1 //===-- ARMLowOverheadLoops.cpp - CodeGen Low-overhead Loops ---*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 /// \file 9 /// Finalize v8.1-m low-overhead loops by converting the associated pseudo 10 /// instructions into machine operations. 11 /// The expectation is that the loop contains three pseudo instructions: 12 /// - t2*LoopStart - placed in the preheader or pre-preheader. The do-loop 13 /// form should be in the preheader, whereas the while form should be in the 14 /// preheaders only predecessor. TODO: Could DoLoopStart get moved into the 15 /// pre-preheader? 16 /// - t2LoopDec - placed within in the loop body. 17 /// - t2LoopEnd - the loop latch terminator. 18 /// 19 //===----------------------------------------------------------------------===// 20 21 #include "ARM.h" 22 #include "ARMBaseInstrInfo.h" 23 #include "ARMBaseRegisterInfo.h" 24 #include "ARMBasicBlockInfo.h" 25 #include "ARMSubtarget.h" 26 #include "llvm/CodeGen/MachineFunctionPass.h" 27 #include "llvm/CodeGen/MachineLoopInfo.h" 28 #include "llvm/CodeGen/MachineRegisterInfo.h" 29 30 using namespace llvm; 31 32 #define DEBUG_TYPE "arm-low-overhead-loops" 33 #define ARM_LOW_OVERHEAD_LOOPS_NAME "ARM Low Overhead Loops pass" 34 35 namespace { 36 37 class ARMLowOverheadLoops : public MachineFunctionPass { 38 const ARMBaseInstrInfo *TII = nullptr; 39 MachineRegisterInfo *MRI = nullptr; 40 std::unique_ptr<ARMBasicBlockUtils> BBUtils = nullptr; 41 42 public: 43 static char ID; 44 45 ARMLowOverheadLoops() : MachineFunctionPass(ID) { } 46 47 void getAnalysisUsage(AnalysisUsage &AU) const override { 48 AU.setPreservesCFG(); 49 AU.addRequired<MachineLoopInfo>(); 50 MachineFunctionPass::getAnalysisUsage(AU); 51 } 52 53 bool runOnMachineFunction(MachineFunction &MF) override; 54 55 bool ProcessLoop(MachineLoop *ML); 56 57 void RevertWhile(MachineInstr *MI) const; 58 59 void RevertLoopDec(MachineInstr *MI) const; 60 61 void RevertLoopEnd(MachineInstr *MI) const; 62 63 void Expand(MachineLoop *ML, MachineInstr *Start, 64 MachineInstr *Dec, MachineInstr *End, bool Revert); 65 66 MachineFunctionProperties getRequiredProperties() const override { 67 return MachineFunctionProperties().set( 68 MachineFunctionProperties::Property::NoVRegs); 69 } 70 71 StringRef getPassName() const override { 72 return ARM_LOW_OVERHEAD_LOOPS_NAME; 73 } 74 }; 75 } 76 77 char ARMLowOverheadLoops::ID = 0; 78 79 INITIALIZE_PASS(ARMLowOverheadLoops, DEBUG_TYPE, ARM_LOW_OVERHEAD_LOOPS_NAME, 80 false, false) 81 82 bool ARMLowOverheadLoops::runOnMachineFunction(MachineFunction &MF) { 83 if (!static_cast<const ARMSubtarget&>(MF.getSubtarget()).hasLOB()) 84 return false; 85 86 LLVM_DEBUG(dbgs() << "ARM Loops on " << MF.getName() << " ------------- \n"); 87 88 auto &MLI = getAnalysis<MachineLoopInfo>(); 89 MRI = &MF.getRegInfo(); 90 TII = static_cast<const ARMBaseInstrInfo*>( 91 MF.getSubtarget().getInstrInfo()); 92 BBUtils = std::unique_ptr<ARMBasicBlockUtils>(new ARMBasicBlockUtils(MF)); 93 BBUtils->computeAllBlockSizes(); 94 95 bool Changed = false; 96 for (auto ML : MLI) { 97 if (!ML->getParentLoop()) 98 Changed |= ProcessLoop(ML); 99 } 100 return Changed; 101 } 102 103 bool ARMLowOverheadLoops::ProcessLoop(MachineLoop *ML) { 104 105 bool Changed = false; 106 107 // Process inner loops first. 108 for (auto I = ML->begin(), E = ML->end(); I != E; ++I) 109 Changed |= ProcessLoop(*I); 110 111 LLVM_DEBUG(dbgs() << "ARM Loops: Processing " << *ML); 112 113 auto IsLoopStart = [](MachineInstr &MI) { 114 return MI.getOpcode() == ARM::t2DoLoopStart || 115 MI.getOpcode() == ARM::t2WhileLoopStart; 116 }; 117 118 // Search the given block for a loop start instruction. If one isn't found, 119 // and there's only one predecessor block, search that one too. 120 std::function<MachineInstr*(MachineBasicBlock*)> SearchForStart = 121 [&IsLoopStart, &SearchForStart](MachineBasicBlock *MBB) -> MachineInstr* { 122 for (auto &MI : *MBB) { 123 if (IsLoopStart(MI)) 124 return &MI; 125 } 126 if (MBB->pred_size() == 1) 127 return SearchForStart(*MBB->pred_begin()); 128 return nullptr; 129 }; 130 131 MachineInstr *Start = nullptr; 132 MachineInstr *Dec = nullptr; 133 MachineInstr *End = nullptr; 134 bool Revert = false; 135 136 // Search the preheader for the start intrinsic, or look through the 137 // predecessors of the header to find exactly one set.iterations intrinsic. 138 // FIXME: I don't see why we shouldn't be supporting multiple predecessors 139 // with potentially multiple set.loop.iterations, so we need to enable this. 140 if (auto *Preheader = ML->getLoopPreheader()) { 141 Start = SearchForStart(Preheader); 142 } else { 143 LLVM_DEBUG(dbgs() << "ARM Loops: Failed to find loop preheader!\n" 144 << " - Performing manual predecessor search.\n"); 145 MachineBasicBlock *Pred = nullptr; 146 for (auto *MBB : ML->getHeader()->predecessors()) { 147 if (!ML->contains(MBB)) { 148 if (Pred) { 149 LLVM_DEBUG(dbgs() << " - Found multiple out-of-loop preds.\n"); 150 Start = nullptr; 151 break; 152 } 153 Pred = MBB; 154 Start = SearchForStart(MBB); 155 } 156 } 157 } 158 159 // Find the low-overhead loop components and decide whether or not to fall 160 // back to a normal loop. 161 for (auto *MBB : reverse(ML->getBlocks())) { 162 for (auto &MI : *MBB) { 163 if (MI.getOpcode() == ARM::t2LoopDec) 164 Dec = &MI; 165 else if (MI.getOpcode() == ARM::t2LoopEnd) 166 End = &MI; 167 else if (MI.getDesc().isCall()) 168 // TODO: Though the call will require LE to execute again, does this 169 // mean we should revert? Always executing LE hopefully should be 170 // faster than performing a sub,cmp,br or even subs,br. 171 Revert = true; 172 173 if (!Dec) 174 continue; 175 176 // If we find that we load/store LR between LoopDec and LoopEnd, expect 177 // that the decremented value has been spilled to the stack. Because 178 // this value isn't actually going to be produced until the latch, by LE, 179 // we would need to generate a real sub. The value is also likely to be 180 // reloaded for use of LoopEnd - in which in case we'd need to perform 181 // an add because it gets negated again by LE! The other option is to 182 // then generate the other form of LE which doesn't perform the sub. 183 if (MI.mayLoad() || MI.mayStore()) 184 Revert = 185 MI.getOperand(0).isReg() && MI.getOperand(0).getReg() == ARM::LR; 186 } 187 188 if (Dec && End && Revert) 189 break; 190 } 191 192 if (!Start && !Dec && !End) { 193 LLVM_DEBUG(dbgs() << "ARM Loops: Not a low-overhead loop.\n"); 194 return Changed; 195 } if (!(Start && Dec && End)) { 196 report_fatal_error("Failed to find all loop components"); 197 } 198 199 if (!End->getOperand(1).isMBB() || 200 End->getOperand(1).getMBB() != ML->getHeader()) 201 report_fatal_error("Expected LoopEnd to target Loop Header"); 202 203 // The LE instructions has 12-bits for the label offset. 204 if (!BBUtils->isBBInRange(End, ML->getHeader(), 4096)) { 205 LLVM_DEBUG(dbgs() << "ARM Loops: Too large for a low-overhead loop!\n"); 206 Revert = true; 207 } 208 209 LLVM_DEBUG(dbgs() << "ARM Loops:\n - Found Loop Start: " << *Start 210 << " - Found Loop Dec: " << *Dec 211 << " - Found Loop End: " << *End); 212 213 Expand(ML, Start, Dec, End, Revert); 214 return true; 215 } 216 217 // WhileLoopStart holds the exit block, so produce a cmp lr, 0 and then a 218 // beq that branches to the exit branch. 219 // FIXME: Need to check that we're not trashing the CPSR when generating the 220 // cmp. We could also try to generate a cbz if the value in LR is also in 221 // another low register. 222 void ARMLowOverheadLoops::RevertWhile(MachineInstr *MI) const { 223 LLVM_DEBUG(dbgs() << "ARM Loops: Reverting to cmp: " << *MI); 224 MachineBasicBlock *MBB = MI->getParent(); 225 MachineInstrBuilder MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), 226 TII->get(ARM::t2CMPri)); 227 MIB.addReg(ARM::LR); 228 MIB.addImm(0); 229 MIB.addImm(ARMCC::AL); 230 MIB.addReg(ARM::CPSR); 231 232 // TODO: Try to use tBcc instead 233 MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(ARM::t2Bcc)); 234 MIB.add(MI->getOperand(1)); // branch target 235 MIB.addImm(ARMCC::EQ); // condition code 236 MIB.addReg(ARM::CPSR); 237 MI->eraseFromParent(); 238 } 239 240 // TODO: Check flags so that we can possibly generate a tSubs or tSub. 241 void ARMLowOverheadLoops::RevertLoopDec(MachineInstr *MI) const { 242 LLVM_DEBUG(dbgs() << "ARM Loops: Reverting to sub: " << *MI); 243 MachineBasicBlock *MBB = MI->getParent(); 244 MachineInstrBuilder MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), 245 TII->get(ARM::t2SUBri)); 246 MIB.addDef(ARM::LR); 247 MIB.add(MI->getOperand(1)); 248 MIB.add(MI->getOperand(2)); 249 MIB.addImm(ARMCC::AL); 250 MIB.addReg(0); 251 MIB.addReg(0); 252 MI->eraseFromParent(); 253 } 254 255 // Generate a subs, or sub and cmp, and a branch instead of an LE. 256 // FIXME: Need to check that we're not trashing the CPSR when generating 257 // the cmp. 258 void ARMLowOverheadLoops::RevertLoopEnd(MachineInstr *MI) const { 259 LLVM_DEBUG(dbgs() << "ARM Loops: Reverting to cmp, br: " << *MI); 260 261 // Create cmp 262 MachineBasicBlock *MBB = MI->getParent(); 263 MachineInstrBuilder MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), 264 TII->get(ARM::t2CMPri)); 265 MIB.addReg(ARM::LR); 266 MIB.addImm(0); 267 MIB.addImm(ARMCC::AL); 268 MIB.addReg(ARM::CPSR); 269 270 // TODO Try to use tBcc instead. 271 // Create bne 272 MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(ARM::t2Bcc)); 273 MIB.add(MI->getOperand(1)); // branch target 274 MIB.addImm(ARMCC::NE); // condition code 275 MIB.addReg(ARM::CPSR); 276 MI->eraseFromParent(); 277 } 278 279 void ARMLowOverheadLoops::Expand(MachineLoop *ML, MachineInstr *Start, 280 MachineInstr *Dec, MachineInstr *End, 281 bool Revert) { 282 283 auto ExpandLoopStart = [this](MachineLoop *ML, MachineInstr *Start) { 284 // The trip count should already been held in LR since the instructions 285 // within the loop can only read and write to LR. So, there should be a 286 // mov to setup the count. WLS/DLS perform this move, so find the original 287 // and delete it - inserting WLS/DLS in its place. 288 MachineBasicBlock *MBB = Start->getParent(); 289 MachineInstr *InsertPt = Start; 290 for (auto &I : MRI->def_instructions(ARM::LR)) { 291 if (I.getParent() != MBB) 292 continue; 293 294 // Always execute. 295 if (!I.getOperand(2).isImm() || I.getOperand(2).getImm() != ARMCC::AL) 296 continue; 297 298 // Only handle move reg, if the trip count it will need moving into a reg 299 // before the setup instruction anyway. 300 if (!I.getDesc().isMoveReg() || 301 !I.getOperand(1).isIdenticalTo(Start->getOperand(0))) 302 continue; 303 InsertPt = &I; 304 break; 305 } 306 307 unsigned Opc = Start->getOpcode() == ARM::t2DoLoopStart ? 308 ARM::t2DLS : ARM::t2WLS; 309 MachineInstrBuilder MIB = 310 BuildMI(*MBB, InsertPt, InsertPt->getDebugLoc(), TII->get(Opc)); 311 312 MIB.addDef(ARM::LR); 313 MIB.add(Start->getOperand(0)); 314 if (Opc == ARM::t2WLS) 315 MIB.add(Start->getOperand(1)); 316 317 if (InsertPt != Start) 318 InsertPt->eraseFromParent(); 319 Start->eraseFromParent(); 320 LLVM_DEBUG(dbgs() << "ARM Loops: Inserted start: " << *MIB); 321 return &*MIB; 322 }; 323 324 // Combine the LoopDec and LoopEnd instructions into LE(TP). 325 auto ExpandLoopEnd = [this](MachineLoop *ML, MachineInstr *Dec, 326 MachineInstr *End) { 327 MachineBasicBlock *MBB = End->getParent(); 328 MachineInstrBuilder MIB = BuildMI(*MBB, End, End->getDebugLoc(), 329 TII->get(ARM::t2LEUpdate)); 330 MIB.addDef(ARM::LR); 331 MIB.add(End->getOperand(0)); 332 MIB.add(End->getOperand(1)); 333 LLVM_DEBUG(dbgs() << "ARM Loops: Inserted LE: " << *MIB); 334 335 End->eraseFromParent(); 336 Dec->eraseFromParent(); 337 return &*MIB; 338 }; 339 340 // TODO: We should be able to automatically remove these branches before we 341 // get here - probably by teaching analyzeBranch about the pseudo 342 // instructions. 343 // If there is an unconditional branch, after I, that just branches to the 344 // next block, remove it. 345 auto RemoveDeadBranch = [](MachineInstr *I) { 346 MachineBasicBlock *BB = I->getParent(); 347 MachineInstr *Terminator = &BB->instr_back(); 348 if (Terminator->isUnconditionalBranch() && I != Terminator) { 349 MachineBasicBlock *Succ = Terminator->getOperand(0).getMBB(); 350 if (BB->isLayoutSuccessor(Succ)) { 351 LLVM_DEBUG(dbgs() << "ARM Loops: Removing branch: " << *Terminator); 352 Terminator->eraseFromParent(); 353 } 354 } 355 }; 356 357 if (Revert) { 358 if (Start->getOpcode() == ARM::t2WhileLoopStart) 359 RevertWhile(Start); 360 else 361 Start->eraseFromParent(); 362 RevertLoopDec(Dec); 363 RevertLoopEnd(End); 364 } else { 365 Start = ExpandLoopStart(ML, Start); 366 RemoveDeadBranch(Start); 367 End = ExpandLoopEnd(ML, Dec, End); 368 RemoveDeadBranch(End); 369 } 370 } 371 372 FunctionPass *llvm::createARMLowOverheadLoopsPass() { 373 return new ARMLowOverheadLoops(); 374 } 375