1 //===-- SystemZISelLowering.h - SystemZ DAG lowering interface --*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file defines the interfaces that SystemZ uses to lower LLVM code into a 10 // selection DAG. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #ifndef LLVM_LIB_TARGET_SYSTEMZ_SYSTEMZISELLOWERING_H 15 #define LLVM_LIB_TARGET_SYSTEMZ_SYSTEMZISELLOWERING_H 16 17 #include "SystemZ.h" 18 #include "SystemZInstrInfo.h" 19 #include "llvm/CodeGen/MachineBasicBlock.h" 20 #include "llvm/CodeGen/SelectionDAG.h" 21 #include "llvm/CodeGen/TargetLowering.h" 22 #include <optional> 23 24 namespace llvm { 25 namespace SystemZISD { 26 enum NodeType : unsigned { 27 FIRST_NUMBER = ISD::BUILTIN_OP_END, 28 29 // Return with a glue operand. Operand 0 is the chain operand. 30 RET_GLUE, 31 32 // Calls a function. Operand 0 is the chain operand and operand 1 33 // is the target address. The arguments start at operand 2. 34 // There is an optional glue operand at the end. 35 CALL, 36 SIBCALL, 37 38 // TLS calls. Like regular calls, except operand 1 is the TLS symbol. 39 // (The call target is implicitly __tls_get_offset.) 40 TLS_GDCALL, 41 TLS_LDCALL, 42 43 // Wraps a TargetGlobalAddress that should be loaded using PC-relative 44 // accesses (LARL). Operand 0 is the address. 45 PCREL_WRAPPER, 46 47 // Used in cases where an offset is applied to a TargetGlobalAddress. 48 // Operand 0 is the full TargetGlobalAddress and operand 1 is a 49 // PCREL_WRAPPER for an anchor point. This is used so that we can 50 // cheaply refer to either the full address or the anchor point 51 // as a register base. 52 PCREL_OFFSET, 53 54 // Integer comparisons. There are three operands: the two values 55 // to compare, and an integer of type SystemZICMP. 56 ICMP, 57 58 // Floating-point comparisons. The two operands are the values to compare. 59 FCMP, 60 61 // Test under mask. The first operand is ANDed with the second operand 62 // and the condition codes are set on the result. The third operand is 63 // a boolean that is true if the condition codes need to distinguish 64 // between CCMASK_TM_MIXED_MSB_0 and CCMASK_TM_MIXED_MSB_1 (which the 65 // register forms do but the memory forms don't). 66 TM, 67 68 // Branches if a condition is true. Operand 0 is the chain operand; 69 // operand 1 is the 4-bit condition-code mask, with bit N in 70 // big-endian order meaning "branch if CC=N"; operand 2 is the 71 // target block and operand 3 is the flag operand. 72 BR_CCMASK, 73 74 // Selects between operand 0 and operand 1. Operand 2 is the 75 // mask of condition-code values for which operand 0 should be 76 // chosen over operand 1; it has the same form as BR_CCMASK. 77 // Operand 3 is the flag operand. 78 SELECT_CCMASK, 79 80 // Evaluates to the gap between the stack pointer and the 81 // base of the dynamically-allocatable area. 82 ADJDYNALLOC, 83 84 // For allocating stack space when using stack clash protector. 85 // Allocation is performed by block, and each block is probed. 86 PROBED_ALLOCA, 87 88 // Count number of bits set in operand 0 per byte. 89 POPCNT, 90 91 // Wrappers around the ISD opcodes of the same name. The output is GR128. 92 // Input operands may be GR64 or GR32, depending on the instruction. 93 SMUL_LOHI, 94 UMUL_LOHI, 95 SDIVREM, 96 UDIVREM, 97 98 // Add/subtract with overflow/carry. These have the same operands as 99 // the corresponding standard operations, except with the carry flag 100 // replaced by a condition code value. 101 SADDO, SSUBO, UADDO, USUBO, ADDCARRY, SUBCARRY, 102 103 // Set the condition code from a boolean value in operand 0. 104 // Operand 1 is a mask of all condition-code values that may result of this 105 // operation, operand 2 is a mask of condition-code values that may result 106 // if the boolean is true. 107 // Note that this operation is always optimized away, we will never 108 // generate any code for it. 109 GET_CCMASK, 110 111 // Use a series of MVCs to copy bytes from one memory location to another. 112 // The operands are: 113 // - the target address 114 // - the source address 115 // - the constant length 116 // 117 // This isn't a memory opcode because we'd need to attach two 118 // MachineMemOperands rather than one. 119 MVC, 120 121 // Similar to MVC, but for logic operations (AND, OR, XOR). 122 NC, 123 OC, 124 XC, 125 126 // Use CLC to compare two blocks of memory, with the same comments 127 // as for MVC. 128 CLC, 129 130 // Use MVC to set a block of memory after storing the first byte. 131 MEMSET_MVC, 132 133 // Use an MVST-based sequence to implement stpcpy(). 134 STPCPY, 135 136 // Use a CLST-based sequence to implement strcmp(). The two input operands 137 // are the addresses of the strings to compare. 138 STRCMP, 139 140 // Use an SRST-based sequence to search a block of memory. The first 141 // operand is the end address, the second is the start, and the third 142 // is the character to search for. CC is set to 1 on success and 2 143 // on failure. 144 SEARCH_STRING, 145 146 // Store the CC value in bits 29 and 28 of an integer. 147 IPM, 148 149 // Transaction begin. The first operand is the chain, the second 150 // the TDB pointer, and the third the immediate control field. 151 // Returns CC value and chain. 152 TBEGIN, 153 TBEGIN_NOFLOAT, 154 155 // Transaction end. Just the chain operand. Returns CC value and chain. 156 TEND, 157 158 // Create a vector constant by filling byte N of the result with bit 159 // 15-N of the single operand. 160 BYTE_MASK, 161 162 // Create a vector constant by replicating an element-sized RISBG-style mask. 163 // The first operand specifies the starting set bit and the second operand 164 // specifies the ending set bit. Both operands count from the MSB of the 165 // element. 166 ROTATE_MASK, 167 168 // Replicate a GPR scalar value into all elements of a vector. 169 REPLICATE, 170 171 // Create a vector from two i64 GPRs. 172 JOIN_DWORDS, 173 174 // Replicate one element of a vector into all elements. The first operand 175 // is the vector and the second is the index of the element to replicate. 176 SPLAT, 177 178 // Interleave elements from the high half of operand 0 and the high half 179 // of operand 1. 180 MERGE_HIGH, 181 182 // Likewise for the low halves. 183 MERGE_LOW, 184 185 // Concatenate the vectors in the first two operands, shift them left 186 // by the third operand, and take the first half of the result. 187 SHL_DOUBLE, 188 189 // Take one element of the first v2i64 operand and the one element of 190 // the second v2i64 operand and concatenate them to form a v2i64 result. 191 // The third operand is a 4-bit value of the form 0A0B, where A and B 192 // are the element selectors for the first operand and second operands 193 // respectively. 194 PERMUTE_DWORDS, 195 196 // Perform a general vector permute on vector operands 0 and 1. 197 // Each byte of operand 2 controls the corresponding byte of the result, 198 // in the same way as a byte-level VECTOR_SHUFFLE mask. 199 PERMUTE, 200 201 // Pack vector operands 0 and 1 into a single vector with half-sized elements. 202 PACK, 203 204 // Likewise, but saturate the result and set CC. PACKS_CC does signed 205 // saturation and PACKLS_CC does unsigned saturation. 206 PACKS_CC, 207 PACKLS_CC, 208 209 // Unpack the first half of vector operand 0 into double-sized elements. 210 // UNPACK_HIGH sign-extends and UNPACKL_HIGH zero-extends. 211 UNPACK_HIGH, 212 UNPACKL_HIGH, 213 214 // Likewise for the second half. 215 UNPACK_LOW, 216 UNPACKL_LOW, 217 218 // Shift/rotate each element of vector operand 0 by the number of bits 219 // specified by scalar operand 1. 220 VSHL_BY_SCALAR, 221 VSRL_BY_SCALAR, 222 VSRA_BY_SCALAR, 223 VROTL_BY_SCALAR, 224 225 // For each element of the output type, sum across all sub-elements of 226 // operand 0 belonging to the corresponding element, and add in the 227 // rightmost sub-element of the corresponding element of operand 1. 228 VSUM, 229 230 // Compute carry/borrow indication for add/subtract. 231 VACC, VSCBI, 232 // Add/subtract with carry/borrow. 233 VAC, VSBI, 234 // Compute carry/borrow indication for add/subtract with carry/borrow. 235 VACCC, VSBCBI, 236 237 // Compare integer vector operands 0 and 1 to produce the usual 0/-1 238 // vector result. VICMPE is for equality, VICMPH for "signed greater than" 239 // and VICMPHL for "unsigned greater than". 240 VICMPE, 241 VICMPH, 242 VICMPHL, 243 244 // Likewise, but also set the condition codes on the result. 245 VICMPES, 246 VICMPHS, 247 VICMPHLS, 248 249 // Compare floating-point vector operands 0 and 1 to produce the usual 0/-1 250 // vector result. VFCMPE is for "ordered and equal", VFCMPH for "ordered and 251 // greater than" and VFCMPHE for "ordered and greater than or equal to". 252 VFCMPE, 253 VFCMPH, 254 VFCMPHE, 255 256 // Likewise, but also set the condition codes on the result. 257 VFCMPES, 258 VFCMPHS, 259 VFCMPHES, 260 261 // Test floating-point data class for vectors. 262 VFTCI, 263 264 // Extend the even f32 elements of vector operand 0 to produce a vector 265 // of f64 elements. 266 VEXTEND, 267 268 // Round the f64 elements of vector operand 0 to f32s and store them in the 269 // even elements of the result. 270 VROUND, 271 272 // AND the two vector operands together and set CC based on the result. 273 VTM, 274 275 // i128 high integer comparisons. 276 SCMP128HI, 277 UCMP128HI, 278 279 // String operations that set CC as a side-effect. 280 VFAE_CC, 281 VFAEZ_CC, 282 VFEE_CC, 283 VFEEZ_CC, 284 VFENE_CC, 285 VFENEZ_CC, 286 VISTR_CC, 287 VSTRC_CC, 288 VSTRCZ_CC, 289 VSTRS_CC, 290 VSTRSZ_CC, 291 292 // Test Data Class. 293 // 294 // Operand 0: the value to test 295 // Operand 1: the bit mask 296 TDC, 297 298 // z/OS XPLINK ADA Entry 299 // Wraps a TargetGlobalAddress that should be loaded from a function's 300 // AssociatedData Area (ADA). Tha ADA is passed to the function by the 301 // caller in the XPLink ABI defined register R5. 302 // Operand 0: the GlobalValue/External Symbol 303 // Operand 1: the ADA register 304 // Operand 2: the offset (0 for the first and 8 for the second element in the 305 // function descriptor) 306 ADA_ENTRY, 307 308 // Strict variants of scalar floating-point comparisons. 309 // Quiet and signaling versions. 310 STRICT_FCMP = ISD::FIRST_TARGET_STRICTFP_OPCODE, 311 STRICT_FCMPS, 312 313 // Strict variants of vector floating-point comparisons. 314 // Quiet and signaling versions. 315 STRICT_VFCMPE, 316 STRICT_VFCMPH, 317 STRICT_VFCMPHE, 318 STRICT_VFCMPES, 319 STRICT_VFCMPHS, 320 STRICT_VFCMPHES, 321 322 // Strict variants of VEXTEND and VROUND. 323 STRICT_VEXTEND, 324 STRICT_VROUND, 325 326 // Wrappers around the inner loop of an 8- or 16-bit ATOMIC_SWAP or 327 // ATOMIC_LOAD_<op>. 328 // 329 // Operand 0: the address of the containing 32-bit-aligned field 330 // Operand 1: the second operand of <op>, in the high bits of an i32 331 // for everything except ATOMIC_SWAPW 332 // Operand 2: how many bits to rotate the i32 left to bring the first 333 // operand into the high bits 334 // Operand 3: the negative of operand 2, for rotating the other way 335 // Operand 4: the width of the field in bits (8 or 16) 336 ATOMIC_SWAPW = ISD::FIRST_TARGET_MEMORY_OPCODE, 337 ATOMIC_LOADW_ADD, 338 ATOMIC_LOADW_SUB, 339 ATOMIC_LOADW_AND, 340 ATOMIC_LOADW_OR, 341 ATOMIC_LOADW_XOR, 342 ATOMIC_LOADW_NAND, 343 ATOMIC_LOADW_MIN, 344 ATOMIC_LOADW_MAX, 345 ATOMIC_LOADW_UMIN, 346 ATOMIC_LOADW_UMAX, 347 348 // A wrapper around the inner loop of an ATOMIC_CMP_SWAP. 349 // 350 // Operand 0: the address of the containing 32-bit-aligned field 351 // Operand 1: the compare value, in the low bits of an i32 352 // Operand 2: the swap value, in the low bits of an i32 353 // Operand 3: how many bits to rotate the i32 left to bring the first 354 // operand into the high bits 355 // Operand 4: the negative of operand 2, for rotating the other way 356 // Operand 5: the width of the field in bits (8 or 16) 357 ATOMIC_CMP_SWAPW, 358 359 // Atomic compare-and-swap returning CC value. 360 // Val, CC, OUTCHAIN = ATOMIC_CMP_SWAP(INCHAIN, ptr, cmp, swap) 361 ATOMIC_CMP_SWAP, 362 363 // 128-bit atomic load. 364 // Val, OUTCHAIN = ATOMIC_LOAD_128(INCHAIN, ptr) 365 ATOMIC_LOAD_128, 366 367 // 128-bit atomic store. 368 // OUTCHAIN = ATOMIC_STORE_128(INCHAIN, val, ptr) 369 ATOMIC_STORE_128, 370 371 // 128-bit atomic compare-and-swap. 372 // Val, CC, OUTCHAIN = ATOMIC_CMP_SWAP(INCHAIN, ptr, cmp, swap) 373 ATOMIC_CMP_SWAP_128, 374 375 // Byte swapping load/store. Same operands as regular load/store. 376 LRV, STRV, 377 378 // Element swapping load/store. Same operands as regular load/store. 379 VLER, VSTER, 380 381 // Prefetch from the second operand using the 4-bit control code in 382 // the first operand. The code is 1 for a load prefetch and 2 for 383 // a store prefetch. 384 PREFETCH 385 }; 386 387 // Return true if OPCODE is some kind of PC-relative address. 388 inline bool isPCREL(unsigned Opcode) { 389 return Opcode == PCREL_WRAPPER || Opcode == PCREL_OFFSET; 390 } 391 } // end namespace SystemZISD 392 393 namespace SystemZICMP { 394 // Describes whether an integer comparison needs to be signed or unsigned, 395 // or whether either type is OK. 396 enum { 397 Any, 398 UnsignedOnly, 399 SignedOnly 400 }; 401 } // end namespace SystemZICMP 402 403 class SystemZSubtarget; 404 405 class SystemZTargetLowering : public TargetLowering { 406 public: 407 explicit SystemZTargetLowering(const TargetMachine &TM, 408 const SystemZSubtarget &STI); 409 410 bool useSoftFloat() const override; 411 412 // Override TargetLowering. 413 MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override { 414 return MVT::i32; 415 } 416 MVT getVectorIdxTy(const DataLayout &DL) const override { 417 // Only the lower 12 bits of an element index are used, so we don't 418 // want to clobber the upper 32 bits of a GPR unnecessarily. 419 return MVT::i32; 420 } 421 TargetLoweringBase::LegalizeTypeAction getPreferredVectorAction(MVT VT) 422 const override { 423 // Widen subvectors to the full width rather than promoting integer 424 // elements. This is better because: 425 // 426 // (a) it means that we can handle the ABI for passing and returning 427 // sub-128 vectors without having to handle them as legal types. 428 // 429 // (b) we don't have instructions to extend on load and truncate on store, 430 // so promoting the integers is less efficient. 431 // 432 // (c) there are no multiplication instructions for the widest integer 433 // type (v2i64). 434 if (VT.getScalarSizeInBits() % 8 == 0) 435 return TypeWidenVector; 436 return TargetLoweringBase::getPreferredVectorAction(VT); 437 } 438 unsigned 439 getNumRegisters(LLVMContext &Context, EVT VT, 440 std::optional<MVT> RegisterVT) const override { 441 // i128 inline assembly operand. 442 if (VT == MVT::i128 && RegisterVT && *RegisterVT == MVT::Untyped) 443 return 1; 444 return TargetLowering::getNumRegisters(Context, VT); 445 } 446 MVT getRegisterTypeForCallingConv(LLVMContext &Context, CallingConv::ID CC, 447 EVT VT) const override { 448 // 128-bit single-element vector types are passed like other vectors, 449 // not like their element type. 450 if (VT.isVector() && VT.getSizeInBits() == 128 && 451 VT.getVectorNumElements() == 1) 452 return MVT::v16i8; 453 return TargetLowering::getRegisterTypeForCallingConv(Context, CC, VT); 454 } 455 bool isCheapToSpeculateCtlz(Type *) const override { return true; } 456 bool isCheapToSpeculateCttz(Type *) const override { return true; } 457 bool preferZeroCompareBranch() const override { return true; } 458 bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override { 459 ConstantInt* Mask = dyn_cast<ConstantInt>(AndI.getOperand(1)); 460 return Mask && Mask->getValue().isIntN(16); 461 } 462 bool convertSetCCLogicToBitwiseLogic(EVT VT) const override { 463 return VT.isScalarInteger(); 464 } 465 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &, 466 EVT) const override; 467 bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF, 468 EVT VT) const override; 469 bool isFPImmLegal(const APFloat &Imm, EVT VT, 470 bool ForCodeSize) const override; 471 bool ShouldShrinkFPConstant(EVT VT) const override { 472 // Do not shrink 64-bit FP constpool entries since LDEB is slower than 473 // LD, and having the full constant in memory enables reg/mem opcodes. 474 return VT != MVT::f64; 475 } 476 bool hasInlineStackProbe(const MachineFunction &MF) const override; 477 AtomicExpansionKind 478 shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const override; 479 bool isLegalICmpImmediate(int64_t Imm) const override; 480 bool isLegalAddImmediate(int64_t Imm) const override; 481 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty, 482 unsigned AS, 483 Instruction *I = nullptr) const override; 484 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AS, Align Alignment, 485 MachineMemOperand::Flags Flags, 486 unsigned *Fast) const override; 487 bool 488 findOptimalMemOpLowering(std::vector<EVT> &MemOps, unsigned Limit, 489 const MemOp &Op, unsigned DstAS, unsigned SrcAS, 490 const AttributeList &FuncAttributes) const override; 491 EVT getOptimalMemOpType(const MemOp &Op, 492 const AttributeList &FuncAttributes) const override; 493 bool isTruncateFree(Type *, Type *) const override; 494 bool isTruncateFree(EVT, EVT) const override; 495 496 bool shouldFormOverflowOp(unsigned Opcode, EVT VT, 497 bool MathUsed) const override { 498 // Form add and sub with overflow intrinsics regardless of any extra 499 // users of the math result. 500 return VT == MVT::i32 || VT == MVT::i64; 501 } 502 503 bool shouldConsiderGEPOffsetSplit() const override { return true; } 504 505 const char *getTargetNodeName(unsigned Opcode) const override; 506 std::pair<unsigned, const TargetRegisterClass *> 507 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, 508 StringRef Constraint, MVT VT) const override; 509 TargetLowering::ConstraintType 510 getConstraintType(StringRef Constraint) const override; 511 TargetLowering::ConstraintWeight 512 getSingleConstraintMatchWeight(AsmOperandInfo &info, 513 const char *constraint) const override; 514 void LowerAsmOperandForConstraint(SDValue Op, StringRef Constraint, 515 std::vector<SDValue> &Ops, 516 SelectionDAG &DAG) const override; 517 518 InlineAsm::ConstraintCode 519 getInlineAsmMemConstraint(StringRef ConstraintCode) const override { 520 if (ConstraintCode.size() == 1) { 521 switch(ConstraintCode[0]) { 522 default: 523 break; 524 case 'o': 525 return InlineAsm::ConstraintCode::o; 526 case 'Q': 527 return InlineAsm::ConstraintCode::Q; 528 case 'R': 529 return InlineAsm::ConstraintCode::R; 530 case 'S': 531 return InlineAsm::ConstraintCode::S; 532 case 'T': 533 return InlineAsm::ConstraintCode::T; 534 } 535 } else if (ConstraintCode.size() == 2 && ConstraintCode[0] == 'Z') { 536 switch (ConstraintCode[1]) { 537 default: 538 break; 539 case 'Q': 540 return InlineAsm::ConstraintCode::ZQ; 541 case 'R': 542 return InlineAsm::ConstraintCode::ZR; 543 case 'S': 544 return InlineAsm::ConstraintCode::ZS; 545 case 'T': 546 return InlineAsm::ConstraintCode::ZT; 547 } 548 } 549 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode); 550 } 551 552 Register getRegisterByName(const char *RegName, LLT VT, 553 const MachineFunction &MF) const override; 554 555 /// If a physical register, this returns the register that receives the 556 /// exception address on entry to an EH pad. 557 Register 558 getExceptionPointerRegister(const Constant *PersonalityFn) const override; 559 560 /// If a physical register, this returns the register that receives the 561 /// exception typeid on entry to a landing pad. 562 Register 563 getExceptionSelectorRegister(const Constant *PersonalityFn) const override; 564 565 /// Override to support customized stack guard loading. 566 bool useLoadStackGuardNode() const override { 567 return true; 568 } 569 void insertSSPDeclarations(Module &M) const override { 570 } 571 572 MachineBasicBlock * 573 EmitInstrWithCustomInserter(MachineInstr &MI, 574 MachineBasicBlock *BB) const override; 575 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override; 576 void LowerOperationWrapper(SDNode *N, SmallVectorImpl<SDValue> &Results, 577 SelectionDAG &DAG) const override; 578 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results, 579 SelectionDAG &DAG) const override; 580 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override; 581 bool allowTruncateForTailCall(Type *, Type *) const override; 582 bool mayBeEmittedAsTailCall(const CallInst *CI) const override; 583 bool splitValueIntoRegisterParts( 584 SelectionDAG & DAG, const SDLoc &DL, SDValue Val, SDValue *Parts, 585 unsigned NumParts, MVT PartVT, std::optional<CallingConv::ID> CC) 586 const override; 587 SDValue joinRegisterPartsIntoValue( 588 SelectionDAG & DAG, const SDLoc &DL, const SDValue *Parts, 589 unsigned NumParts, MVT PartVT, EVT ValueVT, 590 std::optional<CallingConv::ID> CC) const override; 591 SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, 592 bool isVarArg, 593 const SmallVectorImpl<ISD::InputArg> &Ins, 594 const SDLoc &DL, SelectionDAG &DAG, 595 SmallVectorImpl<SDValue> &InVals) const override; 596 SDValue LowerCall(CallLoweringInfo &CLI, 597 SmallVectorImpl<SDValue> &InVals) const override; 598 599 std::pair<SDValue, SDValue> 600 makeExternalCall(SDValue Chain, SelectionDAG &DAG, const char *CalleeName, 601 EVT RetVT, ArrayRef<SDValue> Ops, CallingConv::ID CallConv, 602 bool IsSigned, SDLoc DL, bool DoesNotReturn, 603 bool IsReturnValueUsed) const; 604 605 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF, 606 bool isVarArg, 607 const SmallVectorImpl<ISD::OutputArg> &Outs, 608 LLVMContext &Context) const override; 609 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, 610 const SmallVectorImpl<ISD::OutputArg> &Outs, 611 const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL, 612 SelectionDAG &DAG) const override; 613 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override; 614 615 /// Determine which of the bits specified in Mask are known to be either 616 /// zero or one and return them in the KnownZero/KnownOne bitsets. 617 void computeKnownBitsForTargetNode(const SDValue Op, 618 KnownBits &Known, 619 const APInt &DemandedElts, 620 const SelectionDAG &DAG, 621 unsigned Depth = 0) const override; 622 623 /// Determine the number of bits in the operation that are sign bits. 624 unsigned ComputeNumSignBitsForTargetNode(SDValue Op, 625 const APInt &DemandedElts, 626 const SelectionDAG &DAG, 627 unsigned Depth) const override; 628 629 bool isGuaranteedNotToBeUndefOrPoisonForTargetNode( 630 SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG, 631 bool PoisonOnly, unsigned Depth) const override; 632 633 ISD::NodeType getExtendForAtomicOps() const override { 634 return ISD::ANY_EXTEND; 635 } 636 ISD::NodeType getExtendForAtomicCmpSwapArg() const override { 637 return ISD::ZERO_EXTEND; 638 } 639 640 bool supportSwiftError() const override { 641 return true; 642 } 643 644 unsigned getStackProbeSize(const MachineFunction &MF) const; 645 646 private: 647 const SystemZSubtarget &Subtarget; 648 649 // Implement LowerOperation for individual opcodes. 650 SDValue getVectorCmp(SelectionDAG &DAG, unsigned Opcode, 651 const SDLoc &DL, EVT VT, 652 SDValue CmpOp0, SDValue CmpOp1, SDValue Chain) const; 653 SDValue lowerVectorSETCC(SelectionDAG &DAG, const SDLoc &DL, 654 EVT VT, ISD::CondCode CC, 655 SDValue CmpOp0, SDValue CmpOp1, 656 SDValue Chain = SDValue(), 657 bool IsSignaling = false) const; 658 SDValue lowerSETCC(SDValue Op, SelectionDAG &DAG) const; 659 SDValue lowerSTRICT_FSETCC(SDValue Op, SelectionDAG &DAG, 660 bool IsSignaling) const; 661 SDValue lowerBR_CC(SDValue Op, SelectionDAG &DAG) const; 662 SDValue lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const; 663 SDValue lowerGlobalAddress(GlobalAddressSDNode *Node, 664 SelectionDAG &DAG) const; 665 SDValue lowerTLSGetOffset(GlobalAddressSDNode *Node, 666 SelectionDAG &DAG, unsigned Opcode, 667 SDValue GOTOffset) const; 668 SDValue lowerThreadPointer(const SDLoc &DL, SelectionDAG &DAG) const; 669 SDValue lowerGlobalTLSAddress(GlobalAddressSDNode *Node, 670 SelectionDAG &DAG) const; 671 SDValue lowerBlockAddress(BlockAddressSDNode *Node, 672 SelectionDAG &DAG) const; 673 SDValue lowerJumpTable(JumpTableSDNode *JT, SelectionDAG &DAG) const; 674 SDValue lowerConstantPool(ConstantPoolSDNode *CP, SelectionDAG &DAG) const; 675 SDValue lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const; 676 SDValue lowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const; 677 SDValue lowerVASTART(SDValue Op, SelectionDAG &DAG) const; 678 SDValue lowerVASTART_ELF(SDValue Op, SelectionDAG &DAG) const; 679 SDValue lowerVASTART_XPLINK(SDValue Op, SelectionDAG &DAG) const; 680 SDValue lowerVACOPY(SDValue Op, SelectionDAG &DAG) const; 681 SDValue lowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const; 682 SDValue lowerDYNAMIC_STACKALLOC_ELF(SDValue Op, SelectionDAG &DAG) const; 683 SDValue lowerDYNAMIC_STACKALLOC_XPLINK(SDValue Op, SelectionDAG &DAG) const; 684 SDValue lowerGET_DYNAMIC_AREA_OFFSET(SDValue Op, SelectionDAG &DAG) const; 685 SDValue lowerSMUL_LOHI(SDValue Op, SelectionDAG &DAG) const; 686 SDValue lowerUMUL_LOHI(SDValue Op, SelectionDAG &DAG) const; 687 SDValue lowerSDIVREM(SDValue Op, SelectionDAG &DAG) const; 688 SDValue lowerUDIVREM(SDValue Op, SelectionDAG &DAG) const; 689 SDValue lowerXALUO(SDValue Op, SelectionDAG &DAG) const; 690 SDValue lowerUADDSUBO_CARRY(SDValue Op, SelectionDAG &DAG) const; 691 SDValue lowerBITCAST(SDValue Op, SelectionDAG &DAG) const; 692 SDValue lowerOR(SDValue Op, SelectionDAG &DAG) const; 693 SDValue lowerCTPOP(SDValue Op, SelectionDAG &DAG) const; 694 SDValue lowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG) const; 695 SDValue lowerATOMIC_LOAD(SDValue Op, SelectionDAG &DAG) const; 696 SDValue lowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) const; 697 SDValue lowerATOMIC_LOAD_OP(SDValue Op, SelectionDAG &DAG, 698 unsigned Opcode) const; 699 SDValue lowerATOMIC_LOAD_SUB(SDValue Op, SelectionDAG &DAG) const; 700 SDValue lowerATOMIC_CMP_SWAP(SDValue Op, SelectionDAG &DAG) const; 701 SDValue lowerSTACKSAVE(SDValue Op, SelectionDAG &DAG) const; 702 SDValue lowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG) const; 703 SDValue lowerPREFETCH(SDValue Op, SelectionDAG &DAG) const; 704 SDValue lowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) const; 705 SDValue lowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const; 706 bool isVectorElementLoad(SDValue Op) const; 707 SDValue buildVector(SelectionDAG &DAG, const SDLoc &DL, EVT VT, 708 SmallVectorImpl<SDValue> &Elems) const; 709 SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const; 710 SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const; 711 SDValue lowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const; 712 SDValue lowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const; 713 SDValue lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const; 714 SDValue lowerSIGN_EXTEND_VECTOR_INREG(SDValue Op, SelectionDAG &DAG) const; 715 SDValue lowerZERO_EXTEND_VECTOR_INREG(SDValue Op, SelectionDAG &DAG) const; 716 SDValue lowerShift(SDValue Op, SelectionDAG &DAG, unsigned ByScalar) const; 717 SDValue lowerIS_FPCLASS(SDValue Op, SelectionDAG &DAG) const; 718 SDValue lowerGET_ROUNDING(SDValue Op, SelectionDAG &DAG) const; 719 720 bool canTreatAsByteVector(EVT VT) const; 721 SDValue combineExtract(const SDLoc &DL, EVT ElemVT, EVT VecVT, SDValue OrigOp, 722 unsigned Index, DAGCombinerInfo &DCI, 723 bool Force) const; 724 SDValue combineTruncateExtract(const SDLoc &DL, EVT TruncVT, SDValue Op, 725 DAGCombinerInfo &DCI) const; 726 SDValue combineZERO_EXTEND(SDNode *N, DAGCombinerInfo &DCI) const; 727 SDValue combineSIGN_EXTEND(SDNode *N, DAGCombinerInfo &DCI) const; 728 SDValue combineSIGN_EXTEND_INREG(SDNode *N, DAGCombinerInfo &DCI) const; 729 SDValue combineMERGE(SDNode *N, DAGCombinerInfo &DCI) const; 730 bool canLoadStoreByteSwapped(EVT VT) const; 731 SDValue combineLOAD(SDNode *N, DAGCombinerInfo &DCI) const; 732 SDValue combineSTORE(SDNode *N, DAGCombinerInfo &DCI) const; 733 SDValue combineVECTOR_SHUFFLE(SDNode *N, DAGCombinerInfo &DCI) const; 734 SDValue combineEXTRACT_VECTOR_ELT(SDNode *N, DAGCombinerInfo &DCI) const; 735 SDValue combineJOIN_DWORDS(SDNode *N, DAGCombinerInfo &DCI) const; 736 SDValue combineFP_ROUND(SDNode *N, DAGCombinerInfo &DCI) const; 737 SDValue combineFP_EXTEND(SDNode *N, DAGCombinerInfo &DCI) const; 738 SDValue combineINT_TO_FP(SDNode *N, DAGCombinerInfo &DCI) const; 739 SDValue combineBSWAP(SDNode *N, DAGCombinerInfo &DCI) const; 740 SDValue combineBR_CCMASK(SDNode *N, DAGCombinerInfo &DCI) const; 741 SDValue combineSELECT_CCMASK(SDNode *N, DAGCombinerInfo &DCI) const; 742 SDValue combineGET_CCMASK(SDNode *N, DAGCombinerInfo &DCI) const; 743 SDValue combineIntDIVREM(SDNode *N, DAGCombinerInfo &DCI) const; 744 SDValue combineINTRINSIC(SDNode *N, DAGCombinerInfo &DCI) const; 745 746 SDValue unwrapAddress(SDValue N) const override; 747 748 // If the last instruction before MBBI in MBB was some form of COMPARE, 749 // try to replace it with a COMPARE AND BRANCH just before MBBI. 750 // CCMask and Target are the BRC-like operands for the branch. 751 // Return true if the change was made. 752 bool convertPrevCompareToBranch(MachineBasicBlock *MBB, 753 MachineBasicBlock::iterator MBBI, 754 unsigned CCMask, 755 MachineBasicBlock *Target) const; 756 757 // Implement EmitInstrWithCustomInserter for individual operation types. 758 MachineBasicBlock *emitSelect(MachineInstr &MI, MachineBasicBlock *BB) const; 759 MachineBasicBlock *emitCondStore(MachineInstr &MI, MachineBasicBlock *BB, 760 unsigned StoreOpcode, unsigned STOCOpcode, 761 bool Invert) const; 762 MachineBasicBlock *emitICmp128Hi(MachineInstr &MI, MachineBasicBlock *BB, 763 bool Unsigned) const; 764 MachineBasicBlock *emitPair128(MachineInstr &MI, 765 MachineBasicBlock *MBB) const; 766 MachineBasicBlock *emitExt128(MachineInstr &MI, MachineBasicBlock *MBB, 767 bool ClearEven) const; 768 MachineBasicBlock *emitAtomicLoadBinary(MachineInstr &MI, 769 MachineBasicBlock *BB, 770 unsigned BinOpcode, 771 bool Invert = false) const; 772 MachineBasicBlock *emitAtomicLoadMinMax(MachineInstr &MI, 773 MachineBasicBlock *MBB, 774 unsigned CompareOpcode, 775 unsigned KeepOldMask) const; 776 MachineBasicBlock *emitAtomicCmpSwapW(MachineInstr &MI, 777 MachineBasicBlock *BB) const; 778 MachineBasicBlock *emitMemMemWrapper(MachineInstr &MI, MachineBasicBlock *BB, 779 unsigned Opcode, 780 bool IsMemset = false) const; 781 MachineBasicBlock *emitStringWrapper(MachineInstr &MI, MachineBasicBlock *BB, 782 unsigned Opcode) const; 783 MachineBasicBlock *emitTransactionBegin(MachineInstr &MI, 784 MachineBasicBlock *MBB, 785 unsigned Opcode, bool NoFloat) const; 786 MachineBasicBlock *emitLoadAndTestCmp0(MachineInstr &MI, 787 MachineBasicBlock *MBB, 788 unsigned Opcode) const; 789 MachineBasicBlock *emitProbedAlloca(MachineInstr &MI, 790 MachineBasicBlock *MBB) const; 791 792 SDValue getBackchainAddress(SDValue SP, SelectionDAG &DAG) const; 793 794 MachineMemOperand::Flags 795 getTargetMMOFlags(const Instruction &I) const override; 796 const TargetRegisterClass *getRepRegClassFor(MVT VT) const override; 797 }; 798 799 struct SystemZVectorConstantInfo { 800 private: 801 APInt IntBits; // The 128 bits as an integer. 802 APInt SplatBits; // Smallest splat value. 803 APInt SplatUndef; // Bits correspoding to undef operands of the BVN. 804 unsigned SplatBitSize = 0; 805 bool isFP128 = false; 806 public: 807 unsigned Opcode = 0; 808 SmallVector<unsigned, 2> OpVals; 809 MVT VecVT; 810 SystemZVectorConstantInfo(APInt IntImm); 811 SystemZVectorConstantInfo(APFloat FPImm) 812 : SystemZVectorConstantInfo(FPImm.bitcastToAPInt()) { 813 isFP128 = (&FPImm.getSemantics() == &APFloat::IEEEquad()); 814 } 815 SystemZVectorConstantInfo(BuildVectorSDNode *BVN); 816 bool isVectorConstantLegal(const SystemZSubtarget &Subtarget); 817 }; 818 819 } // end namespace llvm 820 821 #endif 822