SparcISelLowering.cpp revision 20b10abf4e88ca532810fbf749b029ce582d6793
1//===-- SparcISelLowering.cpp - Sparc 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 interfaces that Sparc uses to lower LLVM code into a 11// selection DAG. 12// 13//===----------------------------------------------------------------------===// 14 15#include "SparcISelLowering.h" 16#include "SparcMachineFunctionInfo.h" 17#include "SparcRegisterInfo.h" 18#include "SparcTargetMachine.h" 19#include "MCTargetDesc/SparcBaseInfo.h" 20#include "llvm/CodeGen/CallingConvLower.h" 21#include "llvm/CodeGen/MachineFrameInfo.h" 22#include "llvm/CodeGen/MachineFunction.h" 23#include "llvm/CodeGen/MachineInstrBuilder.h" 24#include "llvm/CodeGen/MachineRegisterInfo.h" 25#include "llvm/CodeGen/SelectionDAG.h" 26#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" 27#include "llvm/IR/DerivedTypes.h" 28#include "llvm/IR/Function.h" 29#include "llvm/IR/Module.h" 30#include "llvm/Support/ErrorHandling.h" 31using namespace llvm; 32 33 34//===----------------------------------------------------------------------===// 35// Calling Convention Implementation 36//===----------------------------------------------------------------------===// 37 38static bool CC_Sparc_Assign_SRet(unsigned &ValNo, MVT &ValVT, 39 MVT &LocVT, CCValAssign::LocInfo &LocInfo, 40 ISD::ArgFlagsTy &ArgFlags, CCState &State) 41{ 42 assert (ArgFlags.isSRet()); 43 44 // Assign SRet argument. 45 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT, 46 0, 47 LocVT, LocInfo)); 48 return true; 49} 50 51static bool CC_Sparc_Assign_f64(unsigned &ValNo, MVT &ValVT, 52 MVT &LocVT, CCValAssign::LocInfo &LocInfo, 53 ISD::ArgFlagsTy &ArgFlags, CCState &State) 54{ 55 static const uint16_t RegList[] = { 56 SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5 57 }; 58 // Try to get first reg. 59 if (unsigned Reg = State.AllocateReg(RegList, 6)) { 60 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 61 } else { 62 // Assign whole thing in stack. 63 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT, 64 State.AllocateStack(8,4), 65 LocVT, LocInfo)); 66 return true; 67 } 68 69 // Try to get second reg. 70 if (unsigned Reg = State.AllocateReg(RegList, 6)) 71 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 72 else 73 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT, 74 State.AllocateStack(4,4), 75 LocVT, LocInfo)); 76 return true; 77} 78 79// Allocate a full-sized argument for the 64-bit ABI. 80static bool CC_Sparc64_Full(unsigned &ValNo, MVT &ValVT, 81 MVT &LocVT, CCValAssign::LocInfo &LocInfo, 82 ISD::ArgFlagsTy &ArgFlags, CCState &State) { 83 assert((LocVT == MVT::f32 || LocVT.getSizeInBits() == 64) && 84 "Can't handle non-64 bits locations"); 85 86 // Stack space is allocated for all arguments starting from [%fp+BIAS+128]. 87 unsigned Offset = State.AllocateStack(8, 8); 88 unsigned Reg = 0; 89 90 if (LocVT == MVT::i64 && Offset < 6*8) 91 // Promote integers to %i0-%i5. 92 Reg = SP::I0 + Offset/8; 93 else if (LocVT == MVT::f64 && Offset < 16*8) 94 // Promote doubles to %d0-%d30. (Which LLVM calls D0-D15). 95 Reg = SP::D0 + Offset/8; 96 else if (LocVT == MVT::f32 && Offset < 16*8) 97 // Promote floats to %f1, %f3, ... 98 Reg = SP::F1 + Offset/4; 99 100 // Promote to register when possible, otherwise use the stack slot. 101 if (Reg) { 102 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 103 return true; 104 } 105 106 // This argument goes on the stack in an 8-byte slot. 107 // When passing floats, LocVT is smaller than 8 bytes. Adjust the offset to 108 // the right-aligned float. The first 4 bytes of the stack slot are undefined. 109 if (LocVT == MVT::f32) 110 Offset += 4; 111 112 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); 113 return true; 114} 115 116// Allocate a half-sized argument for the 64-bit ABI. 117// 118// This is used when passing { float, int } structs by value in registers. 119static bool CC_Sparc64_Half(unsigned &ValNo, MVT &ValVT, 120 MVT &LocVT, CCValAssign::LocInfo &LocInfo, 121 ISD::ArgFlagsTy &ArgFlags, CCState &State) { 122 assert(LocVT.getSizeInBits() == 32 && "Can't handle non-32 bits locations"); 123 unsigned Offset = State.AllocateStack(4, 4); 124 125 if (LocVT == MVT::f32 && Offset < 16*8) { 126 // Promote floats to %f0-%f31. 127 State.addLoc(CCValAssign::getReg(ValNo, ValVT, SP::F0 + Offset/4, 128 LocVT, LocInfo)); 129 return true; 130 } 131 132 if (LocVT == MVT::i32 && Offset < 6*8) { 133 // Promote integers to %i0-%i5, using half the register. 134 unsigned Reg = SP::I0 + Offset/8; 135 LocVT = MVT::i64; 136 LocInfo = CCValAssign::AExt; 137 138 // Set the Custom bit if this i32 goes in the high bits of a register. 139 if (Offset % 8 == 0) 140 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, 141 LocVT, LocInfo)); 142 else 143 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 144 return true; 145 } 146 147 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); 148 return true; 149} 150 151#include "SparcGenCallingConv.inc" 152 153// The calling conventions in SparcCallingConv.td are described in terms of the 154// callee's register window. This function translates registers to the 155// corresponding caller window %o register. 156static unsigned toCallerWindow(unsigned Reg) { 157 assert(SP::I0 + 7 == SP::I7 && SP::O0 + 7 == SP::O7 && "Unexpected enum"); 158 if (Reg >= SP::I0 && Reg <= SP::I7) 159 return Reg - SP::I0 + SP::O0; 160 return Reg; 161} 162 163SDValue 164SparcTargetLowering::LowerReturn(SDValue Chain, 165 CallingConv::ID CallConv, bool IsVarArg, 166 const SmallVectorImpl<ISD::OutputArg> &Outs, 167 const SmallVectorImpl<SDValue> &OutVals, 168 SDLoc DL, SelectionDAG &DAG) const { 169 if (Subtarget->is64Bit()) 170 return LowerReturn_64(Chain, CallConv, IsVarArg, Outs, OutVals, DL, DAG); 171 return LowerReturn_32(Chain, CallConv, IsVarArg, Outs, OutVals, DL, DAG); 172} 173 174SDValue 175SparcTargetLowering::LowerReturn_32(SDValue Chain, 176 CallingConv::ID CallConv, bool IsVarArg, 177 const SmallVectorImpl<ISD::OutputArg> &Outs, 178 const SmallVectorImpl<SDValue> &OutVals, 179 SDLoc DL, SelectionDAG &DAG) const { 180 MachineFunction &MF = DAG.getMachineFunction(); 181 182 // CCValAssign - represent the assignment of the return value to locations. 183 SmallVector<CCValAssign, 16> RVLocs; 184 185 // CCState - Info about the registers and stack slot. 186 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), 187 DAG.getTarget(), RVLocs, *DAG.getContext()); 188 189 // Analyze return values. 190 CCInfo.AnalyzeReturn(Outs, RetCC_Sparc32); 191 192 SDValue Flag; 193 SmallVector<SDValue, 4> RetOps(1, Chain); 194 // Make room for the return address offset. 195 RetOps.push_back(SDValue()); 196 197 // Copy the result values into the output registers. 198 for (unsigned i = 0; i != RVLocs.size(); ++i) { 199 CCValAssign &VA = RVLocs[i]; 200 assert(VA.isRegLoc() && "Can only return in registers!"); 201 202 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), 203 OutVals[i], Flag); 204 205 // Guarantee that all emitted copies are stuck together with flags. 206 Flag = Chain.getValue(1); 207 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 208 } 209 210 unsigned RetAddrOffset = 8; // Call Inst + Delay Slot 211 // If the function returns a struct, copy the SRetReturnReg to I0 212 if (MF.getFunction()->hasStructRetAttr()) { 213 SparcMachineFunctionInfo *SFI = MF.getInfo<SparcMachineFunctionInfo>(); 214 unsigned Reg = SFI->getSRetReturnReg(); 215 if (!Reg) 216 llvm_unreachable("sret virtual register not created in the entry block"); 217 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy()); 218 Chain = DAG.getCopyToReg(Chain, DL, SP::I0, Val, Flag); 219 Flag = Chain.getValue(1); 220 RetOps.push_back(DAG.getRegister(SP::I0, getPointerTy())); 221 RetAddrOffset = 12; // CallInst + Delay Slot + Unimp 222 } 223 224 RetOps[0] = Chain; // Update chain. 225 RetOps[1] = DAG.getConstant(RetAddrOffset, MVT::i32); 226 227 // Add the flag if we have it. 228 if (Flag.getNode()) 229 RetOps.push_back(Flag); 230 231 return DAG.getNode(SPISD::RET_FLAG, DL, MVT::Other, 232 &RetOps[0], RetOps.size()); 233} 234 235// Lower return values for the 64-bit ABI. 236// Return values are passed the exactly the same way as function arguments. 237SDValue 238SparcTargetLowering::LowerReturn_64(SDValue Chain, 239 CallingConv::ID CallConv, bool IsVarArg, 240 const SmallVectorImpl<ISD::OutputArg> &Outs, 241 const SmallVectorImpl<SDValue> &OutVals, 242 SDLoc DL, SelectionDAG &DAG) const { 243 // CCValAssign - represent the assignment of the return value to locations. 244 SmallVector<CCValAssign, 16> RVLocs; 245 246 // CCState - Info about the registers and stack slot. 247 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), 248 DAG.getTarget(), RVLocs, *DAG.getContext()); 249 250 // Analyze return values. 251 CCInfo.AnalyzeReturn(Outs, CC_Sparc64); 252 253 SDValue Flag; 254 SmallVector<SDValue, 4> RetOps(1, Chain); 255 256 // The second operand on the return instruction is the return address offset. 257 // The return address is always %i7+8 with the 64-bit ABI. 258 RetOps.push_back(DAG.getConstant(8, MVT::i32)); 259 260 // Copy the result values into the output registers. 261 for (unsigned i = 0; i != RVLocs.size(); ++i) { 262 CCValAssign &VA = RVLocs[i]; 263 assert(VA.isRegLoc() && "Can only return in registers!"); 264 SDValue OutVal = OutVals[i]; 265 266 // Integer return values must be sign or zero extended by the callee. 267 switch (VA.getLocInfo()) { 268 case CCValAssign::SExt: 269 OutVal = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), OutVal); 270 break; 271 case CCValAssign::ZExt: 272 OutVal = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), OutVal); 273 break; 274 case CCValAssign::AExt: 275 OutVal = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), OutVal); 276 default: 277 break; 278 } 279 280 // The custom bit on an i32 return value indicates that it should be passed 281 // in the high bits of the register. 282 if (VA.getValVT() == MVT::i32 && VA.needsCustom()) { 283 OutVal = DAG.getNode(ISD::SHL, DL, MVT::i64, OutVal, 284 DAG.getConstant(32, MVT::i32)); 285 286 // The next value may go in the low bits of the same register. 287 // Handle both at once. 288 if (i+1 < RVLocs.size() && RVLocs[i+1].getLocReg() == VA.getLocReg()) { 289 SDValue NV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, OutVals[i+1]); 290 OutVal = DAG.getNode(ISD::OR, DL, MVT::i64, OutVal, NV); 291 // Skip the next value, it's already done. 292 ++i; 293 } 294 } 295 296 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), OutVal, Flag); 297 298 // Guarantee that all emitted copies are stuck together with flags. 299 Flag = Chain.getValue(1); 300 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 301 } 302 303 RetOps[0] = Chain; // Update chain. 304 305 // Add the flag if we have it. 306 if (Flag.getNode()) 307 RetOps.push_back(Flag); 308 309 return DAG.getNode(SPISD::RET_FLAG, DL, MVT::Other, 310 &RetOps[0], RetOps.size()); 311} 312 313SDValue SparcTargetLowering:: 314LowerFormalArguments(SDValue Chain, 315 CallingConv::ID CallConv, 316 bool IsVarArg, 317 const SmallVectorImpl<ISD::InputArg> &Ins, 318 SDLoc DL, 319 SelectionDAG &DAG, 320 SmallVectorImpl<SDValue> &InVals) const { 321 if (Subtarget->is64Bit()) 322 return LowerFormalArguments_64(Chain, CallConv, IsVarArg, Ins, 323 DL, DAG, InVals); 324 return LowerFormalArguments_32(Chain, CallConv, IsVarArg, Ins, 325 DL, DAG, InVals); 326} 327 328/// LowerFormalArguments32 - V8 uses a very simple ABI, where all values are 329/// passed in either one or two GPRs, including FP values. TODO: we should 330/// pass FP values in FP registers for fastcc functions. 331SDValue SparcTargetLowering:: 332LowerFormalArguments_32(SDValue Chain, 333 CallingConv::ID CallConv, 334 bool isVarArg, 335 const SmallVectorImpl<ISD::InputArg> &Ins, 336 SDLoc dl, 337 SelectionDAG &DAG, 338 SmallVectorImpl<SDValue> &InVals) const { 339 MachineFunction &MF = DAG.getMachineFunction(); 340 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 341 SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>(); 342 343 // Assign locations to all of the incoming arguments. 344 SmallVector<CCValAssign, 16> ArgLocs; 345 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 346 getTargetMachine(), ArgLocs, *DAG.getContext()); 347 CCInfo.AnalyzeFormalArguments(Ins, CC_Sparc32); 348 349 const unsigned StackOffset = 92; 350 351 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 352 CCValAssign &VA = ArgLocs[i]; 353 354 if (i == 0 && Ins[i].Flags.isSRet()) { 355 // Get SRet from [%fp+64]. 356 int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, 64, true); 357 SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32); 358 SDValue Arg = DAG.getLoad(MVT::i32, dl, Chain, FIPtr, 359 MachinePointerInfo(), 360 false, false, false, 0); 361 InVals.push_back(Arg); 362 continue; 363 } 364 365 if (VA.isRegLoc()) { 366 if (VA.needsCustom()) { 367 assert(VA.getLocVT() == MVT::f64); 368 unsigned VRegHi = RegInfo.createVirtualRegister(&SP::IntRegsRegClass); 369 MF.getRegInfo().addLiveIn(VA.getLocReg(), VRegHi); 370 SDValue HiVal = DAG.getCopyFromReg(Chain, dl, VRegHi, MVT::i32); 371 372 assert(i+1 < e); 373 CCValAssign &NextVA = ArgLocs[++i]; 374 375 SDValue LoVal; 376 if (NextVA.isMemLoc()) { 377 int FrameIdx = MF.getFrameInfo()-> 378 CreateFixedObject(4, StackOffset+NextVA.getLocMemOffset(),true); 379 SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32); 380 LoVal = DAG.getLoad(MVT::i32, dl, Chain, FIPtr, 381 MachinePointerInfo(), 382 false, false, false, 0); 383 } else { 384 unsigned loReg = MF.addLiveIn(NextVA.getLocReg(), 385 &SP::IntRegsRegClass); 386 LoVal = DAG.getCopyFromReg(Chain, dl, loReg, MVT::i32); 387 } 388 SDValue WholeValue = 389 DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, LoVal, HiVal); 390 WholeValue = DAG.getNode(ISD::BITCAST, dl, MVT::f64, WholeValue); 391 InVals.push_back(WholeValue); 392 continue; 393 } 394 unsigned VReg = RegInfo.createVirtualRegister(&SP::IntRegsRegClass); 395 MF.getRegInfo().addLiveIn(VA.getLocReg(), VReg); 396 SDValue Arg = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32); 397 if (VA.getLocVT() == MVT::f32) 398 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Arg); 399 else if (VA.getLocVT() != MVT::i32) { 400 Arg = DAG.getNode(ISD::AssertSext, dl, MVT::i32, Arg, 401 DAG.getValueType(VA.getLocVT())); 402 Arg = DAG.getNode(ISD::TRUNCATE, dl, VA.getLocVT(), Arg); 403 } 404 InVals.push_back(Arg); 405 continue; 406 } 407 408 assert(VA.isMemLoc()); 409 410 unsigned Offset = VA.getLocMemOffset()+StackOffset; 411 412 if (VA.needsCustom()) { 413 assert(VA.getValVT() == MVT::f64); 414 // If it is double-word aligned, just load. 415 if (Offset % 8 == 0) { 416 int FI = MF.getFrameInfo()->CreateFixedObject(8, 417 Offset, 418 true); 419 SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy()); 420 SDValue Load = DAG.getLoad(VA.getValVT(), dl, Chain, FIPtr, 421 MachinePointerInfo(), 422 false,false, false, 0); 423 InVals.push_back(Load); 424 continue; 425 } 426 427 int FI = MF.getFrameInfo()->CreateFixedObject(4, 428 Offset, 429 true); 430 SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy()); 431 SDValue HiVal = DAG.getLoad(MVT::i32, dl, Chain, FIPtr, 432 MachinePointerInfo(), 433 false, false, false, 0); 434 int FI2 = MF.getFrameInfo()->CreateFixedObject(4, 435 Offset+4, 436 true); 437 SDValue FIPtr2 = DAG.getFrameIndex(FI2, getPointerTy()); 438 439 SDValue LoVal = DAG.getLoad(MVT::i32, dl, Chain, FIPtr2, 440 MachinePointerInfo(), 441 false, false, false, 0); 442 443 SDValue WholeValue = 444 DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, LoVal, HiVal); 445 WholeValue = DAG.getNode(ISD::BITCAST, dl, MVT::f64, WholeValue); 446 InVals.push_back(WholeValue); 447 continue; 448 } 449 450 int FI = MF.getFrameInfo()->CreateFixedObject(4, 451 Offset, 452 true); 453 SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy()); 454 SDValue Load ; 455 if (VA.getValVT() == MVT::i32 || VA.getValVT() == MVT::f32) { 456 Load = DAG.getLoad(VA.getValVT(), dl, Chain, FIPtr, 457 MachinePointerInfo(), 458 false, false, false, 0); 459 } else { 460 ISD::LoadExtType LoadOp = ISD::SEXTLOAD; 461 // Sparc is big endian, so add an offset based on the ObjectVT. 462 unsigned Offset = 4-std::max(1U, VA.getValVT().getSizeInBits()/8); 463 FIPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, FIPtr, 464 DAG.getConstant(Offset, MVT::i32)); 465 Load = DAG.getExtLoad(LoadOp, dl, MVT::i32, Chain, FIPtr, 466 MachinePointerInfo(), 467 VA.getValVT(), false, false,0); 468 Load = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Load); 469 } 470 InVals.push_back(Load); 471 } 472 473 if (MF.getFunction()->hasStructRetAttr()) { 474 // Copy the SRet Argument to SRetReturnReg. 475 SparcMachineFunctionInfo *SFI = MF.getInfo<SparcMachineFunctionInfo>(); 476 unsigned Reg = SFI->getSRetReturnReg(); 477 if (!Reg) { 478 Reg = MF.getRegInfo().createVirtualRegister(&SP::IntRegsRegClass); 479 SFI->setSRetReturnReg(Reg); 480 } 481 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]); 482 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain); 483 } 484 485 // Store remaining ArgRegs to the stack if this is a varargs function. 486 if (isVarArg) { 487 static const uint16_t ArgRegs[] = { 488 SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5 489 }; 490 unsigned NumAllocated = CCInfo.getFirstUnallocated(ArgRegs, 6); 491 const uint16_t *CurArgReg = ArgRegs+NumAllocated, *ArgRegEnd = ArgRegs+6; 492 unsigned ArgOffset = CCInfo.getNextStackOffset(); 493 if (NumAllocated == 6) 494 ArgOffset += StackOffset; 495 else { 496 assert(!ArgOffset); 497 ArgOffset = 68+4*NumAllocated; 498 } 499 500 // Remember the vararg offset for the va_start implementation. 501 FuncInfo->setVarArgsFrameOffset(ArgOffset); 502 503 std::vector<SDValue> OutChains; 504 505 for (; CurArgReg != ArgRegEnd; ++CurArgReg) { 506 unsigned VReg = RegInfo.createVirtualRegister(&SP::IntRegsRegClass); 507 MF.getRegInfo().addLiveIn(*CurArgReg, VReg); 508 SDValue Arg = DAG.getCopyFromReg(DAG.getRoot(), dl, VReg, MVT::i32); 509 510 int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, ArgOffset, 511 true); 512 SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32); 513 514 OutChains.push_back(DAG.getStore(DAG.getRoot(), dl, Arg, FIPtr, 515 MachinePointerInfo(), 516 false, false, 0)); 517 ArgOffset += 4; 518 } 519 520 if (!OutChains.empty()) { 521 OutChains.push_back(Chain); 522 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 523 &OutChains[0], OutChains.size()); 524 } 525 } 526 527 return Chain; 528} 529 530// Lower formal arguments for the 64 bit ABI. 531SDValue SparcTargetLowering:: 532LowerFormalArguments_64(SDValue Chain, 533 CallingConv::ID CallConv, 534 bool IsVarArg, 535 const SmallVectorImpl<ISD::InputArg> &Ins, 536 SDLoc DL, 537 SelectionDAG &DAG, 538 SmallVectorImpl<SDValue> &InVals) const { 539 MachineFunction &MF = DAG.getMachineFunction(); 540 541 // Analyze arguments according to CC_Sparc64. 542 SmallVector<CCValAssign, 16> ArgLocs; 543 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), 544 getTargetMachine(), ArgLocs, *DAG.getContext()); 545 CCInfo.AnalyzeFormalArguments(Ins, CC_Sparc64); 546 547 // The argument array begins at %fp+BIAS+128, after the register save area. 548 const unsigned ArgArea = 128; 549 550 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 551 CCValAssign &VA = ArgLocs[i]; 552 if (VA.isRegLoc()) { 553 // This argument is passed in a register. 554 // All integer register arguments are promoted by the caller to i64. 555 556 // Create a virtual register for the promoted live-in value. 557 unsigned VReg = MF.addLiveIn(VA.getLocReg(), 558 getRegClassFor(VA.getLocVT())); 559 SDValue Arg = DAG.getCopyFromReg(Chain, DL, VReg, VA.getLocVT()); 560 561 // Get the high bits for i32 struct elements. 562 if (VA.getValVT() == MVT::i32 && VA.needsCustom()) 563 Arg = DAG.getNode(ISD::SRL, DL, VA.getLocVT(), Arg, 564 DAG.getConstant(32, MVT::i32)); 565 566 // The caller promoted the argument, so insert an Assert?ext SDNode so we 567 // won't promote the value again in this function. 568 switch (VA.getLocInfo()) { 569 case CCValAssign::SExt: 570 Arg = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Arg, 571 DAG.getValueType(VA.getValVT())); 572 break; 573 case CCValAssign::ZExt: 574 Arg = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Arg, 575 DAG.getValueType(VA.getValVT())); 576 break; 577 default: 578 break; 579 } 580 581 // Truncate the register down to the argument type. 582 if (VA.isExtInLoc()) 583 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg); 584 585 InVals.push_back(Arg); 586 continue; 587 } 588 589 // The registers are exhausted. This argument was passed on the stack. 590 assert(VA.isMemLoc()); 591 // The CC_Sparc64_Full/Half functions compute stack offsets relative to the 592 // beginning of the arguments area at %fp+BIAS+128. 593 unsigned Offset = VA.getLocMemOffset() + ArgArea; 594 unsigned ValSize = VA.getValVT().getSizeInBits() / 8; 595 // Adjust offset for extended arguments, SPARC is big-endian. 596 // The caller will have written the full slot with extended bytes, but we 597 // prefer our own extending loads. 598 if (VA.isExtInLoc()) 599 Offset += 8 - ValSize; 600 int FI = MF.getFrameInfo()->CreateFixedObject(ValSize, Offset, true); 601 InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, 602 DAG.getFrameIndex(FI, getPointerTy()), 603 MachinePointerInfo::getFixedStack(FI), 604 false, false, false, 0)); 605 } 606 607 if (!IsVarArg) 608 return Chain; 609 610 // This function takes variable arguments, some of which may have been passed 611 // in registers %i0-%i5. Variable floating point arguments are never passed 612 // in floating point registers. They go on %i0-%i5 or on the stack like 613 // integer arguments. 614 // 615 // The va_start intrinsic needs to know the offset to the first variable 616 // argument. 617 unsigned ArgOffset = CCInfo.getNextStackOffset(); 618 SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>(); 619 // Skip the 128 bytes of register save area. 620 FuncInfo->setVarArgsFrameOffset(ArgOffset + ArgArea + 621 Subtarget->getStackPointerBias()); 622 623 // Save the variable arguments that were passed in registers. 624 // The caller is required to reserve stack space for 6 arguments regardless 625 // of how many arguments were actually passed. 626 SmallVector<SDValue, 8> OutChains; 627 for (; ArgOffset < 6*8; ArgOffset += 8) { 628 unsigned VReg = MF.addLiveIn(SP::I0 + ArgOffset/8, &SP::I64RegsRegClass); 629 SDValue VArg = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64); 630 int FI = MF.getFrameInfo()->CreateFixedObject(8, ArgOffset + ArgArea, true); 631 OutChains.push_back(DAG.getStore(Chain, DL, VArg, 632 DAG.getFrameIndex(FI, getPointerTy()), 633 MachinePointerInfo::getFixedStack(FI), 634 false, false, 0)); 635 } 636 637 if (!OutChains.empty()) 638 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, 639 &OutChains[0], OutChains.size()); 640 641 return Chain; 642} 643 644SDValue 645SparcTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, 646 SmallVectorImpl<SDValue> &InVals) const { 647 if (Subtarget->is64Bit()) 648 return LowerCall_64(CLI, InVals); 649 return LowerCall_32(CLI, InVals); 650} 651 652static bool hasReturnsTwiceAttr(SelectionDAG &DAG, SDValue Callee, 653 ImmutableCallSite *CS) { 654 if (CS) 655 return CS->hasFnAttr(Attribute::ReturnsTwice); 656 657 const Function *CalleeFn = 0; 658 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 659 CalleeFn = dyn_cast<Function>(G->getGlobal()); 660 } else if (ExternalSymbolSDNode *E = 661 dyn_cast<ExternalSymbolSDNode>(Callee)) { 662 const Function *Fn = DAG.getMachineFunction().getFunction(); 663 const Module *M = Fn->getParent(); 664 const char *CalleeName = E->getSymbol(); 665 CalleeFn = M->getFunction(CalleeName); 666 } 667 668 if (!CalleeFn) 669 return false; 670 return CalleeFn->hasFnAttribute(Attribute::ReturnsTwice); 671} 672 673// Lower a call for the 32-bit ABI. 674SDValue 675SparcTargetLowering::LowerCall_32(TargetLowering::CallLoweringInfo &CLI, 676 SmallVectorImpl<SDValue> &InVals) const { 677 SelectionDAG &DAG = CLI.DAG; 678 SDLoc &dl = CLI.DL; 679 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; 680 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; 681 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; 682 SDValue Chain = CLI.Chain; 683 SDValue Callee = CLI.Callee; 684 bool &isTailCall = CLI.IsTailCall; 685 CallingConv::ID CallConv = CLI.CallConv; 686 bool isVarArg = CLI.IsVarArg; 687 688 // Sparc target does not yet support tail call optimization. 689 isTailCall = false; 690 691 // Analyze operands of the call, assigning locations to each operand. 692 SmallVector<CCValAssign, 16> ArgLocs; 693 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 694 DAG.getTarget(), ArgLocs, *DAG.getContext()); 695 CCInfo.AnalyzeCallOperands(Outs, CC_Sparc32); 696 697 // Get the size of the outgoing arguments stack space requirement. 698 unsigned ArgsSize = CCInfo.getNextStackOffset(); 699 700 // Keep stack frames 8-byte aligned. 701 ArgsSize = (ArgsSize+7) & ~7; 702 703 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 704 705 // Create local copies for byval args. 706 SmallVector<SDValue, 8> ByValArgs; 707 for (unsigned i = 0, e = Outs.size(); i != e; ++i) { 708 ISD::ArgFlagsTy Flags = Outs[i].Flags; 709 if (!Flags.isByVal()) 710 continue; 711 712 SDValue Arg = OutVals[i]; 713 unsigned Size = Flags.getByValSize(); 714 unsigned Align = Flags.getByValAlign(); 715 716 int FI = MFI->CreateStackObject(Size, Align, false); 717 SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy()); 718 SDValue SizeNode = DAG.getConstant(Size, MVT::i32); 719 720 Chain = DAG.getMemcpy(Chain, dl, FIPtr, Arg, SizeNode, Align, 721 false, // isVolatile, 722 (Size <= 32), // AlwaysInline if size <= 32 723 MachinePointerInfo(), MachinePointerInfo()); 724 ByValArgs.push_back(FIPtr); 725 } 726 727 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(ArgsSize, true), 728 dl); 729 730 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; 731 SmallVector<SDValue, 8> MemOpChains; 732 733 const unsigned StackOffset = 92; 734 bool hasStructRetAttr = false; 735 // Walk the register/memloc assignments, inserting copies/loads. 736 for (unsigned i = 0, realArgIdx = 0, byvalArgIdx = 0, e = ArgLocs.size(); 737 i != e; 738 ++i, ++realArgIdx) { 739 CCValAssign &VA = ArgLocs[i]; 740 SDValue Arg = OutVals[realArgIdx]; 741 742 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags; 743 744 // Use local copy if it is a byval arg. 745 if (Flags.isByVal()) 746 Arg = ByValArgs[byvalArgIdx++]; 747 748 // Promote the value if needed. 749 switch (VA.getLocInfo()) { 750 default: llvm_unreachable("Unknown loc info!"); 751 case CCValAssign::Full: break; 752 case CCValAssign::SExt: 753 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); 754 break; 755 case CCValAssign::ZExt: 756 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); 757 break; 758 case CCValAssign::AExt: 759 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg); 760 break; 761 case CCValAssign::BCvt: 762 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg); 763 break; 764 } 765 766 if (Flags.isSRet()) { 767 assert(VA.needsCustom()); 768 // store SRet argument in %sp+64 769 SDValue StackPtr = DAG.getRegister(SP::O6, MVT::i32); 770 SDValue PtrOff = DAG.getIntPtrConstant(64); 771 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff); 772 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff, 773 MachinePointerInfo(), 774 false, false, 0)); 775 hasStructRetAttr = true; 776 continue; 777 } 778 779 if (VA.needsCustom()) { 780 assert(VA.getLocVT() == MVT::f64); 781 782 if (VA.isMemLoc()) { 783 unsigned Offset = VA.getLocMemOffset() + StackOffset; 784 // if it is double-word aligned, just store. 785 if (Offset % 8 == 0) { 786 SDValue StackPtr = DAG.getRegister(SP::O6, MVT::i32); 787 SDValue PtrOff = DAG.getIntPtrConstant(Offset); 788 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff); 789 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff, 790 MachinePointerInfo(), 791 false, false, 0)); 792 continue; 793 } 794 } 795 796 SDValue StackPtr = DAG.CreateStackTemporary(MVT::f64, MVT::i32); 797 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, 798 Arg, StackPtr, MachinePointerInfo(), 799 false, false, 0); 800 // Sparc is big-endian, so the high part comes first. 801 SDValue Hi = DAG.getLoad(MVT::i32, dl, Store, StackPtr, 802 MachinePointerInfo(), false, false, false, 0); 803 // Increment the pointer to the other half. 804 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr, 805 DAG.getIntPtrConstant(4)); 806 // Load the low part. 807 SDValue Lo = DAG.getLoad(MVT::i32, dl, Store, StackPtr, 808 MachinePointerInfo(), false, false, false, 0); 809 810 if (VA.isRegLoc()) { 811 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Hi)); 812 assert(i+1 != e); 813 CCValAssign &NextVA = ArgLocs[++i]; 814 if (NextVA.isRegLoc()) { 815 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), Lo)); 816 } else { 817 // Store the low part in stack. 818 unsigned Offset = NextVA.getLocMemOffset() + StackOffset; 819 SDValue StackPtr = DAG.getRegister(SP::O6, MVT::i32); 820 SDValue PtrOff = DAG.getIntPtrConstant(Offset); 821 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff); 822 MemOpChains.push_back(DAG.getStore(Chain, dl, Lo, PtrOff, 823 MachinePointerInfo(), 824 false, false, 0)); 825 } 826 } else { 827 unsigned Offset = VA.getLocMemOffset() + StackOffset; 828 // Store the high part. 829 SDValue StackPtr = DAG.getRegister(SP::O6, MVT::i32); 830 SDValue PtrOff = DAG.getIntPtrConstant(Offset); 831 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff); 832 MemOpChains.push_back(DAG.getStore(Chain, dl, Hi, PtrOff, 833 MachinePointerInfo(), 834 false, false, 0)); 835 // Store the low part. 836 PtrOff = DAG.getIntPtrConstant(Offset+4); 837 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff); 838 MemOpChains.push_back(DAG.getStore(Chain, dl, Lo, PtrOff, 839 MachinePointerInfo(), 840 false, false, 0)); 841 } 842 continue; 843 } 844 845 // Arguments that can be passed on register must be kept at 846 // RegsToPass vector 847 if (VA.isRegLoc()) { 848 if (VA.getLocVT() != MVT::f32) { 849 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 850 continue; 851 } 852 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg); 853 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 854 continue; 855 } 856 857 assert(VA.isMemLoc()); 858 859 // Create a store off the stack pointer for this argument. 860 SDValue StackPtr = DAG.getRegister(SP::O6, MVT::i32); 861 SDValue PtrOff = DAG.getIntPtrConstant(VA.getLocMemOffset()+StackOffset); 862 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff); 863 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff, 864 MachinePointerInfo(), 865 false, false, 0)); 866 } 867 868 869 // Emit all stores, make sure the occur before any copies into physregs. 870 if (!MemOpChains.empty()) 871 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 872 &MemOpChains[0], MemOpChains.size()); 873 874 // Build a sequence of copy-to-reg nodes chained together with token 875 // chain and flag operands which copy the outgoing args into registers. 876 // The InFlag in necessary since all emitted instructions must be 877 // stuck together. 878 SDValue InFlag; 879 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 880 unsigned Reg = toCallerWindow(RegsToPass[i].first); 881 Chain = DAG.getCopyToReg(Chain, dl, Reg, RegsToPass[i].second, InFlag); 882 InFlag = Chain.getValue(1); 883 } 884 885 unsigned SRetArgSize = (hasStructRetAttr)? getSRetArgSize(DAG, Callee):0; 886 bool hasReturnsTwice = hasReturnsTwiceAttr(DAG, Callee, CLI.CS); 887 888 // If the callee is a GlobalAddress node (quite common, every direct call is) 889 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. 890 // Likewise ExternalSymbol -> TargetExternalSymbol. 891 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) 892 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, MVT::i32); 893 else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee)) 894 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), MVT::i32); 895 896 // Returns a chain & a flag for retval copy to use 897 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 898 SmallVector<SDValue, 8> Ops; 899 Ops.push_back(Chain); 900 Ops.push_back(Callee); 901 if (hasStructRetAttr) 902 Ops.push_back(DAG.getTargetConstant(SRetArgSize, MVT::i32)); 903 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 904 Ops.push_back(DAG.getRegister(toCallerWindow(RegsToPass[i].first), 905 RegsToPass[i].second.getValueType())); 906 907 // Add a register mask operand representing the call-preserved registers. 908 const SparcRegisterInfo *TRI = 909 ((const SparcTargetMachine&)getTargetMachine()).getRegisterInfo(); 910 const uint32_t *Mask = ((hasReturnsTwice) 911 ? TRI->getRTCallPreservedMask(CallConv) 912 : TRI->getCallPreservedMask(CallConv)); 913 assert(Mask && "Missing call preserved mask for calling convention"); 914 Ops.push_back(DAG.getRegisterMask(Mask)); 915 916 if (InFlag.getNode()) 917 Ops.push_back(InFlag); 918 919 Chain = DAG.getNode(SPISD::CALL, dl, NodeTys, &Ops[0], Ops.size()); 920 InFlag = Chain.getValue(1); 921 922 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(ArgsSize, true), 923 DAG.getIntPtrConstant(0, true), InFlag, dl); 924 InFlag = Chain.getValue(1); 925 926 // Assign locations to each value returned by this call. 927 SmallVector<CCValAssign, 16> RVLocs; 928 CCState RVInfo(CallConv, isVarArg, DAG.getMachineFunction(), 929 DAG.getTarget(), RVLocs, *DAG.getContext()); 930 931 RVInfo.AnalyzeCallResult(Ins, RetCC_Sparc32); 932 933 // Copy all of the result registers out of their specified physreg. 934 for (unsigned i = 0; i != RVLocs.size(); ++i) { 935 Chain = DAG.getCopyFromReg(Chain, dl, toCallerWindow(RVLocs[i].getLocReg()), 936 RVLocs[i].getValVT(), InFlag).getValue(1); 937 InFlag = Chain.getValue(2); 938 InVals.push_back(Chain.getValue(0)); 939 } 940 941 return Chain; 942} 943 944// This functions returns true if CalleeName is a ABI function that returns 945// a long double (fp128). 946static bool isFP128ABICall(const char *CalleeName) 947{ 948 static const char *const ABICalls[] = 949 { "_Q_add", "_Q_sub", "_Q_mul", "_Q_div", 950 "_Q_sqrt", "_Q_neg", 951 "_Q_itoq", "_Q_stoq", "_Q_dtoq", "_Q_utoq", 952 0 953 }; 954 for (const char * const *I = ABICalls; I != 0; ++I) 955 if (strcmp(CalleeName, *I) == 0) 956 return true; 957 return false; 958} 959 960unsigned 961SparcTargetLowering::getSRetArgSize(SelectionDAG &DAG, SDValue Callee) const 962{ 963 const Function *CalleeFn = 0; 964 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 965 CalleeFn = dyn_cast<Function>(G->getGlobal()); 966 } else if (ExternalSymbolSDNode *E = 967 dyn_cast<ExternalSymbolSDNode>(Callee)) { 968 const Function *Fn = DAG.getMachineFunction().getFunction(); 969 const Module *M = Fn->getParent(); 970 const char *CalleeName = E->getSymbol(); 971 CalleeFn = M->getFunction(CalleeName); 972 if (!CalleeFn && isFP128ABICall(CalleeName)) 973 return 16; // Return sizeof(fp128) 974 } 975 976 if (!CalleeFn) 977 return 0; 978 979 assert(CalleeFn->hasStructRetAttr() && 980 "Callee does not have the StructRet attribute."); 981 982 PointerType *Ty = cast<PointerType>(CalleeFn->arg_begin()->getType()); 983 Type *ElementTy = Ty->getElementType(); 984 return getDataLayout()->getTypeAllocSize(ElementTy); 985} 986 987 988// Fixup floating point arguments in the ... part of a varargs call. 989// 990// The SPARC v9 ABI requires that floating point arguments are treated the same 991// as integers when calling a varargs function. This does not apply to the 992// fixed arguments that are part of the function's prototype. 993// 994// This function post-processes a CCValAssign array created by 995// AnalyzeCallOperands(). 996static void fixupVariableFloatArgs(SmallVectorImpl<CCValAssign> &ArgLocs, 997 ArrayRef<ISD::OutputArg> Outs) { 998 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 999 const CCValAssign &VA = ArgLocs[i]; 1000 // FIXME: What about f32 arguments? C promotes them to f64 when calling 1001 // varargs functions. 1002 if (!VA.isRegLoc() || VA.getLocVT() != MVT::f64) 1003 continue; 1004 // The fixed arguments to a varargs function still go in FP registers. 1005 if (Outs[VA.getValNo()].IsFixed) 1006 continue; 1007 1008 // This floating point argument should be reassigned. 1009 CCValAssign NewVA; 1010 1011 // Determine the offset into the argument array. 1012 unsigned Offset = 8 * (VA.getLocReg() - SP::D0); 1013 assert(Offset < 16*8 && "Offset out of range, bad register enum?"); 1014 1015 if (Offset < 6*8) { 1016 // This argument should go in %i0-%i5. 1017 unsigned IReg = SP::I0 + Offset/8; 1018 // Full register, just bitconvert into i64. 1019 NewVA = CCValAssign::getReg(VA.getValNo(), VA.getValVT(), 1020 IReg, MVT::i64, CCValAssign::BCvt); 1021 } else { 1022 // This needs to go to memory, we're out of integer registers. 1023 NewVA = CCValAssign::getMem(VA.getValNo(), VA.getValVT(), 1024 Offset, VA.getLocVT(), VA.getLocInfo()); 1025 } 1026 ArgLocs[i] = NewVA; 1027 } 1028} 1029 1030// Lower a call for the 64-bit ABI. 1031SDValue 1032SparcTargetLowering::LowerCall_64(TargetLowering::CallLoweringInfo &CLI, 1033 SmallVectorImpl<SDValue> &InVals) const { 1034 SelectionDAG &DAG = CLI.DAG; 1035 SDLoc DL = CLI.DL; 1036 SDValue Chain = CLI.Chain; 1037 1038 // Analyze operands of the call, assigning locations to each operand. 1039 SmallVector<CCValAssign, 16> ArgLocs; 1040 CCState CCInfo(CLI.CallConv, CLI.IsVarArg, DAG.getMachineFunction(), 1041 DAG.getTarget(), ArgLocs, *DAG.getContext()); 1042 CCInfo.AnalyzeCallOperands(CLI.Outs, CC_Sparc64); 1043 1044 // Get the size of the outgoing arguments stack space requirement. 1045 // The stack offset computed by CC_Sparc64 includes all arguments. 1046 // Called functions expect 6 argument words to exist in the stack frame, used 1047 // or not. 1048 unsigned ArgsSize = std::max(6*8u, CCInfo.getNextStackOffset()); 1049 1050 // Keep stack frames 16-byte aligned. 1051 ArgsSize = RoundUpToAlignment(ArgsSize, 16); 1052 1053 // Varargs calls require special treatment. 1054 if (CLI.IsVarArg) 1055 fixupVariableFloatArgs(ArgLocs, CLI.Outs); 1056 1057 // Adjust the stack pointer to make room for the arguments. 1058 // FIXME: Use hasReservedCallFrame to avoid %sp adjustments around all calls 1059 // with more than 6 arguments. 1060 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(ArgsSize, true), 1061 DL); 1062 1063 // Collect the set of registers to pass to the function and their values. 1064 // This will be emitted as a sequence of CopyToReg nodes glued to the call 1065 // instruction. 1066 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; 1067 1068 // Collect chains from all the memory opeations that copy arguments to the 1069 // stack. They must follow the stack pointer adjustment above and precede the 1070 // call instruction itself. 1071 SmallVector<SDValue, 8> MemOpChains; 1072 1073 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 1074 const CCValAssign &VA = ArgLocs[i]; 1075 SDValue Arg = CLI.OutVals[i]; 1076 1077 // Promote the value if needed. 1078 switch (VA.getLocInfo()) { 1079 default: 1080 llvm_unreachable("Unknown location info!"); 1081 case CCValAssign::Full: 1082 break; 1083 case CCValAssign::SExt: 1084 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg); 1085 break; 1086 case CCValAssign::ZExt: 1087 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg); 1088 break; 1089 case CCValAssign::AExt: 1090 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg); 1091 break; 1092 case CCValAssign::BCvt: 1093 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg); 1094 break; 1095 } 1096 1097 if (VA.isRegLoc()) { 1098 // The custom bit on an i32 return value indicates that it should be 1099 // passed in the high bits of the register. 1100 if (VA.getValVT() == MVT::i32 && VA.needsCustom()) { 1101 Arg = DAG.getNode(ISD::SHL, DL, MVT::i64, Arg, 1102 DAG.getConstant(32, MVT::i32)); 1103 1104 // The next value may go in the low bits of the same register. 1105 // Handle both at once. 1106 if (i+1 < ArgLocs.size() && ArgLocs[i+1].isRegLoc() && 1107 ArgLocs[i+1].getLocReg() == VA.getLocReg()) { 1108 SDValue NV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, 1109 CLI.OutVals[i+1]); 1110 Arg = DAG.getNode(ISD::OR, DL, MVT::i64, Arg, NV); 1111 // Skip the next value, it's already done. 1112 ++i; 1113 } 1114 } 1115 RegsToPass.push_back(std::make_pair(toCallerWindow(VA.getLocReg()), Arg)); 1116 continue; 1117 } 1118 1119 assert(VA.isMemLoc()); 1120 1121 // Create a store off the stack pointer for this argument. 1122 SDValue StackPtr = DAG.getRegister(SP::O6, getPointerTy()); 1123 // The argument area starts at %fp+BIAS+128 in the callee frame, 1124 // %sp+BIAS+128 in ours. 1125 SDValue PtrOff = DAG.getIntPtrConstant(VA.getLocMemOffset() + 1126 Subtarget->getStackPointerBias() + 1127 128); 1128 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff); 1129 MemOpChains.push_back(DAG.getStore(Chain, DL, Arg, PtrOff, 1130 MachinePointerInfo(), 1131 false, false, 0)); 1132 } 1133 1134 // Emit all stores, make sure they occur before the call. 1135 if (!MemOpChains.empty()) 1136 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, 1137 &MemOpChains[0], MemOpChains.size()); 1138 1139 // Build a sequence of CopyToReg nodes glued together with token chain and 1140 // glue operands which copy the outgoing args into registers. The InGlue is 1141 // necessary since all emitted instructions must be stuck together in order 1142 // to pass the live physical registers. 1143 SDValue InGlue; 1144 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 1145 Chain = DAG.getCopyToReg(Chain, DL, 1146 RegsToPass[i].first, RegsToPass[i].second, InGlue); 1147 InGlue = Chain.getValue(1); 1148 } 1149 1150 // If the callee is a GlobalAddress node (quite common, every direct call is) 1151 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. 1152 // Likewise ExternalSymbol -> TargetExternalSymbol. 1153 SDValue Callee = CLI.Callee; 1154 bool hasReturnsTwice = hasReturnsTwiceAttr(DAG, Callee, CLI.CS); 1155 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) 1156 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, getPointerTy()); 1157 else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee)) 1158 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), getPointerTy()); 1159 1160 // Build the operands for the call instruction itself. 1161 SmallVector<SDValue, 8> Ops; 1162 Ops.push_back(Chain); 1163 Ops.push_back(Callee); 1164 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 1165 Ops.push_back(DAG.getRegister(RegsToPass[i].first, 1166 RegsToPass[i].second.getValueType())); 1167 1168 // Add a register mask operand representing the call-preserved registers. 1169 const SparcRegisterInfo *TRI = 1170 ((const SparcTargetMachine&)getTargetMachine()).getRegisterInfo(); 1171 const uint32_t *Mask = ((hasReturnsTwice) 1172 ? TRI->getRTCallPreservedMask(CLI.CallConv) 1173 : TRI->getCallPreservedMask(CLI.CallConv)); 1174 assert(Mask && "Missing call preserved mask for calling convention"); 1175 Ops.push_back(DAG.getRegisterMask(Mask)); 1176 1177 // Make sure the CopyToReg nodes are glued to the call instruction which 1178 // consumes the registers. 1179 if (InGlue.getNode()) 1180 Ops.push_back(InGlue); 1181 1182 // Now the call itself. 1183 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 1184 Chain = DAG.getNode(SPISD::CALL, DL, NodeTys, &Ops[0], Ops.size()); 1185 InGlue = Chain.getValue(1); 1186 1187 // Revert the stack pointer immediately after the call. 1188 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(ArgsSize, true), 1189 DAG.getIntPtrConstant(0, true), InGlue, DL); 1190 InGlue = Chain.getValue(1); 1191 1192 // Now extract the return values. This is more or less the same as 1193 // LowerFormalArguments_64. 1194 1195 // Assign locations to each value returned by this call. 1196 SmallVector<CCValAssign, 16> RVLocs; 1197 CCState RVInfo(CLI.CallConv, CLI.IsVarArg, DAG.getMachineFunction(), 1198 DAG.getTarget(), RVLocs, *DAG.getContext()); 1199 RVInfo.AnalyzeCallResult(CLI.Ins, CC_Sparc64); 1200 1201 // Copy all of the result registers out of their specified physreg. 1202 for (unsigned i = 0; i != RVLocs.size(); ++i) { 1203 CCValAssign &VA = RVLocs[i]; 1204 unsigned Reg = toCallerWindow(VA.getLocReg()); 1205 1206 // When returning 'inreg {i32, i32 }', two consecutive i32 arguments can 1207 // reside in the same register in the high and low bits. Reuse the 1208 // CopyFromReg previous node to avoid duplicate copies. 1209 SDValue RV; 1210 if (RegisterSDNode *SrcReg = dyn_cast<RegisterSDNode>(Chain.getOperand(1))) 1211 if (SrcReg->getReg() == Reg && Chain->getOpcode() == ISD::CopyFromReg) 1212 RV = Chain.getValue(0); 1213 1214 // But usually we'll create a new CopyFromReg for a different register. 1215 if (!RV.getNode()) { 1216 RV = DAG.getCopyFromReg(Chain, DL, Reg, RVLocs[i].getLocVT(), InGlue); 1217 Chain = RV.getValue(1); 1218 InGlue = Chain.getValue(2); 1219 } 1220 1221 // Get the high bits for i32 struct elements. 1222 if (VA.getValVT() == MVT::i32 && VA.needsCustom()) 1223 RV = DAG.getNode(ISD::SRL, DL, VA.getLocVT(), RV, 1224 DAG.getConstant(32, MVT::i32)); 1225 1226 // The callee promoted the return value, so insert an Assert?ext SDNode so 1227 // we won't promote the value again in this function. 1228 switch (VA.getLocInfo()) { 1229 case CCValAssign::SExt: 1230 RV = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), RV, 1231 DAG.getValueType(VA.getValVT())); 1232 break; 1233 case CCValAssign::ZExt: 1234 RV = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), RV, 1235 DAG.getValueType(VA.getValVT())); 1236 break; 1237 default: 1238 break; 1239 } 1240 1241 // Truncate the register down to the return value type. 1242 if (VA.isExtInLoc()) 1243 RV = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), RV); 1244 1245 InVals.push_back(RV); 1246 } 1247 1248 return Chain; 1249} 1250 1251//===----------------------------------------------------------------------===// 1252// TargetLowering Implementation 1253//===----------------------------------------------------------------------===// 1254 1255/// IntCondCCodeToICC - Convert a DAG integer condition code to a SPARC ICC 1256/// condition. 1257static SPCC::CondCodes IntCondCCodeToICC(ISD::CondCode CC) { 1258 switch (CC) { 1259 default: llvm_unreachable("Unknown integer condition code!"); 1260 case ISD::SETEQ: return SPCC::ICC_E; 1261 case ISD::SETNE: return SPCC::ICC_NE; 1262 case ISD::SETLT: return SPCC::ICC_L; 1263 case ISD::SETGT: return SPCC::ICC_G; 1264 case ISD::SETLE: return SPCC::ICC_LE; 1265 case ISD::SETGE: return SPCC::ICC_GE; 1266 case ISD::SETULT: return SPCC::ICC_CS; 1267 case ISD::SETULE: return SPCC::ICC_LEU; 1268 case ISD::SETUGT: return SPCC::ICC_GU; 1269 case ISD::SETUGE: return SPCC::ICC_CC; 1270 } 1271} 1272 1273/// FPCondCCodeToFCC - Convert a DAG floatingp oint condition code to a SPARC 1274/// FCC condition. 1275static SPCC::CondCodes FPCondCCodeToFCC(ISD::CondCode CC) { 1276 switch (CC) { 1277 default: llvm_unreachable("Unknown fp condition code!"); 1278 case ISD::SETEQ: 1279 case ISD::SETOEQ: return SPCC::FCC_E; 1280 case ISD::SETNE: 1281 case ISD::SETUNE: return SPCC::FCC_NE; 1282 case ISD::SETLT: 1283 case ISD::SETOLT: return SPCC::FCC_L; 1284 case ISD::SETGT: 1285 case ISD::SETOGT: return SPCC::FCC_G; 1286 case ISD::SETLE: 1287 case ISD::SETOLE: return SPCC::FCC_LE; 1288 case ISD::SETGE: 1289 case ISD::SETOGE: return SPCC::FCC_GE; 1290 case ISD::SETULT: return SPCC::FCC_UL; 1291 case ISD::SETULE: return SPCC::FCC_ULE; 1292 case ISD::SETUGT: return SPCC::FCC_UG; 1293 case ISD::SETUGE: return SPCC::FCC_UGE; 1294 case ISD::SETUO: return SPCC::FCC_U; 1295 case ISD::SETO: return SPCC::FCC_O; 1296 case ISD::SETONE: return SPCC::FCC_LG; 1297 case ISD::SETUEQ: return SPCC::FCC_UE; 1298 } 1299} 1300 1301SparcTargetLowering::SparcTargetLowering(TargetMachine &TM) 1302 : TargetLowering(TM, new TargetLoweringObjectFileELF()) { 1303 Subtarget = &TM.getSubtarget<SparcSubtarget>(); 1304 1305 // Set up the register classes. 1306 addRegisterClass(MVT::i32, &SP::IntRegsRegClass); 1307 addRegisterClass(MVT::f32, &SP::FPRegsRegClass); 1308 addRegisterClass(MVT::f64, &SP::DFPRegsRegClass); 1309 addRegisterClass(MVT::f128, &SP::QFPRegsRegClass); 1310 if (Subtarget->is64Bit()) 1311 addRegisterClass(MVT::i64, &SP::I64RegsRegClass); 1312 1313 // Turn FP extload into load/fextend 1314 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand); 1315 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand); 1316 1317 // Sparc doesn't have i1 sign extending load 1318 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 1319 1320 // Turn FP truncstore into trunc + store. 1321 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 1322 setTruncStoreAction(MVT::f128, MVT::f32, Expand); 1323 setTruncStoreAction(MVT::f128, MVT::f64, Expand); 1324 1325 // Custom legalize GlobalAddress nodes into LO/HI parts. 1326 setOperationAction(ISD::GlobalAddress, getPointerTy(), Custom); 1327 setOperationAction(ISD::GlobalTLSAddress, getPointerTy(), Custom); 1328 setOperationAction(ISD::ConstantPool, getPointerTy(), Custom); 1329 setOperationAction(ISD::BlockAddress, getPointerTy(), Custom); 1330 1331 // Sparc doesn't have sext_inreg, replace them with shl/sra 1332 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 1333 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand); 1334 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand); 1335 1336 // Sparc has no REM or DIVREM operations. 1337 setOperationAction(ISD::UREM, MVT::i32, Expand); 1338 setOperationAction(ISD::SREM, MVT::i32, Expand); 1339 setOperationAction(ISD::SDIVREM, MVT::i32, Expand); 1340 setOperationAction(ISD::UDIVREM, MVT::i32, Expand); 1341 1342 // Custom expand fp<->sint 1343 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 1344 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom); 1345 1346 // Expand fp<->uint 1347 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand); 1348 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand); 1349 1350 setOperationAction(ISD::BITCAST, MVT::f32, Expand); 1351 setOperationAction(ISD::BITCAST, MVT::i32, Expand); 1352 1353 // Sparc has no select or setcc: expand to SELECT_CC. 1354 setOperationAction(ISD::SELECT, MVT::i32, Expand); 1355 setOperationAction(ISD::SELECT, MVT::f32, Expand); 1356 setOperationAction(ISD::SELECT, MVT::f64, Expand); 1357 setOperationAction(ISD::SELECT, MVT::f128, Expand); 1358 1359 setOperationAction(ISD::SETCC, MVT::i32, Expand); 1360 setOperationAction(ISD::SETCC, MVT::f32, Expand); 1361 setOperationAction(ISD::SETCC, MVT::f64, Expand); 1362 setOperationAction(ISD::SETCC, MVT::f128, Expand); 1363 1364 // Sparc doesn't have BRCOND either, it has BR_CC. 1365 setOperationAction(ISD::BRCOND, MVT::Other, Expand); 1366 setOperationAction(ISD::BRIND, MVT::Other, Expand); 1367 setOperationAction(ISD::BR_JT, MVT::Other, Expand); 1368 setOperationAction(ISD::BR_CC, MVT::i32, Custom); 1369 setOperationAction(ISD::BR_CC, MVT::f32, Custom); 1370 setOperationAction(ISD::BR_CC, MVT::f64, Custom); 1371 setOperationAction(ISD::BR_CC, MVT::f128, Custom); 1372 1373 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom); 1374 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 1375 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 1376 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom); 1377 1378 if (Subtarget->is64Bit()) { 1379 setOperationAction(ISD::ADDC, MVT::i64, Custom); 1380 setOperationAction(ISD::ADDE, MVT::i64, Custom); 1381 setOperationAction(ISD::SUBC, MVT::i64, Custom); 1382 setOperationAction(ISD::SUBE, MVT::i64, Custom); 1383 setOperationAction(ISD::BITCAST, MVT::f64, Expand); 1384 setOperationAction(ISD::BITCAST, MVT::i64, Expand); 1385 setOperationAction(ISD::SELECT, MVT::i64, Expand); 1386 setOperationAction(ISD::SETCC, MVT::i64, Expand); 1387 setOperationAction(ISD::BR_CC, MVT::i64, Custom); 1388 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom); 1389 } 1390 1391 // FIXME: There are instructions available for ATOMIC_FENCE 1392 // on SparcV8 and later. 1393 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand); 1394 1395 if (!Subtarget->isV9()) { 1396 // SparcV8 does not have FNEGD and FABSD. 1397 setOperationAction(ISD::FNEG, MVT::f64, Custom); 1398 setOperationAction(ISD::FABS, MVT::f64, Custom); 1399 } 1400 1401 setOperationAction(ISD::FSIN , MVT::f128, Expand); 1402 setOperationAction(ISD::FCOS , MVT::f128, Expand); 1403 setOperationAction(ISD::FSINCOS, MVT::f128, Expand); 1404 setOperationAction(ISD::FREM , MVT::f128, Expand); 1405 setOperationAction(ISD::FMA , MVT::f128, Expand); 1406 setOperationAction(ISD::FSIN , MVT::f64, Expand); 1407 setOperationAction(ISD::FCOS , MVT::f64, Expand); 1408 setOperationAction(ISD::FSINCOS, MVT::f64, Expand); 1409 setOperationAction(ISD::FREM , MVT::f64, Expand); 1410 setOperationAction(ISD::FMA , MVT::f64, Expand); 1411 setOperationAction(ISD::FSIN , MVT::f32, Expand); 1412 setOperationAction(ISD::FCOS , MVT::f32, Expand); 1413 setOperationAction(ISD::FSINCOS, MVT::f32, Expand); 1414 setOperationAction(ISD::FREM , MVT::f32, Expand); 1415 setOperationAction(ISD::FMA , MVT::f32, Expand); 1416 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 1417 setOperationAction(ISD::CTTZ , MVT::i32, Expand); 1418 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand); 1419 setOperationAction(ISD::CTLZ , MVT::i32, Expand); 1420 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand); 1421 setOperationAction(ISD::ROTL , MVT::i32, Expand); 1422 setOperationAction(ISD::ROTR , MVT::i32, Expand); 1423 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 1424 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand); 1425 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand); 1426 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand); 1427 setOperationAction(ISD::FPOW , MVT::f128, Expand); 1428 setOperationAction(ISD::FPOW , MVT::f64, Expand); 1429 setOperationAction(ISD::FPOW , MVT::f32, Expand); 1430 1431 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand); 1432 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand); 1433 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand); 1434 1435 // FIXME: Sparc provides these multiplies, but we don't have them yet. 1436 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand); 1437 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand); 1438 1439 // VASTART needs to be custom lowered to use the VarArgsFrameIndex. 1440 setOperationAction(ISD::VASTART , MVT::Other, Custom); 1441 // VAARG needs to be lowered to not do unaligned accesses for doubles. 1442 setOperationAction(ISD::VAARG , MVT::Other, Custom); 1443 1444 // Use the default implementation. 1445 setOperationAction(ISD::VACOPY , MVT::Other, Expand); 1446 setOperationAction(ISD::VAEND , MVT::Other, Expand); 1447 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand); 1448 setOperationAction(ISD::STACKRESTORE , MVT::Other, Expand); 1449 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom); 1450 1451 setExceptionPointerRegister(SP::I0); 1452 setExceptionSelectorRegister(SP::I1); 1453 1454 setStackPointerRegisterToSaveRestore(SP::O6); 1455 1456 if (Subtarget->isV9()) 1457 setOperationAction(ISD::CTPOP, MVT::i32, Legal); 1458 1459 if (Subtarget->isV9() && Subtarget->hasHardQuad()) { 1460 setOperationAction(ISD::LOAD, MVT::f128, Legal); 1461 setOperationAction(ISD::STORE, MVT::f128, Legal); 1462 } else { 1463 setOperationAction(ISD::LOAD, MVT::f128, Custom); 1464 setOperationAction(ISD::STORE, MVT::f128, Custom); 1465 } 1466 1467 if (Subtarget->hasHardQuad()) { 1468 setOperationAction(ISD::FADD, MVT::f128, Legal); 1469 setOperationAction(ISD::FSUB, MVT::f128, Legal); 1470 setOperationAction(ISD::FMUL, MVT::f128, Legal); 1471 setOperationAction(ISD::FDIV, MVT::f128, Legal); 1472 setOperationAction(ISD::FSQRT, MVT::f128, Legal); 1473 setOperationAction(ISD::FP_EXTEND, MVT::f128, Legal); 1474 setOperationAction(ISD::FP_ROUND, MVT::f64, Legal); 1475 if (Subtarget->isV9()) { 1476 setOperationAction(ISD::FNEG, MVT::f128, Legal); 1477 setOperationAction(ISD::FABS, MVT::f128, Legal); 1478 } else { 1479 setOperationAction(ISD::FNEG, MVT::f128, Custom); 1480 setOperationAction(ISD::FABS, MVT::f128, Custom); 1481 } 1482 } else { 1483 // Custom legalize f128 operations. 1484 1485 setOperationAction(ISD::FADD, MVT::f128, Custom); 1486 setOperationAction(ISD::FSUB, MVT::f128, Custom); 1487 setOperationAction(ISD::FMUL, MVT::f128, Custom); 1488 setOperationAction(ISD::FDIV, MVT::f128, Custom); 1489 setOperationAction(ISD::FSQRT, MVT::f128, Custom); 1490 setOperationAction(ISD::FNEG, MVT::f128, Custom); 1491 setOperationAction(ISD::FABS, MVT::f128, Custom); 1492 1493 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom); 1494 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom); 1495 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom); 1496 1497 // Setup Runtime library names. 1498 if (Subtarget->is64Bit()) { 1499 setLibcallName(RTLIB::ADD_F128, "_Qp_add"); 1500 setLibcallName(RTLIB::SUB_F128, "_Qp_sub"); 1501 setLibcallName(RTLIB::MUL_F128, "_Qp_mul"); 1502 setLibcallName(RTLIB::DIV_F128, "_Qp_div"); 1503 setLibcallName(RTLIB::SQRT_F128, "_Qp_sqrt"); 1504 setLibcallName(RTLIB::FPTOSINT_F128_I32, "_Qp_qtoi"); 1505 setLibcallName(RTLIB::SINTTOFP_I32_F128, "_Qp_itoq"); 1506 setLibcallName(RTLIB::FPEXT_F32_F128, "_Qp_stoq"); 1507 setLibcallName(RTLIB::FPEXT_F64_F128, "_Qp_dtoq"); 1508 setLibcallName(RTLIB::FPROUND_F128_F32, "_Qp_qtos"); 1509 setLibcallName(RTLIB::FPROUND_F128_F64, "_Qp_qtod"); 1510 } else { 1511 setLibcallName(RTLIB::ADD_F128, "_Q_add"); 1512 setLibcallName(RTLIB::SUB_F128, "_Q_sub"); 1513 setLibcallName(RTLIB::MUL_F128, "_Q_mul"); 1514 setLibcallName(RTLIB::DIV_F128, "_Q_div"); 1515 setLibcallName(RTLIB::SQRT_F128, "_Q_sqrt"); 1516 setLibcallName(RTLIB::FPTOSINT_F128_I32, "_Q_qtoi"); 1517 setLibcallName(RTLIB::SINTTOFP_I32_F128, "_Q_itoq"); 1518 setLibcallName(RTLIB::FPEXT_F32_F128, "_Q_stoq"); 1519 setLibcallName(RTLIB::FPEXT_F64_F128, "_Q_dtoq"); 1520 setLibcallName(RTLIB::FPROUND_F128_F32, "_Q_qtos"); 1521 setLibcallName(RTLIB::FPROUND_F128_F64, "_Q_qtod"); 1522 } 1523 } 1524 1525 setMinFunctionAlignment(2); 1526 1527 computeRegisterProperties(); 1528} 1529 1530const char *SparcTargetLowering::getTargetNodeName(unsigned Opcode) const { 1531 switch (Opcode) { 1532 default: return 0; 1533 case SPISD::CMPICC: return "SPISD::CMPICC"; 1534 case SPISD::CMPFCC: return "SPISD::CMPFCC"; 1535 case SPISD::BRICC: return "SPISD::BRICC"; 1536 case SPISD::BRXCC: return "SPISD::BRXCC"; 1537 case SPISD::BRFCC: return "SPISD::BRFCC"; 1538 case SPISD::SELECT_ICC: return "SPISD::SELECT_ICC"; 1539 case SPISD::SELECT_XCC: return "SPISD::SELECT_XCC"; 1540 case SPISD::SELECT_FCC: return "SPISD::SELECT_FCC"; 1541 case SPISD::Hi: return "SPISD::Hi"; 1542 case SPISD::Lo: return "SPISD::Lo"; 1543 case SPISD::FTOI: return "SPISD::FTOI"; 1544 case SPISD::ITOF: return "SPISD::ITOF"; 1545 case SPISD::CALL: return "SPISD::CALL"; 1546 case SPISD::RET_FLAG: return "SPISD::RET_FLAG"; 1547 case SPISD::GLOBAL_BASE_REG: return "SPISD::GLOBAL_BASE_REG"; 1548 case SPISD::FLUSHW: return "SPISD::FLUSHW"; 1549 case SPISD::TLS_ADD: return "SPISD::TLS_ADD"; 1550 case SPISD::TLS_LD: return "SPISD::TLS_LD"; 1551 case SPISD::TLS_CALL: return "SPISD::TLS_CALL"; 1552 } 1553} 1554 1555/// isMaskedValueZeroForTargetNode - Return true if 'Op & Mask' is known to 1556/// be zero. Op is expected to be a target specific node. Used by DAG 1557/// combiner. 1558void SparcTargetLowering::computeMaskedBitsForTargetNode 1559 (const SDValue Op, 1560 APInt &KnownZero, 1561 APInt &KnownOne, 1562 const SelectionDAG &DAG, 1563 unsigned Depth) const { 1564 APInt KnownZero2, KnownOne2; 1565 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0); 1566 1567 switch (Op.getOpcode()) { 1568 default: break; 1569 case SPISD::SELECT_ICC: 1570 case SPISD::SELECT_XCC: 1571 case SPISD::SELECT_FCC: 1572 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZero, KnownOne, Depth+1); 1573 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero2, KnownOne2, Depth+1); 1574 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1575 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1576 1577 // Only known if known in both the LHS and RHS. 1578 KnownOne &= KnownOne2; 1579 KnownZero &= KnownZero2; 1580 break; 1581 } 1582} 1583 1584// Look at LHS/RHS/CC and see if they are a lowered setcc instruction. If so 1585// set LHS/RHS and SPCC to the LHS/RHS of the setcc and SPCC to the condition. 1586static void LookThroughSetCC(SDValue &LHS, SDValue &RHS, 1587 ISD::CondCode CC, unsigned &SPCC) { 1588 if (isa<ConstantSDNode>(RHS) && 1589 cast<ConstantSDNode>(RHS)->isNullValue() && 1590 CC == ISD::SETNE && 1591 (((LHS.getOpcode() == SPISD::SELECT_ICC || 1592 LHS.getOpcode() == SPISD::SELECT_XCC) && 1593 LHS.getOperand(3).getOpcode() == SPISD::CMPICC) || 1594 (LHS.getOpcode() == SPISD::SELECT_FCC && 1595 LHS.getOperand(3).getOpcode() == SPISD::CMPFCC)) && 1596 isa<ConstantSDNode>(LHS.getOperand(0)) && 1597 isa<ConstantSDNode>(LHS.getOperand(1)) && 1598 cast<ConstantSDNode>(LHS.getOperand(0))->isOne() && 1599 cast<ConstantSDNode>(LHS.getOperand(1))->isNullValue()) { 1600 SDValue CMPCC = LHS.getOperand(3); 1601 SPCC = cast<ConstantSDNode>(LHS.getOperand(2))->getZExtValue(); 1602 LHS = CMPCC.getOperand(0); 1603 RHS = CMPCC.getOperand(1); 1604 } 1605} 1606 1607// Convert to a target node and set target flags. 1608SDValue SparcTargetLowering::withTargetFlags(SDValue Op, unsigned TF, 1609 SelectionDAG &DAG) const { 1610 if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) 1611 return DAG.getTargetGlobalAddress(GA->getGlobal(), 1612 SDLoc(GA), 1613 GA->getValueType(0), 1614 GA->getOffset(), TF); 1615 1616 if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op)) 1617 return DAG.getTargetConstantPool(CP->getConstVal(), 1618 CP->getValueType(0), 1619 CP->getAlignment(), 1620 CP->getOffset(), TF); 1621 1622 if (const BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(Op)) 1623 return DAG.getTargetBlockAddress(BA->getBlockAddress(), 1624 Op.getValueType(), 1625 0, 1626 TF); 1627 1628 if (const ExternalSymbolSDNode *ES = dyn_cast<ExternalSymbolSDNode>(Op)) 1629 return DAG.getTargetExternalSymbol(ES->getSymbol(), 1630 ES->getValueType(0), TF); 1631 1632 llvm_unreachable("Unhandled address SDNode"); 1633} 1634 1635// Split Op into high and low parts according to HiTF and LoTF. 1636// Return an ADD node combining the parts. 1637SDValue SparcTargetLowering::makeHiLoPair(SDValue Op, 1638 unsigned HiTF, unsigned LoTF, 1639 SelectionDAG &DAG) const { 1640 SDLoc DL(Op); 1641 EVT VT = Op.getValueType(); 1642 SDValue Hi = DAG.getNode(SPISD::Hi, DL, VT, withTargetFlags(Op, HiTF, DAG)); 1643 SDValue Lo = DAG.getNode(SPISD::Lo, DL, VT, withTargetFlags(Op, LoTF, DAG)); 1644 return DAG.getNode(ISD::ADD, DL, VT, Hi, Lo); 1645} 1646 1647// Build SDNodes for producing an address from a GlobalAddress, ConstantPool, 1648// or ExternalSymbol SDNode. 1649SDValue SparcTargetLowering::makeAddress(SDValue Op, SelectionDAG &DAG) const { 1650 SDLoc DL(Op); 1651 EVT VT = getPointerTy(); 1652 1653 // Handle PIC mode first. 1654 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) { 1655 // This is the pic32 code model, the GOT is known to be smaller than 4GB. 1656 SDValue HiLo = makeHiLoPair(Op, SPII::MO_HI, SPII::MO_LO, DAG); 1657 SDValue GlobalBase = DAG.getNode(SPISD::GLOBAL_BASE_REG, DL, VT); 1658 SDValue AbsAddr = DAG.getNode(ISD::ADD, DL, VT, GlobalBase, HiLo); 1659 // GLOBAL_BASE_REG codegen'ed with call. Inform MFI that this 1660 // function has calls. 1661 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 1662 MFI->setHasCalls(true); 1663 return DAG.getLoad(VT, DL, DAG.getEntryNode(), AbsAddr, 1664 MachinePointerInfo::getGOT(), false, false, false, 0); 1665 } 1666 1667 // This is one of the absolute code models. 1668 switch(getTargetMachine().getCodeModel()) { 1669 default: 1670 llvm_unreachable("Unsupported absolute code model"); 1671 case CodeModel::Small: 1672 // abs32. 1673 return makeHiLoPair(Op, SPII::MO_HI, SPII::MO_LO, DAG); 1674 case CodeModel::Medium: { 1675 // abs44. 1676 SDValue H44 = makeHiLoPair(Op, SPII::MO_H44, SPII::MO_M44, DAG); 1677 H44 = DAG.getNode(ISD::SHL, DL, VT, H44, DAG.getConstant(12, MVT::i32)); 1678 SDValue L44 = withTargetFlags(Op, SPII::MO_L44, DAG); 1679 L44 = DAG.getNode(SPISD::Lo, DL, VT, L44); 1680 return DAG.getNode(ISD::ADD, DL, VT, H44, L44); 1681 } 1682 case CodeModel::Large: { 1683 // abs64. 1684 SDValue Hi = makeHiLoPair(Op, SPII::MO_HH, SPII::MO_HM, DAG); 1685 Hi = DAG.getNode(ISD::SHL, DL, VT, Hi, DAG.getConstant(32, MVT::i32)); 1686 SDValue Lo = makeHiLoPair(Op, SPII::MO_HI, SPII::MO_LO, DAG); 1687 return DAG.getNode(ISD::ADD, DL, VT, Hi, Lo); 1688 } 1689 } 1690} 1691 1692SDValue SparcTargetLowering::LowerGlobalAddress(SDValue Op, 1693 SelectionDAG &DAG) const { 1694 return makeAddress(Op, DAG); 1695} 1696 1697SDValue SparcTargetLowering::LowerConstantPool(SDValue Op, 1698 SelectionDAG &DAG) const { 1699 return makeAddress(Op, DAG); 1700} 1701 1702SDValue SparcTargetLowering::LowerBlockAddress(SDValue Op, 1703 SelectionDAG &DAG) const { 1704 return makeAddress(Op, DAG); 1705} 1706 1707SDValue SparcTargetLowering::LowerGlobalTLSAddress(SDValue Op, 1708 SelectionDAG &DAG) const { 1709 1710 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 1711 SDLoc DL(GA); 1712 const GlobalValue *GV = GA->getGlobal(); 1713 EVT PtrVT = getPointerTy(); 1714 1715 TLSModel::Model model = getTargetMachine().getTLSModel(GV); 1716 1717 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) { 1718 unsigned HiTF = ((model == TLSModel::GeneralDynamic)? SPII::MO_TLS_GD_HI22 1719 : SPII::MO_TLS_LDM_HI22); 1720 unsigned LoTF = ((model == TLSModel::GeneralDynamic)? SPII::MO_TLS_GD_LO10 1721 : SPII::MO_TLS_LDM_LO10); 1722 unsigned addTF = ((model == TLSModel::GeneralDynamic)? SPII::MO_TLS_GD_ADD 1723 : SPII::MO_TLS_LDM_ADD); 1724 unsigned callTF = ((model == TLSModel::GeneralDynamic)? SPII::MO_TLS_GD_CALL 1725 : SPII::MO_TLS_LDM_CALL); 1726 1727 SDValue HiLo = makeHiLoPair(Op, HiTF, LoTF, DAG); 1728 SDValue Base = DAG.getNode(SPISD::GLOBAL_BASE_REG, DL, PtrVT); 1729 SDValue Argument = DAG.getNode(SPISD::TLS_ADD, DL, PtrVT, Base, HiLo, 1730 withTargetFlags(Op, addTF, DAG)); 1731 1732 SDValue Chain = DAG.getEntryNode(); 1733 SDValue InFlag; 1734 1735 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(1, true), DL); 1736 Chain = DAG.getCopyToReg(Chain, DL, SP::O0, Argument, InFlag); 1737 InFlag = Chain.getValue(1); 1738 SDValue Callee = DAG.getTargetExternalSymbol("__tls_get_addr", PtrVT); 1739 SDValue Symbol = withTargetFlags(Op, callTF, DAG); 1740 1741 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 1742 SmallVector<SDValue, 4> Ops; 1743 Ops.push_back(Chain); 1744 Ops.push_back(Callee); 1745 Ops.push_back(Symbol); 1746 Ops.push_back(DAG.getRegister(SP::O0, PtrVT)); 1747 const uint32_t *Mask = getTargetMachine() 1748 .getRegisterInfo()->getCallPreservedMask(CallingConv::C); 1749 assert(Mask && "Missing call preserved mask for calling convention"); 1750 Ops.push_back(DAG.getRegisterMask(Mask)); 1751 Ops.push_back(InFlag); 1752 Chain = DAG.getNode(SPISD::TLS_CALL, DL, NodeTys, &Ops[0], Ops.size()); 1753 InFlag = Chain.getValue(1); 1754 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(1, true), 1755 DAG.getIntPtrConstant(0, true), InFlag, DL); 1756 InFlag = Chain.getValue(1); 1757 SDValue Ret = DAG.getCopyFromReg(Chain, DL, SP::O0, PtrVT, InFlag); 1758 1759 if (model != TLSModel::LocalDynamic) 1760 return Ret; 1761 1762 SDValue Hi = DAG.getNode(SPISD::Hi, DL, PtrVT, 1763 withTargetFlags(Op, SPII::MO_TLS_LDO_HIX22, DAG)); 1764 SDValue Lo = DAG.getNode(SPISD::Lo, DL, PtrVT, 1765 withTargetFlags(Op, SPII::MO_TLS_LDO_LOX10, DAG)); 1766 HiLo = DAG.getNode(ISD::XOR, DL, PtrVT, Hi, Lo); 1767 return DAG.getNode(SPISD::TLS_ADD, DL, PtrVT, Ret, HiLo, 1768 withTargetFlags(Op, SPII::MO_TLS_LDO_ADD, DAG)); 1769 } 1770 1771 if (model == TLSModel::InitialExec) { 1772 unsigned ldTF = ((PtrVT == MVT::i64)? SPII::MO_TLS_IE_LDX 1773 : SPII::MO_TLS_IE_LD); 1774 1775 SDValue Base = DAG.getNode(SPISD::GLOBAL_BASE_REG, DL, PtrVT); 1776 1777 // GLOBAL_BASE_REG codegen'ed with call. Inform MFI that this 1778 // function has calls. 1779 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 1780 MFI->setHasCalls(true); 1781 1782 SDValue TGA = makeHiLoPair(Op, 1783 SPII::MO_TLS_IE_HI22, SPII::MO_TLS_IE_LO10, DAG); 1784 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Base, TGA); 1785 SDValue Offset = DAG.getNode(SPISD::TLS_LD, 1786 DL, PtrVT, Ptr, 1787 withTargetFlags(Op, ldTF, DAG)); 1788 return DAG.getNode(SPISD::TLS_ADD, DL, PtrVT, 1789 DAG.getRegister(SP::G7, PtrVT), Offset, 1790 withTargetFlags(Op, SPII::MO_TLS_IE_ADD, DAG)); 1791 } 1792 1793 assert(model == TLSModel::LocalExec); 1794 SDValue Hi = DAG.getNode(SPISD::Hi, DL, PtrVT, 1795 withTargetFlags(Op, SPII::MO_TLS_LE_HIX22, DAG)); 1796 SDValue Lo = DAG.getNode(SPISD::Lo, DL, PtrVT, 1797 withTargetFlags(Op, SPII::MO_TLS_LE_LOX10, DAG)); 1798 SDValue Offset = DAG.getNode(ISD::XOR, DL, PtrVT, Hi, Lo); 1799 1800 return DAG.getNode(ISD::ADD, DL, PtrVT, 1801 DAG.getRegister(SP::G7, PtrVT), Offset); 1802} 1803 1804SDValue 1805SparcTargetLowering::LowerF128_LibCallArg(SDValue Chain, ArgListTy &Args, 1806 SDValue Arg, SDLoc DL, 1807 SelectionDAG &DAG) const { 1808 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 1809 EVT ArgVT = Arg.getValueType(); 1810 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); 1811 1812 ArgListEntry Entry; 1813 Entry.Node = Arg; 1814 Entry.Ty = ArgTy; 1815 1816 if (ArgTy->isFP128Ty()) { 1817 // Create a stack object and pass the pointer to the library function. 1818 int FI = MFI->CreateStackObject(16, 8, false); 1819 SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy()); 1820 Chain = DAG.getStore(Chain, 1821 DL, 1822 Entry.Node, 1823 FIPtr, 1824 MachinePointerInfo(), 1825 false, 1826 false, 1827 8); 1828 1829 Entry.Node = FIPtr; 1830 Entry.Ty = PointerType::getUnqual(ArgTy); 1831 } 1832 Args.push_back(Entry); 1833 return Chain; 1834} 1835 1836SDValue 1837SparcTargetLowering::LowerF128Op(SDValue Op, SelectionDAG &DAG, 1838 const char *LibFuncName, 1839 unsigned numArgs) const { 1840 1841 ArgListTy Args; 1842 1843 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 1844 1845 SDValue Callee = DAG.getExternalSymbol(LibFuncName, getPointerTy()); 1846 Type *RetTy = Op.getValueType().getTypeForEVT(*DAG.getContext()); 1847 Type *RetTyABI = RetTy; 1848 SDValue Chain = DAG.getEntryNode(); 1849 SDValue RetPtr; 1850 1851 if (RetTy->isFP128Ty()) { 1852 // Create a Stack Object to receive the return value of type f128. 1853 ArgListEntry Entry; 1854 int RetFI = MFI->CreateStackObject(16, 8, false); 1855 RetPtr = DAG.getFrameIndex(RetFI, getPointerTy()); 1856 Entry.Node = RetPtr; 1857 Entry.Ty = PointerType::getUnqual(RetTy); 1858 if (!Subtarget->is64Bit()) 1859 Entry.isSRet = true; 1860 Entry.isReturned = false; 1861 Args.push_back(Entry); 1862 RetTyABI = Type::getVoidTy(*DAG.getContext()); 1863 } 1864 1865 assert(Op->getNumOperands() >= numArgs && "Not enough operands!"); 1866 for (unsigned i = 0, e = numArgs; i != e; ++i) { 1867 Chain = LowerF128_LibCallArg(Chain, Args, Op.getOperand(i), SDLoc(Op), DAG); 1868 } 1869 TargetLowering:: 1870 CallLoweringInfo CLI(Chain, 1871 RetTyABI, 1872 false, false, false, false, 1873 0, CallingConv::C, 1874 false, false, true, 1875 Callee, Args, DAG, SDLoc(Op)); 1876 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI); 1877 1878 // chain is in second result. 1879 if (RetTyABI == RetTy) 1880 return CallInfo.first; 1881 1882 assert (RetTy->isFP128Ty() && "Unexpected return type!"); 1883 1884 Chain = CallInfo.second; 1885 1886 // Load RetPtr to get the return value. 1887 return DAG.getLoad(Op.getValueType(), 1888 SDLoc(Op), 1889 Chain, 1890 RetPtr, 1891 MachinePointerInfo(), 1892 false, false, false, 8); 1893} 1894 1895SDValue 1896SparcTargetLowering::LowerF128Compare(SDValue LHS, SDValue RHS, 1897 unsigned &SPCC, 1898 SDLoc DL, 1899 SelectionDAG &DAG) const { 1900 1901 const char *LibCall = 0; 1902 bool is64Bit = Subtarget->is64Bit(); 1903 switch(SPCC) { 1904 default: llvm_unreachable("Unhandled conditional code!"); 1905 case SPCC::FCC_E : LibCall = is64Bit? "_Qp_feq" : "_Q_feq"; break; 1906 case SPCC::FCC_NE : LibCall = is64Bit? "_Qp_fne" : "_Q_fne"; break; 1907 case SPCC::FCC_L : LibCall = is64Bit? "_Qp_flt" : "_Q_flt"; break; 1908 case SPCC::FCC_G : LibCall = is64Bit? "_Qp_fgt" : "_Q_fgt"; break; 1909 case SPCC::FCC_LE : LibCall = is64Bit? "_Qp_fle" : "_Q_fle"; break; 1910 case SPCC::FCC_GE : LibCall = is64Bit? "_Qp_fge" : "_Q_fge"; break; 1911 case SPCC::FCC_UL : 1912 case SPCC::FCC_ULE: 1913 case SPCC::FCC_UG : 1914 case SPCC::FCC_UGE: 1915 case SPCC::FCC_U : 1916 case SPCC::FCC_O : 1917 case SPCC::FCC_LG : 1918 case SPCC::FCC_UE : LibCall = is64Bit? "_Qp_cmp" : "_Q_cmp"; break; 1919 } 1920 1921 SDValue Callee = DAG.getExternalSymbol(LibCall, getPointerTy()); 1922 Type *RetTy = Type::getInt32Ty(*DAG.getContext()); 1923 ArgListTy Args; 1924 SDValue Chain = DAG.getEntryNode(); 1925 Chain = LowerF128_LibCallArg(Chain, Args, LHS, DL, DAG); 1926 Chain = LowerF128_LibCallArg(Chain, Args, RHS, DL, DAG); 1927 1928 TargetLowering:: 1929 CallLoweringInfo CLI(Chain, 1930 RetTy, 1931 false, false, false, false, 1932 0, CallingConv::C, 1933 false, false, true, 1934 Callee, Args, DAG, DL); 1935 1936 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI); 1937 1938 // result is in first, and chain is in second result. 1939 SDValue Result = CallInfo.first; 1940 1941 switch(SPCC) { 1942 default: { 1943 SDValue RHS = DAG.getTargetConstant(0, Result.getValueType()); 1944 SPCC = SPCC::ICC_NE; 1945 return DAG.getNode(SPISD::CMPICC, DL, MVT::Glue, Result, RHS); 1946 } 1947 case SPCC::FCC_UL : { 1948 SDValue Mask = DAG.getTargetConstant(1, Result.getValueType()); 1949 Result = DAG.getNode(ISD::AND, DL, Result.getValueType(), Result, Mask); 1950 SDValue RHS = DAG.getTargetConstant(0, Result.getValueType()); 1951 SPCC = SPCC::ICC_NE; 1952 return DAG.getNode(SPISD::CMPICC, DL, MVT::Glue, Result, RHS); 1953 } 1954 case SPCC::FCC_ULE: { 1955 SDValue RHS = DAG.getTargetConstant(2, Result.getValueType()); 1956 SPCC = SPCC::ICC_NE; 1957 return DAG.getNode(SPISD::CMPICC, DL, MVT::Glue, Result, RHS); 1958 } 1959 case SPCC::FCC_UG : { 1960 SDValue RHS = DAG.getTargetConstant(1, Result.getValueType()); 1961 SPCC = SPCC::ICC_G; 1962 return DAG.getNode(SPISD::CMPICC, DL, MVT::Glue, Result, RHS); 1963 } 1964 case SPCC::FCC_UGE: { 1965 SDValue RHS = DAG.getTargetConstant(1, Result.getValueType()); 1966 SPCC = SPCC::ICC_NE; 1967 return DAG.getNode(SPISD::CMPICC, DL, MVT::Glue, Result, RHS); 1968 } 1969 1970 case SPCC::FCC_U : { 1971 SDValue RHS = DAG.getTargetConstant(3, Result.getValueType()); 1972 SPCC = SPCC::ICC_E; 1973 return DAG.getNode(SPISD::CMPICC, DL, MVT::Glue, Result, RHS); 1974 } 1975 case SPCC::FCC_O : { 1976 SDValue RHS = DAG.getTargetConstant(3, Result.getValueType()); 1977 SPCC = SPCC::ICC_NE; 1978 return DAG.getNode(SPISD::CMPICC, DL, MVT::Glue, Result, RHS); 1979 } 1980 case SPCC::FCC_LG : { 1981 SDValue Mask = DAG.getTargetConstant(3, Result.getValueType()); 1982 Result = DAG.getNode(ISD::AND, DL, Result.getValueType(), Result, Mask); 1983 SDValue RHS = DAG.getTargetConstant(0, Result.getValueType()); 1984 SPCC = SPCC::ICC_NE; 1985 return DAG.getNode(SPISD::CMPICC, DL, MVT::Glue, Result, RHS); 1986 } 1987 case SPCC::FCC_UE : { 1988 SDValue Mask = DAG.getTargetConstant(3, Result.getValueType()); 1989 Result = DAG.getNode(ISD::AND, DL, Result.getValueType(), Result, Mask); 1990 SDValue RHS = DAG.getTargetConstant(0, Result.getValueType()); 1991 SPCC = SPCC::ICC_E; 1992 return DAG.getNode(SPISD::CMPICC, DL, MVT::Glue, Result, RHS); 1993 } 1994 } 1995} 1996 1997static SDValue 1998LowerF128_FPEXTEND(SDValue Op, SelectionDAG &DAG, 1999 const SparcTargetLowering &TLI) { 2000 2001 if (Op.getOperand(0).getValueType() == MVT::f64) 2002 return TLI.LowerF128Op(Op, DAG, 2003 TLI.getLibcallName(RTLIB::FPEXT_F64_F128), 1); 2004 2005 if (Op.getOperand(0).getValueType() == MVT::f32) 2006 return TLI.LowerF128Op(Op, DAG, 2007 TLI.getLibcallName(RTLIB::FPEXT_F32_F128), 1); 2008 2009 llvm_unreachable("fpextend with non-float operand!"); 2010 return SDValue(0, 0); 2011} 2012 2013static SDValue 2014LowerF128_FPROUND(SDValue Op, SelectionDAG &DAG, 2015 const SparcTargetLowering &TLI) { 2016 // FP_ROUND on f64 and f32 are legal. 2017 if (Op.getOperand(0).getValueType() != MVT::f128) 2018 return Op; 2019 2020 if (Op.getValueType() == MVT::f64) 2021 return TLI.LowerF128Op(Op, DAG, 2022 TLI.getLibcallName(RTLIB::FPROUND_F128_F64), 1); 2023 if (Op.getValueType() == MVT::f32) 2024 return TLI.LowerF128Op(Op, DAG, 2025 TLI.getLibcallName(RTLIB::FPROUND_F128_F32), 1); 2026 2027 llvm_unreachable("fpround to non-float!"); 2028 return SDValue(0, 0); 2029} 2030 2031static SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG, 2032 const SparcTargetLowering &TLI, 2033 bool hasHardQuad) { 2034 SDLoc dl(Op); 2035 // Convert the fp value to integer in an FP register. 2036 assert(Op.getValueType() == MVT::i32); 2037 2038 if (Op.getOperand(0).getValueType() == MVT::f128 && !hasHardQuad) 2039 return TLI.LowerF128Op(Op, DAG, 2040 TLI.getLibcallName(RTLIB::FPTOSINT_F128_I32), 1); 2041 2042 Op = DAG.getNode(SPISD::FTOI, dl, MVT::f32, Op.getOperand(0)); 2043 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); 2044} 2045 2046static SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG, 2047 const SparcTargetLowering &TLI, 2048 bool hasHardQuad) { 2049 SDLoc dl(Op); 2050 assert(Op.getOperand(0).getValueType() == MVT::i32); 2051 SDValue Tmp = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0)); 2052 // Convert the int value to FP in an FP register. 2053 if (Op.getValueType() == MVT::f128 && !hasHardQuad) 2054 return TLI.LowerF128Op(Op, DAG, 2055 TLI.getLibcallName(RTLIB::SINTTOFP_I32_F128), 1); 2056 return DAG.getNode(SPISD::ITOF, dl, Op.getValueType(), Tmp); 2057} 2058 2059static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG, 2060 const SparcTargetLowering &TLI, 2061 bool hasHardQuad) { 2062 SDValue Chain = Op.getOperand(0); 2063 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 2064 SDValue LHS = Op.getOperand(2); 2065 SDValue RHS = Op.getOperand(3); 2066 SDValue Dest = Op.getOperand(4); 2067 SDLoc dl(Op); 2068 unsigned Opc, SPCC = ~0U; 2069 2070 // If this is a br_cc of a "setcc", and if the setcc got lowered into 2071 // an CMP[IF]CC/SELECT_[IF]CC pair, find the original compared values. 2072 LookThroughSetCC(LHS, RHS, CC, SPCC); 2073 2074 // Get the condition flag. 2075 SDValue CompareFlag; 2076 if (LHS.getValueType().isInteger()) { 2077 CompareFlag = DAG.getNode(SPISD::CMPICC, dl, MVT::Glue, LHS, RHS); 2078 if (SPCC == ~0U) SPCC = IntCondCCodeToICC(CC); 2079 // 32-bit compares use the icc flags, 64-bit uses the xcc flags. 2080 Opc = LHS.getValueType() == MVT::i32 ? SPISD::BRICC : SPISD::BRXCC; 2081 } else { 2082 if (!hasHardQuad && LHS.getValueType() == MVT::f128) { 2083 if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC); 2084 CompareFlag = TLI.LowerF128Compare(LHS, RHS, SPCC, dl, DAG); 2085 Opc = SPISD::BRICC; 2086 } else { 2087 CompareFlag = DAG.getNode(SPISD::CMPFCC, dl, MVT::Glue, LHS, RHS); 2088 if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC); 2089 Opc = SPISD::BRFCC; 2090 } 2091 } 2092 return DAG.getNode(Opc, dl, MVT::Other, Chain, Dest, 2093 DAG.getConstant(SPCC, MVT::i32), CompareFlag); 2094} 2095 2096static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG, 2097 const SparcTargetLowering &TLI, 2098 bool hasHardQuad) { 2099 SDValue LHS = Op.getOperand(0); 2100 SDValue RHS = Op.getOperand(1); 2101 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get(); 2102 SDValue TrueVal = Op.getOperand(2); 2103 SDValue FalseVal = Op.getOperand(3); 2104 SDLoc dl(Op); 2105 unsigned Opc, SPCC = ~0U; 2106 2107 // If this is a select_cc of a "setcc", and if the setcc got lowered into 2108 // an CMP[IF]CC/SELECT_[IF]CC pair, find the original compared values. 2109 LookThroughSetCC(LHS, RHS, CC, SPCC); 2110 2111 SDValue CompareFlag; 2112 if (LHS.getValueType().isInteger()) { 2113 CompareFlag = DAG.getNode(SPISD::CMPICC, dl, MVT::Glue, LHS, RHS); 2114 Opc = LHS.getValueType() == MVT::i32 ? 2115 SPISD::SELECT_ICC : SPISD::SELECT_XCC; 2116 if (SPCC == ~0U) SPCC = IntCondCCodeToICC(CC); 2117 } else { 2118 if (!hasHardQuad && LHS.getValueType() == MVT::f128) { 2119 if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC); 2120 CompareFlag = TLI.LowerF128Compare(LHS, RHS, SPCC, dl, DAG); 2121 Opc = SPISD::SELECT_ICC; 2122 } else { 2123 CompareFlag = DAG.getNode(SPISD::CMPFCC, dl, MVT::Glue, LHS, RHS); 2124 Opc = SPISD::SELECT_FCC; 2125 if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC); 2126 } 2127 } 2128 return DAG.getNode(Opc, dl, TrueVal.getValueType(), TrueVal, FalseVal, 2129 DAG.getConstant(SPCC, MVT::i32), CompareFlag); 2130} 2131 2132static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG, 2133 const SparcTargetLowering &TLI) { 2134 MachineFunction &MF = DAG.getMachineFunction(); 2135 SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>(); 2136 2137 // Need frame address to find the address of VarArgsFrameIndex. 2138 MF.getFrameInfo()->setFrameAddressIsTaken(true); 2139 2140 // vastart just stores the address of the VarArgsFrameIndex slot into the 2141 // memory location argument. 2142 SDLoc DL(Op); 2143 SDValue Offset = 2144 DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), 2145 DAG.getRegister(SP::I6, TLI.getPointerTy()), 2146 DAG.getIntPtrConstant(FuncInfo->getVarArgsFrameOffset())); 2147 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 2148 return DAG.getStore(Op.getOperand(0), DL, Offset, Op.getOperand(1), 2149 MachinePointerInfo(SV), false, false, 0); 2150} 2151 2152static SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) { 2153 SDNode *Node = Op.getNode(); 2154 EVT VT = Node->getValueType(0); 2155 SDValue InChain = Node->getOperand(0); 2156 SDValue VAListPtr = Node->getOperand(1); 2157 EVT PtrVT = VAListPtr.getValueType(); 2158 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue(); 2159 SDLoc DL(Node); 2160 SDValue VAList = DAG.getLoad(PtrVT, DL, InChain, VAListPtr, 2161 MachinePointerInfo(SV), false, false, false, 0); 2162 // Increment the pointer, VAList, to the next vaarg. 2163 SDValue NextPtr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList, 2164 DAG.getIntPtrConstant(VT.getSizeInBits()/8)); 2165 // Store the incremented VAList to the legalized pointer. 2166 InChain = DAG.getStore(VAList.getValue(1), DL, NextPtr, 2167 VAListPtr, MachinePointerInfo(SV), false, false, 0); 2168 // Load the actual argument out of the pointer VAList. 2169 // We can't count on greater alignment than the word size. 2170 return DAG.getLoad(VT, DL, InChain, VAList, MachinePointerInfo(), 2171 false, false, false, 2172 std::min(PtrVT.getSizeInBits(), VT.getSizeInBits())/8); 2173} 2174 2175static SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) { 2176 SDValue Chain = Op.getOperand(0); // Legalize the chain. 2177 SDValue Size = Op.getOperand(1); // Legalize the size. 2178 SDLoc dl(Op); 2179 2180 unsigned SPReg = SP::O6; 2181 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, MVT::i32); 2182 SDValue NewSP = DAG.getNode(ISD::SUB, dl, MVT::i32, SP, Size); // Value 2183 Chain = DAG.getCopyToReg(SP.getValue(1), dl, SPReg, NewSP); // Output chain 2184 2185 // The resultant pointer is actually 16 words from the bottom of the stack, 2186 // to provide a register spill area. 2187 SDValue NewVal = DAG.getNode(ISD::ADD, dl, MVT::i32, NewSP, 2188 DAG.getConstant(96, MVT::i32)); 2189 SDValue Ops[2] = { NewVal, Chain }; 2190 return DAG.getMergeValues(Ops, 2, dl); 2191} 2192 2193 2194static SDValue getFLUSHW(SDValue Op, SelectionDAG &DAG) { 2195 SDLoc dl(Op); 2196 SDValue Chain = DAG.getNode(SPISD::FLUSHW, 2197 dl, MVT::Other, DAG.getEntryNode()); 2198 return Chain; 2199} 2200 2201static SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) { 2202 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 2203 MFI->setFrameAddressIsTaken(true); 2204 2205 EVT VT = Op.getValueType(); 2206 SDLoc dl(Op); 2207 unsigned FrameReg = SP::I6; 2208 2209 uint64_t depth = Op.getConstantOperandVal(0); 2210 2211 SDValue FrameAddr; 2212 if (depth == 0) 2213 FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT); 2214 else { 2215 // flush first to make sure the windowed registers' values are in stack 2216 SDValue Chain = getFLUSHW(Op, DAG); 2217 FrameAddr = DAG.getCopyFromReg(Chain, dl, FrameReg, VT); 2218 2219 for (uint64_t i = 0; i != depth; ++i) { 2220 SDValue Ptr = DAG.getNode(ISD::ADD, 2221 dl, MVT::i32, 2222 FrameAddr, DAG.getIntPtrConstant(56)); 2223 FrameAddr = DAG.getLoad(MVT::i32, dl, 2224 Chain, 2225 Ptr, 2226 MachinePointerInfo(), false, false, false, 0); 2227 } 2228 } 2229 return FrameAddr; 2230} 2231 2232static SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG, 2233 const SparcTargetLowering &TLI) { 2234 MachineFunction &MF = DAG.getMachineFunction(); 2235 MachineFrameInfo *MFI = MF.getFrameInfo(); 2236 MFI->setReturnAddressIsTaken(true); 2237 2238 EVT VT = Op.getValueType(); 2239 SDLoc dl(Op); 2240 uint64_t depth = Op.getConstantOperandVal(0); 2241 2242 SDValue RetAddr; 2243 if (depth == 0) { 2244 unsigned RetReg = MF.addLiveIn(SP::I7, 2245 TLI.getRegClassFor(TLI.getPointerTy())); 2246 RetAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, RetReg, VT); 2247 } else { 2248 // Need frame address to find return address of the caller. 2249 MFI->setFrameAddressIsTaken(true); 2250 2251 // flush first to make sure the windowed registers' values are in stack 2252 SDValue Chain = getFLUSHW(Op, DAG); 2253 RetAddr = DAG.getCopyFromReg(Chain, dl, SP::I6, VT); 2254 2255 for (uint64_t i = 0; i != depth; ++i) { 2256 SDValue Ptr = DAG.getNode(ISD::ADD, 2257 dl, MVT::i32, 2258 RetAddr, 2259 DAG.getIntPtrConstant((i == depth-1)?60:56)); 2260 RetAddr = DAG.getLoad(MVT::i32, dl, 2261 Chain, 2262 Ptr, 2263 MachinePointerInfo(), false, false, false, 0); 2264 } 2265 } 2266 return RetAddr; 2267} 2268 2269static SDValue LowerF64Op(SDValue Op, SelectionDAG &DAG, unsigned opcode) 2270{ 2271 SDLoc dl(Op); 2272 2273 assert(Op.getValueType() == MVT::f64 && "LowerF64Op called on non-double!"); 2274 assert(opcode == ISD::FNEG || opcode == ISD::FABS); 2275 2276 // Lower fneg/fabs on f64 to fneg/fabs on f32. 2277 // fneg f64 => fneg f32:sub_even, fmov f32:sub_odd. 2278 // fabs f64 => fabs f32:sub_even, fmov f32:sub_odd. 2279 2280 SDValue SrcReg64 = Op.getOperand(0); 2281 SDValue Hi32 = DAG.getTargetExtractSubreg(SP::sub_even, dl, MVT::f32, 2282 SrcReg64); 2283 SDValue Lo32 = DAG.getTargetExtractSubreg(SP::sub_odd, dl, MVT::f32, 2284 SrcReg64); 2285 2286 Hi32 = DAG.getNode(opcode, dl, MVT::f32, Hi32); 2287 2288 SDValue DstReg64 = SDValue(DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, 2289 dl, MVT::f64), 0); 2290 DstReg64 = DAG.getTargetInsertSubreg(SP::sub_even, dl, MVT::f64, 2291 DstReg64, Hi32); 2292 DstReg64 = DAG.getTargetInsertSubreg(SP::sub_odd, dl, MVT::f64, 2293 DstReg64, Lo32); 2294 return DstReg64; 2295} 2296 2297// Lower a f128 load into two f64 loads. 2298static SDValue LowerF128Load(SDValue Op, SelectionDAG &DAG) 2299{ 2300 SDLoc dl(Op); 2301 LoadSDNode *LdNode = dyn_cast<LoadSDNode>(Op.getNode()); 2302 assert(LdNode && LdNode->getOffset().getOpcode() == ISD::UNDEF 2303 && "Unexpected node type"); 2304 2305 unsigned alignment = LdNode->getAlignment(); 2306 if (alignment > 8) 2307 alignment = 8; 2308 2309 SDValue Hi64 = DAG.getLoad(MVT::f64, 2310 dl, 2311 LdNode->getChain(), 2312 LdNode->getBasePtr(), 2313 LdNode->getPointerInfo(), 2314 false, false, false, alignment); 2315 EVT addrVT = LdNode->getBasePtr().getValueType(); 2316 SDValue LoPtr = DAG.getNode(ISD::ADD, dl, addrVT, 2317 LdNode->getBasePtr(), 2318 DAG.getConstant(8, addrVT)); 2319 SDValue Lo64 = DAG.getLoad(MVT::f64, 2320 dl, 2321 LdNode->getChain(), 2322 LoPtr, 2323 LdNode->getPointerInfo(), 2324 false, false, false, alignment); 2325 2326 SDValue SubRegEven = DAG.getTargetConstant(SP::sub_even64, MVT::i32); 2327 SDValue SubRegOdd = DAG.getTargetConstant(SP::sub_odd64, MVT::i32); 2328 2329 SDNode *InFP128 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, 2330 dl, MVT::f128); 2331 InFP128 = DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, 2332 MVT::f128, 2333 SDValue(InFP128, 0), 2334 Hi64, 2335 SubRegEven); 2336 InFP128 = DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, 2337 MVT::f128, 2338 SDValue(InFP128, 0), 2339 Lo64, 2340 SubRegOdd); 2341 SDValue OutChains[2] = { SDValue(Hi64.getNode(), 1), 2342 SDValue(Lo64.getNode(), 1) }; 2343 SDValue OutChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 2344 &OutChains[0], 2); 2345 SDValue Ops[2] = {SDValue(InFP128,0), OutChain}; 2346 return DAG.getMergeValues(Ops, 2, dl); 2347} 2348 2349// Lower a f128 store into two f64 stores. 2350static SDValue LowerF128Store(SDValue Op, SelectionDAG &DAG) { 2351 SDLoc dl(Op); 2352 StoreSDNode *StNode = dyn_cast<StoreSDNode>(Op.getNode()); 2353 assert(StNode && StNode->getOffset().getOpcode() == ISD::UNDEF 2354 && "Unexpected node type"); 2355 SDValue SubRegEven = DAG.getTargetConstant(SP::sub_even64, MVT::i32); 2356 SDValue SubRegOdd = DAG.getTargetConstant(SP::sub_odd64, MVT::i32); 2357 2358 SDNode *Hi64 = DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, 2359 dl, 2360 MVT::f64, 2361 StNode->getValue(), 2362 SubRegEven); 2363 SDNode *Lo64 = DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, 2364 dl, 2365 MVT::f64, 2366 StNode->getValue(), 2367 SubRegOdd); 2368 2369 unsigned alignment = StNode->getAlignment(); 2370 if (alignment > 8) 2371 alignment = 8; 2372 2373 SDValue OutChains[2]; 2374 OutChains[0] = DAG.getStore(StNode->getChain(), 2375 dl, 2376 SDValue(Hi64, 0), 2377 StNode->getBasePtr(), 2378 MachinePointerInfo(), 2379 false, false, alignment); 2380 EVT addrVT = StNode->getBasePtr().getValueType(); 2381 SDValue LoPtr = DAG.getNode(ISD::ADD, dl, addrVT, 2382 StNode->getBasePtr(), 2383 DAG.getConstant(8, addrVT)); 2384 OutChains[1] = DAG.getStore(StNode->getChain(), 2385 dl, 2386 SDValue(Lo64, 0), 2387 LoPtr, 2388 MachinePointerInfo(), 2389 false, false, alignment); 2390 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 2391 &OutChains[0], 2); 2392} 2393 2394static SDValue LowerFNEG(SDValue Op, SelectionDAG &DAG, 2395 const SparcTargetLowering &TLI, 2396 bool is64Bit) { 2397 if (Op.getValueType() == MVT::f64) 2398 return LowerF64Op(Op, DAG, ISD::FNEG); 2399 if (Op.getValueType() == MVT::f128) 2400 return TLI.LowerF128Op(Op, DAG, ((is64Bit) ? "_Qp_neg" : "_Q_neg"), 1); 2401 return Op; 2402} 2403 2404static SDValue LowerFABS(SDValue Op, SelectionDAG &DAG, bool isV9) { 2405 if (Op.getValueType() == MVT::f64) 2406 return LowerF64Op(Op, DAG, ISD::FABS); 2407 if (Op.getValueType() != MVT::f128) 2408 return Op; 2409 2410 // Lower fabs on f128 to fabs on f64 2411 // fabs f128 => fabs f64:sub_even64, fmov f64:sub_odd64 2412 2413 SDLoc dl(Op); 2414 SDValue SrcReg128 = Op.getOperand(0); 2415 SDValue Hi64 = DAG.getTargetExtractSubreg(SP::sub_even64, dl, MVT::f64, 2416 SrcReg128); 2417 SDValue Lo64 = DAG.getTargetExtractSubreg(SP::sub_odd64, dl, MVT::f64, 2418 SrcReg128); 2419 if (isV9) 2420 Hi64 = DAG.getNode(Op.getOpcode(), dl, MVT::f64, Hi64); 2421 else 2422 Hi64 = LowerF64Op(Hi64, DAG, ISD::FABS); 2423 2424 SDValue DstReg128 = SDValue(DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, 2425 dl, MVT::f128), 0); 2426 DstReg128 = DAG.getTargetInsertSubreg(SP::sub_even64, dl, MVT::f128, 2427 DstReg128, Hi64); 2428 DstReg128 = DAG.getTargetInsertSubreg(SP::sub_odd64, dl, MVT::f128, 2429 DstReg128, Lo64); 2430 return DstReg128; 2431} 2432 2433static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) { 2434 2435 if (Op.getValueType() != MVT::i64) 2436 return Op; 2437 2438 SDLoc dl(Op); 2439 SDValue Src1 = Op.getOperand(0); 2440 SDValue Src1Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Src1); 2441 SDValue Src1Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Src1, 2442 DAG.getConstant(32, MVT::i64)); 2443 Src1Hi = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Src1Hi); 2444 2445 SDValue Src2 = Op.getOperand(1); 2446 SDValue Src2Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Src2); 2447 SDValue Src2Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Src2, 2448 DAG.getConstant(32, MVT::i64)); 2449 Src2Hi = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Src2Hi); 2450 2451 2452 bool hasChain = false; 2453 unsigned hiOpc = Op.getOpcode(); 2454 switch (Op.getOpcode()) { 2455 default: llvm_unreachable("Invalid opcode"); 2456 case ISD::ADDC: hiOpc = ISD::ADDE; break; 2457 case ISD::ADDE: hasChain = true; break; 2458 case ISD::SUBC: hiOpc = ISD::SUBE; break; 2459 case ISD::SUBE: hasChain = true; break; 2460 } 2461 SDValue Lo; 2462 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Glue); 2463 if (hasChain) { 2464 Lo = DAG.getNode(Op.getOpcode(), dl, VTs, Src1Lo, Src2Lo, 2465 Op.getOperand(2)); 2466 } else { 2467 Lo = DAG.getNode(Op.getOpcode(), dl, VTs, Src1Lo, Src2Lo); 2468 } 2469 SDValue Hi = DAG.getNode(hiOpc, dl, VTs, Src1Hi, Src2Hi, Lo.getValue(1)); 2470 SDValue Carry = Hi.getValue(1); 2471 2472 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Lo); 2473 Hi = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Hi); 2474 Hi = DAG.getNode(ISD::SHL, dl, MVT::i64, Hi, 2475 DAG.getConstant(32, MVT::i64)); 2476 2477 SDValue Dst = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, Lo); 2478 SDValue Ops[2] = { Dst, Carry }; 2479 return DAG.getMergeValues(Ops, 2, dl); 2480} 2481 2482SDValue SparcTargetLowering:: 2483LowerOperation(SDValue Op, SelectionDAG &DAG) const { 2484 2485 bool hasHardQuad = Subtarget->hasHardQuad(); 2486 bool is64Bit = Subtarget->is64Bit(); 2487 bool isV9 = Subtarget->isV9(); 2488 2489 switch (Op.getOpcode()) { 2490 default: llvm_unreachable("Should not custom lower this!"); 2491 2492 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG, *this); 2493 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); 2494 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); 2495 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG); 2496 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); 2497 case ISD::ConstantPool: return LowerConstantPool(Op, DAG); 2498 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG, *this, 2499 hasHardQuad); 2500 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG, *this, 2501 hasHardQuad); 2502 case ISD::BR_CC: return LowerBR_CC(Op, DAG, *this, 2503 hasHardQuad); 2504 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, *this, 2505 hasHardQuad); 2506 case ISD::VASTART: return LowerVASTART(Op, DAG, *this); 2507 case ISD::VAARG: return LowerVAARG(Op, DAG); 2508 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG); 2509 2510 case ISD::LOAD: return LowerF128Load(Op, DAG); 2511 case ISD::STORE: return LowerF128Store(Op, DAG); 2512 case ISD::FADD: return LowerF128Op(Op, DAG, 2513 getLibcallName(RTLIB::ADD_F128), 2); 2514 case ISD::FSUB: return LowerF128Op(Op, DAG, 2515 getLibcallName(RTLIB::SUB_F128), 2); 2516 case ISD::FMUL: return LowerF128Op(Op, DAG, 2517 getLibcallName(RTLIB::MUL_F128), 2); 2518 case ISD::FDIV: return LowerF128Op(Op, DAG, 2519 getLibcallName(RTLIB::DIV_F128), 2); 2520 case ISD::FSQRT: return LowerF128Op(Op, DAG, 2521 getLibcallName(RTLIB::SQRT_F128),1); 2522 case ISD::FNEG: return LowerFNEG(Op, DAG, *this, is64Bit); 2523 case ISD::FABS: return LowerFABS(Op, DAG, isV9); 2524 case ISD::FP_EXTEND: return LowerF128_FPEXTEND(Op, DAG, *this); 2525 case ISD::FP_ROUND: return LowerF128_FPROUND(Op, DAG, *this); 2526 case ISD::ADDC: 2527 case ISD::ADDE: 2528 case ISD::SUBC: 2529 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG); 2530 } 2531} 2532 2533MachineBasicBlock * 2534SparcTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 2535 MachineBasicBlock *BB) const { 2536 const TargetInstrInfo &TII = *getTargetMachine().getInstrInfo(); 2537 unsigned BROpcode; 2538 unsigned CC; 2539 DebugLoc dl = MI->getDebugLoc(); 2540 // Figure out the conditional branch opcode to use for this select_cc. 2541 switch (MI->getOpcode()) { 2542 default: llvm_unreachable("Unknown SELECT_CC!"); 2543 case SP::SELECT_CC_Int_ICC: 2544 case SP::SELECT_CC_FP_ICC: 2545 case SP::SELECT_CC_DFP_ICC: 2546 case SP::SELECT_CC_QFP_ICC: 2547 BROpcode = SP::BCOND; 2548 break; 2549 case SP::SELECT_CC_Int_FCC: 2550 case SP::SELECT_CC_FP_FCC: 2551 case SP::SELECT_CC_DFP_FCC: 2552 case SP::SELECT_CC_QFP_FCC: 2553 BROpcode = SP::FBCOND; 2554 break; 2555 } 2556 2557 CC = (SPCC::CondCodes)MI->getOperand(3).getImm(); 2558 2559 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond 2560 // control-flow pattern. The incoming instruction knows the destination vreg 2561 // to set, the condition code register to branch on, the true/false values to 2562 // select between, and a branch opcode to use. 2563 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 2564 MachineFunction::iterator It = BB; 2565 ++It; 2566 2567 // thisMBB: 2568 // ... 2569 // TrueVal = ... 2570 // [f]bCC copy1MBB 2571 // fallthrough --> copy0MBB 2572 MachineBasicBlock *thisMBB = BB; 2573 MachineFunction *F = BB->getParent(); 2574 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 2575 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); 2576 F->insert(It, copy0MBB); 2577 F->insert(It, sinkMBB); 2578 2579 // Transfer the remainder of BB and its successor edges to sinkMBB. 2580 sinkMBB->splice(sinkMBB->begin(), BB, 2581 llvm::next(MachineBasicBlock::iterator(MI)), 2582 BB->end()); 2583 sinkMBB->transferSuccessorsAndUpdatePHIs(BB); 2584 2585 // Add the true and fallthrough blocks as its successors. 2586 BB->addSuccessor(copy0MBB); 2587 BB->addSuccessor(sinkMBB); 2588 2589 BuildMI(BB, dl, TII.get(BROpcode)).addMBB(sinkMBB).addImm(CC); 2590 2591 // copy0MBB: 2592 // %FalseValue = ... 2593 // # fallthrough to sinkMBB 2594 BB = copy0MBB; 2595 2596 // Update machine-CFG edges 2597 BB->addSuccessor(sinkMBB); 2598 2599 // sinkMBB: 2600 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] 2601 // ... 2602 BB = sinkMBB; 2603 BuildMI(*BB, BB->begin(), dl, TII.get(SP::PHI), MI->getOperand(0).getReg()) 2604 .addReg(MI->getOperand(2).getReg()).addMBB(copy0MBB) 2605 .addReg(MI->getOperand(1).getReg()).addMBB(thisMBB); 2606 2607 MI->eraseFromParent(); // The pseudo instruction is gone now. 2608 return BB; 2609} 2610 2611//===----------------------------------------------------------------------===// 2612// Sparc Inline Assembly Support 2613//===----------------------------------------------------------------------===// 2614 2615/// getConstraintType - Given a constraint letter, return the type of 2616/// constraint it is for this target. 2617SparcTargetLowering::ConstraintType 2618SparcTargetLowering::getConstraintType(const std::string &Constraint) const { 2619 if (Constraint.size() == 1) { 2620 switch (Constraint[0]) { 2621 default: break; 2622 case 'r': return C_RegisterClass; 2623 } 2624 } 2625 2626 return TargetLowering::getConstraintType(Constraint); 2627} 2628 2629std::pair<unsigned, const TargetRegisterClass*> 2630SparcTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint, 2631 MVT VT) const { 2632 if (Constraint.size() == 1) { 2633 switch (Constraint[0]) { 2634 case 'r': 2635 return std::make_pair(0U, &SP::IntRegsRegClass); 2636 } 2637 } 2638 2639 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 2640} 2641 2642bool 2643SparcTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 2644 // The Sparc target isn't yet aware of offsets. 2645 return false; 2646} 2647