SystemZISelLowering.cpp revision fb2e752e4175920d0531f2afc93a23d0cdf4db14
1//===-- SystemZISelLowering.cpp - SystemZ 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 SystemZTargetLowering class. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "systemz-lower" 15 16#include "SystemZISelLowering.h" 17#include "SystemZ.h" 18#include "SystemZTargetMachine.h" 19#include "SystemZSubtarget.h" 20#include "llvm/DerivedTypes.h" 21#include "llvm/Function.h" 22#include "llvm/Intrinsics.h" 23#include "llvm/CallingConv.h" 24#include "llvm/GlobalVariable.h" 25#include "llvm/GlobalAlias.h" 26#include "llvm/CodeGen/CallingConvLower.h" 27#include "llvm/CodeGen/MachineFrameInfo.h" 28#include "llvm/CodeGen/MachineFunction.h" 29#include "llvm/CodeGen/MachineInstrBuilder.h" 30#include "llvm/CodeGen/MachineRegisterInfo.h" 31#include "llvm/CodeGen/PseudoSourceValue.h" 32#include "llvm/CodeGen/SelectionDAGISel.h" 33#include "llvm/CodeGen/ValueTypes.h" 34#include "llvm/Target/TargetOptions.h" 35#include "llvm/Target/TargetLoweringObjectFile.h" 36#include "llvm/Support/Debug.h" 37#include "llvm/Support/raw_ostream.h" 38#include "llvm/ADT/VectorExtras.h" 39using namespace llvm; 40 41SystemZTargetLowering::SystemZTargetLowering(SystemZTargetMachine &tm) : 42 TargetLowering(tm, new TargetLoweringObjectFileELF()), 43 Subtarget(*tm.getSubtargetImpl()), TM(tm) { 44 45 RegInfo = TM.getRegisterInfo(); 46 47 // Set up the register classes. 48 addRegisterClass(MVT::i32, SystemZ::GR32RegisterClass); 49 addRegisterClass(MVT::i64, SystemZ::GR64RegisterClass); 50 addRegisterClass(MVT::v2i32,SystemZ::GR64PRegisterClass); 51 addRegisterClass(MVT::v2i64,SystemZ::GR128RegisterClass); 52 53 if (!UseSoftFloat) { 54 addRegisterClass(MVT::f32, SystemZ::FP32RegisterClass); 55 addRegisterClass(MVT::f64, SystemZ::FP64RegisterClass); 56 57 addLegalFPImmediate(APFloat(+0.0)); // lzer 58 addLegalFPImmediate(APFloat(+0.0f)); // lzdr 59 addLegalFPImmediate(APFloat(-0.0)); // lzer + lner 60 addLegalFPImmediate(APFloat(-0.0f)); // lzdr + lndr 61 } 62 63 // Compute derived properties from the register classes 64 computeRegisterProperties(); 65 66 // Set shifts properties 67 setShiftAmountType(MVT::i64); 68 69 // Provide all sorts of operation actions 70 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 71 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote); 72 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote); 73 74 setLoadExtAction(ISD::SEXTLOAD, MVT::f32, Expand); 75 setLoadExtAction(ISD::ZEXTLOAD, MVT::f32, Expand); 76 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand); 77 78 setLoadExtAction(ISD::SEXTLOAD, MVT::f64, Expand); 79 setLoadExtAction(ISD::ZEXTLOAD, MVT::f64, Expand); 80 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand); 81 82 setStackPointerRegisterToSaveRestore(SystemZ::R15D); 83 setSchedulingPreference(SchedulingForLatency); 84 setBooleanContents(ZeroOrOneBooleanContent); 85 86 setOperationAction(ISD::BR_JT, MVT::Other, Expand); 87 setOperationAction(ISD::BRCOND, MVT::Other, Expand); 88 setOperationAction(ISD::BR_CC, MVT::i32, Custom); 89 setOperationAction(ISD::BR_CC, MVT::i64, Custom); 90 setOperationAction(ISD::BR_CC, MVT::f32, Custom); 91 setOperationAction(ISD::BR_CC, MVT::f64, Custom); 92 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 93 setOperationAction(ISD::ConstantPool, MVT::i64, Custom); 94 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); 95 setOperationAction(ISD::JumpTable, MVT::i64, Custom); 96 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand); 97 98 setOperationAction(ISD::SDIV, MVT::i32, Expand); 99 setOperationAction(ISD::UDIV, MVT::i32, Expand); 100 setOperationAction(ISD::SDIV, MVT::i64, Expand); 101 setOperationAction(ISD::UDIV, MVT::i64, Expand); 102 setOperationAction(ISD::SREM, MVT::i32, Expand); 103 setOperationAction(ISD::UREM, MVT::i32, Expand); 104 setOperationAction(ISD::SREM, MVT::i64, Expand); 105 setOperationAction(ISD::UREM, MVT::i64, Expand); 106 107 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 108 109 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 110 setOperationAction(ISD::CTPOP, MVT::i64, Expand); 111 setOperationAction(ISD::CTTZ, MVT::i32, Expand); 112 setOperationAction(ISD::CTTZ, MVT::i64, Expand); 113 setOperationAction(ISD::CTLZ, MVT::i32, Promote); 114 setOperationAction(ISD::CTLZ, MVT::i64, Legal); 115 116 // FIXME: Can we lower these 2 efficiently? 117 setOperationAction(ISD::SETCC, MVT::i32, Expand); 118 setOperationAction(ISD::SETCC, MVT::i64, Expand); 119 setOperationAction(ISD::SETCC, MVT::f32, Expand); 120 setOperationAction(ISD::SETCC, MVT::f64, Expand); 121 setOperationAction(ISD::SELECT, MVT::i32, Expand); 122 setOperationAction(ISD::SELECT, MVT::i64, Expand); 123 setOperationAction(ISD::SELECT, MVT::f32, Expand); 124 setOperationAction(ISD::SELECT, MVT::f64, Expand); 125 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom); 126 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom); 127 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 128 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 129 130 setOperationAction(ISD::MULHS, MVT::i64, Expand); 131 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand); 132 133 // FIXME: Can we support these natively? 134 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand); 135 setOperationAction(ISD::SRL_PARTS, MVT::i64, Expand); 136 setOperationAction(ISD::SHL_PARTS, MVT::i64, Expand); 137 setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand); 138 139 // Lower some FP stuff 140 setOperationAction(ISD::FSIN, MVT::f32, Expand); 141 setOperationAction(ISD::FSIN, MVT::f64, Expand); 142 setOperationAction(ISD::FCOS, MVT::f32, Expand); 143 setOperationAction(ISD::FCOS, MVT::f64, Expand); 144 setOperationAction(ISD::FREM, MVT::f32, Expand); 145 setOperationAction(ISD::FREM, MVT::f64, Expand); 146 147 // We have only 64-bit bitconverts 148 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand); 149 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand); 150 151 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand); 152 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand); 153 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand); 154 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand); 155 156 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 157} 158 159SDValue SystemZTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) { 160 switch (Op.getOpcode()) { 161 case ISD::BR_CC: return LowerBR_CC(Op, DAG); 162 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG); 163 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG); 164 case ISD::JumpTable: return LowerJumpTable(Op, DAG); 165 case ISD::ConstantPool: return LowerConstantPool(Op, DAG); 166 default: 167 llvm_unreachable("Should not custom lower this!"); 168 return SDValue(); 169 } 170} 171 172//===----------------------------------------------------------------------===// 173// SystemZ Inline Assembly Support 174//===----------------------------------------------------------------------===// 175 176/// getConstraintType - Given a constraint letter, return the type of 177/// constraint it is for this target. 178TargetLowering::ConstraintType 179SystemZTargetLowering::getConstraintType(const std::string &Constraint) const { 180 if (Constraint.size() == 1) { 181 switch (Constraint[0]) { 182 case 'r': 183 return C_RegisterClass; 184 default: 185 break; 186 } 187 } 188 return TargetLowering::getConstraintType(Constraint); 189} 190 191std::pair<unsigned, const TargetRegisterClass*> 192SystemZTargetLowering:: 193getRegForInlineAsmConstraint(const std::string &Constraint, 194 EVT VT) const { 195 if (Constraint.size() == 1) { 196 // GCC Constraint Letters 197 switch (Constraint[0]) { 198 default: break; 199 case 'r': // GENERAL_REGS 200 if (VT == MVT::i32) 201 return std::make_pair(0U, SystemZ::GR32RegisterClass); 202 else if (VT == MVT::i128) 203 return std::make_pair(0U, SystemZ::GR128RegisterClass); 204 205 return std::make_pair(0U, SystemZ::GR64RegisterClass); 206 } 207 } 208 209 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 210} 211 212//===----------------------------------------------------------------------===// 213// Calling Convention Implementation 214//===----------------------------------------------------------------------===// 215 216#include "SystemZGenCallingConv.inc" 217 218SDValue 219SystemZTargetLowering::LowerFormalArguments(SDValue Chain, 220 CallingConv::ID CallConv, 221 bool isVarArg, 222 const SmallVectorImpl<ISD::InputArg> 223 &Ins, 224 DebugLoc dl, 225 SelectionDAG &DAG, 226 SmallVectorImpl<SDValue> &InVals) { 227 228 switch (CallConv) { 229 default: 230 llvm_unreachable("Unsupported calling convention"); 231 case CallingConv::C: 232 case CallingConv::Fast: 233 return LowerCCCArguments(Chain, CallConv, isVarArg, Ins, dl, DAG, InVals); 234 } 235} 236 237SDValue 238SystemZTargetLowering::LowerCall(SDValue Chain, SDValue Callee, 239 CallingConv::ID CallConv, bool isVarArg, 240 bool isTailCall, 241 const SmallVectorImpl<ISD::OutputArg> &Outs, 242 const SmallVectorImpl<ISD::InputArg> &Ins, 243 DebugLoc dl, SelectionDAG &DAG, 244 SmallVectorImpl<SDValue> &InVals) { 245 246 switch (CallConv) { 247 default: 248 llvm_unreachable("Unsupported calling convention"); 249 case CallingConv::Fast: 250 case CallingConv::C: 251 return LowerCCCCallTo(Chain, Callee, CallConv, isVarArg, isTailCall, 252 Outs, Ins, dl, DAG, InVals); 253 } 254} 255 256/// LowerCCCArguments - transform physical registers into virtual registers and 257/// generate load operations for arguments places on the stack. 258// FIXME: struct return stuff 259// FIXME: varargs 260SDValue 261SystemZTargetLowering::LowerCCCArguments(SDValue Chain, 262 CallingConv::ID CallConv, 263 bool isVarArg, 264 const SmallVectorImpl<ISD::InputArg> 265 &Ins, 266 DebugLoc dl, 267 SelectionDAG &DAG, 268 SmallVectorImpl<SDValue> &InVals) { 269 270 MachineFunction &MF = DAG.getMachineFunction(); 271 MachineFrameInfo *MFI = MF.getFrameInfo(); 272 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 273 274 // Assign locations to all of the incoming arguments. 275 SmallVector<CCValAssign, 16> ArgLocs; 276 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), 277 ArgLocs, *DAG.getContext()); 278 CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ); 279 280 if (isVarArg) 281 llvm_report_error("Varargs not supported yet"); 282 283 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 284 SDValue ArgValue; 285 CCValAssign &VA = ArgLocs[i]; 286 EVT LocVT = VA.getLocVT(); 287 if (VA.isRegLoc()) { 288 // Arguments passed in registers 289 TargetRegisterClass *RC; 290 switch (LocVT.getSimpleVT().SimpleTy) { 291 default: 292#ifndef NDEBUG 293 errs() << "LowerFormalArguments Unhandled argument type: " 294 << LocVT.getSimpleVT().SimpleTy 295 << "\n"; 296#endif 297 llvm_unreachable(0); 298 case MVT::i64: 299 RC = SystemZ::GR64RegisterClass; 300 break; 301 case MVT::f32: 302 RC = SystemZ::FP32RegisterClass; 303 break; 304 case MVT::f64: 305 RC = SystemZ::FP64RegisterClass; 306 break; 307 } 308 309 unsigned VReg = RegInfo.createVirtualRegister(RC); 310 RegInfo.addLiveIn(VA.getLocReg(), VReg); 311 ArgValue = DAG.getCopyFromReg(Chain, dl, VReg, LocVT); 312 } else { 313 // Sanity check 314 assert(VA.isMemLoc()); 315 316 // Create the nodes corresponding to a load from this parameter slot. 317 // Create the frame index object for this incoming parameter... 318 int FI = MFI->CreateFixedObject(LocVT.getSizeInBits()/8, 319 VA.getLocMemOffset()); 320 321 // Create the SelectionDAG nodes corresponding to a load 322 // from this parameter 323 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 324 ArgValue = DAG.getLoad(LocVT, dl, Chain, FIN, 325 PseudoSourceValue::getFixedStack(FI), 0); 326 } 327 328 // If this is an 8/16/32-bit value, it is really passed promoted to 64 329 // bits. Insert an assert[sz]ext to capture this, then truncate to the 330 // right size. 331 if (VA.getLocInfo() == CCValAssign::SExt) 332 ArgValue = DAG.getNode(ISD::AssertSext, dl, LocVT, ArgValue, 333 DAG.getValueType(VA.getValVT())); 334 else if (VA.getLocInfo() == CCValAssign::ZExt) 335 ArgValue = DAG.getNode(ISD::AssertZext, dl, LocVT, ArgValue, 336 DAG.getValueType(VA.getValVT())); 337 338 if (VA.getLocInfo() != CCValAssign::Full) 339 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 340 341 InVals.push_back(ArgValue); 342 } 343 344 return Chain; 345} 346 347/// LowerCCCCallTo - functions arguments are copied from virtual regs to 348/// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted. 349/// TODO: sret. 350SDValue 351SystemZTargetLowering::LowerCCCCallTo(SDValue Chain, SDValue Callee, 352 CallingConv::ID CallConv, bool isVarArg, 353 bool isTailCall, 354 const SmallVectorImpl<ISD::OutputArg> 355 &Outs, 356 const SmallVectorImpl<ISD::InputArg> &Ins, 357 DebugLoc dl, SelectionDAG &DAG, 358 SmallVectorImpl<SDValue> &InVals) { 359 360 MachineFunction &MF = DAG.getMachineFunction(); 361 362 // Offset to first argument stack slot. 363 const unsigned FirstArgOffset = 160; 364 365 // Analyze operands of the call, assigning locations to each operand. 366 SmallVector<CCValAssign, 16> ArgLocs; 367 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), 368 ArgLocs, *DAG.getContext()); 369 370 CCInfo.AnalyzeCallOperands(Outs, CC_SystemZ); 371 372 // Get a count of how many bytes are to be pushed on the stack. 373 unsigned NumBytes = CCInfo.getNextStackOffset(); 374 375 Chain = DAG.getCALLSEQ_START(Chain ,DAG.getConstant(NumBytes, 376 getPointerTy(), true)); 377 378 SmallVector<std::pair<unsigned, SDValue>, 4> RegsToPass; 379 SmallVector<SDValue, 12> MemOpChains; 380 SDValue StackPtr; 381 382 // Walk the register/memloc assignments, inserting copies/loads. 383 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 384 CCValAssign &VA = ArgLocs[i]; 385 386 SDValue Arg = Outs[i].Val; 387 388 // Promote the value if needed. 389 switch (VA.getLocInfo()) { 390 default: assert(0 && "Unknown loc info!"); 391 case CCValAssign::Full: break; 392 case CCValAssign::SExt: 393 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); 394 break; 395 case CCValAssign::ZExt: 396 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); 397 break; 398 case CCValAssign::AExt: 399 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg); 400 break; 401 } 402 403 // Arguments that can be passed on register must be kept at RegsToPass 404 // vector 405 if (VA.isRegLoc()) { 406 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 407 } else { 408 assert(VA.isMemLoc()); 409 410 if (StackPtr.getNode() == 0) 411 StackPtr = 412 DAG.getCopyFromReg(Chain, dl, 413 (RegInfo->hasFP(MF) ? 414 SystemZ::R11D : SystemZ::R15D), 415 getPointerTy()); 416 417 unsigned Offset = FirstArgOffset + VA.getLocMemOffset(); 418 SDValue PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), 419 StackPtr, 420 DAG.getIntPtrConstant(Offset)); 421 422 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff, 423 PseudoSourceValue::getStack(), Offset)); 424 } 425 } 426 427 // Transform all store nodes into one single node because all store nodes are 428 // independent of each other. 429 if (!MemOpChains.empty()) 430 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 431 &MemOpChains[0], MemOpChains.size()); 432 433 // Build a sequence of copy-to-reg nodes chained together with token chain and 434 // flag operands which copy the outgoing args into registers. The InFlag in 435 // necessary since all emited instructions must be stuck together. 436 SDValue InFlag; 437 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 438 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, 439 RegsToPass[i].second, InFlag); 440 InFlag = Chain.getValue(1); 441 } 442 443 // If the callee is a GlobalAddress node (quite common, every direct call is) 444 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. 445 // Likewise ExternalSymbol -> TargetExternalSymbol. 446 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) 447 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy()); 448 else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee)) 449 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), getPointerTy()); 450 451 // Returns a chain & a flag for retval copy to use. 452 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag); 453 SmallVector<SDValue, 8> Ops; 454 Ops.push_back(Chain); 455 Ops.push_back(Callee); 456 457 // Add argument registers to the end of the list so that they are 458 // known live into the call. 459 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 460 Ops.push_back(DAG.getRegister(RegsToPass[i].first, 461 RegsToPass[i].second.getValueType())); 462 463 if (InFlag.getNode()) 464 Ops.push_back(InFlag); 465 466 Chain = DAG.getNode(SystemZISD::CALL, dl, NodeTys, &Ops[0], Ops.size()); 467 InFlag = Chain.getValue(1); 468 469 // Create the CALLSEQ_END node. 470 Chain = DAG.getCALLSEQ_END(Chain, 471 DAG.getConstant(NumBytes, getPointerTy(), true), 472 DAG.getConstant(0, getPointerTy(), true), 473 InFlag); 474 InFlag = Chain.getValue(1); 475 476 // Handle result values, copying them out of physregs into vregs that we 477 // return. 478 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, 479 DAG, InVals); 480} 481 482/// LowerCallResult - Lower the result values of a call into the 483/// appropriate copies out of appropriate physical registers. 484/// 485SDValue 486SystemZTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, 487 CallingConv::ID CallConv, bool isVarArg, 488 const SmallVectorImpl<ISD::InputArg> 489 &Ins, 490 DebugLoc dl, SelectionDAG &DAG, 491 SmallVectorImpl<SDValue> &InVals) { 492 493 // Assign locations to each value returned by this call. 494 SmallVector<CCValAssign, 16> RVLocs; 495 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs, 496 *DAG.getContext()); 497 498 CCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ); 499 500 // Copy all of the result registers out of their specified physreg. 501 for (unsigned i = 0; i != RVLocs.size(); ++i) { 502 CCValAssign &VA = RVLocs[i]; 503 504 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), 505 VA.getLocVT(), InFlag).getValue(1); 506 SDValue RetValue = Chain.getValue(0); 507 InFlag = Chain.getValue(2); 508 509 // If this is an 8/16/32-bit value, it is really passed promoted to 64 510 // bits. Insert an assert[sz]ext to capture this, then truncate to the 511 // right size. 512 if (VA.getLocInfo() == CCValAssign::SExt) 513 RetValue = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), RetValue, 514 DAG.getValueType(VA.getValVT())); 515 else if (VA.getLocInfo() == CCValAssign::ZExt) 516 RetValue = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), RetValue, 517 DAG.getValueType(VA.getValVT())); 518 519 if (VA.getLocInfo() != CCValAssign::Full) 520 RetValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), RetValue); 521 522 InVals.push_back(RetValue); 523 } 524 525 return Chain; 526} 527 528 529SDValue 530SystemZTargetLowering::LowerReturn(SDValue Chain, 531 CallingConv::ID CallConv, bool isVarArg, 532 const SmallVectorImpl<ISD::OutputArg> &Outs, 533 DebugLoc dl, SelectionDAG &DAG) { 534 535 // CCValAssign - represent the assignment of the return value to a location 536 SmallVector<CCValAssign, 16> RVLocs; 537 538 // CCState - Info about the registers and stack slot. 539 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), 540 RVLocs, *DAG.getContext()); 541 542 // Analize return values. 543 CCInfo.AnalyzeReturn(Outs, RetCC_SystemZ); 544 545 // If this is the first return lowered for this function, add the regs to the 546 // liveout set for the function. 547 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) { 548 for (unsigned i = 0; i != RVLocs.size(); ++i) 549 if (RVLocs[i].isRegLoc()) 550 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg()); 551 } 552 553 SDValue Flag; 554 555 // Copy the result values into the output registers. 556 for (unsigned i = 0; i != RVLocs.size(); ++i) { 557 CCValAssign &VA = RVLocs[i]; 558 SDValue ResValue = Outs[i].Val; 559 assert(VA.isRegLoc() && "Can only return in registers!"); 560 561 // If this is an 8/16/32-bit value, it is really should be passed promoted 562 // to 64 bits. 563 if (VA.getLocInfo() == CCValAssign::SExt) 564 ResValue = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ResValue); 565 else if (VA.getLocInfo() == CCValAssign::ZExt) 566 ResValue = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ResValue); 567 else if (VA.getLocInfo() == CCValAssign::AExt) 568 ResValue = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ResValue); 569 570 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ResValue, Flag); 571 572 // Guarantee that all emitted copies are stuck together, 573 // avoiding something bad. 574 Flag = Chain.getValue(1); 575 } 576 577 if (Flag.getNode()) 578 return DAG.getNode(SystemZISD::RET_FLAG, dl, MVT::Other, Chain, Flag); 579 580 // Return Void 581 return DAG.getNode(SystemZISD::RET_FLAG, dl, MVT::Other, Chain); 582} 583 584SDValue SystemZTargetLowering::EmitCmp(SDValue LHS, SDValue RHS, 585 ISD::CondCode CC, SDValue &SystemZCC, 586 SelectionDAG &DAG) { 587 // FIXME: Emit a test if RHS is zero 588 589 bool isUnsigned = false; 590 SystemZCC::CondCodes TCC; 591 switch (CC) { 592 default: 593 llvm_unreachable("Invalid integer condition!"); 594 case ISD::SETEQ: 595 case ISD::SETOEQ: 596 TCC = SystemZCC::E; 597 break; 598 case ISD::SETUEQ: 599 TCC = SystemZCC::NLH; 600 break; 601 case ISD::SETNE: 602 case ISD::SETONE: 603 TCC = SystemZCC::NE; 604 break; 605 case ISD::SETUNE: 606 TCC = SystemZCC::LH; 607 break; 608 case ISD::SETO: 609 TCC = SystemZCC::O; 610 break; 611 case ISD::SETUO: 612 TCC = SystemZCC::NO; 613 break; 614 case ISD::SETULE: 615 if (LHS.getValueType().isFloatingPoint()) { 616 TCC = SystemZCC::NH; 617 break; 618 } 619 isUnsigned = true; // FALLTHROUGH 620 case ISD::SETLE: 621 case ISD::SETOLE: 622 TCC = SystemZCC::LE; 623 break; 624 case ISD::SETUGE: 625 if (LHS.getValueType().isFloatingPoint()) { 626 TCC = SystemZCC::NL; 627 break; 628 } 629 isUnsigned = true; // FALLTHROUGH 630 case ISD::SETGE: 631 case ISD::SETOGE: 632 TCC = SystemZCC::HE; 633 break; 634 case ISD::SETUGT: 635 if (LHS.getValueType().isFloatingPoint()) { 636 TCC = SystemZCC::NLE; 637 break; 638 } 639 isUnsigned = true; // FALLTHROUGH 640 case ISD::SETGT: 641 case ISD::SETOGT: 642 TCC = SystemZCC::H; 643 break; 644 case ISD::SETULT: 645 if (LHS.getValueType().isFloatingPoint()) { 646 TCC = SystemZCC::NHE; 647 break; 648 } 649 isUnsigned = true; // FALLTHROUGH 650 case ISD::SETLT: 651 case ISD::SETOLT: 652 TCC = SystemZCC::L; 653 break; 654 } 655 656 SystemZCC = DAG.getConstant(TCC, MVT::i32); 657 658 DebugLoc dl = LHS.getDebugLoc(); 659 return DAG.getNode((isUnsigned ? SystemZISD::UCMP : SystemZISD::CMP), 660 dl, MVT::Flag, LHS, RHS); 661} 662 663 664SDValue SystemZTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) { 665 SDValue Chain = Op.getOperand(0); 666 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 667 SDValue LHS = Op.getOperand(2); 668 SDValue RHS = Op.getOperand(3); 669 SDValue Dest = Op.getOperand(4); 670 DebugLoc dl = Op.getDebugLoc(); 671 672 SDValue SystemZCC; 673 SDValue Flag = EmitCmp(LHS, RHS, CC, SystemZCC, DAG); 674 return DAG.getNode(SystemZISD::BRCOND, dl, Op.getValueType(), 675 Chain, Dest, SystemZCC, Flag); 676} 677 678SDValue SystemZTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) { 679 SDValue LHS = Op.getOperand(0); 680 SDValue RHS = Op.getOperand(1); 681 SDValue TrueV = Op.getOperand(2); 682 SDValue FalseV = Op.getOperand(3); 683 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get(); 684 DebugLoc dl = Op.getDebugLoc(); 685 686 SDValue SystemZCC; 687 SDValue Flag = EmitCmp(LHS, RHS, CC, SystemZCC, DAG); 688 689 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag); 690 SmallVector<SDValue, 4> Ops; 691 Ops.push_back(TrueV); 692 Ops.push_back(FalseV); 693 Ops.push_back(SystemZCC); 694 Ops.push_back(Flag); 695 696 return DAG.getNode(SystemZISD::SELECT, dl, VTs, &Ops[0], Ops.size()); 697} 698 699SDValue SystemZTargetLowering::LowerGlobalAddress(SDValue Op, 700 SelectionDAG &DAG) { 701 DebugLoc dl = Op.getDebugLoc(); 702 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 703 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset(); 704 705 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_; 706 bool ExtraLoadRequired = 707 Subtarget.GVRequiresExtraLoad(GV, getTargetMachine(), false); 708 709 SDValue Result; 710 if (!IsPic && !ExtraLoadRequired) { 711 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset); 712 Offset = 0; 713 } else { 714 unsigned char OpFlags = 0; 715 if (ExtraLoadRequired) 716 OpFlags = SystemZII::MO_GOTENT; 717 718 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags); 719 } 720 721 Result = DAG.getNode(SystemZISD::PCRelativeWrapper, dl, 722 getPointerTy(), Result); 723 724 if (ExtraLoadRequired) 725 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result, 726 PseudoSourceValue::getGOT(), 0); 727 728 // If there was a non-zero offset that we didn't fold, create an explicit 729 // addition for it. 730 if (Offset != 0) 731 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result, 732 DAG.getConstant(Offset, getPointerTy())); 733 734 return Result; 735} 736 737// FIXME: PIC here 738SDValue SystemZTargetLowering::LowerJumpTable(SDValue Op, 739 SelectionDAG &DAG) { 740 DebugLoc dl = Op.getDebugLoc(); 741 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op); 742 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy()); 743 744 return DAG.getNode(SystemZISD::PCRelativeWrapper, dl, getPointerTy(), Result); 745} 746 747 748// FIXME: PIC here 749// FIXME: This is just dirty hack. We need to lower cpool properly 750SDValue SystemZTargetLowering::LowerConstantPool(SDValue Op, 751 SelectionDAG &DAG) { 752 DebugLoc dl = Op.getDebugLoc(); 753 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); 754 755 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(), 756 CP->getAlignment(), 757 CP->getOffset()); 758 759 return DAG.getNode(SystemZISD::PCRelativeWrapper, dl, getPointerTy(), Result); 760} 761 762const char *SystemZTargetLowering::getTargetNodeName(unsigned Opcode) const { 763 switch (Opcode) { 764 case SystemZISD::RET_FLAG: return "SystemZISD::RET_FLAG"; 765 case SystemZISD::CALL: return "SystemZISD::CALL"; 766 case SystemZISD::BRCOND: return "SystemZISD::BRCOND"; 767 case SystemZISD::CMP: return "SystemZISD::CMP"; 768 case SystemZISD::UCMP: return "SystemZISD::UCMP"; 769 case SystemZISD::SELECT: return "SystemZISD::SELECT"; 770 case SystemZISD::PCRelativeWrapper: return "SystemZISD::PCRelativeWrapper"; 771 default: return NULL; 772 } 773} 774 775//===----------------------------------------------------------------------===// 776// Other Lowering Code 777//===----------------------------------------------------------------------===// 778 779MachineBasicBlock* 780SystemZTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 781 MachineBasicBlock *BB, 782 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const { 783 const SystemZInstrInfo &TII = *TM.getInstrInfo(); 784 DebugLoc dl = MI->getDebugLoc(); 785 assert((MI->getOpcode() == SystemZ::Select32 || 786 MI->getOpcode() == SystemZ::SelectF32 || 787 MI->getOpcode() == SystemZ::Select64 || 788 MI->getOpcode() == SystemZ::SelectF64) && 789 "Unexpected instr type to insert"); 790 791 // To "insert" a SELECT instruction, we actually have to insert the diamond 792 // control-flow pattern. The incoming instruction knows the destination vreg 793 // to set, the condition code register to branch on, the true/false values to 794 // select between, and a branch opcode to use. 795 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 796 MachineFunction::iterator I = BB; 797 ++I; 798 799 // thisMBB: 800 // ... 801 // TrueVal = ... 802 // cmpTY ccX, r1, r2 803 // jCC copy1MBB 804 // fallthrough --> copy0MBB 805 MachineBasicBlock *thisMBB = BB; 806 MachineFunction *F = BB->getParent(); 807 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 808 MachineBasicBlock *copy1MBB = F->CreateMachineBasicBlock(LLVM_BB); 809 SystemZCC::CondCodes CC = (SystemZCC::CondCodes)MI->getOperand(3).getImm(); 810 BuildMI(BB, dl, TII.getBrCond(CC)).addMBB(copy1MBB); 811 F->insert(I, copy0MBB); 812 F->insert(I, copy1MBB); 813 // Update machine-CFG edges by transferring all successors of the current 814 // block to the new block which will contain the Phi node for the select. 815 copy1MBB->transferSuccessors(BB); 816 // Next, add the true and fallthrough blocks as its successors. 817 BB->addSuccessor(copy0MBB); 818 BB->addSuccessor(copy1MBB); 819 820 // copy0MBB: 821 // %FalseValue = ... 822 // # fallthrough to copy1MBB 823 BB = copy0MBB; 824 825 // Update machine-CFG edges 826 BB->addSuccessor(copy1MBB); 827 828 // copy1MBB: 829 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] 830 // ... 831 BB = copy1MBB; 832 BuildMI(BB, dl, TII.get(SystemZ::PHI), 833 MI->getOperand(0).getReg()) 834 .addReg(MI->getOperand(2).getReg()).addMBB(copy0MBB) 835 .addReg(MI->getOperand(1).getReg()).addMBB(thisMBB); 836 837 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now. 838 return BB; 839} 840