1 //===-- TailDuplication.cpp - Duplicate blocks into predecessors' tails ---===//
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 pass duplicates basic blocks ending in unconditional branches into
11 // the tails of their predecessors.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "tailduplication"
16 #include "llvm/Function.h"
17 #include "llvm/CodeGen/Passes.h"
18 #include "llvm/CodeGen/MachineModuleInfo.h"
19 #include "llvm/CodeGen/MachineFunctionPass.h"
20 #include "llvm/Target/TargetInstrInfo.h"
21 #include "llvm/Support/CommandLine.h"
22 #include "llvm/Support/Debug.h"
23 #include "llvm/Support/raw_ostream.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/SetVector.h"
26 #include "llvm/ADT/Statistic.h"
29 STATISTIC(NumTailDups , "Number of tail duplicated blocks");
30 STATISTIC(NumInstrDups , "Additional instructions due to tail duplication");
31 STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
33 // Heuristic for tail duplication.
34 static cl::opt<unsigned>
35 TailDuplicateSize("tail-dup-size",
36 cl::desc("Maximum instructions to consider tail duplicating"),
37 cl::init(2), cl::Hidden);
40 /// TailDuplicatePass - Perform tail duplication.
41 class TailDuplicatePass : public MachineFunctionPass {
42 const TargetInstrInfo *TII;
43 MachineModuleInfo *MMI;
47 explicit TailDuplicatePass() : MachineFunctionPass(&ID) {}
49 virtual bool runOnMachineFunction(MachineFunction &MF);
50 virtual const char *getPassName() const { return "Tail Duplication"; }
53 bool TailDuplicateBlocks(MachineFunction &MF);
54 bool TailDuplicate(MachineBasicBlock *TailBB, MachineFunction &MF);
55 void RemoveDeadBlock(MachineBasicBlock *MBB);
58 char TailDuplicatePass::ID = 0;
61 FunctionPass *llvm::createTailDuplicatePass() {
62 return new TailDuplicatePass();
65 bool TailDuplicatePass::runOnMachineFunction(MachineFunction &MF) {
66 TII = MF.getTarget().getInstrInfo();
67 MMI = getAnalysisIfAvailable<MachineModuleInfo>();
69 bool MadeChange = false;
70 bool MadeChangeThisIteration = true;
71 while (MadeChangeThisIteration) {
72 MadeChangeThisIteration = false;
73 MadeChangeThisIteration |= TailDuplicateBlocks(MF);
74 MadeChange |= MadeChangeThisIteration;
80 /// TailDuplicateBlocks - Look for small blocks that are unconditionally
81 /// branched to and do not fall through. Tail-duplicate their instructions
82 /// into their predecessors to eliminate (dynamic) branches.
83 bool TailDuplicatePass::TailDuplicateBlocks(MachineFunction &MF) {
84 bool MadeChange = false;
86 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
87 MachineBasicBlock *MBB = I++;
89 // Only duplicate blocks that end with unconditional branches.
90 if (MBB->canFallThrough())
93 MadeChange |= TailDuplicate(MBB, MF);
95 // If it is dead, remove it.
96 if (MBB->pred_empty()) {
97 NumInstrDups -= MBB->size();
106 /// TailDuplicate - If it is profitable, duplicate TailBB's contents in each
107 /// of its predecessors.
108 bool TailDuplicatePass::TailDuplicate(MachineBasicBlock *TailBB,
109 MachineFunction &MF) {
110 // Don't try to tail-duplicate single-block loops.
111 if (TailBB->isSuccessor(TailBB))
114 // Set the limit on the number of instructions to duplicate, with a default
115 // of one less than the tail-merge threshold. When optimizing for size,
116 // duplicate only one, because one branch instruction can be eliminated to
117 // compensate for the duplication.
118 unsigned MaxDuplicateCount;
119 if (!TailBB->empty() && TailBB->back().getDesc().isIndirectBranch())
120 // If the target has hardware branch prediction that can handle indirect
121 // branches, duplicating them can often make them predictable when there
122 // are common paths through the code. The limit needs to be high enough
123 // to allow undoing the effects of tail merging.
124 MaxDuplicateCount = 20;
125 else if (MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize))
126 MaxDuplicateCount = 1;
128 MaxDuplicateCount = TailDuplicateSize;
130 // Check the instructions in the block to determine whether tail-duplication
131 // is invalid or unlikely to be profitable.
132 unsigned InstrCount = 0;
133 bool HasCall = false;
134 for (MachineBasicBlock::iterator I = TailBB->begin();
135 I != TailBB->end(); ++I) {
136 // Non-duplicable things shouldn't be tail-duplicated.
137 if (I->getDesc().isNotDuplicable()) return false;
138 // Don't duplicate more than the threshold.
139 if (InstrCount == MaxDuplicateCount) return false;
140 // Remember if we saw a call.
141 if (I->getDesc().isCall()) HasCall = true;
142 if (I->getOpcode() != TargetInstrInfo::PHI)
145 // Heuristically, don't tail-duplicate calls if it would expand code size,
146 // as it's less likely to be worth the extra cost.
147 if (InstrCount > 1 && HasCall)
150 // Iterate through all the unique predecessors and tail-duplicate this
151 // block into them, if possible. Copying the list ahead of time also
152 // avoids trouble with the predecessor list reallocating.
153 bool Changed = false;
154 SmallSetVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(),
156 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
157 PE = Preds.end(); PI != PE; ++PI) {
158 MachineBasicBlock *PredBB = *PI;
160 assert(TailBB != PredBB &&
161 "Single-block loop should have been rejected earlier!");
162 if (PredBB->succ_size() > 1) continue;
164 MachineBasicBlock *PredTBB, *PredFBB;
165 SmallVector<MachineOperand, 4> PredCond;
166 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
168 if (!PredCond.empty())
170 // EH edges are ignored by AnalyzeBranch.
171 if (PredBB->succ_size() != 1)
173 // Don't duplicate into a fall-through predecessor (at least for now).
174 if (PredBB->isLayoutSuccessor(TailBB) && PredBB->canFallThrough())
177 DEBUG(errs() << "\nTail-duplicating into PredBB: " << *PredBB
178 << "From Succ: " << *TailBB);
180 // Remove PredBB's unconditional branch.
181 TII->RemoveBranch(*PredBB);
182 // Clone the contents of TailBB into PredBB.
183 for (MachineBasicBlock::iterator I = TailBB->begin(), E = TailBB->end();
185 MachineInstr *NewMI = MF.CloneMachineInstr(I);
186 PredBB->insert(PredBB->end(), NewMI);
188 NumInstrDups += TailBB->size() - 1; // subtract one for removed branch
191 PredBB->removeSuccessor(PredBB->succ_begin());
192 assert(PredBB->succ_empty() &&
193 "TailDuplicate called on block with multiple successors!");
194 for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(),
195 E = TailBB->succ_end(); I != E; ++I)
196 PredBB->addSuccessor(*I);
202 // If TailBB was duplicated into all its predecessors except for the prior
203 // block, which falls through unconditionally, move the contents of this
204 // block into the prior block.
205 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(TailBB));
206 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
207 SmallVector<MachineOperand, 4> PriorCond;
208 bool PriorUnAnalyzable =
209 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
210 // This has to check PrevBB->succ_size() because EH edges are ignored by
212 if (!PriorUnAnalyzable && PriorCond.empty() && !PriorTBB &&
213 TailBB->pred_size() == 1 && PrevBB.succ_size() == 1 &&
214 !TailBB->hasAddressTaken()) {
215 DEBUG(errs() << "\nMerging into block: " << PrevBB
216 << "From MBB: " << *TailBB);
217 PrevBB.splice(PrevBB.end(), TailBB, TailBB->begin(), TailBB->end());
218 PrevBB.removeSuccessor(PrevBB.succ_begin());;
219 assert(PrevBB.succ_empty());
220 PrevBB.transferSuccessors(TailBB);
227 /// RemoveDeadBlock - Remove the specified dead machine basic block from the
228 /// function, updating the CFG.
229 void TailDuplicatePass::RemoveDeadBlock(MachineBasicBlock *MBB) {
230 assert(MBB->pred_empty() && "MBB must be dead!");
231 DEBUG(errs() << "\nRemoving MBB: " << *MBB);
233 // Remove all successors.
234 while (!MBB->succ_empty())
235 MBB->removeSuccessor(MBB->succ_end()-1);
237 // If there are any labels in the basic block, unregister them from
238 // MachineModuleInfo.
239 if (MMI && !MBB->empty()) {
240 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
243 // The label ID # is always operand #0, an immediate.
244 MMI->InvalidateLabel(I->getOperand(0).getImm());
249 MBB->eraseFromParent();