), B might be hoisted above A (if B is a
887 // load) or another store might be sunk below A (if B is a store). In
888 // either case, we can't add additional instructions to B's group. B
889 // will only form a group with instructions that it precedes.
890 break;
891 }
892
893 // At this point, we've checked for illegal code motion. If either A or B
894 // isn't strided, there's nothing left to do.
895 if (!isStrided(DesA.Stride) || !isStrided(DesB.Stride))
9
Taking false branch
896 continue;
897
898 // Ignore A if it's already in a group or isn't the same kind of memory
899 // operation as B.
900 // Note that mayReadFromMemory() isn't mutually exclusive to
901 // mayWriteToMemory in the case of atomic loads. We shouldn't see those
902 // here, canVectorizeMemory() should have returned false - except for the
903 // case we asked for optimization remarks.
904 if (isInterleaved(A) ||
10
Assuming the condition is false
13
Taking false branch
905 (A->mayReadFromMemory() != B->mayReadFromMemory()) ||
11
Assuming the condition is false
906 (A->mayWriteToMemory() != B->mayWriteToMemory()))
12
Assuming the condition is false
907 continue;
908
909 // Check rules 1 and 2. Ignore A if its stride or size is different from
910 // that of B.
911 if (DesA.Stride != DesB.Stride || DesA.Size != DesB.Size)
14
Assuming the condition is false
15
Assuming the condition is false
16
Taking false branch
912 continue;
913
914 // Ignore A if the memory object of A and B don't belong to the same
915 // address space
916 if (getLoadStoreAddressSpace(A) != getLoadStoreAddressSpace(B))
17
Assuming the condition is false
18
Taking false branch
917 continue;
918
919 // Calculate the distance from A to B.
920 const SCEVConstant *DistToB = dyn_cast<SCEVConstant>(
921 PSE.getSE()->getMinusSCEV(DesA.Scev, DesB.Scev));
922 if (!DistToB)
19
Taking false branch
923 continue;
924 int64_t DistanceToB = DistToB->getAPInt().getSExtValue();
925
926 // Check rule 3. Ignore A if its distance to B is not a multiple of the
927 // size.
928 if (DistanceToB % static_cast<int64_t>(DesB.Size))
20
Assuming the condition is false
21
Taking false branch
929 continue;
930
931 // All members of a predicated interleave-group must have the same predicate,
932 // and currently must reside in the same BB.
933 BasicBlock *BlockA = A->getParent();
934 BasicBlock *BlockB = B->getParent();
935 if ((isPredicated(BlockA) || isPredicated(BlockB)) &&
22
Assuming the condition is false
23
Assuming the condition is false
936 (!EnablePredicatedInterleavedMemAccesses || BlockA != BlockB))
937 continue;
938
939 // The index of A is the index of B plus A's distance to B in multiples
940 // of the size.
941 int IndexA =
942 Group->getIndex(B) + DistanceToB / static_cast<int64_t>(DesB.Size);
24
Called C++ object pointer is null
943
944 // Try to insert A into B's group.
945 if (Group->insertMember(A, IndexA, DesA.Align)) {
946 LLVM_DEBUG(dbgs() << "LV: Inserted:" << *A << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Inserted:" << *
A << '\n' << " into the interleave group with"
<< *B << '\n'; } } while (false)
947 << " into the interleave group with" << *Bdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Inserted:" << *
A << '\n' << " into the interleave group with"
<< *B << '\n'; } } while (false)
948 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Inserted:" << *
A << '\n' << " into the interleave group with"
<< *B << '\n'; } } while (false)
;
949 InterleaveGroupMap[A] = Group;
950
951 // Set the first load in program order as the insert position.
952 if (A->mayReadFromMemory())
953 Group->setInsertPos(A);
954 }
955 } // Iteration over A accesses.
956 } // Iteration over B accesses.
957
958 // Remove interleaved store groups with gaps.
959 for (auto *Group : StoreGroups)
960 if (Group->getNumMembers() != Group->getFactor()) {
961 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved store group due "
"to gaps.\n"; } } while (false)
962 dbgs() << "LV: Invalidate candidate interleaved store group due "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved store group due "
"to gaps.\n"; } } while (false)
963 "to gaps.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved store group due "
"to gaps.\n"; } } while (false)
;
964 releaseGroup(Group);
965 }
966 // Remove interleaved groups with gaps (currently only loads) whose memory
967 // accesses may wrap around. We have to revisit the getPtrStride analysis,
968 // this time with ShouldCheckWrap=true, since collectConstStrideAccesses does
969 // not check wrapping (see documentation there).
970 // FORNOW we use Assume=false;
971 // TODO: Change to Assume=true but making sure we don't exceed the threshold
972 // of runtime SCEV assumptions checks (thereby potentially failing to
973 // vectorize altogether).
974 // Additional optional optimizations:
975 // TODO: If we are peeling the loop and we know that the first pointer doesn't
976 // wrap then we can deduce that all pointers in the group don't wrap.
977 // This means that we can forcefully peel the loop in order to only have to
978 // check the first pointer for no-wrap. When we'll change to use Assume=true
979 // we'll only need at most one runtime check per interleaved group.
980 for (auto *Group : LoadGroups) {
981 // Case 1: A full group. Can Skip the checks; For full groups, if the wide
982 // load would wrap around the address space we would do a memory access at
983 // nullptr even without the transformation.
984 if (Group->getNumMembers() == Group->getFactor())
985 continue;
986
987 // Case 2: If first and last members of the group don't wrap this implies
988 // that all the pointers in the group don't wrap.
989 // So we check only group member 0 (which is always guaranteed to exist),
990 // and group member Factor - 1; If the latter doesn't exist we rely on
991 // peeling (if it is a non-reveresed accsess -- see Case 3).
992 Value *FirstMemberPtr = getLoadStorePointerOperand(Group->getMember(0));
993 if (!getPtrStride(PSE, FirstMemberPtr, TheLoop, Strides, /*Assume=*/false,
994 /*ShouldCheckWrap=*/true)) {
995 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to "
"first group member potentially pointer-wrapping.\n"; } } while
(false)
996 dbgs() << "LV: Invalidate candidate interleaved group due to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to "
"first group member potentially pointer-wrapping.\n"; } } while
(false)
997 "first group member potentially pointer-wrapping.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to "
"first group member potentially pointer-wrapping.\n"; } } while
(false)
;
998 releaseGroup(Group);
999 continue;
1000 }
1001 Instruction *LastMember = Group->getMember(Group->getFactor() - 1);
1002 if (LastMember) {
1003 Value *LastMemberPtr = getLoadStorePointerOperand(LastMember);
1004 if (!getPtrStride(PSE, LastMemberPtr, TheLoop, Strides, /*Assume=*/false,
1005 /*ShouldCheckWrap=*/true)) {
1006 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to "
"last group member potentially pointer-wrapping.\n"; } } while
(false)
1007 dbgs() << "LV: Invalidate candidate interleaved group due to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to "
"last group member potentially pointer-wrapping.\n"; } } while
(false)
1008 "last group member potentially pointer-wrapping.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to "
"last group member potentially pointer-wrapping.\n"; } } while
(false)
;
1009 releaseGroup(Group);
1010 }
1011 } else {
1012 // Case 3: A non-reversed interleaved load group with gaps: We need
1013 // to execute at least one scalar epilogue iteration. This will ensure
1014 // we don't speculatively access memory out-of-bounds. We only need
1015 // to look for a member at index factor - 1, since every group must have
1016 // a member at index zero.
1017 if (Group->isReverse()) {
1018 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to "
"a reverse access with gaps.\n"; } } while (false)
1019 dbgs() << "LV: Invalidate candidate interleaved group due to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to "
"a reverse access with gaps.\n"; } } while (false)
1020 "a reverse access with gaps.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to "
"a reverse access with gaps.\n"; } } while (false)
;
1021 releaseGroup(Group);
1022 continue;
1023 }
1024 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Interleaved group requires epilogue iteration.\n"
; } } while (false)
1025 dbgs() << "LV: Interleaved group requires epilogue iteration.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Interleaved group requires epilogue iteration.\n"
; } } while (false)
;
1026 RequiresScalarEpilogue = true;
1027 }
1028 }
1029}
1030
1031void InterleavedAccessInfo::invalidateGroupsRequiringScalarEpilogue() {
1032 // If no group had triggered the requirement to create an epilogue loop,
1033 // there is nothing to do.
1034 if (!requiresScalarEpilogue())
1035 return;
1036
1037 // Avoid releasing a Group twice.
1038 SmallPtrSet<InterleaveGroup<Instruction> *, 4> DelSet;
1039 for (auto &I : InterleaveGroupMap) {
1040 InterleaveGroup<Instruction> *Group = I.second;
1041 if (Group->requiresScalarEpilogue())
1042 DelSet.insert(Group);
1043 }
1044 for (auto *Ptr : DelSet) {
1045 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to gaps that "
"require a scalar epilogue (not allowed under optsize) and cannot "
"be masked (not enabled). \n"; } } while (false)
1046 dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to gaps that "
"require a scalar epilogue (not allowed under optsize) and cannot "
"be masked (not enabled). \n"; } } while (false)
1047 << "LV: Invalidate candidate interleaved group due to gaps that "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to gaps that "
"require a scalar epilogue (not allowed under optsize) and cannot "
"be masked (not enabled). \n"; } } while (false)
1048 "require a scalar epilogue (not allowed under optsize) and cannot "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to gaps that "
"require a scalar epilogue (not allowed under optsize) and cannot "
"be masked (not enabled). \n"; } } while (false)
1049 "be masked (not enabled). \n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("vectorutils")) { dbgs() << "LV: Invalidate candidate interleaved group due to gaps that "
"require a scalar epilogue (not allowed under optsize) and cannot "
"be masked (not enabled). \n"; } } while (false)
;
1050 releaseGroup(Ptr);
1051 }
1052
1053 RequiresScalarEpilogue = false;
1054}
1055
1056template <typename InstT>
1057void InterleaveGroup<InstT>::addMetadata(InstT *NewInst) const {
1058 llvm_unreachable("addMetadata can only be used for Instruction")::llvm::llvm_unreachable_internal("addMetadata can only be used for Instruction"
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Analysis/VectorUtils.cpp"
, 1058)
;
1059}
1060
1061namespace llvm {
1062template <>
1063void InterleaveGroup<Instruction>::addMetadata(Instruction *NewInst) const {
1064 SmallVector<Value *, 4> VL;
1065 std::transform(Members.begin(), Members.end(), std::back_inserter(VL),
1066 [](std::pair<int, Instruction *> p) { return p.second; });
1067 propagateMetadata(NewInst, VL);
1068}
1069}