1 //===- LoopUnrollAndJam.cpp - Loop unroll and jam pass --------------------===//
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 pass implements an unroll and jam pass. Most of the work is done by
10 // Utils/UnrollLoopAndJam.cpp.
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/Transforms/Scalar/LoopUnrollAndJamPass.h"
14 #include "llvm/ADT/ArrayRef.h"
15 #include "llvm/ADT/PriorityWorklist.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/ADT/StringRef.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/CodeMetrics.h"
20 #include "llvm/Analysis/DependenceAnalysis.h"
21 #include "llvm/Analysis/LoopAnalysisManager.h"
22 #include "llvm/Analysis/LoopInfo.h"
23 #include "llvm/Analysis/LoopNestAnalysis.h"
24 #include "llvm/Analysis/LoopPass.h"
25 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
26 #include "llvm/Analysis/ScalarEvolution.h"
27 #include "llvm/Analysis/TargetTransformInfo.h"
28 #include "llvm/IR/BasicBlock.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/Metadata.h"
34 #include "llvm/IR/PassManager.h"
35 #include "llvm/Support/Casting.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Compiler.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Transforms/Scalar/LoopPassManager.h"
41 #include "llvm/Transforms/Utils/LoopPeel.h"
42 #include "llvm/Transforms/Utils/LoopUtils.h"
43 #include "llvm/Transforms/Utils/UnrollLoop.h"
44 #include <cassert>
45 #include <cstdint>
46 
47 namespace llvm {
48 class Instruction;
49 class Value;
50 } // namespace llvm
51 
52 using namespace llvm;
53 
54 #define DEBUG_TYPE "loop-unroll-and-jam"
55 
56 /// @{
57 /// Metadata attribute names
58 static const char *const LLVMLoopUnrollAndJamFollowupAll =
59     "llvm.loop.unroll_and_jam.followup_all";
60 static const char *const LLVMLoopUnrollAndJamFollowupInner =
61     "llvm.loop.unroll_and_jam.followup_inner";
62 static const char *const LLVMLoopUnrollAndJamFollowupOuter =
63     "llvm.loop.unroll_and_jam.followup_outer";
64 static const char *const LLVMLoopUnrollAndJamFollowupRemainderInner =
65     "llvm.loop.unroll_and_jam.followup_remainder_inner";
66 static const char *const LLVMLoopUnrollAndJamFollowupRemainderOuter =
67     "llvm.loop.unroll_and_jam.followup_remainder_outer";
68 /// @}
69 
70 static cl::opt<bool>
71     AllowUnrollAndJam("allow-unroll-and-jam", cl::Hidden,
72                       cl::desc("Allows loops to be unroll-and-jammed."));
73 
74 static cl::opt<unsigned> UnrollAndJamCount(
75     "unroll-and-jam-count", cl::Hidden,
76     cl::desc("Use this unroll count for all loops including those with "
77              "unroll_and_jam_count pragma values, for testing purposes"));
78 
79 static cl::opt<unsigned> UnrollAndJamThreshold(
80     "unroll-and-jam-threshold", cl::init(60), cl::Hidden,
81     cl::desc("Threshold to use for inner loop when doing unroll and jam."));
82 
83 static cl::opt<unsigned> PragmaUnrollAndJamThreshold(
84     "pragma-unroll-and-jam-threshold", cl::init(1024), cl::Hidden,
85     cl::desc("Unrolled size limit for loops with an unroll_and_jam(full) or "
86              "unroll_count pragma."));
87 
88 // Returns the loop hint metadata node with the given name (for example,
89 // "llvm.loop.unroll.count").  If no such metadata node exists, then nullptr is
90 // returned.
91 static MDNode *getUnrollMetadataForLoop(const Loop *L, StringRef Name) {
92   if (MDNode *LoopID = L->getLoopID())
93     return GetUnrollMetadata(LoopID, Name);
94   return nullptr;
95 }
96 
97 // Returns true if the loop has any metadata starting with Prefix. For example a
98 // Prefix of "llvm.loop.unroll." returns true if we have any unroll metadata.
99 static bool hasAnyUnrollPragma(const Loop *L, StringRef Prefix) {
100   if (MDNode *LoopID = L->getLoopID()) {
101     // First operand should refer to the loop id itself.
102     assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
103     assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
104 
105     for (unsigned I = 1, E = LoopID->getNumOperands(); I < E; ++I) {
106       MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(I));
107       if (!MD)
108         continue;
109 
110       MDString *S = dyn_cast<MDString>(MD->getOperand(0));
111       if (!S)
112         continue;
113 
114       if (S->getString().starts_with(Prefix))
115         return true;
116     }
117   }
118   return false;
119 }
120 
121 // Returns true if the loop has an unroll_and_jam(enable) pragma.
122 static bool hasUnrollAndJamEnablePragma(const Loop *L) {
123   return getUnrollMetadataForLoop(L, "llvm.loop.unroll_and_jam.enable");
124 }
125 
126 // If loop has an unroll_and_jam_count pragma return the (necessarily
127 // positive) value from the pragma.  Otherwise return 0.
128 static unsigned unrollAndJamCountPragmaValue(const Loop *L) {
129   MDNode *MD = getUnrollMetadataForLoop(L, "llvm.loop.unroll_and_jam.count");
130   if (MD) {
131     assert(MD->getNumOperands() == 2 &&
132            "Unroll count hint metadata should have two operands.");
133     unsigned Count =
134         mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
135     assert(Count >= 1 && "Unroll count must be positive.");
136     return Count;
137   }
138   return 0;
139 }
140 
141 // Returns loop size estimation for unrolled loop.
142 static uint64_t
143 getUnrollAndJammedLoopSize(unsigned LoopSize,
144                            TargetTransformInfo::UnrollingPreferences &UP) {
145   assert(LoopSize >= UP.BEInsns && "LoopSize should not be less than BEInsns!");
146   return static_cast<uint64_t>(LoopSize - UP.BEInsns) * UP.Count + UP.BEInsns;
147 }
148 
149 // Calculates unroll and jam count and writes it to UP.Count. Returns true if
150 // unroll count was set explicitly.
151 static bool computeUnrollAndJamCount(
152     Loop *L, Loop *SubLoop, const TargetTransformInfo &TTI, DominatorTree &DT,
153     LoopInfo *LI, AssumptionCache *AC, ScalarEvolution &SE,
154     const SmallPtrSetImpl<const Value *> &EphValues,
155     OptimizationRemarkEmitter *ORE, unsigned OuterTripCount,
156     unsigned OuterTripMultiple, const UnrollCostEstimator &OuterUCE,
157     unsigned InnerTripCount, unsigned InnerLoopSize,
158     TargetTransformInfo::UnrollingPreferences &UP,
159     TargetTransformInfo::PeelingPreferences &PP) {
160   unsigned OuterLoopSize = OuterUCE.getRolledLoopSize();
161   // First up use computeUnrollCount from the loop unroller to get a count
162   // for unrolling the outer loop, plus any loops requiring explicit
163   // unrolling we leave to the unroller. This uses UP.Threshold /
164   // UP.PartialThreshold / UP.MaxCount to come up with sensible loop values.
165   // We have already checked that the loop has no unroll.* pragmas.
166   unsigned MaxTripCount = 0;
167   bool UseUpperBound = false;
168   bool ExplicitUnroll = computeUnrollCount(
169     L, TTI, DT, LI, AC, SE, EphValues, ORE, OuterTripCount, MaxTripCount,
170       /*MaxOrZero*/ false, OuterTripMultiple, OuterUCE, UP, PP,
171       UseUpperBound);
172   if (ExplicitUnroll || UseUpperBound) {
173     // If the user explicitly set the loop as unrolled, dont UnJ it. Leave it
174     // for the unroller instead.
175     LLVM_DEBUG(dbgs() << "Won't unroll-and-jam; explicit count set by "
176                          "computeUnrollCount\n");
177     UP.Count = 0;
178     return false;
179   }
180 
181   // Override with any explicit Count from the "unroll-and-jam-count" option.
182   bool UserUnrollCount = UnrollAndJamCount.getNumOccurrences() > 0;
183   if (UserUnrollCount) {
184     UP.Count = UnrollAndJamCount;
185     UP.Force = true;
186     if (UP.AllowRemainder &&
187         getUnrollAndJammedLoopSize(OuterLoopSize, UP) < UP.Threshold &&
188         getUnrollAndJammedLoopSize(InnerLoopSize, UP) <
189             UP.UnrollAndJamInnerLoopThreshold)
190       return true;
191   }
192 
193   // Check for unroll_and_jam pragmas
194   unsigned PragmaCount = unrollAndJamCountPragmaValue(L);
195   if (PragmaCount > 0) {
196     UP.Count = PragmaCount;
197     UP.Runtime = true;
198     UP.Force = true;
199     if ((UP.AllowRemainder || (OuterTripMultiple % PragmaCount == 0)) &&
200         getUnrollAndJammedLoopSize(OuterLoopSize, UP) < UP.Threshold &&
201         getUnrollAndJammedLoopSize(InnerLoopSize, UP) <
202             UP.UnrollAndJamInnerLoopThreshold)
203       return true;
204   }
205 
206   bool PragmaEnableUnroll = hasUnrollAndJamEnablePragma(L);
207   bool ExplicitUnrollAndJamCount = PragmaCount > 0 || UserUnrollCount;
208   bool ExplicitUnrollAndJam = PragmaEnableUnroll || ExplicitUnrollAndJamCount;
209 
210   // If the loop has an unrolling pragma, we want to be more aggressive with
211   // unrolling limits.
212   if (ExplicitUnrollAndJam)
213     UP.UnrollAndJamInnerLoopThreshold = PragmaUnrollAndJamThreshold;
214 
215   if (!UP.AllowRemainder && getUnrollAndJammedLoopSize(InnerLoopSize, UP) >=
216                                 UP.UnrollAndJamInnerLoopThreshold) {
217     LLVM_DEBUG(dbgs() << "Won't unroll-and-jam; can't create remainder and "
218                          "inner loop too large\n");
219     UP.Count = 0;
220     return false;
221   }
222 
223   // We have a sensible limit for the outer loop, now adjust it for the inner
224   // loop and UP.UnrollAndJamInnerLoopThreshold. If the outer limit was set
225   // explicitly, we want to stick to it.
226   if (!ExplicitUnrollAndJamCount && UP.AllowRemainder) {
227     while (UP.Count != 0 && getUnrollAndJammedLoopSize(InnerLoopSize, UP) >=
228                                 UP.UnrollAndJamInnerLoopThreshold)
229       UP.Count--;
230   }
231 
232   // If we are explicitly unroll and jamming, we are done. Otherwise there are a
233   // number of extra performance heuristics to check.
234   if (ExplicitUnrollAndJam)
235     return true;
236 
237   // If the inner loop count is known and small, leave the entire loop nest to
238   // be the unroller
239   if (InnerTripCount && InnerLoopSize * InnerTripCount < UP.Threshold) {
240     LLVM_DEBUG(dbgs() << "Won't unroll-and-jam; small inner loop count is "
241                          "being left for the unroller\n");
242     UP.Count = 0;
243     return false;
244   }
245 
246   // Check for situations where UnJ is likely to be unprofitable. Including
247   // subloops with more than 1 block.
248   if (SubLoop->getBlocks().size() != 1) {
249     LLVM_DEBUG(
250         dbgs() << "Won't unroll-and-jam; More than one inner loop block\n");
251     UP.Count = 0;
252     return false;
253   }
254 
255   // Limit to loops where there is something to gain from unrolling and
256   // jamming the loop. In this case, look for loads that are invariant in the
257   // outer loop and can become shared.
258   unsigned NumInvariant = 0;
259   for (BasicBlock *BB : SubLoop->getBlocks()) {
260     for (Instruction &I : *BB) {
261       if (auto *Ld = dyn_cast<LoadInst>(&I)) {
262         Value *V = Ld->getPointerOperand();
263         const SCEV *LSCEV = SE.getSCEVAtScope(V, L);
264         if (SE.isLoopInvariant(LSCEV, L))
265           NumInvariant++;
266       }
267     }
268   }
269   if (NumInvariant == 0) {
270     LLVM_DEBUG(dbgs() << "Won't unroll-and-jam; No loop invariant loads\n");
271     UP.Count = 0;
272     return false;
273   }
274 
275   return false;
276 }
277 
278 static LoopUnrollResult
279 tryToUnrollAndJamLoop(Loop *L, DominatorTree &DT, LoopInfo *LI,
280                       ScalarEvolution &SE, const TargetTransformInfo &TTI,
281                       AssumptionCache &AC, DependenceInfo &DI,
282                       OptimizationRemarkEmitter &ORE, int OptLevel) {
283   TargetTransformInfo::UnrollingPreferences UP = gatherUnrollingPreferences(
284       L, SE, TTI, nullptr, nullptr, ORE, OptLevel, std::nullopt, std::nullopt,
285       std::nullopt, std::nullopt, std::nullopt, std::nullopt);
286   TargetTransformInfo::PeelingPreferences PP =
287       gatherPeelingPreferences(L, SE, TTI, std::nullopt, std::nullopt);
288 
289   TransformationMode EnableMode = hasUnrollAndJamTransformation(L);
290   if (EnableMode & TM_Disable)
291     return LoopUnrollResult::Unmodified;
292   if (EnableMode & TM_ForcedByUser)
293     UP.UnrollAndJam = true;
294 
295   if (AllowUnrollAndJam.getNumOccurrences() > 0)
296     UP.UnrollAndJam = AllowUnrollAndJam;
297   if (UnrollAndJamThreshold.getNumOccurrences() > 0)
298     UP.UnrollAndJamInnerLoopThreshold = UnrollAndJamThreshold;
299   // Exit early if unrolling is disabled.
300   if (!UP.UnrollAndJam || UP.UnrollAndJamInnerLoopThreshold == 0)
301     return LoopUnrollResult::Unmodified;
302 
303   LLVM_DEBUG(dbgs() << "Loop Unroll and Jam: F["
304                     << L->getHeader()->getParent()->getName() << "] Loop %"
305                     << L->getHeader()->getName() << "\n");
306 
307   // A loop with any unroll pragma (enabling/disabling/count/etc) is left for
308   // the unroller, so long as it does not explicitly have unroll_and_jam
309   // metadata. This means #pragma nounroll will disable unroll and jam as well
310   // as unrolling
311   if (hasAnyUnrollPragma(L, "llvm.loop.unroll.") &&
312       !hasAnyUnrollPragma(L, "llvm.loop.unroll_and_jam.")) {
313     LLVM_DEBUG(dbgs() << "  Disabled due to pragma.\n");
314     return LoopUnrollResult::Unmodified;
315   }
316 
317   if (!isSafeToUnrollAndJam(L, SE, DT, DI, *LI)) {
318     LLVM_DEBUG(dbgs() << "  Disabled due to not being safe.\n");
319     return LoopUnrollResult::Unmodified;
320   }
321 
322   // Approximate the loop size and collect useful info
323   SmallPtrSet<const Value *, 32> EphValues;
324   CodeMetrics::collectEphemeralValues(L, &AC, EphValues);
325   Loop *SubLoop = L->getSubLoops()[0];
326   UnrollCostEstimator InnerUCE(SubLoop, TTI, EphValues, UP.BEInsns);
327   UnrollCostEstimator OuterUCE(L, TTI, EphValues, UP.BEInsns);
328 
329   if (!InnerUCE.canUnroll() || !OuterUCE.canUnroll()) {
330     LLVM_DEBUG(dbgs() << "  Not unrolling loop which contains instructions"
331                       << " which cannot be duplicated or have invalid cost.\n");
332     return LoopUnrollResult::Unmodified;
333   }
334 
335   unsigned InnerLoopSize = InnerUCE.getRolledLoopSize();
336   LLVM_DEBUG(dbgs() << "  Outer Loop Size: " << OuterUCE.getRolledLoopSize()
337                     << "\n");
338   LLVM_DEBUG(dbgs() << "  Inner Loop Size: " << InnerLoopSize << "\n");
339 
340   if (InnerUCE.NumInlineCandidates != 0 || OuterUCE.NumInlineCandidates != 0) {
341     LLVM_DEBUG(dbgs() << "  Not unrolling loop with inlinable calls.\n");
342     return LoopUnrollResult::Unmodified;
343   }
344   if (InnerUCE.Convergent || OuterUCE.Convergent) {
345     LLVM_DEBUG(
346         dbgs() << "  Not unrolling loop with convergent instructions.\n");
347     return LoopUnrollResult::Unmodified;
348   }
349 
350   // Save original loop IDs for after the transformation.
351   MDNode *OrigOuterLoopID = L->getLoopID();
352   MDNode *OrigSubLoopID = SubLoop->getLoopID();
353 
354   // To assign the loop id of the epilogue, assign it before unrolling it so it
355   // is applied to every inner loop of the epilogue. We later apply the loop ID
356   // for the jammed inner loop.
357   std::optional<MDNode *> NewInnerEpilogueLoopID = makeFollowupLoopID(
358       OrigOuterLoopID, {LLVMLoopUnrollAndJamFollowupAll,
359                         LLVMLoopUnrollAndJamFollowupRemainderInner});
360   if (NewInnerEpilogueLoopID)
361     SubLoop->setLoopID(*NewInnerEpilogueLoopID);
362 
363   // Find trip count and trip multiple
364   BasicBlock *Latch = L->getLoopLatch();
365   BasicBlock *SubLoopLatch = SubLoop->getLoopLatch();
366   unsigned OuterTripCount = SE.getSmallConstantTripCount(L, Latch);
367   unsigned OuterTripMultiple = SE.getSmallConstantTripMultiple(L, Latch);
368   unsigned InnerTripCount = SE.getSmallConstantTripCount(SubLoop, SubLoopLatch);
369 
370   // Decide if, and by how much, to unroll
371   bool IsCountSetExplicitly = computeUnrollAndJamCount(
372     L, SubLoop, TTI, DT, LI, &AC, SE, EphValues, &ORE, OuterTripCount,
373       OuterTripMultiple, OuterUCE, InnerTripCount, InnerLoopSize, UP, PP);
374   if (UP.Count <= 1)
375     return LoopUnrollResult::Unmodified;
376   // Unroll factor (Count) must be less or equal to TripCount.
377   if (OuterTripCount && UP.Count > OuterTripCount)
378     UP.Count = OuterTripCount;
379 
380   Loop *EpilogueOuterLoop = nullptr;
381   LoopUnrollResult UnrollResult = UnrollAndJamLoop(
382       L, UP.Count, OuterTripCount, OuterTripMultiple, UP.UnrollRemainder, LI,
383       &SE, &DT, &AC, &TTI, &ORE, &EpilogueOuterLoop);
384 
385   // Assign new loop attributes.
386   if (EpilogueOuterLoop) {
387     std::optional<MDNode *> NewOuterEpilogueLoopID = makeFollowupLoopID(
388         OrigOuterLoopID, {LLVMLoopUnrollAndJamFollowupAll,
389                           LLVMLoopUnrollAndJamFollowupRemainderOuter});
390     if (NewOuterEpilogueLoopID)
391       EpilogueOuterLoop->setLoopID(*NewOuterEpilogueLoopID);
392   }
393 
394   std::optional<MDNode *> NewInnerLoopID =
395       makeFollowupLoopID(OrigOuterLoopID, {LLVMLoopUnrollAndJamFollowupAll,
396                                            LLVMLoopUnrollAndJamFollowupInner});
397   if (NewInnerLoopID)
398     SubLoop->setLoopID(*NewInnerLoopID);
399   else
400     SubLoop->setLoopID(OrigSubLoopID);
401 
402   if (UnrollResult == LoopUnrollResult::PartiallyUnrolled) {
403     std::optional<MDNode *> NewOuterLoopID = makeFollowupLoopID(
404         OrigOuterLoopID,
405         {LLVMLoopUnrollAndJamFollowupAll, LLVMLoopUnrollAndJamFollowupOuter});
406     if (NewOuterLoopID) {
407       L->setLoopID(*NewOuterLoopID);
408 
409       // Do not setLoopAlreadyUnrolled if a followup was given.
410       return UnrollResult;
411     }
412   }
413 
414   // If loop has an unroll count pragma or unrolled by explicitly set count
415   // mark loop as unrolled to prevent unrolling beyond that requested.
416   if (UnrollResult != LoopUnrollResult::FullyUnrolled && IsCountSetExplicitly)
417     L->setLoopAlreadyUnrolled();
418 
419   return UnrollResult;
420 }
421 
422 static bool tryToUnrollAndJamLoop(LoopNest &LN, DominatorTree &DT, LoopInfo &LI,
423                                   ScalarEvolution &SE,
424                                   const TargetTransformInfo &TTI,
425                                   AssumptionCache &AC, DependenceInfo &DI,
426                                   OptimizationRemarkEmitter &ORE, int OptLevel,
427                                   LPMUpdater &U) {
428   bool DidSomething = false;
429   ArrayRef<Loop *> Loops = LN.getLoops();
430   Loop *OutmostLoop = &LN.getOutermostLoop();
431 
432   // Add the loop nests in the reverse order of LN. See method
433   // declaration.
434   SmallPriorityWorklist<Loop *, 4> Worklist;
435   appendLoopsToWorklist(Loops, Worklist);
436   while (!Worklist.empty()) {
437     Loop *L = Worklist.pop_back_val();
438     std::string LoopName = std::string(L->getName());
439     LoopUnrollResult Result =
440         tryToUnrollAndJamLoop(L, DT, &LI, SE, TTI, AC, DI, ORE, OptLevel);
441     if (Result != LoopUnrollResult::Unmodified)
442       DidSomething = true;
443     if (L == OutmostLoop && Result == LoopUnrollResult::FullyUnrolled)
444       U.markLoopAsDeleted(*L, LoopName);
445   }
446 
447   return DidSomething;
448 }
449 
450 PreservedAnalyses LoopUnrollAndJamPass::run(LoopNest &LN,
451                                             LoopAnalysisManager &AM,
452                                             LoopStandardAnalysisResults &AR,
453                                             LPMUpdater &U) {
454   Function &F = *LN.getParent();
455 
456   DependenceInfo DI(&F, &AR.AA, &AR.SE, &AR.LI);
457   OptimizationRemarkEmitter ORE(&F);
458 
459   if (!tryToUnrollAndJamLoop(LN, AR.DT, AR.LI, AR.SE, AR.TTI, AR.AC, DI, ORE,
460                              OptLevel, U))
461     return PreservedAnalyses::all();
462 
463   auto PA = getLoopPassPreservedAnalyses();
464   PA.preserve<LoopNestAnalysis>();
465   return PA;
466 }
467