1 | //===-- UnrollLoopRuntime.cpp - Runtime Loop unrolling utilities ----------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file implements some loop unrolling utilities for loops with run-time |
10 | // trip counts. See LoopUnroll.cpp for unrolling loops with compile-time |
11 | // trip counts. |
12 | // |
13 | // The functions in this file are used to generate extra code when the |
14 | // run-time trip count modulo the unroll factor is not 0. When this is the |
15 | // case, we need to generate code to execute these 'left over' iterations. |
16 | // |
17 | // The current strategy generates an if-then-else sequence prior to the |
18 | // unrolled loop to execute the 'left over' iterations before or after the |
19 | // unrolled loop. |
20 | // |
21 | //===----------------------------------------------------------------------===// |
22 | |
23 | #include "llvm/ADT/Statistic.h" |
24 | #include "llvm/Analysis/DomTreeUpdater.h" |
25 | #include "llvm/Analysis/InstructionSimplify.h" |
26 | #include "llvm/Analysis/LoopIterator.h" |
27 | #include "llvm/Analysis/ScalarEvolution.h" |
28 | #include "llvm/Analysis/ValueTracking.h" |
29 | #include "llvm/IR/BasicBlock.h" |
30 | #include "llvm/IR/Dominators.h" |
31 | #include "llvm/IR/MDBuilder.h" |
32 | #include "llvm/IR/Module.h" |
33 | #include "llvm/IR/ProfDataUtils.h" |
34 | #include "llvm/Support/CommandLine.h" |
35 | #include "llvm/Support/Debug.h" |
36 | #include "llvm/Support/raw_ostream.h" |
37 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
38 | #include "llvm/Transforms/Utils/Cloning.h" |
39 | #include "llvm/Transforms/Utils/Local.h" |
40 | #include "llvm/Transforms/Utils/LoopUtils.h" |
41 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" |
42 | #include "llvm/Transforms/Utils/UnrollLoop.h" |
43 | |
44 | using namespace llvm; |
45 | |
46 | #define DEBUG_TYPE "loop-unroll" |
47 | |
48 | STATISTIC(NumRuntimeUnrolled, |
49 | "Number of loops unrolled with run-time trip counts" ); |
50 | static cl::opt<bool> UnrollRuntimeMultiExit( |
51 | "unroll-runtime-multi-exit" , cl::init(Val: false), cl::Hidden, |
52 | cl::desc("Allow runtime unrolling for loops with multiple exits, when " |
53 | "epilog is generated" )); |
54 | static cl::opt<bool> UnrollRuntimeOtherExitPredictable( |
55 | "unroll-runtime-other-exit-predictable" , cl::init(Val: false), cl::Hidden, |
56 | cl::desc("Assume the non latch exit block to be predictable" )); |
57 | |
58 | // Probability that the loop trip count is so small that after the prolog |
59 | // we do not enter the unrolled loop at all. |
60 | // It is unlikely that the loop trip count is smaller than the unroll factor; |
61 | // other than that, the choice of constant is not tuned yet. |
62 | static const uint32_t [] = {1, 127}; |
63 | // Probability that the loop trip count is so small that we skip the unrolled |
64 | // loop completely and immediately enter the epilogue loop. |
65 | // It is unlikely that the loop trip count is smaller than the unroll factor; |
66 | // other than that, the choice of constant is not tuned yet. |
67 | static const uint32_t [] = {1, 127}; |
68 | |
69 | /// Connect the unrolling prolog code to the original loop. |
70 | /// The unrolling prolog code contains code to execute the |
71 | /// 'extra' iterations if the run-time trip count modulo the |
72 | /// unroll count is non-zero. |
73 | /// |
74 | /// This function performs the following: |
75 | /// - Create PHI nodes at prolog end block to combine values |
76 | /// that exit the prolog code and jump around the prolog. |
77 | /// - Add a PHI operand to a PHI node at the loop exit block |
78 | /// for values that exit the prolog and go around the loop. |
79 | /// - Branch around the original loop if the trip count is less |
80 | /// than the unroll factor. |
81 | /// |
82 | static void ConnectProlog(Loop *L, Value *BECount, unsigned Count, |
83 | BasicBlock *PrologExit, |
84 | BasicBlock *OriginalLoopLatchExit, |
85 | BasicBlock *, BasicBlock *, |
86 | ValueToValueMapTy &VMap, DominatorTree *DT, |
87 | LoopInfo *LI, bool PreserveLCSSA, |
88 | ScalarEvolution &SE) { |
89 | // Loop structure should be the following: |
90 | // Preheader |
91 | // PrologHeader |
92 | // ... |
93 | // PrologLatch |
94 | // PrologExit |
95 | // NewPreheader |
96 | // Header |
97 | // ... |
98 | // Latch |
99 | // LatchExit |
100 | BasicBlock *Latch = L->getLoopLatch(); |
101 | assert(Latch && "Loop must have a latch" ); |
102 | BasicBlock *PrologLatch = cast<BasicBlock>(Val&: VMap[Latch]); |
103 | |
104 | // Create a PHI node for each outgoing value from the original loop |
105 | // (which means it is an outgoing value from the prolog code too). |
106 | // The new PHI node is inserted in the prolog end basic block. |
107 | // The new PHI node value is added as an operand of a PHI node in either |
108 | // the loop header or the loop exit block. |
109 | for (BasicBlock *Succ : successors(BB: Latch)) { |
110 | for (PHINode &PN : Succ->phis()) { |
111 | // Add a new PHI node to the prolog end block and add the |
112 | // appropriate incoming values. |
113 | // TODO: This code assumes that the PrologExit (or the LatchExit block for |
114 | // prolog loop) contains only one predecessor from the loop, i.e. the |
115 | // PrologLatch. When supporting multiple-exiting block loops, we can have |
116 | // two or more blocks that have the LatchExit as the target in the |
117 | // original loop. |
118 | PHINode *NewPN = PHINode::Create(Ty: PN.getType(), NumReservedValues: 2, NameStr: PN.getName() + ".unr" ); |
119 | NewPN->insertBefore(InsertPos: PrologExit->getFirstNonPHIIt()); |
120 | // Adding a value to the new PHI node from the original loop preheader. |
121 | // This is the value that skips all the prolog code. |
122 | if (L->contains(Inst: &PN)) { |
123 | // Succ is loop header. |
124 | NewPN->addIncoming(V: PN.getIncomingValueForBlock(BB: NewPreHeader), |
125 | BB: PreHeader); |
126 | } else { |
127 | // Succ is LatchExit. |
128 | NewPN->addIncoming(V: PoisonValue::get(T: PN.getType()), BB: PreHeader); |
129 | } |
130 | |
131 | Value *V = PN.getIncomingValueForBlock(BB: Latch); |
132 | if (Instruction *I = dyn_cast<Instruction>(Val: V)) { |
133 | if (L->contains(Inst: I)) { |
134 | V = VMap.lookup(Val: I); |
135 | } |
136 | } |
137 | // Adding a value to the new PHI node from the last prolog block |
138 | // that was created. |
139 | NewPN->addIncoming(V, BB: PrologLatch); |
140 | |
141 | // Update the existing PHI node operand with the value from the |
142 | // new PHI node. How this is done depends on if the existing |
143 | // PHI node is in the original loop block, or the exit block. |
144 | if (L->contains(Inst: &PN)) |
145 | PN.setIncomingValueForBlock(BB: NewPreHeader, V: NewPN); |
146 | else |
147 | PN.addIncoming(V: NewPN, BB: PrologExit); |
148 | SE.forgetLcssaPhiWithNewPredecessor(L, V: &PN); |
149 | } |
150 | } |
151 | |
152 | // Make sure that created prolog loop is in simplified form |
153 | SmallVector<BasicBlock *, 4> PrologExitPreds; |
154 | Loop *PrologLoop = LI->getLoopFor(BB: PrologLatch); |
155 | if (PrologLoop) { |
156 | for (BasicBlock *PredBB : predecessors(BB: PrologExit)) |
157 | if (PrologLoop->contains(BB: PredBB)) |
158 | PrologExitPreds.push_back(Elt: PredBB); |
159 | |
160 | SplitBlockPredecessors(BB: PrologExit, Preds: PrologExitPreds, Suffix: ".unr-lcssa" , DT, LI, |
161 | MSSAU: nullptr, PreserveLCSSA); |
162 | } |
163 | |
164 | // Create a branch around the original loop, which is taken if there are no |
165 | // iterations remaining to be executed after running the prologue. |
166 | Instruction *InsertPt = PrologExit->getTerminator(); |
167 | IRBuilder<> B(InsertPt); |
168 | |
169 | assert(Count != 0 && "nonsensical Count!" ); |
170 | |
171 | // If BECount <u (Count - 1) then (BECount + 1) % Count == (BECount + 1) |
172 | // This means %xtraiter is (BECount + 1) and all of the iterations of this |
173 | // loop were executed by the prologue. Note that if BECount <u (Count - 1) |
174 | // then (BECount + 1) cannot unsigned-overflow. |
175 | Value *BrLoopExit = |
176 | B.CreateICmpULT(LHS: BECount, RHS: ConstantInt::get(Ty: BECount->getType(), V: Count - 1)); |
177 | // Split the exit to maintain loop canonicalization guarantees |
178 | SmallVector<BasicBlock *, 4> Preds(predecessors(BB: OriginalLoopLatchExit)); |
179 | SplitBlockPredecessors(BB: OriginalLoopLatchExit, Preds, Suffix: ".unr-lcssa" , DT, LI, |
180 | MSSAU: nullptr, PreserveLCSSA); |
181 | // Add the branch to the exit block (around the unrolled loop) |
182 | MDNode *BranchWeights = nullptr; |
183 | if (hasBranchWeightMD(I: *Latch->getTerminator())) { |
184 | // Assume loop is nearly always entered. |
185 | MDBuilder MDB(B.getContext()); |
186 | BranchWeights = MDB.createBranchWeights(Weights: UnrolledLoopHeaderWeights); |
187 | } |
188 | B.CreateCondBr(Cond: BrLoopExit, True: OriginalLoopLatchExit, False: NewPreHeader, |
189 | BranchWeights); |
190 | InsertPt->eraseFromParent(); |
191 | if (DT) { |
192 | auto *NewDom = DT->findNearestCommonDominator(A: OriginalLoopLatchExit, |
193 | B: PrologExit); |
194 | DT->changeImmediateDominator(BB: OriginalLoopLatchExit, NewBB: NewDom); |
195 | } |
196 | } |
197 | |
198 | /// Connect the unrolling epilog code to the original loop. |
199 | /// The unrolling epilog code contains code to execute the |
200 | /// 'extra' iterations if the run-time trip count modulo the |
201 | /// unroll count is non-zero. |
202 | /// |
203 | /// This function performs the following: |
204 | /// - Update PHI nodes at the unrolling loop exit and epilog loop exit |
205 | /// - Create PHI nodes at the unrolling loop exit to combine |
206 | /// values that exit the unrolling loop code and jump around it. |
207 | /// - Update PHI operands in the epilog loop by the new PHI nodes |
208 | /// - Branch around the epilog loop if extra iters (ModVal) is zero. |
209 | /// |
210 | static void ConnectEpilog(Loop *L, Value *ModVal, BasicBlock *NewExit, |
211 | BasicBlock *Exit, BasicBlock *, |
212 | BasicBlock *, BasicBlock *, |
213 | ValueToValueMapTy &VMap, DominatorTree *DT, |
214 | LoopInfo *LI, bool PreserveLCSSA, ScalarEvolution &SE, |
215 | unsigned Count) { |
216 | BasicBlock *Latch = L->getLoopLatch(); |
217 | assert(Latch && "Loop must have a latch" ); |
218 | BasicBlock *EpilogLatch = cast<BasicBlock>(Val&: VMap[Latch]); |
219 | |
220 | // Loop structure should be the following: |
221 | // |
222 | // PreHeader |
223 | // NewPreHeader |
224 | // Header |
225 | // ... |
226 | // Latch |
227 | // NewExit (PN) |
228 | // EpilogPreHeader |
229 | // EpilogHeader |
230 | // ... |
231 | // EpilogLatch |
232 | // Exit (EpilogPN) |
233 | |
234 | // Update PHI nodes at NewExit and Exit. |
235 | for (PHINode &PN : NewExit->phis()) { |
236 | // PN should be used in another PHI located in Exit block as |
237 | // Exit was split by SplitBlockPredecessors into Exit and NewExit |
238 | // Basically it should look like: |
239 | // NewExit: |
240 | // PN = PHI [I, Latch] |
241 | // ... |
242 | // Exit: |
243 | // EpilogPN = PHI [PN, EpilogPreHeader], [X, Exit2], [Y, Exit2.epil] |
244 | // |
245 | // Exits from non-latch blocks point to the original exit block and the |
246 | // epilogue edges have already been added. |
247 | // |
248 | // There is EpilogPreHeader incoming block instead of NewExit as |
249 | // NewExit was spilt 1 more time to get EpilogPreHeader. |
250 | assert(PN.hasOneUse() && "The phi should have 1 use" ); |
251 | PHINode *EpilogPN = cast<PHINode>(Val: PN.use_begin()->getUser()); |
252 | assert(EpilogPN->getParent() == Exit && "EpilogPN should be in Exit block" ); |
253 | |
254 | // Add incoming PreHeader from branch around the Loop |
255 | PN.addIncoming(V: PoisonValue::get(T: PN.getType()), BB: PreHeader); |
256 | SE.forgetValue(V: &PN); |
257 | |
258 | Value *V = PN.getIncomingValueForBlock(BB: Latch); |
259 | Instruction *I = dyn_cast<Instruction>(Val: V); |
260 | if (I && L->contains(Inst: I)) |
261 | // If value comes from an instruction in the loop add VMap value. |
262 | V = VMap.lookup(Val: I); |
263 | // For the instruction out of the loop, constant or undefined value |
264 | // insert value itself. |
265 | EpilogPN->addIncoming(V, BB: EpilogLatch); |
266 | |
267 | assert(EpilogPN->getBasicBlockIndex(EpilogPreHeader) >= 0 && |
268 | "EpilogPN should have EpilogPreHeader incoming block" ); |
269 | // Change EpilogPreHeader incoming block to NewExit. |
270 | EpilogPN->setIncomingBlock(i: EpilogPN->getBasicBlockIndex(BB: EpilogPreHeader), |
271 | BB: NewExit); |
272 | // Now PHIs should look like: |
273 | // NewExit: |
274 | // PN = PHI [I, Latch], [poison, PreHeader] |
275 | // ... |
276 | // Exit: |
277 | // EpilogPN = PHI [PN, NewExit], [VMap[I], EpilogLatch] |
278 | } |
279 | |
280 | // Create PHI nodes at NewExit (from the unrolling loop Latch and PreHeader). |
281 | // Update corresponding PHI nodes in epilog loop. |
282 | for (BasicBlock *Succ : successors(BB: Latch)) { |
283 | // Skip this as we already updated phis in exit blocks. |
284 | if (!L->contains(BB: Succ)) |
285 | continue; |
286 | for (PHINode &PN : Succ->phis()) { |
287 | // Add new PHI nodes to the loop exit block and update epilog |
288 | // PHIs with the new PHI values. |
289 | PHINode *NewPN = PHINode::Create(Ty: PN.getType(), NumReservedValues: 2, NameStr: PN.getName() + ".unr" ); |
290 | NewPN->insertBefore(InsertPos: NewExit->getFirstNonPHIIt()); |
291 | // Adding a value to the new PHI node from the unrolling loop preheader. |
292 | NewPN->addIncoming(V: PN.getIncomingValueForBlock(BB: NewPreHeader), BB: PreHeader); |
293 | // Adding a value to the new PHI node from the unrolling loop latch. |
294 | NewPN->addIncoming(V: PN.getIncomingValueForBlock(BB: Latch), BB: Latch); |
295 | |
296 | // Update the existing PHI node operand with the value from the new PHI |
297 | // node. Corresponding instruction in epilog loop should be PHI. |
298 | PHINode *VPN = cast<PHINode>(Val&: VMap[&PN]); |
299 | VPN->setIncomingValueForBlock(BB: EpilogPreHeader, V: NewPN); |
300 | } |
301 | } |
302 | |
303 | Instruction *InsertPt = NewExit->getTerminator(); |
304 | IRBuilder<> B(InsertPt); |
305 | Value *BrLoopExit = B.CreateIsNotNull(Arg: ModVal, Name: "lcmp.mod" ); |
306 | assert(Exit && "Loop must have a single exit block only" ); |
307 | // Split the epilogue exit to maintain loop canonicalization guarantees |
308 | SmallVector<BasicBlock*, 4> Preds(predecessors(BB: Exit)); |
309 | SplitBlockPredecessors(BB: Exit, Preds, Suffix: ".epilog-lcssa" , DT, LI, MSSAU: nullptr, |
310 | PreserveLCSSA); |
311 | // Add the branch to the exit block (around the unrolling loop) |
312 | MDNode *BranchWeights = nullptr; |
313 | if (hasBranchWeightMD(I: *Latch->getTerminator())) { |
314 | // Assume equal distribution in interval [0, Count). |
315 | MDBuilder MDB(B.getContext()); |
316 | BranchWeights = MDB.createBranchWeights(TrueWeight: 1, FalseWeight: Count - 1); |
317 | } |
318 | B.CreateCondBr(Cond: BrLoopExit, True: EpilogPreHeader, False: Exit, BranchWeights); |
319 | InsertPt->eraseFromParent(); |
320 | if (DT) { |
321 | auto *NewDom = DT->findNearestCommonDominator(A: Exit, B: NewExit); |
322 | DT->changeImmediateDominator(BB: Exit, NewBB: NewDom); |
323 | } |
324 | |
325 | // Split the main loop exit to maintain canonicalization guarantees. |
326 | SmallVector<BasicBlock*, 4> NewExitPreds{Latch}; |
327 | SplitBlockPredecessors(BB: NewExit, Preds: NewExitPreds, Suffix: ".loopexit" , DT, LI, MSSAU: nullptr, |
328 | PreserveLCSSA); |
329 | } |
330 | |
331 | /// Create a clone of the blocks in a loop and connect them together. A new |
332 | /// loop will be created including all cloned blocks, and the iterator of the |
333 | /// new loop switched to count NewIter down to 0. |
334 | /// The cloned blocks should be inserted between InsertTop and InsertBot. |
335 | /// InsertTop should be new preheader, InsertBot new loop exit. |
336 | /// Returns the new cloned loop that is created. |
337 | static Loop * |
338 | CloneLoopBlocks(Loop *L, Value *NewIter, const bool UseEpilogRemainder, |
339 | const bool UnrollRemainder, |
340 | BasicBlock *InsertTop, |
341 | BasicBlock *InsertBot, BasicBlock *, |
342 | std::vector<BasicBlock *> &NewBlocks, |
343 | LoopBlocksDFS &LoopBlocks, ValueToValueMapTy &VMap, |
344 | DominatorTree *DT, LoopInfo *LI, unsigned Count) { |
345 | StringRef suffix = UseEpilogRemainder ? "epil" : "prol" ; |
346 | BasicBlock * = L->getHeader(); |
347 | BasicBlock *Latch = L->getLoopLatch(); |
348 | Function *F = Header->getParent(); |
349 | LoopBlocksDFS::RPOIterator BlockBegin = LoopBlocks.beginRPO(); |
350 | LoopBlocksDFS::RPOIterator BlockEnd = LoopBlocks.endRPO(); |
351 | Loop *ParentLoop = L->getParentLoop(); |
352 | NewLoopsMap NewLoops; |
353 | NewLoops[ParentLoop] = ParentLoop; |
354 | |
355 | // For each block in the original loop, create a new copy, |
356 | // and update the value map with the newly created values. |
357 | for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) { |
358 | BasicBlock *NewBB = CloneBasicBlock(BB: *BB, VMap, NameSuffix: "." + suffix, F); |
359 | NewBlocks.push_back(x: NewBB); |
360 | |
361 | addClonedBlockToLoopInfo(OriginalBB: *BB, ClonedBB: NewBB, LI, NewLoops); |
362 | |
363 | VMap[*BB] = NewBB; |
364 | if (Header == *BB) { |
365 | // For the first block, add a CFG connection to this newly |
366 | // created block. |
367 | InsertTop->getTerminator()->setSuccessor(Idx: 0, BB: NewBB); |
368 | } |
369 | |
370 | if (DT) { |
371 | if (Header == *BB) { |
372 | // The header is dominated by the preheader. |
373 | DT->addNewBlock(BB: NewBB, DomBB: InsertTop); |
374 | } else { |
375 | // Copy information from original loop to unrolled loop. |
376 | BasicBlock *IDomBB = DT->getNode(BB: *BB)->getIDom()->getBlock(); |
377 | DT->addNewBlock(BB: NewBB, DomBB: cast<BasicBlock>(Val&: VMap[IDomBB])); |
378 | } |
379 | } |
380 | |
381 | if (Latch == *BB) { |
382 | // For the last block, create a loop back to cloned head. |
383 | VMap.erase(Val: (*BB)->getTerminator()); |
384 | // Use an incrementing IV. Pre-incr/post-incr is backedge/trip count. |
385 | // Subtle: NewIter can be 0 if we wrapped when computing the trip count, |
386 | // thus we must compare the post-increment (wrapping) value. |
387 | BasicBlock *FirstLoopBB = cast<BasicBlock>(Val&: VMap[Header]); |
388 | BranchInst *LatchBR = cast<BranchInst>(Val: NewBB->getTerminator()); |
389 | IRBuilder<> Builder(LatchBR); |
390 | PHINode *NewIdx = |
391 | PHINode::Create(Ty: NewIter->getType(), NumReservedValues: 2, NameStr: suffix + ".iter" ); |
392 | NewIdx->insertBefore(InsertPos: FirstLoopBB->getFirstNonPHIIt()); |
393 | auto *Zero = ConstantInt::get(Ty: NewIdx->getType(), V: 0); |
394 | auto *One = ConstantInt::get(Ty: NewIdx->getType(), V: 1); |
395 | Value *IdxNext = |
396 | Builder.CreateAdd(LHS: NewIdx, RHS: One, Name: NewIdx->getName() + ".next" ); |
397 | Value *IdxCmp = Builder.CreateICmpNE(LHS: IdxNext, RHS: NewIter, Name: NewIdx->getName() + ".cmp" ); |
398 | MDNode *BranchWeights = nullptr; |
399 | if (hasBranchWeightMD(I: *LatchBR)) { |
400 | uint32_t ExitWeight; |
401 | uint32_t BackEdgeWeight; |
402 | if (Count >= 3) { |
403 | // Note: We do not enter this loop for zero-remainders. The check |
404 | // is at the end of the loop. We assume equal distribution between |
405 | // possible remainders in [1, Count). |
406 | ExitWeight = 1; |
407 | BackEdgeWeight = (Count - 2) / 2; |
408 | } else { |
409 | // Unnecessary backedge, should never be taken. The conditional |
410 | // jump should be optimized away later. |
411 | ExitWeight = 1; |
412 | BackEdgeWeight = 0; |
413 | } |
414 | MDBuilder MDB(Builder.getContext()); |
415 | BranchWeights = MDB.createBranchWeights(TrueWeight: BackEdgeWeight, FalseWeight: ExitWeight); |
416 | } |
417 | Builder.CreateCondBr(Cond: IdxCmp, True: FirstLoopBB, False: InsertBot, BranchWeights); |
418 | NewIdx->addIncoming(V: Zero, BB: InsertTop); |
419 | NewIdx->addIncoming(V: IdxNext, BB: NewBB); |
420 | LatchBR->eraseFromParent(); |
421 | } |
422 | } |
423 | |
424 | // Change the incoming values to the ones defined in the preheader or |
425 | // cloned loop. |
426 | for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(Val: I); ++I) { |
427 | PHINode *NewPHI = cast<PHINode>(Val&: VMap[&*I]); |
428 | unsigned idx = NewPHI->getBasicBlockIndex(BB: Preheader); |
429 | NewPHI->setIncomingBlock(i: idx, BB: InsertTop); |
430 | BasicBlock *NewLatch = cast<BasicBlock>(Val&: VMap[Latch]); |
431 | idx = NewPHI->getBasicBlockIndex(BB: Latch); |
432 | Value *InVal = NewPHI->getIncomingValue(i: idx); |
433 | NewPHI->setIncomingBlock(i: idx, BB: NewLatch); |
434 | if (Value *V = VMap.lookup(Val: InVal)) |
435 | NewPHI->setIncomingValue(i: idx, V); |
436 | } |
437 | |
438 | Loop *NewLoop = NewLoops[L]; |
439 | assert(NewLoop && "L should have been cloned" ); |
440 | MDNode *LoopID = NewLoop->getLoopID(); |
441 | |
442 | // Only add loop metadata if the loop is not going to be completely |
443 | // unrolled. |
444 | if (UnrollRemainder) |
445 | return NewLoop; |
446 | |
447 | std::optional<MDNode *> NewLoopID = makeFollowupLoopID( |
448 | OrigLoopID: LoopID, FollowupAttrs: {LLVMLoopUnrollFollowupAll, LLVMLoopUnrollFollowupRemainder}); |
449 | if (NewLoopID) { |
450 | NewLoop->setLoopID(*NewLoopID); |
451 | |
452 | // Do not setLoopAlreadyUnrolled if loop attributes have been defined |
453 | // explicitly. |
454 | return NewLoop; |
455 | } |
456 | |
457 | // Add unroll disable metadata to disable future unrolling for this loop. |
458 | NewLoop->setLoopAlreadyUnrolled(); |
459 | return NewLoop; |
460 | } |
461 | |
462 | /// Returns true if we can profitably unroll the multi-exit loop L. Currently, |
463 | /// we return true only if UnrollRuntimeMultiExit is set to true. |
464 | static bool canProfitablyRuntimeUnrollMultiExitLoop( |
465 | Loop *L, SmallVectorImpl<BasicBlock *> &OtherExits, BasicBlock *LatchExit, |
466 | bool UseEpilogRemainder) { |
467 | |
468 | // The main pain point with multi-exit loop unrolling is that once unrolled, |
469 | // we will not be able to merge all blocks into a straight line code. |
470 | // There are branches within the unrolled loop that go to the OtherExits. |
471 | // The second point is the increase in code size, but this is true |
472 | // irrespective of multiple exits. |
473 | |
474 | // Note: Both the heuristics below are coarse grained. We are essentially |
475 | // enabling unrolling of loops that have a single side exit other than the |
476 | // normal LatchExit (i.e. exiting into a deoptimize block). |
477 | // The heuristics considered are: |
478 | // 1. low number of branches in the unrolled version. |
479 | // 2. high predictability of these extra branches. |
480 | // We avoid unrolling loops that have more than two exiting blocks. This |
481 | // limits the total number of branches in the unrolled loop to be atmost |
482 | // the unroll factor (since one of the exiting blocks is the latch block). |
483 | SmallVector<BasicBlock*, 4> ExitingBlocks; |
484 | L->getExitingBlocks(ExitingBlocks); |
485 | if (ExitingBlocks.size() > 2) |
486 | return false; |
487 | |
488 | // Allow unrolling of loops with no non latch exit blocks. |
489 | if (OtherExits.size() == 0) |
490 | return true; |
491 | |
492 | // The second heuristic is that L has one exit other than the latchexit and |
493 | // that exit is a deoptimize block. We know that deoptimize blocks are rarely |
494 | // taken, which also implies the branch leading to the deoptimize block is |
495 | // highly predictable. When UnrollRuntimeOtherExitPredictable is specified, we |
496 | // assume the other exit branch is predictable even if it has no deoptimize |
497 | // call. |
498 | return (OtherExits.size() == 1 && |
499 | (UnrollRuntimeOtherExitPredictable || |
500 | OtherExits[0]->getPostdominatingDeoptimizeCall())); |
501 | // TODO: These can be fine-tuned further to consider code size or deopt states |
502 | // that are captured by the deoptimize exit block. |
503 | // Also, we can extend this to support more cases, if we actually |
504 | // know of kinds of multiexit loops that would benefit from unrolling. |
505 | } |
506 | |
507 | /// Calculate ModVal = (BECount + 1) % Count on the abstract integer domain |
508 | /// accounting for the possibility of unsigned overflow in the 2s complement |
509 | /// domain. Preconditions: |
510 | /// 1) TripCount = BECount + 1 (allowing overflow) |
511 | /// 2) Log2(Count) <= BitWidth(BECount) |
512 | static Value *CreateTripRemainder(IRBuilder<> &B, Value *BECount, |
513 | Value *TripCount, unsigned Count) { |
514 | // Note that TripCount is BECount + 1. |
515 | if (isPowerOf2_32(Value: Count)) |
516 | // If the expression is zero, then either: |
517 | // 1. There are no iterations to be run in the prolog/epilog loop. |
518 | // OR |
519 | // 2. The addition computing TripCount overflowed. |
520 | // |
521 | // If (2) is true, we know that TripCount really is (1 << BEWidth) and so |
522 | // the number of iterations that remain to be run in the original loop is a |
523 | // multiple Count == (1 << Log2(Count)) because Log2(Count) <= BEWidth (a |
524 | // precondition of this method). |
525 | return B.CreateAnd(LHS: TripCount, RHS: Count - 1, Name: "xtraiter" ); |
526 | |
527 | // As (BECount + 1) can potentially unsigned overflow we count |
528 | // (BECount % Count) + 1 which is overflow safe as BECount % Count < Count. |
529 | Constant *CountC = ConstantInt::get(Ty: BECount->getType(), V: Count); |
530 | Value *ModValTmp = B.CreateURem(LHS: BECount, RHS: CountC); |
531 | Value *ModValAdd = B.CreateAdd(LHS: ModValTmp, |
532 | RHS: ConstantInt::get(Ty: ModValTmp->getType(), V: 1)); |
533 | // At that point (BECount % Count) + 1 could be equal to Count. |
534 | // To handle this case we need to take mod by Count one more time. |
535 | return B.CreateURem(LHS: ModValAdd, RHS: CountC, Name: "xtraiter" ); |
536 | } |
537 | |
538 | |
539 | /// Insert code in the prolog/epilog code when unrolling a loop with a |
540 | /// run-time trip-count. |
541 | /// |
542 | /// This method assumes that the loop unroll factor is total number |
543 | /// of loop bodies in the loop after unrolling. (Some folks refer |
544 | /// to the unroll factor as the number of *extra* copies added). |
545 | /// We assume also that the loop unroll factor is a power-of-two. So, after |
546 | /// unrolling the loop, the number of loop bodies executed is 2, |
547 | /// 4, 8, etc. Note - LLVM converts the if-then-sequence to a switch |
548 | /// instruction in SimplifyCFG.cpp. Then, the backend decides how code for |
549 | /// the switch instruction is generated. |
550 | /// |
551 | /// ***Prolog case*** |
552 | /// extraiters = tripcount % loopfactor |
553 | /// if (extraiters == 0) jump Loop: |
554 | /// else jump Prol: |
555 | /// Prol: LoopBody; |
556 | /// extraiters -= 1 // Omitted if unroll factor is 2. |
557 | /// if (extraiters != 0) jump Prol: // Omitted if unroll factor is 2. |
558 | /// if (tripcount < loopfactor) jump End: |
559 | /// Loop: |
560 | /// ... |
561 | /// End: |
562 | /// |
563 | /// ***Epilog case*** |
564 | /// extraiters = tripcount % loopfactor |
565 | /// if (tripcount < loopfactor) jump LoopExit: |
566 | /// unroll_iters = tripcount - extraiters |
567 | /// Loop: LoopBody; (executes unroll_iter times); |
568 | /// unroll_iter -= 1 |
569 | /// if (unroll_iter != 0) jump Loop: |
570 | /// LoopExit: |
571 | /// if (extraiters == 0) jump EpilExit: |
572 | /// Epil: LoopBody; (executes extraiters times) |
573 | /// extraiters -= 1 // Omitted if unroll factor is 2. |
574 | /// if (extraiters != 0) jump Epil: // Omitted if unroll factor is 2. |
575 | /// EpilExit: |
576 | |
577 | bool llvm::UnrollRuntimeLoopRemainder( |
578 | Loop *L, unsigned Count, bool AllowExpensiveTripCount, |
579 | bool UseEpilogRemainder, bool UnrollRemainder, bool ForgetAllSCEV, |
580 | LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, AssumptionCache *AC, |
581 | const TargetTransformInfo *TTI, bool PreserveLCSSA, |
582 | unsigned SCEVExpansionBudget, bool RuntimeUnrollMultiExit, |
583 | Loop **ResultLoop) { |
584 | LLVM_DEBUG(dbgs() << "Trying runtime unrolling on Loop: \n" ); |
585 | LLVM_DEBUG(L->dump()); |
586 | LLVM_DEBUG(UseEpilogRemainder ? dbgs() << "Using epilog remainder.\n" |
587 | : dbgs() << "Using prolog remainder.\n" ); |
588 | |
589 | // Make sure the loop is in canonical form. |
590 | if (!L->isLoopSimplifyForm()) { |
591 | LLVM_DEBUG(dbgs() << "Not in simplify form!\n" ); |
592 | return false; |
593 | } |
594 | |
595 | // Guaranteed by LoopSimplifyForm. |
596 | BasicBlock *Latch = L->getLoopLatch(); |
597 | BasicBlock * = L->getHeader(); |
598 | |
599 | BranchInst *LatchBR = cast<BranchInst>(Val: Latch->getTerminator()); |
600 | |
601 | if (!LatchBR || LatchBR->isUnconditional()) { |
602 | // The loop-rotate pass can be helpful to avoid this in many cases. |
603 | LLVM_DEBUG( |
604 | dbgs() |
605 | << "Loop latch not terminated by a conditional branch.\n" ); |
606 | return false; |
607 | } |
608 | |
609 | unsigned ExitIndex = LatchBR->getSuccessor(i: 0) == Header ? 1 : 0; |
610 | BasicBlock *LatchExit = LatchBR->getSuccessor(i: ExitIndex); |
611 | |
612 | if (L->contains(BB: LatchExit)) { |
613 | // Cloning the loop basic blocks (`CloneLoopBlocks`) requires that one of the |
614 | // targets of the Latch be an exit block out of the loop. |
615 | LLVM_DEBUG( |
616 | dbgs() |
617 | << "One of the loop latch successors must be the exit block.\n" ); |
618 | return false; |
619 | } |
620 | |
621 | // These are exit blocks other than the target of the latch exiting block. |
622 | SmallVector<BasicBlock *, 4> OtherExits; |
623 | L->getUniqueNonLatchExitBlocks(ExitBlocks&: OtherExits); |
624 | // Support only single exit and exiting block unless multi-exit loop |
625 | // unrolling is enabled. |
626 | if (!L->getExitingBlock() || OtherExits.size()) { |
627 | // We rely on LCSSA form being preserved when the exit blocks are transformed. |
628 | // (Note that only an off-by-default mode of the old PM disables PreserveLCCA.) |
629 | if (!PreserveLCSSA) |
630 | return false; |
631 | |
632 | // Priority goes to UnrollRuntimeMultiExit if it's supplied. |
633 | if (UnrollRuntimeMultiExit.getNumOccurrences()) { |
634 | if (!UnrollRuntimeMultiExit) |
635 | return false; |
636 | } else { |
637 | // Otherwise perform multi-exit unrolling, if either the target indicates |
638 | // it is profitable or the general profitability heuristics apply. |
639 | if (!RuntimeUnrollMultiExit && |
640 | !canProfitablyRuntimeUnrollMultiExitLoop(L, OtherExits, LatchExit, |
641 | UseEpilogRemainder)) { |
642 | LLVM_DEBUG(dbgs() << "Multiple exit/exiting blocks in loop and " |
643 | "multi-exit unrolling not enabled!\n" ); |
644 | return false; |
645 | } |
646 | } |
647 | } |
648 | // Use Scalar Evolution to compute the trip count. This allows more loops to |
649 | // be unrolled than relying on induction var simplification. |
650 | if (!SE) |
651 | return false; |
652 | |
653 | // Only unroll loops with a computable trip count. |
654 | // We calculate the backedge count by using getExitCount on the Latch block, |
655 | // which is proven to be the only exiting block in this loop. This is same as |
656 | // calculating getBackedgeTakenCount on the loop (which computes SCEV for all |
657 | // exiting blocks). |
658 | const SCEV *BECountSC = SE->getExitCount(L, ExitingBlock: Latch); |
659 | if (isa<SCEVCouldNotCompute>(Val: BECountSC)) { |
660 | LLVM_DEBUG(dbgs() << "Could not compute exit block SCEV\n" ); |
661 | return false; |
662 | } |
663 | |
664 | unsigned BEWidth = cast<IntegerType>(Val: BECountSC->getType())->getBitWidth(); |
665 | |
666 | // Add 1 since the backedge count doesn't include the first loop iteration. |
667 | // (Note that overflow can occur, this is handled explicitly below) |
668 | const SCEV *TripCountSC = |
669 | SE->getAddExpr(LHS: BECountSC, RHS: SE->getConstant(Ty: BECountSC->getType(), V: 1)); |
670 | if (isa<SCEVCouldNotCompute>(Val: TripCountSC)) { |
671 | LLVM_DEBUG(dbgs() << "Could not compute trip count SCEV.\n" ); |
672 | return false; |
673 | } |
674 | |
675 | BasicBlock * = L->getLoopPreheader(); |
676 | BranchInst * = cast<BranchInst>(Val: PreHeader->getTerminator()); |
677 | const DataLayout &DL = Header->getDataLayout(); |
678 | SCEVExpander Expander(*SE, DL, "loop-unroll" ); |
679 | if (!AllowExpensiveTripCount && |
680 | Expander.isHighCostExpansion(Exprs: TripCountSC, L, Budget: SCEVExpansionBudget, TTI, |
681 | At: PreHeaderBR)) { |
682 | LLVM_DEBUG(dbgs() << "High cost for expanding trip count scev!\n" ); |
683 | return false; |
684 | } |
685 | |
686 | // This constraint lets us deal with an overflowing trip count easily; see the |
687 | // comment on ModVal below. |
688 | if (Log2_32(Value: Count) > BEWidth) { |
689 | LLVM_DEBUG( |
690 | dbgs() |
691 | << "Count failed constraint on overflow trip count calculation.\n" ); |
692 | return false; |
693 | } |
694 | |
695 | // Loop structure is the following: |
696 | // |
697 | // PreHeader |
698 | // Header |
699 | // ... |
700 | // Latch |
701 | // LatchExit |
702 | |
703 | BasicBlock *; |
704 | BasicBlock *NewExit = nullptr; |
705 | BasicBlock *PrologExit = nullptr; |
706 | BasicBlock * = nullptr; |
707 | BasicBlock * = nullptr; |
708 | |
709 | if (UseEpilogRemainder) { |
710 | // If epilog remainder |
711 | // Split PreHeader to insert a branch around loop for unrolling. |
712 | NewPreHeader = SplitBlock(Old: PreHeader, SplitPt: PreHeader->getTerminator(), DT, LI); |
713 | NewPreHeader->setName(PreHeader->getName() + ".new" ); |
714 | // Split LatchExit to create phi nodes from branch above. |
715 | NewExit = SplitBlockPredecessors(BB: LatchExit, Preds: {Latch}, Suffix: ".unr-lcssa" , DT, LI, |
716 | MSSAU: nullptr, PreserveLCSSA); |
717 | // NewExit gets its DebugLoc from LatchExit, which is not part of the |
718 | // original Loop. |
719 | // Fix this by setting Loop's DebugLoc to NewExit. |
720 | auto *NewExitTerminator = NewExit->getTerminator(); |
721 | NewExitTerminator->setDebugLoc(Header->getTerminator()->getDebugLoc()); |
722 | // Split NewExit to insert epilog remainder loop. |
723 | EpilogPreHeader = SplitBlock(Old: NewExit, SplitPt: NewExitTerminator, DT, LI); |
724 | EpilogPreHeader->setName(Header->getName() + ".epil.preheader" ); |
725 | |
726 | // If the latch exits from multiple level of nested loops, then |
727 | // by assumption there must be another loop exit which branches to the |
728 | // outer loop and we must adjust the loop for the newly inserted blocks |
729 | // to account for the fact that our epilogue is still in the same outer |
730 | // loop. Note that this leaves loopinfo temporarily out of sync with the |
731 | // CFG until the actual epilogue loop is inserted. |
732 | if (auto *ParentL = L->getParentLoop()) |
733 | if (LI->getLoopFor(BB: LatchExit) != ParentL) { |
734 | LI->removeBlock(BB: NewExit); |
735 | ParentL->addBasicBlockToLoop(NewBB: NewExit, LI&: *LI); |
736 | LI->removeBlock(BB: EpilogPreHeader); |
737 | ParentL->addBasicBlockToLoop(NewBB: EpilogPreHeader, LI&: *LI); |
738 | } |
739 | |
740 | } else { |
741 | // If prolog remainder |
742 | // Split the original preheader twice to insert prolog remainder loop |
743 | PrologPreHeader = SplitEdge(From: PreHeader, To: Header, DT, LI); |
744 | PrologPreHeader->setName(Header->getName() + ".prol.preheader" ); |
745 | PrologExit = SplitBlock(Old: PrologPreHeader, SplitPt: PrologPreHeader->getTerminator(), |
746 | DT, LI); |
747 | PrologExit->setName(Header->getName() + ".prol.loopexit" ); |
748 | // Split PrologExit to get NewPreHeader. |
749 | NewPreHeader = SplitBlock(Old: PrologExit, SplitPt: PrologExit->getTerminator(), DT, LI); |
750 | NewPreHeader->setName(PreHeader->getName() + ".new" ); |
751 | } |
752 | // Loop structure should be the following: |
753 | // Epilog Prolog |
754 | // |
755 | // PreHeader PreHeader |
756 | // *NewPreHeader *PrologPreHeader |
757 | // Header *PrologExit |
758 | // ... *NewPreHeader |
759 | // Latch Header |
760 | // *NewExit ... |
761 | // *EpilogPreHeader Latch |
762 | // LatchExit LatchExit |
763 | |
764 | // Calculate conditions for branch around loop for unrolling |
765 | // in epilog case and around prolog remainder loop in prolog case. |
766 | // Compute the number of extra iterations required, which is: |
767 | // extra iterations = run-time trip count % loop unroll factor |
768 | PreHeaderBR = cast<BranchInst>(Val: PreHeader->getTerminator()); |
769 | IRBuilder<> B(PreHeaderBR); |
770 | Value *TripCount = Expander.expandCodeFor(SH: TripCountSC, Ty: TripCountSC->getType(), |
771 | I: PreHeaderBR); |
772 | Value *BECount; |
773 | // If there are other exits before the latch, that may cause the latch exit |
774 | // branch to never be executed, and the latch exit count may be poison. |
775 | // In this case, freeze the TripCount and base BECount on the frozen |
776 | // TripCount. We will introduce two branches using these values, and it's |
777 | // important that they see a consistent value (which would not be guaranteed |
778 | // if were frozen independently.) |
779 | if ((!OtherExits.empty() || !SE->loopHasNoAbnormalExits(L)) && |
780 | !isGuaranteedNotToBeUndefOrPoison(V: TripCount, AC, CtxI: PreHeaderBR, DT)) { |
781 | TripCount = B.CreateFreeze(V: TripCount); |
782 | BECount = |
783 | B.CreateAdd(LHS: TripCount, RHS: Constant::getAllOnesValue(Ty: TripCount->getType())); |
784 | } else { |
785 | // If we don't need to freeze, use SCEVExpander for BECount as well, to |
786 | // allow slightly better value reuse. |
787 | BECount = |
788 | Expander.expandCodeFor(SH: BECountSC, Ty: BECountSC->getType(), I: PreHeaderBR); |
789 | } |
790 | |
791 | Value * const ModVal = CreateTripRemainder(B, BECount, TripCount, Count); |
792 | |
793 | Value *BranchVal = |
794 | UseEpilogRemainder ? B.CreateICmpULT(LHS: BECount, |
795 | RHS: ConstantInt::get(Ty: BECount->getType(), |
796 | V: Count - 1)) : |
797 | B.CreateIsNotNull(Arg: ModVal, Name: "lcmp.mod" ); |
798 | BasicBlock *RemainderLoop = UseEpilogRemainder ? NewExit : PrologPreHeader; |
799 | BasicBlock *UnrollingLoop = UseEpilogRemainder ? NewPreHeader : PrologExit; |
800 | // Branch to either remainder (extra iterations) loop or unrolling loop. |
801 | MDNode *BranchWeights = nullptr; |
802 | if (hasBranchWeightMD(I: *Latch->getTerminator())) { |
803 | // Assume loop is nearly always entered. |
804 | MDBuilder MDB(B.getContext()); |
805 | BranchWeights = MDB.createBranchWeights(Weights: EpilogHeaderWeights); |
806 | } |
807 | B.CreateCondBr(Cond: BranchVal, True: RemainderLoop, False: UnrollingLoop, BranchWeights); |
808 | PreHeaderBR->eraseFromParent(); |
809 | if (DT) { |
810 | if (UseEpilogRemainder) |
811 | DT->changeImmediateDominator(BB: NewExit, NewBB: PreHeader); |
812 | else |
813 | DT->changeImmediateDominator(BB: PrologExit, NewBB: PreHeader); |
814 | } |
815 | Function *F = Header->getParent(); |
816 | // Get an ordered list of blocks in the loop to help with the ordering of the |
817 | // cloned blocks in the prolog/epilog code |
818 | LoopBlocksDFS LoopBlocks(L); |
819 | LoopBlocks.perform(LI); |
820 | |
821 | // |
822 | // For each extra loop iteration, create a copy of the loop's basic blocks |
823 | // and generate a condition that branches to the copy depending on the |
824 | // number of 'left over' iterations. |
825 | // |
826 | std::vector<BasicBlock *> NewBlocks; |
827 | ValueToValueMapTy VMap; |
828 | |
829 | // Clone all the basic blocks in the loop. If Count is 2, we don't clone |
830 | // the loop, otherwise we create a cloned loop to execute the extra |
831 | // iterations. This function adds the appropriate CFG connections. |
832 | BasicBlock *InsertBot = UseEpilogRemainder ? LatchExit : PrologExit; |
833 | BasicBlock *InsertTop = UseEpilogRemainder ? EpilogPreHeader : PrologPreHeader; |
834 | Loop *remainderLoop = CloneLoopBlocks( |
835 | L, NewIter: ModVal, UseEpilogRemainder, UnrollRemainder, InsertTop, InsertBot, |
836 | Preheader: NewPreHeader, NewBlocks, LoopBlocks, VMap, DT, LI, Count); |
837 | |
838 | // Insert the cloned blocks into the function. |
839 | F->splice(ToIt: InsertBot->getIterator(), FromF: F, FromBeginIt: NewBlocks[0]->getIterator(), FromEndIt: F->end()); |
840 | |
841 | // Now the loop blocks are cloned and the other exiting blocks from the |
842 | // remainder are connected to the original Loop's exit blocks. The remaining |
843 | // work is to update the phi nodes in the original loop, and take in the |
844 | // values from the cloned region. |
845 | for (auto *BB : OtherExits) { |
846 | // Given we preserve LCSSA form, we know that the values used outside the |
847 | // loop will be used through these phi nodes at the exit blocks that are |
848 | // transformed below. |
849 | for (PHINode &PN : BB->phis()) { |
850 | unsigned oldNumOperands = PN.getNumIncomingValues(); |
851 | // Add the incoming values from the remainder code to the end of the phi |
852 | // node. |
853 | for (unsigned i = 0; i < oldNumOperands; i++){ |
854 | auto *PredBB =PN.getIncomingBlock(i); |
855 | if (PredBB == Latch) |
856 | // The latch exit is handled separately, see connectX |
857 | continue; |
858 | if (!L->contains(BB: PredBB)) |
859 | // Even if we had dedicated exits, the code above inserted an |
860 | // extra branch which can reach the latch exit. |
861 | continue; |
862 | |
863 | auto *V = PN.getIncomingValue(i); |
864 | if (Instruction *I = dyn_cast<Instruction>(Val: V)) |
865 | if (L->contains(Inst: I)) |
866 | V = VMap.lookup(Val: I); |
867 | PN.addIncoming(V, BB: cast<BasicBlock>(Val&: VMap[PredBB])); |
868 | } |
869 | } |
870 | #if defined(EXPENSIVE_CHECKS) && !defined(NDEBUG) |
871 | for (BasicBlock *SuccBB : successors(BB)) { |
872 | assert(!(llvm::is_contained(OtherExits, SuccBB) || SuccBB == LatchExit) && |
873 | "Breaks the definition of dedicated exits!" ); |
874 | } |
875 | #endif |
876 | } |
877 | |
878 | // Update the immediate dominator of the exit blocks and blocks that are |
879 | // reachable from the exit blocks. This is needed because we now have paths |
880 | // from both the original loop and the remainder code reaching the exit |
881 | // blocks. While the IDom of these exit blocks were from the original loop, |
882 | // now the IDom is the preheader (which decides whether the original loop or |
883 | // remainder code should run). |
884 | if (DT && !L->getExitingBlock()) { |
885 | SmallVector<BasicBlock *, 16> ChildrenToUpdate; |
886 | // NB! We have to examine the dom children of all loop blocks, not just |
887 | // those which are the IDom of the exit blocks. This is because blocks |
888 | // reachable from the exit blocks can have their IDom as the nearest common |
889 | // dominator of the exit blocks. |
890 | for (auto *BB : L->blocks()) { |
891 | auto *DomNodeBB = DT->getNode(BB); |
892 | for (auto *DomChild : DomNodeBB->children()) { |
893 | auto *DomChildBB = DomChild->getBlock(); |
894 | if (!L->contains(L: LI->getLoopFor(BB: DomChildBB))) |
895 | ChildrenToUpdate.push_back(Elt: DomChildBB); |
896 | } |
897 | } |
898 | for (auto *BB : ChildrenToUpdate) |
899 | DT->changeImmediateDominator(BB, NewBB: PreHeader); |
900 | } |
901 | |
902 | // Loop structure should be the following: |
903 | // Epilog Prolog |
904 | // |
905 | // PreHeader PreHeader |
906 | // NewPreHeader PrologPreHeader |
907 | // Header PrologHeader |
908 | // ... ... |
909 | // Latch PrologLatch |
910 | // NewExit PrologExit |
911 | // EpilogPreHeader NewPreHeader |
912 | // EpilogHeader Header |
913 | // ... ... |
914 | // EpilogLatch Latch |
915 | // LatchExit LatchExit |
916 | |
917 | // Rewrite the cloned instruction operands to use the values created when the |
918 | // clone is created. |
919 | for (BasicBlock *BB : NewBlocks) { |
920 | Module *M = BB->getModule(); |
921 | for (Instruction &I : *BB) { |
922 | RemapInstruction(I: &I, VM&: VMap, |
923 | Flags: RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); |
924 | RemapDbgRecordRange(M, Range: I.getDbgRecordRange(), VM&: VMap, |
925 | Flags: RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); |
926 | } |
927 | } |
928 | |
929 | if (UseEpilogRemainder) { |
930 | // Connect the epilog code to the original loop and update the |
931 | // PHI functions. |
932 | ConnectEpilog(L, ModVal, NewExit, Exit: LatchExit, PreHeader, EpilogPreHeader, |
933 | NewPreHeader, VMap, DT, LI, PreserveLCSSA, SE&: *SE, Count); |
934 | |
935 | // Update counter in loop for unrolling. |
936 | // Use an incrementing IV. Pre-incr/post-incr is backedge/trip count. |
937 | // Subtle: TestVal can be 0 if we wrapped when computing the trip count, |
938 | // thus we must compare the post-increment (wrapping) value. |
939 | IRBuilder<> B2(NewPreHeader->getTerminator()); |
940 | Value *TestVal = B2.CreateSub(LHS: TripCount, RHS: ModVal, Name: "unroll_iter" ); |
941 | BranchInst *LatchBR = cast<BranchInst>(Val: Latch->getTerminator()); |
942 | PHINode *NewIdx = PHINode::Create(Ty: TestVal->getType(), NumReservedValues: 2, NameStr: "niter" ); |
943 | NewIdx->insertBefore(InsertPos: Header->getFirstNonPHIIt()); |
944 | B2.SetInsertPoint(LatchBR); |
945 | auto *Zero = ConstantInt::get(Ty: NewIdx->getType(), V: 0); |
946 | auto *One = ConstantInt::get(Ty: NewIdx->getType(), V: 1); |
947 | Value *IdxNext = B2.CreateAdd(LHS: NewIdx, RHS: One, Name: NewIdx->getName() + ".next" ); |
948 | auto Pred = LatchBR->getSuccessor(i: 0) == Header ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ; |
949 | Value *IdxCmp = B2.CreateICmp(P: Pred, LHS: IdxNext, RHS: TestVal, Name: NewIdx->getName() + ".ncmp" ); |
950 | NewIdx->addIncoming(V: Zero, BB: NewPreHeader); |
951 | NewIdx->addIncoming(V: IdxNext, BB: Latch); |
952 | LatchBR->setCondition(IdxCmp); |
953 | } else { |
954 | // Connect the prolog code to the original loop and update the |
955 | // PHI functions. |
956 | ConnectProlog(L, BECount, Count, PrologExit, OriginalLoopLatchExit: LatchExit, PreHeader, |
957 | NewPreHeader, VMap, DT, LI, PreserveLCSSA, SE&: *SE); |
958 | } |
959 | |
960 | // If this loop is nested, then the loop unroller changes the code in the any |
961 | // of its parent loops, so the Scalar Evolution pass needs to be run again. |
962 | SE->forgetTopmostLoop(L); |
963 | |
964 | // Verify that the Dom Tree and Loop Info are correct. |
965 | #if defined(EXPENSIVE_CHECKS) && !defined(NDEBUG) |
966 | if (DT) { |
967 | assert(DT->verify(DominatorTree::VerificationLevel::Full)); |
968 | LI->verify(*DT); |
969 | } |
970 | #endif |
971 | |
972 | // For unroll factor 2 remainder loop will have 1 iteration. |
973 | if (Count == 2 && DT && LI && SE) { |
974 | // TODO: This code could probably be pulled out into a helper function |
975 | // (e.g. breakLoopBackedgeAndSimplify) and reused in loop-deletion. |
976 | BasicBlock *RemainderLatch = remainderLoop->getLoopLatch(); |
977 | assert(RemainderLatch); |
978 | SmallVector<BasicBlock *> RemainderBlocks(remainderLoop->getBlocks()); |
979 | breakLoopBackedge(L: remainderLoop, DT&: *DT, SE&: *SE, LI&: *LI, MSSA: nullptr); |
980 | remainderLoop = nullptr; |
981 | |
982 | // Simplify loop values after breaking the backedge |
983 | const DataLayout &DL = L->getHeader()->getDataLayout(); |
984 | SmallVector<WeakTrackingVH, 16> DeadInsts; |
985 | for (BasicBlock *BB : RemainderBlocks) { |
986 | for (Instruction &Inst : llvm::make_early_inc_range(Range&: *BB)) { |
987 | if (Value *V = simplifyInstruction(I: &Inst, Q: {DL, nullptr, DT, AC})) |
988 | if (LI->replacementPreservesLCSSAForm(From: &Inst, To: V)) |
989 | Inst.replaceAllUsesWith(V); |
990 | if (isInstructionTriviallyDead(I: &Inst)) |
991 | DeadInsts.emplace_back(Args: &Inst); |
992 | } |
993 | // We can't do recursive deletion until we're done iterating, as we might |
994 | // have a phi which (potentially indirectly) uses instructions later in |
995 | // the block we're iterating through. |
996 | RecursivelyDeleteTriviallyDeadInstructions(DeadInsts); |
997 | } |
998 | |
999 | // Merge latch into exit block. |
1000 | auto *ExitBB = RemainderLatch->getSingleSuccessor(); |
1001 | assert(ExitBB && "required after breaking cond br backedge" ); |
1002 | DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); |
1003 | MergeBlockIntoPredecessor(BB: ExitBB, DTU: &DTU, LI); |
1004 | } |
1005 | |
1006 | // Canonicalize to LoopSimplifyForm both original and remainder loops. We |
1007 | // cannot rely on the LoopUnrollPass to do this because it only does |
1008 | // canonicalization for parent/subloops and not the sibling loops. |
1009 | if (OtherExits.size() > 0) { |
1010 | // Generate dedicated exit blocks for the original loop, to preserve |
1011 | // LoopSimplifyForm. |
1012 | formDedicatedExitBlocks(L, DT, LI, MSSAU: nullptr, PreserveLCSSA); |
1013 | // Generate dedicated exit blocks for the remainder loop if one exists, to |
1014 | // preserve LoopSimplifyForm. |
1015 | if (remainderLoop) |
1016 | formDedicatedExitBlocks(L: remainderLoop, DT, LI, MSSAU: nullptr, PreserveLCSSA); |
1017 | } |
1018 | |
1019 | auto UnrollResult = LoopUnrollResult::Unmodified; |
1020 | if (remainderLoop && UnrollRemainder) { |
1021 | LLVM_DEBUG(dbgs() << "Unrolling remainder loop\n" ); |
1022 | UnrollLoopOptions ULO; |
1023 | ULO.Count = Count - 1; |
1024 | ULO.Force = false; |
1025 | ULO.Runtime = false; |
1026 | ULO.AllowExpensiveTripCount = false; |
1027 | ULO.UnrollRemainder = false; |
1028 | ULO.ForgetAllSCEV = ForgetAllSCEV; |
1029 | assert(!getLoopConvergenceHeart(L) && |
1030 | "A loop with a convergence heart does not allow runtime unrolling." ); |
1031 | UnrollResult = UnrollLoop(L: remainderLoop, ULO, LI, SE, DT, AC, TTI, |
1032 | /*ORE*/ nullptr, PreserveLCSSA); |
1033 | } |
1034 | |
1035 | if (ResultLoop && UnrollResult != LoopUnrollResult::FullyUnrolled) |
1036 | *ResultLoop = remainderLoop; |
1037 | NumRuntimeUnrolled++; |
1038 | return true; |
1039 | } |
1040 | |