SystemZISelLowering.cpp revision 299fdd814f4c2850d44387d24c440980c5377d3e
1//===-- SystemZISelLowering.cpp - SystemZ DAG lowering implementation -----===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements the SystemZTargetLowering class. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "systemz-lower" 15 16#include "SystemZISelLowering.h" 17#include "SystemZCallingConv.h" 18#include "SystemZConstantPoolValue.h" 19#include "SystemZMachineFunctionInfo.h" 20#include "SystemZTargetMachine.h" 21#include "llvm/CodeGen/CallingConvLower.h" 22#include "llvm/CodeGen/MachineInstrBuilder.h" 23#include "llvm/CodeGen/MachineRegisterInfo.h" 24#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" 25 26using namespace llvm; 27 28// Classify VT as either 32 or 64 bit. 29static bool is32Bit(EVT VT) { 30 switch (VT.getSimpleVT().SimpleTy) { 31 case MVT::i32: 32 return true; 33 case MVT::i64: 34 return false; 35 default: 36 llvm_unreachable("Unsupported type"); 37 } 38} 39 40// Return a version of MachineOperand that can be safely used before the 41// final use. 42static MachineOperand earlyUseOperand(MachineOperand Op) { 43 if (Op.isReg()) 44 Op.setIsKill(false); 45 return Op; 46} 47 48SystemZTargetLowering::SystemZTargetLowering(SystemZTargetMachine &tm) 49 : TargetLowering(tm, new TargetLoweringObjectFileELF()), 50 Subtarget(*tm.getSubtargetImpl()), TM(tm) { 51 MVT PtrVT = getPointerTy(); 52 53 // Set up the register classes. 54 addRegisterClass(MVT::i32, &SystemZ::GR32BitRegClass); 55 addRegisterClass(MVT::i64, &SystemZ::GR64BitRegClass); 56 addRegisterClass(MVT::f32, &SystemZ::FP32BitRegClass); 57 addRegisterClass(MVT::f64, &SystemZ::FP64BitRegClass); 58 addRegisterClass(MVT::f128, &SystemZ::FP128BitRegClass); 59 60 // Compute derived properties from the register classes 61 computeRegisterProperties(); 62 63 // Set up special registers. 64 setExceptionPointerRegister(SystemZ::R6D); 65 setExceptionSelectorRegister(SystemZ::R7D); 66 setStackPointerRegisterToSaveRestore(SystemZ::R15D); 67 68 // TODO: It may be better to default to latency-oriented scheduling, however 69 // LLVM's current latency-oriented scheduler can't handle physreg definitions 70 // such as SystemZ has with CC, so set this to the register-pressure 71 // scheduler, because it can. 72 setSchedulingPreference(Sched::RegPressure); 73 74 setBooleanContents(ZeroOrOneBooleanContent); 75 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct? 76 77 // Instructions are strings of 2-byte aligned 2-byte values. 78 setMinFunctionAlignment(2); 79 80 // Handle operations that are handled in a similar way for all types. 81 for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE; 82 I <= MVT::LAST_FP_VALUETYPE; 83 ++I) { 84 MVT VT = MVT::SimpleValueType(I); 85 if (isTypeLegal(VT)) { 86 // Expand SETCC(X, Y, COND) into SELECT_CC(X, Y, 1, 0, COND). 87 setOperationAction(ISD::SETCC, VT, Expand); 88 89 // Expand SELECT(C, A, B) into SELECT_CC(X, 0, A, B, NE). 90 setOperationAction(ISD::SELECT, VT, Expand); 91 92 // Lower SELECT_CC and BR_CC into separate comparisons and branches. 93 setOperationAction(ISD::SELECT_CC, VT, Custom); 94 setOperationAction(ISD::BR_CC, VT, Custom); 95 } 96 } 97 98 // Expand jump table branches as address arithmetic followed by an 99 // indirect jump. 100 setOperationAction(ISD::BR_JT, MVT::Other, Expand); 101 102 // Expand BRCOND into a BR_CC (see above). 103 setOperationAction(ISD::BRCOND, MVT::Other, Expand); 104 105 // Handle integer types. 106 for (unsigned I = MVT::FIRST_INTEGER_VALUETYPE; 107 I <= MVT::LAST_INTEGER_VALUETYPE; 108 ++I) { 109 MVT VT = MVT::SimpleValueType(I); 110 if (isTypeLegal(VT)) { 111 // Expand individual DIV and REMs into DIVREMs. 112 setOperationAction(ISD::SDIV, VT, Expand); 113 setOperationAction(ISD::UDIV, VT, Expand); 114 setOperationAction(ISD::SREM, VT, Expand); 115 setOperationAction(ISD::UREM, VT, Expand); 116 setOperationAction(ISD::SDIVREM, VT, Custom); 117 setOperationAction(ISD::UDIVREM, VT, Custom); 118 119 // Expand ATOMIC_LOAD and ATOMIC_STORE using ATOMIC_CMP_SWAP. 120 // FIXME: probably much too conservative. 121 setOperationAction(ISD::ATOMIC_LOAD, VT, Expand); 122 setOperationAction(ISD::ATOMIC_STORE, VT, Expand); 123 124 // No special instructions for these. 125 setOperationAction(ISD::CTPOP, VT, Expand); 126 setOperationAction(ISD::CTTZ, VT, Expand); 127 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 128 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 129 setOperationAction(ISD::ROTR, VT, Expand); 130 131 // Use *MUL_LOHI where possible instead of MULH*. 132 setOperationAction(ISD::MULHS, VT, Expand); 133 setOperationAction(ISD::MULHU, VT, Expand); 134 setOperationAction(ISD::SMUL_LOHI, VT, Custom); 135 setOperationAction(ISD::UMUL_LOHI, VT, Custom); 136 137 // We have instructions for signed but not unsigned FP conversion. 138 setOperationAction(ISD::FP_TO_UINT, VT, Expand); 139 } 140 } 141 142 // Type legalization will convert 8- and 16-bit atomic operations into 143 // forms that operate on i32s (but still keeping the original memory VT). 144 // Lower them into full i32 operations. 145 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Custom); 146 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Custom); 147 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom); 148 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Custom); 149 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Custom); 150 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Custom); 151 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Custom); 152 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Custom); 153 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Custom); 154 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Custom); 155 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Custom); 156 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom); 157 158 // We have instructions for signed but not unsigned FP conversion. 159 // Handle unsigned 32-bit types as signed 64-bit types. 160 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Promote); 161 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand); 162 163 // We have native support for a 64-bit CTLZ, via FLOGR. 164 setOperationAction(ISD::CTLZ, MVT::i32, Promote); 165 setOperationAction(ISD::CTLZ, MVT::i64, Legal); 166 167 // Give LowerOperation the chance to replace 64-bit ORs with subregs. 168 setOperationAction(ISD::OR, MVT::i64, Custom); 169 170 // FIXME: Can we support these natively? 171 setOperationAction(ISD::SRL_PARTS, MVT::i64, Expand); 172 setOperationAction(ISD::SHL_PARTS, MVT::i64, Expand); 173 setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand); 174 175 // We have native instructions for i8, i16 and i32 extensions, but not i1. 176 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 177 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote); 178 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote); 179 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 180 181 // Handle the various types of symbolic address. 182 setOperationAction(ISD::ConstantPool, PtrVT, Custom); 183 setOperationAction(ISD::GlobalAddress, PtrVT, Custom); 184 setOperationAction(ISD::GlobalTLSAddress, PtrVT, Custom); 185 setOperationAction(ISD::BlockAddress, PtrVT, Custom); 186 setOperationAction(ISD::JumpTable, PtrVT, Custom); 187 188 // We need to handle dynamic allocations specially because of the 189 // 160-byte area at the bottom of the stack. 190 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom); 191 192 // Use custom expanders so that we can force the function to use 193 // a frame pointer. 194 setOperationAction(ISD::STACKSAVE, MVT::Other, Custom); 195 setOperationAction(ISD::STACKRESTORE, MVT::Other, Custom); 196 197 // Handle prefetches with PFD or PFDRL. 198 setOperationAction(ISD::PREFETCH, MVT::Other, Custom); 199 200 // Handle floating-point types. 201 for (unsigned I = MVT::FIRST_FP_VALUETYPE; 202 I <= MVT::LAST_FP_VALUETYPE; 203 ++I) { 204 MVT VT = MVT::SimpleValueType(I); 205 if (isTypeLegal(VT)) { 206 // We can use FI for FRINT. 207 setOperationAction(ISD::FRINT, VT, Legal); 208 209 // We can use the extended form of FI for other rounding operations. 210 if (Subtarget.hasFPExtension()) { 211 setOperationAction(ISD::FNEARBYINT, VT, Legal); 212 setOperationAction(ISD::FFLOOR, VT, Legal); 213 setOperationAction(ISD::FCEIL, VT, Legal); 214 setOperationAction(ISD::FTRUNC, VT, Legal); 215 setOperationAction(ISD::FROUND, VT, Legal); 216 } 217 218 // No special instructions for these. 219 setOperationAction(ISD::FSIN, VT, Expand); 220 setOperationAction(ISD::FCOS, VT, Expand); 221 setOperationAction(ISD::FREM, VT, Expand); 222 } 223 } 224 225 // We have fused multiply-addition for f32 and f64 but not f128. 226 setOperationAction(ISD::FMA, MVT::f32, Legal); 227 setOperationAction(ISD::FMA, MVT::f64, Legal); 228 setOperationAction(ISD::FMA, MVT::f128, Expand); 229 230 // Needed so that we don't try to implement f128 constant loads using 231 // a load-and-extend of a f80 constant (in cases where the constant 232 // would fit in an f80). 233 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand); 234 235 // Floating-point truncation and stores need to be done separately. 236 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 237 setTruncStoreAction(MVT::f128, MVT::f32, Expand); 238 setTruncStoreAction(MVT::f128, MVT::f64, Expand); 239 240 // We have 64-bit FPR<->GPR moves, but need special handling for 241 // 32-bit forms. 242 setOperationAction(ISD::BITCAST, MVT::i32, Custom); 243 setOperationAction(ISD::BITCAST, MVT::f32, Custom); 244 245 // VASTART and VACOPY need to deal with the SystemZ-specific varargs 246 // structure, but VAEND is a no-op. 247 setOperationAction(ISD::VASTART, MVT::Other, Custom); 248 setOperationAction(ISD::VACOPY, MVT::Other, Custom); 249 setOperationAction(ISD::VAEND, MVT::Other, Expand); 250 251 // We want to use MVC in preference to even a single load/store pair. 252 MaxStoresPerMemcpy = 0; 253 MaxStoresPerMemcpyOptSize = 0; 254 255 // The main memset sequence is a byte store followed by an MVC. 256 // Two STC or MV..I stores win over that, but the kind of fused stores 257 // generated by target-independent code don't when the byte value is 258 // variable. E.g. "STC <reg>;MHI <reg>,257;STH <reg>" is not better 259 // than "STC;MVC". Handle the choice in target-specific code instead. 260 MaxStoresPerMemset = 0; 261 MaxStoresPerMemsetOptSize = 0; 262} 263 264bool 265SystemZTargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const { 266 VT = VT.getScalarType(); 267 268 if (!VT.isSimple()) 269 return false; 270 271 switch (VT.getSimpleVT().SimpleTy) { 272 case MVT::f32: 273 case MVT::f64: 274 return true; 275 case MVT::f128: 276 return false; 277 default: 278 break; 279 } 280 281 return false; 282} 283 284bool SystemZTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const { 285 // We can load zero using LZ?R and negative zero using LZ?R;LC?BR. 286 return Imm.isZero() || Imm.isNegZero(); 287} 288 289bool SystemZTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, 290 bool *Fast) const { 291 // Unaligned accesses should never be slower than the expanded version. 292 // We check specifically for aligned accesses in the few cases where 293 // they are required. 294 if (Fast) 295 *Fast = true; 296 return true; 297} 298 299bool SystemZTargetLowering::isLegalAddressingMode(const AddrMode &AM, 300 Type *Ty) const { 301 // Punt on globals for now, although they can be used in limited 302 // RELATIVE LONG cases. 303 if (AM.BaseGV) 304 return false; 305 306 // Require a 20-bit signed offset. 307 if (!isInt<20>(AM.BaseOffs)) 308 return false; 309 310 // Indexing is OK but no scale factor can be applied. 311 return AM.Scale == 0 || AM.Scale == 1; 312} 313 314bool SystemZTargetLowering::isTruncateFree(Type *FromType, Type *ToType) const { 315 if (!FromType->isIntegerTy() || !ToType->isIntegerTy()) 316 return false; 317 unsigned FromBits = FromType->getPrimitiveSizeInBits(); 318 unsigned ToBits = ToType->getPrimitiveSizeInBits(); 319 return FromBits > ToBits; 320} 321 322bool SystemZTargetLowering::isTruncateFree(EVT FromVT, EVT ToVT) const { 323 if (!FromVT.isInteger() || !ToVT.isInteger()) 324 return false; 325 unsigned FromBits = FromVT.getSizeInBits(); 326 unsigned ToBits = ToVT.getSizeInBits(); 327 return FromBits > ToBits; 328} 329 330//===----------------------------------------------------------------------===// 331// Inline asm support 332//===----------------------------------------------------------------------===// 333 334TargetLowering::ConstraintType 335SystemZTargetLowering::getConstraintType(const std::string &Constraint) const { 336 if (Constraint.size() == 1) { 337 switch (Constraint[0]) { 338 case 'a': // Address register 339 case 'd': // Data register (equivalent to 'r') 340 case 'f': // Floating-point register 341 case 'r': // General-purpose register 342 return C_RegisterClass; 343 344 case 'Q': // Memory with base and unsigned 12-bit displacement 345 case 'R': // Likewise, plus an index 346 case 'S': // Memory with base and signed 20-bit displacement 347 case 'T': // Likewise, plus an index 348 case 'm': // Equivalent to 'T'. 349 return C_Memory; 350 351 case 'I': // Unsigned 8-bit constant 352 case 'J': // Unsigned 12-bit constant 353 case 'K': // Signed 16-bit constant 354 case 'L': // Signed 20-bit displacement (on all targets we support) 355 case 'M': // 0x7fffffff 356 return C_Other; 357 358 default: 359 break; 360 } 361 } 362 return TargetLowering::getConstraintType(Constraint); 363} 364 365TargetLowering::ConstraintWeight SystemZTargetLowering:: 366getSingleConstraintMatchWeight(AsmOperandInfo &info, 367 const char *constraint) const { 368 ConstraintWeight weight = CW_Invalid; 369 Value *CallOperandVal = info.CallOperandVal; 370 // If we don't have a value, we can't do a match, 371 // but allow it at the lowest weight. 372 if (CallOperandVal == NULL) 373 return CW_Default; 374 Type *type = CallOperandVal->getType(); 375 // Look at the constraint type. 376 switch (*constraint) { 377 default: 378 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); 379 break; 380 381 case 'a': // Address register 382 case 'd': // Data register (equivalent to 'r') 383 case 'r': // General-purpose register 384 if (CallOperandVal->getType()->isIntegerTy()) 385 weight = CW_Register; 386 break; 387 388 case 'f': // Floating-point register 389 if (type->isFloatingPointTy()) 390 weight = CW_Register; 391 break; 392 393 case 'I': // Unsigned 8-bit constant 394 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) 395 if (isUInt<8>(C->getZExtValue())) 396 weight = CW_Constant; 397 break; 398 399 case 'J': // Unsigned 12-bit constant 400 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) 401 if (isUInt<12>(C->getZExtValue())) 402 weight = CW_Constant; 403 break; 404 405 case 'K': // Signed 16-bit constant 406 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) 407 if (isInt<16>(C->getSExtValue())) 408 weight = CW_Constant; 409 break; 410 411 case 'L': // Signed 20-bit displacement (on all targets we support) 412 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) 413 if (isInt<20>(C->getSExtValue())) 414 weight = CW_Constant; 415 break; 416 417 case 'M': // 0x7fffffff 418 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) 419 if (C->getZExtValue() == 0x7fffffff) 420 weight = CW_Constant; 421 break; 422 } 423 return weight; 424} 425 426// Parse a "{tNNN}" register constraint for which the register type "t" 427// has already been verified. MC is the class associated with "t" and 428// Map maps 0-based register numbers to LLVM register numbers. 429static std::pair<unsigned, const TargetRegisterClass *> 430parseRegisterNumber(const std::string &Constraint, 431 const TargetRegisterClass *RC, const unsigned *Map) { 432 assert(*(Constraint.end()-1) == '}' && "Missing '}'"); 433 if (isdigit(Constraint[2])) { 434 std::string Suffix(Constraint.data() + 2, Constraint.size() - 2); 435 unsigned Index = atoi(Suffix.c_str()); 436 if (Index < 16 && Map[Index]) 437 return std::make_pair(Map[Index], RC); 438 } 439 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0)); 440} 441 442std::pair<unsigned, const TargetRegisterClass *> SystemZTargetLowering:: 443getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const { 444 if (Constraint.size() == 1) { 445 // GCC Constraint Letters 446 switch (Constraint[0]) { 447 default: break; 448 case 'd': // Data register (equivalent to 'r') 449 case 'r': // General-purpose register 450 if (VT == MVT::i64) 451 return std::make_pair(0U, &SystemZ::GR64BitRegClass); 452 else if (VT == MVT::i128) 453 return std::make_pair(0U, &SystemZ::GR128BitRegClass); 454 return std::make_pair(0U, &SystemZ::GR32BitRegClass); 455 456 case 'a': // Address register 457 if (VT == MVT::i64) 458 return std::make_pair(0U, &SystemZ::ADDR64BitRegClass); 459 else if (VT == MVT::i128) 460 return std::make_pair(0U, &SystemZ::ADDR128BitRegClass); 461 return std::make_pair(0U, &SystemZ::ADDR32BitRegClass); 462 463 case 'f': // Floating-point register 464 if (VT == MVT::f64) 465 return std::make_pair(0U, &SystemZ::FP64BitRegClass); 466 else if (VT == MVT::f128) 467 return std::make_pair(0U, &SystemZ::FP128BitRegClass); 468 return std::make_pair(0U, &SystemZ::FP32BitRegClass); 469 } 470 } 471 if (Constraint[0] == '{') { 472 // We need to override the default register parsing for GPRs and FPRs 473 // because the interpretation depends on VT. The internal names of 474 // the registers are also different from the external names 475 // (F0D and F0S instead of F0, etc.). 476 if (Constraint[1] == 'r') { 477 if (VT == MVT::i32) 478 return parseRegisterNumber(Constraint, &SystemZ::GR32BitRegClass, 479 SystemZMC::GR32Regs); 480 if (VT == MVT::i128) 481 return parseRegisterNumber(Constraint, &SystemZ::GR128BitRegClass, 482 SystemZMC::GR128Regs); 483 return parseRegisterNumber(Constraint, &SystemZ::GR64BitRegClass, 484 SystemZMC::GR64Regs); 485 } 486 if (Constraint[1] == 'f') { 487 if (VT == MVT::f32) 488 return parseRegisterNumber(Constraint, &SystemZ::FP32BitRegClass, 489 SystemZMC::FP32Regs); 490 if (VT == MVT::f128) 491 return parseRegisterNumber(Constraint, &SystemZ::FP128BitRegClass, 492 SystemZMC::FP128Regs); 493 return parseRegisterNumber(Constraint, &SystemZ::FP64BitRegClass, 494 SystemZMC::FP64Regs); 495 } 496 } 497 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 498} 499 500void SystemZTargetLowering:: 501LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint, 502 std::vector<SDValue> &Ops, 503 SelectionDAG &DAG) const { 504 // Only support length 1 constraints for now. 505 if (Constraint.length() == 1) { 506 switch (Constraint[0]) { 507 case 'I': // Unsigned 8-bit constant 508 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) 509 if (isUInt<8>(C->getZExtValue())) 510 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(), 511 Op.getValueType())); 512 return; 513 514 case 'J': // Unsigned 12-bit constant 515 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) 516 if (isUInt<12>(C->getZExtValue())) 517 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(), 518 Op.getValueType())); 519 return; 520 521 case 'K': // Signed 16-bit constant 522 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) 523 if (isInt<16>(C->getSExtValue())) 524 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(), 525 Op.getValueType())); 526 return; 527 528 case 'L': // Signed 20-bit displacement (on all targets we support) 529 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) 530 if (isInt<20>(C->getSExtValue())) 531 Ops.push_back(DAG.getTargetConstant(C->getSExtValue(), 532 Op.getValueType())); 533 return; 534 535 case 'M': // 0x7fffffff 536 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) 537 if (C->getZExtValue() == 0x7fffffff) 538 Ops.push_back(DAG.getTargetConstant(C->getZExtValue(), 539 Op.getValueType())); 540 return; 541 } 542 } 543 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); 544} 545 546//===----------------------------------------------------------------------===// 547// Calling conventions 548//===----------------------------------------------------------------------===// 549 550#include "SystemZGenCallingConv.inc" 551 552bool SystemZTargetLowering::allowTruncateForTailCall(Type *FromType, 553 Type *ToType) const { 554 return isTruncateFree(FromType, ToType); 555} 556 557bool SystemZTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const { 558 if (!CI->isTailCall()) 559 return false; 560 return true; 561} 562 563// Value is a value that has been passed to us in the location described by VA 564// (and so has type VA.getLocVT()). Convert Value to VA.getValVT(), chaining 565// any loads onto Chain. 566static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDLoc DL, 567 CCValAssign &VA, SDValue Chain, 568 SDValue Value) { 569 // If the argument has been promoted from a smaller type, insert an 570 // assertion to capture this. 571 if (VA.getLocInfo() == CCValAssign::SExt) 572 Value = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Value, 573 DAG.getValueType(VA.getValVT())); 574 else if (VA.getLocInfo() == CCValAssign::ZExt) 575 Value = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Value, 576 DAG.getValueType(VA.getValVT())); 577 578 if (VA.isExtInLoc()) 579 Value = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Value); 580 else if (VA.getLocInfo() == CCValAssign::Indirect) 581 Value = DAG.getLoad(VA.getValVT(), DL, Chain, Value, 582 MachinePointerInfo(), false, false, false, 0); 583 else 584 assert(VA.getLocInfo() == CCValAssign::Full && "Unsupported getLocInfo"); 585 return Value; 586} 587 588// Value is a value of type VA.getValVT() that we need to copy into 589// the location described by VA. Return a copy of Value converted to 590// VA.getValVT(). The caller is responsible for handling indirect values. 591static SDValue convertValVTToLocVT(SelectionDAG &DAG, SDLoc DL, 592 CCValAssign &VA, SDValue Value) { 593 switch (VA.getLocInfo()) { 594 case CCValAssign::SExt: 595 return DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Value); 596 case CCValAssign::ZExt: 597 return DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Value); 598 case CCValAssign::AExt: 599 return DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Value); 600 case CCValAssign::Full: 601 return Value; 602 default: 603 llvm_unreachable("Unhandled getLocInfo()"); 604 } 605} 606 607SDValue SystemZTargetLowering:: 608LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, 609 const SmallVectorImpl<ISD::InputArg> &Ins, 610 SDLoc DL, SelectionDAG &DAG, 611 SmallVectorImpl<SDValue> &InVals) const { 612 MachineFunction &MF = DAG.getMachineFunction(); 613 MachineFrameInfo *MFI = MF.getFrameInfo(); 614 MachineRegisterInfo &MRI = MF.getRegInfo(); 615 SystemZMachineFunctionInfo *FuncInfo = 616 MF.getInfo<SystemZMachineFunctionInfo>(); 617 const SystemZFrameLowering *TFL = 618 static_cast<const SystemZFrameLowering *>(TM.getFrameLowering()); 619 620 // Assign locations to all of the incoming arguments. 621 SmallVector<CCValAssign, 16> ArgLocs; 622 CCState CCInfo(CallConv, IsVarArg, MF, TM, ArgLocs, *DAG.getContext()); 623 CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ); 624 625 unsigned NumFixedGPRs = 0; 626 unsigned NumFixedFPRs = 0; 627 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) { 628 SDValue ArgValue; 629 CCValAssign &VA = ArgLocs[I]; 630 EVT LocVT = VA.getLocVT(); 631 if (VA.isRegLoc()) { 632 // Arguments passed in registers 633 const TargetRegisterClass *RC; 634 switch (LocVT.getSimpleVT().SimpleTy) { 635 default: 636 // Integers smaller than i64 should be promoted to i64. 637 llvm_unreachable("Unexpected argument type"); 638 case MVT::i32: 639 NumFixedGPRs += 1; 640 RC = &SystemZ::GR32BitRegClass; 641 break; 642 case MVT::i64: 643 NumFixedGPRs += 1; 644 RC = &SystemZ::GR64BitRegClass; 645 break; 646 case MVT::f32: 647 NumFixedFPRs += 1; 648 RC = &SystemZ::FP32BitRegClass; 649 break; 650 case MVT::f64: 651 NumFixedFPRs += 1; 652 RC = &SystemZ::FP64BitRegClass; 653 break; 654 } 655 656 unsigned VReg = MRI.createVirtualRegister(RC); 657 MRI.addLiveIn(VA.getLocReg(), VReg); 658 ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, LocVT); 659 } else { 660 assert(VA.isMemLoc() && "Argument not register or memory"); 661 662 // Create the frame index object for this incoming parameter. 663 int FI = MFI->CreateFixedObject(LocVT.getSizeInBits() / 8, 664 VA.getLocMemOffset(), true); 665 666 // Create the SelectionDAG nodes corresponding to a load 667 // from this parameter. Unpromoted ints and floats are 668 // passed as right-justified 8-byte values. 669 EVT PtrVT = getPointerTy(); 670 SDValue FIN = DAG.getFrameIndex(FI, PtrVT); 671 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32) 672 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4)); 673 ArgValue = DAG.getLoad(LocVT, DL, Chain, FIN, 674 MachinePointerInfo::getFixedStack(FI), 675 false, false, false, 0); 676 } 677 678 // Convert the value of the argument register into the value that's 679 // being passed. 680 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, ArgValue)); 681 } 682 683 if (IsVarArg) { 684 // Save the number of non-varargs registers for later use by va_start, etc. 685 FuncInfo->setVarArgsFirstGPR(NumFixedGPRs); 686 FuncInfo->setVarArgsFirstFPR(NumFixedFPRs); 687 688 // Likewise the address (in the form of a frame index) of where the 689 // first stack vararg would be. The 1-byte size here is arbitrary. 690 int64_t StackSize = CCInfo.getNextStackOffset(); 691 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize, true)); 692 693 // ...and a similar frame index for the caller-allocated save area 694 // that will be used to store the incoming registers. 695 int64_t RegSaveOffset = TFL->getOffsetOfLocalArea(); 696 unsigned RegSaveIndex = MFI->CreateFixedObject(1, RegSaveOffset, true); 697 FuncInfo->setRegSaveFrameIndex(RegSaveIndex); 698 699 // Store the FPR varargs in the reserved frame slots. (We store the 700 // GPRs as part of the prologue.) 701 if (NumFixedFPRs < SystemZ::NumArgFPRs) { 702 SDValue MemOps[SystemZ::NumArgFPRs]; 703 for (unsigned I = NumFixedFPRs; I < SystemZ::NumArgFPRs; ++I) { 704 unsigned Offset = TFL->getRegSpillOffset(SystemZ::ArgFPRs[I]); 705 int FI = MFI->CreateFixedObject(8, RegSaveOffset + Offset, true); 706 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 707 unsigned VReg = MF.addLiveIn(SystemZ::ArgFPRs[I], 708 &SystemZ::FP64BitRegClass); 709 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f64); 710 MemOps[I] = DAG.getStore(ArgValue.getValue(1), DL, ArgValue, FIN, 711 MachinePointerInfo::getFixedStack(FI), 712 false, false, 0); 713 714 } 715 // Join the stores, which are independent of one another. 716 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, 717 &MemOps[NumFixedFPRs], 718 SystemZ::NumArgFPRs - NumFixedFPRs); 719 } 720 } 721 722 return Chain; 723} 724 725static bool canUseSiblingCall(CCState ArgCCInfo, 726 SmallVectorImpl<CCValAssign> &ArgLocs) { 727 // Punt if there are any indirect or stack arguments, or if the call 728 // needs the call-saved argument register R6. 729 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) { 730 CCValAssign &VA = ArgLocs[I]; 731 if (VA.getLocInfo() == CCValAssign::Indirect) 732 return false; 733 if (!VA.isRegLoc()) 734 return false; 735 unsigned Reg = VA.getLocReg(); 736 if (Reg == SystemZ::R6W || Reg == SystemZ::R6D) 737 return false; 738 } 739 return true; 740} 741 742SDValue 743SystemZTargetLowering::LowerCall(CallLoweringInfo &CLI, 744 SmallVectorImpl<SDValue> &InVals) const { 745 SelectionDAG &DAG = CLI.DAG; 746 SDLoc &DL = CLI.DL; 747 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; 748 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; 749 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; 750 SDValue Chain = CLI.Chain; 751 SDValue Callee = CLI.Callee; 752 bool &IsTailCall = CLI.IsTailCall; 753 CallingConv::ID CallConv = CLI.CallConv; 754 bool IsVarArg = CLI.IsVarArg; 755 MachineFunction &MF = DAG.getMachineFunction(); 756 EVT PtrVT = getPointerTy(); 757 758 // Analyze the operands of the call, assigning locations to each operand. 759 SmallVector<CCValAssign, 16> ArgLocs; 760 CCState ArgCCInfo(CallConv, IsVarArg, MF, TM, ArgLocs, *DAG.getContext()); 761 ArgCCInfo.AnalyzeCallOperands(Outs, CC_SystemZ); 762 763 // We don't support GuaranteedTailCallOpt, only automatically-detected 764 // sibling calls. 765 if (IsTailCall && !canUseSiblingCall(ArgCCInfo, ArgLocs)) 766 IsTailCall = false; 767 768 // Get a count of how many bytes are to be pushed on the stack. 769 unsigned NumBytes = ArgCCInfo.getNextStackOffset(); 770 771 // Mark the start of the call. 772 if (!IsTailCall) 773 Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(NumBytes, PtrVT, true), 774 DL); 775 776 // Copy argument values to their designated locations. 777 SmallVector<std::pair<unsigned, SDValue>, 9> RegsToPass; 778 SmallVector<SDValue, 8> MemOpChains; 779 SDValue StackPtr; 780 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) { 781 CCValAssign &VA = ArgLocs[I]; 782 SDValue ArgValue = OutVals[I]; 783 784 if (VA.getLocInfo() == CCValAssign::Indirect) { 785 // Store the argument in a stack slot and pass its address. 786 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT()); 787 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex(); 788 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, SpillSlot, 789 MachinePointerInfo::getFixedStack(FI), 790 false, false, 0)); 791 ArgValue = SpillSlot; 792 } else 793 ArgValue = convertValVTToLocVT(DAG, DL, VA, ArgValue); 794 795 if (VA.isRegLoc()) 796 // Queue up the argument copies and emit them at the end. 797 RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue)); 798 else { 799 assert(VA.isMemLoc() && "Argument not register or memory"); 800 801 // Work out the address of the stack slot. Unpromoted ints and 802 // floats are passed as right-justified 8-byte values. 803 if (!StackPtr.getNode()) 804 StackPtr = DAG.getCopyFromReg(Chain, DL, SystemZ::R15D, PtrVT); 805 unsigned Offset = SystemZMC::CallFrameSize + VA.getLocMemOffset(); 806 if (VA.getLocVT() == MVT::i32 || VA.getLocVT() == MVT::f32) 807 Offset += 4; 808 SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, 809 DAG.getIntPtrConstant(Offset)); 810 811 // Emit the store. 812 MemOpChains.push_back(DAG.getStore(Chain, DL, ArgValue, Address, 813 MachinePointerInfo(), 814 false, false, 0)); 815 } 816 } 817 818 // Join the stores, which are independent of one another. 819 if (!MemOpChains.empty()) 820 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, 821 &MemOpChains[0], MemOpChains.size()); 822 823 // Accept direct calls by converting symbolic call addresses to the 824 // associated Target* opcodes. Force %r1 to be used for indirect 825 // tail calls. 826 SDValue Glue; 827 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 828 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, PtrVT); 829 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee); 830 } else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee)) { 831 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), PtrVT); 832 Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee); 833 } else if (IsTailCall) { 834 Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R1D, Callee, Glue); 835 Glue = Chain.getValue(1); 836 Callee = DAG.getRegister(SystemZ::R1D, Callee.getValueType()); 837 } 838 839 // Build a sequence of copy-to-reg nodes, chained and glued together. 840 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I) { 841 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[I].first, 842 RegsToPass[I].second, Glue); 843 Glue = Chain.getValue(1); 844 } 845 846 // The first call operand is the chain and the second is the target address. 847 SmallVector<SDValue, 8> Ops; 848 Ops.push_back(Chain); 849 Ops.push_back(Callee); 850 851 // Add argument registers to the end of the list so that they are 852 // known live into the call. 853 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I) 854 Ops.push_back(DAG.getRegister(RegsToPass[I].first, 855 RegsToPass[I].second.getValueType())); 856 857 // Glue the call to the argument copies, if any. 858 if (Glue.getNode()) 859 Ops.push_back(Glue); 860 861 // Emit the call. 862 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 863 if (IsTailCall) 864 return DAG.getNode(SystemZISD::SIBCALL, DL, NodeTys, &Ops[0], Ops.size()); 865 Chain = DAG.getNode(SystemZISD::CALL, DL, NodeTys, &Ops[0], Ops.size()); 866 Glue = Chain.getValue(1); 867 868 // Mark the end of the call, which is glued to the call itself. 869 Chain = DAG.getCALLSEQ_END(Chain, 870 DAG.getConstant(NumBytes, PtrVT, true), 871 DAG.getConstant(0, PtrVT, true), 872 Glue, DL); 873 Glue = Chain.getValue(1); 874 875 // Assign locations to each value returned by this call. 876 SmallVector<CCValAssign, 16> RetLocs; 877 CCState RetCCInfo(CallConv, IsVarArg, MF, TM, RetLocs, *DAG.getContext()); 878 RetCCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ); 879 880 // Copy all of the result registers out of their specified physreg. 881 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) { 882 CCValAssign &VA = RetLocs[I]; 883 884 // Copy the value out, gluing the copy to the end of the call sequence. 885 SDValue RetValue = DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), 886 VA.getLocVT(), Glue); 887 Chain = RetValue.getValue(1); 888 Glue = RetValue.getValue(2); 889 890 // Convert the value of the return register into the value that's 891 // being returned. 892 InVals.push_back(convertLocVTToValVT(DAG, DL, VA, Chain, RetValue)); 893 } 894 895 return Chain; 896} 897 898SDValue 899SystemZTargetLowering::LowerReturn(SDValue Chain, 900 CallingConv::ID CallConv, bool IsVarArg, 901 const SmallVectorImpl<ISD::OutputArg> &Outs, 902 const SmallVectorImpl<SDValue> &OutVals, 903 SDLoc DL, SelectionDAG &DAG) const { 904 MachineFunction &MF = DAG.getMachineFunction(); 905 906 // Assign locations to each returned value. 907 SmallVector<CCValAssign, 16> RetLocs; 908 CCState RetCCInfo(CallConv, IsVarArg, MF, TM, RetLocs, *DAG.getContext()); 909 RetCCInfo.AnalyzeReturn(Outs, RetCC_SystemZ); 910 911 // Quick exit for void returns 912 if (RetLocs.empty()) 913 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other, Chain); 914 915 // Copy the result values into the output registers. 916 SDValue Glue; 917 SmallVector<SDValue, 4> RetOps; 918 RetOps.push_back(Chain); 919 for (unsigned I = 0, E = RetLocs.size(); I != E; ++I) { 920 CCValAssign &VA = RetLocs[I]; 921 SDValue RetValue = OutVals[I]; 922 923 // Make the return register live on exit. 924 assert(VA.isRegLoc() && "Can only return in registers!"); 925 926 // Promote the value as required. 927 RetValue = convertValVTToLocVT(DAG, DL, VA, RetValue); 928 929 // Chain and glue the copies together. 930 unsigned Reg = VA.getLocReg(); 931 Chain = DAG.getCopyToReg(Chain, DL, Reg, RetValue, Glue); 932 Glue = Chain.getValue(1); 933 RetOps.push_back(DAG.getRegister(Reg, VA.getLocVT())); 934 } 935 936 // Update chain and glue. 937 RetOps[0] = Chain; 938 if (Glue.getNode()) 939 RetOps.push_back(Glue); 940 941 return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other, 942 RetOps.data(), RetOps.size()); 943} 944 945// CC is a comparison that will be implemented using an integer or 946// floating-point comparison. Return the condition code mask for 947// a branch on true. In the integer case, CCMASK_CMP_UO is set for 948// unsigned comparisons and clear for signed ones. In the floating-point 949// case, CCMASK_CMP_UO has its normal mask meaning (unordered). 950static unsigned CCMaskForCondCode(ISD::CondCode CC) { 951#define CONV(X) \ 952 case ISD::SET##X: return SystemZ::CCMASK_CMP_##X; \ 953 case ISD::SETO##X: return SystemZ::CCMASK_CMP_##X; \ 954 case ISD::SETU##X: return SystemZ::CCMASK_CMP_UO | SystemZ::CCMASK_CMP_##X 955 956 switch (CC) { 957 default: 958 llvm_unreachable("Invalid integer condition!"); 959 960 CONV(EQ); 961 CONV(NE); 962 CONV(GT); 963 CONV(GE); 964 CONV(LT); 965 CONV(LE); 966 967 case ISD::SETO: return SystemZ::CCMASK_CMP_O; 968 case ISD::SETUO: return SystemZ::CCMASK_CMP_UO; 969 } 970#undef CONV 971} 972 973// If a comparison described by IsUnsigned, CCMask, CmpOp0 and CmpOp1 974// can be converted to a comparison against zero, adjust the operands 975// as necessary. 976static void adjustZeroCmp(SelectionDAG &DAG, bool &IsUnsigned, 977 SDValue &CmpOp0, SDValue &CmpOp1, 978 unsigned &CCMask) { 979 if (IsUnsigned) 980 return; 981 982 ConstantSDNode *ConstOp1 = dyn_cast<ConstantSDNode>(CmpOp1.getNode()); 983 if (!ConstOp1) 984 return; 985 986 int64_t Value = ConstOp1->getSExtValue(); 987 if ((Value == -1 && CCMask == SystemZ::CCMASK_CMP_GT) || 988 (Value == -1 && CCMask == SystemZ::CCMASK_CMP_LE) || 989 (Value == 1 && CCMask == SystemZ::CCMASK_CMP_LT) || 990 (Value == 1 && CCMask == SystemZ::CCMASK_CMP_GE)) { 991 CCMask ^= SystemZ::CCMASK_CMP_EQ; 992 CmpOp1 = DAG.getConstant(0, CmpOp1.getValueType()); 993 } 994} 995 996// If a comparison described by IsUnsigned, CCMask, CmpOp0 and CmpOp1 997// is suitable for CLI(Y), CHHSI or CLHHSI, adjust the operands as necessary. 998static void adjustSubwordCmp(SelectionDAG &DAG, bool &IsUnsigned, 999 SDValue &CmpOp0, SDValue &CmpOp1, 1000 unsigned &CCMask) { 1001 // For us to make any changes, it must a comparison between a single-use 1002 // load and a constant. 1003 if (!CmpOp0.hasOneUse() || 1004 CmpOp0.getOpcode() != ISD::LOAD || 1005 CmpOp1.getOpcode() != ISD::Constant) 1006 return; 1007 1008 // We must have an 8- or 16-bit load. 1009 LoadSDNode *Load = cast<LoadSDNode>(CmpOp0); 1010 unsigned NumBits = Load->getMemoryVT().getStoreSizeInBits(); 1011 if (NumBits != 8 && NumBits != 16) 1012 return; 1013 1014 // The load must be an extending one and the constant must be within the 1015 // range of the unextended value. 1016 ConstantSDNode *Constant = cast<ConstantSDNode>(CmpOp1); 1017 uint64_t Value = Constant->getZExtValue(); 1018 uint64_t Mask = (1 << NumBits) - 1; 1019 if (Load->getExtensionType() == ISD::SEXTLOAD) { 1020 int64_t SignedValue = Constant->getSExtValue(); 1021 if (uint64_t(SignedValue) + (1ULL << (NumBits - 1)) > Mask) 1022 return; 1023 // Unsigned comparison between two sign-extended values is equivalent 1024 // to unsigned comparison between two zero-extended values. 1025 if (IsUnsigned) 1026 Value &= Mask; 1027 else if (CCMask == SystemZ::CCMASK_CMP_EQ || 1028 CCMask == SystemZ::CCMASK_CMP_NE) 1029 // Any choice of IsUnsigned is OK for equality comparisons. 1030 // We could use either CHHSI or CLHHSI for 16-bit comparisons, 1031 // but since we use CLHHSI for zero extensions, it seems better 1032 // to be consistent and do the same here. 1033 Value &= Mask, IsUnsigned = true; 1034 else if (NumBits == 8) { 1035 // Try to treat the comparison as unsigned, so that we can use CLI. 1036 // Adjust CCMask and Value as necessary. 1037 if (Value == 0 && CCMask == SystemZ::CCMASK_CMP_LT) 1038 // Test whether the high bit of the byte is set. 1039 Value = 127, CCMask = SystemZ::CCMASK_CMP_GT, IsUnsigned = true; 1040 else if (Value == 0 && CCMask == SystemZ::CCMASK_CMP_GE) 1041 // Test whether the high bit of the byte is clear. 1042 Value = 128, CCMask = SystemZ::CCMASK_CMP_LT, IsUnsigned = true; 1043 else 1044 // No instruction exists for this combination. 1045 return; 1046 } 1047 } else if (Load->getExtensionType() == ISD::ZEXTLOAD) { 1048 if (Value > Mask) 1049 return; 1050 // Signed comparison between two zero-extended values is equivalent 1051 // to unsigned comparison. 1052 IsUnsigned = true; 1053 } else 1054 return; 1055 1056 // Make sure that the first operand is an i32 of the right extension type. 1057 ISD::LoadExtType ExtType = IsUnsigned ? ISD::ZEXTLOAD : ISD::SEXTLOAD; 1058 if (CmpOp0.getValueType() != MVT::i32 || 1059 Load->getExtensionType() != ExtType) 1060 CmpOp0 = DAG.getExtLoad(ExtType, SDLoc(Load), MVT::i32, 1061 Load->getChain(), Load->getBasePtr(), 1062 Load->getPointerInfo(), Load->getMemoryVT(), 1063 Load->isVolatile(), Load->isNonTemporal(), 1064 Load->getAlignment()); 1065 1066 // Make sure that the second operand is an i32 with the right value. 1067 if (CmpOp1.getValueType() != MVT::i32 || 1068 Value != Constant->getZExtValue()) 1069 CmpOp1 = DAG.getConstant(Value, MVT::i32); 1070} 1071 1072// Return true if Op is either an unextended load, or a load suitable 1073// for integer register-memory comparisons of type ICmpType. 1074static bool isNaturalMemoryOperand(SDValue Op, unsigned ICmpType) { 1075 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op.getNode()); 1076 if (Load) { 1077 // There are no instructions to compare a register with a memory byte. 1078 if (Load->getMemoryVT() == MVT::i8) 1079 return false; 1080 // Otherwise decide on extension type. 1081 switch (Load->getExtensionType()) { 1082 case ISD::NON_EXTLOAD: 1083 return true; 1084 case ISD::SEXTLOAD: 1085 return ICmpType != SystemZICMP::UnsignedOnly; 1086 case ISD::ZEXTLOAD: 1087 return ICmpType != SystemZICMP::SignedOnly; 1088 default: 1089 break; 1090 } 1091 } 1092 return false; 1093} 1094 1095// Return true if it is better to swap comparison operands Op0 and Op1. 1096// ICmpType is the type of an integer comparison. 1097static bool shouldSwapCmpOperands(SDValue Op0, SDValue Op1, 1098 unsigned ICmpType) { 1099 // Leave f128 comparisons alone, since they have no memory forms. 1100 if (Op0.getValueType() == MVT::f128) 1101 return false; 1102 1103 // Always keep a floating-point constant second, since comparisons with 1104 // zero can use LOAD TEST and comparisons with other constants make a 1105 // natural memory operand. 1106 if (isa<ConstantFPSDNode>(Op1)) 1107 return false; 1108 1109 // Never swap comparisons with zero since there are many ways to optimize 1110 // those later. 1111 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1); 1112 if (COp1 && COp1->getZExtValue() == 0) 1113 return false; 1114 1115 // Look for cases where Cmp0 is a single-use load and Cmp1 isn't. 1116 // In that case we generally prefer the memory to be second. 1117 if ((isNaturalMemoryOperand(Op0, ICmpType) && Op0.hasOneUse()) && 1118 !(isNaturalMemoryOperand(Op1, ICmpType) && Op1.hasOneUse())) { 1119 // The only exceptions are when the second operand is a constant and 1120 // we can use things like CHHSI. 1121 if (!COp1) 1122 return true; 1123 // The unsigned memory-immediate instructions can handle 16-bit 1124 // unsigned integers. 1125 if (ICmpType != SystemZICMP::SignedOnly && 1126 isUInt<16>(COp1->getZExtValue())) 1127 return false; 1128 // The signed memory-immediate instructions can handle 16-bit 1129 // signed integers. 1130 if (ICmpType != SystemZICMP::UnsignedOnly && 1131 isInt<16>(COp1->getSExtValue())) 1132 return false; 1133 return true; 1134 } 1135 return false; 1136} 1137 1138// Check whether the CC value produced by TEST UNDER MASK is descriptive 1139// enough to handle an AND with Mask followed by a comparison of type Opcode 1140// with CmpVal. CCMask says which comparison result is being tested and 1141// BitSize is the number of bits in the operands. Return the CC mask that 1142// should be used for the TEST UNDER MASK result, or 0 if the condition is 1143// too complex. 1144static unsigned getTestUnderMaskCond(unsigned BitSize, unsigned CCMask, 1145 uint64_t Mask, uint64_t CmpVal, 1146 unsigned ICmpType) { 1147 assert(Mask != 0 && "ANDs with zero should have been removed by now"); 1148 1149 // Work out the masks for the lowest and highest bits. 1150 unsigned HighShift = 63 - countLeadingZeros(Mask); 1151 uint64_t High = uint64_t(1) << HighShift; 1152 uint64_t Low = uint64_t(1) << countTrailingZeros(Mask); 1153 1154 // Signed ordered comparisons are effectively unsigned if the sign 1155 // bit is dropped. 1156 bool EffectivelyUnsigned = (ICmpType != SystemZICMP::SignedOnly); 1157 1158 // Check for equality comparisons with 0, or the equivalent. 1159 if (CmpVal == 0) { 1160 if (CCMask == SystemZ::CCMASK_CMP_EQ) 1161 return SystemZ::CCMASK_TM_ALL_0; 1162 if (CCMask == SystemZ::CCMASK_CMP_NE) 1163 return SystemZ::CCMASK_TM_SOME_1; 1164 } 1165 if (EffectivelyUnsigned && CmpVal <= Low) { 1166 if (CCMask == SystemZ::CCMASK_CMP_LT) 1167 return SystemZ::CCMASK_TM_ALL_0; 1168 if (CCMask == SystemZ::CCMASK_CMP_GE) 1169 return SystemZ::CCMASK_TM_SOME_1; 1170 } 1171 if (EffectivelyUnsigned && CmpVal < Low) { 1172 if (CCMask == SystemZ::CCMASK_CMP_LE) 1173 return SystemZ::CCMASK_TM_ALL_0; 1174 if (CCMask == SystemZ::CCMASK_CMP_GT) 1175 return SystemZ::CCMASK_TM_SOME_1; 1176 } 1177 1178 // Check for equality comparisons with the mask, or the equivalent. 1179 if (CmpVal == Mask) { 1180 if (CCMask == SystemZ::CCMASK_CMP_EQ) 1181 return SystemZ::CCMASK_TM_ALL_1; 1182 if (CCMask == SystemZ::CCMASK_CMP_NE) 1183 return SystemZ::CCMASK_TM_SOME_0; 1184 } 1185 if (EffectivelyUnsigned && CmpVal >= Mask - Low && CmpVal < Mask) { 1186 if (CCMask == SystemZ::CCMASK_CMP_GT) 1187 return SystemZ::CCMASK_TM_ALL_1; 1188 if (CCMask == SystemZ::CCMASK_CMP_LE) 1189 return SystemZ::CCMASK_TM_SOME_0; 1190 } 1191 if (EffectivelyUnsigned && CmpVal > Mask - Low && CmpVal <= Mask) { 1192 if (CCMask == SystemZ::CCMASK_CMP_GE) 1193 return SystemZ::CCMASK_TM_ALL_1; 1194 if (CCMask == SystemZ::CCMASK_CMP_LT) 1195 return SystemZ::CCMASK_TM_SOME_0; 1196 } 1197 1198 // Check for ordered comparisons with the top bit. 1199 if (EffectivelyUnsigned && CmpVal >= Mask - High && CmpVal < High) { 1200 if (CCMask == SystemZ::CCMASK_CMP_LE) 1201 return SystemZ::CCMASK_TM_MSB_0; 1202 if (CCMask == SystemZ::CCMASK_CMP_GT) 1203 return SystemZ::CCMASK_TM_MSB_1; 1204 } 1205 if (EffectivelyUnsigned && CmpVal > Mask - High && CmpVal <= High) { 1206 if (CCMask == SystemZ::CCMASK_CMP_LT) 1207 return SystemZ::CCMASK_TM_MSB_0; 1208 if (CCMask == SystemZ::CCMASK_CMP_GE) 1209 return SystemZ::CCMASK_TM_MSB_1; 1210 } 1211 1212 // If there are just two bits, we can do equality checks for Low and High 1213 // as well. 1214 if (Mask == Low + High) { 1215 if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == Low) 1216 return SystemZ::CCMASK_TM_MIXED_MSB_0; 1217 if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == Low) 1218 return SystemZ::CCMASK_TM_MIXED_MSB_0 ^ SystemZ::CCMASK_ANY; 1219 if (CCMask == SystemZ::CCMASK_CMP_EQ && CmpVal == High) 1220 return SystemZ::CCMASK_TM_MIXED_MSB_1; 1221 if (CCMask == SystemZ::CCMASK_CMP_NE && CmpVal == High) 1222 return SystemZ::CCMASK_TM_MIXED_MSB_1 ^ SystemZ::CCMASK_ANY; 1223 } 1224 1225 // Looks like we've exhausted our options. 1226 return 0; 1227} 1228 1229// See whether the comparison (Opcode CmpOp0, CmpOp1, ICmpType) can be 1230// implemented as a TEST UNDER MASK instruction when the condition being 1231// tested is as described by CCValid and CCMask. Update the arguments 1232// with the TM version if so. 1233static void adjustForTestUnderMask(unsigned &Opcode, SDValue &CmpOp0, 1234 SDValue &CmpOp1, unsigned &CCValid, 1235 unsigned &CCMask, unsigned &ICmpType) { 1236 // Check that we have a comparison with a constant. 1237 ConstantSDNode *ConstCmpOp1 = dyn_cast<ConstantSDNode>(CmpOp1); 1238 if (!ConstCmpOp1) 1239 return; 1240 1241 // Check whether the nonconstant input is an AND with a constant mask. 1242 if (CmpOp0.getOpcode() != ISD::AND) 1243 return; 1244 SDValue AndOp0 = CmpOp0.getOperand(0); 1245 SDValue AndOp1 = CmpOp0.getOperand(1); 1246 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(AndOp1.getNode()); 1247 if (!Mask) 1248 return; 1249 1250 // Check whether the mask is suitable for TMHH, TMHL, TMLH or TMLL. 1251 uint64_t MaskVal = Mask->getZExtValue(); 1252 if (!SystemZ::isImmLL(MaskVal) && !SystemZ::isImmLH(MaskVal) && 1253 !SystemZ::isImmHL(MaskVal) && !SystemZ::isImmHH(MaskVal)) 1254 return; 1255 1256 // Check whether the combination of mask, comparison value and comparison 1257 // type are suitable. 1258 unsigned BitSize = CmpOp0.getValueType().getSizeInBits(); 1259 unsigned NewCCMask = getTestUnderMaskCond(BitSize, CCMask, MaskVal, 1260 ConstCmpOp1->getZExtValue(), 1261 ICmpType); 1262 if (!NewCCMask) 1263 return; 1264 1265 // Go ahead and make the change. 1266 Opcode = SystemZISD::TM; 1267 CmpOp0 = AndOp0; 1268 CmpOp1 = AndOp1; 1269 ICmpType = (bool(NewCCMask & SystemZ::CCMASK_TM_MIXED_MSB_0) != 1270 bool(NewCCMask & SystemZ::CCMASK_TM_MIXED_MSB_1)); 1271 CCValid = SystemZ::CCMASK_TM; 1272 CCMask = NewCCMask; 1273} 1274 1275// Return a target node that compares CmpOp0 with CmpOp1 and stores a 1276// 2-bit result in CC. Set CCValid to the CCMASK_* of all possible 1277// 2-bit results and CCMask to the subset of those results that are 1278// associated with Cond. 1279static SDValue emitCmp(const SystemZTargetMachine &TM, SelectionDAG &DAG, 1280 SDLoc DL, SDValue CmpOp0, SDValue CmpOp1, 1281 ISD::CondCode Cond, unsigned &CCValid, 1282 unsigned &CCMask) { 1283 bool IsUnsigned = false; 1284 CCMask = CCMaskForCondCode(Cond); 1285 unsigned Opcode, ICmpType = 0; 1286 if (CmpOp0.getValueType().isFloatingPoint()) { 1287 CCValid = SystemZ::CCMASK_FCMP; 1288 Opcode = SystemZISD::FCMP; 1289 } else { 1290 IsUnsigned = CCMask & SystemZ::CCMASK_CMP_UO; 1291 CCValid = SystemZ::CCMASK_ICMP; 1292 CCMask &= CCValid; 1293 adjustZeroCmp(DAG, IsUnsigned, CmpOp0, CmpOp1, CCMask); 1294 adjustSubwordCmp(DAG, IsUnsigned, CmpOp0, CmpOp1, CCMask); 1295 Opcode = SystemZISD::ICMP; 1296 // Choose the type of comparison. Equality and inequality tests can 1297 // use either signed or unsigned comparisons. The choice also doesn't 1298 // matter if both sign bits are known to be clear. In those cases we 1299 // want to give the main isel code the freedom to choose whichever 1300 // form fits best. 1301 if (CCMask == SystemZ::CCMASK_CMP_EQ || 1302 CCMask == SystemZ::CCMASK_CMP_NE || 1303 (DAG.SignBitIsZero(CmpOp0) && DAG.SignBitIsZero(CmpOp1))) 1304 ICmpType = SystemZICMP::Any; 1305 else if (IsUnsigned) 1306 ICmpType = SystemZICMP::UnsignedOnly; 1307 else 1308 ICmpType = SystemZICMP::SignedOnly; 1309 } 1310 1311 if (shouldSwapCmpOperands(CmpOp0, CmpOp1, ICmpType)) { 1312 std::swap(CmpOp0, CmpOp1); 1313 CCMask = ((CCMask & SystemZ::CCMASK_CMP_EQ) | 1314 (CCMask & SystemZ::CCMASK_CMP_GT ? SystemZ::CCMASK_CMP_LT : 0) | 1315 (CCMask & SystemZ::CCMASK_CMP_LT ? SystemZ::CCMASK_CMP_GT : 0) | 1316 (CCMask & SystemZ::CCMASK_CMP_UO)); 1317 } 1318 1319 adjustForTestUnderMask(Opcode, CmpOp0, CmpOp1, CCValid, CCMask, ICmpType); 1320 if (Opcode == SystemZISD::ICMP || Opcode == SystemZISD::TM) 1321 return DAG.getNode(Opcode, DL, MVT::Glue, CmpOp0, CmpOp1, 1322 DAG.getConstant(ICmpType, MVT::i32)); 1323 return DAG.getNode(Opcode, DL, MVT::Glue, CmpOp0, CmpOp1); 1324} 1325 1326// Implement a 32-bit *MUL_LOHI operation by extending both operands to 1327// 64 bits. Extend is the extension type to use. Store the high part 1328// in Hi and the low part in Lo. 1329static void lowerMUL_LOHI32(SelectionDAG &DAG, SDLoc DL, 1330 unsigned Extend, SDValue Op0, SDValue Op1, 1331 SDValue &Hi, SDValue &Lo) { 1332 Op0 = DAG.getNode(Extend, DL, MVT::i64, Op0); 1333 Op1 = DAG.getNode(Extend, DL, MVT::i64, Op1); 1334 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, Op0, Op1); 1335 Hi = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul, DAG.getConstant(32, MVT::i64)); 1336 Hi = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Hi); 1337 Lo = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mul); 1338} 1339 1340// Lower a binary operation that produces two VT results, one in each 1341// half of a GR128 pair. Op0 and Op1 are the VT operands to the operation, 1342// Extend extends Op0 to a GR128, and Opcode performs the GR128 operation 1343// on the extended Op0 and (unextended) Op1. Store the even register result 1344// in Even and the odd register result in Odd. 1345static void lowerGR128Binary(SelectionDAG &DAG, SDLoc DL, EVT VT, 1346 unsigned Extend, unsigned Opcode, 1347 SDValue Op0, SDValue Op1, 1348 SDValue &Even, SDValue &Odd) { 1349 SDNode *In128 = DAG.getMachineNode(Extend, DL, MVT::Untyped, Op0); 1350 SDValue Result = DAG.getNode(Opcode, DL, MVT::Untyped, 1351 SDValue(In128, 0), Op1); 1352 bool Is32Bit = is32Bit(VT); 1353 SDValue SubReg0 = DAG.getTargetConstant(SystemZ::even128(Is32Bit), VT); 1354 SDValue SubReg1 = DAG.getTargetConstant(SystemZ::odd128(Is32Bit), VT); 1355 SDNode *Reg0 = DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, 1356 VT, Result, SubReg0); 1357 SDNode *Reg1 = DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, 1358 VT, Result, SubReg1); 1359 Even = SDValue(Reg0, 0); 1360 Odd = SDValue(Reg1, 0); 1361} 1362 1363SDValue SystemZTargetLowering::lowerBR_CC(SDValue Op, SelectionDAG &DAG) const { 1364 SDValue Chain = Op.getOperand(0); 1365 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 1366 SDValue CmpOp0 = Op.getOperand(2); 1367 SDValue CmpOp1 = Op.getOperand(3); 1368 SDValue Dest = Op.getOperand(4); 1369 SDLoc DL(Op); 1370 1371 unsigned CCValid, CCMask; 1372 SDValue Flags = emitCmp(TM, DAG, DL, CmpOp0, CmpOp1, CC, CCValid, CCMask); 1373 return DAG.getNode(SystemZISD::BR_CCMASK, DL, Op.getValueType(), 1374 Chain, DAG.getConstant(CCValid, MVT::i32), 1375 DAG.getConstant(CCMask, MVT::i32), Dest, Flags); 1376} 1377 1378SDValue SystemZTargetLowering::lowerSELECT_CC(SDValue Op, 1379 SelectionDAG &DAG) const { 1380 SDValue CmpOp0 = Op.getOperand(0); 1381 SDValue CmpOp1 = Op.getOperand(1); 1382 SDValue TrueOp = Op.getOperand(2); 1383 SDValue FalseOp = Op.getOperand(3); 1384 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get(); 1385 SDLoc DL(Op); 1386 1387 unsigned CCValid, CCMask; 1388 SDValue Flags = emitCmp(TM, DAG, DL, CmpOp0, CmpOp1, CC, CCValid, CCMask); 1389 1390 SmallVector<SDValue, 5> Ops; 1391 Ops.push_back(TrueOp); 1392 Ops.push_back(FalseOp); 1393 Ops.push_back(DAG.getConstant(CCValid, MVT::i32)); 1394 Ops.push_back(DAG.getConstant(CCMask, MVT::i32)); 1395 Ops.push_back(Flags); 1396 1397 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue); 1398 return DAG.getNode(SystemZISD::SELECT_CCMASK, DL, VTs, &Ops[0], Ops.size()); 1399} 1400 1401SDValue SystemZTargetLowering::lowerGlobalAddress(GlobalAddressSDNode *Node, 1402 SelectionDAG &DAG) const { 1403 SDLoc DL(Node); 1404 const GlobalValue *GV = Node->getGlobal(); 1405 int64_t Offset = Node->getOffset(); 1406 EVT PtrVT = getPointerTy(); 1407 Reloc::Model RM = TM.getRelocationModel(); 1408 CodeModel::Model CM = TM.getCodeModel(); 1409 1410 SDValue Result; 1411 if (Subtarget.isPC32DBLSymbol(GV, RM, CM)) { 1412 // Make sure that the offset is aligned to a halfword. If it isn't, 1413 // create an "anchor" at the previous 12-bit boundary. 1414 // FIXME check whether there is a better way of handling this. 1415 if (Offset & 1) { 1416 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 1417 Offset & ~uint64_t(0xfff)); 1418 Offset &= 0xfff; 1419 } else { 1420 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, Offset); 1421 Offset = 0; 1422 } 1423 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result); 1424 } else { 1425 Result = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, SystemZII::MO_GOT); 1426 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result); 1427 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result, 1428 MachinePointerInfo::getGOT(), false, false, false, 0); 1429 } 1430 1431 // If there was a non-zero offset that we didn't fold, create an explicit 1432 // addition for it. 1433 if (Offset != 0) 1434 Result = DAG.getNode(ISD::ADD, DL, PtrVT, Result, 1435 DAG.getConstant(Offset, PtrVT)); 1436 1437 return Result; 1438} 1439 1440SDValue SystemZTargetLowering::lowerGlobalTLSAddress(GlobalAddressSDNode *Node, 1441 SelectionDAG &DAG) const { 1442 SDLoc DL(Node); 1443 const GlobalValue *GV = Node->getGlobal(); 1444 EVT PtrVT = getPointerTy(); 1445 TLSModel::Model model = TM.getTLSModel(GV); 1446 1447 if (model != TLSModel::LocalExec) 1448 llvm_unreachable("only local-exec TLS mode supported"); 1449 1450 // The high part of the thread pointer is in access register 0. 1451 SDValue TPHi = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32, 1452 DAG.getConstant(0, MVT::i32)); 1453 TPHi = DAG.getNode(ISD::ANY_EXTEND, DL, PtrVT, TPHi); 1454 1455 // The low part of the thread pointer is in access register 1. 1456 SDValue TPLo = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32, 1457 DAG.getConstant(1, MVT::i32)); 1458 TPLo = DAG.getNode(ISD::ZERO_EXTEND, DL, PtrVT, TPLo); 1459 1460 // Merge them into a single 64-bit address. 1461 SDValue TPHiShifted = DAG.getNode(ISD::SHL, DL, PtrVT, TPHi, 1462 DAG.getConstant(32, PtrVT)); 1463 SDValue TP = DAG.getNode(ISD::OR, DL, PtrVT, TPHiShifted, TPLo); 1464 1465 // Get the offset of GA from the thread pointer. 1466 SystemZConstantPoolValue *CPV = 1467 SystemZConstantPoolValue::Create(GV, SystemZCP::NTPOFF); 1468 1469 // Force the offset into the constant pool and load it from there. 1470 SDValue CPAddr = DAG.getConstantPool(CPV, PtrVT, 8); 1471 SDValue Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), 1472 CPAddr, MachinePointerInfo::getConstantPool(), 1473 false, false, false, 0); 1474 1475 // Add the base and offset together. 1476 return DAG.getNode(ISD::ADD, DL, PtrVT, TP, Offset); 1477} 1478 1479SDValue SystemZTargetLowering::lowerBlockAddress(BlockAddressSDNode *Node, 1480 SelectionDAG &DAG) const { 1481 SDLoc DL(Node); 1482 const BlockAddress *BA = Node->getBlockAddress(); 1483 int64_t Offset = Node->getOffset(); 1484 EVT PtrVT = getPointerTy(); 1485 1486 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset); 1487 Result = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result); 1488 return Result; 1489} 1490 1491SDValue SystemZTargetLowering::lowerJumpTable(JumpTableSDNode *JT, 1492 SelectionDAG &DAG) const { 1493 SDLoc DL(JT); 1494 EVT PtrVT = getPointerTy(); 1495 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT); 1496 1497 // Use LARL to load the address of the table. 1498 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result); 1499} 1500 1501SDValue SystemZTargetLowering::lowerConstantPool(ConstantPoolSDNode *CP, 1502 SelectionDAG &DAG) const { 1503 SDLoc DL(CP); 1504 EVT PtrVT = getPointerTy(); 1505 1506 SDValue Result; 1507 if (CP->isMachineConstantPoolEntry()) 1508 Result = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT, 1509 CP->getAlignment()); 1510 else 1511 Result = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, 1512 CP->getAlignment(), CP->getOffset()); 1513 1514 // Use LARL to load the address of the constant pool entry. 1515 return DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Result); 1516} 1517 1518SDValue SystemZTargetLowering::lowerBITCAST(SDValue Op, 1519 SelectionDAG &DAG) const { 1520 SDLoc DL(Op); 1521 SDValue In = Op.getOperand(0); 1522 EVT InVT = In.getValueType(); 1523 EVT ResVT = Op.getValueType(); 1524 1525 SDValue SubReg32 = DAG.getTargetConstant(SystemZ::subreg_32bit, MVT::i64); 1526 SDValue Shift32 = DAG.getConstant(32, MVT::i64); 1527 if (InVT == MVT::i32 && ResVT == MVT::f32) { 1528 SDValue In64 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, In); 1529 SDValue Shift = DAG.getNode(ISD::SHL, DL, MVT::i64, In64, Shift32); 1530 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::f64, Shift); 1531 SDNode *Out = DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, 1532 MVT::f32, Out64, SubReg32); 1533 return SDValue(Out, 0); 1534 } 1535 if (InVT == MVT::f32 && ResVT == MVT::i32) { 1536 SDNode *U64 = DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, MVT::f64); 1537 SDNode *In64 = DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, 1538 MVT::f64, SDValue(U64, 0), In, SubReg32); 1539 SDValue Out64 = DAG.getNode(ISD::BITCAST, DL, MVT::i64, SDValue(In64, 0)); 1540 SDValue Shift = DAG.getNode(ISD::SRL, DL, MVT::i64, Out64, Shift32); 1541 SDValue Out = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Shift); 1542 return Out; 1543 } 1544 llvm_unreachable("Unexpected bitcast combination"); 1545} 1546 1547SDValue SystemZTargetLowering::lowerVASTART(SDValue Op, 1548 SelectionDAG &DAG) const { 1549 MachineFunction &MF = DAG.getMachineFunction(); 1550 SystemZMachineFunctionInfo *FuncInfo = 1551 MF.getInfo<SystemZMachineFunctionInfo>(); 1552 EVT PtrVT = getPointerTy(); 1553 1554 SDValue Chain = Op.getOperand(0); 1555 SDValue Addr = Op.getOperand(1); 1556 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 1557 SDLoc DL(Op); 1558 1559 // The initial values of each field. 1560 const unsigned NumFields = 4; 1561 SDValue Fields[NumFields] = { 1562 DAG.getConstant(FuncInfo->getVarArgsFirstGPR(), PtrVT), 1563 DAG.getConstant(FuncInfo->getVarArgsFirstFPR(), PtrVT), 1564 DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT), 1565 DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT) 1566 }; 1567 1568 // Store each field into its respective slot. 1569 SDValue MemOps[NumFields]; 1570 unsigned Offset = 0; 1571 for (unsigned I = 0; I < NumFields; ++I) { 1572 SDValue FieldAddr = Addr; 1573 if (Offset != 0) 1574 FieldAddr = DAG.getNode(ISD::ADD, DL, PtrVT, FieldAddr, 1575 DAG.getIntPtrConstant(Offset)); 1576 MemOps[I] = DAG.getStore(Chain, DL, Fields[I], FieldAddr, 1577 MachinePointerInfo(SV, Offset), 1578 false, false, 0); 1579 Offset += 8; 1580 } 1581 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps, NumFields); 1582} 1583 1584SDValue SystemZTargetLowering::lowerVACOPY(SDValue Op, 1585 SelectionDAG &DAG) const { 1586 SDValue Chain = Op.getOperand(0); 1587 SDValue DstPtr = Op.getOperand(1); 1588 SDValue SrcPtr = Op.getOperand(2); 1589 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue(); 1590 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue(); 1591 SDLoc DL(Op); 1592 1593 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr, DAG.getIntPtrConstant(32), 1594 /*Align*/8, /*isVolatile*/false, /*AlwaysInline*/false, 1595 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV)); 1596} 1597 1598SDValue SystemZTargetLowering:: 1599lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const { 1600 SDValue Chain = Op.getOperand(0); 1601 SDValue Size = Op.getOperand(1); 1602 SDLoc DL(Op); 1603 1604 unsigned SPReg = getStackPointerRegisterToSaveRestore(); 1605 1606 // Get a reference to the stack pointer. 1607 SDValue OldSP = DAG.getCopyFromReg(Chain, DL, SPReg, MVT::i64); 1608 1609 // Get the new stack pointer value. 1610 SDValue NewSP = DAG.getNode(ISD::SUB, DL, MVT::i64, OldSP, Size); 1611 1612 // Copy the new stack pointer back. 1613 Chain = DAG.getCopyToReg(Chain, DL, SPReg, NewSP); 1614 1615 // The allocated data lives above the 160 bytes allocated for the standard 1616 // frame, plus any outgoing stack arguments. We don't know how much that 1617 // amounts to yet, so emit a special ADJDYNALLOC placeholder. 1618 SDValue ArgAdjust = DAG.getNode(SystemZISD::ADJDYNALLOC, DL, MVT::i64); 1619 SDValue Result = DAG.getNode(ISD::ADD, DL, MVT::i64, NewSP, ArgAdjust); 1620 1621 SDValue Ops[2] = { Result, Chain }; 1622 return DAG.getMergeValues(Ops, 2, DL); 1623} 1624 1625SDValue SystemZTargetLowering::lowerSMUL_LOHI(SDValue Op, 1626 SelectionDAG &DAG) const { 1627 EVT VT = Op.getValueType(); 1628 SDLoc DL(Op); 1629 SDValue Ops[2]; 1630 if (is32Bit(VT)) 1631 // Just do a normal 64-bit multiplication and extract the results. 1632 // We define this so that it can be used for constant division. 1633 lowerMUL_LOHI32(DAG, DL, ISD::SIGN_EXTEND, Op.getOperand(0), 1634 Op.getOperand(1), Ops[1], Ops[0]); 1635 else { 1636 // Do a full 128-bit multiplication based on UMUL_LOHI64: 1637 // 1638 // (ll * rl) + ((lh * rl) << 64) + ((ll * rh) << 64) 1639 // 1640 // but using the fact that the upper halves are either all zeros 1641 // or all ones: 1642 // 1643 // (ll * rl) - ((lh & rl) << 64) - ((ll & rh) << 64) 1644 // 1645 // and grouping the right terms together since they are quicker than the 1646 // multiplication: 1647 // 1648 // (ll * rl) - (((lh & rl) + (ll & rh)) << 64) 1649 SDValue C63 = DAG.getConstant(63, MVT::i64); 1650 SDValue LL = Op.getOperand(0); 1651 SDValue RL = Op.getOperand(1); 1652 SDValue LH = DAG.getNode(ISD::SRA, DL, VT, LL, C63); 1653 SDValue RH = DAG.getNode(ISD::SRA, DL, VT, RL, C63); 1654 // UMUL_LOHI64 returns the low result in the odd register and the high 1655 // result in the even register. SMUL_LOHI is defined to return the 1656 // low half first, so the results are in reverse order. 1657 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64, 1658 LL, RL, Ops[1], Ops[0]); 1659 SDValue NegLLTimesRH = DAG.getNode(ISD::AND, DL, VT, LL, RH); 1660 SDValue NegLHTimesRL = DAG.getNode(ISD::AND, DL, VT, LH, RL); 1661 SDValue NegSum = DAG.getNode(ISD::ADD, DL, VT, NegLLTimesRH, NegLHTimesRL); 1662 Ops[1] = DAG.getNode(ISD::SUB, DL, VT, Ops[1], NegSum); 1663 } 1664 return DAG.getMergeValues(Ops, 2, DL); 1665} 1666 1667SDValue SystemZTargetLowering::lowerUMUL_LOHI(SDValue Op, 1668 SelectionDAG &DAG) const { 1669 EVT VT = Op.getValueType(); 1670 SDLoc DL(Op); 1671 SDValue Ops[2]; 1672 if (is32Bit(VT)) 1673 // Just do a normal 64-bit multiplication and extract the results. 1674 // We define this so that it can be used for constant division. 1675 lowerMUL_LOHI32(DAG, DL, ISD::ZERO_EXTEND, Op.getOperand(0), 1676 Op.getOperand(1), Ops[1], Ops[0]); 1677 else 1678 // UMUL_LOHI64 returns the low result in the odd register and the high 1679 // result in the even register. UMUL_LOHI is defined to return the 1680 // low half first, so the results are in reverse order. 1681 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64, 1682 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]); 1683 return DAG.getMergeValues(Ops, 2, DL); 1684} 1685 1686SDValue SystemZTargetLowering::lowerSDIVREM(SDValue Op, 1687 SelectionDAG &DAG) const { 1688 SDValue Op0 = Op.getOperand(0); 1689 SDValue Op1 = Op.getOperand(1); 1690 EVT VT = Op.getValueType(); 1691 SDLoc DL(Op); 1692 unsigned Opcode; 1693 1694 // We use DSGF for 32-bit division. 1695 if (is32Bit(VT)) { 1696 Op0 = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Op0); 1697 Opcode = SystemZISD::SDIVREM32; 1698 } else if (DAG.ComputeNumSignBits(Op1) > 32) { 1699 Op1 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Op1); 1700 Opcode = SystemZISD::SDIVREM32; 1701 } else 1702 Opcode = SystemZISD::SDIVREM64; 1703 1704 // DSG(F) takes a 64-bit dividend, so the even register in the GR128 1705 // input is "don't care". The instruction returns the remainder in 1706 // the even register and the quotient in the odd register. 1707 SDValue Ops[2]; 1708 lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, Opcode, 1709 Op0, Op1, Ops[1], Ops[0]); 1710 return DAG.getMergeValues(Ops, 2, DL); 1711} 1712 1713SDValue SystemZTargetLowering::lowerUDIVREM(SDValue Op, 1714 SelectionDAG &DAG) const { 1715 EVT VT = Op.getValueType(); 1716 SDLoc DL(Op); 1717 1718 // DL(G) uses a double-width dividend, so we need to clear the even 1719 // register in the GR128 input. The instruction returns the remainder 1720 // in the even register and the quotient in the odd register. 1721 SDValue Ops[2]; 1722 if (is32Bit(VT)) 1723 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_32, SystemZISD::UDIVREM32, 1724 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]); 1725 else 1726 lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_64, SystemZISD::UDIVREM64, 1727 Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]); 1728 return DAG.getMergeValues(Ops, 2, DL); 1729} 1730 1731SDValue SystemZTargetLowering::lowerOR(SDValue Op, SelectionDAG &DAG) const { 1732 assert(Op.getValueType() == MVT::i64 && "Should be 64-bit operation"); 1733 1734 // Get the known-zero masks for each operand. 1735 SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1) }; 1736 APInt KnownZero[2], KnownOne[2]; 1737 DAG.ComputeMaskedBits(Ops[0], KnownZero[0], KnownOne[0]); 1738 DAG.ComputeMaskedBits(Ops[1], KnownZero[1], KnownOne[1]); 1739 1740 // See if the upper 32 bits of one operand and the lower 32 bits of the 1741 // other are known zero. They are the low and high operands respectively. 1742 uint64_t Masks[] = { KnownZero[0].getZExtValue(), 1743 KnownZero[1].getZExtValue() }; 1744 unsigned High, Low; 1745 if ((Masks[0] >> 32) == 0xffffffff && uint32_t(Masks[1]) == 0xffffffff) 1746 High = 1, Low = 0; 1747 else if ((Masks[1] >> 32) == 0xffffffff && uint32_t(Masks[0]) == 0xffffffff) 1748 High = 0, Low = 1; 1749 else 1750 return Op; 1751 1752 SDValue LowOp = Ops[Low]; 1753 SDValue HighOp = Ops[High]; 1754 1755 // If the high part is a constant, we're better off using IILH. 1756 if (HighOp.getOpcode() == ISD::Constant) 1757 return Op; 1758 1759 // If the low part is a constant that is outside the range of LHI, 1760 // then we're better off using IILF. 1761 if (LowOp.getOpcode() == ISD::Constant) { 1762 int64_t Value = int32_t(cast<ConstantSDNode>(LowOp)->getZExtValue()); 1763 if (!isInt<16>(Value)) 1764 return Op; 1765 } 1766 1767 // Check whether the high part is an AND that doesn't change the 1768 // high 32 bits and just masks out low bits. We can skip it if so. 1769 if (HighOp.getOpcode() == ISD::AND && 1770 HighOp.getOperand(1).getOpcode() == ISD::Constant) { 1771 ConstantSDNode *MaskNode = cast<ConstantSDNode>(HighOp.getOperand(1)); 1772 uint64_t Mask = MaskNode->getZExtValue() | Masks[High]; 1773 if ((Mask >> 32) == 0xffffffff) 1774 HighOp = HighOp.getOperand(0); 1775 } 1776 1777 // Take advantage of the fact that all GR32 operations only change the 1778 // low 32 bits by truncating Low to an i32 and inserting it directly 1779 // using a subreg. The interesting cases are those where the truncation 1780 // can be folded. 1781 SDLoc DL(Op); 1782 SDValue Low32 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, LowOp); 1783 SDValue SubReg32 = DAG.getTargetConstant(SystemZ::subreg_32bit, MVT::i64); 1784 SDNode *Result = DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, 1785 MVT::i64, HighOp, Low32, SubReg32); 1786 return SDValue(Result, 0); 1787} 1788 1789// Op is an 8-, 16-bit or 32-bit ATOMIC_LOAD_* operation. Lower the first 1790// two into the fullword ATOMIC_LOADW_* operation given by Opcode. 1791SDValue SystemZTargetLowering::lowerATOMIC_LOAD(SDValue Op, 1792 SelectionDAG &DAG, 1793 unsigned Opcode) const { 1794 AtomicSDNode *Node = cast<AtomicSDNode>(Op.getNode()); 1795 1796 // 32-bit operations need no code outside the main loop. 1797 EVT NarrowVT = Node->getMemoryVT(); 1798 EVT WideVT = MVT::i32; 1799 if (NarrowVT == WideVT) 1800 return Op; 1801 1802 int64_t BitSize = NarrowVT.getSizeInBits(); 1803 SDValue ChainIn = Node->getChain(); 1804 SDValue Addr = Node->getBasePtr(); 1805 SDValue Src2 = Node->getVal(); 1806 MachineMemOperand *MMO = Node->getMemOperand(); 1807 SDLoc DL(Node); 1808 EVT PtrVT = Addr.getValueType(); 1809 1810 // Convert atomic subtracts of constants into additions. 1811 if (Opcode == SystemZISD::ATOMIC_LOADW_SUB) 1812 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Src2)) { 1813 Opcode = SystemZISD::ATOMIC_LOADW_ADD; 1814 Src2 = DAG.getConstant(-Const->getSExtValue(), Src2.getValueType()); 1815 } 1816 1817 // Get the address of the containing word. 1818 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr, 1819 DAG.getConstant(-4, PtrVT)); 1820 1821 // Get the number of bits that the word must be rotated left in order 1822 // to bring the field to the top bits of a GR32. 1823 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr, 1824 DAG.getConstant(3, PtrVT)); 1825 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift); 1826 1827 // Get the complementing shift amount, for rotating a field in the top 1828 // bits back to its proper position. 1829 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT, 1830 DAG.getConstant(0, WideVT), BitShift); 1831 1832 // Extend the source operand to 32 bits and prepare it for the inner loop. 1833 // ATOMIC_SWAPW uses RISBG to rotate the field left, but all other 1834 // operations require the source to be shifted in advance. (This shift 1835 // can be folded if the source is constant.) For AND and NAND, the lower 1836 // bits must be set, while for other opcodes they should be left clear. 1837 if (Opcode != SystemZISD::ATOMIC_SWAPW) 1838 Src2 = DAG.getNode(ISD::SHL, DL, WideVT, Src2, 1839 DAG.getConstant(32 - BitSize, WideVT)); 1840 if (Opcode == SystemZISD::ATOMIC_LOADW_AND || 1841 Opcode == SystemZISD::ATOMIC_LOADW_NAND) 1842 Src2 = DAG.getNode(ISD::OR, DL, WideVT, Src2, 1843 DAG.getConstant(uint32_t(-1) >> BitSize, WideVT)); 1844 1845 // Construct the ATOMIC_LOADW_* node. 1846 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other); 1847 SDValue Ops[] = { ChainIn, AlignedAddr, Src2, BitShift, NegBitShift, 1848 DAG.getConstant(BitSize, WideVT) }; 1849 SDValue AtomicOp = DAG.getMemIntrinsicNode(Opcode, DL, VTList, Ops, 1850 array_lengthof(Ops), 1851 NarrowVT, MMO); 1852 1853 // Rotate the result of the final CS so that the field is in the lower 1854 // bits of a GR32, then truncate it. 1855 SDValue ResultShift = DAG.getNode(ISD::ADD, DL, WideVT, BitShift, 1856 DAG.getConstant(BitSize, WideVT)); 1857 SDValue Result = DAG.getNode(ISD::ROTL, DL, WideVT, AtomicOp, ResultShift); 1858 1859 SDValue RetOps[2] = { Result, AtomicOp.getValue(1) }; 1860 return DAG.getMergeValues(RetOps, 2, DL); 1861} 1862 1863// Node is an 8- or 16-bit ATOMIC_CMP_SWAP operation. Lower the first two 1864// into a fullword ATOMIC_CMP_SWAPW operation. 1865SDValue SystemZTargetLowering::lowerATOMIC_CMP_SWAP(SDValue Op, 1866 SelectionDAG &DAG) const { 1867 AtomicSDNode *Node = cast<AtomicSDNode>(Op.getNode()); 1868 1869 // We have native support for 32-bit compare and swap. 1870 EVT NarrowVT = Node->getMemoryVT(); 1871 EVT WideVT = MVT::i32; 1872 if (NarrowVT == WideVT) 1873 return Op; 1874 1875 int64_t BitSize = NarrowVT.getSizeInBits(); 1876 SDValue ChainIn = Node->getOperand(0); 1877 SDValue Addr = Node->getOperand(1); 1878 SDValue CmpVal = Node->getOperand(2); 1879 SDValue SwapVal = Node->getOperand(3); 1880 MachineMemOperand *MMO = Node->getMemOperand(); 1881 SDLoc DL(Node); 1882 EVT PtrVT = Addr.getValueType(); 1883 1884 // Get the address of the containing word. 1885 SDValue AlignedAddr = DAG.getNode(ISD::AND, DL, PtrVT, Addr, 1886 DAG.getConstant(-4, PtrVT)); 1887 1888 // Get the number of bits that the word must be rotated left in order 1889 // to bring the field to the top bits of a GR32. 1890 SDValue BitShift = DAG.getNode(ISD::SHL, DL, PtrVT, Addr, 1891 DAG.getConstant(3, PtrVT)); 1892 BitShift = DAG.getNode(ISD::TRUNCATE, DL, WideVT, BitShift); 1893 1894 // Get the complementing shift amount, for rotating a field in the top 1895 // bits back to its proper position. 1896 SDValue NegBitShift = DAG.getNode(ISD::SUB, DL, WideVT, 1897 DAG.getConstant(0, WideVT), BitShift); 1898 1899 // Construct the ATOMIC_CMP_SWAPW node. 1900 SDVTList VTList = DAG.getVTList(WideVT, MVT::Other); 1901 SDValue Ops[] = { ChainIn, AlignedAddr, CmpVal, SwapVal, BitShift, 1902 NegBitShift, DAG.getConstant(BitSize, WideVT) }; 1903 SDValue AtomicOp = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_CMP_SWAPW, DL, 1904 VTList, Ops, array_lengthof(Ops), 1905 NarrowVT, MMO); 1906 return AtomicOp; 1907} 1908 1909SDValue SystemZTargetLowering::lowerSTACKSAVE(SDValue Op, 1910 SelectionDAG &DAG) const { 1911 MachineFunction &MF = DAG.getMachineFunction(); 1912 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true); 1913 return DAG.getCopyFromReg(Op.getOperand(0), SDLoc(Op), 1914 SystemZ::R15D, Op.getValueType()); 1915} 1916 1917SDValue SystemZTargetLowering::lowerSTACKRESTORE(SDValue Op, 1918 SelectionDAG &DAG) const { 1919 MachineFunction &MF = DAG.getMachineFunction(); 1920 MF.getInfo<SystemZMachineFunctionInfo>()->setManipulatesSP(true); 1921 return DAG.getCopyToReg(Op.getOperand(0), SDLoc(Op), 1922 SystemZ::R15D, Op.getOperand(1)); 1923} 1924 1925SDValue SystemZTargetLowering::lowerPREFETCH(SDValue Op, 1926 SelectionDAG &DAG) const { 1927 bool IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue(); 1928 if (!IsData) 1929 // Just preserve the chain. 1930 return Op.getOperand(0); 1931 1932 bool IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue(); 1933 unsigned Code = IsWrite ? SystemZ::PFD_WRITE : SystemZ::PFD_READ; 1934 MemIntrinsicSDNode *Node = cast<MemIntrinsicSDNode>(Op.getNode()); 1935 SDValue Ops[] = { 1936 Op.getOperand(0), 1937 DAG.getConstant(Code, MVT::i32), 1938 Op.getOperand(1) 1939 }; 1940 return DAG.getMemIntrinsicNode(SystemZISD::PREFETCH, SDLoc(Op), 1941 Node->getVTList(), Ops, array_lengthof(Ops), 1942 Node->getMemoryVT(), Node->getMemOperand()); 1943} 1944 1945SDValue SystemZTargetLowering::LowerOperation(SDValue Op, 1946 SelectionDAG &DAG) const { 1947 switch (Op.getOpcode()) { 1948 case ISD::BR_CC: 1949 return lowerBR_CC(Op, DAG); 1950 case ISD::SELECT_CC: 1951 return lowerSELECT_CC(Op, DAG); 1952 case ISD::GlobalAddress: 1953 return lowerGlobalAddress(cast<GlobalAddressSDNode>(Op), DAG); 1954 case ISD::GlobalTLSAddress: 1955 return lowerGlobalTLSAddress(cast<GlobalAddressSDNode>(Op), DAG); 1956 case ISD::BlockAddress: 1957 return lowerBlockAddress(cast<BlockAddressSDNode>(Op), DAG); 1958 case ISD::JumpTable: 1959 return lowerJumpTable(cast<JumpTableSDNode>(Op), DAG); 1960 case ISD::ConstantPool: 1961 return lowerConstantPool(cast<ConstantPoolSDNode>(Op), DAG); 1962 case ISD::BITCAST: 1963 return lowerBITCAST(Op, DAG); 1964 case ISD::VASTART: 1965 return lowerVASTART(Op, DAG); 1966 case ISD::VACOPY: 1967 return lowerVACOPY(Op, DAG); 1968 case ISD::DYNAMIC_STACKALLOC: 1969 return lowerDYNAMIC_STACKALLOC(Op, DAG); 1970 case ISD::SMUL_LOHI: 1971 return lowerSMUL_LOHI(Op, DAG); 1972 case ISD::UMUL_LOHI: 1973 return lowerUMUL_LOHI(Op, DAG); 1974 case ISD::SDIVREM: 1975 return lowerSDIVREM(Op, DAG); 1976 case ISD::UDIVREM: 1977 return lowerUDIVREM(Op, DAG); 1978 case ISD::OR: 1979 return lowerOR(Op, DAG); 1980 case ISD::ATOMIC_SWAP: 1981 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_SWAPW); 1982 case ISD::ATOMIC_LOAD_ADD: 1983 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_ADD); 1984 case ISD::ATOMIC_LOAD_SUB: 1985 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_SUB); 1986 case ISD::ATOMIC_LOAD_AND: 1987 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_AND); 1988 case ISD::ATOMIC_LOAD_OR: 1989 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_OR); 1990 case ISD::ATOMIC_LOAD_XOR: 1991 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_XOR); 1992 case ISD::ATOMIC_LOAD_NAND: 1993 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_NAND); 1994 case ISD::ATOMIC_LOAD_MIN: 1995 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_MIN); 1996 case ISD::ATOMIC_LOAD_MAX: 1997 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_MAX); 1998 case ISD::ATOMIC_LOAD_UMIN: 1999 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_UMIN); 2000 case ISD::ATOMIC_LOAD_UMAX: 2001 return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_UMAX); 2002 case ISD::ATOMIC_CMP_SWAP: 2003 return lowerATOMIC_CMP_SWAP(Op, DAG); 2004 case ISD::STACKSAVE: 2005 return lowerSTACKSAVE(Op, DAG); 2006 case ISD::STACKRESTORE: 2007 return lowerSTACKRESTORE(Op, DAG); 2008 case ISD::PREFETCH: 2009 return lowerPREFETCH(Op, DAG); 2010 default: 2011 llvm_unreachable("Unexpected node to lower"); 2012 } 2013} 2014 2015const char *SystemZTargetLowering::getTargetNodeName(unsigned Opcode) const { 2016#define OPCODE(NAME) case SystemZISD::NAME: return "SystemZISD::" #NAME 2017 switch (Opcode) { 2018 OPCODE(RET_FLAG); 2019 OPCODE(CALL); 2020 OPCODE(SIBCALL); 2021 OPCODE(PCREL_WRAPPER); 2022 OPCODE(ICMP); 2023 OPCODE(FCMP); 2024 OPCODE(TM); 2025 OPCODE(BR_CCMASK); 2026 OPCODE(SELECT_CCMASK); 2027 OPCODE(ADJDYNALLOC); 2028 OPCODE(EXTRACT_ACCESS); 2029 OPCODE(UMUL_LOHI64); 2030 OPCODE(SDIVREM64); 2031 OPCODE(UDIVREM32); 2032 OPCODE(UDIVREM64); 2033 OPCODE(MVC); 2034 OPCODE(MVC_LOOP); 2035 OPCODE(NC); 2036 OPCODE(NC_LOOP); 2037 OPCODE(OC); 2038 OPCODE(OC_LOOP); 2039 OPCODE(XC); 2040 OPCODE(XC_LOOP); 2041 OPCODE(CLC); 2042 OPCODE(CLC_LOOP); 2043 OPCODE(STRCMP); 2044 OPCODE(STPCPY); 2045 OPCODE(SEARCH_STRING); 2046 OPCODE(IPM); 2047 OPCODE(ATOMIC_SWAPW); 2048 OPCODE(ATOMIC_LOADW_ADD); 2049 OPCODE(ATOMIC_LOADW_SUB); 2050 OPCODE(ATOMIC_LOADW_AND); 2051 OPCODE(ATOMIC_LOADW_OR); 2052 OPCODE(ATOMIC_LOADW_XOR); 2053 OPCODE(ATOMIC_LOADW_NAND); 2054 OPCODE(ATOMIC_LOADW_MIN); 2055 OPCODE(ATOMIC_LOADW_MAX); 2056 OPCODE(ATOMIC_LOADW_UMIN); 2057 OPCODE(ATOMIC_LOADW_UMAX); 2058 OPCODE(ATOMIC_CMP_SWAPW); 2059 OPCODE(PREFETCH); 2060 } 2061 return NULL; 2062#undef OPCODE 2063} 2064 2065//===----------------------------------------------------------------------===// 2066// Custom insertion 2067//===----------------------------------------------------------------------===// 2068 2069// Create a new basic block after MBB. 2070static MachineBasicBlock *emitBlockAfter(MachineBasicBlock *MBB) { 2071 MachineFunction &MF = *MBB->getParent(); 2072 MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(MBB->getBasicBlock()); 2073 MF.insert(llvm::next(MachineFunction::iterator(MBB)), NewMBB); 2074 return NewMBB; 2075} 2076 2077// Split MBB after MI and return the new block (the one that contains 2078// instructions after MI). 2079static MachineBasicBlock *splitBlockAfter(MachineInstr *MI, 2080 MachineBasicBlock *MBB) { 2081 MachineBasicBlock *NewMBB = emitBlockAfter(MBB); 2082 NewMBB->splice(NewMBB->begin(), MBB, 2083 llvm::next(MachineBasicBlock::iterator(MI)), 2084 MBB->end()); 2085 NewMBB->transferSuccessorsAndUpdatePHIs(MBB); 2086 return NewMBB; 2087} 2088 2089// Split MBB before MI and return the new block (the one that contains MI). 2090static MachineBasicBlock *splitBlockBefore(MachineInstr *MI, 2091 MachineBasicBlock *MBB) { 2092 MachineBasicBlock *NewMBB = emitBlockAfter(MBB); 2093 NewMBB->splice(NewMBB->begin(), MBB, MI, MBB->end()); 2094 NewMBB->transferSuccessorsAndUpdatePHIs(MBB); 2095 return NewMBB; 2096} 2097 2098// Force base value Base into a register before MI. Return the register. 2099static unsigned forceReg(MachineInstr *MI, MachineOperand &Base, 2100 const SystemZInstrInfo *TII) { 2101 if (Base.isReg()) 2102 return Base.getReg(); 2103 2104 MachineBasicBlock *MBB = MI->getParent(); 2105 MachineFunction &MF = *MBB->getParent(); 2106 MachineRegisterInfo &MRI = MF.getRegInfo(); 2107 2108 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass); 2109 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LA), Reg) 2110 .addOperand(Base).addImm(0).addReg(0); 2111 return Reg; 2112} 2113 2114// Implement EmitInstrWithCustomInserter for pseudo Select* instruction MI. 2115MachineBasicBlock * 2116SystemZTargetLowering::emitSelect(MachineInstr *MI, 2117 MachineBasicBlock *MBB) const { 2118 const SystemZInstrInfo *TII = TM.getInstrInfo(); 2119 2120 unsigned DestReg = MI->getOperand(0).getReg(); 2121 unsigned TrueReg = MI->getOperand(1).getReg(); 2122 unsigned FalseReg = MI->getOperand(2).getReg(); 2123 unsigned CCValid = MI->getOperand(3).getImm(); 2124 unsigned CCMask = MI->getOperand(4).getImm(); 2125 DebugLoc DL = MI->getDebugLoc(); 2126 2127 MachineBasicBlock *StartMBB = MBB; 2128 MachineBasicBlock *JoinMBB = splitBlockBefore(MI, MBB); 2129 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB); 2130 2131 // StartMBB: 2132 // BRC CCMask, JoinMBB 2133 // # fallthrough to FalseMBB 2134 MBB = StartMBB; 2135 BuildMI(MBB, DL, TII->get(SystemZ::BRC)) 2136 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB); 2137 MBB->addSuccessor(JoinMBB); 2138 MBB->addSuccessor(FalseMBB); 2139 2140 // FalseMBB: 2141 // # fallthrough to JoinMBB 2142 MBB = FalseMBB; 2143 MBB->addSuccessor(JoinMBB); 2144 2145 // JoinMBB: 2146 // %Result = phi [ %FalseReg, FalseMBB ], [ %TrueReg, StartMBB ] 2147 // ... 2148 MBB = JoinMBB; 2149 BuildMI(*MBB, MI, DL, TII->get(SystemZ::PHI), DestReg) 2150 .addReg(TrueReg).addMBB(StartMBB) 2151 .addReg(FalseReg).addMBB(FalseMBB); 2152 2153 MI->eraseFromParent(); 2154 return JoinMBB; 2155} 2156 2157// Implement EmitInstrWithCustomInserter for pseudo CondStore* instruction MI. 2158// StoreOpcode is the store to use and Invert says whether the store should 2159// happen when the condition is false rather than true. If a STORE ON 2160// CONDITION is available, STOCOpcode is its opcode, otherwise it is 0. 2161MachineBasicBlock * 2162SystemZTargetLowering::emitCondStore(MachineInstr *MI, 2163 MachineBasicBlock *MBB, 2164 unsigned StoreOpcode, unsigned STOCOpcode, 2165 bool Invert) const { 2166 const SystemZInstrInfo *TII = TM.getInstrInfo(); 2167 2168 unsigned SrcReg = MI->getOperand(0).getReg(); 2169 MachineOperand Base = MI->getOperand(1); 2170 int64_t Disp = MI->getOperand(2).getImm(); 2171 unsigned IndexReg = MI->getOperand(3).getReg(); 2172 unsigned CCValid = MI->getOperand(4).getImm(); 2173 unsigned CCMask = MI->getOperand(5).getImm(); 2174 DebugLoc DL = MI->getDebugLoc(); 2175 2176 StoreOpcode = TII->getOpcodeForOffset(StoreOpcode, Disp); 2177 2178 // Use STOCOpcode if possible. We could use different store patterns in 2179 // order to avoid matching the index register, but the performance trade-offs 2180 // might be more complicated in that case. 2181 if (STOCOpcode && !IndexReg && TM.getSubtargetImpl()->hasLoadStoreOnCond()) { 2182 if (Invert) 2183 CCMask ^= CCValid; 2184 BuildMI(*MBB, MI, DL, TII->get(STOCOpcode)) 2185 .addReg(SrcReg).addOperand(Base).addImm(Disp) 2186 .addImm(CCValid).addImm(CCMask); 2187 MI->eraseFromParent(); 2188 return MBB; 2189 } 2190 2191 // Get the condition needed to branch around the store. 2192 if (!Invert) 2193 CCMask ^= CCValid; 2194 2195 MachineBasicBlock *StartMBB = MBB; 2196 MachineBasicBlock *JoinMBB = splitBlockBefore(MI, MBB); 2197 MachineBasicBlock *FalseMBB = emitBlockAfter(StartMBB); 2198 2199 // StartMBB: 2200 // BRC CCMask, JoinMBB 2201 // # fallthrough to FalseMBB 2202 MBB = StartMBB; 2203 BuildMI(MBB, DL, TII->get(SystemZ::BRC)) 2204 .addImm(CCValid).addImm(CCMask).addMBB(JoinMBB); 2205 MBB->addSuccessor(JoinMBB); 2206 MBB->addSuccessor(FalseMBB); 2207 2208 // FalseMBB: 2209 // store %SrcReg, %Disp(%Index,%Base) 2210 // # fallthrough to JoinMBB 2211 MBB = FalseMBB; 2212 BuildMI(MBB, DL, TII->get(StoreOpcode)) 2213 .addReg(SrcReg).addOperand(Base).addImm(Disp).addReg(IndexReg); 2214 MBB->addSuccessor(JoinMBB); 2215 2216 MI->eraseFromParent(); 2217 return JoinMBB; 2218} 2219 2220// Implement EmitInstrWithCustomInserter for pseudo ATOMIC_LOAD{,W}_* 2221// or ATOMIC_SWAP{,W} instruction MI. BinOpcode is the instruction that 2222// performs the binary operation elided by "*", or 0 for ATOMIC_SWAP{,W}. 2223// BitSize is the width of the field in bits, or 0 if this is a partword 2224// ATOMIC_LOADW_* or ATOMIC_SWAPW instruction, in which case the bitsize 2225// is one of the operands. Invert says whether the field should be 2226// inverted after performing BinOpcode (e.g. for NAND). 2227MachineBasicBlock * 2228SystemZTargetLowering::emitAtomicLoadBinary(MachineInstr *MI, 2229 MachineBasicBlock *MBB, 2230 unsigned BinOpcode, 2231 unsigned BitSize, 2232 bool Invert) const { 2233 const SystemZInstrInfo *TII = TM.getInstrInfo(); 2234 MachineFunction &MF = *MBB->getParent(); 2235 MachineRegisterInfo &MRI = MF.getRegInfo(); 2236 bool IsSubWord = (BitSize < 32); 2237 2238 // Extract the operands. Base can be a register or a frame index. 2239 // Src2 can be a register or immediate. 2240 unsigned Dest = MI->getOperand(0).getReg(); 2241 MachineOperand Base = earlyUseOperand(MI->getOperand(1)); 2242 int64_t Disp = MI->getOperand(2).getImm(); 2243 MachineOperand Src2 = earlyUseOperand(MI->getOperand(3)); 2244 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0); 2245 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0); 2246 DebugLoc DL = MI->getDebugLoc(); 2247 if (IsSubWord) 2248 BitSize = MI->getOperand(6).getImm(); 2249 2250 // Subword operations use 32-bit registers. 2251 const TargetRegisterClass *RC = (BitSize <= 32 ? 2252 &SystemZ::GR32BitRegClass : 2253 &SystemZ::GR64BitRegClass); 2254 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG; 2255 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG; 2256 2257 // Get the right opcodes for the displacement. 2258 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp); 2259 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp); 2260 assert(LOpcode && CSOpcode && "Displacement out of range"); 2261 2262 // Create virtual registers for temporary results. 2263 unsigned OrigVal = MRI.createVirtualRegister(RC); 2264 unsigned OldVal = MRI.createVirtualRegister(RC); 2265 unsigned NewVal = (BinOpcode || IsSubWord ? 2266 MRI.createVirtualRegister(RC) : Src2.getReg()); 2267 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal); 2268 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal); 2269 2270 // Insert a basic block for the main loop. 2271 MachineBasicBlock *StartMBB = MBB; 2272 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB); 2273 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB); 2274 2275 // StartMBB: 2276 // ... 2277 // %OrigVal = L Disp(%Base) 2278 // # fall through to LoopMMB 2279 MBB = StartMBB; 2280 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal) 2281 .addOperand(Base).addImm(Disp).addReg(0); 2282 MBB->addSuccessor(LoopMBB); 2283 2284 // LoopMBB: 2285 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, LoopMBB ] 2286 // %RotatedOldVal = RLL %OldVal, 0(%BitShift) 2287 // %RotatedNewVal = OP %RotatedOldVal, %Src2 2288 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift) 2289 // %Dest = CS %OldVal, %NewVal, Disp(%Base) 2290 // JNE LoopMBB 2291 // # fall through to DoneMMB 2292 MBB = LoopMBB; 2293 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal) 2294 .addReg(OrigVal).addMBB(StartMBB) 2295 .addReg(Dest).addMBB(LoopMBB); 2296 if (IsSubWord) 2297 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal) 2298 .addReg(OldVal).addReg(BitShift).addImm(0); 2299 if (Invert) { 2300 // Perform the operation normally and then invert every bit of the field. 2301 unsigned Tmp = MRI.createVirtualRegister(RC); 2302 BuildMI(MBB, DL, TII->get(BinOpcode), Tmp) 2303 .addReg(RotatedOldVal).addOperand(Src2); 2304 if (BitSize < 32) 2305 // XILF with the upper BitSize bits set. 2306 BuildMI(MBB, DL, TII->get(SystemZ::XILF32), RotatedNewVal) 2307 .addReg(Tmp).addImm(uint32_t(~0 << (32 - BitSize))); 2308 else if (BitSize == 32) 2309 // XILF with every bit set. 2310 BuildMI(MBB, DL, TII->get(SystemZ::XILF32), RotatedNewVal) 2311 .addReg(Tmp).addImm(~uint32_t(0)); 2312 else { 2313 // Use LCGR and add -1 to the result, which is more compact than 2314 // an XILF, XILH pair. 2315 unsigned Tmp2 = MRI.createVirtualRegister(RC); 2316 BuildMI(MBB, DL, TII->get(SystemZ::LCGR), Tmp2).addReg(Tmp); 2317 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), RotatedNewVal) 2318 .addReg(Tmp2).addImm(-1); 2319 } 2320 } else if (BinOpcode) 2321 // A simply binary operation. 2322 BuildMI(MBB, DL, TII->get(BinOpcode), RotatedNewVal) 2323 .addReg(RotatedOldVal).addOperand(Src2); 2324 else if (IsSubWord) 2325 // Use RISBG to rotate Src2 into position and use it to replace the 2326 // field in RotatedOldVal. 2327 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedNewVal) 2328 .addReg(RotatedOldVal).addReg(Src2.getReg()) 2329 .addImm(32).addImm(31 + BitSize).addImm(32 - BitSize); 2330 if (IsSubWord) 2331 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal) 2332 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0); 2333 BuildMI(MBB, DL, TII->get(CSOpcode), Dest) 2334 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp); 2335 BuildMI(MBB, DL, TII->get(SystemZ::BRC)) 2336 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB); 2337 MBB->addSuccessor(LoopMBB); 2338 MBB->addSuccessor(DoneMBB); 2339 2340 MI->eraseFromParent(); 2341 return DoneMBB; 2342} 2343 2344// Implement EmitInstrWithCustomInserter for pseudo 2345// ATOMIC_LOAD{,W}_{,U}{MIN,MAX} instruction MI. CompareOpcode is the 2346// instruction that should be used to compare the current field with the 2347// minimum or maximum value. KeepOldMask is the BRC condition-code mask 2348// for when the current field should be kept. BitSize is the width of 2349// the field in bits, or 0 if this is a partword ATOMIC_LOADW_* instruction. 2350MachineBasicBlock * 2351SystemZTargetLowering::emitAtomicLoadMinMax(MachineInstr *MI, 2352 MachineBasicBlock *MBB, 2353 unsigned CompareOpcode, 2354 unsigned KeepOldMask, 2355 unsigned BitSize) const { 2356 const SystemZInstrInfo *TII = TM.getInstrInfo(); 2357 MachineFunction &MF = *MBB->getParent(); 2358 MachineRegisterInfo &MRI = MF.getRegInfo(); 2359 bool IsSubWord = (BitSize < 32); 2360 2361 // Extract the operands. Base can be a register or a frame index. 2362 unsigned Dest = MI->getOperand(0).getReg(); 2363 MachineOperand Base = earlyUseOperand(MI->getOperand(1)); 2364 int64_t Disp = MI->getOperand(2).getImm(); 2365 unsigned Src2 = MI->getOperand(3).getReg(); 2366 unsigned BitShift = (IsSubWord ? MI->getOperand(4).getReg() : 0); 2367 unsigned NegBitShift = (IsSubWord ? MI->getOperand(5).getReg() : 0); 2368 DebugLoc DL = MI->getDebugLoc(); 2369 if (IsSubWord) 2370 BitSize = MI->getOperand(6).getImm(); 2371 2372 // Subword operations use 32-bit registers. 2373 const TargetRegisterClass *RC = (BitSize <= 32 ? 2374 &SystemZ::GR32BitRegClass : 2375 &SystemZ::GR64BitRegClass); 2376 unsigned LOpcode = BitSize <= 32 ? SystemZ::L : SystemZ::LG; 2377 unsigned CSOpcode = BitSize <= 32 ? SystemZ::CS : SystemZ::CSG; 2378 2379 // Get the right opcodes for the displacement. 2380 LOpcode = TII->getOpcodeForOffset(LOpcode, Disp); 2381 CSOpcode = TII->getOpcodeForOffset(CSOpcode, Disp); 2382 assert(LOpcode && CSOpcode && "Displacement out of range"); 2383 2384 // Create virtual registers for temporary results. 2385 unsigned OrigVal = MRI.createVirtualRegister(RC); 2386 unsigned OldVal = MRI.createVirtualRegister(RC); 2387 unsigned NewVal = MRI.createVirtualRegister(RC); 2388 unsigned RotatedOldVal = (IsSubWord ? MRI.createVirtualRegister(RC) : OldVal); 2389 unsigned RotatedAltVal = (IsSubWord ? MRI.createVirtualRegister(RC) : Src2); 2390 unsigned RotatedNewVal = (IsSubWord ? MRI.createVirtualRegister(RC) : NewVal); 2391 2392 // Insert 3 basic blocks for the loop. 2393 MachineBasicBlock *StartMBB = MBB; 2394 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB); 2395 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB); 2396 MachineBasicBlock *UseAltMBB = emitBlockAfter(LoopMBB); 2397 MachineBasicBlock *UpdateMBB = emitBlockAfter(UseAltMBB); 2398 2399 // StartMBB: 2400 // ... 2401 // %OrigVal = L Disp(%Base) 2402 // # fall through to LoopMMB 2403 MBB = StartMBB; 2404 BuildMI(MBB, DL, TII->get(LOpcode), OrigVal) 2405 .addOperand(Base).addImm(Disp).addReg(0); 2406 MBB->addSuccessor(LoopMBB); 2407 2408 // LoopMBB: 2409 // %OldVal = phi [ %OrigVal, StartMBB ], [ %Dest, UpdateMBB ] 2410 // %RotatedOldVal = RLL %OldVal, 0(%BitShift) 2411 // CompareOpcode %RotatedOldVal, %Src2 2412 // BRC KeepOldMask, UpdateMBB 2413 MBB = LoopMBB; 2414 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal) 2415 .addReg(OrigVal).addMBB(StartMBB) 2416 .addReg(Dest).addMBB(UpdateMBB); 2417 if (IsSubWord) 2418 BuildMI(MBB, DL, TII->get(SystemZ::RLL), RotatedOldVal) 2419 .addReg(OldVal).addReg(BitShift).addImm(0); 2420 BuildMI(MBB, DL, TII->get(CompareOpcode)) 2421 .addReg(RotatedOldVal).addReg(Src2); 2422 BuildMI(MBB, DL, TII->get(SystemZ::BRC)) 2423 .addImm(SystemZ::CCMASK_ICMP).addImm(KeepOldMask).addMBB(UpdateMBB); 2424 MBB->addSuccessor(UpdateMBB); 2425 MBB->addSuccessor(UseAltMBB); 2426 2427 // UseAltMBB: 2428 // %RotatedAltVal = RISBG %RotatedOldVal, %Src2, 32, 31 + BitSize, 0 2429 // # fall through to UpdateMMB 2430 MBB = UseAltMBB; 2431 if (IsSubWord) 2432 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RotatedAltVal) 2433 .addReg(RotatedOldVal).addReg(Src2) 2434 .addImm(32).addImm(31 + BitSize).addImm(0); 2435 MBB->addSuccessor(UpdateMBB); 2436 2437 // UpdateMBB: 2438 // %RotatedNewVal = PHI [ %RotatedOldVal, LoopMBB ], 2439 // [ %RotatedAltVal, UseAltMBB ] 2440 // %NewVal = RLL %RotatedNewVal, 0(%NegBitShift) 2441 // %Dest = CS %OldVal, %NewVal, Disp(%Base) 2442 // JNE LoopMBB 2443 // # fall through to DoneMMB 2444 MBB = UpdateMBB; 2445 BuildMI(MBB, DL, TII->get(SystemZ::PHI), RotatedNewVal) 2446 .addReg(RotatedOldVal).addMBB(LoopMBB) 2447 .addReg(RotatedAltVal).addMBB(UseAltMBB); 2448 if (IsSubWord) 2449 BuildMI(MBB, DL, TII->get(SystemZ::RLL), NewVal) 2450 .addReg(RotatedNewVal).addReg(NegBitShift).addImm(0); 2451 BuildMI(MBB, DL, TII->get(CSOpcode), Dest) 2452 .addReg(OldVal).addReg(NewVal).addOperand(Base).addImm(Disp); 2453 BuildMI(MBB, DL, TII->get(SystemZ::BRC)) 2454 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB); 2455 MBB->addSuccessor(LoopMBB); 2456 MBB->addSuccessor(DoneMBB); 2457 2458 MI->eraseFromParent(); 2459 return DoneMBB; 2460} 2461 2462// Implement EmitInstrWithCustomInserter for pseudo ATOMIC_CMP_SWAPW 2463// instruction MI. 2464MachineBasicBlock * 2465SystemZTargetLowering::emitAtomicCmpSwapW(MachineInstr *MI, 2466 MachineBasicBlock *MBB) const { 2467 const SystemZInstrInfo *TII = TM.getInstrInfo(); 2468 MachineFunction &MF = *MBB->getParent(); 2469 MachineRegisterInfo &MRI = MF.getRegInfo(); 2470 2471 // Extract the operands. Base can be a register or a frame index. 2472 unsigned Dest = MI->getOperand(0).getReg(); 2473 MachineOperand Base = earlyUseOperand(MI->getOperand(1)); 2474 int64_t Disp = MI->getOperand(2).getImm(); 2475 unsigned OrigCmpVal = MI->getOperand(3).getReg(); 2476 unsigned OrigSwapVal = MI->getOperand(4).getReg(); 2477 unsigned BitShift = MI->getOperand(5).getReg(); 2478 unsigned NegBitShift = MI->getOperand(6).getReg(); 2479 int64_t BitSize = MI->getOperand(7).getImm(); 2480 DebugLoc DL = MI->getDebugLoc(); 2481 2482 const TargetRegisterClass *RC = &SystemZ::GR32BitRegClass; 2483 2484 // Get the right opcodes for the displacement. 2485 unsigned LOpcode = TII->getOpcodeForOffset(SystemZ::L, Disp); 2486 unsigned CSOpcode = TII->getOpcodeForOffset(SystemZ::CS, Disp); 2487 assert(LOpcode && CSOpcode && "Displacement out of range"); 2488 2489 // Create virtual registers for temporary results. 2490 unsigned OrigOldVal = MRI.createVirtualRegister(RC); 2491 unsigned OldVal = MRI.createVirtualRegister(RC); 2492 unsigned CmpVal = MRI.createVirtualRegister(RC); 2493 unsigned SwapVal = MRI.createVirtualRegister(RC); 2494 unsigned StoreVal = MRI.createVirtualRegister(RC); 2495 unsigned RetryOldVal = MRI.createVirtualRegister(RC); 2496 unsigned RetryCmpVal = MRI.createVirtualRegister(RC); 2497 unsigned RetrySwapVal = MRI.createVirtualRegister(RC); 2498 2499 // Insert 2 basic blocks for the loop. 2500 MachineBasicBlock *StartMBB = MBB; 2501 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB); 2502 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB); 2503 MachineBasicBlock *SetMBB = emitBlockAfter(LoopMBB); 2504 2505 // StartMBB: 2506 // ... 2507 // %OrigOldVal = L Disp(%Base) 2508 // # fall through to LoopMMB 2509 MBB = StartMBB; 2510 BuildMI(MBB, DL, TII->get(LOpcode), OrigOldVal) 2511 .addOperand(Base).addImm(Disp).addReg(0); 2512 MBB->addSuccessor(LoopMBB); 2513 2514 // LoopMBB: 2515 // %OldVal = phi [ %OrigOldVal, EntryBB ], [ %RetryOldVal, SetMBB ] 2516 // %CmpVal = phi [ %OrigCmpVal, EntryBB ], [ %RetryCmpVal, SetMBB ] 2517 // %SwapVal = phi [ %OrigSwapVal, EntryBB ], [ %RetrySwapVal, SetMBB ] 2518 // %Dest = RLL %OldVal, BitSize(%BitShift) 2519 // ^^ The low BitSize bits contain the field 2520 // of interest. 2521 // %RetryCmpVal = RISBG32 %CmpVal, %Dest, 32, 63-BitSize, 0 2522 // ^^ Replace the upper 32-BitSize bits of the 2523 // comparison value with those that we loaded, 2524 // so that we can use a full word comparison. 2525 // CR %Dest, %RetryCmpVal 2526 // JNE DoneMBB 2527 // # Fall through to SetMBB 2528 MBB = LoopMBB; 2529 BuildMI(MBB, DL, TII->get(SystemZ::PHI), OldVal) 2530 .addReg(OrigOldVal).addMBB(StartMBB) 2531 .addReg(RetryOldVal).addMBB(SetMBB); 2532 BuildMI(MBB, DL, TII->get(SystemZ::PHI), CmpVal) 2533 .addReg(OrigCmpVal).addMBB(StartMBB) 2534 .addReg(RetryCmpVal).addMBB(SetMBB); 2535 BuildMI(MBB, DL, TII->get(SystemZ::PHI), SwapVal) 2536 .addReg(OrigSwapVal).addMBB(StartMBB) 2537 .addReg(RetrySwapVal).addMBB(SetMBB); 2538 BuildMI(MBB, DL, TII->get(SystemZ::RLL), Dest) 2539 .addReg(OldVal).addReg(BitShift).addImm(BitSize); 2540 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetryCmpVal) 2541 .addReg(CmpVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0); 2542 BuildMI(MBB, DL, TII->get(SystemZ::CR)) 2543 .addReg(Dest).addReg(RetryCmpVal); 2544 BuildMI(MBB, DL, TII->get(SystemZ::BRC)) 2545 .addImm(SystemZ::CCMASK_ICMP) 2546 .addImm(SystemZ::CCMASK_CMP_NE).addMBB(DoneMBB); 2547 MBB->addSuccessor(DoneMBB); 2548 MBB->addSuccessor(SetMBB); 2549 2550 // SetMBB: 2551 // %RetrySwapVal = RISBG32 %SwapVal, %Dest, 32, 63-BitSize, 0 2552 // ^^ Replace the upper 32-BitSize bits of the new 2553 // value with those that we loaded. 2554 // %StoreVal = RLL %RetrySwapVal, -BitSize(%NegBitShift) 2555 // ^^ Rotate the new field to its proper position. 2556 // %RetryOldVal = CS %Dest, %StoreVal, Disp(%Base) 2557 // JNE LoopMBB 2558 // # fall through to ExitMMB 2559 MBB = SetMBB; 2560 BuildMI(MBB, DL, TII->get(SystemZ::RISBG32), RetrySwapVal) 2561 .addReg(SwapVal).addReg(Dest).addImm(32).addImm(63 - BitSize).addImm(0); 2562 BuildMI(MBB, DL, TII->get(SystemZ::RLL), StoreVal) 2563 .addReg(RetrySwapVal).addReg(NegBitShift).addImm(-BitSize); 2564 BuildMI(MBB, DL, TII->get(CSOpcode), RetryOldVal) 2565 .addReg(OldVal).addReg(StoreVal).addOperand(Base).addImm(Disp); 2566 BuildMI(MBB, DL, TII->get(SystemZ::BRC)) 2567 .addImm(SystemZ::CCMASK_CS).addImm(SystemZ::CCMASK_CS_NE).addMBB(LoopMBB); 2568 MBB->addSuccessor(LoopMBB); 2569 MBB->addSuccessor(DoneMBB); 2570 2571 MI->eraseFromParent(); 2572 return DoneMBB; 2573} 2574 2575// Emit an extension from a GR32 or GR64 to a GR128. ClearEven is true 2576// if the high register of the GR128 value must be cleared or false if 2577// it's "don't care". SubReg is subreg_odd32 when extending a GR32 2578// and subreg_odd when extending a GR64. 2579MachineBasicBlock * 2580SystemZTargetLowering::emitExt128(MachineInstr *MI, 2581 MachineBasicBlock *MBB, 2582 bool ClearEven, unsigned SubReg) const { 2583 const SystemZInstrInfo *TII = TM.getInstrInfo(); 2584 MachineFunction &MF = *MBB->getParent(); 2585 MachineRegisterInfo &MRI = MF.getRegInfo(); 2586 DebugLoc DL = MI->getDebugLoc(); 2587 2588 unsigned Dest = MI->getOperand(0).getReg(); 2589 unsigned Src = MI->getOperand(1).getReg(); 2590 unsigned In128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass); 2591 2592 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::IMPLICIT_DEF), In128); 2593 if (ClearEven) { 2594 unsigned NewIn128 = MRI.createVirtualRegister(&SystemZ::GR128BitRegClass); 2595 unsigned Zero64 = MRI.createVirtualRegister(&SystemZ::GR64BitRegClass); 2596 2597 BuildMI(*MBB, MI, DL, TII->get(SystemZ::LLILL), Zero64) 2598 .addImm(0); 2599 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), NewIn128) 2600 .addReg(In128).addReg(Zero64).addImm(SystemZ::subreg_high); 2601 In128 = NewIn128; 2602 } 2603 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::INSERT_SUBREG), Dest) 2604 .addReg(In128).addReg(Src).addImm(SubReg); 2605 2606 MI->eraseFromParent(); 2607 return MBB; 2608} 2609 2610MachineBasicBlock * 2611SystemZTargetLowering::emitMemMemWrapper(MachineInstr *MI, 2612 MachineBasicBlock *MBB, 2613 unsigned Opcode) const { 2614 const SystemZInstrInfo *TII = TM.getInstrInfo(); 2615 MachineFunction &MF = *MBB->getParent(); 2616 MachineRegisterInfo &MRI = MF.getRegInfo(); 2617 DebugLoc DL = MI->getDebugLoc(); 2618 2619 MachineOperand DestBase = earlyUseOperand(MI->getOperand(0)); 2620 uint64_t DestDisp = MI->getOperand(1).getImm(); 2621 MachineOperand SrcBase = earlyUseOperand(MI->getOperand(2)); 2622 uint64_t SrcDisp = MI->getOperand(3).getImm(); 2623 uint64_t Length = MI->getOperand(4).getImm(); 2624 2625 // When generating more than one CLC, all but the last will need to 2626 // branch to the end when a difference is found. 2627 MachineBasicBlock *EndMBB = (Length > 256 && Opcode == SystemZ::CLC ? 2628 splitBlockAfter(MI, MBB) : 0); 2629 2630 // Check for the loop form, in which operand 5 is the trip count. 2631 if (MI->getNumExplicitOperands() > 5) { 2632 bool HaveSingleBase = DestBase.isIdenticalTo(SrcBase); 2633 2634 uint64_t StartCountReg = MI->getOperand(5).getReg(); 2635 uint64_t StartSrcReg = forceReg(MI, SrcBase, TII); 2636 uint64_t StartDestReg = (HaveSingleBase ? StartSrcReg : 2637 forceReg(MI, DestBase, TII)); 2638 2639 const TargetRegisterClass *RC = &SystemZ::ADDR64BitRegClass; 2640 uint64_t ThisSrcReg = MRI.createVirtualRegister(RC); 2641 uint64_t ThisDestReg = (HaveSingleBase ? ThisSrcReg : 2642 MRI.createVirtualRegister(RC)); 2643 uint64_t NextSrcReg = MRI.createVirtualRegister(RC); 2644 uint64_t NextDestReg = (HaveSingleBase ? NextSrcReg : 2645 MRI.createVirtualRegister(RC)); 2646 2647 RC = &SystemZ::GR64BitRegClass; 2648 uint64_t ThisCountReg = MRI.createVirtualRegister(RC); 2649 uint64_t NextCountReg = MRI.createVirtualRegister(RC); 2650 2651 MachineBasicBlock *StartMBB = MBB; 2652 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB); 2653 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB); 2654 MachineBasicBlock *NextMBB = (EndMBB ? emitBlockAfter(LoopMBB) : LoopMBB); 2655 2656 // StartMBB: 2657 // # fall through to LoopMMB 2658 MBB->addSuccessor(LoopMBB); 2659 2660 // LoopMBB: 2661 // %ThisDestReg = phi [ %StartDestReg, StartMBB ], 2662 // [ %NextDestReg, NextMBB ] 2663 // %ThisSrcReg = phi [ %StartSrcReg, StartMBB ], 2664 // [ %NextSrcReg, NextMBB ] 2665 // %ThisCountReg = phi [ %StartCountReg, StartMBB ], 2666 // [ %NextCountReg, NextMBB ] 2667 // ( PFD 2, 768+DestDisp(%ThisDestReg) ) 2668 // Opcode DestDisp(256,%ThisDestReg), SrcDisp(%ThisSrcReg) 2669 // ( JLH EndMBB ) 2670 // 2671 // The prefetch is used only for MVC. The JLH is used only for CLC. 2672 MBB = LoopMBB; 2673 2674 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisDestReg) 2675 .addReg(StartDestReg).addMBB(StartMBB) 2676 .addReg(NextDestReg).addMBB(NextMBB); 2677 if (!HaveSingleBase) 2678 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisSrcReg) 2679 .addReg(StartSrcReg).addMBB(StartMBB) 2680 .addReg(NextSrcReg).addMBB(NextMBB); 2681 BuildMI(MBB, DL, TII->get(SystemZ::PHI), ThisCountReg) 2682 .addReg(StartCountReg).addMBB(StartMBB) 2683 .addReg(NextCountReg).addMBB(NextMBB); 2684 if (Opcode == SystemZ::MVC) 2685 BuildMI(MBB, DL, TII->get(SystemZ::PFD)) 2686 .addImm(SystemZ::PFD_WRITE) 2687 .addReg(ThisDestReg).addImm(DestDisp + 768).addReg(0); 2688 BuildMI(MBB, DL, TII->get(Opcode)) 2689 .addReg(ThisDestReg).addImm(DestDisp).addImm(256) 2690 .addReg(ThisSrcReg).addImm(SrcDisp); 2691 if (EndMBB) { 2692 BuildMI(MBB, DL, TII->get(SystemZ::BRC)) 2693 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE) 2694 .addMBB(EndMBB); 2695 MBB->addSuccessor(EndMBB); 2696 MBB->addSuccessor(NextMBB); 2697 } 2698 2699 // NextMBB: 2700 // %NextDestReg = LA 256(%ThisDestReg) 2701 // %NextSrcReg = LA 256(%ThisSrcReg) 2702 // %NextCountReg = AGHI %ThisCountReg, -1 2703 // CGHI %NextCountReg, 0 2704 // JLH LoopMBB 2705 // # fall through to DoneMMB 2706 // 2707 // The AGHI, CGHI and JLH should be converted to BRCTG by later passes. 2708 MBB = NextMBB; 2709 2710 BuildMI(MBB, DL, TII->get(SystemZ::LA), NextDestReg) 2711 .addReg(ThisDestReg).addImm(256).addReg(0); 2712 if (!HaveSingleBase) 2713 BuildMI(MBB, DL, TII->get(SystemZ::LA), NextSrcReg) 2714 .addReg(ThisSrcReg).addImm(256).addReg(0); 2715 BuildMI(MBB, DL, TII->get(SystemZ::AGHI), NextCountReg) 2716 .addReg(ThisCountReg).addImm(-1); 2717 BuildMI(MBB, DL, TII->get(SystemZ::CGHI)) 2718 .addReg(NextCountReg).addImm(0); 2719 BuildMI(MBB, DL, TII->get(SystemZ::BRC)) 2720 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE) 2721 .addMBB(LoopMBB); 2722 MBB->addSuccessor(LoopMBB); 2723 MBB->addSuccessor(DoneMBB); 2724 2725 DestBase = MachineOperand::CreateReg(NextDestReg, false); 2726 SrcBase = MachineOperand::CreateReg(NextSrcReg, false); 2727 Length &= 255; 2728 MBB = DoneMBB; 2729 } 2730 // Handle any remaining bytes with straight-line code. 2731 while (Length > 0) { 2732 uint64_t ThisLength = std::min(Length, uint64_t(256)); 2733 // The previous iteration might have created out-of-range displacements. 2734 // Apply them using LAY if so. 2735 if (!isUInt<12>(DestDisp)) { 2736 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass); 2737 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LAY), Reg) 2738 .addOperand(DestBase).addImm(DestDisp).addReg(0); 2739 DestBase = MachineOperand::CreateReg(Reg, false); 2740 DestDisp = 0; 2741 } 2742 if (!isUInt<12>(SrcDisp)) { 2743 unsigned Reg = MRI.createVirtualRegister(&SystemZ::ADDR64BitRegClass); 2744 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(SystemZ::LAY), Reg) 2745 .addOperand(SrcBase).addImm(SrcDisp).addReg(0); 2746 SrcBase = MachineOperand::CreateReg(Reg, false); 2747 SrcDisp = 0; 2748 } 2749 BuildMI(*MBB, MI, DL, TII->get(Opcode)) 2750 .addOperand(DestBase).addImm(DestDisp).addImm(ThisLength) 2751 .addOperand(SrcBase).addImm(SrcDisp); 2752 DestDisp += ThisLength; 2753 SrcDisp += ThisLength; 2754 Length -= ThisLength; 2755 // If there's another CLC to go, branch to the end if a difference 2756 // was found. 2757 if (EndMBB && Length > 0) { 2758 MachineBasicBlock *NextMBB = splitBlockBefore(MI, MBB); 2759 BuildMI(MBB, DL, TII->get(SystemZ::BRC)) 2760 .addImm(SystemZ::CCMASK_ICMP).addImm(SystemZ::CCMASK_CMP_NE) 2761 .addMBB(EndMBB); 2762 MBB->addSuccessor(EndMBB); 2763 MBB->addSuccessor(NextMBB); 2764 MBB = NextMBB; 2765 } 2766 } 2767 if (EndMBB) { 2768 MBB->addSuccessor(EndMBB); 2769 MBB = EndMBB; 2770 MBB->addLiveIn(SystemZ::CC); 2771 } 2772 2773 MI->eraseFromParent(); 2774 return MBB; 2775} 2776 2777// Decompose string pseudo-instruction MI into a loop that continually performs 2778// Opcode until CC != 3. 2779MachineBasicBlock * 2780SystemZTargetLowering::emitStringWrapper(MachineInstr *MI, 2781 MachineBasicBlock *MBB, 2782 unsigned Opcode) const { 2783 const SystemZInstrInfo *TII = TM.getInstrInfo(); 2784 MachineFunction &MF = *MBB->getParent(); 2785 MachineRegisterInfo &MRI = MF.getRegInfo(); 2786 DebugLoc DL = MI->getDebugLoc(); 2787 2788 uint64_t End1Reg = MI->getOperand(0).getReg(); 2789 uint64_t Start1Reg = MI->getOperand(1).getReg(); 2790 uint64_t Start2Reg = MI->getOperand(2).getReg(); 2791 uint64_t CharReg = MI->getOperand(3).getReg(); 2792 2793 const TargetRegisterClass *RC = &SystemZ::GR64BitRegClass; 2794 uint64_t This1Reg = MRI.createVirtualRegister(RC); 2795 uint64_t This2Reg = MRI.createVirtualRegister(RC); 2796 uint64_t End2Reg = MRI.createVirtualRegister(RC); 2797 2798 MachineBasicBlock *StartMBB = MBB; 2799 MachineBasicBlock *DoneMBB = splitBlockBefore(MI, MBB); 2800 MachineBasicBlock *LoopMBB = emitBlockAfter(StartMBB); 2801 2802 // StartMBB: 2803 // # fall through to LoopMMB 2804 MBB->addSuccessor(LoopMBB); 2805 2806 // LoopMBB: 2807 // %This1Reg = phi [ %Start1Reg, StartMBB ], [ %End1Reg, LoopMBB ] 2808 // %This2Reg = phi [ %Start2Reg, StartMBB ], [ %End2Reg, LoopMBB ] 2809 // R0W = %CharReg 2810 // %End1Reg, %End2Reg = CLST %This1Reg, %This2Reg -- uses R0W 2811 // JO LoopMBB 2812 // # fall through to DoneMMB 2813 // 2814 // The load of R0W can be hoisted by post-RA LICM. 2815 MBB = LoopMBB; 2816 2817 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This1Reg) 2818 .addReg(Start1Reg).addMBB(StartMBB) 2819 .addReg(End1Reg).addMBB(LoopMBB); 2820 BuildMI(MBB, DL, TII->get(SystemZ::PHI), This2Reg) 2821 .addReg(Start2Reg).addMBB(StartMBB) 2822 .addReg(End2Reg).addMBB(LoopMBB); 2823 BuildMI(MBB, DL, TII->get(TargetOpcode::COPY), SystemZ::R0W).addReg(CharReg); 2824 BuildMI(MBB, DL, TII->get(Opcode)) 2825 .addReg(End1Reg, RegState::Define).addReg(End2Reg, RegState::Define) 2826 .addReg(This1Reg).addReg(This2Reg); 2827 BuildMI(MBB, DL, TII->get(SystemZ::BRC)) 2828 .addImm(SystemZ::CCMASK_ANY).addImm(SystemZ::CCMASK_3).addMBB(LoopMBB); 2829 MBB->addSuccessor(LoopMBB); 2830 MBB->addSuccessor(DoneMBB); 2831 2832 DoneMBB->addLiveIn(SystemZ::CC); 2833 2834 MI->eraseFromParent(); 2835 return DoneMBB; 2836} 2837 2838MachineBasicBlock *SystemZTargetLowering:: 2839EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const { 2840 switch (MI->getOpcode()) { 2841 case SystemZ::Select32: 2842 case SystemZ::SelectF32: 2843 case SystemZ::Select64: 2844 case SystemZ::SelectF64: 2845 case SystemZ::SelectF128: 2846 return emitSelect(MI, MBB); 2847 2848 case SystemZ::CondStore8_32: 2849 return emitCondStore(MI, MBB, SystemZ::STC32, 0, false); 2850 case SystemZ::CondStore8_32Inv: 2851 return emitCondStore(MI, MBB, SystemZ::STC32, 0, true); 2852 case SystemZ::CondStore16_32: 2853 return emitCondStore(MI, MBB, SystemZ::STH32, 0, false); 2854 case SystemZ::CondStore16_32Inv: 2855 return emitCondStore(MI, MBB, SystemZ::STH32, 0, true); 2856 case SystemZ::CondStore32_32: 2857 return emitCondStore(MI, MBB, SystemZ::ST32, SystemZ::STOC32, false); 2858 case SystemZ::CondStore32_32Inv: 2859 return emitCondStore(MI, MBB, SystemZ::ST32, SystemZ::STOC32, true); 2860 case SystemZ::CondStore8: 2861 return emitCondStore(MI, MBB, SystemZ::STC, 0, false); 2862 case SystemZ::CondStore8Inv: 2863 return emitCondStore(MI, MBB, SystemZ::STC, 0, true); 2864 case SystemZ::CondStore16: 2865 return emitCondStore(MI, MBB, SystemZ::STH, 0, false); 2866 case SystemZ::CondStore16Inv: 2867 return emitCondStore(MI, MBB, SystemZ::STH, 0, true); 2868 case SystemZ::CondStore32: 2869 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, false); 2870 case SystemZ::CondStore32Inv: 2871 return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, true); 2872 case SystemZ::CondStore64: 2873 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, false); 2874 case SystemZ::CondStore64Inv: 2875 return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, true); 2876 case SystemZ::CondStoreF32: 2877 return emitCondStore(MI, MBB, SystemZ::STE, 0, false); 2878 case SystemZ::CondStoreF32Inv: 2879 return emitCondStore(MI, MBB, SystemZ::STE, 0, true); 2880 case SystemZ::CondStoreF64: 2881 return emitCondStore(MI, MBB, SystemZ::STD, 0, false); 2882 case SystemZ::CondStoreF64Inv: 2883 return emitCondStore(MI, MBB, SystemZ::STD, 0, true); 2884 2885 case SystemZ::AEXT128_64: 2886 return emitExt128(MI, MBB, false, SystemZ::subreg_low); 2887 case SystemZ::ZEXT128_32: 2888 return emitExt128(MI, MBB, true, SystemZ::subreg_low32); 2889 case SystemZ::ZEXT128_64: 2890 return emitExt128(MI, MBB, true, SystemZ::subreg_low); 2891 2892 case SystemZ::ATOMIC_SWAPW: 2893 return emitAtomicLoadBinary(MI, MBB, 0, 0); 2894 case SystemZ::ATOMIC_SWAP_32: 2895 return emitAtomicLoadBinary(MI, MBB, 0, 32); 2896 case SystemZ::ATOMIC_SWAP_64: 2897 return emitAtomicLoadBinary(MI, MBB, 0, 64); 2898 2899 case SystemZ::ATOMIC_LOADW_AR: 2900 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 0); 2901 case SystemZ::ATOMIC_LOADW_AFI: 2902 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 0); 2903 case SystemZ::ATOMIC_LOAD_AR: 2904 return emitAtomicLoadBinary(MI, MBB, SystemZ::AR, 32); 2905 case SystemZ::ATOMIC_LOAD_AHI: 2906 return emitAtomicLoadBinary(MI, MBB, SystemZ::AHI, 32); 2907 case SystemZ::ATOMIC_LOAD_AFI: 2908 return emitAtomicLoadBinary(MI, MBB, SystemZ::AFI, 32); 2909 case SystemZ::ATOMIC_LOAD_AGR: 2910 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGR, 64); 2911 case SystemZ::ATOMIC_LOAD_AGHI: 2912 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGHI, 64); 2913 case SystemZ::ATOMIC_LOAD_AGFI: 2914 return emitAtomicLoadBinary(MI, MBB, SystemZ::AGFI, 64); 2915 2916 case SystemZ::ATOMIC_LOADW_SR: 2917 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 0); 2918 case SystemZ::ATOMIC_LOAD_SR: 2919 return emitAtomicLoadBinary(MI, MBB, SystemZ::SR, 32); 2920 case SystemZ::ATOMIC_LOAD_SGR: 2921 return emitAtomicLoadBinary(MI, MBB, SystemZ::SGR, 64); 2922 2923 case SystemZ::ATOMIC_LOADW_NR: 2924 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0); 2925 case SystemZ::ATOMIC_LOADW_NILH: 2926 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH32, 0); 2927 case SystemZ::ATOMIC_LOAD_NR: 2928 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32); 2929 case SystemZ::ATOMIC_LOAD_NILL32: 2930 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL32, 32); 2931 case SystemZ::ATOMIC_LOAD_NILH32: 2932 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH32, 32); 2933 case SystemZ::ATOMIC_LOAD_NILF32: 2934 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF32, 32); 2935 case SystemZ::ATOMIC_LOAD_NGR: 2936 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64); 2937 case SystemZ::ATOMIC_LOAD_NILL: 2938 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 64); 2939 case SystemZ::ATOMIC_LOAD_NILH: 2940 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 64); 2941 case SystemZ::ATOMIC_LOAD_NIHL: 2942 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL, 64); 2943 case SystemZ::ATOMIC_LOAD_NIHH: 2944 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH, 64); 2945 case SystemZ::ATOMIC_LOAD_NILF: 2946 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 64); 2947 case SystemZ::ATOMIC_LOAD_NIHF: 2948 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF, 64); 2949 2950 case SystemZ::ATOMIC_LOADW_OR: 2951 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 0); 2952 case SystemZ::ATOMIC_LOADW_OILH: 2953 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH32, 0); 2954 case SystemZ::ATOMIC_LOAD_OR: 2955 return emitAtomicLoadBinary(MI, MBB, SystemZ::OR, 32); 2956 case SystemZ::ATOMIC_LOAD_OILL32: 2957 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL32, 32); 2958 case SystemZ::ATOMIC_LOAD_OILH32: 2959 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH32, 32); 2960 case SystemZ::ATOMIC_LOAD_OILF32: 2961 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF32, 32); 2962 case SystemZ::ATOMIC_LOAD_OGR: 2963 return emitAtomicLoadBinary(MI, MBB, SystemZ::OGR, 64); 2964 case SystemZ::ATOMIC_LOAD_OILL: 2965 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILL, 64); 2966 case SystemZ::ATOMIC_LOAD_OILH: 2967 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILH, 64); 2968 case SystemZ::ATOMIC_LOAD_OIHL: 2969 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHL, 64); 2970 case SystemZ::ATOMIC_LOAD_OIHH: 2971 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHH, 64); 2972 case SystemZ::ATOMIC_LOAD_OILF: 2973 return emitAtomicLoadBinary(MI, MBB, SystemZ::OILF, 64); 2974 case SystemZ::ATOMIC_LOAD_OIHF: 2975 return emitAtomicLoadBinary(MI, MBB, SystemZ::OIHF, 64); 2976 2977 case SystemZ::ATOMIC_LOADW_XR: 2978 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 0); 2979 case SystemZ::ATOMIC_LOADW_XILF: 2980 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF32, 0); 2981 case SystemZ::ATOMIC_LOAD_XR: 2982 return emitAtomicLoadBinary(MI, MBB, SystemZ::XR, 32); 2983 case SystemZ::ATOMIC_LOAD_XILF32: 2984 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF32, 32); 2985 case SystemZ::ATOMIC_LOAD_XGR: 2986 return emitAtomicLoadBinary(MI, MBB, SystemZ::XGR, 64); 2987 case SystemZ::ATOMIC_LOAD_XILF: 2988 return emitAtomicLoadBinary(MI, MBB, SystemZ::XILF, 64); 2989 case SystemZ::ATOMIC_LOAD_XIHF: 2990 return emitAtomicLoadBinary(MI, MBB, SystemZ::XIHF, 64); 2991 2992 case SystemZ::ATOMIC_LOADW_NRi: 2993 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 0, true); 2994 case SystemZ::ATOMIC_LOADW_NILHi: 2995 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH32, 0, true); 2996 case SystemZ::ATOMIC_LOAD_NRi: 2997 return emitAtomicLoadBinary(MI, MBB, SystemZ::NR, 32, true); 2998 case SystemZ::ATOMIC_LOAD_NILL32i: 2999 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL32, 32, true); 3000 case SystemZ::ATOMIC_LOAD_NILH32i: 3001 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH32, 32, true); 3002 case SystemZ::ATOMIC_LOAD_NILF32i: 3003 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF32, 32, true); 3004 case SystemZ::ATOMIC_LOAD_NGRi: 3005 return emitAtomicLoadBinary(MI, MBB, SystemZ::NGR, 64, true); 3006 case SystemZ::ATOMIC_LOAD_NILLi: 3007 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILL, 64, true); 3008 case SystemZ::ATOMIC_LOAD_NILHi: 3009 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILH, 64, true); 3010 case SystemZ::ATOMIC_LOAD_NIHLi: 3011 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHL, 64, true); 3012 case SystemZ::ATOMIC_LOAD_NIHHi: 3013 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHH, 64, true); 3014 case SystemZ::ATOMIC_LOAD_NILFi: 3015 return emitAtomicLoadBinary(MI, MBB, SystemZ::NILF, 64, true); 3016 case SystemZ::ATOMIC_LOAD_NIHFi: 3017 return emitAtomicLoadBinary(MI, MBB, SystemZ::NIHF, 64, true); 3018 3019 case SystemZ::ATOMIC_LOADW_MIN: 3020 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR, 3021 SystemZ::CCMASK_CMP_LE, 0); 3022 case SystemZ::ATOMIC_LOAD_MIN_32: 3023 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR, 3024 SystemZ::CCMASK_CMP_LE, 32); 3025 case SystemZ::ATOMIC_LOAD_MIN_64: 3026 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR, 3027 SystemZ::CCMASK_CMP_LE, 64); 3028 3029 case SystemZ::ATOMIC_LOADW_MAX: 3030 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR, 3031 SystemZ::CCMASK_CMP_GE, 0); 3032 case SystemZ::ATOMIC_LOAD_MAX_32: 3033 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CR, 3034 SystemZ::CCMASK_CMP_GE, 32); 3035 case SystemZ::ATOMIC_LOAD_MAX_64: 3036 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CGR, 3037 SystemZ::CCMASK_CMP_GE, 64); 3038 3039 case SystemZ::ATOMIC_LOADW_UMIN: 3040 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR, 3041 SystemZ::CCMASK_CMP_LE, 0); 3042 case SystemZ::ATOMIC_LOAD_UMIN_32: 3043 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR, 3044 SystemZ::CCMASK_CMP_LE, 32); 3045 case SystemZ::ATOMIC_LOAD_UMIN_64: 3046 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR, 3047 SystemZ::CCMASK_CMP_LE, 64); 3048 3049 case SystemZ::ATOMIC_LOADW_UMAX: 3050 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR, 3051 SystemZ::CCMASK_CMP_GE, 0); 3052 case SystemZ::ATOMIC_LOAD_UMAX_32: 3053 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLR, 3054 SystemZ::CCMASK_CMP_GE, 32); 3055 case SystemZ::ATOMIC_LOAD_UMAX_64: 3056 return emitAtomicLoadMinMax(MI, MBB, SystemZ::CLGR, 3057 SystemZ::CCMASK_CMP_GE, 64); 3058 3059 case SystemZ::ATOMIC_CMP_SWAPW: 3060 return emitAtomicCmpSwapW(MI, MBB); 3061 case SystemZ::MVCSequence: 3062 case SystemZ::MVCLoop: 3063 return emitMemMemWrapper(MI, MBB, SystemZ::MVC); 3064 case SystemZ::NCSequence: 3065 case SystemZ::NCLoop: 3066 return emitMemMemWrapper(MI, MBB, SystemZ::NC); 3067 case SystemZ::OCSequence: 3068 case SystemZ::OCLoop: 3069 return emitMemMemWrapper(MI, MBB, SystemZ::OC); 3070 case SystemZ::XCSequence: 3071 case SystemZ::XCLoop: 3072 return emitMemMemWrapper(MI, MBB, SystemZ::XC); 3073 case SystemZ::CLCSequence: 3074 case SystemZ::CLCLoop: 3075 return emitMemMemWrapper(MI, MBB, SystemZ::CLC); 3076 case SystemZ::CLSTLoop: 3077 return emitStringWrapper(MI, MBB, SystemZ::CLST); 3078 case SystemZ::MVSTLoop: 3079 return emitStringWrapper(MI, MBB, SystemZ::MVST); 3080 case SystemZ::SRSTLoop: 3081 return emitStringWrapper(MI, MBB, SystemZ::SRST); 3082 default: 3083 llvm_unreachable("Unexpected instr type to insert"); 3084 } 3085} 3086