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