1
//===- LoopRotation.cpp - Loop Rotation Pass ------------------------------===//
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 Loop Rotation Pass.
12
//===----------------------------------------------------------------------===//
14
#define DEBUG_TYPE "loop-rotate"
15
#include "llvm/Transforms/Scalar.h"
16
#include "llvm/Function.h"
17
#include "llvm/IntrinsicInst.h"
18
#include "llvm/Analysis/LoopPass.h"
19
#include "llvm/Analysis/Dominators.h"
20
#include "llvm/Analysis/ScalarEvolution.h"
21
#include "llvm/Transforms/Utils/Local.h"
22
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
23
#include "llvm/Transforms/Utils/SSAUpdater.h"
24
#include "llvm/Support/CommandLine.h"
25
#include "llvm/Support/Debug.h"
26
#include "llvm/ADT/Statistic.h"
27
#include "llvm/ADT/SmallVector.h"
30
#define MAX_HEADER_SIZE 16
32
STATISTIC(NumRotated, "Number of loops rotated");
35
class LoopRotate : public LoopPass {
37
static char ID; // Pass ID, replacement for typeid
38
LoopRotate() : LoopPass(ID) {}
40
// Rotate Loop L as many times as possible. Return true if
41
// loop is rotated at least once.
42
bool runOnLoop(Loop *L, LPPassManager &LPM);
44
// LCSSA form makes instruction renaming easier.
45
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
46
AU.addPreserved<DominatorTree>();
47
AU.addPreserved<DominanceFrontier>();
48
AU.addRequired<LoopInfo>();
49
AU.addPreserved<LoopInfo>();
50
AU.addRequiredID(LoopSimplifyID);
51
AU.addPreservedID(LoopSimplifyID);
52
AU.addRequiredID(LCSSAID);
53
AU.addPreservedID(LCSSAID);
54
AU.addPreserved<ScalarEvolution>();
60
bool rotateLoop(Loop *L, LPPassManager &LPM);
62
/// Initialize local data
65
/// After loop rotation, loop pre-header has multiple sucessors.
66
/// Insert one forwarding basic block to ensure that loop pre-header
67
/// has only one successor.
68
void preserveCanonicalLoopForm(LPPassManager &LPM);
72
BasicBlock *OrigHeader;
73
BasicBlock *OrigPreHeader;
74
BasicBlock *OrigLatch;
75
BasicBlock *NewHeader;
77
LPPassManager *LPM_Ptr;
81
char LoopRotate::ID = 0;
82
INITIALIZE_PASS(LoopRotate, "loop-rotate", "Rotate Loops", false, false);
84
Pass *llvm::createLoopRotatePass() { return new LoopRotate(); }
86
/// Rotate Loop L as many times as possible. Return true if
87
/// the loop is rotated at least once.
88
bool LoopRotate::runOnLoop(Loop *Lp, LPPassManager &LPM) {
90
bool RotatedOneLoop = false;
94
// One loop can be rotated multiple times.
95
while (rotateLoop(Lp,LPM)) {
96
RotatedOneLoop = true;
100
return RotatedOneLoop;
103
/// Rotate loop LP. Return true if the loop is rotated.
104
bool LoopRotate::rotateLoop(Loop *Lp, LPPassManager &LPM) {
107
OrigPreHeader = L->getLoopPreheader();
108
if (!OrigPreHeader) return false;
110
OrigLatch = L->getLoopLatch();
111
if (!OrigLatch) return false;
113
OrigHeader = L->getHeader();
115
// If the loop has only one block then there is not much to rotate.
116
if (L->getBlocks().size() == 1)
119
// If the loop header is not one of the loop exiting blocks then
120
// either this loop is already rotated or it is not
121
// suitable for loop rotation transformations.
122
if (!L->isLoopExiting(OrigHeader))
125
BranchInst *BI = dyn_cast<BranchInst>(OrigHeader->getTerminator());
128
assert(BI->isConditional() && "Branch Instruction is not conditional");
130
// Updating PHInodes in loops with multiple exits adds complexity.
131
// Keep it simple, and restrict loop rotation to loops with one exit only.
132
// In future, lift this restriction and support for multiple exits if
134
SmallVector<BasicBlock*, 8> ExitBlocks;
135
L->getExitBlocks(ExitBlocks);
136
if (ExitBlocks.size() > 1)
139
// Check size of original header and reject
140
// loop if it is very big.
143
// FIXME: Use common api to estimate size.
144
for (BasicBlock::const_iterator OI = OrigHeader->begin(),
145
OE = OrigHeader->end(); OI != OE; ++OI) {
146
if (isa<PHINode>(OI))
147
continue; // PHI nodes don't count.
148
if (isa<DbgInfoIntrinsic>(OI))
149
continue; // Debug intrinsics don't count as size.
153
if (Size > MAX_HEADER_SIZE)
156
// Now, this loop is suitable for rotation.
158
// Anything ScalarEvolution may know about this loop or the PHI nodes
159
// in its header will soon be invalidated.
160
if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
163
// Find new Loop header. NewHeader is a Header's one and only successor
164
// that is inside loop. Header's other successor is outside the
165
// loop. Otherwise loop is not suitable for rotation.
166
Exit = BI->getSuccessor(0);
167
NewHeader = BI->getSuccessor(1);
168
if (L->contains(Exit))
169
std::swap(Exit, NewHeader);
170
assert(NewHeader && "Unable to determine new loop header");
171
assert(L->contains(NewHeader) && !L->contains(Exit) &&
172
"Unable to determine loop header and exit blocks");
174
// This code assumes that the new header has exactly one predecessor.
175
// Remove any single-entry PHI nodes in it.
176
assert(NewHeader->getSinglePredecessor() &&
177
"New header doesn't have one pred!");
178
FoldSingleEntryPHINodes(NewHeader);
180
// Begin by walking OrigHeader and populating ValueMap with an entry for
182
BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end();
183
DenseMap<const Value *, Value *> ValueMap;
185
// For PHI nodes, the value available in OldPreHeader is just the
186
// incoming value from OldPreHeader.
187
for (; PHINode *PN = dyn_cast<PHINode>(I); ++I)
188
ValueMap[PN] = PN->getIncomingValue(PN->getBasicBlockIndex(OrigPreHeader));
190
// For the rest of the instructions, create a clone in the OldPreHeader.
191
TerminatorInst *LoopEntryBranch = OrigPreHeader->getTerminator();
192
for (; I != E; ++I) {
193
Instruction *C = I->clone();
194
C->setName(I->getName());
195
C->insertBefore(LoopEntryBranch);
199
// Along with all the other instructions, we just cloned OrigHeader's
200
// terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's
201
// successors by duplicating their incoming values for OrigHeader.
202
TerminatorInst *TI = OrigHeader->getTerminator();
203
for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
204
for (BasicBlock::iterator BI = TI->getSuccessor(i)->begin();
205
PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
206
PN->addIncoming(PN->getIncomingValueForBlock(OrigHeader), OrigPreHeader);
208
// Now that OrigPreHeader has a clone of OrigHeader's terminator, remove
209
// OrigPreHeader's old terminator (the original branch into the loop), and
210
// remove the corresponding incoming values from the PHI nodes in OrigHeader.
211
LoopEntryBranch->eraseFromParent();
212
for (I = OrigHeader->begin(); PHINode *PN = dyn_cast<PHINode>(I); ++I)
213
PN->removeIncomingValue(PN->getBasicBlockIndex(OrigPreHeader));
215
// Now fix up users of the instructions in OrigHeader, inserting PHI nodes
218
for (I = OrigHeader->begin(); I != E; ++I) {
219
Value *OrigHeaderVal = I;
220
Value *OrigPreHeaderVal = ValueMap[OrigHeaderVal];
222
// The value now exits in two versions: the initial value in the preheader
223
// and the loop "next" value in the original header.
224
SSA.Initialize(OrigHeaderVal->getType(), OrigHeaderVal->getName());
225
SSA.AddAvailableValue(OrigHeader, OrigHeaderVal);
226
SSA.AddAvailableValue(OrigPreHeader, OrigPreHeaderVal);
228
// Visit each use of the OrigHeader instruction.
229
for (Value::use_iterator UI = OrigHeaderVal->use_begin(),
230
UE = OrigHeaderVal->use_end(); UI != UE; ) {
231
// Grab the use before incrementing the iterator.
232
Use &U = UI.getUse();
234
// Increment the iterator before removing the use from the list.
237
// SSAUpdater can't handle a non-PHI use in the same block as an
238
// earlier def. We can easily handle those cases manually.
239
Instruction *UserInst = cast<Instruction>(U.getUser());
240
if (!isa<PHINode>(UserInst)) {
241
BasicBlock *UserBB = UserInst->getParent();
243
// The original users in the OrigHeader are already using the
244
// original definitions.
245
if (UserBB == OrigHeader)
248
// Users in the OrigPreHeader need to use the value to which the
249
// original definitions are mapped.
250
if (UserBB == OrigPreHeader) {
251
U = OrigPreHeaderVal;
256
// Anything else can be handled by SSAUpdater.
261
// NewHeader is now the header of the loop.
262
L->moveToHeader(NewHeader);
264
// Move the original header to the bottom of the loop, where it now more
265
// naturally belongs. This isn't necessary for correctness, and CodeGen can
266
// usually reorder blocks on its own to fix things like this up, but it's
267
// still nice to keep the IR readable.
269
// The original header should have only one predecessor at this point, since
270
// we checked that the loop had a proper preheader and unique backedge before
272
assert(OrigHeader->getSinglePredecessor() &&
273
"Original loop header has too many predecessors after loop rotation!");
274
OrigHeader->moveAfter(OrigHeader->getSinglePredecessor());
276
// Also, since this original header only has one predecessor, zap its
277
// PHI nodes, which are now trivial.
278
FoldSingleEntryPHINodes(OrigHeader);
280
// TODO: We could just go ahead and merge OrigHeader into its predecessor
281
// at this point, if we don't mind updating dominator info.
283
// Establish a new preheader, update dominators, etc.
284
preserveCanonicalLoopForm(LPM);
290
/// Initialize local data
291
void LoopRotate::initialize() {
294
OrigPreHeader = NULL;
299
/// After loop rotation, loop pre-header has multiple sucessors.
300
/// Insert one forwarding basic block to ensure that loop pre-header
301
/// has only one successor.
302
void LoopRotate::preserveCanonicalLoopForm(LPPassManager &LPM) {
304
// Right now original pre-header has two successors, new header and
305
// exit block. Insert new block between original pre-header and
306
// new header such that loop's new pre-header has only one successor.
307
BasicBlock *NewPreHeader = BasicBlock::Create(OrigHeader->getContext(),
309
OrigHeader->getParent(),
311
LoopInfo &LI = getAnalysis<LoopInfo>();
312
if (Loop *PL = LI.getLoopFor(OrigPreHeader))
313
PL->addBasicBlockToLoop(NewPreHeader, LI.getBase());
314
BranchInst::Create(NewHeader, NewPreHeader);
316
BranchInst *OrigPH_BI = cast<BranchInst>(OrigPreHeader->getTerminator());
317
if (OrigPH_BI->getSuccessor(0) == NewHeader)
318
OrigPH_BI->setSuccessor(0, NewPreHeader);
320
assert(OrigPH_BI->getSuccessor(1) == NewHeader &&
321
"Unexpected original pre-header terminator");
322
OrigPH_BI->setSuccessor(1, NewPreHeader);
326
for (BasicBlock::iterator I = NewHeader->begin();
327
(PN = dyn_cast<PHINode>(I)); ++I) {
328
int index = PN->getBasicBlockIndex(OrigPreHeader);
329
assert(index != -1 && "Expected incoming value from Original PreHeader");
330
PN->setIncomingBlock(index, NewPreHeader);
331
assert(PN->getBasicBlockIndex(OrigPreHeader) == -1 &&
332
"Expected only one incoming value from Original PreHeader");
335
if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>()) {
336
DT->addNewBlock(NewPreHeader, OrigPreHeader);
337
DT->changeImmediateDominator(L->getHeader(), NewPreHeader);
338
DT->changeImmediateDominator(Exit, OrigPreHeader);
339
for (Loop::block_iterator BI = L->block_begin(), BE = L->block_end();
342
if (L->getHeader() != B) {
343
DomTreeNode *Node = DT->getNode(B);
344
if (Node && Node->getBlock() == OrigHeader)
345
DT->changeImmediateDominator(*BI, L->getHeader());
348
DT->changeImmediateDominator(OrigHeader, OrigLatch);
351
if (DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>()) {
352
// New Preheader's dominance frontier is Exit block.
353
DominanceFrontier::DomSetType NewPHSet;
354
NewPHSet.insert(Exit);
355
DF->addBasicBlock(NewPreHeader, NewPHSet);
357
// New Header's dominance frontier now includes itself and Exit block
358
DominanceFrontier::iterator HeadI = DF->find(L->getHeader());
359
if (HeadI != DF->end()) {
360
DominanceFrontier::DomSetType & HeaderSet = HeadI->second;
362
HeaderSet.insert(L->getHeader());
363
HeaderSet.insert(Exit);
365
DominanceFrontier::DomSetType HeaderSet;
366
HeaderSet.insert(L->getHeader());
367
HeaderSet.insert(Exit);
368
DF->addBasicBlock(L->getHeader(), HeaderSet);
371
// Original header (new Loop Latch)'s dominance frontier is Exit.
372
DominanceFrontier::iterator LatchI = DF->find(L->getLoopLatch());
373
if (LatchI != DF->end()) {
374
DominanceFrontier::DomSetType &LatchSet = LatchI->second;
375
LatchSet = LatchI->second;
377
LatchSet.insert(Exit);
379
DominanceFrontier::DomSetType LatchSet;
380
LatchSet.insert(Exit);
381
DF->addBasicBlock(L->getHeader(), LatchSet);
384
// If a loop block dominates new loop latch then add to its frontiers
385
// new header and Exit and remove new latch (which is equal to original
387
BasicBlock *NewLatch = L->getLoopLatch();
389
assert(NewLatch == OrigHeader && "NewLatch is inequal to OrigHeader");
391
if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>()) {
392
for (Loop::block_iterator BI = L->block_begin(), BE = L->block_end();
395
if (DT->dominates(B, NewLatch)) {
396
DominanceFrontier::iterator BDFI = DF->find(B);
397
if (BDFI != DF->end()) {
398
DominanceFrontier::DomSetType &BSet = BDFI->second;
399
BSet.erase(NewLatch);
400
BSet.insert(L->getHeader());
403
DominanceFrontier::DomSetType BSet;
404
BSet.insert(L->getHeader());
406
DF->addBasicBlock(B, BSet);
413
// Preserve canonical loop form, which means Exit block should
414
// have only one predecessor.
415
SplitEdge(L->getLoopLatch(), Exit, this);
417
assert(NewHeader && L->getHeader() == NewHeader &&
418
"Invalid loop header after loop rotation");
419
assert(NewPreHeader && L->getLoopPreheader() == NewPreHeader &&
420
"Invalid loop preheader after loop rotation");
421
assert(L->getLoopLatch() &&
422
"Invalid loop latch after loop rotation");