[clangd] Fix erroneous qualification of template type parameters (#116821)
[llvm-project.git] / polly / lib / Analysis / ScopDetection.cpp
blob73c26578005c351eb30f39e5a85f5be4f116c6a9
1 //===- ScopDetection.cpp - Detect Scops -----------------------------------===//
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 // Detect the maximal Scops of a function.
11 // A static control part (Scop) is a subgraph of the control flow graph (CFG)
12 // that only has statically known control flow and can therefore be described
13 // within the polyhedral model.
15 // Every Scop fulfills these restrictions:
17 // * It is a single entry single exit region
19 // * Only affine linear bounds in the loops
21 // Every natural loop in a Scop must have a number of loop iterations that can
22 // be described as an affine linear function in surrounding loop iterators or
23 // parameters. (A parameter is a scalar that does not change its value during
24 // execution of the Scop).
26 // * Only comparisons of affine linear expressions in conditions
28 // * All loops and conditions perfectly nested
30 // The control flow needs to be structured such that it could be written using
31 // just 'for' and 'if' statements, without the need for any 'goto', 'break' or
32 // 'continue'.
34 // * Side effect free functions call
36 // Function calls and intrinsics that do not have side effects (readnone)
37 // or memory intrinsics (memset, memcpy, memmove) are allowed.
39 // The Scop detection finds the largest Scops by checking if the largest
40 // region is a Scop. If this is not the case, its canonical subregions are
41 // checked until a region is a Scop. It is now tried to extend this Scop by
42 // creating a larger non canonical region.
44 //===----------------------------------------------------------------------===//
46 #include "polly/ScopDetection.h"
47 #include "polly/LinkAllPasses.h"
48 #include "polly/Options.h"
49 #include "polly/ScopDetectionDiagnostic.h"
50 #include "polly/Support/SCEVValidator.h"
51 #include "polly/Support/ScopHelper.h"
52 #include "polly/Support/ScopLocation.h"
53 #include "llvm/ADT/SmallPtrSet.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/Analysis/AliasAnalysis.h"
56 #include "llvm/Analysis/Delinearization.h"
57 #include "llvm/Analysis/Loads.h"
58 #include "llvm/Analysis/LoopInfo.h"
59 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
60 #include "llvm/Analysis/RegionInfo.h"
61 #include "llvm/Analysis/ScalarEvolution.h"
62 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
63 #include "llvm/IR/BasicBlock.h"
64 #include "llvm/IR/DebugLoc.h"
65 #include "llvm/IR/DerivedTypes.h"
66 #include "llvm/IR/DiagnosticInfo.h"
67 #include "llvm/IR/DiagnosticPrinter.h"
68 #include "llvm/IR/Dominators.h"
69 #include "llvm/IR/Function.h"
70 #include "llvm/IR/InstrTypes.h"
71 #include "llvm/IR/Instruction.h"
72 #include "llvm/IR/Instructions.h"
73 #include "llvm/IR/IntrinsicInst.h"
74 #include "llvm/IR/Metadata.h"
75 #include "llvm/IR/Module.h"
76 #include "llvm/IR/PassManager.h"
77 #include "llvm/IR/Value.h"
78 #include "llvm/InitializePasses.h"
79 #include "llvm/Pass.h"
80 #include "llvm/Support/Debug.h"
81 #include "llvm/Support/Regex.h"
82 #include "llvm/Support/raw_ostream.h"
83 #include <algorithm>
84 #include <cassert>
85 #include <memory>
86 #include <stack>
87 #include <string>
88 #include <utility>
89 #include <vector>
91 using namespace llvm;
92 using namespace polly;
94 #include "polly/Support/PollyDebug.h"
95 #define DEBUG_TYPE "polly-detect"
97 // This option is set to a very high value, as analyzing such loops increases
98 // compile time on several cases. For experiments that enable this option,
99 // a value of around 40 has been working to avoid run-time regressions with
100 // Polly while still exposing interesting optimization opportunities.
101 static cl::opt<int> ProfitabilityMinPerLoopInstructions(
102 "polly-detect-profitability-min-per-loop-insts",
103 cl::desc("The minimal number of per-loop instructions before a single loop "
104 "region is considered profitable"),
105 cl::Hidden, cl::ValueRequired, cl::init(100000000), cl::cat(PollyCategory));
107 bool polly::PollyProcessUnprofitable;
109 static cl::opt<bool, true> XPollyProcessUnprofitable(
110 "polly-process-unprofitable",
111 cl::desc(
112 "Process scops that are unlikely to benefit from Polly optimizations."),
113 cl::location(PollyProcessUnprofitable), cl::cat(PollyCategory));
115 static cl::list<std::string> OnlyFunctions(
116 "polly-only-func",
117 cl::desc("Only run on functions that match a regex. "
118 "Multiple regexes can be comma separated. "
119 "Scop detection will run on all functions that match "
120 "ANY of the regexes provided."),
121 cl::CommaSeparated, cl::cat(PollyCategory));
123 static cl::list<std::string> IgnoredFunctions(
124 "polly-ignore-func",
125 cl::desc("Ignore functions that match a regex. "
126 "Multiple regexes can be comma separated. "
127 "Scop detection will ignore all functions that match "
128 "ANY of the regexes provided."),
129 cl::CommaSeparated, cl::cat(PollyCategory));
131 bool polly::PollyAllowFullFunction;
133 static cl::opt<bool, true>
134 XAllowFullFunction("polly-detect-full-functions",
135 cl::desc("Allow the detection of full functions"),
136 cl::location(polly::PollyAllowFullFunction),
137 cl::init(false), cl::cat(PollyCategory));
139 static cl::opt<std::string> OnlyRegion(
140 "polly-only-region",
141 cl::desc("Only run on certain regions (The provided identifier must "
142 "appear in the name of the region's entry block"),
143 cl::value_desc("identifier"), cl::ValueRequired, cl::init(""),
144 cl::cat(PollyCategory));
146 static cl::opt<bool>
147 IgnoreAliasing("polly-ignore-aliasing",
148 cl::desc("Ignore possible aliasing of the array bases"),
149 cl::Hidden, cl::cat(PollyCategory));
151 bool polly::PollyAllowUnsignedOperations;
153 static cl::opt<bool, true> XPollyAllowUnsignedOperations(
154 "polly-allow-unsigned-operations",
155 cl::desc("Allow unsigned operations such as comparisons or zero-extends."),
156 cl::location(PollyAllowUnsignedOperations), cl::Hidden, cl::init(true),
157 cl::cat(PollyCategory));
159 bool polly::PollyUseRuntimeAliasChecks;
161 static cl::opt<bool, true> XPollyUseRuntimeAliasChecks(
162 "polly-use-runtime-alias-checks",
163 cl::desc("Use runtime alias checks to resolve possible aliasing."),
164 cl::location(PollyUseRuntimeAliasChecks), cl::Hidden, cl::init(true),
165 cl::cat(PollyCategory));
167 static cl::opt<bool>
168 ReportLevel("polly-report",
169 cl::desc("Print information about the activities of Polly"),
170 cl::cat(PollyCategory));
172 static cl::opt<bool> AllowDifferentTypes(
173 "polly-allow-differing-element-types",
174 cl::desc("Allow different element types for array accesses"), cl::Hidden,
175 cl::init(true), cl::cat(PollyCategory));
177 static cl::opt<bool>
178 AllowNonAffine("polly-allow-nonaffine",
179 cl::desc("Allow non affine access functions in arrays"),
180 cl::Hidden, cl::cat(PollyCategory));
182 static cl::opt<bool>
183 AllowModrefCall("polly-allow-modref-calls",
184 cl::desc("Allow functions with known modref behavior"),
185 cl::Hidden, cl::cat(PollyCategory));
187 static cl::opt<bool> AllowNonAffineSubRegions(
188 "polly-allow-nonaffine-branches",
189 cl::desc("Allow non affine conditions for branches"), cl::Hidden,
190 cl::init(true), cl::cat(PollyCategory));
192 static cl::opt<bool>
193 AllowNonAffineSubLoops("polly-allow-nonaffine-loops",
194 cl::desc("Allow non affine conditions for loops"),
195 cl::Hidden, cl::cat(PollyCategory));
197 static cl::opt<bool, true>
198 TrackFailures("polly-detect-track-failures",
199 cl::desc("Track failure strings in detecting scop regions"),
200 cl::location(PollyTrackFailures), cl::Hidden, cl::init(true),
201 cl::cat(PollyCategory));
203 static cl::opt<bool> KeepGoing("polly-detect-keep-going",
204 cl::desc("Do not fail on the first error."),
205 cl::Hidden, cl::cat(PollyCategory));
207 static cl::opt<bool, true>
208 PollyDelinearizeX("polly-delinearize",
209 cl::desc("Delinearize array access functions"),
210 cl::location(PollyDelinearize), cl::Hidden,
211 cl::init(true), cl::cat(PollyCategory));
213 static cl::opt<bool>
214 VerifyScops("polly-detect-verify",
215 cl::desc("Verify the detected SCoPs after each transformation"),
216 cl::Hidden, cl::cat(PollyCategory));
218 bool polly::PollyInvariantLoadHoisting;
220 static cl::opt<bool, true>
221 XPollyInvariantLoadHoisting("polly-invariant-load-hoisting",
222 cl::desc("Hoist invariant loads."),
223 cl::location(PollyInvariantLoadHoisting),
224 cl::Hidden, cl::cat(PollyCategory));
226 static cl::opt<bool> PollyAllowErrorBlocks(
227 "polly-allow-error-blocks",
228 cl::desc("Allow to speculate on the execution of 'error blocks'."),
229 cl::Hidden, cl::init(true), cl::cat(PollyCategory));
231 /// The minimal trip count under which loops are considered unprofitable.
232 static const unsigned MIN_LOOP_TRIP_COUNT = 8;
234 bool polly::PollyTrackFailures = false;
235 bool polly::PollyDelinearize = false;
236 StringRef polly::PollySkipFnAttr = "polly.skip.fn";
238 //===----------------------------------------------------------------------===//
239 // Statistics.
241 STATISTIC(NumScopRegions, "Number of scops");
242 STATISTIC(NumLoopsInScop, "Number of loops in scops");
243 STATISTIC(NumScopsDepthZero, "Number of scops with maximal loop depth 0");
244 STATISTIC(NumScopsDepthOne, "Number of scops with maximal loop depth 1");
245 STATISTIC(NumScopsDepthTwo, "Number of scops with maximal loop depth 2");
246 STATISTIC(NumScopsDepthThree, "Number of scops with maximal loop depth 3");
247 STATISTIC(NumScopsDepthFour, "Number of scops with maximal loop depth 4");
248 STATISTIC(NumScopsDepthFive, "Number of scops with maximal loop depth 5");
249 STATISTIC(NumScopsDepthLarger,
250 "Number of scops with maximal loop depth 6 and larger");
251 STATISTIC(NumProfScopRegions, "Number of scops (profitable scops only)");
252 STATISTIC(NumLoopsInProfScop,
253 "Number of loops in scops (profitable scops only)");
254 STATISTIC(NumLoopsOverall, "Number of total loops");
255 STATISTIC(NumProfScopsDepthZero,
256 "Number of scops with maximal loop depth 0 (profitable scops only)");
257 STATISTIC(NumProfScopsDepthOne,
258 "Number of scops with maximal loop depth 1 (profitable scops only)");
259 STATISTIC(NumProfScopsDepthTwo,
260 "Number of scops with maximal loop depth 2 (profitable scops only)");
261 STATISTIC(NumProfScopsDepthThree,
262 "Number of scops with maximal loop depth 3 (profitable scops only)");
263 STATISTIC(NumProfScopsDepthFour,
264 "Number of scops with maximal loop depth 4 (profitable scops only)");
265 STATISTIC(NumProfScopsDepthFive,
266 "Number of scops with maximal loop depth 5 (profitable scops only)");
267 STATISTIC(NumProfScopsDepthLarger,
268 "Number of scops with maximal loop depth 6 and larger "
269 "(profitable scops only)");
270 STATISTIC(MaxNumLoopsInScop, "Maximal number of loops in scops");
271 STATISTIC(MaxNumLoopsInProfScop,
272 "Maximal number of loops in scops (profitable scops only)");
274 static void updateLoopCountStatistic(ScopDetection::LoopStats Stats,
275 bool OnlyProfitable);
277 namespace {
279 class DiagnosticScopFound final : public DiagnosticInfo {
280 private:
281 static int PluginDiagnosticKind;
283 Function &F;
284 std::string FileName;
285 unsigned EntryLine, ExitLine;
287 public:
288 DiagnosticScopFound(Function &F, std::string FileName, unsigned EntryLine,
289 unsigned ExitLine)
290 : DiagnosticInfo(PluginDiagnosticKind, DS_Note), F(F), FileName(FileName),
291 EntryLine(EntryLine), ExitLine(ExitLine) {}
293 void print(DiagnosticPrinter &DP) const override;
295 static bool classof(const DiagnosticInfo *DI) {
296 return DI->getKind() == PluginDiagnosticKind;
299 } // namespace
301 int DiagnosticScopFound::PluginDiagnosticKind =
302 getNextAvailablePluginDiagnosticKind();
304 void DiagnosticScopFound::print(DiagnosticPrinter &DP) const {
305 DP << "Polly detected an optimizable loop region (scop) in function '" << F
306 << "'\n";
308 if (FileName.empty()) {
309 DP << "Scop location is unknown. Compile with debug info "
310 "(-g) to get more precise information. ";
311 return;
314 DP << FileName << ":" << EntryLine << ": Start of scop\n";
315 DP << FileName << ":" << ExitLine << ": End of scop";
318 /// Check if a string matches any regex in a list of regexes.
319 /// @param Str the input string to match against.
320 /// @param RegexList a list of strings that are regular expressions.
321 static bool doesStringMatchAnyRegex(StringRef Str,
322 const cl::list<std::string> &RegexList) {
323 for (auto RegexStr : RegexList) {
324 Regex R(RegexStr);
326 std::string Err;
327 if (!R.isValid(Err))
328 report_fatal_error(Twine("invalid regex given as input to polly: ") + Err,
329 true);
331 if (R.match(Str))
332 return true;
334 return false;
337 //===----------------------------------------------------------------------===//
338 // ScopDetection.
340 ScopDetection::ScopDetection(const DominatorTree &DT, ScalarEvolution &SE,
341 LoopInfo &LI, RegionInfo &RI, AAResults &AA,
342 OptimizationRemarkEmitter &ORE)
343 : DT(DT), SE(SE), LI(LI), RI(RI), AA(AA), ORE(ORE) {}
345 void ScopDetection::detect(Function &F) {
346 assert(ValidRegions.empty() && "Detection must run only once");
348 if (!PollyProcessUnprofitable && LI.empty())
349 return;
351 Region *TopRegion = RI.getTopLevelRegion();
353 if (!OnlyFunctions.empty() &&
354 !doesStringMatchAnyRegex(F.getName(), OnlyFunctions))
355 return;
357 if (doesStringMatchAnyRegex(F.getName(), IgnoredFunctions))
358 return;
360 if (!isValidFunction(F))
361 return;
363 findScops(*TopRegion);
365 NumScopRegions += ValidRegions.size();
367 // Prune non-profitable regions.
368 for (auto &DIt : DetectionContextMap) {
369 DetectionContext &DC = *DIt.getSecond().get();
370 if (DC.Log.hasErrors())
371 continue;
372 if (!ValidRegions.count(&DC.CurRegion))
373 continue;
374 LoopStats Stats = countBeneficialLoops(&DC.CurRegion, SE, LI, 0);
375 updateLoopCountStatistic(Stats, false /* OnlyProfitable */);
376 if (isProfitableRegion(DC)) {
377 updateLoopCountStatistic(Stats, true /* OnlyProfitable */);
378 continue;
381 ValidRegions.remove(&DC.CurRegion);
384 NumProfScopRegions += ValidRegions.size();
385 NumLoopsOverall += countBeneficialLoops(TopRegion, SE, LI, 0).NumLoops;
387 // Only makes sense when we tracked errors.
388 if (PollyTrackFailures)
389 emitMissedRemarks(F);
391 if (ReportLevel)
392 printLocations(F);
394 assert(ValidRegions.size() <= DetectionContextMap.size() &&
395 "Cached more results than valid regions");
398 template <class RR, typename... Args>
399 inline bool ScopDetection::invalid(DetectionContext &Context, bool Assert,
400 Args &&...Arguments) const {
401 if (!Context.Verifying) {
402 RejectLog &Log = Context.Log;
403 std::shared_ptr<RR> RejectReason = std::make_shared<RR>(Arguments...);
404 Context.IsInvalid = true;
406 // Log even if PollyTrackFailures is false, the log entries are also used in
407 // canUseISLTripCount().
408 Log.report(RejectReason);
410 POLLY_DEBUG(dbgs() << RejectReason->getMessage());
411 POLLY_DEBUG(dbgs() << "\n");
412 } else {
413 assert(!Assert && "Verification of detected scop failed");
416 return false;
419 bool ScopDetection::isMaxRegionInScop(const Region &R, bool Verify) {
420 if (!ValidRegions.count(&R))
421 return false;
423 if (Verify) {
424 BBPair P = getBBPairForRegion(&R);
425 std::unique_ptr<DetectionContext> &Entry = DetectionContextMap[P];
427 // Free previous DetectionContext for the region and create and verify a new
428 // one. Be sure that the DetectionContext is not still used by a ScopInfop.
429 // Due to changes but CodeGeneration of another Scop, the Region object and
430 // the BBPair might not match anymore.
431 Entry = std::make_unique<DetectionContext>(const_cast<Region &>(R), AA,
432 /*Verifying=*/false);
434 return isValidRegion(*Entry.get());
437 return true;
440 std::string ScopDetection::regionIsInvalidBecause(const Region *R) const {
441 // Get the first error we found. Even in keep-going mode, this is the first
442 // reason that caused the candidate to be rejected.
443 auto *Log = lookupRejectionLog(R);
445 // This can happen when we marked a region invalid, but didn't track
446 // an error for it.
447 if (!Log || !Log->hasErrors())
448 return "";
450 RejectReasonPtr RR = *Log->begin();
451 return RR->getMessage();
454 bool ScopDetection::addOverApproximatedRegion(Region *AR,
455 DetectionContext &Context) const {
456 // If we already know about Ar we can exit.
457 if (!Context.NonAffineSubRegionSet.insert(AR))
458 return true;
460 // All loops in the region have to be overapproximated too if there
461 // are accesses that depend on the iteration count.
463 for (BasicBlock *BB : AR->blocks()) {
464 Loop *L = LI.getLoopFor(BB);
465 if (AR->contains(L))
466 Context.BoxedLoopsSet.insert(L);
469 return (AllowNonAffineSubLoops || Context.BoxedLoopsSet.empty());
472 bool ScopDetection::onlyValidRequiredInvariantLoads(
473 InvariantLoadsSetTy &RequiredILS, DetectionContext &Context) const {
474 Region &CurRegion = Context.CurRegion;
475 const DataLayout &DL = CurRegion.getEntry()->getModule()->getDataLayout();
477 if (!PollyInvariantLoadHoisting && !RequiredILS.empty())
478 return false;
480 for (LoadInst *Load : RequiredILS) {
481 // If we already know a load has been accepted as required invariant, we
482 // already run the validation below once and consequently don't need to
483 // run it again. Hence, we return early. For certain test cases (e.g.,
484 // COSMO this avoids us spending 50% of scop-detection time in this
485 // very function (and its children).
486 if (Context.RequiredILS.count(Load))
487 continue;
488 if (!isHoistableLoad(Load, CurRegion, LI, SE, DT, Context.RequiredILS))
489 return false;
491 for (auto NonAffineRegion : Context.NonAffineSubRegionSet) {
492 if (isSafeToLoadUnconditionally(Load->getPointerOperand(),
493 Load->getType(), Load->getAlign(), DL,
494 nullptr))
495 continue;
497 if (NonAffineRegion->contains(Load) &&
498 Load->getParent() != NonAffineRegion->getEntry())
499 return false;
503 Context.RequiredILS.insert(RequiredILS.begin(), RequiredILS.end());
505 return true;
508 bool ScopDetection::involvesMultiplePtrs(const SCEV *S0, const SCEV *S1,
509 Loop *Scope) const {
510 SetVector<Value *> Values;
511 findValues(S0, SE, Values);
512 if (S1)
513 findValues(S1, SE, Values);
515 SmallPtrSet<Value *, 8> PtrVals;
516 for (auto *V : Values) {
517 if (auto *P2I = dyn_cast<PtrToIntInst>(V))
518 V = P2I->getOperand(0);
520 if (!V->getType()->isPointerTy())
521 continue;
523 auto *PtrSCEV = SE.getSCEVAtScope(V, Scope);
524 if (isa<SCEVConstant>(PtrSCEV))
525 continue;
527 auto *BasePtr = dyn_cast<SCEVUnknown>(SE.getPointerBase(PtrSCEV));
528 if (!BasePtr)
529 return true;
531 auto *BasePtrVal = BasePtr->getValue();
532 if (PtrVals.insert(BasePtrVal).second) {
533 for (auto *PtrVal : PtrVals)
534 if (PtrVal != BasePtrVal && !AA.isNoAlias(PtrVal, BasePtrVal))
535 return true;
539 return false;
542 bool ScopDetection::isAffine(const SCEV *S, Loop *Scope,
543 DetectionContext &Context) const {
544 InvariantLoadsSetTy AccessILS;
545 if (!isAffineExpr(&Context.CurRegion, Scope, S, SE, &AccessILS))
546 return false;
548 if (!onlyValidRequiredInvariantLoads(AccessILS, Context))
549 return false;
551 return true;
554 bool ScopDetection::isValidSwitch(BasicBlock &BB, SwitchInst *SI,
555 Value *Condition, bool IsLoopBranch,
556 DetectionContext &Context) const {
557 Loop *L = LI.getLoopFor(&BB);
558 const SCEV *ConditionSCEV = SE.getSCEVAtScope(Condition, L);
560 if (IsLoopBranch && L->isLoopLatch(&BB))
561 return false;
563 // Check for invalid usage of different pointers in one expression.
564 if (involvesMultiplePtrs(ConditionSCEV, nullptr, L))
565 return false;
567 if (isAffine(ConditionSCEV, L, Context))
568 return true;
570 if (AllowNonAffineSubRegions &&
571 addOverApproximatedRegion(RI.getRegionFor(&BB), Context))
572 return true;
574 return invalid<ReportNonAffBranch>(Context, /*Assert=*/true, &BB,
575 ConditionSCEV, ConditionSCEV, SI);
578 bool ScopDetection::isValidBranch(BasicBlock &BB, BranchInst *BI,
579 Value *Condition, bool IsLoopBranch,
580 DetectionContext &Context) {
581 // Constant integer conditions are always affine.
582 if (isa<ConstantInt>(Condition))
583 return true;
585 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Condition)) {
586 auto Opcode = BinOp->getOpcode();
587 if (Opcode == Instruction::And || Opcode == Instruction::Or) {
588 Value *Op0 = BinOp->getOperand(0);
589 Value *Op1 = BinOp->getOperand(1);
590 return isValidBranch(BB, BI, Op0, IsLoopBranch, Context) &&
591 isValidBranch(BB, BI, Op1, IsLoopBranch, Context);
595 if (auto PHI = dyn_cast<PHINode>(Condition)) {
596 auto *Unique = dyn_cast_or_null<ConstantInt>(
597 getUniqueNonErrorValue(PHI, &Context.CurRegion, this));
598 if (Unique && (Unique->isZero() || Unique->isOne()))
599 return true;
602 if (auto Load = dyn_cast<LoadInst>(Condition))
603 if (!IsLoopBranch && Context.CurRegion.contains(Load)) {
604 Context.RequiredILS.insert(Load);
605 return true;
608 // Non constant conditions of branches need to be ICmpInst.
609 if (!isa<ICmpInst>(Condition)) {
610 if (!IsLoopBranch && AllowNonAffineSubRegions &&
611 addOverApproximatedRegion(RI.getRegionFor(&BB), Context))
612 return true;
613 return invalid<ReportInvalidCond>(Context, /*Assert=*/true, BI, &BB);
616 ICmpInst *ICmp = cast<ICmpInst>(Condition);
618 // Are both operands of the ICmp affine?
619 if (isa<UndefValue>(ICmp->getOperand(0)) ||
620 isa<UndefValue>(ICmp->getOperand(1)))
621 return invalid<ReportUndefOperand>(Context, /*Assert=*/true, &BB, ICmp);
623 Loop *L = LI.getLoopFor(&BB);
624 const SCEV *LHS = SE.getSCEVAtScope(ICmp->getOperand(0), L);
625 const SCEV *RHS = SE.getSCEVAtScope(ICmp->getOperand(1), L);
627 LHS = tryForwardThroughPHI(LHS, Context.CurRegion, SE, this);
628 RHS = tryForwardThroughPHI(RHS, Context.CurRegion, SE, this);
630 // If unsigned operations are not allowed try to approximate the region.
631 if (ICmp->isUnsigned() && !PollyAllowUnsignedOperations)
632 return !IsLoopBranch && AllowNonAffineSubRegions &&
633 addOverApproximatedRegion(RI.getRegionFor(&BB), Context);
635 // Check for invalid usage of different pointers in one expression.
636 if (ICmp->isEquality() && involvesMultiplePtrs(LHS, nullptr, L) &&
637 involvesMultiplePtrs(RHS, nullptr, L))
638 return false;
640 // Check for invalid usage of different pointers in a relational comparison.
641 if (ICmp->isRelational() && involvesMultiplePtrs(LHS, RHS, L))
642 return false;
644 if (isAffine(LHS, L, Context) && isAffine(RHS, L, Context))
645 return true;
647 if (!IsLoopBranch && AllowNonAffineSubRegions &&
648 addOverApproximatedRegion(RI.getRegionFor(&BB), Context))
649 return true;
651 if (IsLoopBranch)
652 return false;
654 return invalid<ReportNonAffBranch>(Context, /*Assert=*/true, &BB, LHS, RHS,
655 ICmp);
658 bool ScopDetection::isValidCFG(BasicBlock &BB, bool IsLoopBranch,
659 bool AllowUnreachable,
660 DetectionContext &Context) {
661 Region &CurRegion = Context.CurRegion;
663 Instruction *TI = BB.getTerminator();
665 if (AllowUnreachable && isa<UnreachableInst>(TI))
666 return true;
668 // Return instructions are only valid if the region is the top level region.
669 if (isa<ReturnInst>(TI) && CurRegion.isTopLevelRegion())
670 return true;
672 Value *Condition = getConditionFromTerminator(TI);
674 if (!Condition)
675 return invalid<ReportInvalidTerminator>(Context, /*Assert=*/true, &BB);
677 // UndefValue is not allowed as condition.
678 if (isa<UndefValue>(Condition))
679 return invalid<ReportUndefCond>(Context, /*Assert=*/true, TI, &BB);
681 if (BranchInst *BI = dyn_cast<BranchInst>(TI))
682 return isValidBranch(BB, BI, Condition, IsLoopBranch, Context);
684 SwitchInst *SI = dyn_cast<SwitchInst>(TI);
685 assert(SI && "Terminator was neither branch nor switch");
687 return isValidSwitch(BB, SI, Condition, IsLoopBranch, Context);
690 bool ScopDetection::isValidCallInst(CallInst &CI,
691 DetectionContext &Context) const {
692 if (CI.doesNotReturn())
693 return false;
695 if (CI.doesNotAccessMemory())
696 return true;
698 if (auto *II = dyn_cast<IntrinsicInst>(&CI))
699 if (isValidIntrinsicInst(*II, Context))
700 return true;
702 Function *CalledFunction = CI.getCalledFunction();
704 // Indirect calls are not supported.
705 if (CalledFunction == nullptr)
706 return false;
708 if (isDebugCall(&CI)) {
709 POLLY_DEBUG(dbgs() << "Allow call to debug function: "
710 << CalledFunction->getName() << '\n');
711 return true;
714 if (AllowModrefCall) {
715 MemoryEffects ME = AA.getMemoryEffects(CalledFunction);
716 if (ME.onlyAccessesArgPointees()) {
717 for (const auto &Arg : CI.args()) {
718 if (!Arg->getType()->isPointerTy())
719 continue;
721 // Bail if a pointer argument has a base address not known to
722 // ScalarEvolution. Note that a zero pointer is acceptable.
723 auto *ArgSCEV = SE.getSCEVAtScope(Arg, LI.getLoopFor(CI.getParent()));
724 if (ArgSCEV->isZero())
725 continue;
727 auto *BP = dyn_cast<SCEVUnknown>(SE.getPointerBase(ArgSCEV));
728 if (!BP)
729 return false;
731 // Implicitly disable delinearization since we have an unknown
732 // accesses with an unknown access function.
733 Context.HasUnknownAccess = true;
736 // Explicitly use addUnknown so we don't put a loop-variant
737 // pointer into the alias set.
738 Context.AST.addUnknown(&CI);
739 return true;
742 if (ME.onlyReadsMemory()) {
743 // Implicitly disable delinearization since we have an unknown
744 // accesses with an unknown access function.
745 Context.HasUnknownAccess = true;
746 // Explicitly use addUnknown so we don't put a loop-variant
747 // pointer into the alias set.
748 Context.AST.addUnknown(&CI);
749 return true;
751 return false;
754 return false;
757 bool ScopDetection::isValidIntrinsicInst(IntrinsicInst &II,
758 DetectionContext &Context) const {
759 if (isIgnoredIntrinsic(&II))
760 return true;
762 // The closest loop surrounding the call instruction.
763 Loop *L = LI.getLoopFor(II.getParent());
765 // The access function and base pointer for memory intrinsics.
766 const SCEV *AF;
767 const SCEVUnknown *BP;
769 switch (II.getIntrinsicID()) {
770 // Memory intrinsics that can be represented are supported.
771 case Intrinsic::memmove:
772 case Intrinsic::memcpy:
773 AF = SE.getSCEVAtScope(cast<MemTransferInst>(II).getSource(), L);
774 if (!AF->isZero()) {
775 BP = dyn_cast<SCEVUnknown>(SE.getPointerBase(AF));
776 // Bail if the source pointer is not valid.
777 if (!isValidAccess(&II, AF, BP, Context))
778 return false;
780 [[fallthrough]];
781 case Intrinsic::memset:
782 AF = SE.getSCEVAtScope(cast<MemIntrinsic>(II).getDest(), L);
783 if (!AF->isZero()) {
784 BP = dyn_cast<SCEVUnknown>(SE.getPointerBase(AF));
785 // Bail if the destination pointer is not valid.
786 if (!isValidAccess(&II, AF, BP, Context))
787 return false;
790 // Bail if the length is not affine.
791 if (!isAffine(SE.getSCEVAtScope(cast<MemIntrinsic>(II).getLength(), L), L,
792 Context))
793 return false;
795 return true;
796 default:
797 break;
800 return false;
803 bool ScopDetection::isInvariant(Value &Val, const Region &Reg,
804 DetectionContext &Ctx) const {
805 // A reference to function argument or constant value is invariant.
806 if (isa<Argument>(Val) || isa<Constant>(Val))
807 return true;
809 Instruction *I = dyn_cast<Instruction>(&Val);
810 if (!I)
811 return false;
813 if (!Reg.contains(I))
814 return true;
816 // Loads within the SCoP may read arbitrary values, need to hoist them. If it
817 // is not hoistable, it will be rejected later, but here we assume it is and
818 // that makes the value invariant.
819 if (auto LI = dyn_cast<LoadInst>(I)) {
820 Ctx.RequiredILS.insert(LI);
821 return true;
824 return false;
827 namespace {
829 /// Remove smax of smax(0, size) expressions from a SCEV expression and
830 /// register the '...' components.
832 /// Array access expressions as they are generated by GFortran contain smax(0,
833 /// size) expressions that confuse the 'normal' delinearization algorithm.
834 /// However, if we extract such expressions before the normal delinearization
835 /// takes place they can actually help to identify array size expressions in
836 /// Fortran accesses. For the subsequently following delinearization the smax(0,
837 /// size) component can be replaced by just 'size'. This is correct as we will
838 /// always add and verify the assumption that for all subscript expressions
839 /// 'exp' the inequality 0 <= exp < size holds. Hence, we will also verify
840 /// that 0 <= size, which means smax(0, size) == size.
841 class SCEVRemoveMax final : public SCEVRewriteVisitor<SCEVRemoveMax> {
842 public:
843 SCEVRemoveMax(ScalarEvolution &SE, std::vector<const SCEV *> *Terms)
844 : SCEVRewriteVisitor(SE), Terms(Terms) {}
846 static const SCEV *rewrite(const SCEV *Scev, ScalarEvolution &SE,
847 std::vector<const SCEV *> *Terms = nullptr) {
848 SCEVRemoveMax Rewriter(SE, Terms);
849 return Rewriter.visit(Scev);
852 const SCEV *visitSMaxExpr(const SCEVSMaxExpr *Expr) {
853 if ((Expr->getNumOperands() == 2) && Expr->getOperand(0)->isZero()) {
854 auto Res = visit(Expr->getOperand(1));
855 if (Terms)
856 (*Terms).push_back(Res);
857 return Res;
860 return Expr;
863 private:
864 std::vector<const SCEV *> *Terms;
866 } // namespace
868 SmallVector<const SCEV *, 4>
869 ScopDetection::getDelinearizationTerms(DetectionContext &Context,
870 const SCEVUnknown *BasePointer) const {
871 SmallVector<const SCEV *, 4> Terms;
872 for (const auto &Pair : Context.Accesses[BasePointer]) {
873 std::vector<const SCEV *> MaxTerms;
874 SCEVRemoveMax::rewrite(Pair.second, SE, &MaxTerms);
875 if (!MaxTerms.empty()) {
876 Terms.insert(Terms.begin(), MaxTerms.begin(), MaxTerms.end());
877 continue;
879 // In case the outermost expression is a plain add, we check if any of its
880 // terms has the form 4 * %inst * %param * %param ..., aka a term that
881 // contains a product between a parameter and an instruction that is
882 // inside the scop. Such instructions, if allowed at all, are instructions
883 // SCEV can not represent, but Polly is still looking through. As a
884 // result, these instructions can depend on induction variables and are
885 // most likely no array sizes. However, terms that are multiplied with
886 // them are likely candidates for array sizes.
887 if (auto *AF = dyn_cast<SCEVAddExpr>(Pair.second)) {
888 for (auto Op : AF->operands()) {
889 if (auto *AF2 = dyn_cast<SCEVAddRecExpr>(Op))
890 collectParametricTerms(SE, AF2, Terms);
891 if (auto *AF2 = dyn_cast<SCEVMulExpr>(Op)) {
892 SmallVector<const SCEV *, 0> Operands;
894 for (auto *MulOp : AF2->operands()) {
895 if (auto *Const = dyn_cast<SCEVConstant>(MulOp))
896 Operands.push_back(Const);
897 if (auto *Unknown = dyn_cast<SCEVUnknown>(MulOp)) {
898 if (auto *Inst = dyn_cast<Instruction>(Unknown->getValue())) {
899 if (!Context.CurRegion.contains(Inst))
900 Operands.push_back(MulOp);
902 } else {
903 Operands.push_back(MulOp);
907 if (Operands.size())
908 Terms.push_back(SE.getMulExpr(Operands));
912 if (Terms.empty())
913 collectParametricTerms(SE, Pair.second, Terms);
915 return Terms;
918 bool ScopDetection::hasValidArraySizes(DetectionContext &Context,
919 SmallVectorImpl<const SCEV *> &Sizes,
920 const SCEVUnknown *BasePointer,
921 Loop *Scope) const {
922 // If no sizes were found, all sizes are trivially valid. We allow this case
923 // to make it possible to pass known-affine accesses to the delinearization to
924 // try to recover some interesting multi-dimensional accesses, but to still
925 // allow the already known to be affine access in case the delinearization
926 // fails. In such situations, the delinearization will just return a Sizes
927 // array of size zero.
928 if (Sizes.size() == 0)
929 return true;
931 Value *BaseValue = BasePointer->getValue();
932 Region &CurRegion = Context.CurRegion;
933 for (const SCEV *DelinearizedSize : Sizes) {
934 // Don't pass down the scope to isAfffine; array dimensions must be
935 // invariant across the entire scop.
936 if (!isAffine(DelinearizedSize, nullptr, Context)) {
937 Sizes.clear();
938 break;
940 if (auto *Unknown = dyn_cast<SCEVUnknown>(DelinearizedSize)) {
941 auto *V = dyn_cast<Value>(Unknown->getValue());
942 if (auto *Load = dyn_cast<LoadInst>(V)) {
943 if (Context.CurRegion.contains(Load) &&
944 isHoistableLoad(Load, CurRegion, LI, SE, DT, Context.RequiredILS))
945 Context.RequiredILS.insert(Load);
946 continue;
949 if (hasScalarDepsInsideRegion(DelinearizedSize, &CurRegion, Scope, false,
950 Context.RequiredILS))
951 return invalid<ReportNonAffineAccess>(
952 Context, /*Assert=*/true, DelinearizedSize,
953 Context.Accesses[BasePointer].front().first, BaseValue);
956 // No array shape derived.
957 if (Sizes.empty()) {
958 if (AllowNonAffine)
959 return true;
961 for (const auto &Pair : Context.Accesses[BasePointer]) {
962 const Instruction *Insn = Pair.first;
963 const SCEV *AF = Pair.second;
965 if (!isAffine(AF, Scope, Context)) {
966 invalid<ReportNonAffineAccess>(Context, /*Assert=*/true, AF, Insn,
967 BaseValue);
968 if (!KeepGoing)
969 return false;
972 return false;
974 return true;
977 // We first store the resulting memory accesses in TempMemoryAccesses. Only
978 // if the access functions for all memory accesses have been successfully
979 // delinearized we continue. Otherwise, we either report a failure or, if
980 // non-affine accesses are allowed, we drop the information. In case the
981 // information is dropped the memory accesses need to be overapproximated
982 // when translated to a polyhedral representation.
983 bool ScopDetection::computeAccessFunctions(
984 DetectionContext &Context, const SCEVUnknown *BasePointer,
985 std::shared_ptr<ArrayShape> Shape) const {
986 Value *BaseValue = BasePointer->getValue();
987 bool BasePtrHasNonAffine = false;
988 MapInsnToMemAcc TempMemoryAccesses;
989 for (const auto &Pair : Context.Accesses[BasePointer]) {
990 const Instruction *Insn = Pair.first;
991 auto *AF = Pair.second;
992 AF = SCEVRemoveMax::rewrite(AF, SE);
993 bool IsNonAffine = false;
994 TempMemoryAccesses.insert(std::make_pair(Insn, MemAcc(Insn, Shape)));
995 MemAcc *Acc = &TempMemoryAccesses.find(Insn)->second;
996 auto *Scope = LI.getLoopFor(Insn->getParent());
998 if (!AF) {
999 if (isAffine(Pair.second, Scope, Context))
1000 Acc->DelinearizedSubscripts.push_back(Pair.second);
1001 else
1002 IsNonAffine = true;
1003 } else {
1004 if (Shape->DelinearizedSizes.size() == 0) {
1005 Acc->DelinearizedSubscripts.push_back(AF);
1006 } else {
1007 llvm::computeAccessFunctions(SE, AF, Acc->DelinearizedSubscripts,
1008 Shape->DelinearizedSizes);
1009 if (Acc->DelinearizedSubscripts.size() == 0)
1010 IsNonAffine = true;
1012 for (const SCEV *S : Acc->DelinearizedSubscripts)
1013 if (!isAffine(S, Scope, Context))
1014 IsNonAffine = true;
1017 // (Possibly) report non affine access
1018 if (IsNonAffine) {
1019 BasePtrHasNonAffine = true;
1020 if (!AllowNonAffine) {
1021 invalid<ReportNonAffineAccess>(Context, /*Assert=*/true, Pair.second,
1022 Insn, BaseValue);
1023 if (!KeepGoing)
1024 return false;
1029 if (!BasePtrHasNonAffine)
1030 Context.InsnToMemAcc.insert(TempMemoryAccesses.begin(),
1031 TempMemoryAccesses.end());
1033 return true;
1036 bool ScopDetection::hasBaseAffineAccesses(DetectionContext &Context,
1037 const SCEVUnknown *BasePointer,
1038 Loop *Scope) const {
1039 auto Shape = std::shared_ptr<ArrayShape>(new ArrayShape(BasePointer));
1041 auto Terms = getDelinearizationTerms(Context, BasePointer);
1043 findArrayDimensions(SE, Terms, Shape->DelinearizedSizes,
1044 Context.ElementSize[BasePointer]);
1046 if (!hasValidArraySizes(Context, Shape->DelinearizedSizes, BasePointer,
1047 Scope))
1048 return false;
1050 return computeAccessFunctions(Context, BasePointer, Shape);
1053 bool ScopDetection::hasAffineMemoryAccesses(DetectionContext &Context) const {
1054 // TODO: If we have an unknown access and other non-affine accesses we do
1055 // not try to delinearize them for now.
1056 if (Context.HasUnknownAccess && !Context.NonAffineAccesses.empty())
1057 return AllowNonAffine;
1059 for (auto &Pair : Context.NonAffineAccesses) {
1060 auto *BasePointer = Pair.first;
1061 auto *Scope = Pair.second;
1062 if (!hasBaseAffineAccesses(Context, BasePointer, Scope)) {
1063 Context.IsInvalid = true;
1064 if (!KeepGoing)
1065 return false;
1068 return true;
1071 bool ScopDetection::isValidAccess(Instruction *Inst, const SCEV *AF,
1072 const SCEVUnknown *BP,
1073 DetectionContext &Context) const {
1075 if (!BP)
1076 return invalid<ReportNoBasePtr>(Context, /*Assert=*/true, Inst);
1078 auto *BV = BP->getValue();
1079 if (isa<UndefValue>(BV))
1080 return invalid<ReportUndefBasePtr>(Context, /*Assert=*/true, Inst);
1082 // FIXME: Think about allowing IntToPtrInst
1083 if (IntToPtrInst *Inst = dyn_cast<IntToPtrInst>(BV))
1084 return invalid<ReportIntToPtr>(Context, /*Assert=*/true, Inst);
1086 // Check that the base address of the access is invariant in the current
1087 // region.
1088 if (!isInvariant(*BV, Context.CurRegion, Context))
1089 return invalid<ReportVariantBasePtr>(Context, /*Assert=*/true, BV, Inst);
1091 AF = SE.getMinusSCEV(AF, BP);
1093 const SCEV *Size;
1094 if (!isa<MemIntrinsic>(Inst)) {
1095 Size = SE.getElementSize(Inst);
1096 } else {
1097 auto *SizeTy =
1098 SE.getEffectiveSCEVType(PointerType::getUnqual(SE.getContext()));
1099 Size = SE.getConstant(SizeTy, 8);
1102 if (Context.ElementSize[BP]) {
1103 if (!AllowDifferentTypes && Context.ElementSize[BP] != Size)
1104 return invalid<ReportDifferentArrayElementSize>(Context, /*Assert=*/true,
1105 Inst, BV);
1107 Context.ElementSize[BP] = SE.getSMinExpr(Size, Context.ElementSize[BP]);
1108 } else {
1109 Context.ElementSize[BP] = Size;
1112 bool IsVariantInNonAffineLoop = false;
1113 SetVector<const Loop *> Loops;
1114 findLoops(AF, Loops);
1115 for (const Loop *L : Loops)
1116 if (Context.BoxedLoopsSet.count(L))
1117 IsVariantInNonAffineLoop = true;
1119 auto *Scope = LI.getLoopFor(Inst->getParent());
1120 bool IsAffine = !IsVariantInNonAffineLoop && isAffine(AF, Scope, Context);
1121 // Do not try to delinearize memory intrinsics and force them to be affine.
1122 if (isa<MemIntrinsic>(Inst) && !IsAffine) {
1123 return invalid<ReportNonAffineAccess>(Context, /*Assert=*/true, AF, Inst,
1124 BV);
1125 } else if (PollyDelinearize && !IsVariantInNonAffineLoop) {
1126 Context.Accesses[BP].push_back({Inst, AF});
1128 if (!IsAffine)
1129 Context.NonAffineAccesses.insert(
1130 std::make_pair(BP, LI.getLoopFor(Inst->getParent())));
1131 } else if (!AllowNonAffine && !IsAffine) {
1132 return invalid<ReportNonAffineAccess>(Context, /*Assert=*/true, AF, Inst,
1133 BV);
1136 if (IgnoreAliasing)
1137 return true;
1139 // Check if the base pointer of the memory access does alias with
1140 // any other pointer. This cannot be handled at the moment.
1141 AAMDNodes AATags = Inst->getAAMetadata();
1142 AliasSet &AS = Context.AST.getAliasSetFor(
1143 MemoryLocation::getBeforeOrAfter(BP->getValue(), AATags));
1145 if (!AS.isMustAlias()) {
1146 if (PollyUseRuntimeAliasChecks) {
1147 bool CanBuildRunTimeCheck = true;
1148 // The run-time alias check places code that involves the base pointer at
1149 // the beginning of the SCoP. This breaks if the base pointer is defined
1150 // inside the scop. Hence, we can only create a run-time check if we are
1151 // sure the base pointer is not an instruction defined inside the scop.
1152 // However, we can ignore loads that will be hoisted.
1154 auto ASPointers = AS.getPointers();
1156 InvariantLoadsSetTy VariantLS, InvariantLS;
1157 // In order to detect loads which are dependent on other invariant loads
1158 // as invariant, we use fixed-point iteration method here i.e we iterate
1159 // over the alias set for arbitrary number of times until it is safe to
1160 // assume that all the invariant loads have been detected
1161 while (true) {
1162 const unsigned int VariantSize = VariantLS.size(),
1163 InvariantSize = InvariantLS.size();
1165 for (const Value *Ptr : ASPointers) {
1166 Instruction *Inst = dyn_cast<Instruction>(const_cast<Value *>(Ptr));
1167 if (Inst && Context.CurRegion.contains(Inst)) {
1168 auto *Load = dyn_cast<LoadInst>(Inst);
1169 if (Load && InvariantLS.count(Load))
1170 continue;
1171 if (Load && isHoistableLoad(Load, Context.CurRegion, LI, SE, DT,
1172 InvariantLS)) {
1173 if (VariantLS.count(Load))
1174 VariantLS.remove(Load);
1175 Context.RequiredILS.insert(Load);
1176 InvariantLS.insert(Load);
1177 } else {
1178 CanBuildRunTimeCheck = false;
1179 VariantLS.insert(Load);
1184 if (InvariantSize == InvariantLS.size() &&
1185 VariantSize == VariantLS.size())
1186 break;
1189 if (CanBuildRunTimeCheck)
1190 return true;
1192 return invalid<ReportAlias>(Context, /*Assert=*/true, Inst, AS);
1195 return true;
1198 bool ScopDetection::isValidMemoryAccess(MemAccInst Inst,
1199 DetectionContext &Context) const {
1200 Value *Ptr = Inst.getPointerOperand();
1201 Loop *L = LI.getLoopFor(Inst->getParent());
1202 const SCEV *AccessFunction = SE.getSCEVAtScope(Ptr, L);
1203 const SCEVUnknown *BasePointer;
1205 BasePointer = dyn_cast<SCEVUnknown>(SE.getPointerBase(AccessFunction));
1207 return isValidAccess(Inst, AccessFunction, BasePointer, Context);
1210 bool ScopDetection::isValidInstruction(Instruction &Inst,
1211 DetectionContext &Context) {
1212 for (auto &Op : Inst.operands()) {
1213 auto *OpInst = dyn_cast<Instruction>(&Op);
1215 if (!OpInst)
1216 continue;
1218 if (isErrorBlock(*OpInst->getParent(), Context.CurRegion)) {
1219 auto *PHI = dyn_cast<PHINode>(OpInst);
1220 if (PHI) {
1221 for (User *U : PHI->users()) {
1222 auto *UI = dyn_cast<Instruction>(U);
1223 if (!UI || !UI->isTerminator())
1224 return false;
1226 } else {
1227 return false;
1232 if (isa<LandingPadInst>(&Inst) || isa<ResumeInst>(&Inst))
1233 return false;
1235 // We only check the call instruction but not invoke instruction.
1236 if (CallInst *CI = dyn_cast<CallInst>(&Inst)) {
1237 if (isValidCallInst(*CI, Context))
1238 return true;
1240 return invalid<ReportFuncCall>(Context, /*Assert=*/true, &Inst);
1243 if (!Inst.mayReadOrWriteMemory()) {
1244 if (!isa<AllocaInst>(Inst))
1245 return true;
1247 return invalid<ReportAlloca>(Context, /*Assert=*/true, &Inst);
1250 // Check the access function.
1251 if (auto MemInst = MemAccInst::dyn_cast(Inst)) {
1252 Context.hasStores |= isa<StoreInst>(MemInst);
1253 Context.hasLoads |= isa<LoadInst>(MemInst);
1254 if (!MemInst.isSimple())
1255 return invalid<ReportNonSimpleMemoryAccess>(Context, /*Assert=*/true,
1256 &Inst);
1258 return isValidMemoryAccess(MemInst, Context);
1261 // We do not know this instruction, therefore we assume it is invalid.
1262 return invalid<ReportUnknownInst>(Context, /*Assert=*/true, &Inst);
1265 /// Check whether @p L has exiting blocks.
1267 /// @param L The loop of interest
1269 /// @return True if the loop has exiting blocks, false otherwise.
1270 static bool hasExitingBlocks(Loop *L) {
1271 SmallVector<BasicBlock *, 4> ExitingBlocks;
1272 L->getExitingBlocks(ExitingBlocks);
1273 return !ExitingBlocks.empty();
1276 bool ScopDetection::canUseISLTripCount(Loop *L, DetectionContext &Context) {
1277 // FIXME: Yes, this is bad. isValidCFG() may call invalid<Reason>() which
1278 // causes the SCoP to be rejected regardless on whether non-ISL trip counts
1279 // could be used. We currently preserve the legacy behaviour of rejecting
1280 // based on Context.Log.size() added by isValidCFG() or before, regardless on
1281 // whether the ISL trip count can be used or can be used as a non-affine
1282 // region. However, we allow rejections by isValidCFG() that do not result in
1283 // an error log entry.
1284 bool OldIsInvalid = Context.IsInvalid;
1286 // Ensure the loop has valid exiting blocks as well as latches, otherwise we
1287 // need to overapproximate it as a boxed loop.
1288 SmallVector<BasicBlock *, 4> LoopControlBlocks;
1289 L->getExitingBlocks(LoopControlBlocks);
1290 L->getLoopLatches(LoopControlBlocks);
1291 for (BasicBlock *ControlBB : LoopControlBlocks) {
1292 if (!isValidCFG(*ControlBB, true, false, Context)) {
1293 Context.IsInvalid = OldIsInvalid || Context.Log.size();
1294 return false;
1298 // We can use ISL to compute the trip count of L.
1299 Context.IsInvalid = OldIsInvalid || Context.Log.size();
1300 return true;
1303 bool ScopDetection::isValidLoop(Loop *L, DetectionContext &Context) {
1304 // Loops that contain part but not all of the blocks of a region cannot be
1305 // handled by the schedule generation. Such loop constructs can happen
1306 // because a region can contain BBs that have no path to the exit block
1307 // (Infinite loops, UnreachableInst), but such blocks are never part of a
1308 // loop.
1310 // _______________
1311 // | Loop Header | <-----------.
1312 // --------------- |
1313 // | |
1314 // _______________ ______________
1315 // | RegionEntry |-----> | RegionExit |----->
1316 // --------------- --------------
1317 // |
1318 // _______________
1319 // | EndlessLoop | <--.
1320 // --------------- |
1321 // | |
1322 // \------------/
1324 // In the example above, the loop (LoopHeader,RegionEntry,RegionExit) is
1325 // neither entirely contained in the region RegionEntry->RegionExit
1326 // (containing RegionEntry,EndlessLoop) nor is the region entirely contained
1327 // in the loop.
1328 // The block EndlessLoop is contained in the region because Region::contains
1329 // tests whether it is not dominated by RegionExit. This is probably to not
1330 // having to query the PostdominatorTree. Instead of an endless loop, a dead
1331 // end can also be formed by an UnreachableInst. This case is already caught
1332 // by isErrorBlock(). We hence only have to reject endless loops here.
1333 if (!hasExitingBlocks(L))
1334 return invalid<ReportLoopHasNoExit>(Context, /*Assert=*/true, L);
1336 // The algorithm for domain construction assumes that loops has only a single
1337 // exit block (and hence corresponds to a subregion). Note that we cannot use
1338 // L->getExitBlock() because it does not check whether all exiting edges point
1339 // to the same BB.
1340 SmallVector<BasicBlock *, 4> ExitBlocks;
1341 L->getExitBlocks(ExitBlocks);
1342 BasicBlock *TheExitBlock = ExitBlocks[0];
1343 for (BasicBlock *ExitBB : ExitBlocks) {
1344 if (TheExitBlock != ExitBB)
1345 return invalid<ReportLoopHasMultipleExits>(Context, /*Assert=*/true, L);
1348 if (canUseISLTripCount(L, Context))
1349 return true;
1351 if (AllowNonAffineSubLoops && AllowNonAffineSubRegions) {
1352 Region *R = RI.getRegionFor(L->getHeader());
1353 while (R != &Context.CurRegion && !R->contains(L))
1354 R = R->getParent();
1356 if (addOverApproximatedRegion(R, Context))
1357 return true;
1360 const SCEV *LoopCount = SE.getBackedgeTakenCount(L);
1361 return invalid<ReportLoopBound>(Context, /*Assert=*/true, L, LoopCount);
1364 /// Return the number of loops in @p L (incl. @p L) that have a trip
1365 /// count that is not known to be less than @MinProfitableTrips.
1366 ScopDetection::LoopStats
1367 ScopDetection::countBeneficialSubLoops(Loop *L, ScalarEvolution &SE,
1368 unsigned MinProfitableTrips) {
1369 auto *TripCount = SE.getBackedgeTakenCount(L);
1371 int NumLoops = 1;
1372 int MaxLoopDepth = 1;
1373 if (MinProfitableTrips > 0)
1374 if (auto *TripCountC = dyn_cast<SCEVConstant>(TripCount))
1375 if (TripCountC->getType()->getScalarSizeInBits() <= 64)
1376 if (TripCountC->getValue()->getZExtValue() <= MinProfitableTrips)
1377 NumLoops -= 1;
1379 for (auto &SubLoop : *L) {
1380 LoopStats Stats = countBeneficialSubLoops(SubLoop, SE, MinProfitableTrips);
1381 NumLoops += Stats.NumLoops;
1382 MaxLoopDepth = std::max(MaxLoopDepth, Stats.MaxDepth + 1);
1385 return {NumLoops, MaxLoopDepth};
1388 ScopDetection::LoopStats
1389 ScopDetection::countBeneficialLoops(Region *R, ScalarEvolution &SE,
1390 LoopInfo &LI, unsigned MinProfitableTrips) {
1391 int LoopNum = 0;
1392 int MaxLoopDepth = 0;
1394 auto L = LI.getLoopFor(R->getEntry());
1396 // If L is fully contained in R, move to first loop surrounding R. Otherwise,
1397 // L is either nullptr or already surrounding R.
1398 if (L && R->contains(L)) {
1399 L = R->outermostLoopInRegion(L);
1400 L = L->getParentLoop();
1403 auto SubLoops =
1404 L ? L->getSubLoopsVector() : std::vector<Loop *>(LI.begin(), LI.end());
1406 for (auto &SubLoop : SubLoops)
1407 if (R->contains(SubLoop)) {
1408 LoopStats Stats =
1409 countBeneficialSubLoops(SubLoop, SE, MinProfitableTrips);
1410 LoopNum += Stats.NumLoops;
1411 MaxLoopDepth = std::max(MaxLoopDepth, Stats.MaxDepth);
1414 return {LoopNum, MaxLoopDepth};
1417 static bool isErrorBlockImpl(BasicBlock &BB, const Region &R, LoopInfo &LI,
1418 const DominatorTree &DT) {
1419 if (isa<UnreachableInst>(BB.getTerminator()))
1420 return true;
1422 if (LI.isLoopHeader(&BB))
1423 return false;
1425 // Don't consider something outside the SCoP as error block. It will precede
1426 // the code versioning runtime check.
1427 if (!R.contains(&BB))
1428 return false;
1430 // Basic blocks that are always executed are not considered error blocks,
1431 // as their execution can not be a rare event.
1432 bool DominatesAllPredecessors = true;
1433 if (R.isTopLevelRegion()) {
1434 for (BasicBlock &I : *R.getEntry()->getParent()) {
1435 if (isa<ReturnInst>(I.getTerminator()) && !DT.dominates(&BB, &I)) {
1436 DominatesAllPredecessors = false;
1437 break;
1440 } else {
1441 for (auto Pred : predecessors(R.getExit())) {
1442 if (R.contains(Pred) && !DT.dominates(&BB, Pred)) {
1443 DominatesAllPredecessors = false;
1444 break;
1449 if (DominatesAllPredecessors)
1450 return false;
1452 for (Instruction &Inst : BB)
1453 if (CallInst *CI = dyn_cast<CallInst>(&Inst)) {
1454 if (isDebugCall(CI))
1455 continue;
1457 if (isIgnoredIntrinsic(CI))
1458 continue;
1460 // memset, memcpy and memmove are modeled intrinsics.
1461 if (isa<MemSetInst>(CI) || isa<MemTransferInst>(CI))
1462 continue;
1464 if (!CI->doesNotAccessMemory())
1465 return true;
1466 if (CI->doesNotReturn())
1467 return true;
1470 return false;
1473 bool ScopDetection::isErrorBlock(llvm::BasicBlock &BB, const llvm::Region &R) {
1474 if (!PollyAllowErrorBlocks)
1475 return false;
1477 auto It = ErrorBlockCache.insert({std::make_pair(&BB, &R), false});
1478 if (!It.second)
1479 return It.first->getSecond();
1481 bool Result = isErrorBlockImpl(BB, R, LI, DT);
1482 It.first->second = Result;
1483 return Result;
1486 Region *ScopDetection::expandRegion(Region &R) {
1487 // Initial no valid region was found (greater than R)
1488 std::unique_ptr<Region> LastValidRegion;
1489 auto ExpandedRegion = std::unique_ptr<Region>(R.getExpandedRegion());
1491 POLLY_DEBUG(dbgs() << "\tExpanding " << R.getNameStr() << "\n");
1493 while (ExpandedRegion) {
1494 BBPair P = getBBPairForRegion(ExpandedRegion.get());
1495 std::unique_ptr<DetectionContext> &Entry = DetectionContextMap[P];
1496 Entry = std::make_unique<DetectionContext>(*ExpandedRegion, AA,
1497 /*Verifying=*/false);
1498 DetectionContext &Context = *Entry.get();
1500 POLLY_DEBUG(dbgs() << "\t\tTrying " << ExpandedRegion->getNameStr()
1501 << "\n");
1502 // Only expand when we did not collect errors.
1504 if (!Context.Log.hasErrors()) {
1505 // If the exit is valid check all blocks
1506 // - if true, a valid region was found => store it + keep expanding
1507 // - if false, .tbd. => stop (should this really end the loop?)
1508 if (!allBlocksValid(Context) || Context.Log.hasErrors()) {
1509 removeCachedResults(*ExpandedRegion);
1510 DetectionContextMap.erase(P);
1511 break;
1514 // Store this region, because it is the greatest valid (encountered so
1515 // far).
1516 if (LastValidRegion) {
1517 removeCachedResults(*LastValidRegion);
1518 DetectionContextMap.erase(P);
1520 LastValidRegion = std::move(ExpandedRegion);
1522 // Create and test the next greater region (if any)
1523 ExpandedRegion =
1524 std::unique_ptr<Region>(LastValidRegion->getExpandedRegion());
1526 } else {
1527 // Create and test the next greater region (if any)
1528 removeCachedResults(*ExpandedRegion);
1529 DetectionContextMap.erase(P);
1530 ExpandedRegion =
1531 std::unique_ptr<Region>(ExpandedRegion->getExpandedRegion());
1535 POLLY_DEBUG({
1536 if (LastValidRegion)
1537 dbgs() << "\tto " << LastValidRegion->getNameStr() << "\n";
1538 else
1539 dbgs() << "\tExpanding " << R.getNameStr() << " failed\n";
1542 return LastValidRegion.release();
1545 static bool regionWithoutLoops(Region &R, LoopInfo &LI) {
1546 for (const BasicBlock *BB : R.blocks())
1547 if (R.contains(LI.getLoopFor(BB)))
1548 return false;
1550 return true;
1553 void ScopDetection::removeCachedResultsRecursively(const Region &R) {
1554 for (auto &SubRegion : R) {
1555 if (ValidRegions.count(SubRegion.get())) {
1556 removeCachedResults(*SubRegion.get());
1557 } else
1558 removeCachedResultsRecursively(*SubRegion);
1562 void ScopDetection::removeCachedResults(const Region &R) {
1563 ValidRegions.remove(&R);
1566 void ScopDetection::findScops(Region &R) {
1567 std::unique_ptr<DetectionContext> &Entry =
1568 DetectionContextMap[getBBPairForRegion(&R)];
1569 Entry = std::make_unique<DetectionContext>(R, AA, /*Verifying=*/false);
1570 DetectionContext &Context = *Entry.get();
1572 bool DidBailout = true;
1573 if (!PollyProcessUnprofitable && regionWithoutLoops(R, LI))
1574 invalid<ReportUnprofitable>(Context, /*Assert=*/true, &R);
1575 else
1576 DidBailout = !isValidRegion(Context);
1578 (void)DidBailout;
1579 if (KeepGoing) {
1580 assert((!DidBailout || Context.IsInvalid) &&
1581 "With -polly-detect-keep-going, it is sufficient that if "
1582 "isValidRegion short-circuited, that SCoP is invalid");
1583 } else {
1584 assert(DidBailout == Context.IsInvalid &&
1585 "isValidRegion must short-circuit iff the ScoP is invalid");
1588 if (Context.IsInvalid) {
1589 removeCachedResults(R);
1590 } else {
1591 ValidRegions.insert(&R);
1592 return;
1595 for (auto &SubRegion : R)
1596 findScops(*SubRegion);
1598 // Try to expand regions.
1600 // As the region tree normally only contains canonical regions, non canonical
1601 // regions that form a Scop are not found. Therefore, those non canonical
1602 // regions are checked by expanding the canonical ones.
1604 std::vector<Region *> ToExpand;
1606 for (auto &SubRegion : R)
1607 ToExpand.push_back(SubRegion.get());
1609 for (Region *CurrentRegion : ToExpand) {
1610 // Skip invalid regions. Regions may become invalid, if they are element of
1611 // an already expanded region.
1612 if (!ValidRegions.count(CurrentRegion))
1613 continue;
1615 // Skip regions that had errors.
1616 bool HadErrors = lookupRejectionLog(CurrentRegion)->hasErrors();
1617 if (HadErrors)
1618 continue;
1620 Region *ExpandedR = expandRegion(*CurrentRegion);
1622 if (!ExpandedR)
1623 continue;
1625 R.addSubRegion(ExpandedR, true);
1626 ValidRegions.insert(ExpandedR);
1627 removeCachedResults(*CurrentRegion);
1628 removeCachedResultsRecursively(*ExpandedR);
1632 bool ScopDetection::allBlocksValid(DetectionContext &Context) {
1633 Region &CurRegion = Context.CurRegion;
1635 for (const BasicBlock *BB : CurRegion.blocks()) {
1636 Loop *L = LI.getLoopFor(BB);
1637 if (L && L->getHeader() == BB) {
1638 if (CurRegion.contains(L)) {
1639 if (!isValidLoop(L, Context)) {
1640 Context.IsInvalid = true;
1641 if (!KeepGoing)
1642 return false;
1644 } else {
1645 SmallVector<BasicBlock *, 1> Latches;
1646 L->getLoopLatches(Latches);
1647 for (BasicBlock *Latch : Latches)
1648 if (CurRegion.contains(Latch))
1649 return invalid<ReportLoopOnlySomeLatches>(Context, /*Assert=*/true,
1655 for (BasicBlock *BB : CurRegion.blocks()) {
1656 bool IsErrorBlock = isErrorBlock(*BB, CurRegion);
1658 // Also check exception blocks (and possibly register them as non-affine
1659 // regions). Even though exception blocks are not modeled, we use them
1660 // to forward-propagate domain constraints during ScopInfo construction.
1661 if (!isValidCFG(*BB, false, IsErrorBlock, Context) && !KeepGoing)
1662 return false;
1664 if (IsErrorBlock)
1665 continue;
1667 for (BasicBlock::iterator I = BB->begin(), E = --BB->end(); I != E; ++I)
1668 if (!isValidInstruction(*I, Context)) {
1669 Context.IsInvalid = true;
1670 if (!KeepGoing)
1671 return false;
1675 if (!hasAffineMemoryAccesses(Context))
1676 return false;
1678 return true;
1681 bool ScopDetection::hasSufficientCompute(DetectionContext &Context,
1682 int NumLoops) const {
1683 int InstCount = 0;
1685 if (NumLoops == 0)
1686 return false;
1688 for (auto *BB : Context.CurRegion.blocks())
1689 if (Context.CurRegion.contains(LI.getLoopFor(BB)))
1690 InstCount += BB->size();
1692 InstCount = InstCount / NumLoops;
1694 return InstCount >= ProfitabilityMinPerLoopInstructions;
1697 bool ScopDetection::hasPossiblyDistributableLoop(
1698 DetectionContext &Context) const {
1699 for (auto *BB : Context.CurRegion.blocks()) {
1700 auto *L = LI.getLoopFor(BB);
1701 if (!L)
1702 continue;
1703 if (!Context.CurRegion.contains(L))
1704 continue;
1705 if (Context.BoxedLoopsSet.count(L))
1706 continue;
1707 unsigned StmtsWithStoresInLoops = 0;
1708 for (auto *LBB : L->blocks()) {
1709 bool MemStore = false;
1710 for (auto &I : *LBB)
1711 MemStore |= isa<StoreInst>(&I);
1712 StmtsWithStoresInLoops += MemStore;
1714 return (StmtsWithStoresInLoops > 1);
1716 return false;
1719 bool ScopDetection::isProfitableRegion(DetectionContext &Context) const {
1720 Region &CurRegion = Context.CurRegion;
1722 if (PollyProcessUnprofitable)
1723 return true;
1725 // We can probably not do a lot on scops that only write or only read
1726 // data.
1727 if (!Context.hasStores || !Context.hasLoads)
1728 return invalid<ReportUnprofitable>(Context, /*Assert=*/true, &CurRegion);
1730 int NumLoops =
1731 countBeneficialLoops(&CurRegion, SE, LI, MIN_LOOP_TRIP_COUNT).NumLoops;
1732 int NumAffineLoops = NumLoops - Context.BoxedLoopsSet.size();
1734 // Scops with at least two loops may allow either loop fusion or tiling and
1735 // are consequently interesting to look at.
1736 if (NumAffineLoops >= 2)
1737 return true;
1739 // A loop with multiple non-trivial blocks might be amendable to distribution.
1740 if (NumAffineLoops == 1 && hasPossiblyDistributableLoop(Context))
1741 return true;
1743 // Scops that contain a loop with a non-trivial amount of computation per
1744 // loop-iteration are interesting as we may be able to parallelize such
1745 // loops. Individual loops that have only a small amount of computation
1746 // per-iteration are performance-wise very fragile as any change to the
1747 // loop induction variables may affect performance. To not cause spurious
1748 // performance regressions, we do not consider such loops.
1749 if (NumAffineLoops == 1 && hasSufficientCompute(Context, NumLoops))
1750 return true;
1752 return invalid<ReportUnprofitable>(Context, /*Assert=*/true, &CurRegion);
1755 bool ScopDetection::isValidRegion(DetectionContext &Context) {
1756 Region &CurRegion = Context.CurRegion;
1758 POLLY_DEBUG(dbgs() << "Checking region: " << CurRegion.getNameStr()
1759 << "\n\t");
1761 if (!PollyAllowFullFunction && CurRegion.isTopLevelRegion()) {
1762 POLLY_DEBUG(dbgs() << "Top level region is invalid\n");
1763 Context.IsInvalid = true;
1764 return false;
1767 DebugLoc DbgLoc;
1768 if (CurRegion.getExit() &&
1769 isa<UnreachableInst>(CurRegion.getExit()->getTerminator())) {
1770 POLLY_DEBUG(dbgs() << "Unreachable in exit\n");
1771 return invalid<ReportUnreachableInExit>(Context, /*Assert=*/true,
1772 CurRegion.getExit(), DbgLoc);
1775 if (!OnlyRegion.empty() &&
1776 !CurRegion.getEntry()->getName().count(OnlyRegion)) {
1777 POLLY_DEBUG({
1778 dbgs() << "Region entry does not match -polly-only-region";
1779 dbgs() << "\n";
1781 Context.IsInvalid = true;
1782 return false;
1785 for (BasicBlock *Pred : predecessors(CurRegion.getEntry())) {
1786 Instruction *PredTerm = Pred->getTerminator();
1787 if (isa<IndirectBrInst>(PredTerm) || isa<CallBrInst>(PredTerm))
1788 return invalid<ReportIndirectPredecessor>(
1789 Context, /*Assert=*/true, PredTerm, PredTerm->getDebugLoc());
1792 // SCoP cannot contain the entry block of the function, because we need
1793 // to insert alloca instruction there when translate scalar to array.
1794 if (!PollyAllowFullFunction &&
1795 CurRegion.getEntry() ==
1796 &(CurRegion.getEntry()->getParent()->getEntryBlock()))
1797 return invalid<ReportEntry>(Context, /*Assert=*/true, CurRegion.getEntry());
1799 if (!allBlocksValid(Context)) {
1800 // TODO: Every failure condition within allBlocksValid should call
1801 // invalid<Reason>(). Otherwise we reject SCoPs without giving feedback to
1802 // the user.
1803 Context.IsInvalid = true;
1804 return false;
1807 if (!isReducibleRegion(CurRegion, DbgLoc))
1808 return invalid<ReportIrreducibleRegion>(Context, /*Assert=*/true,
1809 &CurRegion, DbgLoc);
1811 POLLY_DEBUG(dbgs() << "OK\n");
1812 return true;
1815 void ScopDetection::markFunctionAsInvalid(Function *F) {
1816 F->addFnAttr(PollySkipFnAttr);
1819 bool ScopDetection::isValidFunction(Function &F) {
1820 return !F.hasFnAttribute(PollySkipFnAttr);
1823 void ScopDetection::printLocations(Function &F) {
1824 for (const Region *R : *this) {
1825 unsigned LineEntry, LineExit;
1826 std::string FileName;
1828 getDebugLocation(R, LineEntry, LineExit, FileName);
1829 DiagnosticScopFound Diagnostic(F, FileName, LineEntry, LineExit);
1830 F.getContext().diagnose(Diagnostic);
1834 void ScopDetection::emitMissedRemarks(const Function &F) {
1835 for (auto &DIt : DetectionContextMap) {
1836 DetectionContext &DC = *DIt.getSecond().get();
1837 if (DC.Log.hasErrors())
1838 emitRejectionRemarks(DIt.getFirst(), DC.Log, ORE);
1842 bool ScopDetection::isReducibleRegion(Region &R, DebugLoc &DbgLoc) const {
1843 /// Enum for coloring BBs in Region.
1845 /// WHITE - Unvisited BB in DFS walk.
1846 /// GREY - BBs which are currently on the DFS stack for processing.
1847 /// BLACK - Visited and completely processed BB.
1848 enum Color { WHITE, GREY, BLACK };
1850 BasicBlock *REntry = R.getEntry();
1851 BasicBlock *RExit = R.getExit();
1852 // Map to match the color of a BasicBlock during the DFS walk.
1853 DenseMap<const BasicBlock *, Color> BBColorMap;
1854 // Stack keeping track of current BB and index of next child to be processed.
1855 std::stack<std::pair<BasicBlock *, unsigned>> DFSStack;
1857 unsigned AdjacentBlockIndex = 0;
1858 BasicBlock *CurrBB, *SuccBB;
1859 CurrBB = REntry;
1861 // Initialize the map for all BB with WHITE color.
1862 for (auto *BB : R.blocks())
1863 BBColorMap[BB] = WHITE;
1865 // Process the entry block of the Region.
1866 BBColorMap[CurrBB] = GREY;
1867 DFSStack.push(std::make_pair(CurrBB, 0));
1869 while (!DFSStack.empty()) {
1870 // Get next BB on stack to be processed.
1871 CurrBB = DFSStack.top().first;
1872 AdjacentBlockIndex = DFSStack.top().second;
1873 DFSStack.pop();
1875 // Loop to iterate over the successors of current BB.
1876 const Instruction *TInst = CurrBB->getTerminator();
1877 unsigned NSucc = TInst->getNumSuccessors();
1878 for (unsigned I = AdjacentBlockIndex; I < NSucc;
1879 ++I, ++AdjacentBlockIndex) {
1880 SuccBB = TInst->getSuccessor(I);
1882 // Checks for region exit block and self-loops in BB.
1883 if (SuccBB == RExit || SuccBB == CurrBB)
1884 continue;
1886 // WHITE indicates an unvisited BB in DFS walk.
1887 if (BBColorMap[SuccBB] == WHITE) {
1888 // Push the current BB and the index of the next child to be visited.
1889 DFSStack.push(std::make_pair(CurrBB, I + 1));
1890 // Push the next BB to be processed.
1891 DFSStack.push(std::make_pair(SuccBB, 0));
1892 // First time the BB is being processed.
1893 BBColorMap[SuccBB] = GREY;
1894 break;
1895 } else if (BBColorMap[SuccBB] == GREY) {
1896 // GREY indicates a loop in the control flow.
1897 // If the destination dominates the source, it is a natural loop
1898 // else, an irreducible control flow in the region is detected.
1899 if (!DT.dominates(SuccBB, CurrBB)) {
1900 // Get debug info of instruction which causes irregular control flow.
1901 DbgLoc = TInst->getDebugLoc();
1902 return false;
1907 // If all children of current BB have been processed,
1908 // then mark that BB as fully processed.
1909 if (AdjacentBlockIndex == NSucc)
1910 BBColorMap[CurrBB] = BLACK;
1913 return true;
1916 static void updateLoopCountStatistic(ScopDetection::LoopStats Stats,
1917 bool OnlyProfitable) {
1918 if (!OnlyProfitable) {
1919 NumLoopsInScop += Stats.NumLoops;
1920 MaxNumLoopsInScop =
1921 std::max(MaxNumLoopsInScop.getValue(), (uint64_t)Stats.NumLoops);
1922 if (Stats.MaxDepth == 0)
1923 NumScopsDepthZero++;
1924 else if (Stats.MaxDepth == 1)
1925 NumScopsDepthOne++;
1926 else if (Stats.MaxDepth == 2)
1927 NumScopsDepthTwo++;
1928 else if (Stats.MaxDepth == 3)
1929 NumScopsDepthThree++;
1930 else if (Stats.MaxDepth == 4)
1931 NumScopsDepthFour++;
1932 else if (Stats.MaxDepth == 5)
1933 NumScopsDepthFive++;
1934 else
1935 NumScopsDepthLarger++;
1936 } else {
1937 NumLoopsInProfScop += Stats.NumLoops;
1938 MaxNumLoopsInProfScop =
1939 std::max(MaxNumLoopsInProfScop.getValue(), (uint64_t)Stats.NumLoops);
1940 if (Stats.MaxDepth == 0)
1941 NumProfScopsDepthZero++;
1942 else if (Stats.MaxDepth == 1)
1943 NumProfScopsDepthOne++;
1944 else if (Stats.MaxDepth == 2)
1945 NumProfScopsDepthTwo++;
1946 else if (Stats.MaxDepth == 3)
1947 NumProfScopsDepthThree++;
1948 else if (Stats.MaxDepth == 4)
1949 NumProfScopsDepthFour++;
1950 else if (Stats.MaxDepth == 5)
1951 NumProfScopsDepthFive++;
1952 else
1953 NumProfScopsDepthLarger++;
1957 ScopDetection::DetectionContext *
1958 ScopDetection::getDetectionContext(const Region *R) const {
1959 auto DCMIt = DetectionContextMap.find(getBBPairForRegion(R));
1960 if (DCMIt == DetectionContextMap.end())
1961 return nullptr;
1962 return DCMIt->second.get();
1965 const RejectLog *ScopDetection::lookupRejectionLog(const Region *R) const {
1966 const DetectionContext *DC = getDetectionContext(R);
1967 return DC ? &DC->Log : nullptr;
1970 void ScopDetection::verifyRegion(const Region &R) {
1971 assert(isMaxRegionInScop(R) && "Expect R is a valid region.");
1973 DetectionContext Context(const_cast<Region &>(R), AA, true /*verifying*/);
1974 isValidRegion(Context);
1977 void ScopDetection::verifyAnalysis() {
1978 if (!VerifyScops)
1979 return;
1981 for (const Region *R : ValidRegions)
1982 verifyRegion(*R);
1985 bool ScopDetectionWrapperPass::runOnFunction(Function &F) {
1986 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1987 auto &RI = getAnalysis<RegionInfoPass>().getRegionInfo();
1988 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
1989 auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1990 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1991 auto &ORE = getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
1993 Result = std::make_unique<ScopDetection>(DT, SE, LI, RI, AA, ORE);
1994 Result->detect(F);
1995 return false;
1998 void ScopDetectionWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1999 AU.addRequired<LoopInfoWrapperPass>();
2000 AU.addRequiredTransitive<ScalarEvolutionWrapperPass>();
2001 AU.addRequired<DominatorTreeWrapperPass>();
2002 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
2003 // We also need AA and RegionInfo when we are verifying analysis.
2004 AU.addRequiredTransitive<AAResultsWrapperPass>();
2005 AU.addRequiredTransitive<RegionInfoPass>();
2006 AU.setPreservesAll();
2009 void ScopDetectionWrapperPass::print(raw_ostream &OS, const Module *) const {
2010 for (const Region *R : Result->ValidRegions)
2011 OS << "Valid Region for Scop: " << R->getNameStr() << '\n';
2013 OS << "\n";
2016 ScopDetectionWrapperPass::ScopDetectionWrapperPass() : FunctionPass(ID) {
2017 // Disable runtime alias checks if we ignore aliasing all together.
2018 if (IgnoreAliasing)
2019 PollyUseRuntimeAliasChecks = false;
2022 ScopAnalysis::ScopAnalysis() {
2023 // Disable runtime alias checks if we ignore aliasing all together.
2024 if (IgnoreAliasing)
2025 PollyUseRuntimeAliasChecks = false;
2028 void ScopDetectionWrapperPass::releaseMemory() { Result.reset(); }
2030 char ScopDetectionWrapperPass::ID;
2032 AnalysisKey ScopAnalysis::Key;
2034 ScopDetection ScopAnalysis::run(Function &F, FunctionAnalysisManager &FAM) {
2035 auto &LI = FAM.getResult<LoopAnalysis>(F);
2036 auto &RI = FAM.getResult<RegionInfoAnalysis>(F);
2037 auto &AA = FAM.getResult<AAManager>(F);
2038 auto &SE = FAM.getResult<ScalarEvolutionAnalysis>(F);
2039 auto &DT = FAM.getResult<DominatorTreeAnalysis>(F);
2040 auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
2042 ScopDetection Result(DT, SE, LI, RI, AA, ORE);
2043 Result.detect(F);
2044 return Result;
2047 PreservedAnalyses ScopAnalysisPrinterPass::run(Function &F,
2048 FunctionAnalysisManager &FAM) {
2049 OS << "Detected Scops in Function " << F.getName() << "\n";
2050 auto &SD = FAM.getResult<ScopAnalysis>(F);
2051 for (const Region *R : SD.ValidRegions)
2052 OS << "Valid Region for Scop: " << R->getNameStr() << '\n';
2054 OS << "\n";
2055 return PreservedAnalyses::all();
2058 Pass *polly::createScopDetectionWrapperPassPass() {
2059 return new ScopDetectionWrapperPass();
2062 INITIALIZE_PASS_BEGIN(ScopDetectionWrapperPass, "polly-detect",
2063 "Polly - Detect static control parts (SCoPs)", false,
2064 false);
2065 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass);
2066 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass);
2067 INITIALIZE_PASS_DEPENDENCY(RegionInfoPass);
2068 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass);
2069 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass);
2070 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass);
2071 INITIALIZE_PASS_END(ScopDetectionWrapperPass, "polly-detect",
2072 "Polly - Detect static control parts (SCoPs)", false, false)
2074 //===----------------------------------------------------------------------===//
2076 namespace {
2077 /// Print result from ScopDetectionWrapperPass.
2078 class ScopDetectionPrinterLegacyPass final : public FunctionPass {
2079 public:
2080 static char ID;
2082 ScopDetectionPrinterLegacyPass() : ScopDetectionPrinterLegacyPass(outs()) {}
2084 explicit ScopDetectionPrinterLegacyPass(llvm::raw_ostream &OS)
2085 : FunctionPass(ID), OS(OS) {}
2087 bool runOnFunction(Function &F) override {
2088 ScopDetectionWrapperPass &P = getAnalysis<ScopDetectionWrapperPass>();
2090 OS << "Printing analysis '" << P.getPassName() << "' for function '"
2091 << F.getName() << "':\n";
2092 P.print(OS);
2094 return false;
2097 void getAnalysisUsage(AnalysisUsage &AU) const override {
2098 FunctionPass::getAnalysisUsage(AU);
2099 AU.addRequired<ScopDetectionWrapperPass>();
2100 AU.setPreservesAll();
2103 private:
2104 llvm::raw_ostream &OS;
2107 char ScopDetectionPrinterLegacyPass::ID = 0;
2108 } // namespace
2110 Pass *polly::createScopDetectionPrinterLegacyPass(raw_ostream &OS) {
2111 return new ScopDetectionPrinterLegacyPass(OS);
2114 INITIALIZE_PASS_BEGIN(ScopDetectionPrinterLegacyPass, "polly-print-detect",
2115 "Polly - Print static control parts (SCoPs)", false,
2116 false);
2117 INITIALIZE_PASS_DEPENDENCY(ScopDetectionWrapperPass);
2118 INITIALIZE_PASS_END(ScopDetectionPrinterLegacyPass, "polly-print-detect",
2119 "Polly - Print static control parts (SCoPs)", false, false)