1 //===-- PTXInstrInfo.cpp - PTX Instruction Information --------------------===// 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 file contains the PTX implementation of the TargetInstrInfo class. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #define DEBUG_TYPE "ptx-instrinfo" 15 16 #include "PTXInstrInfo.h" 17 #include "PTX.h" 18 #include "llvm/CodeGen/MachineInstrBuilder.h" 19 #include "llvm/CodeGen/MachineRegisterInfo.h" 20 #include "llvm/CodeGen/SelectionDAG.h" 21 #include "llvm/CodeGen/SelectionDAGNodes.h" 22 #include "llvm/Support/Debug.h" 23 #include "llvm/Support/TargetRegistry.h" 24 #include "llvm/Support/raw_ostream.h" 25 26 #define GET_INSTRINFO_CTOR 27 #include "PTXGenInstrInfo.inc" 28 29 using namespace llvm; 30 31 PTXInstrInfo::PTXInstrInfo(PTXTargetMachine &_TM) 32 : PTXGenInstrInfo(), 33 RI(_TM, *this), TM(_TM) {} 34 35 static const struct map_entry { 36 const TargetRegisterClass *cls; 37 const int opcode; 38 } map[] = { 39 { &PTX::RegI16RegClass, PTX::MOVU16rr }, 40 { &PTX::RegI32RegClass, PTX::MOVU32rr }, 41 { &PTX::RegI64RegClass, PTX::MOVU64rr }, 42 { &PTX::RegF32RegClass, PTX::MOVF32rr }, 43 { &PTX::RegF64RegClass, PTX::MOVF64rr }, 44 { &PTX::RegPredRegClass, PTX::MOVPREDrr } 45 }; 46 47 void PTXInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 48 MachineBasicBlock::iterator I, DebugLoc DL, 49 unsigned DstReg, unsigned SrcReg, 50 bool KillSrc) const { 51 52 const MachineRegisterInfo& MRI = MBB.getParent()->getRegInfo(); 53 //assert(MRI.getRegClass(SrcReg) == MRI.getRegClass(DstReg) && 54 // "Invalid register copy between two register classes"); 55 56 for (int i = 0, e = sizeof(map)/sizeof(map[0]); i != e; ++i) { 57 if (map[i].cls == MRI.getRegClass(DstReg)) { 58 const MCInstrDesc &MCID = get(map[i].opcode); 59 MachineInstr *MI = BuildMI(MBB, I, DL, MCID, DstReg). 60 addReg(SrcReg, getKillRegState(KillSrc)); 61 AddDefaultPredicate(MI); 62 return; 63 } 64 } 65 66 llvm_unreachable("Impossible reg-to-reg copy"); 67 } 68 69 bool PTXInstrInfo::copyRegToReg(MachineBasicBlock &MBB, 70 MachineBasicBlock::iterator I, 71 unsigned DstReg, unsigned SrcReg, 72 const TargetRegisterClass *DstRC, 73 const TargetRegisterClass *SrcRC, 74 DebugLoc DL) const { 75 if (DstRC != SrcRC) 76 return false; 77 78 for (int i = 0, e = sizeof(map)/sizeof(map[0]); i != e; ++ i) 79 if (DstRC == map[i].cls) { 80 const MCInstrDesc &MCID = get(map[i].opcode); 81 MachineInstr *MI = BuildMI(MBB, I, DL, MCID, DstReg).addReg(SrcReg); 82 AddDefaultPredicate(MI); 83 return true; 84 } 85 86 return false; 87 } 88 89 bool PTXInstrInfo::isMoveInstr(const MachineInstr& MI, 90 unsigned &SrcReg, unsigned &DstReg, 91 unsigned &SrcSubIdx, unsigned &DstSubIdx) const { 92 switch (MI.getOpcode()) { 93 default: 94 return false; 95 case PTX::MOVU16rr: 96 case PTX::MOVU32rr: 97 case PTX::MOVU64rr: 98 case PTX::MOVF32rr: 99 case PTX::MOVF64rr: 100 case PTX::MOVPREDrr: 101 assert(MI.getNumOperands() >= 2 && 102 MI.getOperand(0).isReg() && MI.getOperand(1).isReg() && 103 "Invalid register-register move instruction"); 104 SrcSubIdx = DstSubIdx = 0; // No sub-registers 105 DstReg = MI.getOperand(0).getReg(); 106 SrcReg = MI.getOperand(1).getReg(); 107 return true; 108 } 109 } 110 111 // predicate support 112 113 bool PTXInstrInfo::isPredicated(const MachineInstr *MI) const { 114 int i = MI->findFirstPredOperandIdx(); 115 return i != -1 && MI->getOperand(i).getReg() != PTX::NoRegister; 116 } 117 118 bool PTXInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const { 119 return !isPredicated(MI) && MI->isTerminator(); 120 } 121 122 bool PTXInstrInfo:: 123 PredicateInstruction(MachineInstr *MI, 124 const SmallVectorImpl<MachineOperand> &Pred) const { 125 if (Pred.size() < 2) 126 llvm_unreachable("lesser than 2 predicate operands are provided"); 127 128 int i = MI->findFirstPredOperandIdx(); 129 if (i == -1) 130 llvm_unreachable("missing predicate operand"); 131 132 MI->getOperand(i).setReg(Pred[0].getReg()); 133 MI->getOperand(i+1).setImm(Pred[1].getImm()); 134 135 return true; 136 } 137 138 bool PTXInstrInfo:: 139 SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1, 140 const SmallVectorImpl<MachineOperand> &Pred2) const { 141 const MachineOperand &PredReg1 = Pred1[0]; 142 const MachineOperand &PredReg2 = Pred2[0]; 143 if (PredReg1.getReg() != PredReg2.getReg()) 144 return false; 145 146 const MachineOperand &PredOp1 = Pred1[1]; 147 const MachineOperand &PredOp2 = Pred2[1]; 148 if (PredOp1.getImm() != PredOp2.getImm()) 149 return false; 150 151 return true; 152 } 153 154 bool PTXInstrInfo:: 155 DefinesPredicate(MachineInstr *MI, 156 std::vector<MachineOperand> &Pred) const { 157 // If an instruction sets a predicate register, it defines a predicate. 158 159 // TODO supprot 5-operand format of setp instruction 160 161 if (MI->getNumOperands() < 1) 162 return false; 163 164 const MachineOperand &MO = MI->getOperand(0); 165 166 if (!MO.isReg() || RI.getRegClass(MO.getReg()) != &PTX::RegPredRegClass) 167 return false; 168 169 Pred.push_back(MO); 170 Pred.push_back(MachineOperand::CreateImm(PTXPredicate::None)); 171 return true; 172 } 173 174 // branch support 175 176 bool PTXInstrInfo:: 177 AnalyzeBranch(MachineBasicBlock &MBB, 178 MachineBasicBlock *&TBB, 179 MachineBasicBlock *&FBB, 180 SmallVectorImpl<MachineOperand> &Cond, 181 bool AllowModify) const { 182 // TODO implement cases when AllowModify is true 183 184 if (MBB.empty()) 185 return true; 186 187 MachineBasicBlock::iterator iter = MBB.end(); 188 const MachineInstr& instLast1 = *--iter; 189 // for special case that MBB has only 1 instruction 190 const bool IsSizeOne = MBB.size() == 1; 191 // if IsSizeOne is true, *--iter and instLast2 are invalid 192 // we put a dummy value in instLast2 and desc2 since they are used 193 const MachineInstr& instLast2 = IsSizeOne ? instLast1 : *--iter; 194 195 DEBUG(dbgs() << "\n"); 196 DEBUG(dbgs() << "AnalyzeBranch: opcode: " << instLast1.getOpcode() << "\n"); 197 DEBUG(dbgs() << "AnalyzeBranch: MBB: " << MBB.getName().str() << "\n"); 198 DEBUG(dbgs() << "AnalyzeBranch: TBB: " << TBB << "\n"); 199 DEBUG(dbgs() << "AnalyzeBranch: FBB: " << FBB << "\n"); 200 201 // this block ends with no branches 202 if (!IsAnyKindOfBranch(instLast1)) { 203 DEBUG(dbgs() << "AnalyzeBranch: ends with no branch\n"); 204 return false; 205 } 206 207 // this block ends with only an unconditional branch 208 if (instLast1.isUnconditionalBranch() && 209 // when IsSizeOne is true, it "absorbs" the evaluation of instLast2 210 (IsSizeOne || !IsAnyKindOfBranch(instLast2))) { 211 DEBUG(dbgs() << "AnalyzeBranch: ends with only uncond branch\n"); 212 TBB = GetBranchTarget(instLast1); 213 return false; 214 } 215 216 // this block ends with a conditional branch and 217 // it falls through to a successor block 218 if (instLast1.isConditionalBranch() && 219 IsAnySuccessorAlsoLayoutSuccessor(MBB)) { 220 DEBUG(dbgs() << "AnalyzeBranch: ends with cond branch and fall through\n"); 221 TBB = GetBranchTarget(instLast1); 222 int i = instLast1.findFirstPredOperandIdx(); 223 Cond.push_back(instLast1.getOperand(i)); 224 Cond.push_back(instLast1.getOperand(i+1)); 225 return false; 226 } 227 228 // when IsSizeOne is true, we are done 229 if (IsSizeOne) 230 return true; 231 232 // this block ends with a conditional branch 233 // followed by an unconditional branch 234 if (instLast2.isConditionalBranch() && 235 instLast1.isUnconditionalBranch()) { 236 DEBUG(dbgs() << "AnalyzeBranch: ends with cond and uncond branch\n"); 237 TBB = GetBranchTarget(instLast2); 238 FBB = GetBranchTarget(instLast1); 239 int i = instLast2.findFirstPredOperandIdx(); 240 Cond.push_back(instLast2.getOperand(i)); 241 Cond.push_back(instLast2.getOperand(i+1)); 242 return false; 243 } 244 245 // branch cannot be understood 246 DEBUG(dbgs() << "AnalyzeBranch: cannot be understood\n"); 247 return true; 248 } 249 250 unsigned PTXInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const { 251 unsigned count = 0; 252 while (!MBB.empty()) 253 if (IsAnyKindOfBranch(MBB.back())) { 254 MBB.pop_back(); 255 ++count; 256 } else 257 break; 258 DEBUG(dbgs() << "RemoveBranch: MBB: " << MBB.getName().str() << "\n"); 259 DEBUG(dbgs() << "RemoveBranch: remove " << count << " branch inst\n"); 260 return count; 261 } 262 263 unsigned PTXInstrInfo:: 264 InsertBranch(MachineBasicBlock &MBB, 265 MachineBasicBlock *TBB, 266 MachineBasicBlock *FBB, 267 const SmallVectorImpl<MachineOperand> &Cond, 268 DebugLoc DL) const { 269 DEBUG(dbgs() << "InsertBranch: MBB: " << MBB.getName().str() << "\n"); 270 DEBUG(if (TBB) dbgs() << "InsertBranch: TBB: " << TBB->getName().str() 271 << "\n"; 272 else dbgs() << "InsertBranch: TBB: (NULL)\n"); 273 DEBUG(if (FBB) dbgs() << "InsertBranch: FBB: " << FBB->getName().str() 274 << "\n"; 275 else dbgs() << "InsertBranch: FBB: (NULL)\n"); 276 DEBUG(dbgs() << "InsertBranch: Cond size: " << Cond.size() << "\n"); 277 278 assert(TBB && "TBB is NULL"); 279 280 if (FBB) { 281 BuildMI(&MBB, DL, get(PTX::BRAdp)) 282 .addMBB(TBB).addReg(Cond[0].getReg()).addImm(Cond[1].getImm()); 283 BuildMI(&MBB, DL, get(PTX::BRAd)) 284 .addMBB(FBB).addReg(PTX::NoRegister).addImm(PTXPredicate::None); 285 return 2; 286 } else if (Cond.size()) { 287 BuildMI(&MBB, DL, get(PTX::BRAdp)) 288 .addMBB(TBB).addReg(Cond[0].getReg()).addImm(Cond[1].getImm()); 289 return 1; 290 } else { 291 BuildMI(&MBB, DL, get(PTX::BRAd)) 292 .addMBB(TBB).addReg(PTX::NoRegister).addImm(PTXPredicate::None); 293 return 1; 294 } 295 } 296 297 // Memory operand folding for spills 298 void PTXInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, 299 MachineBasicBlock::iterator MII, 300 unsigned SrcReg, bool isKill, int FrameIdx, 301 const TargetRegisterClass *RC, 302 const TargetRegisterInfo *TRI) const { 303 llvm_unreachable("storeRegToStackSlot should not be called for PTX"); 304 } 305 306 void PTXInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, 307 MachineBasicBlock::iterator MII, 308 unsigned DestReg, int FrameIdx, 309 const TargetRegisterClass *RC, 310 const TargetRegisterInfo *TRI) const { 311 llvm_unreachable("loadRegFromStackSlot should not be called for PTX"); 312 } 313 314 // static helper routines 315 316 MachineSDNode *PTXInstrInfo:: 317 GetPTXMachineNode(SelectionDAG *DAG, unsigned Opcode, 318 DebugLoc dl, EVT VT, SDValue Op1) { 319 SDValue predReg = DAG->getRegister(PTX::NoRegister, MVT::i1); 320 SDValue predOp = DAG->getTargetConstant(PTXPredicate::None, MVT::i32); 321 SDValue ops[] = { Op1, predReg, predOp }; 322 return DAG->getMachineNode(Opcode, dl, VT, ops, array_lengthof(ops)); 323 } 324 325 MachineSDNode *PTXInstrInfo:: 326 GetPTXMachineNode(SelectionDAG *DAG, unsigned Opcode, 327 DebugLoc dl, EVT VT, SDValue Op1, SDValue Op2) { 328 SDValue predReg = DAG->getRegister(PTX::NoRegister, MVT::i1); 329 SDValue predOp = DAG->getTargetConstant(PTXPredicate::None, MVT::i32); 330 SDValue ops[] = { Op1, Op2, predReg, predOp }; 331 return DAG->getMachineNode(Opcode, dl, VT, ops, array_lengthof(ops)); 332 } 333 334 void PTXInstrInfo::AddDefaultPredicate(MachineInstr *MI) { 335 if (MI->findFirstPredOperandIdx() == -1) { 336 MI->addOperand(MachineOperand::CreateReg(PTX::NoRegister, /*IsDef=*/false)); 337 MI->addOperand(MachineOperand::CreateImm(PTXPredicate::None)); 338 } 339 } 340 341 bool PTXInstrInfo::IsAnyKindOfBranch(const MachineInstr& inst) { 342 return inst.isTerminator() || inst.isBranch() || inst.isIndirectBranch(); 343 } 344 345 bool PTXInstrInfo:: 346 IsAnySuccessorAlsoLayoutSuccessor(const MachineBasicBlock& MBB) { 347 for (MachineBasicBlock::const_succ_iterator 348 i = MBB.succ_begin(), e = MBB.succ_end(); i != e; ++i) 349 if (MBB.isLayoutSuccessor((const MachineBasicBlock*) &*i)) 350 return true; 351 return false; 352 } 353 354 MachineBasicBlock *PTXInstrInfo::GetBranchTarget(const MachineInstr& inst) { 355 // FIXME So far all branch instructions put destination in 1st operand 356 const MachineOperand& target = inst.getOperand(0); 357 assert(target.isMBB() && "FIXME: detect branch target operand"); 358 return target.getMBB(); 359 } 360