1 //===-- XCoreISelLowering.cpp - XCore DAG Lowering Implementation ---------===// 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 implements the XCoreTargetLowering class. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #define DEBUG_TYPE "xcore-lower" 15 16 #include "XCoreISelLowering.h" 17 #include "XCoreMachineFunctionInfo.h" 18 #include "XCore.h" 19 #include "XCoreTargetObjectFile.h" 20 #include "XCoreTargetMachine.h" 21 #include "XCoreSubtarget.h" 22 #include "llvm/DerivedTypes.h" 23 #include "llvm/Function.h" 24 #include "llvm/Intrinsics.h" 25 #include "llvm/CallingConv.h" 26 #include "llvm/GlobalVariable.h" 27 #include "llvm/GlobalAlias.h" 28 #include "llvm/CodeGen/CallingConvLower.h" 29 #include "llvm/CodeGen/MachineFrameInfo.h" 30 #include "llvm/CodeGen/MachineFunction.h" 31 #include "llvm/CodeGen/MachineInstrBuilder.h" 32 #include "llvm/CodeGen/MachineJumpTableInfo.h" 33 #include "llvm/CodeGen/MachineRegisterInfo.h" 34 #include "llvm/CodeGen/SelectionDAGISel.h" 35 #include "llvm/CodeGen/ValueTypes.h" 36 #include "llvm/Support/Debug.h" 37 #include "llvm/Support/ErrorHandling.h" 38 #include "llvm/Support/raw_ostream.h" 39 using namespace llvm; 40 41 const char *XCoreTargetLowering:: 42 getTargetNodeName(unsigned Opcode) const 43 { 44 switch (Opcode) 45 { 46 case XCoreISD::BL : return "XCoreISD::BL"; 47 case XCoreISD::PCRelativeWrapper : return "XCoreISD::PCRelativeWrapper"; 48 case XCoreISD::DPRelativeWrapper : return "XCoreISD::DPRelativeWrapper"; 49 case XCoreISD::CPRelativeWrapper : return "XCoreISD::CPRelativeWrapper"; 50 case XCoreISD::STWSP : return "XCoreISD::STWSP"; 51 case XCoreISD::RETSP : return "XCoreISD::RETSP"; 52 case XCoreISD::LADD : return "XCoreISD::LADD"; 53 case XCoreISD::LSUB : return "XCoreISD::LSUB"; 54 case XCoreISD::LMUL : return "XCoreISD::LMUL"; 55 case XCoreISD::MACCU : return "XCoreISD::MACCU"; 56 case XCoreISD::MACCS : return "XCoreISD::MACCS"; 57 case XCoreISD::BR_JT : return "XCoreISD::BR_JT"; 58 case XCoreISD::BR_JT32 : return "XCoreISD::BR_JT32"; 59 default : return NULL; 60 } 61 } 62 63 XCoreTargetLowering::XCoreTargetLowering(XCoreTargetMachine &XTM) 64 : TargetLowering(XTM, new XCoreTargetObjectFile()), 65 TM(XTM), 66 Subtarget(*XTM.getSubtargetImpl()) { 67 68 // Set up the register classes. 69 addRegisterClass(MVT::i32, XCore::GRRegsRegisterClass); 70 71 // Compute derived properties from the register classes 72 computeRegisterProperties(); 73 74 // Division is expensive 75 setIntDivIsCheap(false); 76 77 setStackPointerRegisterToSaveRestore(XCore::SP); 78 79 setSchedulingPreference(Sched::RegPressure); 80 81 // Use i32 for setcc operations results (slt, sgt, ...). 82 setBooleanContents(ZeroOrOneBooleanContent); 83 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct? 84 85 // XCore does not have the NodeTypes below. 86 setOperationAction(ISD::BR_CC, MVT::Other, Expand); 87 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom); 88 setOperationAction(ISD::ADDC, MVT::i32, Expand); 89 setOperationAction(ISD::ADDE, MVT::i32, Expand); 90 setOperationAction(ISD::SUBC, MVT::i32, Expand); 91 setOperationAction(ISD::SUBE, MVT::i32, Expand); 92 93 // Stop the combiner recombining select and set_cc 94 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand); 95 96 // 64bit 97 setOperationAction(ISD::ADD, MVT::i64, Custom); 98 setOperationAction(ISD::SUB, MVT::i64, Custom); 99 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Custom); 100 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Custom); 101 setOperationAction(ISD::MULHS, MVT::i32, Expand); 102 setOperationAction(ISD::MULHU, MVT::i32, Expand); 103 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand); 104 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand); 105 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand); 106 107 // Bit Manipulation 108 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 109 setOperationAction(ISD::ROTL , MVT::i32, Expand); 110 setOperationAction(ISD::ROTR , MVT::i32, Expand); 111 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand); 112 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand); 113 114 setOperationAction(ISD::TRAP, MVT::Other, Legal); 115 116 // Jump tables. 117 setOperationAction(ISD::BR_JT, MVT::Other, Custom); 118 119 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 120 setOperationAction(ISD::BlockAddress, MVT::i32 , Custom); 121 122 // Thread Local Storage 123 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 124 125 // Conversion of i64 -> double produces constantpool nodes 126 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 127 128 // Loads 129 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote); 130 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote); 131 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 132 133 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand); 134 setLoadExtAction(ISD::ZEXTLOAD, MVT::i16, Expand); 135 136 // Custom expand misaligned loads / stores. 137 setOperationAction(ISD::LOAD, MVT::i32, Custom); 138 setOperationAction(ISD::STORE, MVT::i32, Custom); 139 140 // Varargs 141 setOperationAction(ISD::VAEND, MVT::Other, Expand); 142 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 143 setOperationAction(ISD::VAARG, MVT::Other, Custom); 144 setOperationAction(ISD::VASTART, MVT::Other, Custom); 145 146 // Dynamic stack 147 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 148 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 149 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 150 151 // TRAMPOLINE is custom lowered. 152 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom); 153 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom); 154 155 maxStoresPerMemset = maxStoresPerMemsetOptSize = 4; 156 maxStoresPerMemmove = maxStoresPerMemmoveOptSize 157 = maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 2; 158 159 // We have target-specific dag combine patterns for the following nodes: 160 setTargetDAGCombine(ISD::STORE); 161 setTargetDAGCombine(ISD::ADD); 162 163 setMinFunctionAlignment(1); 164 } 165 166 SDValue XCoreTargetLowering:: 167 LowerOperation(SDValue Op, SelectionDAG &DAG) const { 168 switch (Op.getOpcode()) 169 { 170 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG); 171 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); 172 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); 173 case ISD::ConstantPool: return LowerConstantPool(Op, DAG); 174 case ISD::BR_JT: return LowerBR_JT(Op, DAG); 175 case ISD::LOAD: return LowerLOAD(Op, DAG); 176 case ISD::STORE: return LowerSTORE(Op, DAG); 177 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG); 178 case ISD::VAARG: return LowerVAARG(Op, DAG); 179 case ISD::VASTART: return LowerVASTART(Op, DAG); 180 case ISD::SMUL_LOHI: return LowerSMUL_LOHI(Op, DAG); 181 case ISD::UMUL_LOHI: return LowerUMUL_LOHI(Op, DAG); 182 // FIXME: Remove these when LegalizeDAGTypes lands. 183 case ISD::ADD: 184 case ISD::SUB: return ExpandADDSUB(Op.getNode(), DAG); 185 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); 186 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG); 187 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG); 188 default: 189 llvm_unreachable("unimplemented operand"); 190 } 191 } 192 193 /// ReplaceNodeResults - Replace the results of node with an illegal result 194 /// type with new values built out of custom code. 195 void XCoreTargetLowering::ReplaceNodeResults(SDNode *N, 196 SmallVectorImpl<SDValue>&Results, 197 SelectionDAG &DAG) const { 198 switch (N->getOpcode()) { 199 default: 200 llvm_unreachable("Don't know how to custom expand this!"); 201 case ISD::ADD: 202 case ISD::SUB: 203 Results.push_back(ExpandADDSUB(N, DAG)); 204 return; 205 } 206 } 207 208 //===----------------------------------------------------------------------===// 209 // Misc Lower Operation implementation 210 //===----------------------------------------------------------------------===// 211 212 SDValue XCoreTargetLowering:: 213 LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const 214 { 215 DebugLoc dl = Op.getDebugLoc(); 216 SDValue Cond = DAG.getNode(ISD::SETCC, dl, MVT::i32, Op.getOperand(2), 217 Op.getOperand(3), Op.getOperand(4)); 218 return DAG.getNode(ISD::SELECT, dl, MVT::i32, Cond, Op.getOperand(0), 219 Op.getOperand(1)); 220 } 221 222 SDValue XCoreTargetLowering:: 223 getGlobalAddressWrapper(SDValue GA, const GlobalValue *GV, 224 SelectionDAG &DAG) const 225 { 226 // FIXME there is no actual debug info here 227 DebugLoc dl = GA.getDebugLoc(); 228 if (isa<Function>(GV)) { 229 return DAG.getNode(XCoreISD::PCRelativeWrapper, dl, MVT::i32, GA); 230 } 231 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV); 232 if (!GVar) { 233 // If GV is an alias then use the aliasee to determine constness 234 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV)) 235 GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal()); 236 } 237 bool isConst = GVar && GVar->isConstant(); 238 if (isConst) { 239 return DAG.getNode(XCoreISD::CPRelativeWrapper, dl, MVT::i32, GA); 240 } 241 return DAG.getNode(XCoreISD::DPRelativeWrapper, dl, MVT::i32, GA); 242 } 243 244 SDValue XCoreTargetLowering:: 245 LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const 246 { 247 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 248 SDValue GA = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(), MVT::i32); 249 return getGlobalAddressWrapper(GA, GV, DAG); 250 } 251 252 static inline SDValue BuildGetId(SelectionDAG &DAG, DebugLoc dl) { 253 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32, 254 DAG.getConstant(Intrinsic::xcore_getid, MVT::i32)); 255 } 256 257 static inline bool isZeroLengthArray(Type *Ty) { 258 ArrayType *AT = dyn_cast_or_null<ArrayType>(Ty); 259 return AT && (AT->getNumElements() == 0); 260 } 261 262 SDValue XCoreTargetLowering:: 263 LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const 264 { 265 // FIXME there isn't really debug info here 266 DebugLoc dl = Op.getDebugLoc(); 267 // transform to label + getid() * size 268 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 269 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32); 270 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV); 271 if (!GVar) { 272 // If GV is an alias then use the aliasee to determine size 273 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV)) 274 GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal()); 275 } 276 if (!GVar) { 277 llvm_unreachable("Thread local object not a GlobalVariable?"); 278 } 279 Type *Ty = cast<PointerType>(GV->getType())->getElementType(); 280 if (!Ty->isSized() || isZeroLengthArray(Ty)) { 281 #ifndef NDEBUG 282 errs() << "Size of thread local object " << GVar->getName() 283 << " is unknown\n"; 284 #endif 285 llvm_unreachable(0); 286 } 287 SDValue base = getGlobalAddressWrapper(GA, GV, DAG); 288 const TargetData *TD = TM.getTargetData(); 289 unsigned Size = TD->getTypeAllocSize(Ty); 290 SDValue offset = DAG.getNode(ISD::MUL, dl, MVT::i32, BuildGetId(DAG, dl), 291 DAG.getConstant(Size, MVT::i32)); 292 return DAG.getNode(ISD::ADD, dl, MVT::i32, base, offset); 293 } 294 295 SDValue XCoreTargetLowering:: 296 LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const 297 { 298 DebugLoc DL = Op.getDebugLoc(); 299 300 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress(); 301 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(), /*isTarget=*/true); 302 303 return DAG.getNode(XCoreISD::PCRelativeWrapper, DL, getPointerTy(), Result); 304 } 305 306 SDValue XCoreTargetLowering:: 307 LowerConstantPool(SDValue Op, SelectionDAG &DAG) const 308 { 309 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); 310 // FIXME there isn't really debug info here 311 DebugLoc dl = CP->getDebugLoc(); 312 EVT PtrVT = Op.getValueType(); 313 SDValue Res; 314 if (CP->isMachineConstantPoolEntry()) { 315 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT, 316 CP->getAlignment()); 317 } else { 318 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, 319 CP->getAlignment()); 320 } 321 return DAG.getNode(XCoreISD::CPRelativeWrapper, dl, MVT::i32, Res); 322 } 323 324 unsigned XCoreTargetLowering::getJumpTableEncoding() const { 325 return MachineJumpTableInfo::EK_Inline; 326 } 327 328 SDValue XCoreTargetLowering:: 329 LowerBR_JT(SDValue Op, SelectionDAG &DAG) const 330 { 331 SDValue Chain = Op.getOperand(0); 332 SDValue Table = Op.getOperand(1); 333 SDValue Index = Op.getOperand(2); 334 DebugLoc dl = Op.getDebugLoc(); 335 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table); 336 unsigned JTI = JT->getIndex(); 337 MachineFunction &MF = DAG.getMachineFunction(); 338 const MachineJumpTableInfo *MJTI = MF.getJumpTableInfo(); 339 SDValue TargetJT = DAG.getTargetJumpTable(JT->getIndex(), MVT::i32); 340 341 unsigned NumEntries = MJTI->getJumpTables()[JTI].MBBs.size(); 342 if (NumEntries <= 32) { 343 return DAG.getNode(XCoreISD::BR_JT, dl, MVT::Other, Chain, TargetJT, Index); 344 } 345 assert((NumEntries >> 31) == 0); 346 SDValue ScaledIndex = DAG.getNode(ISD::SHL, dl, MVT::i32, Index, 347 DAG.getConstant(1, MVT::i32)); 348 return DAG.getNode(XCoreISD::BR_JT32, dl, MVT::Other, Chain, TargetJT, 349 ScaledIndex); 350 } 351 352 static bool 353 IsWordAlignedBasePlusConstantOffset(SDValue Addr, SDValue &AlignedBase, 354 int64_t &Offset) 355 { 356 if (Addr.getOpcode() != ISD::ADD) { 357 return false; 358 } 359 ConstantSDNode *CN = 0; 360 if (!(CN = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) { 361 return false; 362 } 363 int64_t off = CN->getSExtValue(); 364 const SDValue &Base = Addr.getOperand(0); 365 const SDValue *Root = &Base; 366 if (Base.getOpcode() == ISD::ADD && 367 Base.getOperand(1).getOpcode() == ISD::SHL) { 368 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Base.getOperand(1) 369 .getOperand(1)); 370 if (CN && (CN->getSExtValue() >= 2)) { 371 Root = &Base.getOperand(0); 372 } 373 } 374 if (isa<FrameIndexSDNode>(*Root)) { 375 // All frame indicies are word aligned 376 AlignedBase = Base; 377 Offset = off; 378 return true; 379 } 380 if (Root->getOpcode() == XCoreISD::DPRelativeWrapper || 381 Root->getOpcode() == XCoreISD::CPRelativeWrapper) { 382 // All dp / cp relative addresses are word aligned 383 AlignedBase = Base; 384 Offset = off; 385 return true; 386 } 387 // Check for an aligned global variable. 388 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(*Root)) { 389 const GlobalValue *GV = GA->getGlobal(); 390 if (GA->getOffset() == 0 && GV->getAlignment() >= 4) { 391 AlignedBase = Base; 392 Offset = off; 393 return true; 394 } 395 } 396 return false; 397 } 398 399 SDValue XCoreTargetLowering:: 400 LowerLOAD(SDValue Op, SelectionDAG &DAG) const { 401 LoadSDNode *LD = cast<LoadSDNode>(Op); 402 assert(LD->getExtensionType() == ISD::NON_EXTLOAD && 403 "Unexpected extension type"); 404 assert(LD->getMemoryVT() == MVT::i32 && "Unexpected load EVT"); 405 if (allowsUnalignedMemoryAccesses(LD->getMemoryVT())) 406 return SDValue(); 407 408 unsigned ABIAlignment = getTargetData()-> 409 getABITypeAlignment(LD->getMemoryVT().getTypeForEVT(*DAG.getContext())); 410 // Leave aligned load alone. 411 if (LD->getAlignment() >= ABIAlignment) 412 return SDValue(); 413 414 SDValue Chain = LD->getChain(); 415 SDValue BasePtr = LD->getBasePtr(); 416 DebugLoc DL = Op.getDebugLoc(); 417 418 SDValue Base; 419 int64_t Offset; 420 if (!LD->isVolatile() && 421 IsWordAlignedBasePlusConstantOffset(BasePtr, Base, Offset)) { 422 if (Offset % 4 == 0) { 423 // We've managed to infer better alignment information than the load 424 // already has. Use an aligned load. 425 // 426 return DAG.getLoad(getPointerTy(), DL, Chain, BasePtr, 427 MachinePointerInfo(), 428 false, false, false, 0); 429 } 430 // Lower to 431 // ldw low, base[offset >> 2] 432 // ldw high, base[(offset >> 2) + 1] 433 // shr low_shifted, low, (offset & 0x3) * 8 434 // shl high_shifted, high, 32 - (offset & 0x3) * 8 435 // or result, low_shifted, high_shifted 436 SDValue LowOffset = DAG.getConstant(Offset & ~0x3, MVT::i32); 437 SDValue HighOffset = DAG.getConstant((Offset & ~0x3) + 4, MVT::i32); 438 SDValue LowShift = DAG.getConstant((Offset & 0x3) * 8, MVT::i32); 439 SDValue HighShift = DAG.getConstant(32 - (Offset & 0x3) * 8, MVT::i32); 440 441 SDValue LowAddr = DAG.getNode(ISD::ADD, DL, MVT::i32, Base, LowOffset); 442 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, MVT::i32, Base, HighOffset); 443 444 SDValue Low = DAG.getLoad(getPointerTy(), DL, Chain, 445 LowAddr, MachinePointerInfo(), 446 false, false, false, 0); 447 SDValue High = DAG.getLoad(getPointerTy(), DL, Chain, 448 HighAddr, MachinePointerInfo(), 449 false, false, false, 0); 450 SDValue LowShifted = DAG.getNode(ISD::SRL, DL, MVT::i32, Low, LowShift); 451 SDValue HighShifted = DAG.getNode(ISD::SHL, DL, MVT::i32, High, HighShift); 452 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i32, LowShifted, HighShifted); 453 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Low.getValue(1), 454 High.getValue(1)); 455 SDValue Ops[] = { Result, Chain }; 456 return DAG.getMergeValues(Ops, 2, DL); 457 } 458 459 if (LD->getAlignment() == 2) { 460 SDValue Low = DAG.getExtLoad(ISD::ZEXTLOAD, DL, MVT::i32, Chain, 461 BasePtr, LD->getPointerInfo(), MVT::i16, 462 LD->isVolatile(), LD->isNonTemporal(), 2); 463 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr, 464 DAG.getConstant(2, MVT::i32)); 465 SDValue High = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain, 466 HighAddr, 467 LD->getPointerInfo().getWithOffset(2), 468 MVT::i16, LD->isVolatile(), 469 LD->isNonTemporal(), 2); 470 SDValue HighShifted = DAG.getNode(ISD::SHL, DL, MVT::i32, High, 471 DAG.getConstant(16, MVT::i32)); 472 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i32, Low, HighShifted); 473 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Low.getValue(1), 474 High.getValue(1)); 475 SDValue Ops[] = { Result, Chain }; 476 return DAG.getMergeValues(Ops, 2, DL); 477 } 478 479 // Lower to a call to __misaligned_load(BasePtr). 480 Type *IntPtrTy = getTargetData()->getIntPtrType(*DAG.getContext()); 481 TargetLowering::ArgListTy Args; 482 TargetLowering::ArgListEntry Entry; 483 484 Entry.Ty = IntPtrTy; 485 Entry.Node = BasePtr; 486 Args.push_back(Entry); 487 488 std::pair<SDValue, SDValue> CallResult = 489 LowerCallTo(Chain, IntPtrTy, false, false, 490 false, false, 0, CallingConv::C, /*isTailCall=*/false, 491 /*doesNotRet=*/false, /*isReturnValueUsed=*/true, 492 DAG.getExternalSymbol("__misaligned_load", getPointerTy()), 493 Args, DAG, DL); 494 495 SDValue Ops[] = 496 { CallResult.first, CallResult.second }; 497 498 return DAG.getMergeValues(Ops, 2, DL); 499 } 500 501 SDValue XCoreTargetLowering:: 502 LowerSTORE(SDValue Op, SelectionDAG &DAG) const 503 { 504 StoreSDNode *ST = cast<StoreSDNode>(Op); 505 assert(!ST->isTruncatingStore() && "Unexpected store type"); 506 assert(ST->getMemoryVT() == MVT::i32 && "Unexpected store EVT"); 507 if (allowsUnalignedMemoryAccesses(ST->getMemoryVT())) { 508 return SDValue(); 509 } 510 unsigned ABIAlignment = getTargetData()-> 511 getABITypeAlignment(ST->getMemoryVT().getTypeForEVT(*DAG.getContext())); 512 // Leave aligned store alone. 513 if (ST->getAlignment() >= ABIAlignment) { 514 return SDValue(); 515 } 516 SDValue Chain = ST->getChain(); 517 SDValue BasePtr = ST->getBasePtr(); 518 SDValue Value = ST->getValue(); 519 DebugLoc dl = Op.getDebugLoc(); 520 521 if (ST->getAlignment() == 2) { 522 SDValue Low = Value; 523 SDValue High = DAG.getNode(ISD::SRL, dl, MVT::i32, Value, 524 DAG.getConstant(16, MVT::i32)); 525 SDValue StoreLow = DAG.getTruncStore(Chain, dl, Low, BasePtr, 526 ST->getPointerInfo(), MVT::i16, 527 ST->isVolatile(), ST->isNonTemporal(), 528 2); 529 SDValue HighAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, BasePtr, 530 DAG.getConstant(2, MVT::i32)); 531 SDValue StoreHigh = DAG.getTruncStore(Chain, dl, High, HighAddr, 532 ST->getPointerInfo().getWithOffset(2), 533 MVT::i16, ST->isVolatile(), 534 ST->isNonTemporal(), 2); 535 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, StoreLow, StoreHigh); 536 } 537 538 // Lower to a call to __misaligned_store(BasePtr, Value). 539 Type *IntPtrTy = getTargetData()->getIntPtrType(*DAG.getContext()); 540 TargetLowering::ArgListTy Args; 541 TargetLowering::ArgListEntry Entry; 542 543 Entry.Ty = IntPtrTy; 544 Entry.Node = BasePtr; 545 Args.push_back(Entry); 546 547 Entry.Node = Value; 548 Args.push_back(Entry); 549 550 std::pair<SDValue, SDValue> CallResult = 551 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()), false, false, 552 false, false, 0, CallingConv::C, /*isTailCall=*/false, 553 /*doesNotRet=*/false, /*isReturnValueUsed=*/true, 554 DAG.getExternalSymbol("__misaligned_store", getPointerTy()), 555 Args, DAG, dl); 556 557 return CallResult.second; 558 } 559 560 SDValue XCoreTargetLowering:: 561 LowerSMUL_LOHI(SDValue Op, SelectionDAG &DAG) const 562 { 563 assert(Op.getValueType() == MVT::i32 && Op.getOpcode() == ISD::SMUL_LOHI && 564 "Unexpected operand to lower!"); 565 DebugLoc dl = Op.getDebugLoc(); 566 SDValue LHS = Op.getOperand(0); 567 SDValue RHS = Op.getOperand(1); 568 SDValue Zero = DAG.getConstant(0, MVT::i32); 569 SDValue Hi = DAG.getNode(XCoreISD::MACCS, dl, 570 DAG.getVTList(MVT::i32, MVT::i32), Zero, Zero, 571 LHS, RHS); 572 SDValue Lo(Hi.getNode(), 1); 573 SDValue Ops[] = { Lo, Hi }; 574 return DAG.getMergeValues(Ops, 2, dl); 575 } 576 577 SDValue XCoreTargetLowering:: 578 LowerUMUL_LOHI(SDValue Op, SelectionDAG &DAG) const 579 { 580 assert(Op.getValueType() == MVT::i32 && Op.getOpcode() == ISD::UMUL_LOHI && 581 "Unexpected operand to lower!"); 582 DebugLoc dl = Op.getDebugLoc(); 583 SDValue LHS = Op.getOperand(0); 584 SDValue RHS = Op.getOperand(1); 585 SDValue Zero = DAG.getConstant(0, MVT::i32); 586 SDValue Hi = DAG.getNode(XCoreISD::LMUL, dl, 587 DAG.getVTList(MVT::i32, MVT::i32), LHS, RHS, 588 Zero, Zero); 589 SDValue Lo(Hi.getNode(), 1); 590 SDValue Ops[] = { Lo, Hi }; 591 return DAG.getMergeValues(Ops, 2, dl); 592 } 593 594 /// isADDADDMUL - Return whether Op is in a form that is equivalent to 595 /// add(add(mul(x,y),a),b). If requireIntermediatesHaveOneUse is true then 596 /// each intermediate result in the calculation must also have a single use. 597 /// If the Op is in the correct form the constituent parts are written to Mul0, 598 /// Mul1, Addend0 and Addend1. 599 static bool 600 isADDADDMUL(SDValue Op, SDValue &Mul0, SDValue &Mul1, SDValue &Addend0, 601 SDValue &Addend1, bool requireIntermediatesHaveOneUse) 602 { 603 if (Op.getOpcode() != ISD::ADD) 604 return false; 605 SDValue N0 = Op.getOperand(0); 606 SDValue N1 = Op.getOperand(1); 607 SDValue AddOp; 608 SDValue OtherOp; 609 if (N0.getOpcode() == ISD::ADD) { 610 AddOp = N0; 611 OtherOp = N1; 612 } else if (N1.getOpcode() == ISD::ADD) { 613 AddOp = N1; 614 OtherOp = N0; 615 } else { 616 return false; 617 } 618 if (requireIntermediatesHaveOneUse && !AddOp.hasOneUse()) 619 return false; 620 if (OtherOp.getOpcode() == ISD::MUL) { 621 // add(add(a,b),mul(x,y)) 622 if (requireIntermediatesHaveOneUse && !OtherOp.hasOneUse()) 623 return false; 624 Mul0 = OtherOp.getOperand(0); 625 Mul1 = OtherOp.getOperand(1); 626 Addend0 = AddOp.getOperand(0); 627 Addend1 = AddOp.getOperand(1); 628 return true; 629 } 630 if (AddOp.getOperand(0).getOpcode() == ISD::MUL) { 631 // add(add(mul(x,y),a),b) 632 if (requireIntermediatesHaveOneUse && !AddOp.getOperand(0).hasOneUse()) 633 return false; 634 Mul0 = AddOp.getOperand(0).getOperand(0); 635 Mul1 = AddOp.getOperand(0).getOperand(1); 636 Addend0 = AddOp.getOperand(1); 637 Addend1 = OtherOp; 638 return true; 639 } 640 if (AddOp.getOperand(1).getOpcode() == ISD::MUL) { 641 // add(add(a,mul(x,y)),b) 642 if (requireIntermediatesHaveOneUse && !AddOp.getOperand(1).hasOneUse()) 643 return false; 644 Mul0 = AddOp.getOperand(1).getOperand(0); 645 Mul1 = AddOp.getOperand(1).getOperand(1); 646 Addend0 = AddOp.getOperand(0); 647 Addend1 = OtherOp; 648 return true; 649 } 650 return false; 651 } 652 653 SDValue XCoreTargetLowering:: 654 TryExpandADDWithMul(SDNode *N, SelectionDAG &DAG) const 655 { 656 SDValue Mul; 657 SDValue Other; 658 if (N->getOperand(0).getOpcode() == ISD::MUL) { 659 Mul = N->getOperand(0); 660 Other = N->getOperand(1); 661 } else if (N->getOperand(1).getOpcode() == ISD::MUL) { 662 Mul = N->getOperand(1); 663 Other = N->getOperand(0); 664 } else { 665 return SDValue(); 666 } 667 DebugLoc dl = N->getDebugLoc(); 668 SDValue LL, RL, AddendL, AddendH; 669 LL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 670 Mul.getOperand(0), DAG.getConstant(0, MVT::i32)); 671 RL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 672 Mul.getOperand(1), DAG.getConstant(0, MVT::i32)); 673 AddendL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 674 Other, DAG.getConstant(0, MVT::i32)); 675 AddendH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 676 Other, DAG.getConstant(1, MVT::i32)); 677 APInt HighMask = APInt::getHighBitsSet(64, 32); 678 unsigned LHSSB = DAG.ComputeNumSignBits(Mul.getOperand(0)); 679 unsigned RHSSB = DAG.ComputeNumSignBits(Mul.getOperand(1)); 680 if (DAG.MaskedValueIsZero(Mul.getOperand(0), HighMask) && 681 DAG.MaskedValueIsZero(Mul.getOperand(1), HighMask)) { 682 // The inputs are both zero-extended. 683 SDValue Hi = DAG.getNode(XCoreISD::MACCU, dl, 684 DAG.getVTList(MVT::i32, MVT::i32), AddendH, 685 AddendL, LL, RL); 686 SDValue Lo(Hi.getNode(), 1); 687 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi); 688 } 689 if (LHSSB > 32 && RHSSB > 32) { 690 // The inputs are both sign-extended. 691 SDValue Hi = DAG.getNode(XCoreISD::MACCS, dl, 692 DAG.getVTList(MVT::i32, MVT::i32), AddendH, 693 AddendL, LL, RL); 694 SDValue Lo(Hi.getNode(), 1); 695 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi); 696 } 697 SDValue LH, RH; 698 LH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 699 Mul.getOperand(0), DAG.getConstant(1, MVT::i32)); 700 RH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 701 Mul.getOperand(1), DAG.getConstant(1, MVT::i32)); 702 SDValue Hi = DAG.getNode(XCoreISD::MACCU, dl, 703 DAG.getVTList(MVT::i32, MVT::i32), AddendH, 704 AddendL, LL, RL); 705 SDValue Lo(Hi.getNode(), 1); 706 RH = DAG.getNode(ISD::MUL, dl, MVT::i32, LL, RH); 707 LH = DAG.getNode(ISD::MUL, dl, MVT::i32, LH, RL); 708 Hi = DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, RH); 709 Hi = DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, LH); 710 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi); 711 } 712 713 SDValue XCoreTargetLowering:: 714 ExpandADDSUB(SDNode *N, SelectionDAG &DAG) const 715 { 716 assert(N->getValueType(0) == MVT::i64 && 717 (N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) && 718 "Unknown operand to lower!"); 719 720 if (N->getOpcode() == ISD::ADD) { 721 SDValue Result = TryExpandADDWithMul(N, DAG); 722 if (Result.getNode() != 0) 723 return Result; 724 } 725 726 DebugLoc dl = N->getDebugLoc(); 727 728 // Extract components 729 SDValue LHSL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 730 N->getOperand(0), DAG.getConstant(0, MVT::i32)); 731 SDValue LHSH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 732 N->getOperand(0), DAG.getConstant(1, MVT::i32)); 733 SDValue RHSL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 734 N->getOperand(1), DAG.getConstant(0, MVT::i32)); 735 SDValue RHSH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 736 N->getOperand(1), DAG.getConstant(1, MVT::i32)); 737 738 // Expand 739 unsigned Opcode = (N->getOpcode() == ISD::ADD) ? XCoreISD::LADD : 740 XCoreISD::LSUB; 741 SDValue Zero = DAG.getConstant(0, MVT::i32); 742 SDValue Carry = DAG.getNode(Opcode, dl, DAG.getVTList(MVT::i32, MVT::i32), 743 LHSL, RHSL, Zero); 744 SDValue Lo(Carry.getNode(), 1); 745 746 SDValue Ignored = DAG.getNode(Opcode, dl, DAG.getVTList(MVT::i32, MVT::i32), 747 LHSH, RHSH, Carry); 748 SDValue Hi(Ignored.getNode(), 1); 749 // Merge the pieces 750 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi); 751 } 752 753 SDValue XCoreTargetLowering:: 754 LowerVAARG(SDValue Op, SelectionDAG &DAG) const 755 { 756 llvm_unreachable("unimplemented"); 757 // FIXME Arguments passed by reference need a extra dereference. 758 SDNode *Node = Op.getNode(); 759 DebugLoc dl = Node->getDebugLoc(); 760 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue(); 761 EVT VT = Node->getValueType(0); 762 SDValue VAList = DAG.getLoad(getPointerTy(), dl, Node->getOperand(0), 763 Node->getOperand(1), MachinePointerInfo(V), 764 false, false, false, 0); 765 // Increment the pointer, VAList, to the next vararg 766 SDValue Tmp3 = DAG.getNode(ISD::ADD, dl, getPointerTy(), VAList, 767 DAG.getConstant(VT.getSizeInBits(), 768 getPointerTy())); 769 // Store the incremented VAList to the legalized pointer 770 Tmp3 = DAG.getStore(VAList.getValue(1), dl, Tmp3, Node->getOperand(1), 771 MachinePointerInfo(V), false, false, 0); 772 // Load the actual argument out of the pointer VAList 773 return DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(), 774 false, false, false, 0); 775 } 776 777 SDValue XCoreTargetLowering:: 778 LowerVASTART(SDValue Op, SelectionDAG &DAG) const 779 { 780 DebugLoc dl = Op.getDebugLoc(); 781 // vastart stores the address of the VarArgsFrameIndex slot into the 782 // memory location argument 783 MachineFunction &MF = DAG.getMachineFunction(); 784 XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>(); 785 SDValue Addr = DAG.getFrameIndex(XFI->getVarArgsFrameIndex(), MVT::i32); 786 return DAG.getStore(Op.getOperand(0), dl, Addr, Op.getOperand(1), 787 MachinePointerInfo(), false, false, 0); 788 } 789 790 SDValue XCoreTargetLowering::LowerFRAMEADDR(SDValue Op, 791 SelectionDAG &DAG) const { 792 DebugLoc dl = Op.getDebugLoc(); 793 // Depths > 0 not supported yet! 794 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0) 795 return SDValue(); 796 797 MachineFunction &MF = DAG.getMachineFunction(); 798 const TargetRegisterInfo *RegInfo = getTargetMachine().getRegisterInfo(); 799 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, 800 RegInfo->getFrameRegister(MF), MVT::i32); 801 } 802 803 SDValue XCoreTargetLowering:: 804 LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const { 805 return Op.getOperand(0); 806 } 807 808 SDValue XCoreTargetLowering:: 809 LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const { 810 SDValue Chain = Op.getOperand(0); 811 SDValue Trmp = Op.getOperand(1); // trampoline 812 SDValue FPtr = Op.getOperand(2); // nested function 813 SDValue Nest = Op.getOperand(3); // 'nest' parameter value 814 815 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue(); 816 817 // .align 4 818 // LDAPF_u10 r11, nest 819 // LDW_2rus r11, r11[0] 820 // STWSP_ru6 r11, sp[0] 821 // LDAPF_u10 r11, fptr 822 // LDW_2rus r11, r11[0] 823 // BAU_1r r11 824 // nest: 825 // .word nest 826 // fptr: 827 // .word fptr 828 SDValue OutChains[5]; 829 830 SDValue Addr = Trmp; 831 832 DebugLoc dl = Op.getDebugLoc(); 833 OutChains[0] = DAG.getStore(Chain, dl, DAG.getConstant(0x0a3cd805, MVT::i32), 834 Addr, MachinePointerInfo(TrmpAddr), false, false, 835 0); 836 837 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, 838 DAG.getConstant(4, MVT::i32)); 839 OutChains[1] = DAG.getStore(Chain, dl, DAG.getConstant(0xd80456c0, MVT::i32), 840 Addr, MachinePointerInfo(TrmpAddr, 4), false, 841 false, 0); 842 843 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, 844 DAG.getConstant(8, MVT::i32)); 845 OutChains[2] = DAG.getStore(Chain, dl, DAG.getConstant(0x27fb0a3c, MVT::i32), 846 Addr, MachinePointerInfo(TrmpAddr, 8), false, 847 false, 0); 848 849 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, 850 DAG.getConstant(12, MVT::i32)); 851 OutChains[3] = DAG.getStore(Chain, dl, Nest, Addr, 852 MachinePointerInfo(TrmpAddr, 12), false, false, 853 0); 854 855 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, 856 DAG.getConstant(16, MVT::i32)); 857 OutChains[4] = DAG.getStore(Chain, dl, FPtr, Addr, 858 MachinePointerInfo(TrmpAddr, 16), false, false, 859 0); 860 861 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 5); 862 } 863 864 //===----------------------------------------------------------------------===// 865 // Calling Convention Implementation 866 //===----------------------------------------------------------------------===// 867 868 #include "XCoreGenCallingConv.inc" 869 870 //===----------------------------------------------------------------------===// 871 // Call Calling Convention Implementation 872 //===----------------------------------------------------------------------===// 873 874 /// XCore call implementation 875 SDValue 876 XCoreTargetLowering::LowerCall(SDValue Chain, SDValue Callee, 877 CallingConv::ID CallConv, bool isVarArg, 878 bool doesNotRet, bool &isTailCall, 879 const SmallVectorImpl<ISD::OutputArg> &Outs, 880 const SmallVectorImpl<SDValue> &OutVals, 881 const SmallVectorImpl<ISD::InputArg> &Ins, 882 DebugLoc dl, SelectionDAG &DAG, 883 SmallVectorImpl<SDValue> &InVals) const { 884 // XCore target does not yet support tail call optimization. 885 isTailCall = false; 886 887 // For now, only CallingConv::C implemented 888 switch (CallConv) 889 { 890 default: 891 llvm_unreachable("Unsupported calling convention"); 892 case CallingConv::Fast: 893 case CallingConv::C: 894 return LowerCCCCallTo(Chain, Callee, CallConv, isVarArg, isTailCall, 895 Outs, OutVals, Ins, dl, DAG, InVals); 896 } 897 } 898 899 /// LowerCCCCallTo - functions arguments are copied from virtual 900 /// regs to (physical regs)/(stack frame), CALLSEQ_START and 901 /// CALLSEQ_END are emitted. 902 /// TODO: isTailCall, sret. 903 SDValue 904 XCoreTargetLowering::LowerCCCCallTo(SDValue Chain, SDValue Callee, 905 CallingConv::ID CallConv, bool isVarArg, 906 bool isTailCall, 907 const SmallVectorImpl<ISD::OutputArg> &Outs, 908 const SmallVectorImpl<SDValue> &OutVals, 909 const SmallVectorImpl<ISD::InputArg> &Ins, 910 DebugLoc dl, SelectionDAG &DAG, 911 SmallVectorImpl<SDValue> &InVals) const { 912 913 // Analyze operands of the call, assigning locations to each operand. 914 SmallVector<CCValAssign, 16> ArgLocs; 915 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 916 getTargetMachine(), ArgLocs, *DAG.getContext()); 917 918 // The ABI dictates there should be one stack slot available to the callee 919 // on function entry (for saving lr). 920 CCInfo.AllocateStack(4, 4); 921 922 CCInfo.AnalyzeCallOperands(Outs, CC_XCore); 923 924 // Get a count of how many bytes are to be pushed on the stack. 925 unsigned NumBytes = CCInfo.getNextStackOffset(); 926 927 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, 928 getPointerTy(), true)); 929 930 SmallVector<std::pair<unsigned, SDValue>, 4> RegsToPass; 931 SmallVector<SDValue, 12> MemOpChains; 932 933 // Walk the register/memloc assignments, inserting copies/loads. 934 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 935 CCValAssign &VA = ArgLocs[i]; 936 SDValue Arg = OutVals[i]; 937 938 // Promote the value if needed. 939 switch (VA.getLocInfo()) { 940 default: llvm_unreachable("Unknown loc info!"); 941 case CCValAssign::Full: break; 942 case CCValAssign::SExt: 943 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); 944 break; 945 case CCValAssign::ZExt: 946 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); 947 break; 948 case CCValAssign::AExt: 949 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg); 950 break; 951 } 952 953 // Arguments that can be passed on register must be kept at 954 // RegsToPass vector 955 if (VA.isRegLoc()) { 956 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 957 } else { 958 assert(VA.isMemLoc()); 959 960 int Offset = VA.getLocMemOffset(); 961 962 MemOpChains.push_back(DAG.getNode(XCoreISD::STWSP, dl, MVT::Other, 963 Chain, Arg, 964 DAG.getConstant(Offset/4, MVT::i32))); 965 } 966 } 967 968 // Transform all store nodes into one single node because 969 // all store nodes are independent of each other. 970 if (!MemOpChains.empty()) 971 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 972 &MemOpChains[0], MemOpChains.size()); 973 974 // Build a sequence of copy-to-reg nodes chained together with token 975 // chain and flag operands which copy the outgoing args into registers. 976 // The InFlag in necessary since all emitted instructions must be 977 // stuck together. 978 SDValue InFlag; 979 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 980 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, 981 RegsToPass[i].second, InFlag); 982 InFlag = Chain.getValue(1); 983 } 984 985 // If the callee is a GlobalAddress node (quite common, every direct call is) 986 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. 987 // Likewise ExternalSymbol -> TargetExternalSymbol. 988 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) 989 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, MVT::i32); 990 else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee)) 991 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), MVT::i32); 992 993 // XCoreBranchLink = #chain, #target_address, #opt_in_flags... 994 // = Chain, Callee, Reg#1, Reg#2, ... 995 // 996 // Returns a chain & a flag for retval copy to use. 997 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 998 SmallVector<SDValue, 8> Ops; 999 Ops.push_back(Chain); 1000 Ops.push_back(Callee); 1001 1002 // Add argument registers to the end of the list so that they are 1003 // known live into the call. 1004 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 1005 Ops.push_back(DAG.getRegister(RegsToPass[i].first, 1006 RegsToPass[i].second.getValueType())); 1007 1008 if (InFlag.getNode()) 1009 Ops.push_back(InFlag); 1010 1011 Chain = DAG.getNode(XCoreISD::BL, dl, NodeTys, &Ops[0], Ops.size()); 1012 InFlag = Chain.getValue(1); 1013 1014 // Create the CALLSEQ_END node. 1015 Chain = DAG.getCALLSEQ_END(Chain, 1016 DAG.getConstant(NumBytes, getPointerTy(), true), 1017 DAG.getConstant(0, getPointerTy(), true), 1018 InFlag); 1019 InFlag = Chain.getValue(1); 1020 1021 // Handle result values, copying them out of physregs into vregs that we 1022 // return. 1023 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, 1024 Ins, dl, DAG, InVals); 1025 } 1026 1027 /// LowerCallResult - Lower the result values of a call into the 1028 /// appropriate copies out of appropriate physical registers. 1029 SDValue 1030 XCoreTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, 1031 CallingConv::ID CallConv, bool isVarArg, 1032 const SmallVectorImpl<ISD::InputArg> &Ins, 1033 DebugLoc dl, SelectionDAG &DAG, 1034 SmallVectorImpl<SDValue> &InVals) const { 1035 1036 // Assign locations to each value returned by this call. 1037 SmallVector<CCValAssign, 16> RVLocs; 1038 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1039 getTargetMachine(), RVLocs, *DAG.getContext()); 1040 1041 CCInfo.AnalyzeCallResult(Ins, RetCC_XCore); 1042 1043 // Copy all of the result registers out of their specified physreg. 1044 for (unsigned i = 0; i != RVLocs.size(); ++i) { 1045 Chain = DAG.getCopyFromReg(Chain, dl, RVLocs[i].getLocReg(), 1046 RVLocs[i].getValVT(), InFlag).getValue(1); 1047 InFlag = Chain.getValue(2); 1048 InVals.push_back(Chain.getValue(0)); 1049 } 1050 1051 return Chain; 1052 } 1053 1054 //===----------------------------------------------------------------------===// 1055 // Formal Arguments Calling Convention Implementation 1056 //===----------------------------------------------------------------------===// 1057 1058 /// XCore formal arguments implementation 1059 SDValue 1060 XCoreTargetLowering::LowerFormalArguments(SDValue Chain, 1061 CallingConv::ID CallConv, 1062 bool isVarArg, 1063 const SmallVectorImpl<ISD::InputArg> &Ins, 1064 DebugLoc dl, 1065 SelectionDAG &DAG, 1066 SmallVectorImpl<SDValue> &InVals) 1067 const { 1068 switch (CallConv) 1069 { 1070 default: 1071 llvm_unreachable("Unsupported calling convention"); 1072 case CallingConv::C: 1073 case CallingConv::Fast: 1074 return LowerCCCArguments(Chain, CallConv, isVarArg, 1075 Ins, dl, DAG, InVals); 1076 } 1077 } 1078 1079 /// LowerCCCArguments - transform physical registers into 1080 /// virtual registers and generate load operations for 1081 /// arguments places on the stack. 1082 /// TODO: sret 1083 SDValue 1084 XCoreTargetLowering::LowerCCCArguments(SDValue Chain, 1085 CallingConv::ID CallConv, 1086 bool isVarArg, 1087 const SmallVectorImpl<ISD::InputArg> 1088 &Ins, 1089 DebugLoc dl, 1090 SelectionDAG &DAG, 1091 SmallVectorImpl<SDValue> &InVals) const { 1092 MachineFunction &MF = DAG.getMachineFunction(); 1093 MachineFrameInfo *MFI = MF.getFrameInfo(); 1094 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 1095 1096 // Assign locations to all of the incoming arguments. 1097 SmallVector<CCValAssign, 16> ArgLocs; 1098 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1099 getTargetMachine(), ArgLocs, *DAG.getContext()); 1100 1101 CCInfo.AnalyzeFormalArguments(Ins, CC_XCore); 1102 1103 unsigned StackSlotSize = XCoreFrameLowering::stackSlotSize(); 1104 1105 unsigned LRSaveSize = StackSlotSize; 1106 1107 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 1108 1109 CCValAssign &VA = ArgLocs[i]; 1110 1111 if (VA.isRegLoc()) { 1112 // Arguments passed in registers 1113 EVT RegVT = VA.getLocVT(); 1114 switch (RegVT.getSimpleVT().SimpleTy) { 1115 default: 1116 { 1117 #ifndef NDEBUG 1118 errs() << "LowerFormalArguments Unhandled argument type: " 1119 << RegVT.getSimpleVT().SimpleTy << "\n"; 1120 #endif 1121 llvm_unreachable(0); 1122 } 1123 case MVT::i32: 1124 unsigned VReg = RegInfo.createVirtualRegister( 1125 XCore::GRRegsRegisterClass); 1126 RegInfo.addLiveIn(VA.getLocReg(), VReg); 1127 InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT)); 1128 } 1129 } else { 1130 // sanity check 1131 assert(VA.isMemLoc()); 1132 // Load the argument to a virtual register 1133 unsigned ObjSize = VA.getLocVT().getSizeInBits()/8; 1134 if (ObjSize > StackSlotSize) { 1135 errs() << "LowerFormalArguments Unhandled argument type: " 1136 << EVT(VA.getLocVT()).getEVTString() 1137 << "\n"; 1138 } 1139 // Create the frame index object for this incoming parameter... 1140 int FI = MFI->CreateFixedObject(ObjSize, 1141 LRSaveSize + VA.getLocMemOffset(), 1142 true); 1143 1144 // Create the SelectionDAG nodes corresponding to a load 1145 //from this parameter 1146 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32); 1147 InVals.push_back(DAG.getLoad(VA.getLocVT(), dl, Chain, FIN, 1148 MachinePointerInfo::getFixedStack(FI), 1149 false, false, false, 0)); 1150 } 1151 } 1152 1153 if (isVarArg) { 1154 /* Argument registers */ 1155 static const uint16_t ArgRegs[] = { 1156 XCore::R0, XCore::R1, XCore::R2, XCore::R3 1157 }; 1158 XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>(); 1159 unsigned FirstVAReg = CCInfo.getFirstUnallocated(ArgRegs, 1160 array_lengthof(ArgRegs)); 1161 if (FirstVAReg < array_lengthof(ArgRegs)) { 1162 SmallVector<SDValue, 4> MemOps; 1163 int offset = 0; 1164 // Save remaining registers, storing higher register numbers at a higher 1165 // address 1166 for (int i = array_lengthof(ArgRegs) - 1; i >= (int)FirstVAReg; --i) { 1167 // Create a stack slot 1168 int FI = MFI->CreateFixedObject(4, offset, true); 1169 if (i == (int)FirstVAReg) { 1170 XFI->setVarArgsFrameIndex(FI); 1171 } 1172 offset -= StackSlotSize; 1173 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32); 1174 // Move argument from phys reg -> virt reg 1175 unsigned VReg = RegInfo.createVirtualRegister( 1176 XCore::GRRegsRegisterClass); 1177 RegInfo.addLiveIn(ArgRegs[i], VReg); 1178 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32); 1179 // Move argument from virt reg -> stack 1180 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, 1181 MachinePointerInfo(), false, false, 0); 1182 MemOps.push_back(Store); 1183 } 1184 if (!MemOps.empty()) 1185 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 1186 &MemOps[0], MemOps.size()); 1187 } else { 1188 // This will point to the next argument passed via stack. 1189 XFI->setVarArgsFrameIndex( 1190 MFI->CreateFixedObject(4, LRSaveSize + CCInfo.getNextStackOffset(), 1191 true)); 1192 } 1193 } 1194 1195 return Chain; 1196 } 1197 1198 //===----------------------------------------------------------------------===// 1199 // Return Value Calling Convention Implementation 1200 //===----------------------------------------------------------------------===// 1201 1202 bool XCoreTargetLowering:: 1203 CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF, 1204 bool isVarArg, 1205 const SmallVectorImpl<ISD::OutputArg> &Outs, 1206 LLVMContext &Context) const { 1207 SmallVector<CCValAssign, 16> RVLocs; 1208 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context); 1209 return CCInfo.CheckReturn(Outs, RetCC_XCore); 1210 } 1211 1212 SDValue 1213 XCoreTargetLowering::LowerReturn(SDValue Chain, 1214 CallingConv::ID CallConv, bool isVarArg, 1215 const SmallVectorImpl<ISD::OutputArg> &Outs, 1216 const SmallVectorImpl<SDValue> &OutVals, 1217 DebugLoc dl, SelectionDAG &DAG) const { 1218 1219 // CCValAssign - represent the assignment of 1220 // the return value to a location 1221 SmallVector<CCValAssign, 16> RVLocs; 1222 1223 // CCState - Info about the registers and stack slot. 1224 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1225 getTargetMachine(), RVLocs, *DAG.getContext()); 1226 1227 // Analyze return values. 1228 CCInfo.AnalyzeReturn(Outs, RetCC_XCore); 1229 1230 // If this is the first return lowered for this function, add 1231 // the regs to the liveout set for the function. 1232 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) { 1233 for (unsigned i = 0; i != RVLocs.size(); ++i) 1234 if (RVLocs[i].isRegLoc()) 1235 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg()); 1236 } 1237 1238 SDValue Flag; 1239 1240 // Copy the result values into the output registers. 1241 for (unsigned i = 0; i != RVLocs.size(); ++i) { 1242 CCValAssign &VA = RVLocs[i]; 1243 assert(VA.isRegLoc() && "Can only return in registers!"); 1244 1245 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), 1246 OutVals[i], Flag); 1247 1248 // guarantee that all emitted copies are 1249 // stuck together, avoiding something bad 1250 Flag = Chain.getValue(1); 1251 } 1252 1253 // Return on XCore is always a "retsp 0" 1254 if (Flag.getNode()) 1255 return DAG.getNode(XCoreISD::RETSP, dl, MVT::Other, 1256 Chain, DAG.getConstant(0, MVT::i32), Flag); 1257 else // Return Void 1258 return DAG.getNode(XCoreISD::RETSP, dl, MVT::Other, 1259 Chain, DAG.getConstant(0, MVT::i32)); 1260 } 1261 1262 //===----------------------------------------------------------------------===// 1263 // Other Lowering Code 1264 //===----------------------------------------------------------------------===// 1265 1266 MachineBasicBlock * 1267 XCoreTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 1268 MachineBasicBlock *BB) const { 1269 const TargetInstrInfo &TII = *getTargetMachine().getInstrInfo(); 1270 DebugLoc dl = MI->getDebugLoc(); 1271 assert((MI->getOpcode() == XCore::SELECT_CC) && 1272 "Unexpected instr type to insert"); 1273 1274 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond 1275 // control-flow pattern. The incoming instruction knows the destination vreg 1276 // to set, the condition code register to branch on, the true/false values to 1277 // select between, and a branch opcode to use. 1278 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1279 MachineFunction::iterator It = BB; 1280 ++It; 1281 1282 // thisMBB: 1283 // ... 1284 // TrueVal = ... 1285 // cmpTY ccX, r1, r2 1286 // bCC copy1MBB 1287 // fallthrough --> copy0MBB 1288 MachineBasicBlock *thisMBB = BB; 1289 MachineFunction *F = BB->getParent(); 1290 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 1291 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); 1292 F->insert(It, copy0MBB); 1293 F->insert(It, sinkMBB); 1294 1295 // Transfer the remainder of BB and its successor edges to sinkMBB. 1296 sinkMBB->splice(sinkMBB->begin(), BB, 1297 llvm::next(MachineBasicBlock::iterator(MI)), 1298 BB->end()); 1299 sinkMBB->transferSuccessorsAndUpdatePHIs(BB); 1300 1301 // Next, add the true and fallthrough blocks as its successors. 1302 BB->addSuccessor(copy0MBB); 1303 BB->addSuccessor(sinkMBB); 1304 1305 BuildMI(BB, dl, TII.get(XCore::BRFT_lru6)) 1306 .addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB); 1307 1308 // copy0MBB: 1309 // %FalseValue = ... 1310 // # fallthrough to sinkMBB 1311 BB = copy0MBB; 1312 1313 // Update machine-CFG edges 1314 BB->addSuccessor(sinkMBB); 1315 1316 // sinkMBB: 1317 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] 1318 // ... 1319 BB = sinkMBB; 1320 BuildMI(*BB, BB->begin(), dl, 1321 TII.get(XCore::PHI), MI->getOperand(0).getReg()) 1322 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB) 1323 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB); 1324 1325 MI->eraseFromParent(); // The pseudo instruction is gone now. 1326 return BB; 1327 } 1328 1329 //===----------------------------------------------------------------------===// 1330 // Target Optimization Hooks 1331 //===----------------------------------------------------------------------===// 1332 1333 SDValue XCoreTargetLowering::PerformDAGCombine(SDNode *N, 1334 DAGCombinerInfo &DCI) const { 1335 SelectionDAG &DAG = DCI.DAG; 1336 DebugLoc dl = N->getDebugLoc(); 1337 switch (N->getOpcode()) { 1338 default: break; 1339 case XCoreISD::LADD: { 1340 SDValue N0 = N->getOperand(0); 1341 SDValue N1 = N->getOperand(1); 1342 SDValue N2 = N->getOperand(2); 1343 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1344 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1345 EVT VT = N0.getValueType(); 1346 1347 // canonicalize constant to RHS 1348 if (N0C && !N1C) 1349 return DAG.getNode(XCoreISD::LADD, dl, DAG.getVTList(VT, VT), N1, N0, N2); 1350 1351 // fold (ladd 0, 0, x) -> 0, x & 1 1352 if (N0C && N0C->isNullValue() && N1C && N1C->isNullValue()) { 1353 SDValue Carry = DAG.getConstant(0, VT); 1354 SDValue Result = DAG.getNode(ISD::AND, dl, VT, N2, 1355 DAG.getConstant(1, VT)); 1356 SDValue Ops [] = { Carry, Result }; 1357 return DAG.getMergeValues(Ops, 2, dl); 1358 } 1359 1360 // fold (ladd x, 0, y) -> 0, add x, y iff carry is unused and y has only the 1361 // low bit set 1362 if (N1C && N1C->isNullValue() && N->hasNUsesOfValue(0, 0)) { 1363 APInt KnownZero, KnownOne; 1364 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(), 1365 VT.getSizeInBits() - 1); 1366 DAG.ComputeMaskedBits(N2, KnownZero, KnownOne); 1367 if ((KnownZero & Mask) == Mask) { 1368 SDValue Carry = DAG.getConstant(0, VT); 1369 SDValue Result = DAG.getNode(ISD::ADD, dl, VT, N0, N2); 1370 SDValue Ops [] = { Carry, Result }; 1371 return DAG.getMergeValues(Ops, 2, dl); 1372 } 1373 } 1374 } 1375 break; 1376 case XCoreISD::LSUB: { 1377 SDValue N0 = N->getOperand(0); 1378 SDValue N1 = N->getOperand(1); 1379 SDValue N2 = N->getOperand(2); 1380 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1381 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1382 EVT VT = N0.getValueType(); 1383 1384 // fold (lsub 0, 0, x) -> x, -x iff x has only the low bit set 1385 if (N0C && N0C->isNullValue() && N1C && N1C->isNullValue()) { 1386 APInt KnownZero, KnownOne; 1387 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(), 1388 VT.getSizeInBits() - 1); 1389 DAG.ComputeMaskedBits(N2, KnownZero, KnownOne); 1390 if ((KnownZero & Mask) == Mask) { 1391 SDValue Borrow = N2; 1392 SDValue Result = DAG.getNode(ISD::SUB, dl, VT, 1393 DAG.getConstant(0, VT), N2); 1394 SDValue Ops [] = { Borrow, Result }; 1395 return DAG.getMergeValues(Ops, 2, dl); 1396 } 1397 } 1398 1399 // fold (lsub x, 0, y) -> 0, sub x, y iff borrow is unused and y has only the 1400 // low bit set 1401 if (N1C && N1C->isNullValue() && N->hasNUsesOfValue(0, 0)) { 1402 APInt KnownZero, KnownOne; 1403 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(), 1404 VT.getSizeInBits() - 1); 1405 DAG.ComputeMaskedBits(N2, KnownZero, KnownOne); 1406 if ((KnownZero & Mask) == Mask) { 1407 SDValue Borrow = DAG.getConstant(0, VT); 1408 SDValue Result = DAG.getNode(ISD::SUB, dl, VT, N0, N2); 1409 SDValue Ops [] = { Borrow, Result }; 1410 return DAG.getMergeValues(Ops, 2, dl); 1411 } 1412 } 1413 } 1414 break; 1415 case XCoreISD::LMUL: { 1416 SDValue N0 = N->getOperand(0); 1417 SDValue N1 = N->getOperand(1); 1418 SDValue N2 = N->getOperand(2); 1419 SDValue N3 = N->getOperand(3); 1420 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1421 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1422 EVT VT = N0.getValueType(); 1423 // Canonicalize multiplicative constant to RHS. If both multiplicative 1424 // operands are constant canonicalize smallest to RHS. 1425 if ((N0C && !N1C) || 1426 (N0C && N1C && N0C->getZExtValue() < N1C->getZExtValue())) 1427 return DAG.getNode(XCoreISD::LMUL, dl, DAG.getVTList(VT, VT), 1428 N1, N0, N2, N3); 1429 1430 // lmul(x, 0, a, b) 1431 if (N1C && N1C->isNullValue()) { 1432 // If the high result is unused fold to add(a, b) 1433 if (N->hasNUsesOfValue(0, 0)) { 1434 SDValue Lo = DAG.getNode(ISD::ADD, dl, VT, N2, N3); 1435 SDValue Ops [] = { Lo, Lo }; 1436 return DAG.getMergeValues(Ops, 2, dl); 1437 } 1438 // Otherwise fold to ladd(a, b, 0) 1439 return DAG.getNode(XCoreISD::LADD, dl, DAG.getVTList(VT, VT), N2, N3, N1); 1440 } 1441 } 1442 break; 1443 case ISD::ADD: { 1444 // Fold 32 bit expressions such as add(add(mul(x,y),a),b) -> 1445 // lmul(x, y, a, b). The high result of lmul will be ignored. 1446 // This is only profitable if the intermediate results are unused 1447 // elsewhere. 1448 SDValue Mul0, Mul1, Addend0, Addend1; 1449 if (N->getValueType(0) == MVT::i32 && 1450 isADDADDMUL(SDValue(N, 0), Mul0, Mul1, Addend0, Addend1, true)) { 1451 SDValue Ignored = DAG.getNode(XCoreISD::LMUL, dl, 1452 DAG.getVTList(MVT::i32, MVT::i32), Mul0, 1453 Mul1, Addend0, Addend1); 1454 SDValue Result(Ignored.getNode(), 1); 1455 return Result; 1456 } 1457 APInt HighMask = APInt::getHighBitsSet(64, 32); 1458 // Fold 64 bit expression such as add(add(mul(x,y),a),b) -> 1459 // lmul(x, y, a, b) if all operands are zero-extended. We do this 1460 // before type legalization as it is messy to match the operands after 1461 // that. 1462 if (N->getValueType(0) == MVT::i64 && 1463 isADDADDMUL(SDValue(N, 0), Mul0, Mul1, Addend0, Addend1, false) && 1464 DAG.MaskedValueIsZero(Mul0, HighMask) && 1465 DAG.MaskedValueIsZero(Mul1, HighMask) && 1466 DAG.MaskedValueIsZero(Addend0, HighMask) && 1467 DAG.MaskedValueIsZero(Addend1, HighMask)) { 1468 SDValue Mul0L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 1469 Mul0, DAG.getConstant(0, MVT::i32)); 1470 SDValue Mul1L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 1471 Mul1, DAG.getConstant(0, MVT::i32)); 1472 SDValue Addend0L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 1473 Addend0, DAG.getConstant(0, MVT::i32)); 1474 SDValue Addend1L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 1475 Addend1, DAG.getConstant(0, MVT::i32)); 1476 SDValue Hi = DAG.getNode(XCoreISD::LMUL, dl, 1477 DAG.getVTList(MVT::i32, MVT::i32), Mul0L, Mul1L, 1478 Addend0L, Addend1L); 1479 SDValue Lo(Hi.getNode(), 1); 1480 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi); 1481 } 1482 } 1483 break; 1484 case ISD::STORE: { 1485 // Replace unaligned store of unaligned load with memmove. 1486 StoreSDNode *ST = cast<StoreSDNode>(N); 1487 if (!DCI.isBeforeLegalize() || 1488 allowsUnalignedMemoryAccesses(ST->getMemoryVT()) || 1489 ST->isVolatile() || ST->isIndexed()) { 1490 break; 1491 } 1492 SDValue Chain = ST->getChain(); 1493 1494 unsigned StoreBits = ST->getMemoryVT().getStoreSizeInBits(); 1495 if (StoreBits % 8) { 1496 break; 1497 } 1498 unsigned ABIAlignment = getTargetData()->getABITypeAlignment( 1499 ST->getMemoryVT().getTypeForEVT(*DCI.DAG.getContext())); 1500 unsigned Alignment = ST->getAlignment(); 1501 if (Alignment >= ABIAlignment) { 1502 break; 1503 } 1504 1505 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(ST->getValue())) { 1506 if (LD->hasNUsesOfValue(1, 0) && ST->getMemoryVT() == LD->getMemoryVT() && 1507 LD->getAlignment() == Alignment && 1508 !LD->isVolatile() && !LD->isIndexed() && 1509 Chain.reachesChainWithoutSideEffects(SDValue(LD, 1))) { 1510 return DAG.getMemmove(Chain, dl, ST->getBasePtr(), 1511 LD->getBasePtr(), 1512 DAG.getConstant(StoreBits/8, MVT::i32), 1513 Alignment, false, ST->getPointerInfo(), 1514 LD->getPointerInfo()); 1515 } 1516 } 1517 break; 1518 } 1519 } 1520 return SDValue(); 1521 } 1522 1523 void XCoreTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 1524 APInt &KnownZero, 1525 APInt &KnownOne, 1526 const SelectionDAG &DAG, 1527 unsigned Depth) const { 1528 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0); 1529 switch (Op.getOpcode()) { 1530 default: break; 1531 case XCoreISD::LADD: 1532 case XCoreISD::LSUB: 1533 if (Op.getResNo() == 0) { 1534 // Top bits of carry / borrow are clear. 1535 KnownZero = APInt::getHighBitsSet(KnownZero.getBitWidth(), 1536 KnownZero.getBitWidth() - 1); 1537 } 1538 break; 1539 } 1540 } 1541 1542 //===----------------------------------------------------------------------===// 1543 // Addressing mode description hooks 1544 //===----------------------------------------------------------------------===// 1545 1546 static inline bool isImmUs(int64_t val) 1547 { 1548 return (val >= 0 && val <= 11); 1549 } 1550 1551 static inline bool isImmUs2(int64_t val) 1552 { 1553 return (val%2 == 0 && isImmUs(val/2)); 1554 } 1555 1556 static inline bool isImmUs4(int64_t val) 1557 { 1558 return (val%4 == 0 && isImmUs(val/4)); 1559 } 1560 1561 /// isLegalAddressingMode - Return true if the addressing mode represented 1562 /// by AM is legal for this target, for a load/store of the specified type. 1563 bool 1564 XCoreTargetLowering::isLegalAddressingMode(const AddrMode &AM, 1565 Type *Ty) const { 1566 if (Ty->getTypeID() == Type::VoidTyID) 1567 return AM.Scale == 0 && isImmUs(AM.BaseOffs) && isImmUs4(AM.BaseOffs); 1568 1569 const TargetData *TD = TM.getTargetData(); 1570 unsigned Size = TD->getTypeAllocSize(Ty); 1571 if (AM.BaseGV) { 1572 return Size >= 4 && !AM.HasBaseReg && AM.Scale == 0 && 1573 AM.BaseOffs%4 == 0; 1574 } 1575 1576 switch (Size) { 1577 case 1: 1578 // reg + imm 1579 if (AM.Scale == 0) { 1580 return isImmUs(AM.BaseOffs); 1581 } 1582 // reg + reg 1583 return AM.Scale == 1 && AM.BaseOffs == 0; 1584 case 2: 1585 case 3: 1586 // reg + imm 1587 if (AM.Scale == 0) { 1588 return isImmUs2(AM.BaseOffs); 1589 } 1590 // reg + reg<<1 1591 return AM.Scale == 2 && AM.BaseOffs == 0; 1592 default: 1593 // reg + imm 1594 if (AM.Scale == 0) { 1595 return isImmUs4(AM.BaseOffs); 1596 } 1597 // reg + reg<<2 1598 return AM.Scale == 4 && AM.BaseOffs == 0; 1599 } 1600 } 1601 1602 //===----------------------------------------------------------------------===// 1603 // XCore Inline Assembly Support 1604 //===----------------------------------------------------------------------===// 1605 1606 std::pair<unsigned, const TargetRegisterClass*> 1607 XCoreTargetLowering:: 1608 getRegForInlineAsmConstraint(const std::string &Constraint, 1609 EVT VT) const { 1610 if (Constraint.size() == 1) { 1611 switch (Constraint[0]) { 1612 default : break; 1613 case 'r': 1614 return std::make_pair(0U, XCore::GRRegsRegisterClass); 1615 } 1616 } 1617 // Use the default implementation in TargetLowering to convert the register 1618 // constraint into a member of a register class. 1619 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 1620 } 1621