1//===-- BranchFolding.cpp - Fold machine code branch instructions ---------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass forwards branches to unconditional branches to make them branch 11// directly to the target block. This pass often results in dead MBB's, which 12// it then removes. 13// 14// Note that this pass must be run after register allocation, it cannot handle 15// SSA form. 16// 17//===----------------------------------------------------------------------===// 18 19#define DEBUG_TYPE "branchfolding" 20#include "BranchFolding.h" 21#include "llvm/Function.h" 22#include "llvm/CodeGen/Passes.h" 23#include "llvm/CodeGen/MachineModuleInfo.h" 24#include "llvm/CodeGen/MachineFunctionPass.h" 25#include "llvm/CodeGen/MachineJumpTableInfo.h" 26#include "llvm/CodeGen/RegisterScavenging.h" 27#include "llvm/Target/TargetInstrInfo.h" 28#include "llvm/Target/TargetMachine.h" 29#include "llvm/Target/TargetRegisterInfo.h" 30#include "llvm/Support/CommandLine.h" 31#include "llvm/Support/Debug.h" 32#include "llvm/Support/ErrorHandling.h" 33#include "llvm/Support/raw_ostream.h" 34#include "llvm/ADT/SmallSet.h" 35#include "llvm/ADT/SetVector.h" 36#include "llvm/ADT/Statistic.h" 37#include "llvm/ADT/STLExtras.h" 38#include <algorithm> 39using namespace llvm; 40 41STATISTIC(NumDeadBlocks, "Number of dead blocks removed"); 42STATISTIC(NumBranchOpts, "Number of branches optimized"); 43STATISTIC(NumTailMerge , "Number of block tails merged");
| 1//===-- BranchFolding.cpp - Fold machine code branch instructions ---------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass forwards branches to unconditional branches to make them branch 11// directly to the target block. This pass often results in dead MBB's, which 12// it then removes. 13// 14// Note that this pass must be run after register allocation, it cannot handle 15// SSA form. 16// 17//===----------------------------------------------------------------------===// 18 19#define DEBUG_TYPE "branchfolding" 20#include "BranchFolding.h" 21#include "llvm/Function.h" 22#include "llvm/CodeGen/Passes.h" 23#include "llvm/CodeGen/MachineModuleInfo.h" 24#include "llvm/CodeGen/MachineFunctionPass.h" 25#include "llvm/CodeGen/MachineJumpTableInfo.h" 26#include "llvm/CodeGen/RegisterScavenging.h" 27#include "llvm/Target/TargetInstrInfo.h" 28#include "llvm/Target/TargetMachine.h" 29#include "llvm/Target/TargetRegisterInfo.h" 30#include "llvm/Support/CommandLine.h" 31#include "llvm/Support/Debug.h" 32#include "llvm/Support/ErrorHandling.h" 33#include "llvm/Support/raw_ostream.h" 34#include "llvm/ADT/SmallSet.h" 35#include "llvm/ADT/SetVector.h" 36#include "llvm/ADT/Statistic.h" 37#include "llvm/ADT/STLExtras.h" 38#include <algorithm> 39using namespace llvm; 40 41STATISTIC(NumDeadBlocks, "Number of dead blocks removed"); 42STATISTIC(NumBranchOpts, "Number of branches optimized"); 43STATISTIC(NumTailMerge , "Number of block tails merged");
|
| 44STATISTIC(NumTailDups , "Number of tail duplicated blocks"); 45STATISTIC(NumInstrDups , "Additional instructions due to tail duplication"); 46
|
44static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", 45 cl::init(cl::BOU_UNSET), cl::Hidden);
| 47static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", 48 cl::init(cl::BOU_UNSET), cl::Hidden);
|
| 49
|
46// Throttle for huge numbers of predecessors (compile speed problems) 47static cl::opt<unsigned> 48TailMergeThreshold("tail-merge-threshold", 49 cl::desc("Max number of predecessors to consider tail merging"), 50 cl::init(150), cl::Hidden); 51 52// Heuristic for tail merging (and, inversely, tail duplication). 53// TODO: This should be replaced with a target query. 54static cl::opt<unsigned> 55TailMergeSize("tail-merge-size", 56 cl::desc("Min number of instructions to consider tail merging"), 57 cl::init(3), cl::Hidden); 58 59namespace { 60 /// BranchFolderPass - Wrap branch folder in a machine function pass. 61 class BranchFolderPass : public MachineFunctionPass, 62 public BranchFolder { 63 public: 64 static char ID; 65 explicit BranchFolderPass(bool defaultEnableTailMerge) 66 : MachineFunctionPass(&ID), BranchFolder(defaultEnableTailMerge) {} 67 68 virtual bool runOnMachineFunction(MachineFunction &MF); 69 virtual const char *getPassName() const { return "Control Flow Optimizer"; } 70 }; 71} 72 73char BranchFolderPass::ID = 0; 74 75FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) { 76 return new BranchFolderPass(DefaultEnableTailMerge); 77} 78 79bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) { 80 return OptimizeFunction(MF, 81 MF.getTarget().getInstrInfo(), 82 MF.getTarget().getRegisterInfo(), 83 getAnalysisIfAvailable<MachineModuleInfo>()); 84} 85 86 87BranchFolder::BranchFolder(bool defaultEnableTailMerge) { 88 switch (FlagEnableTailMerge) { 89 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 90 case cl::BOU_TRUE: EnableTailMerge = true; break; 91 case cl::BOU_FALSE: EnableTailMerge = false; break; 92 } 93} 94 95/// RemoveDeadBlock - Remove the specified dead machine basic block from the 96/// function, updating the CFG. 97void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 98 assert(MBB->pred_empty() && "MBB must be dead!"); 99 DEBUG(errs() << "\nRemoving MBB: " << *MBB); 100 101 MachineFunction *MF = MBB->getParent(); 102 // drop all successors. 103 while (!MBB->succ_empty()) 104 MBB->removeSuccessor(MBB->succ_end()-1); 105 106 // If there are any labels in the basic block, unregister them from 107 // MachineModuleInfo. 108 if (MMI && !MBB->empty()) { 109 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 110 I != E; ++I) { 111 if (I->isLabel()) 112 // The label ID # is always operand #0, an immediate. 113 MMI->InvalidateLabel(I->getOperand(0).getImm()); 114 } 115 } 116 117 // Remove the block. 118 MF->erase(MBB); 119} 120 121/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 122/// followed by terminators, and if the implicitly defined registers are not 123/// used by the terminators, remove those implicit_def's. e.g. 124/// BB1: 125/// r0 = implicit_def 126/// r1 = implicit_def 127/// br 128/// This block can be optimized away later if the implicit instructions are 129/// removed. 130bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 131 SmallSet<unsigned, 4> ImpDefRegs; 132 MachineBasicBlock::iterator I = MBB->begin(); 133 while (I != MBB->end()) { 134 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) 135 break; 136 unsigned Reg = I->getOperand(0).getReg(); 137 ImpDefRegs.insert(Reg); 138 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg); 139 unsigned SubReg = *SubRegs; ++SubRegs) 140 ImpDefRegs.insert(SubReg); 141 ++I; 142 } 143 if (ImpDefRegs.empty()) 144 return false; 145 146 MachineBasicBlock::iterator FirstTerm = I; 147 while (I != MBB->end()) { 148 if (!TII->isUnpredicatedTerminator(I)) 149 return false; 150 // See if it uses any of the implicitly defined registers. 151 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 152 MachineOperand &MO = I->getOperand(i); 153 if (!MO.isReg() || !MO.isUse()) 154 continue; 155 unsigned Reg = MO.getReg(); 156 if (ImpDefRegs.count(Reg)) 157 return false; 158 } 159 ++I; 160 } 161 162 I = MBB->begin(); 163 while (I != FirstTerm) { 164 MachineInstr *ImpDefMI = &*I; 165 ++I; 166 MBB->erase(ImpDefMI); 167 } 168 169 return true; 170} 171 172/// OptimizeFunction - Perhaps branch folding, tail merging and other 173/// CFG optimizations on the given function. 174bool BranchFolder::OptimizeFunction(MachineFunction &MF, 175 const TargetInstrInfo *tii, 176 const TargetRegisterInfo *tri, 177 MachineModuleInfo *mmi) { 178 if (!tii) return false; 179 180 TII = tii; 181 TRI = tri; 182 MMI = mmi; 183 184 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 185 186 // Fix CFG. The later algorithms expect it to be right. 187 bool MadeChange = false; 188 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 189 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 190 SmallVector<MachineOperand, 4> Cond; 191 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true)) 192 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 193 MadeChange |= OptimizeImpDefsBlock(MBB); 194 } 195
| 50// Throttle for huge numbers of predecessors (compile speed problems) 51static cl::opt<unsigned> 52TailMergeThreshold("tail-merge-threshold", 53 cl::desc("Max number of predecessors to consider tail merging"), 54 cl::init(150), cl::Hidden); 55 56// Heuristic for tail merging (and, inversely, tail duplication). 57// TODO: This should be replaced with a target query. 58static cl::opt<unsigned> 59TailMergeSize("tail-merge-size", 60 cl::desc("Min number of instructions to consider tail merging"), 61 cl::init(3), cl::Hidden); 62 63namespace { 64 /// BranchFolderPass - Wrap branch folder in a machine function pass. 65 class BranchFolderPass : public MachineFunctionPass, 66 public BranchFolder { 67 public: 68 static char ID; 69 explicit BranchFolderPass(bool defaultEnableTailMerge) 70 : MachineFunctionPass(&ID), BranchFolder(defaultEnableTailMerge) {} 71 72 virtual bool runOnMachineFunction(MachineFunction &MF); 73 virtual const char *getPassName() const { return "Control Flow Optimizer"; } 74 }; 75} 76 77char BranchFolderPass::ID = 0; 78 79FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) { 80 return new BranchFolderPass(DefaultEnableTailMerge); 81} 82 83bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) { 84 return OptimizeFunction(MF, 85 MF.getTarget().getInstrInfo(), 86 MF.getTarget().getRegisterInfo(), 87 getAnalysisIfAvailable<MachineModuleInfo>()); 88} 89 90 91BranchFolder::BranchFolder(bool defaultEnableTailMerge) { 92 switch (FlagEnableTailMerge) { 93 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 94 case cl::BOU_TRUE: EnableTailMerge = true; break; 95 case cl::BOU_FALSE: EnableTailMerge = false; break; 96 } 97} 98 99/// RemoveDeadBlock - Remove the specified dead machine basic block from the 100/// function, updating the CFG. 101void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 102 assert(MBB->pred_empty() && "MBB must be dead!"); 103 DEBUG(errs() << "\nRemoving MBB: " << *MBB); 104 105 MachineFunction *MF = MBB->getParent(); 106 // drop all successors. 107 while (!MBB->succ_empty()) 108 MBB->removeSuccessor(MBB->succ_end()-1); 109 110 // If there are any labels in the basic block, unregister them from 111 // MachineModuleInfo. 112 if (MMI && !MBB->empty()) { 113 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 114 I != E; ++I) { 115 if (I->isLabel()) 116 // The label ID # is always operand #0, an immediate. 117 MMI->InvalidateLabel(I->getOperand(0).getImm()); 118 } 119 } 120 121 // Remove the block. 122 MF->erase(MBB); 123} 124 125/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 126/// followed by terminators, and if the implicitly defined registers are not 127/// used by the terminators, remove those implicit_def's. e.g. 128/// BB1: 129/// r0 = implicit_def 130/// r1 = implicit_def 131/// br 132/// This block can be optimized away later if the implicit instructions are 133/// removed. 134bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 135 SmallSet<unsigned, 4> ImpDefRegs; 136 MachineBasicBlock::iterator I = MBB->begin(); 137 while (I != MBB->end()) { 138 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) 139 break; 140 unsigned Reg = I->getOperand(0).getReg(); 141 ImpDefRegs.insert(Reg); 142 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg); 143 unsigned SubReg = *SubRegs; ++SubRegs) 144 ImpDefRegs.insert(SubReg); 145 ++I; 146 } 147 if (ImpDefRegs.empty()) 148 return false; 149 150 MachineBasicBlock::iterator FirstTerm = I; 151 while (I != MBB->end()) { 152 if (!TII->isUnpredicatedTerminator(I)) 153 return false; 154 // See if it uses any of the implicitly defined registers. 155 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 156 MachineOperand &MO = I->getOperand(i); 157 if (!MO.isReg() || !MO.isUse()) 158 continue; 159 unsigned Reg = MO.getReg(); 160 if (ImpDefRegs.count(Reg)) 161 return false; 162 } 163 ++I; 164 } 165 166 I = MBB->begin(); 167 while (I != FirstTerm) { 168 MachineInstr *ImpDefMI = &*I; 169 ++I; 170 MBB->erase(ImpDefMI); 171 } 172 173 return true; 174} 175 176/// OptimizeFunction - Perhaps branch folding, tail merging and other 177/// CFG optimizations on the given function. 178bool BranchFolder::OptimizeFunction(MachineFunction &MF, 179 const TargetInstrInfo *tii, 180 const TargetRegisterInfo *tri, 181 MachineModuleInfo *mmi) { 182 if (!tii) return false; 183 184 TII = tii; 185 TRI = tri; 186 MMI = mmi; 187 188 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 189 190 // Fix CFG. The later algorithms expect it to be right. 191 bool MadeChange = false; 192 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 193 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 194 SmallVector<MachineOperand, 4> Cond; 195 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true)) 196 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 197 MadeChange |= OptimizeImpDefsBlock(MBB); 198 } 199
|
196
| |
197 bool MadeChangeThisIteration = true; 198 while (MadeChangeThisIteration) { 199 MadeChangeThisIteration = false; 200 MadeChangeThisIteration |= TailMergeBlocks(MF); 201 MadeChangeThisIteration |= OptimizeBranches(MF); 202 MadeChange |= MadeChangeThisIteration; 203 } 204
| 200 bool MadeChangeThisIteration = true; 201 while (MadeChangeThisIteration) { 202 MadeChangeThisIteration = false; 203 MadeChangeThisIteration |= TailMergeBlocks(MF); 204 MadeChangeThisIteration |= OptimizeBranches(MF); 205 MadeChange |= MadeChangeThisIteration; 206 } 207
|
205 // Do tail duplication once after tail merging is done. Otherwise it is
| 208 // Do tail duplication after tail merging is done. Otherwise it is
|
206 // tough to avoid situations where tail duplication and tail merging undo 207 // each other's transformations ad infinitum.
| 209 // tough to avoid situations where tail duplication and tail merging undo 210 // each other's transformations ad infinitum.
|
208 MadeChange |= TailDuplicateBlocks(MF);
| 211 MadeChangeThisIteration = true; 212 while (MadeChangeThisIteration) { 213 MadeChangeThisIteration = false; 214 MadeChangeThisIteration |= TailDuplicateBlocks(MF); 215 MadeChange |= MadeChangeThisIteration; 216 }
|
209 210 // See if any jump tables have become mergable or dead as the code generator 211 // did its thing. 212 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 213 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 214 if (!JTs.empty()) { 215 // Figure out how these jump tables should be merged. 216 std::vector<unsigned> JTMapping; 217 JTMapping.reserve(JTs.size()); 218 219 // We always keep the 0th jump table. 220 JTMapping.push_back(0); 221 222 // Scan the jump tables, seeing if there are any duplicates. Note that this 223 // is N^2, which should be fixed someday. 224 for (unsigned i = 1, e = JTs.size(); i != e; ++i) { 225 if (JTs[i].MBBs.empty()) 226 JTMapping.push_back(i); 227 else 228 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 229 } 230 231 // If a jump table was merge with another one, walk the function rewriting 232 // references to jump tables to reference the new JT ID's. Keep track of 233 // whether we see a jump table idx, if not, we can delete the JT. 234 BitVector JTIsLive(JTs.size()); 235 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 236 BB != E; ++BB) { 237 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 238 I != E; ++I) 239 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 240 MachineOperand &Op = I->getOperand(op); 241 if (!Op.isJTI()) continue; 242 unsigned NewIdx = JTMapping[Op.getIndex()]; 243 Op.setIndex(NewIdx); 244 245 // Remember that this JT is live. 246 JTIsLive.set(NewIdx); 247 } 248 } 249 250 // Finally, remove dead jump tables. This happens either because the 251 // indirect jump was unreachable (and thus deleted) or because the jump 252 // table was merged with some other one. 253 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 254 if (!JTIsLive.test(i)) { 255 JTI->RemoveJumpTable(i); 256 MadeChange = true; 257 } 258 } 259 260 delete RS; 261 return MadeChange; 262} 263 264//===----------------------------------------------------------------------===// 265// Tail Merging of Blocks 266//===----------------------------------------------------------------------===// 267 268/// HashMachineInstr - Compute a hash value for MI and its operands. 269static unsigned HashMachineInstr(const MachineInstr *MI) { 270 unsigned Hash = MI->getOpcode(); 271 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 272 const MachineOperand &Op = MI->getOperand(i); 273 274 // Merge in bits from the operand if easy. 275 unsigned OperandHash = 0; 276 switch (Op.getType()) { 277 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 278 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 279 case MachineOperand::MO_MachineBasicBlock: 280 OperandHash = Op.getMBB()->getNumber(); 281 break; 282 case MachineOperand::MO_FrameIndex: 283 case MachineOperand::MO_ConstantPoolIndex: 284 case MachineOperand::MO_JumpTableIndex: 285 OperandHash = Op.getIndex(); 286 break; 287 case MachineOperand::MO_GlobalAddress: 288 case MachineOperand::MO_ExternalSymbol: 289 // Global address / external symbol are too hard, don't bother, but do 290 // pull in the offset. 291 OperandHash = Op.getOffset(); 292 break; 293 default: break; 294 } 295 296 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 297 } 298 return Hash; 299} 300 301/// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks 302/// with no successors, we hash two instructions, because cross-jumping 303/// only saves code when at least two instructions are removed (since a 304/// branch must be inserted). For blocks with a successor, one of the 305/// two blocks to be tail-merged will end with a branch already, so 306/// it gains to cross-jump even for one instruction. 307static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, 308 unsigned minCommonTailLength) { 309 MachineBasicBlock::const_iterator I = MBB->end(); 310 if (I == MBB->begin()) 311 return 0; // Empty MBB. 312 313 --I; 314 unsigned Hash = HashMachineInstr(I); 315 316 if (I == MBB->begin() || minCommonTailLength == 1) 317 return Hash; // Single instr MBB. 318 319 --I; 320 // Hash in the second-to-last instruction. 321 Hash ^= HashMachineInstr(I) << 2; 322 return Hash; 323} 324 325/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 326/// of instructions they actually have in common together at their end. Return 327/// iterators for the first shared instruction in each block. 328static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 329 MachineBasicBlock *MBB2, 330 MachineBasicBlock::iterator &I1, 331 MachineBasicBlock::iterator &I2) { 332 I1 = MBB1->end(); 333 I2 = MBB2->end(); 334 335 unsigned TailLen = 0; 336 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 337 --I1; --I2; 338 if (!I1->isIdenticalTo(I2) || 339 // FIXME: This check is dubious. It's used to get around a problem where 340 // people incorrectly expect inline asm directives to remain in the same 341 // relative order. This is untenable because normal compiler 342 // optimizations (like this one) may reorder and/or merge these 343 // directives. 344 I1->getOpcode() == TargetInstrInfo::INLINEASM) { 345 ++I1; ++I2; 346 break; 347 } 348 ++TailLen; 349 } 350 return TailLen; 351} 352 353/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 354/// after it, replacing it with an unconditional branch to NewDest. This 355/// returns true if OldInst's block is modified, false if NewDest is modified. 356void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 357 MachineBasicBlock *NewDest) { 358 MachineBasicBlock *OldBB = OldInst->getParent(); 359 360 // Remove all the old successors of OldBB from the CFG. 361 while (!OldBB->succ_empty()) 362 OldBB->removeSuccessor(OldBB->succ_begin()); 363 364 // Remove all the dead instructions from the end of OldBB. 365 OldBB->erase(OldInst, OldBB->end()); 366 367 // If OldBB isn't immediately before OldBB, insert a branch to it. 368 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 369 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>()); 370 OldBB->addSuccessor(NewDest); 371 ++NumTailMerge; 372} 373 374/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 375/// MBB so that the part before the iterator falls into the part starting at the 376/// iterator. This returns the new MBB. 377MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 378 MachineBasicBlock::iterator BBI1) { 379 MachineFunction &MF = *CurMBB.getParent(); 380 381 // Create the fall-through block. 382 MachineFunction::iterator MBBI = &CurMBB; 383 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 384 CurMBB.getParent()->insert(++MBBI, NewMBB); 385 386 // Move all the successors of this block to the specified block. 387 NewMBB->transferSuccessors(&CurMBB); 388 389 // Add an edge from CurMBB to NewMBB for the fall-through. 390 CurMBB.addSuccessor(NewMBB); 391 392 // Splice the code over. 393 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 394 395 // For targets that use the register scavenger, we must maintain LiveIns. 396 if (RS) { 397 RS->enterBasicBlock(&CurMBB); 398 if (!CurMBB.empty()) 399 RS->forward(prior(CurMBB.end())); 400 BitVector RegsLiveAtExit(TRI->getNumRegs()); 401 RS->getRegsUsed(RegsLiveAtExit, false); 402 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++) 403 if (RegsLiveAtExit[i]) 404 NewMBB->addLiveIn(i); 405 } 406 407 return NewMBB; 408} 409 410/// EstimateRuntime - Make a rough estimate for how long it will take to run 411/// the specified code. 412static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 413 MachineBasicBlock::iterator E) { 414 unsigned Time = 0; 415 for (; I != E; ++I) { 416 const TargetInstrDesc &TID = I->getDesc(); 417 if (TID.isCall()) 418 Time += 10; 419 else if (TID.mayLoad() || TID.mayStore()) 420 Time += 2; 421 else 422 ++Time; 423 } 424 return Time; 425} 426 427// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 428// branches temporarily for tail merging). In the case where CurMBB ends 429// with a conditional branch to the next block, optimize by reversing the 430// test and conditionally branching to SuccMBB instead. 431static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, 432 const TargetInstrInfo *TII) { 433 MachineFunction *MF = CurMBB->getParent(); 434 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); 435 MachineBasicBlock *TBB = 0, *FBB = 0; 436 SmallVector<MachineOperand, 4> Cond; 437 if (I != MF->end() && 438 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 439 MachineBasicBlock *NextBB = I; 440 if (TBB == NextBB && !Cond.empty() && !FBB) { 441 if (!TII->ReverseBranchCondition(Cond)) { 442 TII->RemoveBranch(*CurMBB); 443 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); 444 return; 445 } 446 } 447 } 448 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>()); 449} 450 451bool 452BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { 453 if (getHash() < o.getHash()) 454 return true; 455 else if (getHash() > o.getHash()) 456 return false; 457 else if (getBlock()->getNumber() < o.getBlock()->getNumber()) 458 return true; 459 else if (getBlock()->getNumber() > o.getBlock()->getNumber()) 460 return false; 461 else { 462 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 463 // an object with itself. 464#ifndef _GLIBCXX_DEBUG 465 llvm_unreachable("Predecessor appears twice"); 466#endif 467 return false; 468 } 469} 470 471/// CountTerminators - Count the number of terminators in the given 472/// block and set I to the position of the first non-terminator, if there 473/// is one, or MBB->end() otherwise. 474static unsigned CountTerminators(MachineBasicBlock *MBB, 475 MachineBasicBlock::iterator &I) { 476 I = MBB->end(); 477 unsigned NumTerms = 0; 478 for (;;) { 479 if (I == MBB->begin()) { 480 I = MBB->end(); 481 break; 482 } 483 --I; 484 if (!I->getDesc().isTerminator()) break; 485 ++NumTerms; 486 } 487 return NumTerms; 488} 489 490/// ProfitableToMerge - Check if two machine basic blocks have a common tail 491/// and decide if it would be profitable to merge those tails. Return the 492/// length of the common tail and iterators to the first common instruction 493/// in each block. 494static bool ProfitableToMerge(MachineBasicBlock *MBB1, 495 MachineBasicBlock *MBB2, 496 unsigned minCommonTailLength, 497 unsigned &CommonTailLen, 498 MachineBasicBlock::iterator &I1, 499 MachineBasicBlock::iterator &I2, 500 MachineBasicBlock *SuccBB, 501 MachineBasicBlock *PredBB) { 502 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 503 MachineFunction *MF = MBB1->getParent(); 504 505 if (CommonTailLen == 0) 506 return false; 507 508 // It's almost always profitable to merge any number of non-terminator 509 // instructions with the block that falls through into the common successor. 510 if (MBB1 == PredBB || MBB2 == PredBB) { 511 MachineBasicBlock::iterator I; 512 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 513 if (CommonTailLen > NumTerms) 514 return true; 515 } 516 517 // If one of the blocks can be completely merged and happens to be in 518 // a position where the other could fall through into it, merge any number 519 // of instructions, because it can be done without a branch. 520 // TODO: If the blocks are not adjacent, move one of them so that they are? 521 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) 522 return true; 523 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) 524 return true; 525 526 // If both blocks have an unconditional branch temporarily stripped out, 527 // count that as an additional common instruction for the following 528 // heuristics. 529 unsigned EffectiveTailLen = CommonTailLen; 530 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 531 !MBB1->back().getDesc().isBarrier() && 532 !MBB2->back().getDesc().isBarrier()) 533 ++EffectiveTailLen; 534 535 // Check if the common tail is long enough to be worthwhile. 536 if (EffectiveTailLen >= minCommonTailLength) 537 return true; 538 539 // If we are optimizing for code size, 2 instructions in common is enough if 540 // we don't have to split a block. At worst we will be introducing 1 new 541 // branch instruction, which is likely to be smaller than the 2 542 // instructions that would be deleted in the merge. 543 if (EffectiveTailLen >= 2 && 544 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 545 (I1 == MBB1->begin() || I2 == MBB2->begin())) 546 return true; 547 548 return false; 549} 550 551/// ComputeSameTails - Look through all the blocks in MergePotentials that have 552/// hash CurHash (guaranteed to match the last element). Build the vector 553/// SameTails of all those that have the (same) largest number of instructions 554/// in common of any pair of these blocks. SameTails entries contain an 555/// iterator into MergePotentials (from which the MachineBasicBlock can be 556/// found) and a MachineBasicBlock::iterator into that MBB indicating the 557/// instruction where the matching code sequence begins. 558/// Order of elements in SameTails is the reverse of the order in which 559/// those blocks appear in MergePotentials (where they are not necessarily 560/// consecutive). 561unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 562 unsigned minCommonTailLength, 563 MachineBasicBlock *SuccBB, 564 MachineBasicBlock *PredBB) { 565 unsigned maxCommonTailLength = 0U; 566 SameTails.clear(); 567 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 568 MPIterator HighestMPIter = prior(MergePotentials.end()); 569 for (MPIterator CurMPIter = prior(MergePotentials.end()), 570 B = MergePotentials.begin(); 571 CurMPIter != B && CurMPIter->getHash() == CurHash; 572 --CurMPIter) { 573 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) { 574 unsigned CommonTailLen; 575 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 576 minCommonTailLength, 577 CommonTailLen, TrialBBI1, TrialBBI2, 578 SuccBB, PredBB)) { 579 if (CommonTailLen > maxCommonTailLength) { 580 SameTails.clear(); 581 maxCommonTailLength = CommonTailLen; 582 HighestMPIter = CurMPIter; 583 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 584 } 585 if (HighestMPIter == CurMPIter && 586 CommonTailLen == maxCommonTailLength) 587 SameTails.push_back(SameTailElt(I, TrialBBI2)); 588 } 589 if (I == B) 590 break; 591 } 592 } 593 return maxCommonTailLength; 594} 595 596/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 597/// MergePotentials, restoring branches at ends of blocks as appropriate. 598void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 599 MachineBasicBlock *SuccBB, 600 MachineBasicBlock *PredBB) { 601 MPIterator CurMPIter, B; 602 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 603 CurMPIter->getHash() == CurHash; 604 --CurMPIter) { 605 // Put the unconditional branch back, if we need one. 606 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 607 if (SuccBB && CurMBB != PredBB) 608 FixTail(CurMBB, SuccBB, TII); 609 if (CurMPIter == B) 610 break; 611 } 612 if (CurMPIter->getHash() != CurHash) 613 CurMPIter++; 614 MergePotentials.erase(CurMPIter, MergePotentials.end()); 615} 616 617/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 618/// only of the common tail. Create a block that does by splitting one. 619unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 620 unsigned maxCommonTailLength) { 621 unsigned commonTailIndex = 0; 622 unsigned TimeEstimate = ~0U; 623 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 624 // Use PredBB if possible; that doesn't require a new branch. 625 if (SameTails[i].getBlock() == PredBB) { 626 commonTailIndex = i; 627 break; 628 } 629 // Otherwise, make a (fairly bogus) choice based on estimate of 630 // how long it will take the various blocks to execute. 631 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 632 SameTails[i].getTailStartPos()); 633 if (t <= TimeEstimate) { 634 TimeEstimate = t; 635 commonTailIndex = i; 636 } 637 } 638 639 MachineBasicBlock::iterator BBI = 640 SameTails[commonTailIndex].getTailStartPos(); 641 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 642 643 DEBUG(errs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 644 << maxCommonTailLength); 645 646 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 647 SameTails[commonTailIndex].setBlock(newMBB); 648 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 649 650 // If we split PredBB, newMBB is the new predecessor. 651 if (PredBB == MBB) 652 PredBB = newMBB; 653 654 return commonTailIndex; 655} 656 657// See if any of the blocks in MergePotentials (which all have a common single 658// successor, or all have no successor) can be tail-merged. If there is a 659// successor, any blocks in MergePotentials that are not tail-merged and 660// are not immediately before Succ must have an unconditional branch to 661// Succ added (but the predecessor/successor lists need no adjustment). 662// The lone predecessor of Succ that falls through into Succ, 663// if any, is given in PredBB. 664 665bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 666 MachineBasicBlock *PredBB) { 667 bool MadeChange = false; 668 669 // Except for the special cases below, tail-merge if there are at least 670 // this many instructions in common. 671 unsigned minCommonTailLength = TailMergeSize; 672 673 DEBUG(errs() << "\nTryTailMergeBlocks: "; 674 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 675 errs() << "BB#" << MergePotentials[i].getBlock()->getNumber() 676 << (i == e-1 ? "" : ", "); 677 errs() << "\n"; 678 if (SuccBB) { 679 errs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 680 if (PredBB) 681 errs() << " which has fall-through from BB#" 682 << PredBB->getNumber() << "\n"; 683 } 684 errs() << "Looking for common tails of at least " 685 << minCommonTailLength << " instruction" 686 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 687 ); 688 689 // Sort by hash value so that blocks with identical end sequences sort 690 // together. 691 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 692 693 // Walk through equivalence sets looking for actual exact matches. 694 while (MergePotentials.size() > 1) { 695 unsigned CurHash = MergePotentials.back().getHash(); 696 697 // Build SameTails, identifying the set of blocks with this hash code 698 // and with the maximum number of instructions in common. 699 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 700 minCommonTailLength, 701 SuccBB, PredBB); 702 703 // If we didn't find any pair that has at least minCommonTailLength 704 // instructions in common, remove all blocks with this hash code and retry. 705 if (SameTails.empty()) { 706 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 707 continue; 708 } 709 710 // If one of the blocks is the entire common tail (and not the entry 711 // block, which we can't jump to), we can treat all blocks with this same 712 // tail at once. Use PredBB if that is one of the possibilities, as that 713 // will not introduce any extra branches. 714 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()-> 715 getParent()->begin(); 716 unsigned commonTailIndex = SameTails.size(); 717 // If there are two blocks, check to see if one can be made to fall through 718 // into the other. 719 if (SameTails.size() == 2 && 720 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 721 SameTails[1].tailIsWholeBlock()) 722 commonTailIndex = 1; 723 else if (SameTails.size() == 2 && 724 SameTails[1].getBlock()->isLayoutSuccessor( 725 SameTails[0].getBlock()) && 726 SameTails[0].tailIsWholeBlock()) 727 commonTailIndex = 0; 728 else { 729 // Otherwise just pick one, favoring the fall-through predecessor if 730 // there is one. 731 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 732 MachineBasicBlock *MBB = SameTails[i].getBlock(); 733 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) 734 continue; 735 if (MBB == PredBB) { 736 commonTailIndex = i; 737 break; 738 } 739 if (SameTails[i].tailIsWholeBlock()) 740 commonTailIndex = i; 741 } 742 } 743 744 if (commonTailIndex == SameTails.size() || 745 (SameTails[commonTailIndex].getBlock() == PredBB && 746 !SameTails[commonTailIndex].tailIsWholeBlock())) { 747 // None of the blocks consist entirely of the common tail. 748 // Split a block so that one does. 749 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength); 750 } 751 752 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 753 // MBB is common tail. Adjust all other BB's to jump to this one. 754 // Traversal must be forwards so erases work. 755 DEBUG(errs() << "\nUsing common tail in BB#" << MBB->getNumber() 756 << " for "); 757 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 758 if (commonTailIndex == i) 759 continue; 760 DEBUG(errs() << "BB#" << SameTails[i].getBlock()->getNumber() 761 << (i == e-1 ? "" : ", ")); 762 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 763 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); 764 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 765 MergePotentials.erase(SameTails[i].getMPIter()); 766 } 767 DEBUG(errs() << "\n"); 768 // We leave commonTailIndex in the worklist in case there are other blocks 769 // that match it with a smaller number of instructions. 770 MadeChange = true; 771 } 772 return MadeChange; 773} 774 775bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 776 777 if (!EnableTailMerge) return false; 778 779 bool MadeChange = false; 780 781 // First find blocks with no successors. 782 MergePotentials.clear(); 783 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 784 if (I->succ_empty()) 785 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I)); 786 } 787 788 // See if we can do any tail merging on those. 789 if (MergePotentials.size() < TailMergeThreshold && 790 MergePotentials.size() >= 2) 791 MadeChange |= TryTailMergeBlocks(NULL, NULL); 792 793 // Look at blocks (IBB) with multiple predecessors (PBB). 794 // We change each predecessor to a canonical form, by 795 // (1) temporarily removing any unconditional branch from the predecessor 796 // to IBB, and 797 // (2) alter conditional branches so they branch to the other block 798 // not IBB; this may require adding back an unconditional branch to IBB 799 // later, where there wasn't one coming in. E.g. 800 // Bcc IBB 801 // fallthrough to QBB 802 // here becomes 803 // Bncc QBB 804 // with a conceptual B to IBB after that, which never actually exists. 805 // With those changes, we see whether the predecessors' tails match, 806 // and merge them if so. We change things out of canonical form and 807 // back to the way they were later in the process. (OptimizeBranches 808 // would undo some of this, but we can't use it, because we'd get into 809 // a compile-time infinite loop repeatedly doing and undoing the same 810 // transformations.) 811 812 for (MachineFunction::iterator I = next(MF.begin()), E = MF.end(); 813 I != E; ++I) { 814 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { 815 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 816 MachineBasicBlock *IBB = I; 817 MachineBasicBlock *PredBB = prior(I); 818 MergePotentials.clear(); 819 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 820 E2 = I->pred_end(); 821 P != E2; ++P) { 822 MachineBasicBlock *PBB = *P; 823 // Skip blocks that loop to themselves, can't tail merge these. 824 if (PBB == IBB) 825 continue; 826 // Visit each predecessor only once. 827 if (!UniquePreds.insert(PBB)) 828 continue; 829 MachineBasicBlock *TBB = 0, *FBB = 0; 830 SmallVector<MachineOperand, 4> Cond; 831 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 832 // Failing case: IBB is the target of a cbr, and 833 // we cannot reverse the branch. 834 SmallVector<MachineOperand, 4> NewCond(Cond); 835 if (!Cond.empty() && TBB == IBB) { 836 if (TII->ReverseBranchCondition(NewCond)) 837 continue; 838 // This is the QBB case described above 839 if (!FBB) 840 FBB = next(MachineFunction::iterator(PBB)); 841 } 842 // Failing case: the only way IBB can be reached from PBB is via 843 // exception handling. Happens for landing pads. Would be nice 844 // to have a bit in the edge so we didn't have to do all this. 845 if (IBB->isLandingPad()) { 846 MachineFunction::iterator IP = PBB; IP++; 847 MachineBasicBlock *PredNextBB = NULL; 848 if (IP != MF.end()) 849 PredNextBB = IP; 850 if (TBB == NULL) { 851 if (IBB != PredNextBB) // fallthrough 852 continue; 853 } else if (FBB) { 854 if (TBB != IBB && FBB != IBB) // cbr then ubr 855 continue; 856 } else if (Cond.empty()) { 857 if (TBB != IBB) // ubr 858 continue; 859 } else { 860 if (TBB != IBB && IBB != PredNextBB) // cbr 861 continue; 862 } 863 } 864 // Remove the unconditional branch at the end, if any. 865 if (TBB && (Cond.empty() || FBB)) { 866 TII->RemoveBranch(*PBB); 867 if (!Cond.empty()) 868 // reinsert conditional branch only, for now 869 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond); 870 } 871 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U), 872 *P)); 873 } 874 } 875 if (MergePotentials.size() >= 2) 876 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 877 // Reinsert an unconditional branch if needed. 878 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks. 879 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks 880 if (MergePotentials.size() == 1 && 881 MergePotentials.begin()->getBlock() != PredBB) 882 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 883 } 884 } 885 return MadeChange; 886} 887 888//===----------------------------------------------------------------------===// 889// Branch Optimization 890//===----------------------------------------------------------------------===// 891 892bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 893 bool MadeChange = false; 894 895 // Make sure blocks are numbered in order 896 MF.RenumberBlocks(); 897 898 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 899 MachineBasicBlock *MBB = I++; 900 MadeChange |= OptimizeBlock(MBB); 901 902 // If it is dead, remove it. 903 if (MBB->pred_empty()) { 904 RemoveDeadBlock(MBB); 905 MadeChange = true; 906 ++NumDeadBlocks; 907 } 908 } 909 return MadeChange; 910} 911 912 913/// CanFallThrough - Return true if the specified block (with the specified 914/// branch condition) can implicitly transfer control to the block after it by 915/// falling off the end of it. This should return false if it can reach the 916/// block after it, but it uses an explicit branch to do so (e.g. a table jump). 917/// 918/// True is a conservative answer. 919/// 920bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, 921 bool BranchUnAnalyzable, 922 MachineBasicBlock *TBB, 923 MachineBasicBlock *FBB, 924 const SmallVectorImpl<MachineOperand> &Cond) { 925 MachineFunction::iterator Fallthrough = CurBB; 926 ++Fallthrough; 927 // If FallthroughBlock is off the end of the function, it can't fall through. 928 if (Fallthrough == CurBB->getParent()->end()) 929 return false; 930 931 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible. 932 if (!CurBB->isSuccessor(Fallthrough)) 933 return false; 934 935 // If we couldn't analyze the branch, examine the last instruction. 936 // If the block doesn't end in a known control barrier, assume fallthrough 937 // is possible. The isPredicable check is needed because this code can be 938 // called during IfConversion, where an instruction which is normally a 939 // Barrier is predicated and thus no longer an actual control barrier. This 940 // is over-conservative though, because if an instruction isn't actually 941 // predicated we could still treat it like a barrier. 942 if (BranchUnAnalyzable) 943 return CurBB->empty() || !CurBB->back().getDesc().isBarrier() || 944 CurBB->back().getDesc().isPredicable(); 945 946 // If there is no branch, control always falls through. 947 if (TBB == 0) return true; 948 949 // If there is some explicit branch to the fallthrough block, it can obviously 950 // reach, even though the branch should get folded to fall through implicitly. 951 if (MachineFunction::iterator(TBB) == Fallthrough || 952 MachineFunction::iterator(FBB) == Fallthrough) 953 return true; 954 955 // If it's an unconditional branch to some block not the fall through, it 956 // doesn't fall through. 957 if (Cond.empty()) return false; 958 959 // Otherwise, if it is conditional and has no explicit false block, it falls 960 // through. 961 return FBB == 0; 962} 963 964/// CanFallThrough - Return true if the specified can implicitly transfer 965/// control to the block after it by falling off the end of it. This should 966/// return false if it can reach the block after it, but it uses an explicit 967/// branch to do so (e.g. a table jump). 968/// 969/// True is a conservative answer. 970/// 971bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { 972 MachineBasicBlock *TBB = 0, *FBB = 0; 973 SmallVector<MachineOperand, 4> Cond; 974 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true); 975 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); 976} 977 978/// IsBetterFallthrough - Return true if it would be clearly better to 979/// fall-through to MBB1 than to fall through into MBB2. This has to return 980/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 981/// result in infinite loops. 982static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 983 MachineBasicBlock *MBB2) { 984 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 985 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 986 // optimize branches that branch to either a return block or an assert block 987 // into a fallthrough to the return. 988 if (MBB1->empty() || MBB2->empty()) return false; 989 990 // If there is a clear successor ordering we make sure that one block 991 // will fall through to the next 992 if (MBB1->isSuccessor(MBB2)) return true; 993 if (MBB2->isSuccessor(MBB1)) return false; 994 995 MachineInstr *MBB1I = --MBB1->end(); 996 MachineInstr *MBB2I = --MBB2->end(); 997 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); 998} 999 1000/// TailDuplicateBlocks - Look for small blocks that are unconditionally 1001/// branched to and do not fall through. Tail-duplicate their instructions 1002/// into their predecessors to eliminate (dynamic) branches. 1003bool BranchFolder::TailDuplicateBlocks(MachineFunction &MF) { 1004 bool MadeChange = false; 1005
| 217 218 // See if any jump tables have become mergable or dead as the code generator 219 // did its thing. 220 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 221 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 222 if (!JTs.empty()) { 223 // Figure out how these jump tables should be merged. 224 std::vector<unsigned> JTMapping; 225 JTMapping.reserve(JTs.size()); 226 227 // We always keep the 0th jump table. 228 JTMapping.push_back(0); 229 230 // Scan the jump tables, seeing if there are any duplicates. Note that this 231 // is N^2, which should be fixed someday. 232 for (unsigned i = 1, e = JTs.size(); i != e; ++i) { 233 if (JTs[i].MBBs.empty()) 234 JTMapping.push_back(i); 235 else 236 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 237 } 238 239 // If a jump table was merge with another one, walk the function rewriting 240 // references to jump tables to reference the new JT ID's. Keep track of 241 // whether we see a jump table idx, if not, we can delete the JT. 242 BitVector JTIsLive(JTs.size()); 243 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 244 BB != E; ++BB) { 245 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 246 I != E; ++I) 247 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 248 MachineOperand &Op = I->getOperand(op); 249 if (!Op.isJTI()) continue; 250 unsigned NewIdx = JTMapping[Op.getIndex()]; 251 Op.setIndex(NewIdx); 252 253 // Remember that this JT is live. 254 JTIsLive.set(NewIdx); 255 } 256 } 257 258 // Finally, remove dead jump tables. This happens either because the 259 // indirect jump was unreachable (and thus deleted) or because the jump 260 // table was merged with some other one. 261 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 262 if (!JTIsLive.test(i)) { 263 JTI->RemoveJumpTable(i); 264 MadeChange = true; 265 } 266 } 267 268 delete RS; 269 return MadeChange; 270} 271 272//===----------------------------------------------------------------------===// 273// Tail Merging of Blocks 274//===----------------------------------------------------------------------===// 275 276/// HashMachineInstr - Compute a hash value for MI and its operands. 277static unsigned HashMachineInstr(const MachineInstr *MI) { 278 unsigned Hash = MI->getOpcode(); 279 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 280 const MachineOperand &Op = MI->getOperand(i); 281 282 // Merge in bits from the operand if easy. 283 unsigned OperandHash = 0; 284 switch (Op.getType()) { 285 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 286 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 287 case MachineOperand::MO_MachineBasicBlock: 288 OperandHash = Op.getMBB()->getNumber(); 289 break; 290 case MachineOperand::MO_FrameIndex: 291 case MachineOperand::MO_ConstantPoolIndex: 292 case MachineOperand::MO_JumpTableIndex: 293 OperandHash = Op.getIndex(); 294 break; 295 case MachineOperand::MO_GlobalAddress: 296 case MachineOperand::MO_ExternalSymbol: 297 // Global address / external symbol are too hard, don't bother, but do 298 // pull in the offset. 299 OperandHash = Op.getOffset(); 300 break; 301 default: break; 302 } 303 304 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 305 } 306 return Hash; 307} 308 309/// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks 310/// with no successors, we hash two instructions, because cross-jumping 311/// only saves code when at least two instructions are removed (since a 312/// branch must be inserted). For blocks with a successor, one of the 313/// two blocks to be tail-merged will end with a branch already, so 314/// it gains to cross-jump even for one instruction. 315static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, 316 unsigned minCommonTailLength) { 317 MachineBasicBlock::const_iterator I = MBB->end(); 318 if (I == MBB->begin()) 319 return 0; // Empty MBB. 320 321 --I; 322 unsigned Hash = HashMachineInstr(I); 323 324 if (I == MBB->begin() || minCommonTailLength == 1) 325 return Hash; // Single instr MBB. 326 327 --I; 328 // Hash in the second-to-last instruction. 329 Hash ^= HashMachineInstr(I) << 2; 330 return Hash; 331} 332 333/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 334/// of instructions they actually have in common together at their end. Return 335/// iterators for the first shared instruction in each block. 336static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 337 MachineBasicBlock *MBB2, 338 MachineBasicBlock::iterator &I1, 339 MachineBasicBlock::iterator &I2) { 340 I1 = MBB1->end(); 341 I2 = MBB2->end(); 342 343 unsigned TailLen = 0; 344 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 345 --I1; --I2; 346 if (!I1->isIdenticalTo(I2) || 347 // FIXME: This check is dubious. It's used to get around a problem where 348 // people incorrectly expect inline asm directives to remain in the same 349 // relative order. This is untenable because normal compiler 350 // optimizations (like this one) may reorder and/or merge these 351 // directives. 352 I1->getOpcode() == TargetInstrInfo::INLINEASM) { 353 ++I1; ++I2; 354 break; 355 } 356 ++TailLen; 357 } 358 return TailLen; 359} 360 361/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 362/// after it, replacing it with an unconditional branch to NewDest. This 363/// returns true if OldInst's block is modified, false if NewDest is modified. 364void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 365 MachineBasicBlock *NewDest) { 366 MachineBasicBlock *OldBB = OldInst->getParent(); 367 368 // Remove all the old successors of OldBB from the CFG. 369 while (!OldBB->succ_empty()) 370 OldBB->removeSuccessor(OldBB->succ_begin()); 371 372 // Remove all the dead instructions from the end of OldBB. 373 OldBB->erase(OldInst, OldBB->end()); 374 375 // If OldBB isn't immediately before OldBB, insert a branch to it. 376 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 377 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>()); 378 OldBB->addSuccessor(NewDest); 379 ++NumTailMerge; 380} 381 382/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 383/// MBB so that the part before the iterator falls into the part starting at the 384/// iterator. This returns the new MBB. 385MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 386 MachineBasicBlock::iterator BBI1) { 387 MachineFunction &MF = *CurMBB.getParent(); 388 389 // Create the fall-through block. 390 MachineFunction::iterator MBBI = &CurMBB; 391 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 392 CurMBB.getParent()->insert(++MBBI, NewMBB); 393 394 // Move all the successors of this block to the specified block. 395 NewMBB->transferSuccessors(&CurMBB); 396 397 // Add an edge from CurMBB to NewMBB for the fall-through. 398 CurMBB.addSuccessor(NewMBB); 399 400 // Splice the code over. 401 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 402 403 // For targets that use the register scavenger, we must maintain LiveIns. 404 if (RS) { 405 RS->enterBasicBlock(&CurMBB); 406 if (!CurMBB.empty()) 407 RS->forward(prior(CurMBB.end())); 408 BitVector RegsLiveAtExit(TRI->getNumRegs()); 409 RS->getRegsUsed(RegsLiveAtExit, false); 410 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++) 411 if (RegsLiveAtExit[i]) 412 NewMBB->addLiveIn(i); 413 } 414 415 return NewMBB; 416} 417 418/// EstimateRuntime - Make a rough estimate for how long it will take to run 419/// the specified code. 420static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 421 MachineBasicBlock::iterator E) { 422 unsigned Time = 0; 423 for (; I != E; ++I) { 424 const TargetInstrDesc &TID = I->getDesc(); 425 if (TID.isCall()) 426 Time += 10; 427 else if (TID.mayLoad() || TID.mayStore()) 428 Time += 2; 429 else 430 ++Time; 431 } 432 return Time; 433} 434 435// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 436// branches temporarily for tail merging). In the case where CurMBB ends 437// with a conditional branch to the next block, optimize by reversing the 438// test and conditionally branching to SuccMBB instead. 439static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, 440 const TargetInstrInfo *TII) { 441 MachineFunction *MF = CurMBB->getParent(); 442 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); 443 MachineBasicBlock *TBB = 0, *FBB = 0; 444 SmallVector<MachineOperand, 4> Cond; 445 if (I != MF->end() && 446 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 447 MachineBasicBlock *NextBB = I; 448 if (TBB == NextBB && !Cond.empty() && !FBB) { 449 if (!TII->ReverseBranchCondition(Cond)) { 450 TII->RemoveBranch(*CurMBB); 451 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); 452 return; 453 } 454 } 455 } 456 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>()); 457} 458 459bool 460BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { 461 if (getHash() < o.getHash()) 462 return true; 463 else if (getHash() > o.getHash()) 464 return false; 465 else if (getBlock()->getNumber() < o.getBlock()->getNumber()) 466 return true; 467 else if (getBlock()->getNumber() > o.getBlock()->getNumber()) 468 return false; 469 else { 470 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 471 // an object with itself. 472#ifndef _GLIBCXX_DEBUG 473 llvm_unreachable("Predecessor appears twice"); 474#endif 475 return false; 476 } 477} 478 479/// CountTerminators - Count the number of terminators in the given 480/// block and set I to the position of the first non-terminator, if there 481/// is one, or MBB->end() otherwise. 482static unsigned CountTerminators(MachineBasicBlock *MBB, 483 MachineBasicBlock::iterator &I) { 484 I = MBB->end(); 485 unsigned NumTerms = 0; 486 for (;;) { 487 if (I == MBB->begin()) { 488 I = MBB->end(); 489 break; 490 } 491 --I; 492 if (!I->getDesc().isTerminator()) break; 493 ++NumTerms; 494 } 495 return NumTerms; 496} 497 498/// ProfitableToMerge - Check if two machine basic blocks have a common tail 499/// and decide if it would be profitable to merge those tails. Return the 500/// length of the common tail and iterators to the first common instruction 501/// in each block. 502static bool ProfitableToMerge(MachineBasicBlock *MBB1, 503 MachineBasicBlock *MBB2, 504 unsigned minCommonTailLength, 505 unsigned &CommonTailLen, 506 MachineBasicBlock::iterator &I1, 507 MachineBasicBlock::iterator &I2, 508 MachineBasicBlock *SuccBB, 509 MachineBasicBlock *PredBB) { 510 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 511 MachineFunction *MF = MBB1->getParent(); 512 513 if (CommonTailLen == 0) 514 return false; 515 516 // It's almost always profitable to merge any number of non-terminator 517 // instructions with the block that falls through into the common successor. 518 if (MBB1 == PredBB || MBB2 == PredBB) { 519 MachineBasicBlock::iterator I; 520 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 521 if (CommonTailLen > NumTerms) 522 return true; 523 } 524 525 // If one of the blocks can be completely merged and happens to be in 526 // a position where the other could fall through into it, merge any number 527 // of instructions, because it can be done without a branch. 528 // TODO: If the blocks are not adjacent, move one of them so that they are? 529 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) 530 return true; 531 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) 532 return true; 533 534 // If both blocks have an unconditional branch temporarily stripped out, 535 // count that as an additional common instruction for the following 536 // heuristics. 537 unsigned EffectiveTailLen = CommonTailLen; 538 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 539 !MBB1->back().getDesc().isBarrier() && 540 !MBB2->back().getDesc().isBarrier()) 541 ++EffectiveTailLen; 542 543 // Check if the common tail is long enough to be worthwhile. 544 if (EffectiveTailLen >= minCommonTailLength) 545 return true; 546 547 // If we are optimizing for code size, 2 instructions in common is enough if 548 // we don't have to split a block. At worst we will be introducing 1 new 549 // branch instruction, which is likely to be smaller than the 2 550 // instructions that would be deleted in the merge. 551 if (EffectiveTailLen >= 2 && 552 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 553 (I1 == MBB1->begin() || I2 == MBB2->begin())) 554 return true; 555 556 return false; 557} 558 559/// ComputeSameTails - Look through all the blocks in MergePotentials that have 560/// hash CurHash (guaranteed to match the last element). Build the vector 561/// SameTails of all those that have the (same) largest number of instructions 562/// in common of any pair of these blocks. SameTails entries contain an 563/// iterator into MergePotentials (from which the MachineBasicBlock can be 564/// found) and a MachineBasicBlock::iterator into that MBB indicating the 565/// instruction where the matching code sequence begins. 566/// Order of elements in SameTails is the reverse of the order in which 567/// those blocks appear in MergePotentials (where they are not necessarily 568/// consecutive). 569unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 570 unsigned minCommonTailLength, 571 MachineBasicBlock *SuccBB, 572 MachineBasicBlock *PredBB) { 573 unsigned maxCommonTailLength = 0U; 574 SameTails.clear(); 575 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 576 MPIterator HighestMPIter = prior(MergePotentials.end()); 577 for (MPIterator CurMPIter = prior(MergePotentials.end()), 578 B = MergePotentials.begin(); 579 CurMPIter != B && CurMPIter->getHash() == CurHash; 580 --CurMPIter) { 581 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) { 582 unsigned CommonTailLen; 583 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 584 minCommonTailLength, 585 CommonTailLen, TrialBBI1, TrialBBI2, 586 SuccBB, PredBB)) { 587 if (CommonTailLen > maxCommonTailLength) { 588 SameTails.clear(); 589 maxCommonTailLength = CommonTailLen; 590 HighestMPIter = CurMPIter; 591 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 592 } 593 if (HighestMPIter == CurMPIter && 594 CommonTailLen == maxCommonTailLength) 595 SameTails.push_back(SameTailElt(I, TrialBBI2)); 596 } 597 if (I == B) 598 break; 599 } 600 } 601 return maxCommonTailLength; 602} 603 604/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 605/// MergePotentials, restoring branches at ends of blocks as appropriate. 606void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 607 MachineBasicBlock *SuccBB, 608 MachineBasicBlock *PredBB) { 609 MPIterator CurMPIter, B; 610 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 611 CurMPIter->getHash() == CurHash; 612 --CurMPIter) { 613 // Put the unconditional branch back, if we need one. 614 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 615 if (SuccBB && CurMBB != PredBB) 616 FixTail(CurMBB, SuccBB, TII); 617 if (CurMPIter == B) 618 break; 619 } 620 if (CurMPIter->getHash() != CurHash) 621 CurMPIter++; 622 MergePotentials.erase(CurMPIter, MergePotentials.end()); 623} 624 625/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 626/// only of the common tail. Create a block that does by splitting one. 627unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 628 unsigned maxCommonTailLength) { 629 unsigned commonTailIndex = 0; 630 unsigned TimeEstimate = ~0U; 631 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 632 // Use PredBB if possible; that doesn't require a new branch. 633 if (SameTails[i].getBlock() == PredBB) { 634 commonTailIndex = i; 635 break; 636 } 637 // Otherwise, make a (fairly bogus) choice based on estimate of 638 // how long it will take the various blocks to execute. 639 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 640 SameTails[i].getTailStartPos()); 641 if (t <= TimeEstimate) { 642 TimeEstimate = t; 643 commonTailIndex = i; 644 } 645 } 646 647 MachineBasicBlock::iterator BBI = 648 SameTails[commonTailIndex].getTailStartPos(); 649 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 650 651 DEBUG(errs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 652 << maxCommonTailLength); 653 654 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 655 SameTails[commonTailIndex].setBlock(newMBB); 656 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 657 658 // If we split PredBB, newMBB is the new predecessor. 659 if (PredBB == MBB) 660 PredBB = newMBB; 661 662 return commonTailIndex; 663} 664 665// See if any of the blocks in MergePotentials (which all have a common single 666// successor, or all have no successor) can be tail-merged. If there is a 667// successor, any blocks in MergePotentials that are not tail-merged and 668// are not immediately before Succ must have an unconditional branch to 669// Succ added (but the predecessor/successor lists need no adjustment). 670// The lone predecessor of Succ that falls through into Succ, 671// if any, is given in PredBB. 672 673bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 674 MachineBasicBlock *PredBB) { 675 bool MadeChange = false; 676 677 // Except for the special cases below, tail-merge if there are at least 678 // this many instructions in common. 679 unsigned minCommonTailLength = TailMergeSize; 680 681 DEBUG(errs() << "\nTryTailMergeBlocks: "; 682 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 683 errs() << "BB#" << MergePotentials[i].getBlock()->getNumber() 684 << (i == e-1 ? "" : ", "); 685 errs() << "\n"; 686 if (SuccBB) { 687 errs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 688 if (PredBB) 689 errs() << " which has fall-through from BB#" 690 << PredBB->getNumber() << "\n"; 691 } 692 errs() << "Looking for common tails of at least " 693 << minCommonTailLength << " instruction" 694 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 695 ); 696 697 // Sort by hash value so that blocks with identical end sequences sort 698 // together. 699 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 700 701 // Walk through equivalence sets looking for actual exact matches. 702 while (MergePotentials.size() > 1) { 703 unsigned CurHash = MergePotentials.back().getHash(); 704 705 // Build SameTails, identifying the set of blocks with this hash code 706 // and with the maximum number of instructions in common. 707 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 708 minCommonTailLength, 709 SuccBB, PredBB); 710 711 // If we didn't find any pair that has at least minCommonTailLength 712 // instructions in common, remove all blocks with this hash code and retry. 713 if (SameTails.empty()) { 714 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 715 continue; 716 } 717 718 // If one of the blocks is the entire common tail (and not the entry 719 // block, which we can't jump to), we can treat all blocks with this same 720 // tail at once. Use PredBB if that is one of the possibilities, as that 721 // will not introduce any extra branches. 722 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()-> 723 getParent()->begin(); 724 unsigned commonTailIndex = SameTails.size(); 725 // If there are two blocks, check to see if one can be made to fall through 726 // into the other. 727 if (SameTails.size() == 2 && 728 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 729 SameTails[1].tailIsWholeBlock()) 730 commonTailIndex = 1; 731 else if (SameTails.size() == 2 && 732 SameTails[1].getBlock()->isLayoutSuccessor( 733 SameTails[0].getBlock()) && 734 SameTails[0].tailIsWholeBlock()) 735 commonTailIndex = 0; 736 else { 737 // Otherwise just pick one, favoring the fall-through predecessor if 738 // there is one. 739 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 740 MachineBasicBlock *MBB = SameTails[i].getBlock(); 741 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) 742 continue; 743 if (MBB == PredBB) { 744 commonTailIndex = i; 745 break; 746 } 747 if (SameTails[i].tailIsWholeBlock()) 748 commonTailIndex = i; 749 } 750 } 751 752 if (commonTailIndex == SameTails.size() || 753 (SameTails[commonTailIndex].getBlock() == PredBB && 754 !SameTails[commonTailIndex].tailIsWholeBlock())) { 755 // None of the blocks consist entirely of the common tail. 756 // Split a block so that one does. 757 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength); 758 } 759 760 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 761 // MBB is common tail. Adjust all other BB's to jump to this one. 762 // Traversal must be forwards so erases work. 763 DEBUG(errs() << "\nUsing common tail in BB#" << MBB->getNumber() 764 << " for "); 765 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 766 if (commonTailIndex == i) 767 continue; 768 DEBUG(errs() << "BB#" << SameTails[i].getBlock()->getNumber() 769 << (i == e-1 ? "" : ", ")); 770 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 771 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); 772 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 773 MergePotentials.erase(SameTails[i].getMPIter()); 774 } 775 DEBUG(errs() << "\n"); 776 // We leave commonTailIndex in the worklist in case there are other blocks 777 // that match it with a smaller number of instructions. 778 MadeChange = true; 779 } 780 return MadeChange; 781} 782 783bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 784 785 if (!EnableTailMerge) return false; 786 787 bool MadeChange = false; 788 789 // First find blocks with no successors. 790 MergePotentials.clear(); 791 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 792 if (I->succ_empty()) 793 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I)); 794 } 795 796 // See if we can do any tail merging on those. 797 if (MergePotentials.size() < TailMergeThreshold && 798 MergePotentials.size() >= 2) 799 MadeChange |= TryTailMergeBlocks(NULL, NULL); 800 801 // Look at blocks (IBB) with multiple predecessors (PBB). 802 // We change each predecessor to a canonical form, by 803 // (1) temporarily removing any unconditional branch from the predecessor 804 // to IBB, and 805 // (2) alter conditional branches so they branch to the other block 806 // not IBB; this may require adding back an unconditional branch to IBB 807 // later, where there wasn't one coming in. E.g. 808 // Bcc IBB 809 // fallthrough to QBB 810 // here becomes 811 // Bncc QBB 812 // with a conceptual B to IBB after that, which never actually exists. 813 // With those changes, we see whether the predecessors' tails match, 814 // and merge them if so. We change things out of canonical form and 815 // back to the way they were later in the process. (OptimizeBranches 816 // would undo some of this, but we can't use it, because we'd get into 817 // a compile-time infinite loop repeatedly doing and undoing the same 818 // transformations.) 819 820 for (MachineFunction::iterator I = next(MF.begin()), E = MF.end(); 821 I != E; ++I) { 822 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { 823 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 824 MachineBasicBlock *IBB = I; 825 MachineBasicBlock *PredBB = prior(I); 826 MergePotentials.clear(); 827 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 828 E2 = I->pred_end(); 829 P != E2; ++P) { 830 MachineBasicBlock *PBB = *P; 831 // Skip blocks that loop to themselves, can't tail merge these. 832 if (PBB == IBB) 833 continue; 834 // Visit each predecessor only once. 835 if (!UniquePreds.insert(PBB)) 836 continue; 837 MachineBasicBlock *TBB = 0, *FBB = 0; 838 SmallVector<MachineOperand, 4> Cond; 839 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 840 // Failing case: IBB is the target of a cbr, and 841 // we cannot reverse the branch. 842 SmallVector<MachineOperand, 4> NewCond(Cond); 843 if (!Cond.empty() && TBB == IBB) { 844 if (TII->ReverseBranchCondition(NewCond)) 845 continue; 846 // This is the QBB case described above 847 if (!FBB) 848 FBB = next(MachineFunction::iterator(PBB)); 849 } 850 // Failing case: the only way IBB can be reached from PBB is via 851 // exception handling. Happens for landing pads. Would be nice 852 // to have a bit in the edge so we didn't have to do all this. 853 if (IBB->isLandingPad()) { 854 MachineFunction::iterator IP = PBB; IP++; 855 MachineBasicBlock *PredNextBB = NULL; 856 if (IP != MF.end()) 857 PredNextBB = IP; 858 if (TBB == NULL) { 859 if (IBB != PredNextBB) // fallthrough 860 continue; 861 } else if (FBB) { 862 if (TBB != IBB && FBB != IBB) // cbr then ubr 863 continue; 864 } else if (Cond.empty()) { 865 if (TBB != IBB) // ubr 866 continue; 867 } else { 868 if (TBB != IBB && IBB != PredNextBB) // cbr 869 continue; 870 } 871 } 872 // Remove the unconditional branch at the end, if any. 873 if (TBB && (Cond.empty() || FBB)) { 874 TII->RemoveBranch(*PBB); 875 if (!Cond.empty()) 876 // reinsert conditional branch only, for now 877 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond); 878 } 879 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U), 880 *P)); 881 } 882 } 883 if (MergePotentials.size() >= 2) 884 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 885 // Reinsert an unconditional branch if needed. 886 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks. 887 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks 888 if (MergePotentials.size() == 1 && 889 MergePotentials.begin()->getBlock() != PredBB) 890 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 891 } 892 } 893 return MadeChange; 894} 895 896//===----------------------------------------------------------------------===// 897// Branch Optimization 898//===----------------------------------------------------------------------===// 899 900bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 901 bool MadeChange = false; 902 903 // Make sure blocks are numbered in order 904 MF.RenumberBlocks(); 905 906 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 907 MachineBasicBlock *MBB = I++; 908 MadeChange |= OptimizeBlock(MBB); 909 910 // If it is dead, remove it. 911 if (MBB->pred_empty()) { 912 RemoveDeadBlock(MBB); 913 MadeChange = true; 914 ++NumDeadBlocks; 915 } 916 } 917 return MadeChange; 918} 919 920 921/// CanFallThrough - Return true if the specified block (with the specified 922/// branch condition) can implicitly transfer control to the block after it by 923/// falling off the end of it. This should return false if it can reach the 924/// block after it, but it uses an explicit branch to do so (e.g. a table jump). 925/// 926/// True is a conservative answer. 927/// 928bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, 929 bool BranchUnAnalyzable, 930 MachineBasicBlock *TBB, 931 MachineBasicBlock *FBB, 932 const SmallVectorImpl<MachineOperand> &Cond) { 933 MachineFunction::iterator Fallthrough = CurBB; 934 ++Fallthrough; 935 // If FallthroughBlock is off the end of the function, it can't fall through. 936 if (Fallthrough == CurBB->getParent()->end()) 937 return false; 938 939 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible. 940 if (!CurBB->isSuccessor(Fallthrough)) 941 return false; 942 943 // If we couldn't analyze the branch, examine the last instruction. 944 // If the block doesn't end in a known control barrier, assume fallthrough 945 // is possible. The isPredicable check is needed because this code can be 946 // called during IfConversion, where an instruction which is normally a 947 // Barrier is predicated and thus no longer an actual control barrier. This 948 // is over-conservative though, because if an instruction isn't actually 949 // predicated we could still treat it like a barrier. 950 if (BranchUnAnalyzable) 951 return CurBB->empty() || !CurBB->back().getDesc().isBarrier() || 952 CurBB->back().getDesc().isPredicable(); 953 954 // If there is no branch, control always falls through. 955 if (TBB == 0) return true; 956 957 // If there is some explicit branch to the fallthrough block, it can obviously 958 // reach, even though the branch should get folded to fall through implicitly. 959 if (MachineFunction::iterator(TBB) == Fallthrough || 960 MachineFunction::iterator(FBB) == Fallthrough) 961 return true; 962 963 // If it's an unconditional branch to some block not the fall through, it 964 // doesn't fall through. 965 if (Cond.empty()) return false; 966 967 // Otherwise, if it is conditional and has no explicit false block, it falls 968 // through. 969 return FBB == 0; 970} 971 972/// CanFallThrough - Return true if the specified can implicitly transfer 973/// control to the block after it by falling off the end of it. This should 974/// return false if it can reach the block after it, but it uses an explicit 975/// branch to do so (e.g. a table jump). 976/// 977/// True is a conservative answer. 978/// 979bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { 980 MachineBasicBlock *TBB = 0, *FBB = 0; 981 SmallVector<MachineOperand, 4> Cond; 982 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true); 983 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); 984} 985 986/// IsBetterFallthrough - Return true if it would be clearly better to 987/// fall-through to MBB1 than to fall through into MBB2. This has to return 988/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 989/// result in infinite loops. 990static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 991 MachineBasicBlock *MBB2) { 992 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 993 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 994 // optimize branches that branch to either a return block or an assert block 995 // into a fallthrough to the return. 996 if (MBB1->empty() || MBB2->empty()) return false; 997 998 // If there is a clear successor ordering we make sure that one block 999 // will fall through to the next 1000 if (MBB1->isSuccessor(MBB2)) return true; 1001 if (MBB2->isSuccessor(MBB1)) return false; 1002 1003 MachineInstr *MBB1I = --MBB1->end(); 1004 MachineInstr *MBB2I = --MBB2->end(); 1005 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); 1006} 1007 1008/// TailDuplicateBlocks - Look for small blocks that are unconditionally 1009/// branched to and do not fall through. Tail-duplicate their instructions 1010/// into their predecessors to eliminate (dynamic) branches. 1011bool BranchFolder::TailDuplicateBlocks(MachineFunction &MF) { 1012 bool MadeChange = false; 1013
|
1006 // Make sure blocks are numbered in order 1007 MF.RenumberBlocks(); 1008
| |
1009 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 1010 MachineBasicBlock *MBB = I++; 1011 1012 // Only duplicate blocks that end with unconditional branches. 1013 if (CanFallThrough(MBB)) 1014 continue; 1015 1016 MadeChange |= TailDuplicate(MBB, MF); 1017 1018 // If it is dead, remove it. 1019 if (MBB->pred_empty()) {
| 1014 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 1015 MachineBasicBlock *MBB = I++; 1016 1017 // Only duplicate blocks that end with unconditional branches. 1018 if (CanFallThrough(MBB)) 1019 continue; 1020 1021 MadeChange |= TailDuplicate(MBB, MF); 1022 1023 // If it is dead, remove it. 1024 if (MBB->pred_empty()) {
|
| 1025 NumInstrDups -= MBB->size();
|
1020 RemoveDeadBlock(MBB); 1021 MadeChange = true; 1022 ++NumDeadBlocks; 1023 } 1024 } 1025 return MadeChange; 1026} 1027 1028/// TailDuplicate - If it is profitable, duplicate TailBB's contents in each 1029/// of its predecessors. 1030bool BranchFolder::TailDuplicate(MachineBasicBlock *TailBB, 1031 MachineFunction &MF) { 1032 // Don't try to tail-duplicate single-block loops. 1033 if (TailBB->isSuccessor(TailBB)) 1034 return false; 1035 1036 // Set the limit on the number of instructions to duplicate, with a default 1037 // of one less than the tail-merge threshold. When optimizing for size, 1038 // duplicate only one, because one branch instruction can be eliminated to 1039 // compensate for the duplication. 1040 unsigned MaxDuplicateCount = 1041 MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) ? 1042 1 : TII->TailDuplicationLimit(*TailBB, TailMergeSize - 1); 1043 1044 // Check the instructions in the block to determine whether tail-duplication 1045 // is invalid or unlikely to be profitable. 1046 unsigned i = 0; 1047 bool HasCall = false; 1048 for (MachineBasicBlock::iterator I = TailBB->begin(); 1049 I != TailBB->end(); ++I, ++i) { 1050 // Non-duplicable things shouldn't be tail-duplicated. 1051 if (I->getDesc().isNotDuplicable()) return false; 1052 // Don't duplicate more than the threshold. 1053 if (i == MaxDuplicateCount) return false; 1054 // Remember if we saw a call. 1055 if (I->getDesc().isCall()) HasCall = true; 1056 } 1057 // Heuristically, don't tail-duplicate calls if it would expand code size, 1058 // as it's less likely to be worth the extra cost. 1059 if (i > 1 && HasCall) 1060 return false; 1061 1062 // Iterate through all the unique predecessors and tail-duplicate this 1063 // block into them, if possible. Copying the list ahead of time also 1064 // avoids trouble with the predecessor list reallocating. 1065 bool Changed = false; 1066 SmallSetVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(), 1067 TailBB->pred_end()); 1068 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(), 1069 PE = Preds.end(); PI != PE; ++PI) { 1070 MachineBasicBlock *PredBB = *PI; 1071 1072 assert(TailBB != PredBB && 1073 "Single-block loop should have been rejected earlier!"); 1074 if (PredBB->succ_size() > 1) continue; 1075 1076 MachineBasicBlock *PredTBB, *PredFBB; 1077 SmallVector<MachineOperand, 4> PredCond; 1078 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)) 1079 continue; 1080 if (!PredCond.empty()) 1081 continue; 1082 // EH edges are ignored by AnalyzeBranch. 1083 if (PredBB->succ_size() != 1) 1084 continue; 1085 // Don't duplicate into a fall-through predecessor (at least for now). 1086 if (PredBB->isLayoutSuccessor(TailBB) && CanFallThrough(PredBB)) 1087 continue; 1088 1089 DEBUG(errs() << "\nTail-duplicating into PredBB: " << *PredBB 1090 << "From Succ: " << *TailBB); 1091 1092 // Remove PredBB's unconditional branch. 1093 TII->RemoveBranch(*PredBB); 1094 // Clone the contents of TailBB into PredBB. 1095 for (MachineBasicBlock::iterator I = TailBB->begin(), E = TailBB->end(); 1096 I != E; ++I) { 1097 MachineInstr *NewMI = MF.CloneMachineInstr(I); 1098 PredBB->insert(PredBB->end(), NewMI); 1099 }
| 1026 RemoveDeadBlock(MBB); 1027 MadeChange = true; 1028 ++NumDeadBlocks; 1029 } 1030 } 1031 return MadeChange; 1032} 1033 1034/// TailDuplicate - If it is profitable, duplicate TailBB's contents in each 1035/// of its predecessors. 1036bool BranchFolder::TailDuplicate(MachineBasicBlock *TailBB, 1037 MachineFunction &MF) { 1038 // Don't try to tail-duplicate single-block loops. 1039 if (TailBB->isSuccessor(TailBB)) 1040 return false; 1041 1042 // Set the limit on the number of instructions to duplicate, with a default 1043 // of one less than the tail-merge threshold. When optimizing for size, 1044 // duplicate only one, because one branch instruction can be eliminated to 1045 // compensate for the duplication. 1046 unsigned MaxDuplicateCount = 1047 MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) ? 1048 1 : TII->TailDuplicationLimit(*TailBB, TailMergeSize - 1); 1049 1050 // Check the instructions in the block to determine whether tail-duplication 1051 // is invalid or unlikely to be profitable. 1052 unsigned i = 0; 1053 bool HasCall = false; 1054 for (MachineBasicBlock::iterator I = TailBB->begin(); 1055 I != TailBB->end(); ++I, ++i) { 1056 // Non-duplicable things shouldn't be tail-duplicated. 1057 if (I->getDesc().isNotDuplicable()) return false; 1058 // Don't duplicate more than the threshold. 1059 if (i == MaxDuplicateCount) return false; 1060 // Remember if we saw a call. 1061 if (I->getDesc().isCall()) HasCall = true; 1062 } 1063 // Heuristically, don't tail-duplicate calls if it would expand code size, 1064 // as it's less likely to be worth the extra cost. 1065 if (i > 1 && HasCall) 1066 return false; 1067 1068 // Iterate through all the unique predecessors and tail-duplicate this 1069 // block into them, if possible. Copying the list ahead of time also 1070 // avoids trouble with the predecessor list reallocating. 1071 bool Changed = false; 1072 SmallSetVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(), 1073 TailBB->pred_end()); 1074 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(), 1075 PE = Preds.end(); PI != PE; ++PI) { 1076 MachineBasicBlock *PredBB = *PI; 1077 1078 assert(TailBB != PredBB && 1079 "Single-block loop should have been rejected earlier!"); 1080 if (PredBB->succ_size() > 1) continue; 1081 1082 MachineBasicBlock *PredTBB, *PredFBB; 1083 SmallVector<MachineOperand, 4> PredCond; 1084 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)) 1085 continue; 1086 if (!PredCond.empty()) 1087 continue; 1088 // EH edges are ignored by AnalyzeBranch. 1089 if (PredBB->succ_size() != 1) 1090 continue; 1091 // Don't duplicate into a fall-through predecessor (at least for now). 1092 if (PredBB->isLayoutSuccessor(TailBB) && CanFallThrough(PredBB)) 1093 continue; 1094 1095 DEBUG(errs() << "\nTail-duplicating into PredBB: " << *PredBB 1096 << "From Succ: " << *TailBB); 1097 1098 // Remove PredBB's unconditional branch. 1099 TII->RemoveBranch(*PredBB); 1100 // Clone the contents of TailBB into PredBB. 1101 for (MachineBasicBlock::iterator I = TailBB->begin(), E = TailBB->end(); 1102 I != E; ++I) { 1103 MachineInstr *NewMI = MF.CloneMachineInstr(I); 1104 PredBB->insert(PredBB->end(), NewMI); 1105 }
|
| 1106 NumInstrDups += TailBB->size() - 1; // subtract one for removed branch
|
1100 1101 // Update the CFG. 1102 PredBB->removeSuccessor(PredBB->succ_begin()); 1103 assert(PredBB->succ_empty() && 1104 "TailDuplicate called on block with multiple successors!"); 1105 for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(), 1106 E = TailBB->succ_end(); I != E; ++I) 1107 PredBB->addSuccessor(*I); 1108 1109 Changed = true;
| 1107 1108 // Update the CFG. 1109 PredBB->removeSuccessor(PredBB->succ_begin()); 1110 assert(PredBB->succ_empty() && 1111 "TailDuplicate called on block with multiple successors!"); 1112 for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(), 1113 E = TailBB->succ_end(); I != E; ++I) 1114 PredBB->addSuccessor(*I); 1115 1116 Changed = true;
|
| 1117 ++NumTailDups;
|
1110 } 1111 1112 // If TailBB was duplicated into all its predecessors except for the prior 1113 // block, which falls through unconditionally, move the contents of this 1114 // block into the prior block. 1115 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(TailBB)); 1116 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1117 SmallVector<MachineOperand, 4> PriorCond; 1118 bool PriorUnAnalyzable = 1119 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1120 // This has to check PrevBB->succ_size() because EH edges are ignored by 1121 // AnalyzeBranch. 1122 if (!PriorUnAnalyzable && PriorCond.empty() && !PriorTBB && 1123 TailBB->pred_size() == 1 && PrevBB.succ_size() == 1 && 1124 !TailBB->hasAddressTaken()) { 1125 DEBUG(errs() << "\nMerging into block: " << PrevBB 1126 << "From MBB: " << *TailBB); 1127 PrevBB.splice(PrevBB.end(), TailBB, TailBB->begin(), TailBB->end()); 1128 PrevBB.removeSuccessor(PrevBB.succ_begin());; 1129 assert(PrevBB.succ_empty()); 1130 PrevBB.transferSuccessors(TailBB); 1131 Changed = true; 1132 } 1133 1134 return Changed; 1135} 1136 1137/// OptimizeBlock - Analyze and optimize control flow related to the specified 1138/// block. This is never called on the entry block. 1139bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 1140 bool MadeChange = false; 1141 MachineFunction &MF = *MBB->getParent(); 1142ReoptimizeBlock: 1143 1144 MachineFunction::iterator FallThrough = MBB; 1145 ++FallThrough; 1146 1147 // If this block is empty, make everyone use its fall-through, not the block 1148 // explicitly. Landing pads should not do this since the landing-pad table 1149 // points to this block. Blocks with their addresses taken shouldn't be 1150 // optimized away. 1151 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 1152 // Dead block? Leave for cleanup later. 1153 if (MBB->pred_empty()) return MadeChange; 1154 1155 if (FallThrough == MF.end()) { 1156 // TODO: Simplify preds to not branch here if possible! 1157 } else { 1158 // Rewrite all predecessors of the old block to go to the fallthrough 1159 // instead. 1160 while (!MBB->pred_empty()) { 1161 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1162 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 1163 } 1164 // If MBB was the target of a jump table, update jump tables to go to the 1165 // fallthrough instead. 1166 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough); 1167 MadeChange = true; 1168 } 1169 return MadeChange; 1170 } 1171 1172 // Check to see if we can simplify the terminator of the block before this 1173 // one. 1174 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 1175 1176 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1177 SmallVector<MachineOperand, 4> PriorCond; 1178 bool PriorUnAnalyzable = 1179 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1180 if (!PriorUnAnalyzable) { 1181 // If the CFG for the prior block has extra edges, remove them. 1182 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 1183 !PriorCond.empty()); 1184 1185 // If the previous branch is conditional and both conditions go to the same 1186 // destination, remove the branch, replacing it with an unconditional one or 1187 // a fall-through. 1188 if (PriorTBB && PriorTBB == PriorFBB) { 1189 TII->RemoveBranch(PrevBB); 1190 PriorCond.clear(); 1191 if (PriorTBB != MBB) 1192 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 1193 MadeChange = true; 1194 ++NumBranchOpts; 1195 goto ReoptimizeBlock; 1196 } 1197 1198 // If the previous block unconditionally falls through to this block and 1199 // this block has no other predecessors, move the contents of this block 1200 // into the prior block. This doesn't usually happen when SimplifyCFG 1201 // has been used, but it can happen if tail merging splits a fall-through 1202 // predecessor of a block. 1203 // This has to check PrevBB->succ_size() because EH edges are ignored by 1204 // AnalyzeBranch. 1205 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1206 PrevBB.succ_size() == 1 && 1207 !MBB->hasAddressTaken()) { 1208 DEBUG(errs() << "\nMerging into block: " << PrevBB 1209 << "From MBB: " << *MBB); 1210 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1211 PrevBB.removeSuccessor(PrevBB.succ_begin());; 1212 assert(PrevBB.succ_empty()); 1213 PrevBB.transferSuccessors(MBB); 1214 MadeChange = true; 1215 return MadeChange; 1216 } 1217 1218 // If the previous branch *only* branches to *this* block (conditional or 1219 // not) remove the branch. 1220 if (PriorTBB == MBB && PriorFBB == 0) { 1221 TII->RemoveBranch(PrevBB); 1222 MadeChange = true; 1223 ++NumBranchOpts; 1224 goto ReoptimizeBlock; 1225 } 1226 1227 // If the prior block branches somewhere else on the condition and here if 1228 // the condition is false, remove the uncond second branch. 1229 if (PriorFBB == MBB) { 1230 TII->RemoveBranch(PrevBB); 1231 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 1232 MadeChange = true; 1233 ++NumBranchOpts; 1234 goto ReoptimizeBlock; 1235 } 1236 1237 // If the prior block branches here on true and somewhere else on false, and 1238 // if the branch condition is reversible, reverse the branch to create a 1239 // fall-through. 1240 if (PriorTBB == MBB) { 1241 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1242 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1243 TII->RemoveBranch(PrevBB); 1244 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 1245 MadeChange = true; 1246 ++NumBranchOpts; 1247 goto ReoptimizeBlock; 1248 } 1249 } 1250 1251 // If this block has no successors (e.g. it is a return block or ends with 1252 // a call to a no-return function like abort or __cxa_throw) and if the pred 1253 // falls through into this block, and if it would otherwise fall through 1254 // into the block after this, move this block to the end of the function. 1255 // 1256 // We consider it more likely that execution will stay in the function (e.g. 1257 // due to loops) than it is to exit it. This asserts in loops etc, moving 1258 // the assert condition out of the loop body. 1259 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 1260 MachineFunction::iterator(PriorTBB) == FallThrough && 1261 !CanFallThrough(MBB)) { 1262 bool DoTransform = true; 1263 1264 // We have to be careful that the succs of PredBB aren't both no-successor 1265 // blocks. If neither have successors and if PredBB is the second from 1266 // last block in the function, we'd just keep swapping the two blocks for 1267 // last. Only do the swap if one is clearly better to fall through than 1268 // the other. 1269 if (FallThrough == --MF.end() && 1270 !IsBetterFallthrough(PriorTBB, MBB)) 1271 DoTransform = false; 1272 1273 // We don't want to do this transformation if we have control flow like: 1274 // br cond BB2 1275 // BB1: 1276 // .. 1277 // jmp BBX 1278 // BB2: 1279 // .. 1280 // ret 1281 // 1282 // In this case, we could actually be moving the return block *into* a 1283 // loop! 1284 if (DoTransform && !MBB->succ_empty() && 1285 (!CanFallThrough(PriorTBB) || PriorTBB->empty())) 1286 DoTransform = false; 1287 1288 1289 if (DoTransform) { 1290 // Reverse the branch so we will fall through on the previous true cond. 1291 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1292 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1293 DEBUG(errs() << "\nMoving MBB: " << *MBB 1294 << "To make fallthrough to: " << *PriorTBB << "\n"); 1295 1296 TII->RemoveBranch(PrevBB); 1297 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 1298 1299 // Move this block to the end of the function. 1300 MBB->moveAfter(--MF.end()); 1301 MadeChange = true; 1302 ++NumBranchOpts; 1303 return MadeChange; 1304 } 1305 } 1306 } 1307 } 1308 1309 // Analyze the branch in the current block. 1310 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1311 SmallVector<MachineOperand, 4> CurCond; 1312 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1313 if (!CurUnAnalyzable) { 1314 // If the CFG for the prior block has extra edges, remove them. 1315 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1316 1317 // If this is a two-way branch, and the FBB branches to this block, reverse 1318 // the condition so the single-basic-block loop is faster. Instead of: 1319 // Loop: xxx; jcc Out; jmp Loop 1320 // we want: 1321 // Loop: xxx; jncc Loop; jmp Out 1322 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1323 SmallVector<MachineOperand, 4> NewCond(CurCond); 1324 if (!TII->ReverseBranchCondition(NewCond)) { 1325 TII->RemoveBranch(*MBB); 1326 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 1327 MadeChange = true; 1328 ++NumBranchOpts; 1329 goto ReoptimizeBlock; 1330 } 1331 } 1332 1333 // If this branch is the only thing in its block, see if we can forward 1334 // other blocks across it. 1335 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1336 MBB->begin()->getDesc().isBranch() && CurTBB != MBB && 1337 !MBB->hasAddressTaken()) { 1338 // This block may contain just an unconditional branch. Because there can 1339 // be 'non-branch terminators' in the block, try removing the branch and 1340 // then seeing if the block is empty. 1341 TII->RemoveBranch(*MBB); 1342 1343 // If this block is just an unconditional branch to CurTBB, we can 1344 // usually completely eliminate the block. The only case we cannot 1345 // completely eliminate the block is when the block before this one 1346 // falls through into MBB and we can't understand the prior block's branch 1347 // condition. 1348 if (MBB->empty()) { 1349 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); 1350 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1351 !PrevBB.isSuccessor(MBB)) { 1352 // If the prior block falls through into us, turn it into an 1353 // explicit branch to us to make updates simpler. 1354 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1355 PriorTBB != MBB && PriorFBB != MBB) { 1356 if (PriorTBB == 0) { 1357 assert(PriorCond.empty() && PriorFBB == 0 && 1358 "Bad branch analysis"); 1359 PriorTBB = MBB; 1360 } else { 1361 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1362 PriorFBB = MBB; 1363 } 1364 TII->RemoveBranch(PrevBB); 1365 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 1366 } 1367 1368 // Iterate through all the predecessors, revectoring each in-turn. 1369 size_t PI = 0; 1370 bool DidChange = false; 1371 bool HasBranchToSelf = false; 1372 while(PI != MBB->pred_size()) { 1373 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1374 if (PMBB == MBB) { 1375 // If this block has an uncond branch to itself, leave it. 1376 ++PI; 1377 HasBranchToSelf = true; 1378 } else { 1379 DidChange = true; 1380 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1381 // If this change resulted in PMBB ending in a conditional 1382 // branch where both conditions go to the same destination, 1383 // change this to an unconditional branch (and fix the CFG). 1384 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1385 SmallVector<MachineOperand, 4> NewCurCond; 1386 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1387 NewCurFBB, NewCurCond, true); 1388 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1389 TII->RemoveBranch(*PMBB); 1390 NewCurCond.clear(); 1391 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond); 1392 MadeChange = true; 1393 ++NumBranchOpts; 1394 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); 1395 } 1396 } 1397 } 1398 1399 // Change any jumptables to go to the new MBB. 1400 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB); 1401 if (DidChange) { 1402 ++NumBranchOpts; 1403 MadeChange = true; 1404 if (!HasBranchToSelf) return MadeChange; 1405 } 1406 } 1407 } 1408 1409 // Add the branch back if the block is more than just an uncond branch. 1410 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 1411 } 1412 } 1413 1414 // If the prior block doesn't fall through into this block, and if this 1415 // block doesn't fall through into some other block, see if we can find a 1416 // place to move this block where a fall-through will happen. 1417 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable, 1418 PriorTBB, PriorFBB, PriorCond)) { 1419 1420 // Now we know that there was no fall-through into this block, check to 1421 // see if it has a fall-through into its successor. 1422 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, 1423 CurCond); 1424 1425 if (!MBB->isLandingPad()) { 1426 // Check all the predecessors of this block. If one of them has no fall 1427 // throughs, move this block right after it. 1428 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1429 E = MBB->pred_end(); PI != E; ++PI) { 1430 // Analyze the branch at the end of the pred. 1431 MachineBasicBlock *PredBB = *PI; 1432 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1433 MachineBasicBlock *PredTBB, *PredFBB; 1434 SmallVector<MachineOperand, 4> PredCond; 1435 if (PredBB != MBB && !CanFallThrough(PredBB) && 1436 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) 1437 && (!CurFallsThru || !CurTBB || !CurFBB) 1438 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1439 // If the current block doesn't fall through, just move it. 1440 // If the current block can fall through and does not end with a 1441 // conditional branch, we need to append an unconditional jump to 1442 // the (current) next block. To avoid a possible compile-time 1443 // infinite loop, move blocks only backward in this case. 1444 // Also, if there are already 2 branches here, we cannot add a third; 1445 // this means we have the case 1446 // Bcc next 1447 // B elsewhere 1448 // next: 1449 if (CurFallsThru) { 1450 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); 1451 CurCond.clear(); 1452 TII->InsertBranch(*MBB, NextBB, 0, CurCond); 1453 } 1454 MBB->moveAfter(PredBB); 1455 MadeChange = true; 1456 goto ReoptimizeBlock; 1457 } 1458 } 1459 } 1460 1461 if (!CurFallsThru) { 1462 // Check all successors to see if we can move this block before it. 1463 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1464 E = MBB->succ_end(); SI != E; ++SI) { 1465 // Analyze the branch at the end of the block before the succ. 1466 MachineBasicBlock *SuccBB = *SI; 1467 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1468 1469 // If this block doesn't already fall-through to that successor, and if 1470 // the succ doesn't already have a block that can fall through into it, 1471 // and if the successor isn't an EH destination, we can arrange for the 1472 // fallthrough to happen. 1473 if (SuccBB != MBB && &*SuccPrev != MBB && 1474 !CanFallThrough(SuccPrev) && !CurUnAnalyzable && 1475 !SuccBB->isLandingPad()) { 1476 MBB->moveBefore(SuccBB); 1477 MadeChange = true; 1478 goto ReoptimizeBlock; 1479 } 1480 } 1481 1482 // Okay, there is no really great place to put this block. If, however, 1483 // the block before this one would be a fall-through if this block were 1484 // removed, move this block to the end of the function. 1485 MachineBasicBlock *PrevTBB, *PrevFBB; 1486 SmallVector<MachineOperand, 4> PrevCond; 1487 if (FallThrough != MF.end() && 1488 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1489 PrevBB.isSuccessor(FallThrough)) { 1490 MBB->moveAfter(--MF.end()); 1491 MadeChange = true; 1492 return MadeChange; 1493 } 1494 } 1495 } 1496 1497 return MadeChange; 1498}
| 1118 } 1119 1120 // If TailBB was duplicated into all its predecessors except for the prior 1121 // block, which falls through unconditionally, move the contents of this 1122 // block into the prior block. 1123 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(TailBB)); 1124 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1125 SmallVector<MachineOperand, 4> PriorCond; 1126 bool PriorUnAnalyzable = 1127 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1128 // This has to check PrevBB->succ_size() because EH edges are ignored by 1129 // AnalyzeBranch. 1130 if (!PriorUnAnalyzable && PriorCond.empty() && !PriorTBB && 1131 TailBB->pred_size() == 1 && PrevBB.succ_size() == 1 && 1132 !TailBB->hasAddressTaken()) { 1133 DEBUG(errs() << "\nMerging into block: " << PrevBB 1134 << "From MBB: " << *TailBB); 1135 PrevBB.splice(PrevBB.end(), TailBB, TailBB->begin(), TailBB->end()); 1136 PrevBB.removeSuccessor(PrevBB.succ_begin());; 1137 assert(PrevBB.succ_empty()); 1138 PrevBB.transferSuccessors(TailBB); 1139 Changed = true; 1140 } 1141 1142 return Changed; 1143} 1144 1145/// OptimizeBlock - Analyze and optimize control flow related to the specified 1146/// block. This is never called on the entry block. 1147bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 1148 bool MadeChange = false; 1149 MachineFunction &MF = *MBB->getParent(); 1150ReoptimizeBlock: 1151 1152 MachineFunction::iterator FallThrough = MBB; 1153 ++FallThrough; 1154 1155 // If this block is empty, make everyone use its fall-through, not the block 1156 // explicitly. Landing pads should not do this since the landing-pad table 1157 // points to this block. Blocks with their addresses taken shouldn't be 1158 // optimized away. 1159 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 1160 // Dead block? Leave for cleanup later. 1161 if (MBB->pred_empty()) return MadeChange; 1162 1163 if (FallThrough == MF.end()) { 1164 // TODO: Simplify preds to not branch here if possible! 1165 } else { 1166 // Rewrite all predecessors of the old block to go to the fallthrough 1167 // instead. 1168 while (!MBB->pred_empty()) { 1169 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1170 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 1171 } 1172 // If MBB was the target of a jump table, update jump tables to go to the 1173 // fallthrough instead. 1174 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough); 1175 MadeChange = true; 1176 } 1177 return MadeChange; 1178 } 1179 1180 // Check to see if we can simplify the terminator of the block before this 1181 // one. 1182 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 1183 1184 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1185 SmallVector<MachineOperand, 4> PriorCond; 1186 bool PriorUnAnalyzable = 1187 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1188 if (!PriorUnAnalyzable) { 1189 // If the CFG for the prior block has extra edges, remove them. 1190 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 1191 !PriorCond.empty()); 1192 1193 // If the previous branch is conditional and both conditions go to the same 1194 // destination, remove the branch, replacing it with an unconditional one or 1195 // a fall-through. 1196 if (PriorTBB && PriorTBB == PriorFBB) { 1197 TII->RemoveBranch(PrevBB); 1198 PriorCond.clear(); 1199 if (PriorTBB != MBB) 1200 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 1201 MadeChange = true; 1202 ++NumBranchOpts; 1203 goto ReoptimizeBlock; 1204 } 1205 1206 // If the previous block unconditionally falls through to this block and 1207 // this block has no other predecessors, move the contents of this block 1208 // into the prior block. This doesn't usually happen when SimplifyCFG 1209 // has been used, but it can happen if tail merging splits a fall-through 1210 // predecessor of a block. 1211 // This has to check PrevBB->succ_size() because EH edges are ignored by 1212 // AnalyzeBranch. 1213 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1214 PrevBB.succ_size() == 1 && 1215 !MBB->hasAddressTaken()) { 1216 DEBUG(errs() << "\nMerging into block: " << PrevBB 1217 << "From MBB: " << *MBB); 1218 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1219 PrevBB.removeSuccessor(PrevBB.succ_begin());; 1220 assert(PrevBB.succ_empty()); 1221 PrevBB.transferSuccessors(MBB); 1222 MadeChange = true; 1223 return MadeChange; 1224 } 1225 1226 // If the previous branch *only* branches to *this* block (conditional or 1227 // not) remove the branch. 1228 if (PriorTBB == MBB && PriorFBB == 0) { 1229 TII->RemoveBranch(PrevBB); 1230 MadeChange = true; 1231 ++NumBranchOpts; 1232 goto ReoptimizeBlock; 1233 } 1234 1235 // If the prior block branches somewhere else on the condition and here if 1236 // the condition is false, remove the uncond second branch. 1237 if (PriorFBB == MBB) { 1238 TII->RemoveBranch(PrevBB); 1239 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 1240 MadeChange = true; 1241 ++NumBranchOpts; 1242 goto ReoptimizeBlock; 1243 } 1244 1245 // If the prior block branches here on true and somewhere else on false, and 1246 // if the branch condition is reversible, reverse the branch to create a 1247 // fall-through. 1248 if (PriorTBB == MBB) { 1249 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1250 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1251 TII->RemoveBranch(PrevBB); 1252 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 1253 MadeChange = true; 1254 ++NumBranchOpts; 1255 goto ReoptimizeBlock; 1256 } 1257 } 1258 1259 // If this block has no successors (e.g. it is a return block or ends with 1260 // a call to a no-return function like abort or __cxa_throw) and if the pred 1261 // falls through into this block, and if it would otherwise fall through 1262 // into the block after this, move this block to the end of the function. 1263 // 1264 // We consider it more likely that execution will stay in the function (e.g. 1265 // due to loops) than it is to exit it. This asserts in loops etc, moving 1266 // the assert condition out of the loop body. 1267 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 1268 MachineFunction::iterator(PriorTBB) == FallThrough && 1269 !CanFallThrough(MBB)) { 1270 bool DoTransform = true; 1271 1272 // We have to be careful that the succs of PredBB aren't both no-successor 1273 // blocks. If neither have successors and if PredBB is the second from 1274 // last block in the function, we'd just keep swapping the two blocks for 1275 // last. Only do the swap if one is clearly better to fall through than 1276 // the other. 1277 if (FallThrough == --MF.end() && 1278 !IsBetterFallthrough(PriorTBB, MBB)) 1279 DoTransform = false; 1280 1281 // We don't want to do this transformation if we have control flow like: 1282 // br cond BB2 1283 // BB1: 1284 // .. 1285 // jmp BBX 1286 // BB2: 1287 // .. 1288 // ret 1289 // 1290 // In this case, we could actually be moving the return block *into* a 1291 // loop! 1292 if (DoTransform && !MBB->succ_empty() && 1293 (!CanFallThrough(PriorTBB) || PriorTBB->empty())) 1294 DoTransform = false; 1295 1296 1297 if (DoTransform) { 1298 // Reverse the branch so we will fall through on the previous true cond. 1299 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1300 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1301 DEBUG(errs() << "\nMoving MBB: " << *MBB 1302 << "To make fallthrough to: " << *PriorTBB << "\n"); 1303 1304 TII->RemoveBranch(PrevBB); 1305 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 1306 1307 // Move this block to the end of the function. 1308 MBB->moveAfter(--MF.end()); 1309 MadeChange = true; 1310 ++NumBranchOpts; 1311 return MadeChange; 1312 } 1313 } 1314 } 1315 } 1316 1317 // Analyze the branch in the current block. 1318 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1319 SmallVector<MachineOperand, 4> CurCond; 1320 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1321 if (!CurUnAnalyzable) { 1322 // If the CFG for the prior block has extra edges, remove them. 1323 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1324 1325 // If this is a two-way branch, and the FBB branches to this block, reverse 1326 // the condition so the single-basic-block loop is faster. Instead of: 1327 // Loop: xxx; jcc Out; jmp Loop 1328 // we want: 1329 // Loop: xxx; jncc Loop; jmp Out 1330 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1331 SmallVector<MachineOperand, 4> NewCond(CurCond); 1332 if (!TII->ReverseBranchCondition(NewCond)) { 1333 TII->RemoveBranch(*MBB); 1334 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 1335 MadeChange = true; 1336 ++NumBranchOpts; 1337 goto ReoptimizeBlock; 1338 } 1339 } 1340 1341 // If this branch is the only thing in its block, see if we can forward 1342 // other blocks across it. 1343 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1344 MBB->begin()->getDesc().isBranch() && CurTBB != MBB && 1345 !MBB->hasAddressTaken()) { 1346 // This block may contain just an unconditional branch. Because there can 1347 // be 'non-branch terminators' in the block, try removing the branch and 1348 // then seeing if the block is empty. 1349 TII->RemoveBranch(*MBB); 1350 1351 // If this block is just an unconditional branch to CurTBB, we can 1352 // usually completely eliminate the block. The only case we cannot 1353 // completely eliminate the block is when the block before this one 1354 // falls through into MBB and we can't understand the prior block's branch 1355 // condition. 1356 if (MBB->empty()) { 1357 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); 1358 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1359 !PrevBB.isSuccessor(MBB)) { 1360 // If the prior block falls through into us, turn it into an 1361 // explicit branch to us to make updates simpler. 1362 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1363 PriorTBB != MBB && PriorFBB != MBB) { 1364 if (PriorTBB == 0) { 1365 assert(PriorCond.empty() && PriorFBB == 0 && 1366 "Bad branch analysis"); 1367 PriorTBB = MBB; 1368 } else { 1369 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1370 PriorFBB = MBB; 1371 } 1372 TII->RemoveBranch(PrevBB); 1373 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 1374 } 1375 1376 // Iterate through all the predecessors, revectoring each in-turn. 1377 size_t PI = 0; 1378 bool DidChange = false; 1379 bool HasBranchToSelf = false; 1380 while(PI != MBB->pred_size()) { 1381 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1382 if (PMBB == MBB) { 1383 // If this block has an uncond branch to itself, leave it. 1384 ++PI; 1385 HasBranchToSelf = true; 1386 } else { 1387 DidChange = true; 1388 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1389 // If this change resulted in PMBB ending in a conditional 1390 // branch where both conditions go to the same destination, 1391 // change this to an unconditional branch (and fix the CFG). 1392 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1393 SmallVector<MachineOperand, 4> NewCurCond; 1394 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1395 NewCurFBB, NewCurCond, true); 1396 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1397 TII->RemoveBranch(*PMBB); 1398 NewCurCond.clear(); 1399 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond); 1400 MadeChange = true; 1401 ++NumBranchOpts; 1402 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); 1403 } 1404 } 1405 } 1406 1407 // Change any jumptables to go to the new MBB. 1408 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB); 1409 if (DidChange) { 1410 ++NumBranchOpts; 1411 MadeChange = true; 1412 if (!HasBranchToSelf) return MadeChange; 1413 } 1414 } 1415 } 1416 1417 // Add the branch back if the block is more than just an uncond branch. 1418 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 1419 } 1420 } 1421 1422 // If the prior block doesn't fall through into this block, and if this 1423 // block doesn't fall through into some other block, see if we can find a 1424 // place to move this block where a fall-through will happen. 1425 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable, 1426 PriorTBB, PriorFBB, PriorCond)) { 1427 1428 // Now we know that there was no fall-through into this block, check to 1429 // see if it has a fall-through into its successor. 1430 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, 1431 CurCond); 1432 1433 if (!MBB->isLandingPad()) { 1434 // Check all the predecessors of this block. If one of them has no fall 1435 // throughs, move this block right after it. 1436 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1437 E = MBB->pred_end(); PI != E; ++PI) { 1438 // Analyze the branch at the end of the pred. 1439 MachineBasicBlock *PredBB = *PI; 1440 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1441 MachineBasicBlock *PredTBB, *PredFBB; 1442 SmallVector<MachineOperand, 4> PredCond; 1443 if (PredBB != MBB && !CanFallThrough(PredBB) && 1444 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) 1445 && (!CurFallsThru || !CurTBB || !CurFBB) 1446 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1447 // If the current block doesn't fall through, just move it. 1448 // If the current block can fall through and does not end with a 1449 // conditional branch, we need to append an unconditional jump to 1450 // the (current) next block. To avoid a possible compile-time 1451 // infinite loop, move blocks only backward in this case. 1452 // Also, if there are already 2 branches here, we cannot add a third; 1453 // this means we have the case 1454 // Bcc next 1455 // B elsewhere 1456 // next: 1457 if (CurFallsThru) { 1458 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); 1459 CurCond.clear(); 1460 TII->InsertBranch(*MBB, NextBB, 0, CurCond); 1461 } 1462 MBB->moveAfter(PredBB); 1463 MadeChange = true; 1464 goto ReoptimizeBlock; 1465 } 1466 } 1467 } 1468 1469 if (!CurFallsThru) { 1470 // Check all successors to see if we can move this block before it. 1471 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1472 E = MBB->succ_end(); SI != E; ++SI) { 1473 // Analyze the branch at the end of the block before the succ. 1474 MachineBasicBlock *SuccBB = *SI; 1475 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1476 1477 // If this block doesn't already fall-through to that successor, and if 1478 // the succ doesn't already have a block that can fall through into it, 1479 // and if the successor isn't an EH destination, we can arrange for the 1480 // fallthrough to happen. 1481 if (SuccBB != MBB && &*SuccPrev != MBB && 1482 !CanFallThrough(SuccPrev) && !CurUnAnalyzable && 1483 !SuccBB->isLandingPad()) { 1484 MBB->moveBefore(SuccBB); 1485 MadeChange = true; 1486 goto ReoptimizeBlock; 1487 } 1488 } 1489 1490 // Okay, there is no really great place to put this block. If, however, 1491 // the block before this one would be a fall-through if this block were 1492 // removed, move this block to the end of the function. 1493 MachineBasicBlock *PrevTBB, *PrevFBB; 1494 SmallVector<MachineOperand, 4> PrevCond; 1495 if (FallThrough != MF.end() && 1496 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1497 PrevBB.isSuccessor(FallThrough)) { 1498 MBB->moveAfter(--MF.end()); 1499 MadeChange = true; 1500 return MadeChange; 1501 } 1502 } 1503 } 1504 1505 return MadeChange; 1506}
|