1 //===-- UnrollLoopRuntime.cpp - Runtime Loop unrolling utilities ----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements some loop unrolling utilities for loops with run-time
11 // trip counts. See LoopUnroll.cpp for unrolling loops with compile-time
14 // The functions in this file are used to generate extra code when the
15 // run-time trip count modulo the unroll factor is not 0. When this is the
16 // case, we need to generate code to execute these 'left over' iterations.
18 // The current strategy generates an if-then-else sequence prior to the
19 // unrolled loop to execute the 'left over' iterations before or after the
22 //===----------------------------------------------------------------------===//
24 #include "llvm/Transforms/Utils/UnrollLoop.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Analysis/LoopIterator.h"
28 #include "llvm/Analysis/LoopPass.h"
29 #include "llvm/Analysis/ScalarEvolution.h"
30 #include "llvm/Analysis/ScalarEvolutionExpander.h"
31 #include "llvm/IR/BasicBlock.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/Metadata.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
39 #include "llvm/Transforms/Utils/Cloning.h"
44 #define DEBUG_TYPE "loop-unroll"
46 STATISTIC(NumRuntimeUnrolled,
47 "Number of loops unrolled with run-time trip counts");
49 /// Connect the unrolling prolog code to the original loop.
50 /// The unrolling prolog code contains code to execute the
51 /// 'extra' iterations if the run-time trip count modulo the
52 /// unroll count is non-zero.
54 /// This function performs the following:
55 /// - Create PHI nodes at prolog end block to combine values
56 /// that exit the prolog code and jump around the prolog.
57 /// - Add a PHI operand to a PHI node at the loop exit block
58 /// for values that exit the prolog and go around the loop.
59 /// - Branch around the original loop if the trip count is less
60 /// than the unroll factor.
62 static void ConnectProlog(Loop *L, Value *BECount, unsigned Count,
63 BasicBlock *PrologExit, BasicBlock *PreHeader,
64 BasicBlock *NewPreHeader, ValueToValueMapTy &VMap,
65 DominatorTree *DT, LoopInfo *LI, bool PreserveLCSSA) {
66 BasicBlock *Latch = L->getLoopLatch();
67 assert(Latch && "Loop must have a latch");
68 BasicBlock *PrologLatch = cast<BasicBlock>(VMap[Latch]);
70 // Create a PHI node for each outgoing value from the original loop
71 // (which means it is an outgoing value from the prolog code too).
72 // The new PHI node is inserted in the prolog end basic block.
73 // The new PHI node value is added as an operand of a PHI node in either
74 // the loop header or the loop exit block.
75 for (BasicBlock *Succ : successors(Latch)) {
76 for (Instruction &BBI : *Succ) {
77 PHINode *PN = dyn_cast<PHINode>(&BBI);
78 // Exit when we passed all PHI nodes.
81 // Add a new PHI node to the prolog end block and add the
82 // appropriate incoming values.
83 PHINode *NewPN = PHINode::Create(PN->getType(), 2, PN->getName() + ".unr",
84 PrologExit->getFirstNonPHI());
85 // Adding a value to the new PHI node from the original loop preheader.
86 // This is the value that skips all the prolog code.
87 if (L->contains(PN)) {
88 NewPN->addIncoming(PN->getIncomingValueForBlock(NewPreHeader),
91 NewPN->addIncoming(UndefValue::get(PN->getType()), PreHeader);
94 Value *V = PN->getIncomingValueForBlock(Latch);
95 if (Instruction *I = dyn_cast<Instruction>(V)) {
100 // Adding a value to the new PHI node from the last prolog block
102 NewPN->addIncoming(V, PrologLatch);
104 // Update the existing PHI node operand with the value from the
105 // new PHI node. How this is done depends on if the existing
106 // PHI node is in the original loop block, or the exit block.
107 if (L->contains(PN)) {
108 PN->setIncomingValue(PN->getBasicBlockIndex(NewPreHeader), NewPN);
110 PN->addIncoming(NewPN, PrologExit);
115 // Make sure that created prolog loop is in simplified form
116 SmallVector<BasicBlock *, 4> PrologExitPreds;
117 Loop *PrologLoop = LI->getLoopFor(PrologLatch);
119 for (BasicBlock *PredBB : predecessors(PrologExit))
120 if (PrologLoop->contains(PredBB))
121 PrologExitPreds.push_back(PredBB);
123 SplitBlockPredecessors(PrologExit, PrologExitPreds, ".unr-lcssa", DT, LI,
127 // Create a branch around the original loop, which is taken if there are no
128 // iterations remaining to be executed after running the prologue.
129 Instruction *InsertPt = PrologExit->getTerminator();
130 IRBuilder<> B(InsertPt);
132 assert(Count != 0 && "nonsensical Count!");
134 // If BECount <u (Count - 1) then (BECount + 1) % Count == (BECount + 1)
135 // This means %xtraiter is (BECount + 1) and all of the iterations of this
136 // loop were executed by the prologue. Note that if BECount <u (Count - 1)
137 // then (BECount + 1) cannot unsigned-overflow.
139 B.CreateICmpULT(BECount, ConstantInt::get(BECount->getType(), Count - 1));
140 BasicBlock *Exit = L->getUniqueExitBlock();
141 assert(Exit && "Loop must have a single exit block only");
142 // Split the exit to maintain loop canonicalization guarantees
143 SmallVector<BasicBlock*, 4> Preds(predecessors(Exit));
144 SplitBlockPredecessors(Exit, Preds, ".unr-lcssa", DT, LI,
146 // Add the branch to the exit block (around the unrolled loop)
147 B.CreateCondBr(BrLoopExit, Exit, NewPreHeader);
148 InsertPt->eraseFromParent();
151 /// Connect the unrolling epilog code to the original loop.
152 /// The unrolling epilog code contains code to execute the
153 /// 'extra' iterations if the run-time trip count modulo the
154 /// unroll count is non-zero.
156 /// This function performs the following:
157 /// - Update PHI nodes at the unrolling loop exit and epilog loop exit
158 /// - Create PHI nodes at the unrolling loop exit to combine
159 /// values that exit the unrolling loop code and jump around it.
160 /// - Update PHI operands in the epilog loop by the new PHI nodes
161 /// - Branch around the epilog loop if extra iters (ModVal) is zero.
163 static void ConnectEpilog(Loop *L, Value *ModVal, BasicBlock *NewExit,
164 BasicBlock *Exit, BasicBlock *PreHeader,
165 BasicBlock *EpilogPreHeader, BasicBlock *NewPreHeader,
166 ValueToValueMapTy &VMap, DominatorTree *DT,
167 LoopInfo *LI, bool PreserveLCSSA) {
168 BasicBlock *Latch = L->getLoopLatch();
169 assert(Latch && "Loop must have a latch");
170 BasicBlock *EpilogLatch = cast<BasicBlock>(VMap[Latch]);
172 // Loop structure should be the following:
186 // Update PHI nodes at NewExit and Exit.
187 for (Instruction &BBI : *NewExit) {
188 PHINode *PN = dyn_cast<PHINode>(&BBI);
189 // Exit when we passed all PHI nodes.
192 // PN should be used in another PHI located in Exit block as
193 // Exit was split by SplitBlockPredecessors into Exit and NewExit
194 // Basicaly it should look like:
196 // PN = PHI [I, Latch]
199 // EpilogPN = PHI [PN, EpilogPreHeader]
201 // There is EpilogPreHeader incoming block instead of NewExit as
202 // NewExit was spilt 1 more time to get EpilogPreHeader.
203 assert(PN->hasOneUse() && "The phi should have 1 use");
204 PHINode *EpilogPN = cast<PHINode> (PN->use_begin()->getUser());
205 assert(EpilogPN->getParent() == Exit && "EpilogPN should be in Exit block");
207 // Add incoming PreHeader from branch around the Loop
208 PN->addIncoming(UndefValue::get(PN->getType()), PreHeader);
210 Value *V = PN->getIncomingValueForBlock(Latch);
211 Instruction *I = dyn_cast<Instruction>(V);
212 if (I && L->contains(I))
213 // If value comes from an instruction in the loop add VMap value.
215 // For the instruction out of the loop, constant or undefined value
216 // insert value itself.
217 EpilogPN->addIncoming(V, EpilogLatch);
219 assert(EpilogPN->getBasicBlockIndex(EpilogPreHeader) >= 0 &&
220 "EpilogPN should have EpilogPreHeader incoming block");
221 // Change EpilogPreHeader incoming block to NewExit.
222 EpilogPN->setIncomingBlock(EpilogPN->getBasicBlockIndex(EpilogPreHeader),
224 // Now PHIs should look like:
226 // PN = PHI [I, Latch], [undef, PreHeader]
229 // EpilogPN = PHI [PN, NewExit], [VMap[I], EpilogLatch]
232 // Create PHI nodes at NewExit (from the unrolling loop Latch and PreHeader).
233 // Update corresponding PHI nodes in epilog loop.
234 for (BasicBlock *Succ : successors(Latch)) {
235 // Skip this as we already updated phis in exit blocks.
236 if (!L->contains(Succ))
238 for (Instruction &BBI : *Succ) {
239 PHINode *PN = dyn_cast<PHINode>(&BBI);
240 // Exit when we passed all PHI nodes.
243 // Add new PHI nodes to the loop exit block and update epilog
244 // PHIs with the new PHI values.
245 PHINode *NewPN = PHINode::Create(PN->getType(), 2, PN->getName() + ".unr",
246 NewExit->getFirstNonPHI());
247 // Adding a value to the new PHI node from the unrolling loop preheader.
248 NewPN->addIncoming(PN->getIncomingValueForBlock(NewPreHeader), PreHeader);
249 // Adding a value to the new PHI node from the unrolling loop latch.
250 NewPN->addIncoming(PN->getIncomingValueForBlock(Latch), Latch);
252 // Update the existing PHI node operand with the value from the new PHI
253 // node. Corresponding instruction in epilog loop should be PHI.
254 PHINode *VPN = cast<PHINode>(VMap[&BBI]);
255 VPN->setIncomingValue(VPN->getBasicBlockIndex(EpilogPreHeader), NewPN);
259 Instruction *InsertPt = NewExit->getTerminator();
260 IRBuilder<> B(InsertPt);
261 Value *BrLoopExit = B.CreateIsNotNull(ModVal, "lcmp.mod");
262 assert(Exit && "Loop must have a single exit block only");
263 // Split the exit to maintain loop canonicalization guarantees
264 SmallVector<BasicBlock*, 4> Preds(predecessors(Exit));
265 SplitBlockPredecessors(Exit, Preds, ".epilog-lcssa", DT, LI,
267 // Add the branch to the exit block (around the unrolling loop)
268 B.CreateCondBr(BrLoopExit, EpilogPreHeader, Exit);
269 InsertPt->eraseFromParent();
272 /// Create a clone of the blocks in a loop and connect them together.
273 /// If CreateRemainderLoop is false, loop structure will not be cloned,
274 /// otherwise a new loop will be created including all cloned blocks, and the
275 /// iterator of it switches to count NewIter down to 0.
276 /// The cloned blocks should be inserted between InsertTop and InsertBot.
277 /// If loop structure is cloned InsertTop should be new preheader, InsertBot
280 static void CloneLoopBlocks(Loop *L, Value *NewIter,
281 const bool CreateRemainderLoop,
282 const bool UseEpilogRemainder,
283 BasicBlock *InsertTop, BasicBlock *InsertBot,
284 BasicBlock *Preheader,
285 std::vector<BasicBlock *> &NewBlocks,
286 LoopBlocksDFS &LoopBlocks, ValueToValueMapTy &VMap,
288 StringRef suffix = UseEpilogRemainder ? "epil" : "prol";
289 BasicBlock *Header = L->getHeader();
290 BasicBlock *Latch = L->getLoopLatch();
291 Function *F = Header->getParent();
292 LoopBlocksDFS::RPOIterator BlockBegin = LoopBlocks.beginRPO();
293 LoopBlocksDFS::RPOIterator BlockEnd = LoopBlocks.endRPO();
294 Loop *NewLoop = nullptr;
295 Loop *ParentLoop = L->getParentLoop();
296 if (CreateRemainderLoop) {
297 NewLoop = new Loop();
299 ParentLoop->addChildLoop(NewLoop);
301 LI->addTopLevelLoop(NewLoop);
304 NewLoopsMap NewLoops;
305 NewLoops[L] = NewLoop;
306 // For each block in the original loop, create a new copy,
307 // and update the value map with the newly created values.
308 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
309 BasicBlock *NewBB = CloneBasicBlock(*BB, VMap, "." + suffix, F);
310 NewBlocks.push_back(NewBB);
313 addClonedBlockToLoopInfo(*BB, NewBB, LI, NewLoops);
314 } else if (ParentLoop)
315 ParentLoop->addBasicBlockToLoop(NewBB, *LI);
319 // For the first block, add a CFG connection to this newly
321 InsertTop->getTerminator()->setSuccessor(0, NewBB);
325 // For the last block, if CreateRemainderLoop is false, create a direct
326 // jump to InsertBot. If not, create a loop back to cloned head.
327 VMap.erase((*BB)->getTerminator());
328 BasicBlock *FirstLoopBB = cast<BasicBlock>(VMap[Header]);
329 BranchInst *LatchBR = cast<BranchInst>(NewBB->getTerminator());
330 IRBuilder<> Builder(LatchBR);
331 if (!CreateRemainderLoop) {
332 Builder.CreateBr(InsertBot);
334 PHINode *NewIdx = PHINode::Create(NewIter->getType(), 2,
336 FirstLoopBB->getFirstNonPHI());
338 Builder.CreateSub(NewIdx, ConstantInt::get(NewIdx->getType(), 1),
339 NewIdx->getName() + ".sub");
341 Builder.CreateIsNotNull(IdxSub, NewIdx->getName() + ".cmp");
342 Builder.CreateCondBr(IdxCmp, FirstLoopBB, InsertBot);
343 NewIdx->addIncoming(NewIter, InsertTop);
344 NewIdx->addIncoming(IdxSub, NewBB);
346 LatchBR->eraseFromParent();
350 // Change the incoming values to the ones defined in the preheader or
352 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
353 PHINode *NewPHI = cast<PHINode>(VMap[&*I]);
354 if (!CreateRemainderLoop) {
355 if (UseEpilogRemainder) {
356 unsigned idx = NewPHI->getBasicBlockIndex(Preheader);
357 NewPHI->setIncomingBlock(idx, InsertTop);
358 NewPHI->removeIncomingValue(Latch, false);
360 VMap[&*I] = NewPHI->getIncomingValueForBlock(Preheader);
361 cast<BasicBlock>(VMap[Header])->getInstList().erase(NewPHI);
364 unsigned idx = NewPHI->getBasicBlockIndex(Preheader);
365 NewPHI->setIncomingBlock(idx, InsertTop);
366 BasicBlock *NewLatch = cast<BasicBlock>(VMap[Latch]);
367 idx = NewPHI->getBasicBlockIndex(Latch);
368 Value *InVal = NewPHI->getIncomingValue(idx);
369 NewPHI->setIncomingBlock(idx, NewLatch);
370 if (Value *V = VMap.lookup(InVal))
371 NewPHI->setIncomingValue(idx, V);
375 // Add unroll disable metadata to disable future unrolling for this loop.
376 SmallVector<Metadata *, 4> MDs;
377 // Reserve first location for self reference to the LoopID metadata node.
378 MDs.push_back(nullptr);
379 MDNode *LoopID = NewLoop->getLoopID();
381 // First remove any existing loop unrolling metadata.
382 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
383 bool IsUnrollMetadata = false;
384 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
386 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
387 IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll.");
389 if (!IsUnrollMetadata)
390 MDs.push_back(LoopID->getOperand(i));
394 LLVMContext &Context = NewLoop->getHeader()->getContext();
395 SmallVector<Metadata *, 1> DisableOperands;
396 DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable"));
397 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
398 MDs.push_back(DisableNode);
400 MDNode *NewLoopID = MDNode::get(Context, MDs);
401 // Set operand 0 to refer to the loop id itself.
402 NewLoopID->replaceOperandWith(0, NewLoopID);
403 NewLoop->setLoopID(NewLoopID);
407 /// Insert code in the prolog/epilog code when unrolling a loop with a
408 /// run-time trip-count.
410 /// This method assumes that the loop unroll factor is total number
411 /// of loop bodies in the loop after unrolling. (Some folks refer
412 /// to the unroll factor as the number of *extra* copies added).
413 /// We assume also that the loop unroll factor is a power-of-two. So, after
414 /// unrolling the loop, the number of loop bodies executed is 2,
415 /// 4, 8, etc. Note - LLVM converts the if-then-sequence to a switch
416 /// instruction in SimplifyCFG.cpp. Then, the backend decides how code for
417 /// the switch instruction is generated.
419 /// ***Prolog case***
420 /// extraiters = tripcount % loopfactor
421 /// if (extraiters == 0) jump Loop:
424 /// extraiters -= 1 // Omitted if unroll factor is 2.
425 /// if (extraiters != 0) jump Prol: // Omitted if unroll factor is 2.
426 /// if (tripcount < loopfactor) jump End:
431 /// ***Epilog case***
432 /// extraiters = tripcount % loopfactor
433 /// if (tripcount < loopfactor) jump LoopExit:
434 /// unroll_iters = tripcount - extraiters
435 /// Loop: LoopBody; (executes unroll_iter times);
437 /// if (unroll_iter != 0) jump Loop:
439 /// if (extraiters == 0) jump EpilExit:
440 /// Epil: LoopBody; (executes extraiters times)
441 /// extraiters -= 1 // Omitted if unroll factor is 2.
442 /// if (extraiters != 0) jump Epil: // Omitted if unroll factor is 2.
445 bool llvm::UnrollRuntimeLoopRemainder(Loop *L, unsigned Count,
446 bool AllowExpensiveTripCount,
447 bool UseEpilogRemainder,
448 LoopInfo *LI, ScalarEvolution *SE,
449 DominatorTree *DT, bool PreserveLCSSA) {
450 // for now, only unroll loops that contain a single exit
451 if (!L->getExitingBlock())
454 // Make sure the loop is in canonical form, and there is a single
456 if (!L->isLoopSimplifyForm())
458 BasicBlock *Exit = L->getUniqueExitBlock(); // successor out of loop
462 // Use Scalar Evolution to compute the trip count. This allows more loops to
463 // be unrolled than relying on induction var simplification.
467 // Only unroll loops with a computable trip count, and the trip count needs
468 // to be an int value (allowing a pointer type is a TODO item).
469 const SCEV *BECountSC = SE->getBackedgeTakenCount(L);
470 if (isa<SCEVCouldNotCompute>(BECountSC) ||
471 !BECountSC->getType()->isIntegerTy())
474 unsigned BEWidth = cast<IntegerType>(BECountSC->getType())->getBitWidth();
476 // Add 1 since the backedge count doesn't include the first loop iteration.
477 const SCEV *TripCountSC =
478 SE->getAddExpr(BECountSC, SE->getConstant(BECountSC->getType(), 1));
479 if (isa<SCEVCouldNotCompute>(TripCountSC))
482 BasicBlock *Header = L->getHeader();
483 BasicBlock *PreHeader = L->getLoopPreheader();
484 BranchInst *PreHeaderBR = cast<BranchInst>(PreHeader->getTerminator());
485 const DataLayout &DL = Header->getModule()->getDataLayout();
486 SCEVExpander Expander(*SE, DL, "loop-unroll");
487 if (!AllowExpensiveTripCount &&
488 Expander.isHighCostExpansion(TripCountSC, L, PreHeaderBR))
491 // This constraint lets us deal with an overflowing trip count easily; see the
492 // comment on ModVal below.
493 if (Log2_32(Count) > BEWidth)
496 BasicBlock *Latch = L->getLoopLatch();
498 // Loop structure is the following:
506 BasicBlock *NewPreHeader;
507 BasicBlock *NewExit = nullptr;
508 BasicBlock *PrologExit = nullptr;
509 BasicBlock *EpilogPreHeader = nullptr;
510 BasicBlock *PrologPreHeader = nullptr;
512 if (UseEpilogRemainder) {
513 // If epilog remainder
514 // Split PreHeader to insert a branch around loop for unrolling.
515 NewPreHeader = SplitBlock(PreHeader, PreHeader->getTerminator(), DT, LI);
516 NewPreHeader->setName(PreHeader->getName() + ".new");
517 // Split Exit to create phi nodes from branch above.
518 SmallVector<BasicBlock*, 4> Preds(predecessors(Exit));
519 NewExit = SplitBlockPredecessors(Exit, Preds, ".unr-lcssa",
520 DT, LI, PreserveLCSSA);
521 // Split NewExit to insert epilog remainder loop.
522 EpilogPreHeader = SplitBlock(NewExit, NewExit->getTerminator(), DT, LI);
523 EpilogPreHeader->setName(Header->getName() + ".epil.preheader");
525 // If prolog remainder
526 // Split the original preheader twice to insert prolog remainder loop
527 PrologPreHeader = SplitEdge(PreHeader, Header, DT, LI);
528 PrologPreHeader->setName(Header->getName() + ".prol.preheader");
529 PrologExit = SplitBlock(PrologPreHeader, PrologPreHeader->getTerminator(),
531 PrologExit->setName(Header->getName() + ".prol.loopexit");
532 // Split PrologExit to get NewPreHeader.
533 NewPreHeader = SplitBlock(PrologExit, PrologExit->getTerminator(), DT, LI);
534 NewPreHeader->setName(PreHeader->getName() + ".new");
536 // Loop structure should be the following:
539 // PreHeader PreHeader
540 // *NewPreHeader *PrologPreHeader
541 // Header *PrologExit
545 // *EpilogPreHeader Latch
548 // Calculate conditions for branch around loop for unrolling
549 // in epilog case and around prolog remainder loop in prolog case.
550 // Compute the number of extra iterations required, which is:
551 // extra iterations = run-time trip count % loop unroll factor
552 PreHeaderBR = cast<BranchInst>(PreHeader->getTerminator());
553 Value *TripCount = Expander.expandCodeFor(TripCountSC, TripCountSC->getType(),
555 Value *BECount = Expander.expandCodeFor(BECountSC, BECountSC->getType(),
557 IRBuilder<> B(PreHeaderBR);
559 // Calculate ModVal = (BECount + 1) % Count.
560 // Note that TripCount is BECount + 1.
561 if (isPowerOf2_32(Count)) {
562 // When Count is power of 2 we don't BECount for epilog case, however we'll
563 // need it for a branch around unrolling loop for prolog case.
564 ModVal = B.CreateAnd(TripCount, Count - 1, "xtraiter");
565 // 1. There are no iterations to be run in the prolog/epilog loop.
567 // 2. The addition computing TripCount overflowed.
569 // If (2) is true, we know that TripCount really is (1 << BEWidth) and so
570 // the number of iterations that remain to be run in the original loop is a
571 // multiple Count == (1 << Log2(Count)) because Log2(Count) <= BEWidth (we
572 // explicitly check this above).
574 // As (BECount + 1) can potentially unsigned overflow we count
575 // (BECount % Count) + 1 which is overflow safe as BECount % Count < Count.
576 Value *ModValTmp = B.CreateURem(BECount,
577 ConstantInt::get(BECount->getType(),
579 Value *ModValAdd = B.CreateAdd(ModValTmp,
580 ConstantInt::get(ModValTmp->getType(), 1));
581 // At that point (BECount % Count) + 1 could be equal to Count.
582 // To handle this case we need to take mod by Count one more time.
583 ModVal = B.CreateURem(ModValAdd,
584 ConstantInt::get(BECount->getType(), Count),
588 UseEpilogRemainder ? B.CreateICmpULT(BECount,
589 ConstantInt::get(BECount->getType(),
591 B.CreateIsNotNull(ModVal, "lcmp.mod");
592 BasicBlock *RemainderLoop = UseEpilogRemainder ? NewExit : PrologPreHeader;
593 BasicBlock *UnrollingLoop = UseEpilogRemainder ? NewPreHeader : PrologExit;
594 // Branch to either remainder (extra iterations) loop or unrolling loop.
595 B.CreateCondBr(BranchVal, RemainderLoop, UnrollingLoop);
596 PreHeaderBR->eraseFromParent();
597 Function *F = Header->getParent();
598 // Get an ordered list of blocks in the loop to help with the ordering of the
599 // cloned blocks in the prolog/epilog code
600 LoopBlocksDFS LoopBlocks(L);
601 LoopBlocks.perform(LI);
604 // For each extra loop iteration, create a copy of the loop's basic blocks
605 // and generate a condition that branches to the copy depending on the
606 // number of 'left over' iterations.
608 std::vector<BasicBlock *> NewBlocks;
609 ValueToValueMapTy VMap;
611 // For unroll factor 2 remainder loop will have 1 iterations.
612 // Do not create 1 iteration loop.
613 bool CreateRemainderLoop = (Count != 2);
615 // Clone all the basic blocks in the loop. If Count is 2, we don't clone
616 // the loop, otherwise we create a cloned loop to execute the extra
617 // iterations. This function adds the appropriate CFG connections.
618 BasicBlock *InsertBot = UseEpilogRemainder ? Exit : PrologExit;
619 BasicBlock *InsertTop = UseEpilogRemainder ? EpilogPreHeader : PrologPreHeader;
620 CloneLoopBlocks(L, ModVal, CreateRemainderLoop, UseEpilogRemainder, InsertTop,
621 InsertBot, NewPreHeader, NewBlocks, LoopBlocks, VMap, LI);
623 // Insert the cloned blocks into the function.
624 F->getBasicBlockList().splice(InsertBot->getIterator(),
625 F->getBasicBlockList(),
626 NewBlocks[0]->getIterator(),
629 // Loop structure should be the following:
632 // PreHeader PreHeader
633 // NewPreHeader PrologPreHeader
634 // Header PrologHeader
637 // NewExit PrologExit
638 // EpilogPreHeader NewPreHeader
639 // EpilogHeader Header
644 // Rewrite the cloned instruction operands to use the values created when the
646 for (BasicBlock *BB : NewBlocks) {
647 for (Instruction &I : *BB) {
648 RemapInstruction(&I, VMap,
649 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
653 if (UseEpilogRemainder) {
654 // Connect the epilog code to the original loop and update the
656 ConnectEpilog(L, ModVal, NewExit, Exit, PreHeader,
657 EpilogPreHeader, NewPreHeader, VMap, DT, LI,
660 // Update counter in loop for unrolling.
661 // I should be multiply of Count.
662 IRBuilder<> B2(NewPreHeader->getTerminator());
663 Value *TestVal = B2.CreateSub(TripCount, ModVal, "unroll_iter");
664 BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
665 B2.SetInsertPoint(LatchBR);
666 PHINode *NewIdx = PHINode::Create(TestVal->getType(), 2, "niter",
667 Header->getFirstNonPHI());
669 B2.CreateSub(NewIdx, ConstantInt::get(NewIdx->getType(), 1),
670 NewIdx->getName() + ".nsub");
672 if (LatchBR->getSuccessor(0) == Header)
673 IdxCmp = B2.CreateIsNotNull(IdxSub, NewIdx->getName() + ".ncmp");
675 IdxCmp = B2.CreateIsNull(IdxSub, NewIdx->getName() + ".ncmp");
676 NewIdx->addIncoming(TestVal, NewPreHeader);
677 NewIdx->addIncoming(IdxSub, Latch);
678 LatchBR->setCondition(IdxCmp);
680 // Connect the prolog code to the original loop and update the
682 ConnectProlog(L, BECount, Count, PrologExit, PreHeader, NewPreHeader,
683 VMap, DT, LI, PreserveLCSSA);
686 // If this loop is nested, then the loop unroller changes the code in the
687 // parent loop, so the Scalar Evolution pass needs to be run again.
688 if (Loop *ParentLoop = L->getParentLoop())
689 SE->forgetLoop(ParentLoop);
691 NumRuntimeUnrolled++;