1 | //===- LoopVectorizationLegality.cpp --------------------------------------===// |
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 provides loop vectorization legality analysis. Original code |
10 | // resided in LoopVectorize.cpp for a long time. |
11 | // |
12 | // At this point, it is implemented as a utility class, not as an analysis |
13 | // pass. It should be easy to create an analysis pass around it if there |
14 | // is a need (but D45420 needs to happen first). |
15 | // |
16 | |
17 | #include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h" |
18 | #include "llvm/Analysis/Loads.h" |
19 | #include "llvm/Analysis/LoopInfo.h" |
20 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
21 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
22 | #include "llvm/Analysis/TargetLibraryInfo.h" |
23 | #include "llvm/Analysis/TargetTransformInfo.h" |
24 | #include "llvm/Analysis/ValueTracking.h" |
25 | #include "llvm/Analysis/VectorUtils.h" |
26 | #include "llvm/IR/IntrinsicInst.h" |
27 | #include "llvm/IR/PatternMatch.h" |
28 | #include "llvm/Transforms/Utils/SizeOpts.h" |
29 | #include "llvm/Transforms/Vectorize/LoopVectorize.h" |
30 | |
31 | using namespace llvm; |
32 | using namespace PatternMatch; |
33 | |
34 | #define LV_NAME "loop-vectorize" |
35 | #define DEBUG_TYPE LV_NAME |
36 | |
37 | static cl::opt<bool> |
38 | EnableIfConversion("enable-if-conversion" , cl::init(Val: true), cl::Hidden, |
39 | cl::desc("Enable if-conversion during vectorization." )); |
40 | |
41 | static cl::opt<bool> |
42 | AllowStridedPointerIVs("lv-strided-pointer-ivs" , cl::init(Val: false), cl::Hidden, |
43 | cl::desc("Enable recognition of non-constant strided " |
44 | "pointer induction variables." )); |
45 | |
46 | static cl::opt<bool> |
47 | HintsAllowReordering("hints-allow-reordering" , cl::init(Val: true), cl::Hidden, |
48 | cl::desc("Allow enabling loop hints to reorder " |
49 | "FP operations during vectorization." )); |
50 | |
51 | // TODO: Move size-based thresholds out of legality checking, make cost based |
52 | // decisions instead of hard thresholds. |
53 | static cl::opt<unsigned> VectorizeSCEVCheckThreshold( |
54 | "vectorize-scev-check-threshold" , cl::init(Val: 16), cl::Hidden, |
55 | cl::desc("The maximum number of SCEV checks allowed." )); |
56 | |
57 | static cl::opt<unsigned> PragmaVectorizeSCEVCheckThreshold( |
58 | "pragma-vectorize-scev-check-threshold" , cl::init(Val: 128), cl::Hidden, |
59 | cl::desc("The maximum number of SCEV checks allowed with a " |
60 | "vectorize(enable) pragma" )); |
61 | |
62 | static cl::opt<LoopVectorizeHints::ScalableForceKind> |
63 | ForceScalableVectorization( |
64 | "scalable-vectorization" , cl::init(Val: LoopVectorizeHints::SK_Unspecified), |
65 | cl::Hidden, |
66 | cl::desc("Control whether the compiler can use scalable vectors to " |
67 | "vectorize a loop" ), |
68 | cl::values( |
69 | clEnumValN(LoopVectorizeHints::SK_FixedWidthOnly, "off" , |
70 | "Scalable vectorization is disabled." ), |
71 | clEnumValN( |
72 | LoopVectorizeHints::SK_PreferScalable, "preferred" , |
73 | "Scalable vectorization is available and favored when the " |
74 | "cost is inconclusive." ), |
75 | clEnumValN( |
76 | LoopVectorizeHints::SK_PreferScalable, "on" , |
77 | "Scalable vectorization is available and favored when the " |
78 | "cost is inconclusive." ))); |
79 | |
80 | static cl::opt<bool> EnableHistogramVectorization( |
81 | "enable-histogram-loop-vectorization" , cl::init(Val: false), cl::Hidden, |
82 | cl::desc("Enables autovectorization of some loops containing histograms" )); |
83 | |
84 | /// Maximum vectorization interleave count. |
85 | static const unsigned MaxInterleaveFactor = 16; |
86 | |
87 | namespace llvm { |
88 | |
89 | bool LoopVectorizeHints::Hint::validate(unsigned Val) { |
90 | switch (Kind) { |
91 | case HK_WIDTH: |
92 | return isPowerOf2_32(Value: Val) && Val <= VectorizerParams::MaxVectorWidth; |
93 | case HK_INTERLEAVE: |
94 | return isPowerOf2_32(Value: Val) && Val <= MaxInterleaveFactor; |
95 | case HK_FORCE: |
96 | return (Val <= 1); |
97 | case HK_ISVECTORIZED: |
98 | case HK_PREDICATE: |
99 | case HK_SCALABLE: |
100 | return (Val == 0 || Val == 1); |
101 | } |
102 | return false; |
103 | } |
104 | |
105 | LoopVectorizeHints::(const Loop *L, |
106 | bool InterleaveOnlyWhenForced, |
107 | OptimizationRemarkEmitter &ORE, |
108 | const TargetTransformInfo *TTI) |
109 | : Width("vectorize.width" , VectorizerParams::VectorizationFactor, HK_WIDTH), |
110 | Interleave("interleave.count" , InterleaveOnlyWhenForced, HK_INTERLEAVE), |
111 | Force("vectorize.enable" , FK_Undefined, HK_FORCE), |
112 | IsVectorized("isvectorized" , 0, HK_ISVECTORIZED), |
113 | Predicate("vectorize.predicate.enable" , FK_Undefined, HK_PREDICATE), |
114 | Scalable("vectorize.scalable.enable" , SK_Unspecified, HK_SCALABLE), |
115 | TheLoop(L), ORE(ORE) { |
116 | // Populate values with existing loop metadata. |
117 | getHintsFromMetadata(); |
118 | |
119 | // force-vector-interleave overrides DisableInterleaving. |
120 | if (VectorizerParams::isInterleaveForced()) |
121 | Interleave.Value = VectorizerParams::VectorizationInterleave; |
122 | |
123 | // If the metadata doesn't explicitly specify whether to enable scalable |
124 | // vectorization, then decide based on the following criteria (increasing |
125 | // level of priority): |
126 | // - Target default |
127 | // - Metadata width |
128 | // - Force option (always overrides) |
129 | if ((LoopVectorizeHints::ScalableForceKind)Scalable.Value == SK_Unspecified) { |
130 | if (TTI) |
131 | Scalable.Value = TTI->enableScalableVectorization() ? SK_PreferScalable |
132 | : SK_FixedWidthOnly; |
133 | |
134 | if (Width.Value) |
135 | // If the width is set, but the metadata says nothing about the scalable |
136 | // property, then assume it concerns only a fixed-width UserVF. |
137 | // If width is not set, the flag takes precedence. |
138 | Scalable.Value = SK_FixedWidthOnly; |
139 | } |
140 | |
141 | // If the flag is set to force any use of scalable vectors, override the loop |
142 | // hints. |
143 | if (ForceScalableVectorization.getValue() != |
144 | LoopVectorizeHints::SK_Unspecified) |
145 | Scalable.Value = ForceScalableVectorization.getValue(); |
146 | |
147 | // Scalable vectorization is disabled if no preference is specified. |
148 | if ((LoopVectorizeHints::ScalableForceKind)Scalable.Value == SK_Unspecified) |
149 | Scalable.Value = SK_FixedWidthOnly; |
150 | |
151 | if (IsVectorized.Value != 1) |
152 | // If the vectorization width and interleaving count are both 1 then |
153 | // consider the loop to have been already vectorized because there's |
154 | // nothing more that we can do. |
155 | IsVectorized.Value = |
156 | getWidth() == ElementCount::getFixed(MinVal: 1) && getInterleave() == 1; |
157 | LLVM_DEBUG(if (InterleaveOnlyWhenForced && getInterleave() == 1) dbgs() |
158 | << "LV: Interleaving disabled by the pass manager\n" ); |
159 | } |
160 | |
161 | void LoopVectorizeHints::setAlreadyVectorized() { |
162 | LLVMContext &Context = TheLoop->getHeader()->getContext(); |
163 | |
164 | MDNode *IsVectorizedMD = MDNode::get( |
165 | Context, |
166 | MDs: {MDString::get(Context, Str: "llvm.loop.isvectorized" ), |
167 | ConstantAsMetadata::get(C: ConstantInt::get(Context, V: APInt(32, 1)))}); |
168 | MDNode *LoopID = TheLoop->getLoopID(); |
169 | MDNode *NewLoopID = |
170 | makePostTransformationMetadata(Context, OrigLoopID: LoopID, |
171 | RemovePrefixes: {Twine(Prefix(), "vectorize." ).str(), |
172 | Twine(Prefix(), "interleave." ).str()}, |
173 | AddAttrs: {IsVectorizedMD}); |
174 | TheLoop->setLoopID(NewLoopID); |
175 | |
176 | // Update internal cache. |
177 | IsVectorized.Value = 1; |
178 | } |
179 | |
180 | bool LoopVectorizeHints::allowVectorization( |
181 | Function *F, Loop *L, bool VectorizeOnlyWhenForced) const { |
182 | if (getForce() == LoopVectorizeHints::FK_Disabled) { |
183 | LLVM_DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n" ); |
184 | emitRemarkWithHints(); |
185 | return false; |
186 | } |
187 | |
188 | if (VectorizeOnlyWhenForced && getForce() != LoopVectorizeHints::FK_Enabled) { |
189 | LLVM_DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n" ); |
190 | emitRemarkWithHints(); |
191 | return false; |
192 | } |
193 | |
194 | if (getIsVectorized() == 1) { |
195 | LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n" ); |
196 | // FIXME: Add interleave.disable metadata. This will allow |
197 | // vectorize.disable to be used without disabling the pass and errors |
198 | // to differentiate between disabled vectorization and a width of 1. |
199 | ORE.emit(RemarkBuilder: [&]() { |
200 | return OptimizationRemarkAnalysis(vectorizeAnalysisPassName(), |
201 | "AllDisabled" , L->getStartLoc(), |
202 | L->getHeader()) |
203 | << "loop not vectorized: vectorization and interleaving are " |
204 | "explicitly disabled, or the loop has already been " |
205 | "vectorized" ; |
206 | }); |
207 | return false; |
208 | } |
209 | |
210 | return true; |
211 | } |
212 | |
213 | void LoopVectorizeHints::() const { |
214 | using namespace ore; |
215 | |
216 | ORE.emit(RemarkBuilder: [&]() { |
217 | if (Force.Value == LoopVectorizeHints::FK_Disabled) |
218 | return OptimizationRemarkMissed(LV_NAME, "MissedExplicitlyDisabled" , |
219 | TheLoop->getStartLoc(), |
220 | TheLoop->getHeader()) |
221 | << "loop not vectorized: vectorization is explicitly disabled" ; |
222 | |
223 | OptimizationRemarkMissed R(LV_NAME, "MissedDetails" , TheLoop->getStartLoc(), |
224 | TheLoop->getHeader()); |
225 | R << "loop not vectorized" ; |
226 | if (Force.Value == LoopVectorizeHints::FK_Enabled) { |
227 | R << " (Force=" << NV("Force" , true); |
228 | if (Width.Value != 0) |
229 | R << ", Vector Width=" << NV("VectorWidth" , getWidth()); |
230 | if (getInterleave() != 0) |
231 | R << ", Interleave Count=" << NV("InterleaveCount" , getInterleave()); |
232 | R << ")" ; |
233 | } |
234 | return R; |
235 | }); |
236 | } |
237 | |
238 | const char *LoopVectorizeHints::vectorizeAnalysisPassName() const { |
239 | if (getWidth() == ElementCount::getFixed(MinVal: 1)) |
240 | return LV_NAME; |
241 | if (getForce() == LoopVectorizeHints::FK_Disabled) |
242 | return LV_NAME; |
243 | if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth().isZero()) |
244 | return LV_NAME; |
245 | return OptimizationRemarkAnalysis::AlwaysPrint; |
246 | } |
247 | |
248 | bool LoopVectorizeHints::allowReordering() const { |
249 | // Allow the vectorizer to change the order of operations if enabling |
250 | // loop hints are provided |
251 | ElementCount EC = getWidth(); |
252 | return HintsAllowReordering && |
253 | (getForce() == LoopVectorizeHints::FK_Enabled || |
254 | EC.getKnownMinValue() > 1); |
255 | } |
256 | |
257 | void LoopVectorizeHints::getHintsFromMetadata() { |
258 | MDNode *LoopID = TheLoop->getLoopID(); |
259 | if (!LoopID) |
260 | return; |
261 | |
262 | // First operand should refer to the loop id itself. |
263 | assert(LoopID->getNumOperands() > 0 && "requires at least one operand" ); |
264 | assert(LoopID->getOperand(0) == LoopID && "invalid loop id" ); |
265 | |
266 | for (const MDOperand &MDO : llvm::drop_begin(RangeOrContainer: LoopID->operands())) { |
267 | const MDString *S = nullptr; |
268 | SmallVector<Metadata *, 4> Args; |
269 | |
270 | // The expected hint is either a MDString or a MDNode with the first |
271 | // operand a MDString. |
272 | if (const MDNode *MD = dyn_cast<MDNode>(Val: MDO)) { |
273 | if (!MD || MD->getNumOperands() == 0) |
274 | continue; |
275 | S = dyn_cast<MDString>(Val: MD->getOperand(I: 0)); |
276 | for (unsigned Idx = 1; Idx < MD->getNumOperands(); ++Idx) |
277 | Args.push_back(Elt: MD->getOperand(I: Idx)); |
278 | } else { |
279 | S = dyn_cast<MDString>(Val: MDO); |
280 | assert(Args.size() == 0 && "too many arguments for MDString" ); |
281 | } |
282 | |
283 | if (!S) |
284 | continue; |
285 | |
286 | // Check if the hint starts with the loop metadata prefix. |
287 | StringRef Name = S->getString(); |
288 | if (Args.size() == 1) |
289 | setHint(Name, Arg: Args[0]); |
290 | } |
291 | } |
292 | |
293 | void LoopVectorizeHints::setHint(StringRef Name, Metadata *Arg) { |
294 | if (!Name.starts_with(Prefix: Prefix())) |
295 | return; |
296 | Name = Name.substr(Start: Prefix().size(), N: StringRef::npos); |
297 | |
298 | const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(MD&: Arg); |
299 | if (!C) |
300 | return; |
301 | unsigned Val = C->getZExtValue(); |
302 | |
303 | Hint *Hints[] = {&Width, &Interleave, &Force, |
304 | &IsVectorized, &Predicate, &Scalable}; |
305 | for (auto *H : Hints) { |
306 | if (Name == H->Name) { |
307 | if (H->validate(Val)) |
308 | H->Value = Val; |
309 | else |
310 | LLVM_DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n" ); |
311 | break; |
312 | } |
313 | } |
314 | } |
315 | |
316 | // Return true if the inner loop \p Lp is uniform with regard to the outer loop |
317 | // \p OuterLp (i.e., if the outer loop is vectorized, all the vector lanes |
318 | // executing the inner loop will execute the same iterations). This check is |
319 | // very constrained for now but it will be relaxed in the future. \p Lp is |
320 | // considered uniform if it meets all the following conditions: |
321 | // 1) it has a canonical IV (starting from 0 and with stride 1), |
322 | // 2) its latch terminator is a conditional branch and, |
323 | // 3) its latch condition is a compare instruction whose operands are the |
324 | // canonical IV and an OuterLp invariant. |
325 | // This check doesn't take into account the uniformity of other conditions not |
326 | // related to the loop latch because they don't affect the loop uniformity. |
327 | // |
328 | // NOTE: We decided to keep all these checks and its associated documentation |
329 | // together so that we can easily have a picture of the current supported loop |
330 | // nests. However, some of the current checks don't depend on \p OuterLp and |
331 | // would be redundantly executed for each \p Lp if we invoked this function for |
332 | // different candidate outer loops. This is not the case for now because we |
333 | // don't currently have the infrastructure to evaluate multiple candidate outer |
334 | // loops and \p OuterLp will be a fixed parameter while we only support explicit |
335 | // outer loop vectorization. It's also very likely that these checks go away |
336 | // before introducing the aforementioned infrastructure. However, if this is not |
337 | // the case, we should move the \p OuterLp independent checks to a separate |
338 | // function that is only executed once for each \p Lp. |
339 | static bool isUniformLoop(Loop *Lp, Loop *OuterLp) { |
340 | assert(Lp->getLoopLatch() && "Expected loop with a single latch." ); |
341 | |
342 | // If Lp is the outer loop, it's uniform by definition. |
343 | if (Lp == OuterLp) |
344 | return true; |
345 | assert(OuterLp->contains(Lp) && "OuterLp must contain Lp." ); |
346 | |
347 | // 1. |
348 | PHINode *IV = Lp->getCanonicalInductionVariable(); |
349 | if (!IV) { |
350 | LLVM_DEBUG(dbgs() << "LV: Canonical IV not found.\n" ); |
351 | return false; |
352 | } |
353 | |
354 | // 2. |
355 | BasicBlock *Latch = Lp->getLoopLatch(); |
356 | auto *LatchBr = dyn_cast<BranchInst>(Val: Latch->getTerminator()); |
357 | if (!LatchBr || LatchBr->isUnconditional()) { |
358 | LLVM_DEBUG(dbgs() << "LV: Unsupported loop latch branch.\n" ); |
359 | return false; |
360 | } |
361 | |
362 | // 3. |
363 | auto *LatchCmp = dyn_cast<CmpInst>(Val: LatchBr->getCondition()); |
364 | if (!LatchCmp) { |
365 | LLVM_DEBUG( |
366 | dbgs() << "LV: Loop latch condition is not a compare instruction.\n" ); |
367 | return false; |
368 | } |
369 | |
370 | Value *CondOp0 = LatchCmp->getOperand(i_nocapture: 0); |
371 | Value *CondOp1 = LatchCmp->getOperand(i_nocapture: 1); |
372 | Value *IVUpdate = IV->getIncomingValueForBlock(BB: Latch); |
373 | if (!(CondOp0 == IVUpdate && OuterLp->isLoopInvariant(V: CondOp1)) && |
374 | !(CondOp1 == IVUpdate && OuterLp->isLoopInvariant(V: CondOp0))) { |
375 | LLVM_DEBUG(dbgs() << "LV: Loop latch condition is not uniform.\n" ); |
376 | return false; |
377 | } |
378 | |
379 | return true; |
380 | } |
381 | |
382 | // Return true if \p Lp and all its nested loops are uniform with regard to \p |
383 | // OuterLp. |
384 | static bool isUniformLoopNest(Loop *Lp, Loop *OuterLp) { |
385 | if (!isUniformLoop(Lp, OuterLp)) |
386 | return false; |
387 | |
388 | // Check if nested loops are uniform. |
389 | for (Loop *SubLp : *Lp) |
390 | if (!isUniformLoopNest(Lp: SubLp, OuterLp)) |
391 | return false; |
392 | |
393 | return true; |
394 | } |
395 | |
396 | static IntegerType *getInductionIntegerTy(const DataLayout &DL, Type *Ty) { |
397 | assert(Ty->isIntOrPtrTy() && "Expected integer or pointer type" ); |
398 | |
399 | if (Ty->isPointerTy()) |
400 | return DL.getIntPtrType(C&: Ty->getContext(), AddressSpace: Ty->getPointerAddressSpace()); |
401 | |
402 | // It is possible that char's or short's overflow when we ask for the loop's |
403 | // trip count, work around this by changing the type size. |
404 | if (Ty->getScalarSizeInBits() < 32) |
405 | return Type::getInt32Ty(C&: Ty->getContext()); |
406 | |
407 | return cast<IntegerType>(Val: Ty); |
408 | } |
409 | |
410 | static IntegerType *getWiderInductionTy(const DataLayout &DL, Type *Ty0, |
411 | Type *Ty1) { |
412 | IntegerType *TyA = getInductionIntegerTy(DL, Ty: Ty0); |
413 | IntegerType *TyB = getInductionIntegerTy(DL, Ty: Ty1); |
414 | return TyA->getScalarSizeInBits() > TyB->getScalarSizeInBits() ? TyA : TyB; |
415 | } |
416 | |
417 | /// Check that the instruction has outside loop users and is not an |
418 | /// identified reduction variable. |
419 | static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst, |
420 | SmallPtrSetImpl<Value *> &AllowedExit) { |
421 | // Reductions, Inductions and non-header phis are allowed to have exit users. All |
422 | // other instructions must not have external users. |
423 | if (!AllowedExit.count(Ptr: Inst)) |
424 | // Check that all of the users of the loop are inside the BB. |
425 | for (User *U : Inst->users()) { |
426 | Instruction *UI = cast<Instruction>(Val: U); |
427 | // This user may be a reduction exit value. |
428 | if (!TheLoop->contains(Inst: UI)) { |
429 | LLVM_DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n'); |
430 | return true; |
431 | } |
432 | } |
433 | return false; |
434 | } |
435 | |
436 | /// Returns true if A and B have same pointer operands or same SCEVs addresses |
437 | static bool storeToSameAddress(ScalarEvolution *SE, StoreInst *A, |
438 | StoreInst *B) { |
439 | // Compare store |
440 | if (A == B) |
441 | return true; |
442 | |
443 | // Otherwise Compare pointers |
444 | Value *APtr = A->getPointerOperand(); |
445 | Value *BPtr = B->getPointerOperand(); |
446 | if (APtr == BPtr) |
447 | return true; |
448 | |
449 | // Otherwise compare address SCEVs |
450 | return SE->getSCEV(V: APtr) == SE->getSCEV(V: BPtr); |
451 | } |
452 | |
453 | int LoopVectorizationLegality::isConsecutivePtr(Type *AccessTy, |
454 | Value *Ptr) const { |
455 | // FIXME: Currently, the set of symbolic strides is sometimes queried before |
456 | // it's collected. This happens from canVectorizeWithIfConvert, when the |
457 | // pointer is checked to reference consecutive elements suitable for a |
458 | // masked access. |
459 | const auto &Strides = |
460 | LAI ? LAI->getSymbolicStrides() : DenseMap<Value *, const SCEV *>(); |
461 | |
462 | bool CanAddPredicate = !llvm::shouldOptimizeForSize( |
463 | BB: TheLoop->getHeader(), PSI, BFI, QueryType: PGSOQueryType::IRPass); |
464 | int Stride = getPtrStride(PSE, AccessTy, Ptr, Lp: TheLoop, StridesMap: Strides, |
465 | Assume: CanAddPredicate, ShouldCheckWrap: false).value_or(u: 0); |
466 | if (Stride == 1 || Stride == -1) |
467 | return Stride; |
468 | return 0; |
469 | } |
470 | |
471 | bool LoopVectorizationLegality::isInvariant(Value *V) const { |
472 | return LAI->isInvariant(V); |
473 | } |
474 | |
475 | namespace { |
476 | /// A rewriter to build the SCEVs for each of the VF lanes in the expected |
477 | /// vectorized loop, which can then be compared to detect their uniformity. This |
478 | /// is done by replacing the AddRec SCEVs of the original scalar loop (TheLoop) |
479 | /// with new AddRecs where the step is multiplied by StepMultiplier and Offset * |
480 | /// Step is added. Also checks if all sub-expressions are analyzable w.r.t. |
481 | /// uniformity. |
482 | class SCEVAddRecForUniformityRewriter |
483 | : public SCEVRewriteVisitor<SCEVAddRecForUniformityRewriter> { |
484 | /// Multiplier to be applied to the step of AddRecs in TheLoop. |
485 | unsigned StepMultiplier; |
486 | |
487 | /// Offset to be added to the AddRecs in TheLoop. |
488 | unsigned Offset; |
489 | |
490 | /// Loop for which to rewrite AddRecsFor. |
491 | Loop *TheLoop; |
492 | |
493 | /// Is any sub-expressions not analyzable w.r.t. uniformity? |
494 | bool CannotAnalyze = false; |
495 | |
496 | bool canAnalyze() const { return !CannotAnalyze; } |
497 | |
498 | public: |
499 | SCEVAddRecForUniformityRewriter(ScalarEvolution &SE, unsigned StepMultiplier, |
500 | unsigned Offset, Loop *TheLoop) |
501 | : SCEVRewriteVisitor(SE), StepMultiplier(StepMultiplier), Offset(Offset), |
502 | TheLoop(TheLoop) {} |
503 | |
504 | const SCEV *visitAddRecExpr(const SCEVAddRecExpr *Expr) { |
505 | assert(Expr->getLoop() == TheLoop && |
506 | "addrec outside of TheLoop must be invariant and should have been " |
507 | "handled earlier" ); |
508 | // Build a new AddRec by multiplying the step by StepMultiplier and |
509 | // incrementing the start by Offset * step. |
510 | Type *Ty = Expr->getType(); |
511 | const SCEV *Step = Expr->getStepRecurrence(SE); |
512 | if (!SE.isLoopInvariant(S: Step, L: TheLoop)) { |
513 | CannotAnalyze = true; |
514 | return Expr; |
515 | } |
516 | const SCEV *NewStep = |
517 | SE.getMulExpr(LHS: Step, RHS: SE.getConstant(Ty, V: StepMultiplier)); |
518 | const SCEV *ScaledOffset = SE.getMulExpr(LHS: Step, RHS: SE.getConstant(Ty, V: Offset)); |
519 | const SCEV *NewStart = SE.getAddExpr(LHS: Expr->getStart(), RHS: ScaledOffset); |
520 | return SE.getAddRecExpr(Start: NewStart, Step: NewStep, L: TheLoop, Flags: SCEV::FlagAnyWrap); |
521 | } |
522 | |
523 | const SCEV *visit(const SCEV *S) { |
524 | if (CannotAnalyze || SE.isLoopInvariant(S, L: TheLoop)) |
525 | return S; |
526 | return SCEVRewriteVisitor<SCEVAddRecForUniformityRewriter>::visit(S); |
527 | } |
528 | |
529 | const SCEV *visitUnknown(const SCEVUnknown *S) { |
530 | if (SE.isLoopInvariant(S, L: TheLoop)) |
531 | return S; |
532 | // The value could vary across iterations. |
533 | CannotAnalyze = true; |
534 | return S; |
535 | } |
536 | |
537 | const SCEV *visitCouldNotCompute(const SCEVCouldNotCompute *S) { |
538 | // Could not analyze the expression. |
539 | CannotAnalyze = true; |
540 | return S; |
541 | } |
542 | |
543 | static const SCEV *rewrite(const SCEV *S, ScalarEvolution &SE, |
544 | unsigned StepMultiplier, unsigned Offset, |
545 | Loop *TheLoop) { |
546 | /// Bail out if the expression does not contain an UDiv expression. |
547 | /// Uniform values which are not loop invariant require operations to strip |
548 | /// out the lowest bits. For now just look for UDivs and use it to avoid |
549 | /// re-writing UDIV-free expressions for other lanes to limit compile time. |
550 | if (!SCEVExprContains(Root: S, |
551 | Pred: [](const SCEV *S) { return isa<SCEVUDivExpr>(Val: S); })) |
552 | return SE.getCouldNotCompute(); |
553 | |
554 | SCEVAddRecForUniformityRewriter Rewriter(SE, StepMultiplier, Offset, |
555 | TheLoop); |
556 | const SCEV *Result = Rewriter.visit(S); |
557 | |
558 | if (Rewriter.canAnalyze()) |
559 | return Result; |
560 | return SE.getCouldNotCompute(); |
561 | } |
562 | }; |
563 | |
564 | } // namespace |
565 | |
566 | bool LoopVectorizationLegality::isUniform(Value *V, ElementCount VF) const { |
567 | if (isInvariant(V)) |
568 | return true; |
569 | if (VF.isScalable()) |
570 | return false; |
571 | if (VF.isScalar()) |
572 | return true; |
573 | |
574 | // Since we rely on SCEV for uniformity, if the type is not SCEVable, it is |
575 | // never considered uniform. |
576 | auto *SE = PSE.getSE(); |
577 | if (!SE->isSCEVable(Ty: V->getType())) |
578 | return false; |
579 | const SCEV *S = SE->getSCEV(V); |
580 | |
581 | // Rewrite AddRecs in TheLoop to step by VF and check if the expression for |
582 | // lane 0 matches the expressions for all other lanes. |
583 | unsigned FixedVF = VF.getKnownMinValue(); |
584 | const SCEV *FirstLaneExpr = |
585 | SCEVAddRecForUniformityRewriter::rewrite(S, SE&: *SE, StepMultiplier: FixedVF, Offset: 0, TheLoop); |
586 | if (isa<SCEVCouldNotCompute>(Val: FirstLaneExpr)) |
587 | return false; |
588 | |
589 | // Make sure the expressions for lanes FixedVF-1..1 match the expression for |
590 | // lane 0. We check lanes in reverse order for compile-time, as frequently |
591 | // checking the last lane is sufficient to rule out uniformity. |
592 | return all_of(Range: reverse(C: seq<unsigned>(Begin: 1, End: FixedVF)), P: [&](unsigned I) { |
593 | const SCEV *IthLaneExpr = |
594 | SCEVAddRecForUniformityRewriter::rewrite(S, SE&: *SE, StepMultiplier: FixedVF, Offset: I, TheLoop); |
595 | return FirstLaneExpr == IthLaneExpr; |
596 | }); |
597 | } |
598 | |
599 | bool LoopVectorizationLegality::isUniformMemOp(Instruction &I, |
600 | ElementCount VF) const { |
601 | Value *Ptr = getLoadStorePointerOperand(V: &I); |
602 | if (!Ptr) |
603 | return false; |
604 | // Note: There's nothing inherent which prevents predicated loads and |
605 | // stores from being uniform. The current lowering simply doesn't handle |
606 | // it; in particular, the cost model distinguishes scatter/gather from |
607 | // scalar w/predication, and we currently rely on the scalar path. |
608 | return isUniform(V: Ptr, VF) && !blockNeedsPredication(BB: I.getParent()); |
609 | } |
610 | |
611 | bool LoopVectorizationLegality::canVectorizeOuterLoop() { |
612 | assert(!TheLoop->isInnermost() && "We are not vectorizing an outer loop." ); |
613 | // Store the result and return it at the end instead of exiting early, in case |
614 | // allowExtraAnalysis is used to report multiple reasons for not vectorizing. |
615 | bool Result = true; |
616 | bool = ORE->allowExtraAnalysis(DEBUG_TYPE); |
617 | |
618 | for (BasicBlock *BB : TheLoop->blocks()) { |
619 | // Check whether the BB terminator is a BranchInst. Any other terminator is |
620 | // not supported yet. |
621 | auto *Br = dyn_cast<BranchInst>(Val: BB->getTerminator()); |
622 | if (!Br) { |
623 | reportVectorizationFailure(DebugMsg: "Unsupported basic block terminator" , |
624 | OREMsg: "loop control flow is not understood by vectorizer" , |
625 | ORETag: "CFGNotUnderstood" , ORE, TheLoop); |
626 | if (DoExtraAnalysis) |
627 | Result = false; |
628 | else |
629 | return false; |
630 | } |
631 | |
632 | // Check whether the BranchInst is a supported one. Only unconditional |
633 | // branches, conditional branches with an outer loop invariant condition or |
634 | // backedges are supported. |
635 | // FIXME: We skip these checks when VPlan predication is enabled as we |
636 | // want to allow divergent branches. This whole check will be removed |
637 | // once VPlan predication is on by default. |
638 | if (Br && Br->isConditional() && |
639 | !TheLoop->isLoopInvariant(V: Br->getCondition()) && |
640 | !LI->isLoopHeader(BB: Br->getSuccessor(i: 0)) && |
641 | !LI->isLoopHeader(BB: Br->getSuccessor(i: 1))) { |
642 | reportVectorizationFailure(DebugMsg: "Unsupported conditional branch" , |
643 | OREMsg: "loop control flow is not understood by vectorizer" , |
644 | ORETag: "CFGNotUnderstood" , ORE, TheLoop); |
645 | if (DoExtraAnalysis) |
646 | Result = false; |
647 | else |
648 | return false; |
649 | } |
650 | } |
651 | |
652 | // Check whether inner loops are uniform. At this point, we only support |
653 | // simple outer loops scenarios with uniform nested loops. |
654 | if (!isUniformLoopNest(Lp: TheLoop /*loop nest*/, |
655 | OuterLp: TheLoop /*context outer loop*/)) { |
656 | reportVectorizationFailure(DebugMsg: "Outer loop contains divergent loops" , |
657 | OREMsg: "loop control flow is not understood by vectorizer" , |
658 | ORETag: "CFGNotUnderstood" , ORE, TheLoop); |
659 | if (DoExtraAnalysis) |
660 | Result = false; |
661 | else |
662 | return false; |
663 | } |
664 | |
665 | // Check whether we are able to set up outer loop induction. |
666 | if (!setupOuterLoopInductions()) { |
667 | reportVectorizationFailure(DebugMsg: "Unsupported outer loop Phi(s)" , |
668 | ORETag: "UnsupportedPhi" , ORE, TheLoop); |
669 | if (DoExtraAnalysis) |
670 | Result = false; |
671 | else |
672 | return false; |
673 | } |
674 | |
675 | return Result; |
676 | } |
677 | |
678 | void LoopVectorizationLegality::addInductionPhi( |
679 | PHINode *Phi, const InductionDescriptor &ID, |
680 | SmallPtrSetImpl<Value *> &AllowedExit) { |
681 | Inductions[Phi] = ID; |
682 | |
683 | // In case this induction also comes with casts that we know we can ignore |
684 | // in the vectorized loop body, record them here. All casts could be recorded |
685 | // here for ignoring, but suffices to record only the first (as it is the |
686 | // only one that may bw used outside the cast sequence). |
687 | const SmallVectorImpl<Instruction *> &Casts = ID.getCastInsts(); |
688 | if (!Casts.empty()) |
689 | InductionCastsToIgnore.insert(Ptr: *Casts.begin()); |
690 | |
691 | Type *PhiTy = Phi->getType(); |
692 | const DataLayout &DL = Phi->getDataLayout(); |
693 | |
694 | assert((PhiTy->isIntOrPtrTy() || PhiTy->isFloatingPointTy()) && |
695 | "Expected int, ptr, or FP induction phi type" ); |
696 | |
697 | // Get the widest type. |
698 | if (PhiTy->isIntOrPtrTy()) { |
699 | if (!WidestIndTy) |
700 | WidestIndTy = getInductionIntegerTy(DL, Ty: PhiTy); |
701 | else |
702 | WidestIndTy = getWiderInductionTy(DL, Ty0: PhiTy, Ty1: WidestIndTy); |
703 | } |
704 | |
705 | // Int inductions are special because we only allow one IV. |
706 | if (ID.getKind() == InductionDescriptor::IK_IntInduction && |
707 | ID.getConstIntStepValue() && ID.getConstIntStepValue()->isOne() && |
708 | isa<Constant>(Val: ID.getStartValue()) && |
709 | cast<Constant>(Val: ID.getStartValue())->isNullValue()) { |
710 | |
711 | // Use the phi node with the widest type as induction. Use the last |
712 | // one if there are multiple (no good reason for doing this other |
713 | // than it is expedient). We've checked that it begins at zero and |
714 | // steps by one, so this is a canonical induction variable. |
715 | if (!PrimaryInduction || PhiTy == WidestIndTy) |
716 | PrimaryInduction = Phi; |
717 | } |
718 | |
719 | // Both the PHI node itself, and the "post-increment" value feeding |
720 | // back into the PHI node may have external users. |
721 | // We can allow those uses, except if the SCEVs we have for them rely |
722 | // on predicates that only hold within the loop, since allowing the exit |
723 | // currently means re-using this SCEV outside the loop (see PR33706 for more |
724 | // details). |
725 | if (PSE.getPredicate().isAlwaysTrue()) { |
726 | AllowedExit.insert(Ptr: Phi); |
727 | AllowedExit.insert(Ptr: Phi->getIncomingValueForBlock(BB: TheLoop->getLoopLatch())); |
728 | } |
729 | |
730 | LLVM_DEBUG(dbgs() << "LV: Found an induction variable.\n" ); |
731 | } |
732 | |
733 | bool LoopVectorizationLegality::setupOuterLoopInductions() { |
734 | BasicBlock * = TheLoop->getHeader(); |
735 | |
736 | // Returns true if a given Phi is a supported induction. |
737 | auto IsSupportedPhi = [&](PHINode &Phi) -> bool { |
738 | InductionDescriptor ID; |
739 | if (InductionDescriptor::isInductionPHI(Phi: &Phi, L: TheLoop, PSE, D&: ID) && |
740 | ID.getKind() == InductionDescriptor::IK_IntInduction) { |
741 | addInductionPhi(Phi: &Phi, ID, AllowedExit); |
742 | return true; |
743 | } |
744 | // Bail out for any Phi in the outer loop header that is not a supported |
745 | // induction. |
746 | LLVM_DEBUG( |
747 | dbgs() << "LV: Found unsupported PHI for outer loop vectorization.\n" ); |
748 | return false; |
749 | }; |
750 | |
751 | return llvm::all_of(Range: Header->phis(), P: IsSupportedPhi); |
752 | } |
753 | |
754 | /// Checks if a function is scalarizable according to the TLI, in |
755 | /// the sense that it should be vectorized and then expanded in |
756 | /// multiple scalar calls. This is represented in the |
757 | /// TLI via mappings that do not specify a vector name, as in the |
758 | /// following example: |
759 | /// |
760 | /// const VecDesc VecIntrinsics[] = { |
761 | /// {"llvm.phx.abs.i32", "", 4} |
762 | /// }; |
763 | static bool isTLIScalarize(const TargetLibraryInfo &TLI, const CallInst &CI) { |
764 | const StringRef ScalarName = CI.getCalledFunction()->getName(); |
765 | bool Scalarize = TLI.isFunctionVectorizable(F: ScalarName); |
766 | // Check that all known VFs are not associated to a vector |
767 | // function, i.e. the vector name is emty. |
768 | if (Scalarize) { |
769 | ElementCount WidestFixedVF, WidestScalableVF; |
770 | TLI.getWidestVF(ScalarF: ScalarName, FixedVF&: WidestFixedVF, ScalableVF&: WidestScalableVF); |
771 | for (ElementCount VF = ElementCount::getFixed(MinVal: 2); |
772 | ElementCount::isKnownLE(LHS: VF, RHS: WidestFixedVF); VF *= 2) |
773 | Scalarize &= !TLI.isFunctionVectorizable(F: ScalarName, VF); |
774 | for (ElementCount VF = ElementCount::getScalable(MinVal: 1); |
775 | ElementCount::isKnownLE(LHS: VF, RHS: WidestScalableVF); VF *= 2) |
776 | Scalarize &= !TLI.isFunctionVectorizable(F: ScalarName, VF); |
777 | assert((WidestScalableVF.isZero() || !Scalarize) && |
778 | "Caller may decide to scalarize a variant using a scalable VF" ); |
779 | } |
780 | return Scalarize; |
781 | } |
782 | |
783 | /// Returns true if the call return type `Ty` can be widened by the loop |
784 | /// vectorizer. |
785 | static bool canWidenCallReturnType(Type *Ty) { |
786 | auto *StructTy = dyn_cast<StructType>(Val: Ty); |
787 | // TODO: Remove the homogeneous types restriction. This is just an initial |
788 | // simplification. When we want to support things like the overflow intrinsics |
789 | // we will have to lift this restriction. |
790 | if (StructTy && !StructTy->containsHomogeneousTypes()) |
791 | return false; |
792 | return canVectorizeTy(Ty: StructTy); |
793 | } |
794 | |
795 | bool LoopVectorizationLegality::canVectorizeInstrs() { |
796 | BasicBlock * = TheLoop->getHeader(); |
797 | |
798 | // For each block in the loop. |
799 | for (BasicBlock *BB : TheLoop->blocks()) { |
800 | // Scan the instructions in the block and look for hazards. |
801 | for (Instruction &I : *BB) { |
802 | if (auto *Phi = dyn_cast<PHINode>(Val: &I)) { |
803 | Type *PhiTy = Phi->getType(); |
804 | // Check that this PHI type is allowed. |
805 | if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() && |
806 | !PhiTy->isPointerTy()) { |
807 | reportVectorizationFailure(DebugMsg: "Found a non-int non-pointer PHI" , |
808 | OREMsg: "loop control flow is not understood by vectorizer" , |
809 | ORETag: "CFGNotUnderstood" , ORE, TheLoop); |
810 | return false; |
811 | } |
812 | |
813 | // If this PHINode is not in the header block, then we know that we |
814 | // can convert it to select during if-conversion. No need to check if |
815 | // the PHIs in this block are induction or reduction variables. |
816 | if (BB != Header) { |
817 | // Non-header phi nodes that have outside uses can be vectorized. Add |
818 | // them to the list of allowed exits. |
819 | // Unsafe cyclic dependencies with header phis are identified during |
820 | // legalization for reduction, induction and fixed order |
821 | // recurrences. |
822 | AllowedExit.insert(Ptr: &I); |
823 | continue; |
824 | } |
825 | |
826 | // We only allow if-converted PHIs with exactly two incoming values. |
827 | if (Phi->getNumIncomingValues() != 2) { |
828 | reportVectorizationFailure(DebugMsg: "Found an invalid PHI" , |
829 | OREMsg: "loop control flow is not understood by vectorizer" , |
830 | ORETag: "CFGNotUnderstood" , ORE, TheLoop, I: Phi); |
831 | return false; |
832 | } |
833 | |
834 | RecurrenceDescriptor RedDes; |
835 | if (RecurrenceDescriptor::isReductionPHI(Phi, TheLoop, RedDes, DB, AC, |
836 | DT, SE: PSE.getSE())) { |
837 | Requirements->addExactFPMathInst(I: RedDes.getExactFPMathInst()); |
838 | AllowedExit.insert(Ptr: RedDes.getLoopExitInstr()); |
839 | Reductions[Phi] = RedDes; |
840 | continue; |
841 | } |
842 | |
843 | // We prevent matching non-constant strided pointer IVS to preserve |
844 | // historical vectorizer behavior after a generalization of the |
845 | // IVDescriptor code. The intent is to remove this check, but we |
846 | // have to fix issues around code quality for such loops first. |
847 | auto IsDisallowedStridedPointerInduction = |
848 | [](const InductionDescriptor &ID) { |
849 | if (AllowStridedPointerIVs) |
850 | return false; |
851 | return ID.getKind() == InductionDescriptor::IK_PtrInduction && |
852 | ID.getConstIntStepValue() == nullptr; |
853 | }; |
854 | |
855 | // TODO: Instead of recording the AllowedExit, it would be good to |
856 | // record the complementary set: NotAllowedExit. These include (but may |
857 | // not be limited to): |
858 | // 1. Reduction phis as they represent the one-before-last value, which |
859 | // is not available when vectorized |
860 | // 2. Induction phis and increment when SCEV predicates cannot be used |
861 | // outside the loop - see addInductionPhi |
862 | // 3. Non-Phis with outside uses when SCEV predicates cannot be used |
863 | // outside the loop - see call to hasOutsideLoopUser in the non-phi |
864 | // handling below |
865 | // 4. FixedOrderRecurrence phis that can possibly be handled by |
866 | // extraction. |
867 | // By recording these, we can then reason about ways to vectorize each |
868 | // of these NotAllowedExit. |
869 | InductionDescriptor ID; |
870 | if (InductionDescriptor::isInductionPHI(Phi, L: TheLoop, PSE, D&: ID) && |
871 | !IsDisallowedStridedPointerInduction(ID)) { |
872 | addInductionPhi(Phi, ID, AllowedExit); |
873 | Requirements->addExactFPMathInst(I: ID.getExactFPMathInst()); |
874 | continue; |
875 | } |
876 | |
877 | if (RecurrenceDescriptor::isFixedOrderRecurrence(Phi, TheLoop, DT)) { |
878 | AllowedExit.insert(Ptr: Phi); |
879 | FixedOrderRecurrences.insert(Ptr: Phi); |
880 | continue; |
881 | } |
882 | |
883 | // As a last resort, coerce the PHI to a AddRec expression |
884 | // and re-try classifying it a an induction PHI. |
885 | if (InductionDescriptor::isInductionPHI(Phi, L: TheLoop, PSE, D&: ID, Assume: true) && |
886 | !IsDisallowedStridedPointerInduction(ID)) { |
887 | addInductionPhi(Phi, ID, AllowedExit); |
888 | continue; |
889 | } |
890 | |
891 | reportVectorizationFailure(DebugMsg: "Found an unidentified PHI" , |
892 | OREMsg: "value that could not be identified as " |
893 | "reduction is used outside the loop" , |
894 | ORETag: "NonReductionValueUsedOutsideLoop" , ORE, TheLoop, I: Phi); |
895 | return false; |
896 | } // end of PHI handling |
897 | |
898 | // We handle calls that: |
899 | // * Have a mapping to an IR intrinsic. |
900 | // * Have a vector version available. |
901 | auto *CI = dyn_cast<CallInst>(Val: &I); |
902 | |
903 | if (CI && !getVectorIntrinsicIDForCall(CI, TLI) && |
904 | !(CI->getCalledFunction() && TLI && |
905 | (!VFDatabase::getMappings(CI: *CI).empty() || |
906 | isTLIScalarize(TLI: *TLI, CI: *CI)))) { |
907 | // If the call is a recognized math libary call, it is likely that |
908 | // we can vectorize it given loosened floating-point constraints. |
909 | LibFunc Func; |
910 | bool IsMathLibCall = |
911 | TLI && CI->getCalledFunction() && |
912 | CI->getType()->isFloatingPointTy() && |
913 | TLI->getLibFunc(funcName: CI->getCalledFunction()->getName(), F&: Func) && |
914 | TLI->hasOptimizedCodeGen(F: Func); |
915 | |
916 | if (IsMathLibCall) { |
917 | // TODO: Ideally, we should not use clang-specific language here, |
918 | // but it's hard to provide meaningful yet generic advice. |
919 | // Also, should this be guarded by allowExtraAnalysis() and/or be part |
920 | // of the returned info from isFunctionVectorizable()? |
921 | reportVectorizationFailure( |
922 | DebugMsg: "Found a non-intrinsic callsite" , |
923 | OREMsg: "library call cannot be vectorized. " |
924 | "Try compiling with -fno-math-errno, -ffast-math, " |
925 | "or similar flags" , |
926 | ORETag: "CantVectorizeLibcall" , ORE, TheLoop, I: CI); |
927 | } else { |
928 | reportVectorizationFailure(DebugMsg: "Found a non-intrinsic callsite" , |
929 | OREMsg: "call instruction cannot be vectorized" , |
930 | ORETag: "CantVectorizeLibcall" , ORE, TheLoop, I: CI); |
931 | } |
932 | return false; |
933 | } |
934 | |
935 | // Some intrinsics have scalar arguments and should be same in order for |
936 | // them to be vectorized (i.e. loop invariant). |
937 | if (CI) { |
938 | auto *SE = PSE.getSE(); |
939 | Intrinsic::ID IntrinID = getVectorIntrinsicIDForCall(CI, TLI); |
940 | for (unsigned Idx = 0; Idx < CI->arg_size(); ++Idx) |
941 | if (isVectorIntrinsicWithScalarOpAtArg(ID: IntrinID, ScalarOpdIdx: Idx, TTI)) { |
942 | if (!SE->isLoopInvariant(S: PSE.getSCEV(V: CI->getOperand(i_nocapture: Idx)), |
943 | L: TheLoop)) { |
944 | reportVectorizationFailure(DebugMsg: "Found unvectorizable intrinsic" , |
945 | OREMsg: "intrinsic instruction cannot be vectorized" , |
946 | ORETag: "CantVectorizeIntrinsic" , ORE, TheLoop, I: CI); |
947 | return false; |
948 | } |
949 | } |
950 | } |
951 | |
952 | // If we found a vectorized variant of a function, note that so LV can |
953 | // make better decisions about maximum VF. |
954 | if (CI && !VFDatabase::getMappings(CI: *CI).empty()) |
955 | VecCallVariantsFound = true; |
956 | |
957 | auto CanWidenInstructionTy = [](Instruction const &Inst) { |
958 | Type *InstTy = Inst.getType(); |
959 | if (!isa<StructType>(Val: InstTy)) |
960 | return canVectorizeTy(Ty: InstTy); |
961 | |
962 | // For now, we only recognize struct values returned from calls where |
963 | // all users are extractvalue as vectorizable. All element types of the |
964 | // struct must be types that can be widened. |
965 | return isa<CallInst>(Val: Inst) && canWidenCallReturnType(Ty: InstTy) && |
966 | all_of(Range: Inst.users(), P: IsaPred<ExtractValueInst>); |
967 | }; |
968 | |
969 | // Check that the instruction return type is vectorizable. |
970 | // We can't vectorize casts from vector type to scalar type. |
971 | // Also, we can't vectorize extractelement instructions. |
972 | if (!CanWidenInstructionTy(I) || |
973 | (isa<CastInst>(Val: I) && |
974 | !VectorType::isValidElementType(ElemTy: I.getOperand(i: 0)->getType())) || |
975 | isa<ExtractElementInst>(Val: I)) { |
976 | reportVectorizationFailure(DebugMsg: "Found unvectorizable type" , |
977 | OREMsg: "instruction return type cannot be vectorized" , |
978 | ORETag: "CantVectorizeInstructionReturnType" , ORE, TheLoop, I: &I); |
979 | return false; |
980 | } |
981 | |
982 | // Check that the stored type is vectorizable. |
983 | if (auto *ST = dyn_cast<StoreInst>(Val: &I)) { |
984 | Type *T = ST->getValueOperand()->getType(); |
985 | if (!VectorType::isValidElementType(ElemTy: T)) { |
986 | reportVectorizationFailure(DebugMsg: "Store instruction cannot be vectorized" , |
987 | ORETag: "CantVectorizeStore" , ORE, TheLoop, I: ST); |
988 | return false; |
989 | } |
990 | |
991 | // For nontemporal stores, check that a nontemporal vector version is |
992 | // supported on the target. |
993 | if (ST->getMetadata(KindID: LLVMContext::MD_nontemporal)) { |
994 | // Arbitrarily try a vector of 2 elements. |
995 | auto *VecTy = FixedVectorType::get(ElementType: T, /*NumElts=*/2); |
996 | assert(VecTy && "did not find vectorized version of stored type" ); |
997 | if (!TTI->isLegalNTStore(DataType: VecTy, Alignment: ST->getAlign())) { |
998 | reportVectorizationFailure( |
999 | DebugMsg: "nontemporal store instruction cannot be vectorized" , |
1000 | ORETag: "CantVectorizeNontemporalStore" , ORE, TheLoop, I: ST); |
1001 | return false; |
1002 | } |
1003 | } |
1004 | |
1005 | } else if (auto *LD = dyn_cast<LoadInst>(Val: &I)) { |
1006 | if (LD->getMetadata(KindID: LLVMContext::MD_nontemporal)) { |
1007 | // For nontemporal loads, check that a nontemporal vector version is |
1008 | // supported on the target (arbitrarily try a vector of 2 elements). |
1009 | auto *VecTy = FixedVectorType::get(ElementType: I.getType(), /*NumElts=*/2); |
1010 | assert(VecTy && "did not find vectorized version of load type" ); |
1011 | if (!TTI->isLegalNTLoad(DataType: VecTy, Alignment: LD->getAlign())) { |
1012 | reportVectorizationFailure( |
1013 | DebugMsg: "nontemporal load instruction cannot be vectorized" , |
1014 | ORETag: "CantVectorizeNontemporalLoad" , ORE, TheLoop, I: LD); |
1015 | return false; |
1016 | } |
1017 | } |
1018 | |
1019 | // FP instructions can allow unsafe algebra, thus vectorizable by |
1020 | // non-IEEE-754 compliant SIMD units. |
1021 | // This applies to floating-point math operations and calls, not memory |
1022 | // operations, shuffles, or casts, as they don't change precision or |
1023 | // semantics. |
1024 | } else if (I.getType()->isFloatingPointTy() && (CI || I.isBinaryOp()) && |
1025 | !I.isFast()) { |
1026 | LLVM_DEBUG(dbgs() << "LV: Found FP op with unsafe algebra.\n" ); |
1027 | Hints->setPotentiallyUnsafe(); |
1028 | } |
1029 | |
1030 | // Reduction instructions are allowed to have exit users. |
1031 | // All other instructions must not have external users. |
1032 | if (hasOutsideLoopUser(TheLoop, Inst: &I, AllowedExit)) { |
1033 | // We can safely vectorize loops where instructions within the loop are |
1034 | // used outside the loop only if the SCEV predicates within the loop is |
1035 | // same as outside the loop. Allowing the exit means reusing the SCEV |
1036 | // outside the loop. |
1037 | if (PSE.getPredicate().isAlwaysTrue()) { |
1038 | AllowedExit.insert(Ptr: &I); |
1039 | continue; |
1040 | } |
1041 | reportVectorizationFailure(DebugMsg: "Value cannot be used outside the loop" , |
1042 | ORETag: "ValueUsedOutsideLoop" , ORE, TheLoop, I: &I); |
1043 | return false; |
1044 | } |
1045 | } // next instr. |
1046 | } |
1047 | |
1048 | if (!PrimaryInduction) { |
1049 | if (Inductions.empty()) { |
1050 | reportVectorizationFailure(DebugMsg: "Did not find one integer induction var" , |
1051 | OREMsg: "loop induction variable could not be identified" , |
1052 | ORETag: "NoInductionVariable" , ORE, TheLoop); |
1053 | return false; |
1054 | } |
1055 | if (!WidestIndTy) { |
1056 | reportVectorizationFailure(DebugMsg: "Did not find one integer induction var" , |
1057 | OREMsg: "integer loop induction variable could not be identified" , |
1058 | ORETag: "NoIntegerInductionVariable" , ORE, TheLoop); |
1059 | return false; |
1060 | } |
1061 | LLVM_DEBUG(dbgs() << "LV: Did not find one integer induction var.\n" ); |
1062 | } |
1063 | |
1064 | // Now we know the widest induction type, check if our found induction |
1065 | // is the same size. If it's not, unset it here and InnerLoopVectorizer |
1066 | // will create another. |
1067 | if (PrimaryInduction && WidestIndTy != PrimaryInduction->getType()) |
1068 | PrimaryInduction = nullptr; |
1069 | |
1070 | return true; |
1071 | } |
1072 | |
1073 | /// Find histogram operations that match high-level code in loops: |
1074 | /// \code |
1075 | /// buckets[indices[i]]+=step; |
1076 | /// \endcode |
1077 | /// |
1078 | /// It matches a pattern starting from \p HSt, which Stores to the 'buckets' |
1079 | /// array the computed histogram. It uses a BinOp to sum all counts, storing |
1080 | /// them using a loop-variant index Load from the 'indices' input array. |
1081 | /// |
1082 | /// On successful matches it updates the STATISTIC 'HistogramsDetected', |
1083 | /// regardless of hardware support. When there is support, it additionally |
1084 | /// stores the BinOp/Load pairs in \p HistogramCounts, as well the pointers |
1085 | /// used to update histogram in \p HistogramPtrs. |
1086 | static bool findHistogram(LoadInst *LI, StoreInst *HSt, Loop *TheLoop, |
1087 | const PredicatedScalarEvolution &PSE, |
1088 | SmallVectorImpl<HistogramInfo> &Histograms) { |
1089 | |
1090 | // Store value must come from a Binary Operation. |
1091 | Instruction *HPtrInstr = nullptr; |
1092 | BinaryOperator *HBinOp = nullptr; |
1093 | if (!match(V: HSt, P: m_Store(ValueOp: m_BinOp(I&: HBinOp), PointerOp: m_Instruction(I&: HPtrInstr)))) |
1094 | return false; |
1095 | |
1096 | // BinOp must be an Add or a Sub modifying the bucket value by a |
1097 | // loop invariant amount. |
1098 | // FIXME: We assume the loop invariant term is on the RHS. |
1099 | // Fine for an immediate/constant, but maybe not a generic value? |
1100 | Value *HIncVal = nullptr; |
1101 | if (!match(V: HBinOp, P: m_Add(L: m_Load(Op: m_Specific(V: HPtrInstr)), R: m_Value(V&: HIncVal))) && |
1102 | !match(V: HBinOp, P: m_Sub(L: m_Load(Op: m_Specific(V: HPtrInstr)), R: m_Value(V&: HIncVal)))) |
1103 | return false; |
1104 | |
1105 | // Make sure the increment value is loop invariant. |
1106 | if (!TheLoop->isLoopInvariant(V: HIncVal)) |
1107 | return false; |
1108 | |
1109 | // The address to store is calculated through a GEP Instruction. |
1110 | GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Val: HPtrInstr); |
1111 | if (!GEP) |
1112 | return false; |
1113 | |
1114 | // Restrict address calculation to constant indices except for the last term. |
1115 | Value *HIdx = nullptr; |
1116 | for (Value *Index : GEP->indices()) { |
1117 | if (HIdx) |
1118 | return false; |
1119 | if (!isa<ConstantInt>(Val: Index)) |
1120 | HIdx = Index; |
1121 | } |
1122 | |
1123 | if (!HIdx) |
1124 | return false; |
1125 | |
1126 | // Check that the index is calculated by loading from another array. Ignore |
1127 | // any extensions. |
1128 | // FIXME: Support indices from other sources than a linear load from memory? |
1129 | // We're currently trying to match an operation looping over an array |
1130 | // of indices, but there could be additional levels of indirection |
1131 | // in place, or possibly some additional calculation to form the index |
1132 | // from the loaded data. |
1133 | Value *VPtrVal; |
1134 | if (!match(V: HIdx, P: m_ZExtOrSExtOrSelf(Op: m_Load(Op: m_Value(V&: VPtrVal))))) |
1135 | return false; |
1136 | |
1137 | // Make sure the index address varies in this loop, not an outer loop. |
1138 | const auto *AR = dyn_cast<SCEVAddRecExpr>(Val: PSE.getSE()->getSCEV(V: VPtrVal)); |
1139 | if (!AR || AR->getLoop() != TheLoop) |
1140 | return false; |
1141 | |
1142 | // Ensure we'll have the same mask by checking that all parts of the histogram |
1143 | // (gather load, update, scatter store) are in the same block. |
1144 | LoadInst *IndexedLoad = cast<LoadInst>(Val: HBinOp->getOperand(i_nocapture: 0)); |
1145 | BasicBlock *LdBB = IndexedLoad->getParent(); |
1146 | if (LdBB != HBinOp->getParent() || LdBB != HSt->getParent()) |
1147 | return false; |
1148 | |
1149 | LLVM_DEBUG(dbgs() << "LV: Found histogram for: " << *HSt << "\n" ); |
1150 | |
1151 | // Store the operations that make up the histogram. |
1152 | Histograms.emplace_back(Args&: IndexedLoad, Args&: HBinOp, Args&: HSt); |
1153 | return true; |
1154 | } |
1155 | |
1156 | bool LoopVectorizationLegality::canVectorizeIndirectUnsafeDependences() { |
1157 | // For now, we only support an IndirectUnsafe dependency that calculates |
1158 | // a histogram |
1159 | if (!EnableHistogramVectorization) |
1160 | return false; |
1161 | |
1162 | // Find a single IndirectUnsafe dependency. |
1163 | const MemoryDepChecker::Dependence *IUDep = nullptr; |
1164 | const MemoryDepChecker &DepChecker = LAI->getDepChecker(); |
1165 | const auto *Deps = DepChecker.getDependences(); |
1166 | // If there were too many dependences, LAA abandons recording them. We can't |
1167 | // proceed safely if we don't know what the dependences are. |
1168 | if (!Deps) |
1169 | return false; |
1170 | |
1171 | for (const MemoryDepChecker::Dependence &Dep : *Deps) { |
1172 | // Ignore dependencies that are either known to be safe or can be |
1173 | // checked at runtime. |
1174 | if (MemoryDepChecker::Dependence::isSafeForVectorization(Type: Dep.Type) != |
1175 | MemoryDepChecker::VectorizationSafetyStatus::Unsafe) |
1176 | continue; |
1177 | |
1178 | // We're only interested in IndirectUnsafe dependencies here, where the |
1179 | // address might come from a load from memory. We also only want to handle |
1180 | // one such dependency, at least for now. |
1181 | if (Dep.Type != MemoryDepChecker::Dependence::IndirectUnsafe || IUDep) |
1182 | return false; |
1183 | |
1184 | IUDep = &Dep; |
1185 | } |
1186 | if (!IUDep) |
1187 | return false; |
1188 | |
1189 | // For now only normal loads and stores are supported. |
1190 | LoadInst *LI = dyn_cast<LoadInst>(Val: IUDep->getSource(DepChecker)); |
1191 | StoreInst *SI = dyn_cast<StoreInst>(Val: IUDep->getDestination(DepChecker)); |
1192 | |
1193 | if (!LI || !SI) |
1194 | return false; |
1195 | |
1196 | LLVM_DEBUG(dbgs() << "LV: Checking for a histogram on: " << *SI << "\n" ); |
1197 | return findHistogram(LI, HSt: SI, TheLoop, PSE: LAI->getPSE(), Histograms); |
1198 | } |
1199 | |
1200 | bool LoopVectorizationLegality::canVectorizeMemory() { |
1201 | LAI = &LAIs.getInfo(L&: *TheLoop); |
1202 | const OptimizationRemarkAnalysis *LAR = LAI->getReport(); |
1203 | if (LAR) { |
1204 | ORE->emit(RemarkBuilder: [&]() { |
1205 | return OptimizationRemarkAnalysis(Hints->vectorizeAnalysisPassName(), |
1206 | "loop not vectorized: " , *LAR); |
1207 | }); |
1208 | } |
1209 | |
1210 | if (!LAI->canVectorizeMemory()) |
1211 | return canVectorizeIndirectUnsafeDependences(); |
1212 | |
1213 | if (LAI->hasLoadStoreDependenceInvolvingLoopInvariantAddress()) { |
1214 | reportVectorizationFailure(DebugMsg: "We don't allow storing to uniform addresses" , |
1215 | OREMsg: "write to a loop invariant address could not " |
1216 | "be vectorized" , |
1217 | ORETag: "CantVectorizeStoreToLoopInvariantAddress" , ORE, |
1218 | TheLoop); |
1219 | return false; |
1220 | } |
1221 | |
1222 | // We can vectorize stores to invariant address when final reduction value is |
1223 | // guaranteed to be stored at the end of the loop. Also, if decision to |
1224 | // vectorize loop is made, runtime checks are added so as to make sure that |
1225 | // invariant address won't alias with any other objects. |
1226 | if (!LAI->getStoresToInvariantAddresses().empty()) { |
1227 | // For each invariant address, check if last stored value is unconditional |
1228 | // and the address is not calculated inside the loop. |
1229 | for (StoreInst *SI : LAI->getStoresToInvariantAddresses()) { |
1230 | if (!isInvariantStoreOfReduction(SI)) |
1231 | continue; |
1232 | |
1233 | if (blockNeedsPredication(BB: SI->getParent())) { |
1234 | reportVectorizationFailure( |
1235 | DebugMsg: "We don't allow storing to uniform addresses" , |
1236 | OREMsg: "write of conditional recurring variant value to a loop " |
1237 | "invariant address could not be vectorized" , |
1238 | ORETag: "CantVectorizeStoreToLoopInvariantAddress" , ORE, TheLoop); |
1239 | return false; |
1240 | } |
1241 | |
1242 | // Invariant address should be defined outside of loop. LICM pass usually |
1243 | // makes sure it happens, but in rare cases it does not, we do not want |
1244 | // to overcomplicate vectorization to support this case. |
1245 | if (Instruction *Ptr = dyn_cast<Instruction>(Val: SI->getPointerOperand())) { |
1246 | if (TheLoop->contains(Inst: Ptr)) { |
1247 | reportVectorizationFailure( |
1248 | DebugMsg: "Invariant address is calculated inside the loop" , |
1249 | OREMsg: "write to a loop invariant address could not " |
1250 | "be vectorized" , |
1251 | ORETag: "CantVectorizeStoreToLoopInvariantAddress" , ORE, TheLoop); |
1252 | return false; |
1253 | } |
1254 | } |
1255 | } |
1256 | |
1257 | if (LAI->hasStoreStoreDependenceInvolvingLoopInvariantAddress()) { |
1258 | // For each invariant address, check its last stored value is the result |
1259 | // of one of our reductions. |
1260 | // |
1261 | // We do not check if dependence with loads exists because that is already |
1262 | // checked via hasLoadStoreDependenceInvolvingLoopInvariantAddress. |
1263 | ScalarEvolution *SE = PSE.getSE(); |
1264 | SmallVector<StoreInst *, 4> UnhandledStores; |
1265 | for (StoreInst *SI : LAI->getStoresToInvariantAddresses()) { |
1266 | if (isInvariantStoreOfReduction(SI)) { |
1267 | // Earlier stores to this address are effectively deadcode. |
1268 | // With opaque pointers it is possible for one pointer to be used with |
1269 | // different sizes of stored values: |
1270 | // store i32 0, ptr %x |
1271 | // store i8 0, ptr %x |
1272 | // The latest store doesn't complitely overwrite the first one in the |
1273 | // example. That is why we have to make sure that types of stored |
1274 | // values are same. |
1275 | // TODO: Check that bitwidth of unhandled store is smaller then the |
1276 | // one that overwrites it and add a test. |
1277 | erase_if(C&: UnhandledStores, P: [SE, SI](StoreInst *I) { |
1278 | return storeToSameAddress(SE, A: SI, B: I) && |
1279 | I->getValueOperand()->getType() == |
1280 | SI->getValueOperand()->getType(); |
1281 | }); |
1282 | continue; |
1283 | } |
1284 | UnhandledStores.push_back(Elt: SI); |
1285 | } |
1286 | |
1287 | bool IsOK = UnhandledStores.empty(); |
1288 | // TODO: we should also validate against InvariantMemSets. |
1289 | if (!IsOK) { |
1290 | reportVectorizationFailure( |
1291 | DebugMsg: "We don't allow storing to uniform addresses" , |
1292 | OREMsg: "write to a loop invariant address could not " |
1293 | "be vectorized" , |
1294 | ORETag: "CantVectorizeStoreToLoopInvariantAddress" , ORE, TheLoop); |
1295 | return false; |
1296 | } |
1297 | } |
1298 | } |
1299 | |
1300 | PSE.addPredicate(Pred: LAI->getPSE().getPredicate()); |
1301 | return true; |
1302 | } |
1303 | |
1304 | bool LoopVectorizationLegality::canVectorizeFPMath( |
1305 | bool EnableStrictReductions) { |
1306 | |
1307 | // First check if there is any ExactFP math or if we allow reassociations |
1308 | if (!Requirements->getExactFPInst() || Hints->allowReordering()) |
1309 | return true; |
1310 | |
1311 | // If the above is false, we have ExactFPMath & do not allow reordering. |
1312 | // If the EnableStrictReductions flag is set, first check if we have any |
1313 | // Exact FP induction vars, which we cannot vectorize. |
1314 | if (!EnableStrictReductions || |
1315 | any_of(Range: getInductionVars(), P: [&](auto &Induction) -> bool { |
1316 | InductionDescriptor IndDesc = Induction.second; |
1317 | return IndDesc.getExactFPMathInst(); |
1318 | })) |
1319 | return false; |
1320 | |
1321 | // We can now only vectorize if all reductions with Exact FP math also |
1322 | // have the isOrdered flag set, which indicates that we can move the |
1323 | // reduction operations in-loop. |
1324 | return (all_of(Range: getReductionVars(), P: [&](auto &Reduction) -> bool { |
1325 | const RecurrenceDescriptor &RdxDesc = Reduction.second; |
1326 | return !RdxDesc.hasExactFPMath() || RdxDesc.isOrdered(); |
1327 | })); |
1328 | } |
1329 | |
1330 | bool LoopVectorizationLegality::isInvariantStoreOfReduction(StoreInst *SI) { |
1331 | return any_of(Range: getReductionVars(), P: [&](auto &Reduction) -> bool { |
1332 | const RecurrenceDescriptor &RdxDesc = Reduction.second; |
1333 | return RdxDesc.IntermediateStore == SI; |
1334 | }); |
1335 | } |
1336 | |
1337 | bool LoopVectorizationLegality::isInvariantAddressOfReduction(Value *V) { |
1338 | return any_of(Range: getReductionVars(), P: [&](auto &Reduction) -> bool { |
1339 | const RecurrenceDescriptor &RdxDesc = Reduction.second; |
1340 | if (!RdxDesc.IntermediateStore) |
1341 | return false; |
1342 | |
1343 | ScalarEvolution *SE = PSE.getSE(); |
1344 | Value *InvariantAddress = RdxDesc.IntermediateStore->getPointerOperand(); |
1345 | return V == InvariantAddress || |
1346 | SE->getSCEV(V) == SE->getSCEV(V: InvariantAddress); |
1347 | }); |
1348 | } |
1349 | |
1350 | bool LoopVectorizationLegality::isInductionPhi(const Value *V) const { |
1351 | Value *In0 = const_cast<Value *>(V); |
1352 | PHINode *PN = dyn_cast_or_null<PHINode>(Val: In0); |
1353 | if (!PN) |
1354 | return false; |
1355 | |
1356 | return Inductions.count(Key: PN); |
1357 | } |
1358 | |
1359 | const InductionDescriptor * |
1360 | LoopVectorizationLegality::getIntOrFpInductionDescriptor(PHINode *Phi) const { |
1361 | if (!isInductionPhi(V: Phi)) |
1362 | return nullptr; |
1363 | auto &ID = getInductionVars().find(Key: Phi)->second; |
1364 | if (ID.getKind() == InductionDescriptor::IK_IntInduction || |
1365 | ID.getKind() == InductionDescriptor::IK_FpInduction) |
1366 | return &ID; |
1367 | return nullptr; |
1368 | } |
1369 | |
1370 | const InductionDescriptor * |
1371 | LoopVectorizationLegality::getPointerInductionDescriptor(PHINode *Phi) const { |
1372 | if (!isInductionPhi(V: Phi)) |
1373 | return nullptr; |
1374 | auto &ID = getInductionVars().find(Key: Phi)->second; |
1375 | if (ID.getKind() == InductionDescriptor::IK_PtrInduction) |
1376 | return &ID; |
1377 | return nullptr; |
1378 | } |
1379 | |
1380 | bool LoopVectorizationLegality::isCastedInductionVariable( |
1381 | const Value *V) const { |
1382 | auto *Inst = dyn_cast<Instruction>(Val: V); |
1383 | return (Inst && InductionCastsToIgnore.count(Ptr: Inst)); |
1384 | } |
1385 | |
1386 | bool LoopVectorizationLegality::isInductionVariable(const Value *V) const { |
1387 | return isInductionPhi(V) || isCastedInductionVariable(V); |
1388 | } |
1389 | |
1390 | bool LoopVectorizationLegality::isFixedOrderRecurrence( |
1391 | const PHINode *Phi) const { |
1392 | return FixedOrderRecurrences.count(Ptr: Phi); |
1393 | } |
1394 | |
1395 | bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) const { |
1396 | // When vectorizing early exits, create predicates for the latch block only. |
1397 | // The early exiting block must be a direct predecessor of the latch at the |
1398 | // moment. |
1399 | BasicBlock *Latch = TheLoop->getLoopLatch(); |
1400 | if (hasUncountableEarlyExit()) { |
1401 | assert( |
1402 | is_contained(predecessors(Latch), getUncountableEarlyExitingBlock()) && |
1403 | "Uncountable exiting block must be a direct predecessor of latch" ); |
1404 | return BB == Latch; |
1405 | } |
1406 | return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT); |
1407 | } |
1408 | |
1409 | bool LoopVectorizationLegality::blockCanBePredicated( |
1410 | BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs, |
1411 | SmallPtrSetImpl<const Instruction *> &MaskedOp) const { |
1412 | for (Instruction &I : *BB) { |
1413 | // We can predicate blocks with calls to assume, as long as we drop them in |
1414 | // case we flatten the CFG via predication. |
1415 | if (match(V: &I, P: m_Intrinsic<Intrinsic::assume>())) { |
1416 | MaskedOp.insert(Ptr: &I); |
1417 | continue; |
1418 | } |
1419 | |
1420 | // Do not let llvm.experimental.noalias.scope.decl block the vectorization. |
1421 | // TODO: there might be cases that it should block the vectorization. Let's |
1422 | // ignore those for now. |
1423 | if (isa<NoAliasScopeDeclInst>(Val: &I)) |
1424 | continue; |
1425 | |
1426 | // We can allow masked calls if there's at least one vector variant, even |
1427 | // if we end up scalarizing due to the cost model calculations. |
1428 | // TODO: Allow other calls if they have appropriate attributes... readonly |
1429 | // and argmemonly? |
1430 | if (CallInst *CI = dyn_cast<CallInst>(Val: &I)) |
1431 | if (VFDatabase::hasMaskedVariant(CI: *CI)) { |
1432 | MaskedOp.insert(Ptr: CI); |
1433 | continue; |
1434 | } |
1435 | |
1436 | // Loads are handled via masking (or speculated if safe to do so.) |
1437 | if (auto *LI = dyn_cast<LoadInst>(Val: &I)) { |
1438 | if (!SafePtrs.count(Ptr: LI->getPointerOperand())) |
1439 | MaskedOp.insert(Ptr: LI); |
1440 | continue; |
1441 | } |
1442 | |
1443 | // Predicated store requires some form of masking: |
1444 | // 1) masked store HW instruction, |
1445 | // 2) emulation via load-blend-store (only if safe and legal to do so, |
1446 | // be aware on the race conditions), or |
1447 | // 3) element-by-element predicate check and scalar store. |
1448 | if (auto *SI = dyn_cast<StoreInst>(Val: &I)) { |
1449 | MaskedOp.insert(Ptr: SI); |
1450 | continue; |
1451 | } |
1452 | |
1453 | if (I.mayReadFromMemory() || I.mayWriteToMemory() || I.mayThrow()) |
1454 | return false; |
1455 | } |
1456 | |
1457 | return true; |
1458 | } |
1459 | |
1460 | bool LoopVectorizationLegality::canVectorizeWithIfConvert() { |
1461 | if (!EnableIfConversion) { |
1462 | reportVectorizationFailure(DebugMsg: "If-conversion is disabled" , |
1463 | ORETag: "IfConversionDisabled" , ORE, TheLoop); |
1464 | return false; |
1465 | } |
1466 | |
1467 | assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable" ); |
1468 | |
1469 | // A list of pointers which are known to be dereferenceable within scope of |
1470 | // the loop body for each iteration of the loop which executes. That is, |
1471 | // the memory pointed to can be dereferenced (with the access size implied by |
1472 | // the value's type) unconditionally within the loop header without |
1473 | // introducing a new fault. |
1474 | SmallPtrSet<Value *, 8> SafePointers; |
1475 | |
1476 | // Collect safe addresses. |
1477 | for (BasicBlock *BB : TheLoop->blocks()) { |
1478 | if (!blockNeedsPredication(BB)) { |
1479 | for (Instruction &I : *BB) |
1480 | if (auto *Ptr = getLoadStorePointerOperand(V: &I)) |
1481 | SafePointers.insert(Ptr); |
1482 | continue; |
1483 | } |
1484 | |
1485 | // For a block which requires predication, a address may be safe to access |
1486 | // in the loop w/o predication if we can prove dereferenceability facts |
1487 | // sufficient to ensure it'll never fault within the loop. For the moment, |
1488 | // we restrict this to loads; stores are more complicated due to |
1489 | // concurrency restrictions. |
1490 | ScalarEvolution &SE = *PSE.getSE(); |
1491 | SmallVector<const SCEVPredicate *, 4> Predicates; |
1492 | for (Instruction &I : *BB) { |
1493 | LoadInst *LI = dyn_cast<LoadInst>(Val: &I); |
1494 | |
1495 | // Make sure we can execute all computations feeding into Ptr in the loop |
1496 | // w/o triggering UB and that none of the out-of-loop operands are poison. |
1497 | // We do not need to check if operations inside the loop can produce |
1498 | // poison due to flags (e.g. due to an inbounds GEP going out of bounds), |
1499 | // because flags will be dropped when executing them unconditionally. |
1500 | // TODO: Results could be improved by considering poison-propagation |
1501 | // properties of visited ops. |
1502 | auto CanSpeculatePointerOp = [this](Value *Ptr) { |
1503 | SmallVector<Value *> Worklist = {Ptr}; |
1504 | SmallPtrSet<Value *, 4> Visited; |
1505 | while (!Worklist.empty()) { |
1506 | Value *CurrV = Worklist.pop_back_val(); |
1507 | if (!Visited.insert(Ptr: CurrV).second) |
1508 | continue; |
1509 | |
1510 | auto *CurrI = dyn_cast<Instruction>(Val: CurrV); |
1511 | if (!CurrI || !TheLoop->contains(Inst: CurrI)) { |
1512 | // If operands from outside the loop may be poison then Ptr may also |
1513 | // be poison. |
1514 | if (!isGuaranteedNotToBePoison(V: CurrV, AC, |
1515 | CtxI: TheLoop->getLoopPredecessor() |
1516 | ->getTerminator() |
1517 | ->getIterator())) |
1518 | return false; |
1519 | continue; |
1520 | } |
1521 | |
1522 | // A loaded value may be poison, independent of any flags. |
1523 | if (isa<LoadInst>(Val: CurrI) && !isGuaranteedNotToBePoison(V: CurrV, AC)) |
1524 | return false; |
1525 | |
1526 | // For other ops, assume poison can only be introduced via flags, |
1527 | // which can be dropped. |
1528 | if (!isa<PHINode>(Val: CurrI) && !isSafeToSpeculativelyExecute(I: CurrI)) |
1529 | return false; |
1530 | append_range(C&: Worklist, R: CurrI->operands()); |
1531 | } |
1532 | return true; |
1533 | }; |
1534 | // Pass the Predicates pointer to isDereferenceableAndAlignedInLoop so |
1535 | // that it will consider loops that need guarding by SCEV checks. The |
1536 | // vectoriser will generate these checks if we decide to vectorise. |
1537 | if (LI && !LI->getType()->isVectorTy() && !mustSuppressSpeculation(LI: *LI) && |
1538 | CanSpeculatePointerOp(LI->getPointerOperand()) && |
1539 | isDereferenceableAndAlignedInLoop(LI, L: TheLoop, SE, DT&: *DT, AC, |
1540 | Predicates: &Predicates)) |
1541 | SafePointers.insert(Ptr: LI->getPointerOperand()); |
1542 | Predicates.clear(); |
1543 | } |
1544 | } |
1545 | |
1546 | // Collect the blocks that need predication. |
1547 | for (BasicBlock *BB : TheLoop->blocks()) { |
1548 | // We support only branches and switch statements as terminators inside the |
1549 | // loop. |
1550 | if (isa<SwitchInst>(Val: BB->getTerminator())) { |
1551 | if (TheLoop->isLoopExiting(BB)) { |
1552 | reportVectorizationFailure(DebugMsg: "Loop contains an unsupported switch" , |
1553 | ORETag: "LoopContainsUnsupportedSwitch" , ORE, |
1554 | TheLoop, I: BB->getTerminator()); |
1555 | return false; |
1556 | } |
1557 | } else if (!isa<BranchInst>(Val: BB->getTerminator())) { |
1558 | reportVectorizationFailure(DebugMsg: "Loop contains an unsupported terminator" , |
1559 | ORETag: "LoopContainsUnsupportedTerminator" , ORE, |
1560 | TheLoop, I: BB->getTerminator()); |
1561 | return false; |
1562 | } |
1563 | |
1564 | // We must be able to predicate all blocks that need to be predicated. |
1565 | if (blockNeedsPredication(BB) && |
1566 | !blockCanBePredicated(BB, SafePtrs&: SafePointers, MaskedOp)) { |
1567 | reportVectorizationFailure( |
1568 | DebugMsg: "Control flow cannot be substituted for a select" , ORETag: "NoCFGForSelect" , |
1569 | ORE, TheLoop, I: BB->getTerminator()); |
1570 | return false; |
1571 | } |
1572 | } |
1573 | |
1574 | // We can if-convert this loop. |
1575 | return true; |
1576 | } |
1577 | |
1578 | // Helper function to canVectorizeLoopNestCFG. |
1579 | bool LoopVectorizationLegality::canVectorizeLoopCFG(Loop *Lp, |
1580 | bool UseVPlanNativePath) { |
1581 | assert((UseVPlanNativePath || Lp->isInnermost()) && |
1582 | "VPlan-native path is not enabled." ); |
1583 | |
1584 | // TODO: ORE should be improved to show more accurate information when an |
1585 | // outer loop can't be vectorized because a nested loop is not understood or |
1586 | // legal. Something like: "outer_loop_location: loop not vectorized: |
1587 | // (inner_loop_location) loop control flow is not understood by vectorizer". |
1588 | |
1589 | // Store the result and return it at the end instead of exiting early, in case |
1590 | // allowExtraAnalysis is used to report multiple reasons for not vectorizing. |
1591 | bool Result = true; |
1592 | bool = ORE->allowExtraAnalysis(DEBUG_TYPE); |
1593 | |
1594 | // We must have a loop in canonical form. Loops with indirectbr in them cannot |
1595 | // be canonicalized. |
1596 | if (!Lp->getLoopPreheader()) { |
1597 | reportVectorizationFailure(DebugMsg: "Loop doesn't have a legal pre-header" , |
1598 | OREMsg: "loop control flow is not understood by vectorizer" , |
1599 | ORETag: "CFGNotUnderstood" , ORE, TheLoop); |
1600 | if (DoExtraAnalysis) |
1601 | Result = false; |
1602 | else |
1603 | return false; |
1604 | } |
1605 | |
1606 | // We must have a single backedge. |
1607 | if (Lp->getNumBackEdges() != 1) { |
1608 | reportVectorizationFailure(DebugMsg: "The loop must have a single backedge" , |
1609 | OREMsg: "loop control flow is not understood by vectorizer" , |
1610 | ORETag: "CFGNotUnderstood" , ORE, TheLoop); |
1611 | if (DoExtraAnalysis) |
1612 | Result = false; |
1613 | else |
1614 | return false; |
1615 | } |
1616 | |
1617 | return Result; |
1618 | } |
1619 | |
1620 | bool LoopVectorizationLegality::canVectorizeLoopNestCFG( |
1621 | Loop *Lp, bool UseVPlanNativePath) { |
1622 | // Store the result and return it at the end instead of exiting early, in case |
1623 | // allowExtraAnalysis is used to report multiple reasons for not vectorizing. |
1624 | bool Result = true; |
1625 | bool = ORE->allowExtraAnalysis(DEBUG_TYPE); |
1626 | if (!canVectorizeLoopCFG(Lp, UseVPlanNativePath)) { |
1627 | if (DoExtraAnalysis) |
1628 | Result = false; |
1629 | else |
1630 | return false; |
1631 | } |
1632 | |
1633 | // Recursively check whether the loop control flow of nested loops is |
1634 | // understood. |
1635 | for (Loop *SubLp : *Lp) |
1636 | if (!canVectorizeLoopNestCFG(Lp: SubLp, UseVPlanNativePath)) { |
1637 | if (DoExtraAnalysis) |
1638 | Result = false; |
1639 | else |
1640 | return false; |
1641 | } |
1642 | |
1643 | return Result; |
1644 | } |
1645 | |
1646 | bool LoopVectorizationLegality::isVectorizableEarlyExitLoop() { |
1647 | BasicBlock *LatchBB = TheLoop->getLoopLatch(); |
1648 | if (!LatchBB) { |
1649 | reportVectorizationFailure(DebugMsg: "Loop does not have a latch" , |
1650 | OREMsg: "Cannot vectorize early exit loop" , |
1651 | ORETag: "NoLatchEarlyExit" , ORE, TheLoop); |
1652 | return false; |
1653 | } |
1654 | |
1655 | if (Reductions.size() || FixedOrderRecurrences.size()) { |
1656 | reportVectorizationFailure( |
1657 | DebugMsg: "Found reductions or recurrences in early-exit loop" , |
1658 | OREMsg: "Cannot vectorize early exit loop with reductions or recurrences" , |
1659 | ORETag: "RecurrencesInEarlyExitLoop" , ORE, TheLoop); |
1660 | return false; |
1661 | } |
1662 | |
1663 | SmallVector<BasicBlock *, 8> ExitingBlocks; |
1664 | TheLoop->getExitingBlocks(ExitingBlocks); |
1665 | |
1666 | // Keep a record of all the exiting blocks. |
1667 | SmallVector<const SCEVPredicate *, 4> Predicates; |
1668 | std::optional<std::pair<BasicBlock *, BasicBlock *>> SingleUncountableEdge; |
1669 | for (BasicBlock *BB : ExitingBlocks) { |
1670 | const SCEV *EC = |
1671 | PSE.getSE()->getPredicatedExitCount(L: TheLoop, ExitingBlock: BB, Predicates: &Predicates); |
1672 | if (isa<SCEVCouldNotCompute>(Val: EC)) { |
1673 | SmallVector<BasicBlock *, 2> Succs(successors(BB)); |
1674 | if (Succs.size() != 2) { |
1675 | reportVectorizationFailure( |
1676 | DebugMsg: "Early exiting block does not have exactly two successors" , |
1677 | OREMsg: "Incorrect number of successors from early exiting block" , |
1678 | ORETag: "EarlyExitTooManySuccessors" , ORE, TheLoop); |
1679 | return false; |
1680 | } |
1681 | |
1682 | BasicBlock *ExitBlock; |
1683 | if (!TheLoop->contains(BB: Succs[0])) |
1684 | ExitBlock = Succs[0]; |
1685 | else { |
1686 | assert(!TheLoop->contains(Succs[1])); |
1687 | ExitBlock = Succs[1]; |
1688 | } |
1689 | |
1690 | if (SingleUncountableEdge) { |
1691 | reportVectorizationFailure( |
1692 | DebugMsg: "Loop has too many uncountable exits" , |
1693 | OREMsg: "Cannot vectorize early exit loop with more than one early exit" , |
1694 | ORETag: "TooManyUncountableEarlyExits" , ORE, TheLoop); |
1695 | return false; |
1696 | } |
1697 | |
1698 | SingleUncountableEdge = {BB, ExitBlock}; |
1699 | } else |
1700 | CountableExitingBlocks.push_back(Elt: BB); |
1701 | } |
1702 | // We can safely ignore the predicates here because when vectorizing the loop |
1703 | // the PredicatatedScalarEvolution class will keep track of all predicates |
1704 | // for each exiting block anyway. This happens when calling |
1705 | // PSE.getSymbolicMaxBackedgeTakenCount() below. |
1706 | Predicates.clear(); |
1707 | |
1708 | if (!SingleUncountableEdge) { |
1709 | LLVM_DEBUG(dbgs() << "LV: Cound not find any uncountable exits" ); |
1710 | return false; |
1711 | } |
1712 | |
1713 | // The only supported early exit loops so far are ones where the early |
1714 | // exiting block is a unique predecessor of the latch block. |
1715 | BasicBlock *LatchPredBB = LatchBB->getUniquePredecessor(); |
1716 | if (LatchPredBB != SingleUncountableEdge->first) { |
1717 | reportVectorizationFailure(DebugMsg: "Early exit is not the latch predecessor" , |
1718 | OREMsg: "Cannot vectorize early exit loop" , |
1719 | ORETag: "EarlyExitNotLatchPredecessor" , ORE, TheLoop); |
1720 | return false; |
1721 | } |
1722 | |
1723 | // The latch block must have a countable exit. |
1724 | if (isa<SCEVCouldNotCompute>( |
1725 | Val: PSE.getSE()->getPredicatedExitCount(L: TheLoop, ExitingBlock: LatchBB, Predicates: &Predicates))) { |
1726 | reportVectorizationFailure( |
1727 | DebugMsg: "Cannot determine exact exit count for latch block" , |
1728 | OREMsg: "Cannot vectorize early exit loop" , |
1729 | ORETag: "UnknownLatchExitCountEarlyExitLoop" , ORE, TheLoop); |
1730 | return false; |
1731 | } |
1732 | assert(llvm::is_contained(CountableExitingBlocks, LatchBB) && |
1733 | "Latch block not found in list of countable exits!" ); |
1734 | |
1735 | // Check to see if there are instructions that could potentially generate |
1736 | // exceptions or have side-effects. |
1737 | auto IsSafeOperation = [](Instruction *I) -> bool { |
1738 | switch (I->getOpcode()) { |
1739 | case Instruction::Load: |
1740 | case Instruction::Store: |
1741 | case Instruction::PHI: |
1742 | case Instruction::Br: |
1743 | // These are checked separately. |
1744 | return true; |
1745 | default: |
1746 | return isSafeToSpeculativelyExecute(I); |
1747 | } |
1748 | }; |
1749 | |
1750 | for (auto *BB : TheLoop->blocks()) |
1751 | for (auto &I : *BB) { |
1752 | if (I.mayWriteToMemory()) { |
1753 | // We don't support writes to memory. |
1754 | reportVectorizationFailure( |
1755 | DebugMsg: "Writes to memory unsupported in early exit loops" , |
1756 | OREMsg: "Cannot vectorize early exit loop with writes to memory" , |
1757 | ORETag: "WritesInEarlyExitLoop" , ORE, TheLoop); |
1758 | return false; |
1759 | } else if (!IsSafeOperation(&I)) { |
1760 | reportVectorizationFailure(DebugMsg: "Early exit loop contains operations that " |
1761 | "cannot be speculatively executed" , |
1762 | ORETag: "UnsafeOperationsEarlyExitLoop" , ORE, |
1763 | TheLoop); |
1764 | return false; |
1765 | } |
1766 | } |
1767 | |
1768 | // The vectoriser cannot handle loads that occur after the early exit block. |
1769 | assert(LatchBB->getUniquePredecessor() == SingleUncountableEdge->first && |
1770 | "Expected latch predecessor to be the early exiting block" ); |
1771 | |
1772 | // TODO: Handle loops that may fault. |
1773 | Predicates.clear(); |
1774 | if (!isDereferenceableReadOnlyLoop(L: TheLoop, SE: PSE.getSE(), DT, AC, |
1775 | Predicates: &Predicates)) { |
1776 | reportVectorizationFailure( |
1777 | DebugMsg: "Loop may fault" , |
1778 | OREMsg: "Cannot vectorize potentially faulting early exit loop" , |
1779 | ORETag: "PotentiallyFaultingEarlyExitLoop" , ORE, TheLoop); |
1780 | return false; |
1781 | } |
1782 | |
1783 | [[maybe_unused]] const SCEV *SymbolicMaxBTC = |
1784 | PSE.getSymbolicMaxBackedgeTakenCount(); |
1785 | // Since we have an exact exit count for the latch and the early exit |
1786 | // dominates the latch, then this should guarantee a computed SCEV value. |
1787 | assert(!isa<SCEVCouldNotCompute>(SymbolicMaxBTC) && |
1788 | "Failed to get symbolic expression for backedge taken count" ); |
1789 | LLVM_DEBUG(dbgs() << "LV: Found an early exit loop with symbolic max " |
1790 | "backedge taken count: " |
1791 | << *SymbolicMaxBTC << '\n'); |
1792 | UncountableEdge = SingleUncountableEdge; |
1793 | return true; |
1794 | } |
1795 | |
1796 | bool LoopVectorizationLegality::canVectorize(bool UseVPlanNativePath) { |
1797 | // Store the result and return it at the end instead of exiting early, in case |
1798 | // allowExtraAnalysis is used to report multiple reasons for not vectorizing. |
1799 | bool Result = true; |
1800 | |
1801 | bool = ORE->allowExtraAnalysis(DEBUG_TYPE); |
1802 | // Check whether the loop-related control flow in the loop nest is expected by |
1803 | // vectorizer. |
1804 | if (!canVectorizeLoopNestCFG(Lp: TheLoop, UseVPlanNativePath)) { |
1805 | if (DoExtraAnalysis) { |
1806 | LLVM_DEBUG(dbgs() << "LV: legality check failed: loop nest" ); |
1807 | Result = false; |
1808 | } else { |
1809 | return false; |
1810 | } |
1811 | } |
1812 | |
1813 | // We need to have a loop header. |
1814 | LLVM_DEBUG(dbgs() << "LV: Found a loop: " << TheLoop->getHeader()->getName() |
1815 | << '\n'); |
1816 | |
1817 | // Specific checks for outer loops. We skip the remaining legal checks at this |
1818 | // point because they don't support outer loops. |
1819 | if (!TheLoop->isInnermost()) { |
1820 | assert(UseVPlanNativePath && "VPlan-native path is not enabled." ); |
1821 | |
1822 | if (!canVectorizeOuterLoop()) { |
1823 | reportVectorizationFailure(DebugMsg: "Unsupported outer loop" , |
1824 | ORETag: "UnsupportedOuterLoop" , ORE, TheLoop); |
1825 | // TODO: Implement DoExtraAnalysis when subsequent legal checks support |
1826 | // outer loops. |
1827 | return false; |
1828 | } |
1829 | |
1830 | LLVM_DEBUG(dbgs() << "LV: We can vectorize this outer loop!\n" ); |
1831 | return Result; |
1832 | } |
1833 | |
1834 | assert(TheLoop->isInnermost() && "Inner loop expected." ); |
1835 | // Check if we can if-convert non-single-bb loops. |
1836 | unsigned NumBlocks = TheLoop->getNumBlocks(); |
1837 | if (NumBlocks != 1 && !canVectorizeWithIfConvert()) { |
1838 | LLVM_DEBUG(dbgs() << "LV: Can't if-convert the loop.\n" ); |
1839 | if (DoExtraAnalysis) |
1840 | Result = false; |
1841 | else |
1842 | return false; |
1843 | } |
1844 | |
1845 | // Check if we can vectorize the instructions and CFG in this loop. |
1846 | if (!canVectorizeInstrs()) { |
1847 | LLVM_DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n" ); |
1848 | if (DoExtraAnalysis) |
1849 | Result = false; |
1850 | else |
1851 | return false; |
1852 | } |
1853 | |
1854 | if (isa<SCEVCouldNotCompute>(Val: PSE.getBackedgeTakenCount())) { |
1855 | if (TheLoop->getExitingBlock()) { |
1856 | reportVectorizationFailure(DebugMsg: "Cannot vectorize uncountable loop" , |
1857 | ORETag: "UnsupportedUncountableLoop" , ORE, TheLoop); |
1858 | if (DoExtraAnalysis) |
1859 | Result = false; |
1860 | else |
1861 | return false; |
1862 | } else { |
1863 | if (!isVectorizableEarlyExitLoop()) { |
1864 | UncountableEdge = std::nullopt; |
1865 | if (DoExtraAnalysis) |
1866 | Result = false; |
1867 | else |
1868 | return false; |
1869 | } |
1870 | } |
1871 | } |
1872 | |
1873 | // Go over each instruction and look at memory deps. |
1874 | if (!canVectorizeMemory()) { |
1875 | LLVM_DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n" ); |
1876 | if (DoExtraAnalysis) |
1877 | Result = false; |
1878 | else |
1879 | return false; |
1880 | } |
1881 | |
1882 | if (Result) { |
1883 | LLVM_DEBUG(dbgs() << "LV: We can vectorize this loop" |
1884 | << (LAI->getRuntimePointerChecking()->Need |
1885 | ? " (with a runtime bound check)" |
1886 | : "" ) |
1887 | << "!\n" ); |
1888 | } |
1889 | |
1890 | unsigned SCEVThreshold = VectorizeSCEVCheckThreshold; |
1891 | if (Hints->getForce() == LoopVectorizeHints::FK_Enabled) |
1892 | SCEVThreshold = PragmaVectorizeSCEVCheckThreshold; |
1893 | |
1894 | if (PSE.getPredicate().getComplexity() > SCEVThreshold) { |
1895 | LLVM_DEBUG(dbgs() << "LV: Vectorization not profitable " |
1896 | "due to SCEVThreshold" ); |
1897 | reportVectorizationFailure(DebugMsg: "Too many SCEV checks needed" , |
1898 | OREMsg: "Too many SCEV assumptions need to be made and checked at runtime" , |
1899 | ORETag: "TooManySCEVRunTimeChecks" , ORE, TheLoop); |
1900 | if (DoExtraAnalysis) |
1901 | Result = false; |
1902 | else |
1903 | return false; |
1904 | } |
1905 | |
1906 | // Okay! We've done all the tests. If any have failed, return false. Otherwise |
1907 | // we can vectorize, and at this point we don't have any other mem analysis |
1908 | // which may limit our maximum vectorization factor, so just return true with |
1909 | // no restrictions. |
1910 | return Result; |
1911 | } |
1912 | |
1913 | bool LoopVectorizationLegality::canFoldTailByMasking() const { |
1914 | // The only loops we can vectorize without a scalar epilogue, are loops with |
1915 | // a bottom-test and a single exiting block. We'd have to handle the fact |
1916 | // that not every instruction executes on the last iteration. This will |
1917 | // require a lane mask which varies through the vector loop body. (TODO) |
1918 | if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) { |
1919 | LLVM_DEBUG( |
1920 | dbgs() |
1921 | << "LV: Cannot fold tail by masking. Requires a singe latch exit\n" ); |
1922 | return false; |
1923 | } |
1924 | |
1925 | LLVM_DEBUG(dbgs() << "LV: checking if tail can be folded by masking.\n" ); |
1926 | |
1927 | SmallPtrSet<const Value *, 8> ReductionLiveOuts; |
1928 | |
1929 | for (const auto &Reduction : getReductionVars()) |
1930 | ReductionLiveOuts.insert(Ptr: Reduction.second.getLoopExitInstr()); |
1931 | |
1932 | // TODO: handle non-reduction outside users when tail is folded by masking. |
1933 | for (auto *AE : AllowedExit) { |
1934 | // Check that all users of allowed exit values are inside the loop or |
1935 | // are the live-out of a reduction. |
1936 | if (ReductionLiveOuts.count(Ptr: AE)) |
1937 | continue; |
1938 | for (User *U : AE->users()) { |
1939 | Instruction *UI = cast<Instruction>(Val: U); |
1940 | if (TheLoop->contains(Inst: UI)) |
1941 | continue; |
1942 | LLVM_DEBUG( |
1943 | dbgs() |
1944 | << "LV: Cannot fold tail by masking, loop has an outside user for " |
1945 | << *UI << "\n" ); |
1946 | return false; |
1947 | } |
1948 | } |
1949 | |
1950 | for (const auto &Entry : getInductionVars()) { |
1951 | PHINode *OrigPhi = Entry.first; |
1952 | for (User *U : OrigPhi->users()) { |
1953 | auto *UI = cast<Instruction>(Val: U); |
1954 | if (!TheLoop->contains(Inst: UI)) { |
1955 | LLVM_DEBUG(dbgs() << "LV: Cannot fold tail by masking, loop IV has an " |
1956 | "outside user for " |
1957 | << *UI << "\n" ); |
1958 | return false; |
1959 | } |
1960 | } |
1961 | } |
1962 | |
1963 | // The list of pointers that we can safely read and write to remains empty. |
1964 | SmallPtrSet<Value *, 8> SafePointers; |
1965 | |
1966 | // Check all blocks for predication, including those that ordinarily do not |
1967 | // need predication such as the header block. |
1968 | SmallPtrSet<const Instruction *, 8> TmpMaskedOp; |
1969 | for (BasicBlock *BB : TheLoop->blocks()) { |
1970 | if (!blockCanBePredicated(BB, SafePtrs&: SafePointers, MaskedOp&: TmpMaskedOp)) { |
1971 | LLVM_DEBUG(dbgs() << "LV: Cannot fold tail by masking.\n" ); |
1972 | return false; |
1973 | } |
1974 | } |
1975 | |
1976 | LLVM_DEBUG(dbgs() << "LV: can fold tail by masking.\n" ); |
1977 | |
1978 | return true; |
1979 | } |
1980 | |
1981 | void LoopVectorizationLegality::prepareToFoldTailByMasking() { |
1982 | // The list of pointers that we can safely read and write to remains empty. |
1983 | SmallPtrSet<Value *, 8> SafePointers; |
1984 | |
1985 | // Mark all blocks for predication, including those that ordinarily do not |
1986 | // need predication such as the header block. |
1987 | for (BasicBlock *BB : TheLoop->blocks()) { |
1988 | [[maybe_unused]] bool R = blockCanBePredicated(BB, SafePtrs&: SafePointers, MaskedOp); |
1989 | assert(R && "Must be able to predicate block when tail-folding." ); |
1990 | } |
1991 | } |
1992 | |
1993 | } // namespace llvm |
1994 | |