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/ADT/Statistic.h"
25 #include "llvm/Analysis/AliasAnalysis.h"
26 #include "llvm/Analysis/LoopIterator.h"
27 #include "llvm/Analysis/LoopPass.h"
28 #include "llvm/Analysis/ScalarEvolution.h"
29 #include "llvm/Analysis/ScalarEvolutionExpander.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/Dominators.h"
32 #include "llvm/IR/Metadata.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Transforms/Scalar.h"
37 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
38 #include "llvm/Transforms/Utils/Cloning.h"
39 #include "llvm/Transforms/Utils/UnrollLoop.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();
150 DT->changeImmediateDominator(Exit, PrologExit);
153 /// Connect the unrolling epilog code to the original loop.
154 /// The unrolling epilog code contains code to execute the
155 /// 'extra' iterations if the run-time trip count modulo the
156 /// unroll count is non-zero.
158 /// This function performs the following:
159 /// - Update PHI nodes at the unrolling loop exit and epilog loop exit
160 /// - Create PHI nodes at the unrolling loop exit to combine
161 /// values that exit the unrolling loop code and jump around it.
162 /// - Update PHI operands in the epilog loop by the new PHI nodes
163 /// - Branch around the epilog loop if extra iters (ModVal) is zero.
165 static void ConnectEpilog(Loop *L, Value *ModVal, BasicBlock *NewExit,
166 BasicBlock *Exit, BasicBlock *PreHeader,
167 BasicBlock *EpilogPreHeader, BasicBlock *NewPreHeader,
168 ValueToValueMapTy &VMap, DominatorTree *DT,
169 LoopInfo *LI, bool PreserveLCSSA) {
170 BasicBlock *Latch = L->getLoopLatch();
171 assert(Latch && "Loop must have a latch");
172 BasicBlock *EpilogLatch = cast<BasicBlock>(VMap[Latch]);
174 // Loop structure should be the following:
188 // Update PHI nodes at NewExit and Exit.
189 for (Instruction &BBI : *NewExit) {
190 PHINode *PN = dyn_cast<PHINode>(&BBI);
191 // Exit when we passed all PHI nodes.
194 // PN should be used in another PHI located in Exit block as
195 // Exit was split by SplitBlockPredecessors into Exit and NewExit
196 // Basicaly it should look like:
198 // PN = PHI [I, Latch]
201 // EpilogPN = PHI [PN, EpilogPreHeader]
203 // There is EpilogPreHeader incoming block instead of NewExit as
204 // NewExit was spilt 1 more time to get EpilogPreHeader.
205 assert(PN->hasOneUse() && "The phi should have 1 use");
206 PHINode *EpilogPN = cast<PHINode> (PN->use_begin()->getUser());
207 assert(EpilogPN->getParent() == Exit && "EpilogPN should be in Exit block");
209 // Add incoming PreHeader from branch around the Loop
210 PN->addIncoming(UndefValue::get(PN->getType()), PreHeader);
212 Value *V = PN->getIncomingValueForBlock(Latch);
213 Instruction *I = dyn_cast<Instruction>(V);
214 if (I && L->contains(I))
215 // If value comes from an instruction in the loop add VMap value.
217 // For the instruction out of the loop, constant or undefined value
218 // insert value itself.
219 EpilogPN->addIncoming(V, EpilogLatch);
221 assert(EpilogPN->getBasicBlockIndex(EpilogPreHeader) >= 0 &&
222 "EpilogPN should have EpilogPreHeader incoming block");
223 // Change EpilogPreHeader incoming block to NewExit.
224 EpilogPN->setIncomingBlock(EpilogPN->getBasicBlockIndex(EpilogPreHeader),
226 // Now PHIs should look like:
228 // PN = PHI [I, Latch], [undef, PreHeader]
231 // EpilogPN = PHI [PN, NewExit], [VMap[I], EpilogLatch]
234 // Create PHI nodes at NewExit (from the unrolling loop Latch and PreHeader).
235 // Update corresponding PHI nodes in epilog loop.
236 for (BasicBlock *Succ : successors(Latch)) {
237 // Skip this as we already updated phis in exit blocks.
238 if (!L->contains(Succ))
240 for (Instruction &BBI : *Succ) {
241 PHINode *PN = dyn_cast<PHINode>(&BBI);
242 // Exit when we passed all PHI nodes.
245 // Add new PHI nodes to the loop exit block and update epilog
246 // PHIs with the new PHI values.
247 PHINode *NewPN = PHINode::Create(PN->getType(), 2, PN->getName() + ".unr",
248 NewExit->getFirstNonPHI());
249 // Adding a value to the new PHI node from the unrolling loop preheader.
250 NewPN->addIncoming(PN->getIncomingValueForBlock(NewPreHeader), PreHeader);
251 // Adding a value to the new PHI node from the unrolling loop latch.
252 NewPN->addIncoming(PN->getIncomingValueForBlock(Latch), Latch);
254 // Update the existing PHI node operand with the value from the new PHI
255 // node. Corresponding instruction in epilog loop should be PHI.
256 PHINode *VPN = cast<PHINode>(VMap[&BBI]);
257 VPN->setIncomingValue(VPN->getBasicBlockIndex(EpilogPreHeader), NewPN);
261 Instruction *InsertPt = NewExit->getTerminator();
262 IRBuilder<> B(InsertPt);
263 Value *BrLoopExit = B.CreateIsNotNull(ModVal, "lcmp.mod");
264 assert(Exit && "Loop must have a single exit block only");
265 // Split the epilogue exit to maintain loop canonicalization guarantees
266 SmallVector<BasicBlock*, 4> Preds(predecessors(Exit));
267 SplitBlockPredecessors(Exit, Preds, ".epilog-lcssa", DT, LI,
269 // Add the branch to the exit block (around the unrolling loop)
270 B.CreateCondBr(BrLoopExit, EpilogPreHeader, Exit);
271 InsertPt->eraseFromParent();
273 DT->changeImmediateDominator(Exit, NewExit);
275 // Split the main loop exit to maintain canonicalization guarantees.
276 SmallVector<BasicBlock*, 4> NewExitPreds{Latch};
277 SplitBlockPredecessors(NewExit, NewExitPreds, ".loopexit", DT, LI,
281 /// Create a clone of the blocks in a loop and connect them together.
282 /// If CreateRemainderLoop is false, loop structure will not be cloned,
283 /// otherwise a new loop will be created including all cloned blocks, and the
284 /// iterator of it switches to count NewIter down to 0.
285 /// The cloned blocks should be inserted between InsertTop and InsertBot.
286 /// If loop structure is cloned InsertTop should be new preheader, InsertBot
289 static void CloneLoopBlocks(Loop *L, Value *NewIter,
290 const bool CreateRemainderLoop,
291 const bool UseEpilogRemainder,
292 BasicBlock *InsertTop, BasicBlock *InsertBot,
293 BasicBlock *Preheader,
294 std::vector<BasicBlock *> &NewBlocks,
295 LoopBlocksDFS &LoopBlocks, ValueToValueMapTy &VMap,
296 DominatorTree *DT, LoopInfo *LI) {
297 StringRef suffix = UseEpilogRemainder ? "epil" : "prol";
298 BasicBlock *Header = L->getHeader();
299 BasicBlock *Latch = L->getLoopLatch();
300 Function *F = Header->getParent();
301 LoopBlocksDFS::RPOIterator BlockBegin = LoopBlocks.beginRPO();
302 LoopBlocksDFS::RPOIterator BlockEnd = LoopBlocks.endRPO();
303 Loop *ParentLoop = L->getParentLoop();
304 NewLoopsMap NewLoops;
305 NewLoops[ParentLoop] = ParentLoop;
306 if (!CreateRemainderLoop)
307 NewLoops[L] = ParentLoop;
309 // For each block in the original loop, create a new copy,
310 // and update the value map with the newly created values.
311 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
312 BasicBlock *NewBB = CloneBasicBlock(*BB, VMap, "." + suffix, F);
313 NewBlocks.push_back(NewBB);
315 // If we're unrolling the outermost loop, there's no remainder loop,
316 // and this block isn't in a nested loop, then the new block is not
317 // in any loop. Otherwise, add it to loopinfo.
318 if (CreateRemainderLoop || LI->getLoopFor(*BB) != L || ParentLoop)
319 addClonedBlockToLoopInfo(*BB, NewBB, LI, NewLoops);
323 // For the first block, add a CFG connection to this newly
325 InsertTop->getTerminator()->setSuccessor(0, NewBB);
330 // The header is dominated by the preheader.
331 DT->addNewBlock(NewBB, InsertTop);
333 // Copy information from original loop to unrolled loop.
334 BasicBlock *IDomBB = DT->getNode(*BB)->getIDom()->getBlock();
335 DT->addNewBlock(NewBB, cast<BasicBlock>(VMap[IDomBB]));
340 // For the last block, if CreateRemainderLoop is false, create a direct
341 // jump to InsertBot. If not, create a loop back to cloned head.
342 VMap.erase((*BB)->getTerminator());
343 BasicBlock *FirstLoopBB = cast<BasicBlock>(VMap[Header]);
344 BranchInst *LatchBR = cast<BranchInst>(NewBB->getTerminator());
345 IRBuilder<> Builder(LatchBR);
346 if (!CreateRemainderLoop) {
347 Builder.CreateBr(InsertBot);
349 PHINode *NewIdx = PHINode::Create(NewIter->getType(), 2,
351 FirstLoopBB->getFirstNonPHI());
353 Builder.CreateSub(NewIdx, ConstantInt::get(NewIdx->getType(), 1),
354 NewIdx->getName() + ".sub");
356 Builder.CreateIsNotNull(IdxSub, NewIdx->getName() + ".cmp");
357 Builder.CreateCondBr(IdxCmp, FirstLoopBB, InsertBot);
358 NewIdx->addIncoming(NewIter, InsertTop);
359 NewIdx->addIncoming(IdxSub, NewBB);
361 LatchBR->eraseFromParent();
365 // Change the incoming values to the ones defined in the preheader or
367 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
368 PHINode *NewPHI = cast<PHINode>(VMap[&*I]);
369 if (!CreateRemainderLoop) {
370 if (UseEpilogRemainder) {
371 unsigned idx = NewPHI->getBasicBlockIndex(Preheader);
372 NewPHI->setIncomingBlock(idx, InsertTop);
373 NewPHI->removeIncomingValue(Latch, false);
375 VMap[&*I] = NewPHI->getIncomingValueForBlock(Preheader);
376 cast<BasicBlock>(VMap[Header])->getInstList().erase(NewPHI);
379 unsigned idx = NewPHI->getBasicBlockIndex(Preheader);
380 NewPHI->setIncomingBlock(idx, InsertTop);
381 BasicBlock *NewLatch = cast<BasicBlock>(VMap[Latch]);
382 idx = NewPHI->getBasicBlockIndex(Latch);
383 Value *InVal = NewPHI->getIncomingValue(idx);
384 NewPHI->setIncomingBlock(idx, NewLatch);
385 if (Value *V = VMap.lookup(InVal))
386 NewPHI->setIncomingValue(idx, V);
389 if (CreateRemainderLoop) {
390 Loop *NewLoop = NewLoops[L];
391 assert(NewLoop && "L should have been cloned");
392 // Add unroll disable metadata to disable future unrolling for this loop.
393 SmallVector<Metadata *, 4> MDs;
394 // Reserve first location for self reference to the LoopID metadata node.
395 MDs.push_back(nullptr);
396 MDNode *LoopID = NewLoop->getLoopID();
398 // First remove any existing loop unrolling metadata.
399 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
400 bool IsUnrollMetadata = false;
401 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
403 const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
404 IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll.");
406 if (!IsUnrollMetadata)
407 MDs.push_back(LoopID->getOperand(i));
411 LLVMContext &Context = NewLoop->getHeader()->getContext();
412 SmallVector<Metadata *, 1> DisableOperands;
413 DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable"));
414 MDNode *DisableNode = MDNode::get(Context, DisableOperands);
415 MDs.push_back(DisableNode);
417 MDNode *NewLoopID = MDNode::get(Context, MDs);
418 // Set operand 0 to refer to the loop id itself.
419 NewLoopID->replaceOperandWith(0, NewLoopID);
420 NewLoop->setLoopID(NewLoopID);
424 /// Insert code in the prolog/epilog code when unrolling a loop with a
425 /// run-time trip-count.
427 /// This method assumes that the loop unroll factor is total number
428 /// of loop bodies in the loop after unrolling. (Some folks refer
429 /// to the unroll factor as the number of *extra* copies added).
430 /// We assume also that the loop unroll factor is a power-of-two. So, after
431 /// unrolling the loop, the number of loop bodies executed is 2,
432 /// 4, 8, etc. Note - LLVM converts the if-then-sequence to a switch
433 /// instruction in SimplifyCFG.cpp. Then, the backend decides how code for
434 /// the switch instruction is generated.
436 /// ***Prolog case***
437 /// extraiters = tripcount % loopfactor
438 /// if (extraiters == 0) jump Loop:
441 /// extraiters -= 1 // Omitted if unroll factor is 2.
442 /// if (extraiters != 0) jump Prol: // Omitted if unroll factor is 2.
443 /// if (tripcount < loopfactor) jump End:
448 /// ***Epilog case***
449 /// extraiters = tripcount % loopfactor
450 /// if (tripcount < loopfactor) jump LoopExit:
451 /// unroll_iters = tripcount - extraiters
452 /// Loop: LoopBody; (executes unroll_iter times);
454 /// if (unroll_iter != 0) jump Loop:
456 /// if (extraiters == 0) jump EpilExit:
457 /// Epil: LoopBody; (executes extraiters times)
458 /// extraiters -= 1 // Omitted if unroll factor is 2.
459 /// if (extraiters != 0) jump Epil: // Omitted if unroll factor is 2.
462 bool llvm::UnrollRuntimeLoopRemainder(Loop *L, unsigned Count,
463 bool AllowExpensiveTripCount,
464 bool UseEpilogRemainder,
465 LoopInfo *LI, ScalarEvolution *SE,
466 DominatorTree *DT, bool PreserveLCSSA) {
467 // for now, only unroll loops that contain a single exit
468 if (!L->getExitingBlock())
471 // Make sure the loop is in canonical form, and there is a single
473 if (!L->isLoopSimplifyForm())
475 BasicBlock *Exit = L->getUniqueExitBlock(); // successor out of loop
479 // Use Scalar Evolution to compute the trip count. This allows more loops to
480 // be unrolled than relying on induction var simplification.
484 // Only unroll loops with a computable trip count, and the trip count needs
485 // to be an int value (allowing a pointer type is a TODO item).
486 const SCEV *BECountSC = SE->getBackedgeTakenCount(L);
487 if (isa<SCEVCouldNotCompute>(BECountSC) ||
488 !BECountSC->getType()->isIntegerTy())
491 unsigned BEWidth = cast<IntegerType>(BECountSC->getType())->getBitWidth();
493 // Add 1 since the backedge count doesn't include the first loop iteration.
494 const SCEV *TripCountSC =
495 SE->getAddExpr(BECountSC, SE->getConstant(BECountSC->getType(), 1));
496 if (isa<SCEVCouldNotCompute>(TripCountSC))
499 BasicBlock *Header = L->getHeader();
500 BasicBlock *PreHeader = L->getLoopPreheader();
501 BranchInst *PreHeaderBR = cast<BranchInst>(PreHeader->getTerminator());
502 const DataLayout &DL = Header->getModule()->getDataLayout();
503 SCEVExpander Expander(*SE, DL, "loop-unroll");
504 if (!AllowExpensiveTripCount &&
505 Expander.isHighCostExpansion(TripCountSC, L, PreHeaderBR))
508 // This constraint lets us deal with an overflowing trip count easily; see the
509 // comment on ModVal below.
510 if (Log2_32(Count) > BEWidth)
513 BasicBlock *Latch = L->getLoopLatch();
515 // Cloning the loop basic blocks (`CloneLoopBlocks`) requires that one of the
516 // targets of the Latch be the single exit block out of the loop. This needs
517 // to be guaranteed by the callers of UnrollRuntimeLoopRemainder.
518 BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
520 (LatchBR->getSuccessor(0) == Exit || LatchBR->getSuccessor(1) == Exit) &&
521 "one of the loop latch successors should be "
523 // Avoid warning of unused `LatchBR` variable in release builds.
525 // Loop structure is the following:
533 BasicBlock *NewPreHeader;
534 BasicBlock *NewExit = nullptr;
535 BasicBlock *PrologExit = nullptr;
536 BasicBlock *EpilogPreHeader = nullptr;
537 BasicBlock *PrologPreHeader = nullptr;
539 if (UseEpilogRemainder) {
540 // If epilog remainder
541 // Split PreHeader to insert a branch around loop for unrolling.
542 NewPreHeader = SplitBlock(PreHeader, PreHeader->getTerminator(), DT, LI);
543 NewPreHeader->setName(PreHeader->getName() + ".new");
544 // Split Exit to create phi nodes from branch above.
545 SmallVector<BasicBlock*, 4> Preds(predecessors(Exit));
546 NewExit = SplitBlockPredecessors(Exit, Preds, ".unr-lcssa",
547 DT, LI, PreserveLCSSA);
548 // Split NewExit to insert epilog remainder loop.
549 EpilogPreHeader = SplitBlock(NewExit, NewExit->getTerminator(), DT, LI);
550 EpilogPreHeader->setName(Header->getName() + ".epil.preheader");
552 // If prolog remainder
553 // Split the original preheader twice to insert prolog remainder loop
554 PrologPreHeader = SplitEdge(PreHeader, Header, DT, LI);
555 PrologPreHeader->setName(Header->getName() + ".prol.preheader");
556 PrologExit = SplitBlock(PrologPreHeader, PrologPreHeader->getTerminator(),
558 PrologExit->setName(Header->getName() + ".prol.loopexit");
559 // Split PrologExit to get NewPreHeader.
560 NewPreHeader = SplitBlock(PrologExit, PrologExit->getTerminator(), DT, LI);
561 NewPreHeader->setName(PreHeader->getName() + ".new");
563 // Loop structure should be the following:
566 // PreHeader PreHeader
567 // *NewPreHeader *PrologPreHeader
568 // Header *PrologExit
572 // *EpilogPreHeader Latch
575 // Calculate conditions for branch around loop for unrolling
576 // in epilog case and around prolog remainder loop in prolog case.
577 // Compute the number of extra iterations required, which is:
578 // extra iterations = run-time trip count % loop unroll factor
579 PreHeaderBR = cast<BranchInst>(PreHeader->getTerminator());
580 Value *TripCount = Expander.expandCodeFor(TripCountSC, TripCountSC->getType(),
582 Value *BECount = Expander.expandCodeFor(BECountSC, BECountSC->getType(),
584 IRBuilder<> B(PreHeaderBR);
586 // Calculate ModVal = (BECount + 1) % Count.
587 // Note that TripCount is BECount + 1.
588 if (isPowerOf2_32(Count)) {
589 // When Count is power of 2 we don't BECount for epilog case, however we'll
590 // need it for a branch around unrolling loop for prolog case.
591 ModVal = B.CreateAnd(TripCount, Count - 1, "xtraiter");
592 // 1. There are no iterations to be run in the prolog/epilog loop.
594 // 2. The addition computing TripCount overflowed.
596 // If (2) is true, we know that TripCount really is (1 << BEWidth) and so
597 // the number of iterations that remain to be run in the original loop is a
598 // multiple Count == (1 << Log2(Count)) because Log2(Count) <= BEWidth (we
599 // explicitly check this above).
601 // As (BECount + 1) can potentially unsigned overflow we count
602 // (BECount % Count) + 1 which is overflow safe as BECount % Count < Count.
603 Value *ModValTmp = B.CreateURem(BECount,
604 ConstantInt::get(BECount->getType(),
606 Value *ModValAdd = B.CreateAdd(ModValTmp,
607 ConstantInt::get(ModValTmp->getType(), 1));
608 // At that point (BECount % Count) + 1 could be equal to Count.
609 // To handle this case we need to take mod by Count one more time.
610 ModVal = B.CreateURem(ModValAdd,
611 ConstantInt::get(BECount->getType(), Count),
615 UseEpilogRemainder ? B.CreateICmpULT(BECount,
616 ConstantInt::get(BECount->getType(),
618 B.CreateIsNotNull(ModVal, "lcmp.mod");
619 BasicBlock *RemainderLoop = UseEpilogRemainder ? NewExit : PrologPreHeader;
620 BasicBlock *UnrollingLoop = UseEpilogRemainder ? NewPreHeader : PrologExit;
621 // Branch to either remainder (extra iterations) loop or unrolling loop.
622 B.CreateCondBr(BranchVal, RemainderLoop, UnrollingLoop);
623 PreHeaderBR->eraseFromParent();
625 if (UseEpilogRemainder)
626 DT->changeImmediateDominator(NewExit, PreHeader);
628 DT->changeImmediateDominator(PrologExit, PreHeader);
630 Function *F = Header->getParent();
631 // Get an ordered list of blocks in the loop to help with the ordering of the
632 // cloned blocks in the prolog/epilog code
633 LoopBlocksDFS LoopBlocks(L);
634 LoopBlocks.perform(LI);
637 // For each extra loop iteration, create a copy of the loop's basic blocks
638 // and generate a condition that branches to the copy depending on the
639 // number of 'left over' iterations.
641 std::vector<BasicBlock *> NewBlocks;
642 ValueToValueMapTy VMap;
644 // For unroll factor 2 remainder loop will have 1 iterations.
645 // Do not create 1 iteration loop.
646 bool CreateRemainderLoop = (Count != 2);
648 // Clone all the basic blocks in the loop. If Count is 2, we don't clone
649 // the loop, otherwise we create a cloned loop to execute the extra
650 // iterations. This function adds the appropriate CFG connections.
651 BasicBlock *InsertBot = UseEpilogRemainder ? Exit : PrologExit;
652 BasicBlock *InsertTop = UseEpilogRemainder ? EpilogPreHeader : PrologPreHeader;
653 CloneLoopBlocks(L, ModVal, CreateRemainderLoop, UseEpilogRemainder, InsertTop,
654 InsertBot, NewPreHeader, NewBlocks, LoopBlocks, VMap, DT, LI);
656 // Insert the cloned blocks into the function.
657 F->getBasicBlockList().splice(InsertBot->getIterator(),
658 F->getBasicBlockList(),
659 NewBlocks[0]->getIterator(),
662 // Loop structure should be the following:
665 // PreHeader PreHeader
666 // NewPreHeader PrologPreHeader
667 // Header PrologHeader
670 // NewExit PrologExit
671 // EpilogPreHeader NewPreHeader
672 // EpilogHeader Header
677 // Rewrite the cloned instruction operands to use the values created when the
679 for (BasicBlock *BB : NewBlocks) {
680 for (Instruction &I : *BB) {
681 RemapInstruction(&I, VMap,
682 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
686 if (UseEpilogRemainder) {
687 // Connect the epilog code to the original loop and update the
689 ConnectEpilog(L, ModVal, NewExit, Exit, PreHeader,
690 EpilogPreHeader, NewPreHeader, VMap, DT, LI,
693 // Update counter in loop for unrolling.
694 // I should be multiply of Count.
695 IRBuilder<> B2(NewPreHeader->getTerminator());
696 Value *TestVal = B2.CreateSub(TripCount, ModVal, "unroll_iter");
697 BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
698 B2.SetInsertPoint(LatchBR);
699 PHINode *NewIdx = PHINode::Create(TestVal->getType(), 2, "niter",
700 Header->getFirstNonPHI());
702 B2.CreateSub(NewIdx, ConstantInt::get(NewIdx->getType(), 1),
703 NewIdx->getName() + ".nsub");
705 if (LatchBR->getSuccessor(0) == Header)
706 IdxCmp = B2.CreateIsNotNull(IdxSub, NewIdx->getName() + ".ncmp");
708 IdxCmp = B2.CreateIsNull(IdxSub, NewIdx->getName() + ".ncmp");
709 NewIdx->addIncoming(TestVal, NewPreHeader);
710 NewIdx->addIncoming(IdxSub, Latch);
711 LatchBR->setCondition(IdxCmp);
713 // Connect the prolog code to the original loop and update the
715 ConnectProlog(L, BECount, Count, PrologExit, PreHeader, NewPreHeader,
716 VMap, DT, LI, PreserveLCSSA);
719 // If this loop is nested, then the loop unroller changes the code in the
720 // parent loop, so the Scalar Evolution pass needs to be run again.
721 if (Loop *ParentLoop = L->getParentLoop())
722 SE->forgetLoop(ParentLoop);
724 NumRuntimeUnrolled++;