File: | llvm/lib/Transforms/Scalar/LoopFlatten.cpp |
Warning: | line 541, column 24 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- LoopFlatten.cpp - Loop flattening 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 flattens pairs nested loops into a single loop. | |||
10 | // | |||
11 | // The intention is to optimise loop nests like this, which together access an | |||
12 | // array linearly: | |||
13 | // for (int i = 0; i < N; ++i) | |||
14 | // for (int j = 0; j < M; ++j) | |||
15 | // f(A[i*M+j]); | |||
16 | // into one loop: | |||
17 | // for (int i = 0; i < (N*M); ++i) | |||
18 | // f(A[i]); | |||
19 | // | |||
20 | // It can also flatten loops where the induction variables are not used in the | |||
21 | // loop. This is only worth doing if the induction variables are only used in an | |||
22 | // expression like i*M+j. If they had any other uses, we would have to insert a | |||
23 | // div/mod to reconstruct the original values, so this wouldn't be profitable. | |||
24 | // | |||
25 | // We also need to prove that N*M will not overflow. | |||
26 | // | |||
27 | //===----------------------------------------------------------------------===// | |||
28 | ||||
29 | #include "llvm/Transforms/Scalar/LoopFlatten.h" | |||
30 | ||||
31 | #include "llvm/ADT/Statistic.h" | |||
32 | #include "llvm/Analysis/AssumptionCache.h" | |||
33 | #include "llvm/Analysis/LoopInfo.h" | |||
34 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | |||
35 | #include "llvm/Analysis/ScalarEvolution.h" | |||
36 | #include "llvm/Analysis/TargetTransformInfo.h" | |||
37 | #include "llvm/Analysis/ValueTracking.h" | |||
38 | #include "llvm/IR/Dominators.h" | |||
39 | #include "llvm/IR/Function.h" | |||
40 | #include "llvm/IR/IRBuilder.h" | |||
41 | #include "llvm/IR/Module.h" | |||
42 | #include "llvm/IR/PatternMatch.h" | |||
43 | #include "llvm/IR/Verifier.h" | |||
44 | #include "llvm/InitializePasses.h" | |||
45 | #include "llvm/Pass.h" | |||
46 | #include "llvm/Support/Debug.h" | |||
47 | #include "llvm/Support/raw_ostream.h" | |||
48 | #include "llvm/Transforms/Scalar.h" | |||
49 | #include "llvm/Transforms/Utils/Local.h" | |||
50 | #include "llvm/Transforms/Utils/LoopUtils.h" | |||
51 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" | |||
52 | #include "llvm/Transforms/Utils/SimplifyIndVar.h" | |||
53 | ||||
54 | using namespace llvm; | |||
55 | using namespace llvm::PatternMatch; | |||
56 | ||||
57 | #define DEBUG_TYPE"loop-flatten" "loop-flatten" | |||
58 | ||||
59 | STATISTIC(NumFlattened, "Number of loops flattened")static llvm::Statistic NumFlattened = {"loop-flatten", "NumFlattened" , "Number of loops flattened"}; | |||
60 | ||||
61 | static cl::opt<unsigned> RepeatedInstructionThreshold( | |||
62 | "loop-flatten-cost-threshold", cl::Hidden, cl::init(2), | |||
63 | cl::desc("Limit on the cost of instructions that can be repeated due to " | |||
64 | "loop flattening")); | |||
65 | ||||
66 | static cl::opt<bool> | |||
67 | AssumeNoOverflow("loop-flatten-assume-no-overflow", cl::Hidden, | |||
68 | cl::init(false), | |||
69 | cl::desc("Assume that the product of the two iteration " | |||
70 | "trip counts will never overflow")); | |||
71 | ||||
72 | static cl::opt<bool> | |||
73 | WidenIV("loop-flatten-widen-iv", cl::Hidden, | |||
74 | cl::init(true), | |||
75 | cl::desc("Widen the loop induction variables, if possible, so " | |||
76 | "overflow checks won't reject flattening")); | |||
77 | ||||
78 | struct FlattenInfo { | |||
79 | Loop *OuterLoop = nullptr; | |||
80 | Loop *InnerLoop = nullptr; | |||
81 | // These PHINodes correspond to loop induction variables, which are expected | |||
82 | // to start at zero and increment by one on each loop. | |||
83 | PHINode *InnerInductionPHI = nullptr; | |||
84 | PHINode *OuterInductionPHI = nullptr; | |||
85 | Value *InnerTripCount = nullptr; | |||
86 | Value *OuterTripCount = nullptr; | |||
87 | BinaryOperator *InnerIncrement = nullptr; | |||
88 | BinaryOperator *OuterIncrement = nullptr; | |||
89 | BranchInst *InnerBranch = nullptr; | |||
90 | BranchInst *OuterBranch = nullptr; | |||
91 | SmallPtrSet<Value *, 4> LinearIVUses; | |||
92 | SmallPtrSet<PHINode *, 4> InnerPHIsToTransform; | |||
93 | ||||
94 | // Whether this holds the flatten info before or after widening. | |||
95 | bool Widened = false; | |||
96 | ||||
97 | // Holds the old/narrow induction phis, i.e. the Phis before IV widening has | |||
98 | // been applied. This bookkeeping is used so we can skip some checks on these | |||
99 | // phi nodes. | |||
100 | PHINode *NarrowInnerInductionPHI = nullptr; | |||
101 | PHINode *NarrowOuterInductionPHI = nullptr; | |||
102 | ||||
103 | FlattenInfo(Loop *OL, Loop *IL) : OuterLoop(OL), InnerLoop(IL) {}; | |||
104 | ||||
105 | bool isNarrowInductionPhi(PHINode *Phi) { | |||
106 | // This can't be the narrow phi if we haven't widened the IV first. | |||
107 | if (!Widened) | |||
108 | return false; | |||
109 | return NarrowInnerInductionPHI == Phi || NarrowOuterInductionPHI == Phi; | |||
110 | } | |||
111 | }; | |||
112 | ||||
113 | static bool | |||
114 | setLoopComponents(Value *&TC, Value *&TripCount, BinaryOperator *&Increment, | |||
115 | SmallPtrSetImpl<Instruction *> &IterationInstructions) { | |||
116 | TripCount = TC; | |||
117 | IterationInstructions.insert(Increment); | |||
118 | LLVM_DEBUG(dbgs() << "Found Increment: "; Increment->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Found Increment: "; Increment ->dump(); } } while (false); | |||
119 | LLVM_DEBUG(dbgs() << "Found trip count: "; TripCount->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Found trip count: "; TripCount ->dump(); } } while (false); | |||
120 | LLVM_DEBUG(dbgs() << "Successfully found all loop components\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Successfully found all loop components\n" ; } } while (false); | |||
121 | return true; | |||
122 | } | |||
123 | ||||
124 | // Finds the induction variable, increment and trip count for a simple loop that | |||
125 | // we can flatten. | |||
126 | static bool findLoopComponents( | |||
127 | Loop *L, SmallPtrSetImpl<Instruction *> &IterationInstructions, | |||
128 | PHINode *&InductionPHI, Value *&TripCount, BinaryOperator *&Increment, | |||
129 | BranchInst *&BackBranch, ScalarEvolution *SE, bool IsWidened) { | |||
130 | LLVM_DEBUG(dbgs() << "Finding components of loop: " << L->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Finding components of loop: " << L->getName() << "\n"; } } while (false); | |||
131 | ||||
132 | if (!L->isLoopSimplifyForm()) { | |||
133 | LLVM_DEBUG(dbgs() << "Loop is not in normal form\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Loop is not in normal form\n" ; } } while (false); | |||
134 | return false; | |||
135 | } | |||
136 | ||||
137 | // Currently, to simplify the implementation, the Loop induction variable must | |||
138 | // start at zero and increment with a step size of one. | |||
139 | if (!L->isCanonical(*SE)) { | |||
140 | LLVM_DEBUG(dbgs() << "Loop is not canonical\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Loop is not canonical\n" ; } } while (false); | |||
141 | return false; | |||
142 | } | |||
143 | ||||
144 | // There must be exactly one exiting block, and it must be the same at the | |||
145 | // latch. | |||
146 | BasicBlock *Latch = L->getLoopLatch(); | |||
147 | if (L->getExitingBlock() != Latch) { | |||
148 | LLVM_DEBUG(dbgs() << "Exiting and latch block are different\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Exiting and latch block are different\n" ; } } while (false); | |||
149 | return false; | |||
150 | } | |||
151 | ||||
152 | // Find the induction PHI. If there is no induction PHI, we can't do the | |||
153 | // transformation. TODO: could other variables trigger this? Do we have to | |||
154 | // search for the best one? | |||
155 | InductionPHI = L->getInductionVariable(*SE); | |||
156 | if (!InductionPHI) { | |||
157 | LLVM_DEBUG(dbgs() << "Could not find induction PHI\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Could not find induction PHI\n" ; } } while (false); | |||
158 | return false; | |||
159 | } | |||
160 | LLVM_DEBUG(dbgs() << "Found induction PHI: "; InductionPHI->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Found induction PHI: "; InductionPHI ->dump(); } } while (false); | |||
161 | ||||
162 | bool ContinueOnTrue = L->contains(Latch->getTerminator()->getSuccessor(0)); | |||
163 | auto IsValidPredicate = [&](ICmpInst::Predicate Pred) { | |||
164 | if (ContinueOnTrue) | |||
165 | return Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT; | |||
166 | else | |||
167 | return Pred == CmpInst::ICMP_EQ; | |||
168 | }; | |||
169 | ||||
170 | // Find Compare and make sure it is valid. getLatchCmpInst checks that the | |||
171 | // back branch of the latch is conditional. | |||
172 | ICmpInst *Compare = L->getLatchCmpInst(); | |||
173 | if (!Compare || !IsValidPredicate(Compare->getUnsignedPredicate()) || | |||
174 | Compare->hasNUsesOrMore(2)) { | |||
175 | LLVM_DEBUG(dbgs() << "Could not find valid comparison\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Could not find valid comparison\n" ; } } while (false); | |||
176 | return false; | |||
177 | } | |||
178 | BackBranch = cast<BranchInst>(Latch->getTerminator()); | |||
179 | IterationInstructions.insert(BackBranch); | |||
180 | LLVM_DEBUG(dbgs() << "Found back branch: "; BackBranch->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Found back branch: "; BackBranch ->dump(); } } while (false); | |||
181 | IterationInstructions.insert(Compare); | |||
182 | LLVM_DEBUG(dbgs() << "Found comparison: "; Compare->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Found comparison: "; Compare ->dump(); } } while (false); | |||
183 | ||||
184 | // Find increment and trip count. | |||
185 | // There are exactly 2 incoming values to the induction phi; one from the | |||
186 | // pre-header and one from the latch. The incoming latch value is the | |||
187 | // increment variable. | |||
188 | Increment = | |||
189 | dyn_cast<BinaryOperator>(InductionPHI->getIncomingValueForBlock(Latch)); | |||
190 | if (Increment->hasNUsesOrMore(3)) { | |||
191 | LLVM_DEBUG(dbgs() << "Could not find valid increment\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Could not find valid increment\n" ; } } while (false); | |||
192 | return false; | |||
193 | } | |||
194 | // The trip count is the RHS of the compare. If this doesn't match the trip | |||
195 | // count computed by SCEV then this is because the trip count variable | |||
196 | // has been widened so the types don't match, or because it is a constant and | |||
197 | // another transformation has changed the compare (e.g. icmp ult %inc, | |||
198 | // tripcount -> icmp ult %j, tripcount-1), or both. | |||
199 | Value *RHS = Compare->getOperand(1); | |||
200 | const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); | |||
201 | if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) { | |||
202 | LLVM_DEBUG(dbgs() << "Backedge-taken count is not predictable\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Backedge-taken count is not predictable\n" ; } } while (false); | |||
203 | return false; | |||
204 | } | |||
205 | // The use of the Extend=false flag on getTripCountFromExitCount was added | |||
206 | // during a refactoring to preserve existing behavior. However, there's | |||
207 | // nothing obvious in the surrounding code when handles the overflow case. | |||
208 | // FIXME: audit code to establish whether there's a latent bug here. | |||
209 | const SCEV *SCEVTripCount = | |||
210 | SE->getTripCountFromExitCount(BackedgeTakenCount, false); | |||
211 | const SCEV *SCEVRHS = SE->getSCEV(RHS); | |||
212 | if (SCEVRHS == SCEVTripCount) | |||
213 | return setLoopComponents(RHS, TripCount, Increment, IterationInstructions); | |||
214 | ConstantInt *ConstantRHS = dyn_cast<ConstantInt>(RHS); | |||
215 | if (ConstantRHS) { | |||
216 | const SCEV *BackedgeTCExt = nullptr; | |||
217 | if (IsWidened) { | |||
218 | const SCEV *SCEVTripCountExt; | |||
219 | // Find the extended backedge taken count and extended trip count using | |||
220 | // SCEV. One of these should now match the RHS of the compare. | |||
221 | BackedgeTCExt = SE->getZeroExtendExpr(BackedgeTakenCount, RHS->getType()); | |||
222 | SCEVTripCountExt = SE->getTripCountFromExitCount(BackedgeTCExt, false); | |||
223 | if (SCEVRHS != BackedgeTCExt && SCEVRHS != SCEVTripCountExt) { | |||
224 | LLVM_DEBUG(dbgs() << "Could not find valid trip count\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Could not find valid trip count\n" ; } } while (false); | |||
225 | return false; | |||
226 | } | |||
227 | } | |||
228 | // If the RHS of the compare is equal to the backedge taken count we need | |||
229 | // to add one to get the trip count. | |||
230 | if (SCEVRHS == BackedgeTCExt || SCEVRHS == BackedgeTakenCount) { | |||
231 | ConstantInt *One = ConstantInt::get(ConstantRHS->getType(), 1); | |||
232 | Value *NewRHS = ConstantInt::get( | |||
233 | ConstantRHS->getContext(), ConstantRHS->getValue() + One->getValue()); | |||
234 | return setLoopComponents(NewRHS, TripCount, Increment, | |||
235 | IterationInstructions); | |||
236 | } | |||
237 | return setLoopComponents(RHS, TripCount, Increment, IterationInstructions); | |||
238 | } | |||
239 | // If the RHS isn't a constant then check that the reason it doesn't match | |||
240 | // the SCEV trip count is because the RHS is a ZExt or SExt instruction | |||
241 | // (and take the trip count to be the RHS). | |||
242 | if (!IsWidened) { | |||
243 | LLVM_DEBUG(dbgs() << "Could not find valid trip count\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Could not find valid trip count\n" ; } } while (false); | |||
244 | return false; | |||
245 | } | |||
246 | auto *TripCountInst = dyn_cast<Instruction>(RHS); | |||
247 | if (!TripCountInst) { | |||
248 | LLVM_DEBUG(dbgs() << "Could not find valid trip count\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Could not find valid trip count\n" ; } } while (false); | |||
249 | return false; | |||
250 | } | |||
251 | if ((!isa<ZExtInst>(TripCountInst) && !isa<SExtInst>(TripCountInst)) || | |||
252 | SE->getSCEV(TripCountInst->getOperand(0)) != SCEVTripCount) { | |||
253 | LLVM_DEBUG(dbgs() << "Could not find valid extended trip count\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Could not find valid extended trip count\n" ; } } while (false); | |||
254 | return false; | |||
255 | } | |||
256 | return setLoopComponents(RHS, TripCount, Increment, IterationInstructions); | |||
257 | } | |||
258 | ||||
259 | static bool checkPHIs(FlattenInfo &FI, const TargetTransformInfo *TTI) { | |||
260 | // All PHIs in the inner and outer headers must either be: | |||
261 | // - The induction PHI, which we are going to rewrite as one induction in | |||
262 | // the new loop. This is already checked by findLoopComponents. | |||
263 | // - An outer header PHI with all incoming values from outside the loop. | |||
264 | // LoopSimplify guarantees we have a pre-header, so we don't need to | |||
265 | // worry about that here. | |||
266 | // - Pairs of PHIs in the inner and outer headers, which implement a | |||
267 | // loop-carried dependency that will still be valid in the new loop. To | |||
268 | // be valid, this variable must be modified only in the inner loop. | |||
269 | ||||
270 | // The set of PHI nodes in the outer loop header that we know will still be | |||
271 | // valid after the transformation. These will not need to be modified (with | |||
272 | // the exception of the induction variable), but we do need to check that | |||
273 | // there are no unsafe PHI nodes. | |||
274 | SmallPtrSet<PHINode *, 4> SafeOuterPHIs; | |||
275 | SafeOuterPHIs.insert(FI.OuterInductionPHI); | |||
276 | ||||
277 | // Check that all PHI nodes in the inner loop header match one of the valid | |||
278 | // patterns. | |||
279 | for (PHINode &InnerPHI : FI.InnerLoop->getHeader()->phis()) { | |||
280 | // The induction PHIs break these rules, and that's OK because we treat | |||
281 | // them specially when doing the transformation. | |||
282 | if (&InnerPHI == FI.InnerInductionPHI) | |||
283 | continue; | |||
284 | if (FI.isNarrowInductionPhi(&InnerPHI)) | |||
285 | continue; | |||
286 | ||||
287 | // Each inner loop PHI node must have two incoming values/blocks - one | |||
288 | // from the pre-header, and one from the latch. | |||
289 | assert(InnerPHI.getNumIncomingValues() == 2)(static_cast <bool> (InnerPHI.getNumIncomingValues() == 2) ? void (0) : __assert_fail ("InnerPHI.getNumIncomingValues() == 2" , "llvm/lib/Transforms/Scalar/LoopFlatten.cpp", 289, __extension__ __PRETTY_FUNCTION__)); | |||
290 | Value *PreHeaderValue = | |||
291 | InnerPHI.getIncomingValueForBlock(FI.InnerLoop->getLoopPreheader()); | |||
292 | Value *LatchValue = | |||
293 | InnerPHI.getIncomingValueForBlock(FI.InnerLoop->getLoopLatch()); | |||
294 | ||||
295 | // The incoming value from the outer loop must be the PHI node in the | |||
296 | // outer loop header, with no modifications made in the top of the outer | |||
297 | // loop. | |||
298 | PHINode *OuterPHI = dyn_cast<PHINode>(PreHeaderValue); | |||
299 | if (!OuterPHI || OuterPHI->getParent() != FI.OuterLoop->getHeader()) { | |||
300 | LLVM_DEBUG(dbgs() << "value modified in top of outer loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "value modified in top of outer loop\n" ; } } while (false); | |||
301 | return false; | |||
302 | } | |||
303 | ||||
304 | // The other incoming value must come from the inner loop, without any | |||
305 | // modifications in the tail end of the outer loop. We are in LCSSA form, | |||
306 | // so this will actually be a PHI in the inner loop's exit block, which | |||
307 | // only uses values from inside the inner loop. | |||
308 | PHINode *LCSSAPHI = dyn_cast<PHINode>( | |||
309 | OuterPHI->getIncomingValueForBlock(FI.OuterLoop->getLoopLatch())); | |||
310 | if (!LCSSAPHI) { | |||
311 | LLVM_DEBUG(dbgs() << "could not find LCSSA PHI\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "could not find LCSSA PHI\n" ; } } while (false); | |||
312 | return false; | |||
313 | } | |||
314 | ||||
315 | // The value used by the LCSSA PHI must be the same one that the inner | |||
316 | // loop's PHI uses. | |||
317 | if (LCSSAPHI->hasConstantValue() != LatchValue) { | |||
318 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "LCSSA PHI incoming value does not match latch value\n" ; } } while (false) | |||
319 | dbgs() << "LCSSA PHI incoming value does not match latch value\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "LCSSA PHI incoming value does not match latch value\n" ; } } while (false); | |||
320 | return false; | |||
321 | } | |||
322 | ||||
323 | LLVM_DEBUG(dbgs() << "PHI pair is safe:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "PHI pair is safe:\n"; } } while (false); | |||
324 | LLVM_DEBUG(dbgs() << " Inner: "; InnerPHI.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << " Inner: "; InnerPHI.dump (); } } while (false); | |||
325 | LLVM_DEBUG(dbgs() << " Outer: "; OuterPHI->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << " Outer: "; OuterPHI-> dump(); } } while (false); | |||
326 | SafeOuterPHIs.insert(OuterPHI); | |||
327 | FI.InnerPHIsToTransform.insert(&InnerPHI); | |||
328 | } | |||
329 | ||||
330 | for (PHINode &OuterPHI : FI.OuterLoop->getHeader()->phis()) { | |||
331 | if (FI.isNarrowInductionPhi(&OuterPHI)) | |||
332 | continue; | |||
333 | if (!SafeOuterPHIs.count(&OuterPHI)) { | |||
334 | LLVM_DEBUG(dbgs() << "found unsafe PHI in outer loop: "; OuterPHI.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "found unsafe PHI in outer loop: " ; OuterPHI.dump(); } } while (false); | |||
335 | return false; | |||
336 | } | |||
337 | } | |||
338 | ||||
339 | LLVM_DEBUG(dbgs() << "checkPHIs: OK\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "checkPHIs: OK\n"; } } while (false); | |||
340 | return true; | |||
341 | } | |||
342 | ||||
343 | static bool | |||
344 | checkOuterLoopInsts(FlattenInfo &FI, | |||
345 | SmallPtrSetImpl<Instruction *> &IterationInstructions, | |||
346 | const TargetTransformInfo *TTI) { | |||
347 | // Check for instructions in the outer but not inner loop. If any of these | |||
348 | // have side-effects then this transformation is not legal, and if there is | |||
349 | // a significant amount of code here which can't be optimised out that it's | |||
350 | // not profitable (as these instructions would get executed for each | |||
351 | // iteration of the inner loop). | |||
352 | InstructionCost RepeatedInstrCost = 0; | |||
353 | for (auto *B : FI.OuterLoop->getBlocks()) { | |||
354 | if (FI.InnerLoop->contains(B)) | |||
355 | continue; | |||
356 | ||||
357 | for (auto &I : *B) { | |||
358 | if (!isa<PHINode>(&I) && !I.isTerminator() && | |||
359 | !isSafeToSpeculativelyExecute(&I)) { | |||
360 | LLVM_DEBUG(dbgs() << "Cannot flatten because instruction may have "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Cannot flatten because instruction may have " "side effects: "; I.dump(); } } while (false) | |||
361 | "side effects: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Cannot flatten because instruction may have " "side effects: "; I.dump(); } } while (false) | |||
362 | I.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Cannot flatten because instruction may have " "side effects: "; I.dump(); } } while (false); | |||
363 | return false; | |||
364 | } | |||
365 | // The execution count of the outer loop's iteration instructions | |||
366 | // (increment, compare and branch) will be increased, but the | |||
367 | // equivalent instructions will be removed from the inner loop, so | |||
368 | // they make a net difference of zero. | |||
369 | if (IterationInstructions.count(&I)) | |||
370 | continue; | |||
371 | // The uncoditional branch to the inner loop's header will turn into | |||
372 | // a fall-through, so adds no cost. | |||
373 | BranchInst *Br = dyn_cast<BranchInst>(&I); | |||
374 | if (Br && Br->isUnconditional() && | |||
375 | Br->getSuccessor(0) == FI.InnerLoop->getHeader()) | |||
376 | continue; | |||
377 | // Multiplies of the outer iteration variable and inner iteration | |||
378 | // count will be optimised out. | |||
379 | if (match(&I, m_c_Mul(m_Specific(FI.OuterInductionPHI), | |||
380 | m_Specific(FI.InnerTripCount)))) | |||
381 | continue; | |||
382 | InstructionCost Cost = | |||
383 | TTI->getUserCost(&I, TargetTransformInfo::TCK_SizeAndLatency); | |||
384 | LLVM_DEBUG(dbgs() << "Cost " << Cost << ": "; I.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Cost " << Cost << ": "; I.dump(); } } while (false); | |||
385 | RepeatedInstrCost += Cost; | |||
386 | } | |||
387 | } | |||
388 | ||||
389 | LLVM_DEBUG(dbgs() << "Cost of instructions that will be repeated: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Cost of instructions that will be repeated: " << RepeatedInstrCost << "\n"; } } while (false) | |||
390 | << RepeatedInstrCost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Cost of instructions that will be repeated: " << RepeatedInstrCost << "\n"; } } while (false); | |||
391 | // Bail out if flattening the loops would cause instructions in the outer | |||
392 | // loop but not in the inner loop to be executed extra times. | |||
393 | if (RepeatedInstrCost > RepeatedInstructionThreshold) { | |||
394 | LLVM_DEBUG(dbgs() << "checkOuterLoopInsts: not profitable, bailing.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "checkOuterLoopInsts: not profitable, bailing.\n" ; } } while (false); | |||
395 | return false; | |||
396 | } | |||
397 | ||||
398 | LLVM_DEBUG(dbgs() << "checkOuterLoopInsts: OK\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "checkOuterLoopInsts: OK\n" ; } } while (false); | |||
399 | return true; | |||
400 | } | |||
401 | ||||
402 | static bool checkIVUsers(FlattenInfo &FI) { | |||
403 | // We require all uses of both induction variables to match this pattern: | |||
404 | // | |||
405 | // (OuterPHI * InnerTripCount) + InnerPHI | |||
406 | // | |||
407 | // Any uses of the induction variables not matching that pattern would | |||
408 | // require a div/mod to reconstruct in the flattened loop, so the | |||
409 | // transformation wouldn't be profitable. | |||
410 | ||||
411 | Value *InnerTripCount = FI.InnerTripCount; | |||
412 | if (FI.Widened && | |||
413 | (isa<SExtInst>(InnerTripCount) || isa<ZExtInst>(InnerTripCount))) | |||
414 | InnerTripCount = cast<Instruction>(InnerTripCount)->getOperand(0); | |||
415 | ||||
416 | // Check that all uses of the inner loop's induction variable match the | |||
417 | // expected pattern, recording the uses of the outer IV. | |||
418 | SmallPtrSet<Value *, 4> ValidOuterPHIUses; | |||
419 | for (User *U : FI.InnerInductionPHI->users()) { | |||
420 | if (U == FI.InnerIncrement) | |||
421 | continue; | |||
422 | ||||
423 | // After widening the IVs, a trunc instruction might have been introduced, | |||
424 | // so look through truncs. | |||
425 | if (isa<TruncInst>(U)) { | |||
426 | if (!U->hasOneUse()) | |||
427 | return false; | |||
428 | U = *U->user_begin(); | |||
429 | } | |||
430 | ||||
431 | // If the use is in the compare (which is also the condition of the inner | |||
432 | // branch) then the compare has been altered by another transformation e.g | |||
433 | // icmp ult %inc, tripcount -> icmp ult %j, tripcount-1, where tripcount is | |||
434 | // a constant. Ignore this use as the compare gets removed later anyway. | |||
435 | if (U == FI.InnerBranch->getCondition()) | |||
436 | continue; | |||
437 | ||||
438 | LLVM_DEBUG(dbgs() << "Found use of inner induction variable: "; U->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Found use of inner induction variable: " ; U->dump(); } } while (false); | |||
439 | ||||
440 | Value *MatchedMul = nullptr; | |||
441 | Value *MatchedItCount = nullptr; | |||
442 | bool IsAdd = match(U, m_c_Add(m_Specific(FI.InnerInductionPHI), | |||
443 | m_Value(MatchedMul))) && | |||
444 | match(MatchedMul, m_c_Mul(m_Specific(FI.OuterInductionPHI), | |||
445 | m_Value(MatchedItCount))); | |||
446 | ||||
447 | // Matches the same pattern as above, except it also looks for truncs | |||
448 | // on the phi, which can be the result of widening the induction variables. | |||
449 | bool IsAddTrunc = | |||
450 | match(U, m_c_Add(m_Trunc(m_Specific(FI.InnerInductionPHI)), | |||
451 | m_Value(MatchedMul))) && | |||
452 | match(MatchedMul, m_c_Mul(m_Trunc(m_Specific(FI.OuterInductionPHI)), | |||
453 | m_Value(MatchedItCount))); | |||
454 | ||||
455 | if (!MatchedItCount) | |||
456 | return false; | |||
457 | // Look through extends if the IV has been widened. | |||
458 | if (FI.Widened && | |||
459 | (isa<SExtInst>(MatchedItCount) || isa<ZExtInst>(MatchedItCount))) { | |||
460 | assert(MatchedItCount->getType() == FI.InnerInductionPHI->getType() &&(static_cast <bool> (MatchedItCount->getType() == FI .InnerInductionPHI->getType() && "Unexpected type mismatch in types after widening" ) ? void (0) : __assert_fail ("MatchedItCount->getType() == FI.InnerInductionPHI->getType() && \"Unexpected type mismatch in types after widening\"" , "llvm/lib/Transforms/Scalar/LoopFlatten.cpp", 461, __extension__ __PRETTY_FUNCTION__)) | |||
461 | "Unexpected type mismatch in types after widening")(static_cast <bool> (MatchedItCount->getType() == FI .InnerInductionPHI->getType() && "Unexpected type mismatch in types after widening" ) ? void (0) : __assert_fail ("MatchedItCount->getType() == FI.InnerInductionPHI->getType() && \"Unexpected type mismatch in types after widening\"" , "llvm/lib/Transforms/Scalar/LoopFlatten.cpp", 461, __extension__ __PRETTY_FUNCTION__)); | |||
462 | MatchedItCount = isa<SExtInst>(MatchedItCount) | |||
463 | ? dyn_cast<SExtInst>(MatchedItCount)->getOperand(0) | |||
464 | : dyn_cast<ZExtInst>(MatchedItCount)->getOperand(0); | |||
465 | } | |||
466 | ||||
467 | if ((IsAdd || IsAddTrunc) && MatchedItCount == InnerTripCount) { | |||
468 | LLVM_DEBUG(dbgs() << "Use is optimisable\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Use is optimisable\n"; } } while (false); | |||
469 | ValidOuterPHIUses.insert(MatchedMul); | |||
470 | FI.LinearIVUses.insert(U); | |||
471 | } else { | |||
472 | LLVM_DEBUG(dbgs() << "Did not match expected pattern, bailing\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Did not match expected pattern, bailing\n" ; } } while (false); | |||
473 | return false; | |||
474 | } | |||
475 | } | |||
476 | ||||
477 | // Check that there are no uses of the outer IV other than the ones found | |||
478 | // as part of the pattern above. | |||
479 | for (User *U : FI.OuterInductionPHI->users()) { | |||
480 | if (U == FI.OuterIncrement) | |||
481 | continue; | |||
482 | ||||
483 | auto IsValidOuterPHIUses = [&] (User *U) -> bool { | |||
484 | LLVM_DEBUG(dbgs() << "Found use of outer induction variable: "; U->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Found use of outer induction variable: " ; U->dump(); } } while (false); | |||
485 | if (!ValidOuterPHIUses.count(U)) { | |||
486 | LLVM_DEBUG(dbgs() << "Did not match expected pattern, bailing\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Did not match expected pattern, bailing\n" ; } } while (false); | |||
487 | return false; | |||
488 | } | |||
489 | LLVM_DEBUG(dbgs() << "Use is optimisable\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Use is optimisable\n"; } } while (false); | |||
490 | return true; | |||
491 | }; | |||
492 | ||||
493 | if (auto *V = dyn_cast<TruncInst>(U)) { | |||
494 | for (auto *K : V->users()) { | |||
495 | if (!IsValidOuterPHIUses(K)) | |||
496 | return false; | |||
497 | } | |||
498 | continue; | |||
499 | } | |||
500 | ||||
501 | if (!IsValidOuterPHIUses(U)) | |||
502 | return false; | |||
503 | } | |||
504 | ||||
505 | LLVM_DEBUG(dbgs() << "checkIVUsers: OK\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "checkIVUsers: OK\n"; dbgs () << "Found " << FI.LinearIVUses.size() << " value(s) that can be replaced:\n"; for (Value *V : FI.LinearIVUses ) { dbgs() << " "; V->dump(); }; } } while (false) | |||
506 | dbgs() << "Found " << FI.LinearIVUses.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "checkIVUsers: OK\n"; dbgs () << "Found " << FI.LinearIVUses.size() << " value(s) that can be replaced:\n"; for (Value *V : FI.LinearIVUses ) { dbgs() << " "; V->dump(); }; } } while (false) | |||
507 | << " value(s) that can be replaced:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "checkIVUsers: OK\n"; dbgs () << "Found " << FI.LinearIVUses.size() << " value(s) that can be replaced:\n"; for (Value *V : FI.LinearIVUses ) { dbgs() << " "; V->dump(); }; } } while (false) | |||
508 | for (Value *V : FI.LinearIVUses) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "checkIVUsers: OK\n"; dbgs () << "Found " << FI.LinearIVUses.size() << " value(s) that can be replaced:\n"; for (Value *V : FI.LinearIVUses ) { dbgs() << " "; V->dump(); }; } } while (false) | |||
509 | dbgs() << " ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "checkIVUsers: OK\n"; dbgs () << "Found " << FI.LinearIVUses.size() << " value(s) that can be replaced:\n"; for (Value *V : FI.LinearIVUses ) { dbgs() << " "; V->dump(); }; } } while (false) | |||
510 | V->dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "checkIVUsers: OK\n"; dbgs () << "Found " << FI.LinearIVUses.size() << " value(s) that can be replaced:\n"; for (Value *V : FI.LinearIVUses ) { dbgs() << " "; V->dump(); }; } } while (false) | |||
511 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "checkIVUsers: OK\n"; dbgs () << "Found " << FI.LinearIVUses.size() << " value(s) that can be replaced:\n"; for (Value *V : FI.LinearIVUses ) { dbgs() << " "; V->dump(); }; } } while (false); | |||
512 | return true; | |||
513 | } | |||
514 | ||||
515 | // Return an OverflowResult dependant on if overflow of the multiplication of | |||
516 | // InnerTripCount and OuterTripCount can be assumed not to happen. | |||
517 | static OverflowResult checkOverflow(FlattenInfo &FI, DominatorTree *DT, | |||
518 | AssumptionCache *AC) { | |||
519 | Function *F = FI.OuterLoop->getHeader()->getParent(); | |||
520 | const DataLayout &DL = F->getParent()->getDataLayout(); | |||
521 | ||||
522 | // For debugging/testing. | |||
523 | if (AssumeNoOverflow) | |||
524 | return OverflowResult::NeverOverflows; | |||
525 | ||||
526 | // Check if the multiply could not overflow due to known ranges of the | |||
527 | // input values. | |||
528 | OverflowResult OR = computeOverflowForUnsignedMul( | |||
529 | FI.InnerTripCount, FI.OuterTripCount, DL, AC, | |||
530 | FI.OuterLoop->getLoopPreheader()->getTerminator(), DT); | |||
531 | if (OR != OverflowResult::MayOverflow) | |||
532 | return OR; | |||
533 | ||||
534 | for (Value *V : FI.LinearIVUses) { | |||
535 | for (Value *U : V->users()) { | |||
536 | if (auto *GEP
| |||
537 | for (Value *GEPUser : U->users()) { | |||
538 | Instruction *GEPUserInst = dyn_cast<Instruction>(GEPUser); | |||
539 | if (!isa<LoadInst>(GEPUserInst) && | |||
540 | !(isa<StoreInst>(GEPUserInst) && | |||
541 | GEP == GEPUserInst->getOperand(1))) | |||
| ||||
542 | continue; | |||
543 | if (!isGuaranteedToExecuteForEveryIteration(GEPUserInst, | |||
544 | FI.InnerLoop)) | |||
545 | continue; | |||
546 | // The IV is used as the operand of a GEP which dominates the loop | |||
547 | // latch, and the IV is at least as wide as the address space of the | |||
548 | // GEP. In this case, the GEP would wrap around the address space | |||
549 | // before the IV increment wraps, which would be UB. | |||
550 | if (GEP->isInBounds() && | |||
551 | V->getType()->getIntegerBitWidth() >= | |||
552 | DL.getPointerTypeSizeInBits(GEP->getType())) { | |||
553 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "use of linear IV would be UB if overflow occurred: " ; GEP->dump(); } } while (false) | |||
554 | dbgs() << "use of linear IV would be UB if overflow occurred: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "use of linear IV would be UB if overflow occurred: " ; GEP->dump(); } } while (false) | |||
555 | GEP->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "use of linear IV would be UB if overflow occurred: " ; GEP->dump(); } } while (false); | |||
556 | return OverflowResult::NeverOverflows; | |||
557 | } | |||
558 | } | |||
559 | } | |||
560 | } | |||
561 | } | |||
562 | ||||
563 | return OverflowResult::MayOverflow; | |||
564 | } | |||
565 | ||||
566 | static bool CanFlattenLoopPair(FlattenInfo &FI, DominatorTree *DT, LoopInfo *LI, | |||
567 | ScalarEvolution *SE, AssumptionCache *AC, | |||
568 | const TargetTransformInfo *TTI) { | |||
569 | SmallPtrSet<Instruction *, 8> IterationInstructions; | |||
570 | if (!findLoopComponents(FI.InnerLoop, IterationInstructions, | |||
571 | FI.InnerInductionPHI, FI.InnerTripCount, | |||
572 | FI.InnerIncrement, FI.InnerBranch, SE, FI.Widened)) | |||
573 | return false; | |||
574 | if (!findLoopComponents(FI.OuterLoop, IterationInstructions, | |||
575 | FI.OuterInductionPHI, FI.OuterTripCount, | |||
576 | FI.OuterIncrement, FI.OuterBranch, SE, FI.Widened)) | |||
577 | return false; | |||
578 | ||||
579 | // Both of the loop trip count values must be invariant in the outer loop | |||
580 | // (non-instructions are all inherently invariant). | |||
581 | if (!FI.OuterLoop->isLoopInvariant(FI.InnerTripCount)) { | |||
582 | LLVM_DEBUG(dbgs() << "inner loop trip count not invariant\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "inner loop trip count not invariant\n" ; } } while (false); | |||
583 | return false; | |||
584 | } | |||
585 | if (!FI.OuterLoop->isLoopInvariant(FI.OuterTripCount)) { | |||
586 | LLVM_DEBUG(dbgs() << "outer loop trip count not invariant\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "outer loop trip count not invariant\n" ; } } while (false); | |||
587 | return false; | |||
588 | } | |||
589 | ||||
590 | if (!checkPHIs(FI, TTI)) | |||
591 | return false; | |||
592 | ||||
593 | // FIXME: it should be possible to handle different types correctly. | |||
594 | if (FI.InnerInductionPHI->getType() != FI.OuterInductionPHI->getType()) | |||
595 | return false; | |||
596 | ||||
597 | if (!checkOuterLoopInsts(FI, IterationInstructions, TTI)) | |||
598 | return false; | |||
599 | ||||
600 | // Find the values in the loop that can be replaced with the linearized | |||
601 | // induction variable, and check that there are no other uses of the inner | |||
602 | // or outer induction variable. If there were, we could still do this | |||
603 | // transformation, but we'd have to insert a div/mod to calculate the | |||
604 | // original IVs, so it wouldn't be profitable. | |||
605 | if (!checkIVUsers(FI)) | |||
606 | return false; | |||
607 | ||||
608 | LLVM_DEBUG(dbgs() << "CanFlattenLoopPair: OK\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "CanFlattenLoopPair: OK\n" ; } } while (false); | |||
609 | return true; | |||
610 | } | |||
611 | ||||
612 | static bool DoFlattenLoopPair(FlattenInfo &FI, DominatorTree *DT, LoopInfo *LI, | |||
613 | ScalarEvolution *SE, AssumptionCache *AC, | |||
614 | const TargetTransformInfo *TTI, LPMUpdater *U) { | |||
615 | Function *F = FI.OuterLoop->getHeader()->getParent(); | |||
616 | LLVM_DEBUG(dbgs() << "Checks all passed, doing the transformation\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Checks all passed, doing the transformation\n" ; } } while (false); | |||
617 | { | |||
618 | using namespace ore; | |||
619 | OptimizationRemark Remark(DEBUG_TYPE"loop-flatten", "Flattened", FI.InnerLoop->getStartLoc(), | |||
620 | FI.InnerLoop->getHeader()); | |||
621 | OptimizationRemarkEmitter ORE(F); | |||
622 | Remark << "Flattened into outer loop"; | |||
623 | ORE.emit(Remark); | |||
624 | } | |||
625 | ||||
626 | Value *NewTripCount = BinaryOperator::CreateMul( | |||
627 | FI.InnerTripCount, FI.OuterTripCount, "flatten.tripcount", | |||
628 | FI.OuterLoop->getLoopPreheader()->getTerminator()); | |||
629 | LLVM_DEBUG(dbgs() << "Created new trip count in preheader: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Created new trip count in preheader: " ; NewTripCount->dump(); } } while (false) | |||
630 | NewTripCount->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Created new trip count in preheader: " ; NewTripCount->dump(); } } while (false); | |||
631 | ||||
632 | // Fix up PHI nodes that take values from the inner loop back-edge, which | |||
633 | // we are about to remove. | |||
634 | FI.InnerInductionPHI->removeIncomingValue(FI.InnerLoop->getLoopLatch()); | |||
635 | ||||
636 | // The old Phi will be optimised away later, but for now we can't leave | |||
637 | // leave it in an invalid state, so are updating them too. | |||
638 | for (PHINode *PHI : FI.InnerPHIsToTransform) | |||
639 | PHI->removeIncomingValue(FI.InnerLoop->getLoopLatch()); | |||
640 | ||||
641 | // Modify the trip count of the outer loop to be the product of the two | |||
642 | // trip counts. | |||
643 | cast<User>(FI.OuterBranch->getCondition())->setOperand(1, NewTripCount); | |||
644 | ||||
645 | // Replace the inner loop backedge with an unconditional branch to the exit. | |||
646 | BasicBlock *InnerExitBlock = FI.InnerLoop->getExitBlock(); | |||
647 | BasicBlock *InnerExitingBlock = FI.InnerLoop->getExitingBlock(); | |||
648 | InnerExitingBlock->getTerminator()->eraseFromParent(); | |||
649 | BranchInst::Create(InnerExitBlock, InnerExitingBlock); | |||
650 | DT->deleteEdge(InnerExitingBlock, FI.InnerLoop->getHeader()); | |||
651 | ||||
652 | // Replace all uses of the polynomial calculated from the two induction | |||
653 | // variables with the one new one. | |||
654 | IRBuilder<> Builder(FI.OuterInductionPHI->getParent()->getTerminator()); | |||
655 | for (Value *V : FI.LinearIVUses) { | |||
656 | Value *OuterValue = FI.OuterInductionPHI; | |||
657 | if (FI.Widened) | |||
658 | OuterValue = Builder.CreateTrunc(FI.OuterInductionPHI, V->getType(), | |||
659 | "flatten.trunciv"); | |||
660 | ||||
661 | LLVM_DEBUG(dbgs() << "Replacing: "; V->dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Replacing: "; V->dump (); dbgs() << "with: "; OuterValue->dump(); } } while (false) | |||
662 | dbgs() << "with: "; OuterValue->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Replacing: "; V->dump (); dbgs() << "with: "; OuterValue->dump(); } } while (false); | |||
663 | V->replaceAllUsesWith(OuterValue); | |||
664 | } | |||
665 | ||||
666 | // Tell LoopInfo, SCEV and the pass manager that the inner loop has been | |||
667 | // deleted, and any information that have about the outer loop invalidated. | |||
668 | SE->forgetLoop(FI.OuterLoop); | |||
669 | SE->forgetLoop(FI.InnerLoop); | |||
670 | if (U) | |||
671 | U->markLoopAsDeleted(*FI.InnerLoop, FI.InnerLoop->getName()); | |||
672 | LI->erase(FI.InnerLoop); | |||
673 | ||||
674 | // Increment statistic value. | |||
675 | NumFlattened++; | |||
676 | ||||
677 | return true; | |||
678 | } | |||
679 | ||||
680 | static bool CanWidenIV(FlattenInfo &FI, DominatorTree *DT, LoopInfo *LI, | |||
681 | ScalarEvolution *SE, AssumptionCache *AC, | |||
682 | const TargetTransformInfo *TTI) { | |||
683 | if (!WidenIV) { | |||
684 | LLVM_DEBUG(dbgs() << "Widening the IVs is disabled\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Widening the IVs is disabled\n" ; } } while (false); | |||
685 | return false; | |||
686 | } | |||
687 | ||||
688 | LLVM_DEBUG(dbgs() << "Try widening the IVs\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Try widening the IVs\n"; } } while (false); | |||
689 | Module *M = FI.InnerLoop->getHeader()->getParent()->getParent(); | |||
690 | auto &DL = M->getDataLayout(); | |||
691 | auto *InnerType = FI.InnerInductionPHI->getType(); | |||
692 | auto *OuterType = FI.OuterInductionPHI->getType(); | |||
693 | unsigned MaxLegalSize = DL.getLargestLegalIntTypeSizeInBits(); | |||
694 | auto *MaxLegalType = DL.getLargestLegalIntType(M->getContext()); | |||
695 | ||||
696 | // If both induction types are less than the maximum legal integer width, | |||
697 | // promote both to the widest type available so we know calculating | |||
698 | // (OuterTripCount * InnerTripCount) as the new trip count is safe. | |||
699 | if (InnerType != OuterType || | |||
700 | InnerType->getScalarSizeInBits() >= MaxLegalSize || | |||
701 | MaxLegalType->getScalarSizeInBits() < InnerType->getScalarSizeInBits() * 2) { | |||
702 | LLVM_DEBUG(dbgs() << "Can't widen the IV\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Can't widen the IV\n"; } } while (false); | |||
703 | return false; | |||
704 | } | |||
705 | ||||
706 | SCEVExpander Rewriter(*SE, DL, "loopflatten"); | |||
707 | SmallVector<WeakTrackingVH, 4> DeadInsts; | |||
708 | unsigned ElimExt = 0; | |||
709 | unsigned Widened = 0; | |||
710 | ||||
711 | auto CreateWideIV = [&] (WideIVInfo WideIV, bool &Deleted) -> bool { | |||
712 | PHINode *WidePhi = createWideIV(WideIV, LI, SE, Rewriter, DT, DeadInsts, | |||
713 | ElimExt, Widened, true /* HasGuards */, | |||
714 | true /* UsePostIncrementRanges */); | |||
715 | if (!WidePhi) | |||
716 | return false; | |||
717 | LLVM_DEBUG(dbgs() << "Created wide phi: "; WidePhi->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Created wide phi: "; WidePhi ->dump(); } } while (false); | |||
718 | LLVM_DEBUG(dbgs() << "Deleting old phi: "; WideIV.NarrowIV->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Deleting old phi: "; WideIV .NarrowIV->dump(); } } while (false); | |||
719 | Deleted = RecursivelyDeleteDeadPHINode(WideIV.NarrowIV); | |||
720 | return true; | |||
721 | }; | |||
722 | ||||
723 | bool Deleted; | |||
724 | if (!CreateWideIV({FI.InnerInductionPHI, MaxLegalType, false }, Deleted)) | |||
725 | return false; | |||
726 | // Add the narrow phi to list, so that it will be adjusted later when the | |||
727 | // the transformation is performed. | |||
728 | if (!Deleted) | |||
729 | FI.InnerPHIsToTransform.insert(FI.InnerInductionPHI); | |||
730 | ||||
731 | if (!CreateWideIV({FI.OuterInductionPHI, MaxLegalType, false }, Deleted)) | |||
732 | return false; | |||
733 | ||||
734 | assert(Widened && "Widened IV expected")(static_cast <bool> (Widened && "Widened IV expected" ) ? void (0) : __assert_fail ("Widened && \"Widened IV expected\"" , "llvm/lib/Transforms/Scalar/LoopFlatten.cpp", 734, __extension__ __PRETTY_FUNCTION__)); | |||
735 | FI.Widened = true; | |||
736 | ||||
737 | // Save the old/narrow induction phis, which we need to ignore in CheckPHIs. | |||
738 | FI.NarrowInnerInductionPHI = FI.InnerInductionPHI; | |||
739 | FI.NarrowOuterInductionPHI = FI.OuterInductionPHI; | |||
740 | ||||
741 | // After widening, rediscover all the loop components. | |||
742 | return CanFlattenLoopPair(FI, DT, LI, SE, AC, TTI); | |||
743 | } | |||
744 | ||||
745 | static bool FlattenLoopPair(FlattenInfo &FI, DominatorTree *DT, LoopInfo *LI, | |||
746 | ScalarEvolution *SE, AssumptionCache *AC, | |||
747 | const TargetTransformInfo *TTI, LPMUpdater *U) { | |||
748 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Loop flattening running on outer loop " << FI.OuterLoop->getHeader()->getName() << " and inner loop " << FI.InnerLoop->getHeader()-> getName() << " in " << FI.OuterLoop->getHeader ()->getParent()->getName() << "\n"; } } while (false ) | |||
749 | dbgs() << "Loop flattening running on outer loop "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Loop flattening running on outer loop " << FI.OuterLoop->getHeader()->getName() << " and inner loop " << FI.InnerLoop->getHeader()-> getName() << " in " << FI.OuterLoop->getHeader ()->getParent()->getName() << "\n"; } } while (false ) | |||
750 | << FI.OuterLoop->getHeader()->getName() << " and inner loop "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Loop flattening running on outer loop " << FI.OuterLoop->getHeader()->getName() << " and inner loop " << FI.InnerLoop->getHeader()-> getName() << " in " << FI.OuterLoop->getHeader ()->getParent()->getName() << "\n"; } } while (false ) | |||
751 | << FI.InnerLoop->getHeader()->getName() << " in "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Loop flattening running on outer loop " << FI.OuterLoop->getHeader()->getName() << " and inner loop " << FI.InnerLoop->getHeader()-> getName() << " in " << FI.OuterLoop->getHeader ()->getParent()->getName() << "\n"; } } while (false ) | |||
752 | << FI.OuterLoop->getHeader()->getParent()->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Loop flattening running on outer loop " << FI.OuterLoop->getHeader()->getName() << " and inner loop " << FI.InnerLoop->getHeader()-> getName() << " in " << FI.OuterLoop->getHeader ()->getParent()->getName() << "\n"; } } while (false ); | |||
753 | ||||
754 | if (!CanFlattenLoopPair(FI, DT, LI, SE, AC, TTI)) | |||
755 | return false; | |||
756 | ||||
757 | // Check if we can widen the induction variables to avoid overflow checks. | |||
758 | bool CanFlatten = CanWidenIV(FI, DT, LI, SE, AC, TTI); | |||
759 | ||||
760 | // It can happen that after widening of the IV, flattening may not be | |||
761 | // possible/happening, e.g. when it is deemed unprofitable. So bail here if | |||
762 | // that is the case. | |||
763 | // TODO: IV widening without performing the actual flattening transformation | |||
764 | // is not ideal. While this codegen change should not matter much, it is an | |||
765 | // unnecessary change which is better to avoid. It's unlikely this happens | |||
766 | // often, because if it's unprofitibale after widening, it should be | |||
767 | // unprofitabe before widening as checked in the first round of checks. But | |||
768 | // 'RepeatedInstructionThreshold' is set to only 2, which can probably be | |||
769 | // relaxed. Because this is making a code change (the IV widening, but not | |||
770 | // the flattening), we return true here. | |||
771 | if (FI.Widened
| |||
772 | return true; | |||
773 | ||||
774 | // If we have widened and can perform the transformation, do that here. | |||
775 | if (CanFlatten
| |||
776 | return DoFlattenLoopPair(FI, DT, LI, SE, AC, TTI, U); | |||
777 | ||||
778 | // Otherwise, if we haven't widened the IV, check if the new iteration | |||
779 | // variable might overflow. In this case, we need to version the loop, and | |||
780 | // select the original version at runtime if the iteration space is too | |||
781 | // large. | |||
782 | // TODO: We currently don't version the loop. | |||
783 | OverflowResult OR = checkOverflow(FI, DT, AC); | |||
784 | if (OR == OverflowResult::AlwaysOverflowsHigh || | |||
785 | OR == OverflowResult::AlwaysOverflowsLow) { | |||
786 | LLVM_DEBUG(dbgs() << "Multiply would always overflow, so not profitable\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Multiply would always overflow, so not profitable\n" ; } } while (false); | |||
787 | return false; | |||
788 | } else if (OR == OverflowResult::MayOverflow) { | |||
789 | LLVM_DEBUG(dbgs() << "Multiply might overflow, not flattening\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Multiply might overflow, not flattening\n" ; } } while (false); | |||
790 | return false; | |||
791 | } | |||
792 | ||||
793 | LLVM_DEBUG(dbgs() << "Multiply cannot overflow, modifying loop in-place\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-flatten")) { dbgs() << "Multiply cannot overflow, modifying loop in-place\n" ; } } while (false); | |||
794 | return DoFlattenLoopPair(FI, DT, LI, SE, AC, TTI, U); | |||
795 | } | |||
796 | ||||
797 | bool Flatten(LoopNest &LN, DominatorTree *DT, LoopInfo *LI, ScalarEvolution *SE, | |||
798 | AssumptionCache *AC, TargetTransformInfo *TTI, LPMUpdater *U) { | |||
799 | bool Changed = false; | |||
800 | for (Loop *InnerLoop : LN.getLoops()) { | |||
801 | auto *OuterLoop = InnerLoop->getParentLoop(); | |||
802 | if (!OuterLoop) | |||
803 | continue; | |||
804 | FlattenInfo FI(OuterLoop, InnerLoop); | |||
805 | Changed |= FlattenLoopPair(FI, DT, LI, SE, AC, TTI, U); | |||
806 | } | |||
807 | return Changed; | |||
808 | } | |||
809 | ||||
810 | PreservedAnalyses LoopFlattenPass::run(LoopNest &LN, LoopAnalysisManager &LAM, | |||
811 | LoopStandardAnalysisResults &AR, | |||
812 | LPMUpdater &U) { | |||
813 | ||||
814 | bool Changed = false; | |||
815 | ||||
816 | // The loop flattening pass requires loops to be | |||
817 | // in simplified form, and also needs LCSSA. Running | |||
818 | // this pass will simplify all loops that contain inner loops, | |||
819 | // regardless of whether anything ends up being flattened. | |||
820 | Changed |= Flatten(LN, &AR.DT, &AR.LI, &AR.SE, &AR.AC, &AR.TTI, &U); | |||
821 | ||||
822 | if (!Changed) | |||
823 | return PreservedAnalyses::all(); | |||
824 | ||||
825 | return getLoopPassPreservedAnalyses(); | |||
826 | } | |||
827 | ||||
828 | namespace { | |||
829 | class LoopFlattenLegacyPass : public FunctionPass { | |||
830 | public: | |||
831 | static char ID; // Pass ID, replacement for typeid | |||
832 | LoopFlattenLegacyPass() : FunctionPass(ID) { | |||
833 | initializeLoopFlattenLegacyPassPass(*PassRegistry::getPassRegistry()); | |||
834 | } | |||
835 | ||||
836 | // Possibly flatten loop L into its child. | |||
837 | bool runOnFunction(Function &F) override; | |||
838 | ||||
839 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
840 | getLoopAnalysisUsage(AU); | |||
841 | AU.addRequired<TargetTransformInfoWrapperPass>(); | |||
842 | AU.addPreserved<TargetTransformInfoWrapperPass>(); | |||
843 | AU.addRequired<AssumptionCacheTracker>(); | |||
844 | AU.addPreserved<AssumptionCacheTracker>(); | |||
845 | } | |||
846 | }; | |||
847 | } // namespace | |||
848 | ||||
849 | char LoopFlattenLegacyPass::ID = 0; | |||
850 | INITIALIZE_PASS_BEGIN(LoopFlattenLegacyPass, "loop-flatten", "Flattens loops",static void *initializeLoopFlattenLegacyPassPassOnce(PassRegistry &Registry) { | |||
851 | false, false)static void *initializeLoopFlattenLegacyPassPassOnce(PassRegistry &Registry) { | |||
852 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | |||
853 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | |||
854 | INITIALIZE_PASS_END(LoopFlattenLegacyPass, "loop-flatten", "Flattens loops",PassInfo *PI = new PassInfo( "Flattens loops", "loop-flatten" , &LoopFlattenLegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <LoopFlattenLegacyPass>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeLoopFlattenLegacyPassPassFlag ; void llvm::initializeLoopFlattenLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeLoopFlattenLegacyPassPassFlag , initializeLoopFlattenLegacyPassPassOnce, std::ref(Registry) ); } | |||
855 | false, false)PassInfo *PI = new PassInfo( "Flattens loops", "loop-flatten" , &LoopFlattenLegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <LoopFlattenLegacyPass>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeLoopFlattenLegacyPassPassFlag ; void llvm::initializeLoopFlattenLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeLoopFlattenLegacyPassPassFlag , initializeLoopFlattenLegacyPassPassOnce, std::ref(Registry) ); } | |||
856 | ||||
857 | FunctionPass *llvm::createLoopFlattenPass() { return new LoopFlattenLegacyPass(); } | |||
858 | ||||
859 | bool LoopFlattenLegacyPass::runOnFunction(Function &F) { | |||
860 | ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | |||
861 | LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | |||
862 | auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>(); | |||
863 | DominatorTree *DT = DTWP ? &DTWP->getDomTree() : nullptr; | |||
| ||||
864 | auto &TTIP = getAnalysis<TargetTransformInfoWrapperPass>(); | |||
865 | auto *TTI = &TTIP.getTTI(F); | |||
866 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | |||
867 | bool Changed = false; | |||
868 | for (Loop *L : *LI) { | |||
869 | auto LN = LoopNest::getLoopNest(*L, *SE); | |||
870 | Changed |= Flatten(*LN, DT, LI, SE, AC, TTI, nullptr); | |||
871 | } | |||
872 | return Changed; | |||
873 | } |