SystemZISelLowering.cpp revision 3c2734c82bbd9fd6524a541a85c4a228c6648b9d
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/ADT/VectorExtras.h" 38using namespace llvm; 39 40SystemZTargetLowering::SystemZTargetLowering(SystemZTargetMachine &tm) : 41 TargetLowering(tm, new TargetLoweringObjectFileELF()), 42 Subtarget(*tm.getSubtargetImpl()), TM(tm) { 43 44 RegInfo = TM.getRegisterInfo(); 45 46 // Set up the register classes. 47 addRegisterClass(MVT::i32, SystemZ::GR32RegisterClass); 48 addRegisterClass(MVT::i64, SystemZ::GR64RegisterClass); 49 addRegisterClass(MVT::v2i32,SystemZ::GR64PRegisterClass); 50 addRegisterClass(MVT::v2i64,SystemZ::GR128RegisterClass); 51 52 if (!UseSoftFloat) { 53 addRegisterClass(MVT::f32, SystemZ::FP32RegisterClass); 54 addRegisterClass(MVT::f64, SystemZ::FP64RegisterClass); 55 56 addLegalFPImmediate(APFloat(+0.0)); // lzer 57 addLegalFPImmediate(APFloat(+0.0f)); // lzdr 58 addLegalFPImmediate(APFloat(-0.0)); // lzer + lner 59 addLegalFPImmediate(APFloat(-0.0f)); // lzdr + lndr 60 } 61 62 // Compute derived properties from the register classes 63 computeRegisterProperties(); 64 65 // Set shifts properties 66 setShiftAmountType(MVT::i64); 67 68 // Provide all sorts of operation actions 69 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 70 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote); 71 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote); 72 73 setLoadExtAction(ISD::SEXTLOAD, MVT::f32, Expand); 74 setLoadExtAction(ISD::ZEXTLOAD, MVT::f32, Expand); 75 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand); 76 77 setLoadExtAction(ISD::SEXTLOAD, MVT::f64, Expand); 78 setLoadExtAction(ISD::ZEXTLOAD, MVT::f64, Expand); 79 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand); 80 81 setStackPointerRegisterToSaveRestore(SystemZ::R15D); 82 setSchedulingPreference(SchedulingForLatency); 83 setBooleanContents(ZeroOrOneBooleanContent); 84 85 setOperationAction(ISD::BR_JT, MVT::Other, Expand); 86 setOperationAction(ISD::BRCOND, MVT::Other, Expand); 87 setOperationAction(ISD::BR_CC, MVT::i32, Custom); 88 setOperationAction(ISD::BR_CC, MVT::i64, Custom); 89 setOperationAction(ISD::BR_CC, MVT::f32, Custom); 90 setOperationAction(ISD::BR_CC, MVT::f64, Custom); 91 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 92 setOperationAction(ISD::ConstantPool, MVT::i64, Custom); 93 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); 94 setOperationAction(ISD::JumpTable, MVT::i64, Custom); 95 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand); 96 97 setOperationAction(ISD::SDIV, MVT::i32, Expand); 98 setOperationAction(ISD::UDIV, MVT::i32, Expand); 99 setOperationAction(ISD::SDIV, MVT::i64, Expand); 100 setOperationAction(ISD::UDIV, MVT::i64, Expand); 101 setOperationAction(ISD::SREM, MVT::i32, Expand); 102 setOperationAction(ISD::UREM, MVT::i32, Expand); 103 setOperationAction(ISD::SREM, MVT::i64, Expand); 104 setOperationAction(ISD::UREM, MVT::i64, Expand); 105 106 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 107 108 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 109 setOperationAction(ISD::CTPOP, MVT::i64, Expand); 110 setOperationAction(ISD::CTTZ, MVT::i32, Expand); 111 setOperationAction(ISD::CTTZ, MVT::i64, Expand); 112 setOperationAction(ISD::CTLZ, MVT::i32, Promote); 113 setOperationAction(ISD::CTLZ, MVT::i64, Legal); 114 115 // FIXME: Can we lower these 2 efficiently? 116 setOperationAction(ISD::SETCC, MVT::i32, Expand); 117 setOperationAction(ISD::SETCC, MVT::i64, Expand); 118 setOperationAction(ISD::SETCC, MVT::f32, Expand); 119 setOperationAction(ISD::SETCC, MVT::f64, Expand); 120 setOperationAction(ISD::SELECT, MVT::i32, Expand); 121 setOperationAction(ISD::SELECT, MVT::i64, Expand); 122 setOperationAction(ISD::SELECT, MVT::f32, Expand); 123 setOperationAction(ISD::SELECT, MVT::f64, Expand); 124 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom); 125 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom); 126 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 127 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 128 129 // Funny enough: we don't have 64-bit signed versions of these stuff, but have 130 // unsigned. 131 setOperationAction(ISD::MULHS, MVT::i64, Expand); 132 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand); 133 134 // Lower some FP stuff 135 setOperationAction(ISD::FSIN, MVT::f32, Expand); 136 setOperationAction(ISD::FSIN, MVT::f64, Expand); 137 setOperationAction(ISD::FCOS, MVT::f32, Expand); 138 setOperationAction(ISD::FCOS, MVT::f64, Expand); 139 setOperationAction(ISD::FREM, MVT::f32, Expand); 140 setOperationAction(ISD::FREM, MVT::f64, Expand); 141 142 // We have only 64-bit bitconverts 143 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand); 144 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand); 145 146 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand); 147 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand); 148 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand); 149 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand); 150 151 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 152} 153 154SDValue SystemZTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) { 155 switch (Op.getOpcode()) { 156 case ISD::BR_CC: return LowerBR_CC(Op, DAG); 157 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG); 158 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG); 159 case ISD::JumpTable: return LowerJumpTable(Op, DAG); 160 case ISD::ConstantPool: return LowerConstantPool(Op, DAG); 161 default: 162 llvm_unreachable("Should not custom lower this!"); 163 return SDValue(); 164 } 165} 166 167//===----------------------------------------------------------------------===// 168// SystemZ Inline Assembly Support 169//===----------------------------------------------------------------------===// 170 171/// getConstraintType - Given a constraint letter, return the type of 172/// constraint it is for this target. 173TargetLowering::ConstraintType 174SystemZTargetLowering::getConstraintType(const std::string &Constraint) const { 175 if (Constraint.size() == 1) { 176 switch (Constraint[0]) { 177 case 'r': 178 return C_RegisterClass; 179 default: 180 break; 181 } 182 } 183 return TargetLowering::getConstraintType(Constraint); 184} 185 186std::pair<unsigned, const TargetRegisterClass*> 187SystemZTargetLowering:: 188getRegForInlineAsmConstraint(const std::string &Constraint, 189 EVT VT) const { 190 if (Constraint.size() == 1) { 191 // GCC Constraint Letters 192 switch (Constraint[0]) { 193 default: break; 194 case 'r': // GENERAL_REGS 195 if (VT == MVT::i32) 196 return std::make_pair(0U, SystemZ::GR32RegisterClass); 197 else if (VT == MVT::i128) 198 return std::make_pair(0U, SystemZ::GR128RegisterClass); 199 200 return std::make_pair(0U, SystemZ::GR64RegisterClass); 201 } 202 } 203 204 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 205} 206 207//===----------------------------------------------------------------------===// 208// Calling Convention Implementation 209//===----------------------------------------------------------------------===// 210 211#include "SystemZGenCallingConv.inc" 212 213SDValue 214SystemZTargetLowering::LowerFormalArguments(SDValue Chain, 215 unsigned CallConv, 216 bool isVarArg, 217 const SmallVectorImpl<ISD::InputArg> 218 &Ins, 219 DebugLoc dl, 220 SelectionDAG &DAG, 221 SmallVectorImpl<SDValue> &InVals) { 222 223 switch (CallConv) { 224 default: 225 llvm_unreachable("Unsupported calling convention"); 226 case CallingConv::C: 227 case CallingConv::Fast: 228 return LowerCCCArguments(Chain, CallConv, isVarArg, Ins, dl, DAG, InVals); 229 } 230} 231 232SDValue 233SystemZTargetLowering::LowerCall(SDValue Chain, SDValue Callee, 234 unsigned CallConv, bool isVarArg, 235 bool isTailCall, 236 const SmallVectorImpl<ISD::OutputArg> &Outs, 237 const SmallVectorImpl<ISD::InputArg> &Ins, 238 DebugLoc dl, SelectionDAG &DAG, 239 SmallVectorImpl<SDValue> &InVals) { 240 241 switch (CallConv) { 242 default: 243 llvm_unreachable("Unsupported calling convention"); 244 case CallingConv::Fast: 245 case CallingConv::C: 246 return LowerCCCCallTo(Chain, Callee, CallConv, isVarArg, isTailCall, 247 Outs, Ins, dl, DAG, InVals); 248 } 249} 250 251/// LowerCCCArguments - transform physical registers into virtual registers and 252/// generate load operations for arguments places on the stack. 253// FIXME: struct return stuff 254// FIXME: varargs 255SDValue 256SystemZTargetLowering::LowerCCCArguments(SDValue Chain, 257 unsigned CallConv, 258 bool isVarArg, 259 const SmallVectorImpl<ISD::InputArg> 260 &Ins, 261 DebugLoc dl, 262 SelectionDAG &DAG, 263 SmallVectorImpl<SDValue> &InVals) { 264 265 MachineFunction &MF = DAG.getMachineFunction(); 266 MachineFrameInfo *MFI = MF.getFrameInfo(); 267 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 268 269 // Assign locations to all of the incoming arguments. 270 SmallVector<CCValAssign, 16> ArgLocs; 271 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), 272 ArgLocs, *DAG.getContext()); 273 CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ); 274 275 if (isVarArg) 276 llvm_report_error("Varargs not supported yet"); 277 278 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 279 SDValue ArgValue; 280 CCValAssign &VA = ArgLocs[i]; 281 EVT LocVT = VA.getLocVT(); 282 if (VA.isRegLoc()) { 283 // Arguments passed in registers 284 TargetRegisterClass *RC; 285 switch (LocVT.getSimpleVT().SimpleTy) { 286 default: 287#ifndef NDEBUG 288 cerr << "LowerFormalArguments Unhandled argument type: " 289 << LocVT.getSimpleVT().SimpleTy 290 << "\n"; 291#endif 292 llvm_unreachable(0); 293 case MVT::i64: 294 RC = SystemZ::GR64RegisterClass; 295 break; 296 case MVT::f32: 297 RC = SystemZ::FP32RegisterClass; 298 break; 299 case MVT::f64: 300 RC = SystemZ::FP64RegisterClass; 301 break; 302 } 303 304 unsigned VReg = RegInfo.createVirtualRegister(RC); 305 RegInfo.addLiveIn(VA.getLocReg(), VReg); 306 ArgValue = DAG.getCopyFromReg(Chain, dl, VReg, LocVT); 307 } else { 308 // Sanity check 309 assert(VA.isMemLoc()); 310 311 // Create the nodes corresponding to a load from this parameter slot. 312 // Create the frame index object for this incoming parameter... 313 int FI = MFI->CreateFixedObject(LocVT.getSizeInBits()/8, 314 VA.getLocMemOffset()); 315 316 // Create the SelectionDAG nodes corresponding to a load 317 // from this parameter 318 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 319 ArgValue = DAG.getLoad(LocVT, dl, Chain, FIN, 320 PseudoSourceValue::getFixedStack(FI), 0); 321 } 322 323 // If this is an 8/16/32-bit value, it is really passed promoted to 64 324 // bits. Insert an assert[sz]ext to capture this, then truncate to the 325 // right size. 326 if (VA.getLocInfo() == CCValAssign::SExt) 327 ArgValue = DAG.getNode(ISD::AssertSext, dl, LocVT, ArgValue, 328 DAG.getValueType(VA.getValVT())); 329 else if (VA.getLocInfo() == CCValAssign::ZExt) 330 ArgValue = DAG.getNode(ISD::AssertZext, dl, LocVT, ArgValue, 331 DAG.getValueType(VA.getValVT())); 332 333 if (VA.getLocInfo() != CCValAssign::Full) 334 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 335 336 InVals.push_back(ArgValue); 337 } 338 339 return Chain; 340} 341 342/// LowerCCCCallTo - functions arguments are copied from virtual regs to 343/// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted. 344/// TODO: sret. 345SDValue 346SystemZTargetLowering::LowerCCCCallTo(SDValue Chain, SDValue Callee, 347 unsigned CallConv, bool isVarArg, 348 bool isTailCall, 349 const SmallVectorImpl<ISD::OutputArg> 350 &Outs, 351 const SmallVectorImpl<ISD::InputArg> &Ins, 352 DebugLoc dl, SelectionDAG &DAG, 353 SmallVectorImpl<SDValue> &InVals) { 354 355 MachineFunction &MF = DAG.getMachineFunction(); 356 357 // Offset to first argument stack slot. 358 const unsigned FirstArgOffset = 160; 359 360 // Analyze operands of the call, assigning locations to each operand. 361 SmallVector<CCValAssign, 16> ArgLocs; 362 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), 363 ArgLocs, *DAG.getContext()); 364 365 CCInfo.AnalyzeCallOperands(Outs, CC_SystemZ); 366 367 // Get a count of how many bytes are to be pushed on the stack. 368 unsigned NumBytes = CCInfo.getNextStackOffset(); 369 370 Chain = DAG.getCALLSEQ_START(Chain ,DAG.getConstant(NumBytes, 371 getPointerTy(), true)); 372 373 SmallVector<std::pair<unsigned, SDValue>, 4> RegsToPass; 374 SmallVector<SDValue, 12> MemOpChains; 375 SDValue StackPtr; 376 377 // Walk the register/memloc assignments, inserting copies/loads. 378 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 379 CCValAssign &VA = ArgLocs[i]; 380 381 SDValue Arg = Outs[i].Val; 382 383 // Promote the value if needed. 384 switch (VA.getLocInfo()) { 385 default: assert(0 && "Unknown loc info!"); 386 case CCValAssign::Full: break; 387 case CCValAssign::SExt: 388 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); 389 break; 390 case CCValAssign::ZExt: 391 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); 392 break; 393 case CCValAssign::AExt: 394 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg); 395 break; 396 } 397 398 // Arguments that can be passed on register must be kept at RegsToPass 399 // vector 400 if (VA.isRegLoc()) { 401 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 402 } else { 403 assert(VA.isMemLoc()); 404 405 if (StackPtr.getNode() == 0) 406 StackPtr = 407 DAG.getCopyFromReg(Chain, dl, 408 (RegInfo->hasFP(MF) ? 409 SystemZ::R11D : SystemZ::R15D), 410 getPointerTy()); 411 412 unsigned Offset = FirstArgOffset + VA.getLocMemOffset(); 413 SDValue PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), 414 StackPtr, 415 DAG.getIntPtrConstant(Offset)); 416 417 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff, 418 PseudoSourceValue::getStack(), Offset)); 419 } 420 } 421 422 // Transform all store nodes into one single node because all store nodes are 423 // independent of each other. 424 if (!MemOpChains.empty()) 425 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 426 &MemOpChains[0], MemOpChains.size()); 427 428 // Build a sequence of copy-to-reg nodes chained together with token chain and 429 // flag operands which copy the outgoing args into registers. The InFlag in 430 // necessary since all emited instructions must be stuck together. 431 SDValue InFlag; 432 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 433 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, 434 RegsToPass[i].second, InFlag); 435 InFlag = Chain.getValue(1); 436 } 437 438 // If the callee is a GlobalAddress node (quite common, every direct call is) 439 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. 440 // Likewise ExternalSymbol -> TargetExternalSymbol. 441 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) 442 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy()); 443 else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee)) 444 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), getPointerTy()); 445 446 // Returns a chain & a flag for retval copy to use. 447 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag); 448 SmallVector<SDValue, 8> Ops; 449 Ops.push_back(Chain); 450 Ops.push_back(Callee); 451 452 // Add argument registers to the end of the list so that they are 453 // known live into the call. 454 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 455 Ops.push_back(DAG.getRegister(RegsToPass[i].first, 456 RegsToPass[i].second.getValueType())); 457 458 if (InFlag.getNode()) 459 Ops.push_back(InFlag); 460 461 Chain = DAG.getNode(SystemZISD::CALL, dl, NodeTys, &Ops[0], Ops.size()); 462 InFlag = Chain.getValue(1); 463 464 // Create the CALLSEQ_END node. 465 Chain = DAG.getCALLSEQ_END(Chain, 466 DAG.getConstant(NumBytes, getPointerTy(), true), 467 DAG.getConstant(0, getPointerTy(), true), 468 InFlag); 469 InFlag = Chain.getValue(1); 470 471 // Handle result values, copying them out of physregs into vregs that we 472 // return. 473 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, 474 DAG, InVals); 475} 476 477/// LowerCallResult - Lower the result values of a call into the 478/// appropriate copies out of appropriate physical registers. 479/// 480SDValue 481SystemZTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, 482 unsigned CallConv, bool isVarArg, 483 const SmallVectorImpl<ISD::InputArg> 484 &Ins, 485 DebugLoc dl, SelectionDAG &DAG, 486 SmallVectorImpl<SDValue> &InVals) { 487 488 // Assign locations to each value returned by this call. 489 SmallVector<CCValAssign, 16> RVLocs; 490 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs, 491 *DAG.getContext()); 492 493 CCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ); 494 495 // Copy all of the result registers out of their specified physreg. 496 for (unsigned i = 0; i != RVLocs.size(); ++i) { 497 CCValAssign &VA = RVLocs[i]; 498 499 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), 500 VA.getLocVT(), InFlag).getValue(1); 501 SDValue RetValue = Chain.getValue(0); 502 InFlag = Chain.getValue(2); 503 504 // If this is an 8/16/32-bit value, it is really passed promoted to 64 505 // bits. Insert an assert[sz]ext to capture this, then truncate to the 506 // right size. 507 if (VA.getLocInfo() == CCValAssign::SExt) 508 RetValue = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), RetValue, 509 DAG.getValueType(VA.getValVT())); 510 else if (VA.getLocInfo() == CCValAssign::ZExt) 511 RetValue = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), RetValue, 512 DAG.getValueType(VA.getValVT())); 513 514 if (VA.getLocInfo() != CCValAssign::Full) 515 RetValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), RetValue); 516 517 InVals.push_back(RetValue); 518 } 519 520 return Chain; 521} 522 523 524SDValue 525SystemZTargetLowering::LowerReturn(SDValue Chain, 526 unsigned CallConv, bool isVarArg, 527 const SmallVectorImpl<ISD::OutputArg> &Outs, 528 DebugLoc dl, SelectionDAG &DAG) { 529 530 // CCValAssign - represent the assignment of the return value to a location 531 SmallVector<CCValAssign, 16> RVLocs; 532 533 // CCState - Info about the registers and stack slot. 534 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), 535 RVLocs, *DAG.getContext()); 536 537 // Analize return values. 538 CCInfo.AnalyzeReturn(Outs, RetCC_SystemZ); 539 540 // If this is the first return lowered for this function, add the regs to the 541 // liveout set for the function. 542 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) { 543 for (unsigned i = 0; i != RVLocs.size(); ++i) 544 if (RVLocs[i].isRegLoc()) 545 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg()); 546 } 547 548 SDValue Flag; 549 550 // Copy the result values into the output registers. 551 for (unsigned i = 0; i != RVLocs.size(); ++i) { 552 CCValAssign &VA = RVLocs[i]; 553 SDValue ResValue = Outs[i].Val; 554 assert(VA.isRegLoc() && "Can only return in registers!"); 555 556 // If this is an 8/16/32-bit value, it is really should be passed promoted 557 // to 64 bits. 558 if (VA.getLocInfo() == CCValAssign::SExt) 559 ResValue = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ResValue); 560 else if (VA.getLocInfo() == CCValAssign::ZExt) 561 ResValue = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ResValue); 562 else if (VA.getLocInfo() == CCValAssign::AExt) 563 ResValue = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ResValue); 564 565 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ResValue, Flag); 566 567 // Guarantee that all emitted copies are stuck together, 568 // avoiding something bad. 569 Flag = Chain.getValue(1); 570 } 571 572 if (Flag.getNode()) 573 return DAG.getNode(SystemZISD::RET_FLAG, dl, MVT::Other, Chain, Flag); 574 575 // Return Void 576 return DAG.getNode(SystemZISD::RET_FLAG, dl, MVT::Other, Chain); 577} 578 579SDValue SystemZTargetLowering::EmitCmp(SDValue LHS, SDValue RHS, 580 ISD::CondCode CC, SDValue &SystemZCC, 581 SelectionDAG &DAG) { 582 // FIXME: Emit a test if RHS is zero 583 584 bool isUnsigned = false; 585 SystemZCC::CondCodes TCC; 586 switch (CC) { 587 default: 588 llvm_unreachable("Invalid integer condition!"); 589 case ISD::SETEQ: 590 case ISD::SETOEQ: 591 TCC = SystemZCC::E; 592 break; 593 case ISD::SETUEQ: 594 TCC = SystemZCC::NLH; 595 break; 596 case ISD::SETNE: 597 case ISD::SETONE: 598 TCC = SystemZCC::NE; 599 break; 600 case ISD::SETUNE: 601 TCC = SystemZCC::LH; 602 break; 603 case ISD::SETO: 604 TCC = SystemZCC::O; 605 break; 606 case ISD::SETUO: 607 TCC = SystemZCC::NO; 608 break; 609 case ISD::SETULE: 610 if (LHS.getValueType().isFloatingPoint()) { 611 TCC = SystemZCC::NH; 612 break; 613 } 614 isUnsigned = true; // FALLTHROUGH 615 case ISD::SETLE: 616 case ISD::SETOLE: 617 TCC = SystemZCC::LE; 618 break; 619 case ISD::SETUGE: 620 if (LHS.getValueType().isFloatingPoint()) { 621 TCC = SystemZCC::NL; 622 break; 623 } 624 isUnsigned = true; // FALLTHROUGH 625 case ISD::SETGE: 626 case ISD::SETOGE: 627 TCC = SystemZCC::HE; 628 break; 629 case ISD::SETUGT: 630 if (LHS.getValueType().isFloatingPoint()) { 631 TCC = SystemZCC::NLE; 632 break; 633 } 634 isUnsigned = true; // FALLTHROUGH 635 case ISD::SETGT: 636 case ISD::SETOGT: 637 TCC = SystemZCC::H; 638 break; 639 case ISD::SETULT: 640 if (LHS.getValueType().isFloatingPoint()) { 641 TCC = SystemZCC::NHE; 642 break; 643 } 644 isUnsigned = true; // FALLTHROUGH 645 case ISD::SETLT: 646 case ISD::SETOLT: 647 TCC = SystemZCC::L; 648 break; 649 } 650 651 SystemZCC = DAG.getConstant(TCC, MVT::i32); 652 653 DebugLoc dl = LHS.getDebugLoc(); 654 return DAG.getNode((isUnsigned ? SystemZISD::UCMP : SystemZISD::CMP), 655 dl, MVT::Flag, LHS, RHS); 656} 657 658 659SDValue SystemZTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) { 660 SDValue Chain = Op.getOperand(0); 661 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 662 SDValue LHS = Op.getOperand(2); 663 SDValue RHS = Op.getOperand(3); 664 SDValue Dest = Op.getOperand(4); 665 DebugLoc dl = Op.getDebugLoc(); 666 667 SDValue SystemZCC; 668 SDValue Flag = EmitCmp(LHS, RHS, CC, SystemZCC, DAG); 669 return DAG.getNode(SystemZISD::BRCOND, dl, Op.getValueType(), 670 Chain, Dest, SystemZCC, Flag); 671} 672 673SDValue SystemZTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) { 674 SDValue LHS = Op.getOperand(0); 675 SDValue RHS = Op.getOperand(1); 676 SDValue TrueV = Op.getOperand(2); 677 SDValue FalseV = Op.getOperand(3); 678 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get(); 679 DebugLoc dl = Op.getDebugLoc(); 680 681 SDValue SystemZCC; 682 SDValue Flag = EmitCmp(LHS, RHS, CC, SystemZCC, DAG); 683 684 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag); 685 SmallVector<SDValue, 4> Ops; 686 Ops.push_back(TrueV); 687 Ops.push_back(FalseV); 688 Ops.push_back(SystemZCC); 689 Ops.push_back(Flag); 690 691 return DAG.getNode(SystemZISD::SELECT, dl, VTs, &Ops[0], Ops.size()); 692} 693 694SDValue SystemZTargetLowering::LowerGlobalAddress(SDValue Op, 695 SelectionDAG &DAG) { 696 DebugLoc dl = Op.getDebugLoc(); 697 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 698 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset(); 699 700 bool IsPic = getTargetMachine().getRelocationModel() == Reloc::PIC_; 701 bool ExtraLoadRequired = 702 Subtarget.GVRequiresExtraLoad(GV, getTargetMachine(), false); 703 704 SDValue Result; 705 if (!IsPic && !ExtraLoadRequired) { 706 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset); 707 Offset = 0; 708 } else { 709 unsigned char OpFlags = 0; 710 if (ExtraLoadRequired) 711 OpFlags = SystemZII::MO_GOTENT; 712 713 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags); 714 } 715 716 Result = DAG.getNode(SystemZISD::PCRelativeWrapper, dl, 717 getPointerTy(), Result); 718 719 if (ExtraLoadRequired) 720 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result, 721 PseudoSourceValue::getGOT(), 0); 722 723 // If there was a non-zero offset that we didn't fold, create an explicit 724 // addition for it. 725 if (Offset != 0) 726 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result, 727 DAG.getConstant(Offset, getPointerTy())); 728 729 return Result; 730} 731 732// FIXME: PIC here 733SDValue SystemZTargetLowering::LowerJumpTable(SDValue Op, 734 SelectionDAG &DAG) { 735 DebugLoc dl = Op.getDebugLoc(); 736 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op); 737 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy()); 738 739 return DAG.getNode(SystemZISD::PCRelativeWrapper, dl, getPointerTy(), Result); 740} 741 742 743// FIXME: PIC here 744// FIXME: This is just dirty hack. We need to lower cpool properly 745SDValue SystemZTargetLowering::LowerConstantPool(SDValue Op, 746 SelectionDAG &DAG) { 747 DebugLoc dl = Op.getDebugLoc(); 748 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); 749 750 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(), 751 CP->getAlignment(), 752 CP->getOffset()); 753 754 return DAG.getNode(SystemZISD::PCRelativeWrapper, dl, getPointerTy(), Result); 755} 756 757const char *SystemZTargetLowering::getTargetNodeName(unsigned Opcode) const { 758 switch (Opcode) { 759 case SystemZISD::RET_FLAG: return "SystemZISD::RET_FLAG"; 760 case SystemZISD::CALL: return "SystemZISD::CALL"; 761 case SystemZISD::BRCOND: return "SystemZISD::BRCOND"; 762 case SystemZISD::CMP: return "SystemZISD::CMP"; 763 case SystemZISD::UCMP: return "SystemZISD::UCMP"; 764 case SystemZISD::SELECT: return "SystemZISD::SELECT"; 765 case SystemZISD::PCRelativeWrapper: return "SystemZISD::PCRelativeWrapper"; 766 default: return NULL; 767 } 768} 769 770//===----------------------------------------------------------------------===// 771// Other Lowering Code 772//===----------------------------------------------------------------------===// 773 774MachineBasicBlock* 775SystemZTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 776 MachineBasicBlock *BB) const { 777 const SystemZInstrInfo &TII = *TM.getInstrInfo(); 778 DebugLoc dl = MI->getDebugLoc(); 779 assert((MI->getOpcode() == SystemZ::Select32 || 780 MI->getOpcode() == SystemZ::SelectF32 || 781 MI->getOpcode() == SystemZ::Select64 || 782 MI->getOpcode() == SystemZ::SelectF64) && 783 "Unexpected instr type to insert"); 784 785 // To "insert" a SELECT instruction, we actually have to insert the diamond 786 // control-flow pattern. The incoming instruction knows the destination vreg 787 // to set, the condition code register to branch on, the true/false values to 788 // select between, and a branch opcode to use. 789 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 790 MachineFunction::iterator I = BB; 791 ++I; 792 793 // thisMBB: 794 // ... 795 // TrueVal = ... 796 // cmpTY ccX, r1, r2 797 // jCC copy1MBB 798 // fallthrough --> copy0MBB 799 MachineBasicBlock *thisMBB = BB; 800 MachineFunction *F = BB->getParent(); 801 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 802 MachineBasicBlock *copy1MBB = F->CreateMachineBasicBlock(LLVM_BB); 803 SystemZCC::CondCodes CC = (SystemZCC::CondCodes)MI->getOperand(3).getImm(); 804 BuildMI(BB, dl, TII.getBrCond(CC)).addMBB(copy1MBB); 805 F->insert(I, copy0MBB); 806 F->insert(I, copy1MBB); 807 // Update machine-CFG edges by transferring all successors of the current 808 // block to the new block which will contain the Phi node for the select. 809 copy1MBB->transferSuccessors(BB); 810 // Next, add the true and fallthrough blocks as its successors. 811 BB->addSuccessor(copy0MBB); 812 BB->addSuccessor(copy1MBB); 813 814 // copy0MBB: 815 // %FalseValue = ... 816 // # fallthrough to copy1MBB 817 BB = copy0MBB; 818 819 // Update machine-CFG edges 820 BB->addSuccessor(copy1MBB); 821 822 // copy1MBB: 823 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] 824 // ... 825 BB = copy1MBB; 826 BuildMI(BB, dl, TII.get(SystemZ::PHI), 827 MI->getOperand(0).getReg()) 828 .addReg(MI->getOperand(2).getReg()).addMBB(copy0MBB) 829 .addReg(MI->getOperand(1).getReg()).addMBB(thisMBB); 830 831 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now. 832 return BB; 833} 834