1 //===- ARMISelLowering.h - ARM 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 ARM uses to lower LLVM code into a 10 // selection DAG. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #ifndef LLVM_LIB_TARGET_ARM_ARMISELLOWERING_H 15 #define LLVM_LIB_TARGET_ARM_ARMISELLOWERING_H 16 17 #include "MCTargetDesc/ARMBaseInfo.h" 18 #include "llvm/ADT/SmallVector.h" 19 #include "llvm/ADT/StringRef.h" 20 #include "llvm/CodeGen/CallingConvLower.h" 21 #include "llvm/CodeGen/ISDOpcodes.h" 22 #include "llvm/CodeGen/MachineFunction.h" 23 #include "llvm/CodeGen/MachineValueType.h" 24 #include "llvm/CodeGen/SelectionDAGNodes.h" 25 #include "llvm/CodeGen/TargetLowering.h" 26 #include "llvm/CodeGen/ValueTypes.h" 27 #include "llvm/IR/Attributes.h" 28 #include "llvm/IR/CallingConv.h" 29 #include "llvm/IR/Function.h" 30 #include "llvm/IR/IRBuilder.h" 31 #include "llvm/IR/InlineAsm.h" 32 #include "llvm/Support/CodeGen.h" 33 #include <optional> 34 #include <utility> 35 36 namespace llvm { 37 38 class ARMSubtarget; 39 class DataLayout; 40 class FastISel; 41 class FunctionLoweringInfo; 42 class GlobalValue; 43 class InstrItineraryData; 44 class Instruction; 45 class MachineBasicBlock; 46 class MachineInstr; 47 class SelectionDAG; 48 class TargetLibraryInfo; 49 class TargetMachine; 50 class TargetRegisterInfo; 51 class VectorType; 52 53 namespace ARMISD { 54 55 // ARM Specific DAG Nodes 56 enum NodeType : unsigned { 57 // Start the numbering where the builtin ops and target ops leave off. 58 FIRST_NUMBER = ISD::BUILTIN_OP_END, 59 60 Wrapper, // Wrapper - A wrapper node for TargetConstantPool, 61 // TargetExternalSymbol, and TargetGlobalAddress. 62 WrapperPIC, // WrapperPIC - A wrapper node for TargetGlobalAddress in 63 // PIC mode. 64 WrapperJT, // WrapperJT - A wrapper node for TargetJumpTable 65 66 // Add pseudo op to model memcpy for struct byval. 67 COPY_STRUCT_BYVAL, 68 69 CALL, // Function call. 70 CALL_PRED, // Function call that's predicable. 71 CALL_NOLINK, // Function call with branch not branch-and-link. 72 tSECALL, // CMSE non-secure function call. 73 t2CALL_BTI, // Thumb function call followed by BTI instruction. 74 BRCOND, // Conditional branch. 75 BR_JT, // Jumptable branch. 76 BR2_JT, // Jumptable branch (2 level - jumptable entry is a jump). 77 RET_GLUE, // Return with a flag operand. 78 SERET_GLUE, // CMSE Entry function return with a flag operand. 79 INTRET_GLUE, // Interrupt return with an LR-offset and a flag operand. 80 81 PIC_ADD, // Add with a PC operand and a PIC label. 82 83 ASRL, // MVE long arithmetic shift right. 84 LSRL, // MVE long shift right. 85 LSLL, // MVE long shift left. 86 87 CMP, // ARM compare instructions. 88 CMN, // ARM CMN instructions. 89 CMPZ, // ARM compare that sets only Z flag. 90 CMPFP, // ARM VFP compare instruction, sets FPSCR. 91 CMPFPE, // ARM VFP signalling compare instruction, sets FPSCR. 92 CMPFPw0, // ARM VFP compare against zero instruction, sets FPSCR. 93 CMPFPEw0, // ARM VFP signalling compare against zero instruction, sets 94 // FPSCR. 95 FMSTAT, // ARM fmstat instruction. 96 97 CMOV, // ARM conditional move instructions. 98 SUBS, // Flag-setting subtraction. 99 100 SSAT, // Signed saturation 101 USAT, // Unsigned saturation 102 103 BCC_i64, 104 105 SRL_GLUE, // V,Flag = srl_flag X -> srl X, 1 + save carry out. 106 SRA_GLUE, // V,Flag = sra_flag X -> sra X, 1 + save carry out. 107 RRX, // V = RRX X, Flag -> srl X, 1 + shift in carry flag. 108 109 ADDC, // Add with carry 110 ADDE, // Add using carry 111 SUBC, // Sub with carry 112 SUBE, // Sub using carry 113 LSLS, // Shift left producing carry 114 115 VMOVRRD, // double to two gprs. 116 VMOVDRR, // Two gprs to double. 117 VMOVSR, // move gpr to single, used for f32 literal constructed in a gpr 118 119 EH_SJLJ_SETJMP, // SjLj exception handling setjmp. 120 EH_SJLJ_LONGJMP, // SjLj exception handling longjmp. 121 EH_SJLJ_SETUP_DISPATCH, // SjLj exception handling setup_dispatch. 122 123 TC_RETURN, // Tail call return pseudo. 124 125 THREAD_POINTER, 126 127 DYN_ALLOC, // Dynamic allocation on the stack. 128 129 MEMBARRIER_MCR, // Memory barrier (MCR) 130 131 PRELOAD, // Preload 132 133 WIN__CHKSTK, // Windows' __chkstk call to do stack probing. 134 WIN__DBZCHK, // Windows' divide by zero check 135 136 WLS, // Low-overhead loops, While Loop Start branch. See t2WhileLoopStart 137 WLSSETUP, // Setup for the iteration count of a WLS. See t2WhileLoopSetup. 138 LOOP_DEC, // Really a part of LE, performs the sub 139 LE, // Low-overhead loops, Loop End 140 141 PREDICATE_CAST, // Predicate cast for MVE i1 types 142 VECTOR_REG_CAST, // Reinterpret the current contents of a vector register 143 144 MVESEXT, // Legalization aids for extending a vector into two/four vectors. 145 MVEZEXT, // or truncating two/four vectors into one. Eventually becomes 146 MVETRUNC, // stack store/load sequence, if not optimized to anything else. 147 148 VCMP, // Vector compare. 149 VCMPZ, // Vector compare to zero. 150 VTST, // Vector test bits. 151 152 // Vector shift by vector 153 VSHLs, // ...left/right by signed 154 VSHLu, // ...left/right by unsigned 155 156 // Vector shift by immediate: 157 VSHLIMM, // ...left 158 VSHRsIMM, // ...right (signed) 159 VSHRuIMM, // ...right (unsigned) 160 161 // Vector rounding shift by immediate: 162 VRSHRsIMM, // ...right (signed) 163 VRSHRuIMM, // ...right (unsigned) 164 VRSHRNIMM, // ...right narrow 165 166 // Vector saturating shift by immediate: 167 VQSHLsIMM, // ...left (signed) 168 VQSHLuIMM, // ...left (unsigned) 169 VQSHLsuIMM, // ...left (signed to unsigned) 170 VQSHRNsIMM, // ...right narrow (signed) 171 VQSHRNuIMM, // ...right narrow (unsigned) 172 VQSHRNsuIMM, // ...right narrow (signed to unsigned) 173 174 // Vector saturating rounding shift by immediate: 175 VQRSHRNsIMM, // ...right narrow (signed) 176 VQRSHRNuIMM, // ...right narrow (unsigned) 177 VQRSHRNsuIMM, // ...right narrow (signed to unsigned) 178 179 // Vector shift and insert: 180 VSLIIMM, // ...left 181 VSRIIMM, // ...right 182 183 // Vector get lane (VMOV scalar to ARM core register) 184 // (These are used for 8- and 16-bit element types only.) 185 VGETLANEu, // zero-extend vector extract element 186 VGETLANEs, // sign-extend vector extract element 187 188 // Vector move immediate and move negated immediate: 189 VMOVIMM, 190 VMVNIMM, 191 192 // Vector move f32 immediate: 193 VMOVFPIMM, 194 195 // Move H <-> R, clearing top 16 bits 196 VMOVrh, 197 VMOVhr, 198 199 // Vector duplicate: 200 VDUP, 201 VDUPLANE, 202 203 // Vector shuffles: 204 VEXT, // extract 205 VREV64, // reverse elements within 64-bit doublewords 206 VREV32, // reverse elements within 32-bit words 207 VREV16, // reverse elements within 16-bit halfwords 208 VZIP, // zip (interleave) 209 VUZP, // unzip (deinterleave) 210 VTRN, // transpose 211 VTBL1, // 1-register shuffle with mask 212 VTBL2, // 2-register shuffle with mask 213 VMOVN, // MVE vmovn 214 215 // MVE Saturating truncates 216 VQMOVNs, // Vector (V) Saturating (Q) Move and Narrow (N), signed (s) 217 VQMOVNu, // Vector (V) Saturating (Q) Move and Narrow (N), unsigned (u) 218 219 // MVE float <> half converts 220 VCVTN, // MVE vcvt f32 -> f16, truncating into either the bottom or top 221 // lanes 222 VCVTL, // MVE vcvt f16 -> f32, extending from either the bottom or top lanes 223 224 // MVE VIDUP instruction, taking a start value and increment. 225 VIDUP, 226 227 // Vector multiply long: 228 VMULLs, // ...signed 229 VMULLu, // ...unsigned 230 231 VQDMULH, // MVE vqdmulh instruction 232 233 // MVE reductions 234 VADDVs, // sign- or zero-extend the elements of a vector to i32, 235 VADDVu, // add them all together, and return an i32 of their sum 236 VADDVps, // Same as VADDV[su] but with a v4i1 predicate mask 237 VADDVpu, 238 VADDLVs, // sign- or zero-extend elements to i64 and sum, returning 239 VADDLVu, // the low and high 32-bit halves of the sum 240 VADDLVAs, // Same as VADDLV[su] but also add an input accumulator 241 VADDLVAu, // provided as low and high halves 242 VADDLVps, // Same as VADDLV[su] but with a v4i1 predicate mask 243 VADDLVpu, 244 VADDLVAps, // Same as VADDLVp[su] but with a v4i1 predicate mask 245 VADDLVApu, 246 VMLAVs, // sign- or zero-extend the elements of two vectors to i32, multiply 247 VMLAVu, // them and add the results together, returning an i32 of their sum 248 VMLAVps, // Same as VMLAV[su] with a v4i1 predicate mask 249 VMLAVpu, 250 VMLALVs, // Same as VMLAV but with i64, returning the low and 251 VMLALVu, // high 32-bit halves of the sum 252 VMLALVps, // Same as VMLALV[su] with a v4i1 predicate mask 253 VMLALVpu, 254 VMLALVAs, // Same as VMLALV but also add an input accumulator 255 VMLALVAu, // provided as low and high halves 256 VMLALVAps, // Same as VMLALVA[su] with a v4i1 predicate mask 257 VMLALVApu, 258 VMINVu, // Find minimum unsigned value of a vector and register 259 VMINVs, // Find minimum signed value of a vector and register 260 VMAXVu, // Find maximum unsigned value of a vector and register 261 VMAXVs, // Find maximum signed value of a vector and register 262 263 SMULWB, // Signed multiply word by half word, bottom 264 SMULWT, // Signed multiply word by half word, top 265 UMLAL, // 64bit Unsigned Accumulate Multiply 266 SMLAL, // 64bit Signed Accumulate Multiply 267 UMAAL, // 64-bit Unsigned Accumulate Accumulate Multiply 268 SMLALBB, // 64-bit signed accumulate multiply bottom, bottom 16 269 SMLALBT, // 64-bit signed accumulate multiply bottom, top 16 270 SMLALTB, // 64-bit signed accumulate multiply top, bottom 16 271 SMLALTT, // 64-bit signed accumulate multiply top, top 16 272 SMLALD, // Signed multiply accumulate long dual 273 SMLALDX, // Signed multiply accumulate long dual exchange 274 SMLSLD, // Signed multiply subtract long dual 275 SMLSLDX, // Signed multiply subtract long dual exchange 276 SMMLAR, // Signed multiply long, round and add 277 SMMLSR, // Signed multiply long, subtract and round 278 279 // Single Lane QADD8 and QADD16. Only the bottom lane. That's what the b 280 // stands for. 281 QADD8b, 282 QSUB8b, 283 QADD16b, 284 QSUB16b, 285 UQADD8b, 286 UQSUB8b, 287 UQADD16b, 288 UQSUB16b, 289 290 // Operands of the standard BUILD_VECTOR node are not legalized, which 291 // is fine if BUILD_VECTORs are always lowered to shuffles or other 292 // operations, but for ARM some BUILD_VECTORs are legal as-is and their 293 // operands need to be legalized. Define an ARM-specific version of 294 // BUILD_VECTOR for this purpose. 295 BUILD_VECTOR, 296 297 // Bit-field insert 298 BFI, 299 300 // Vector OR with immediate 301 VORRIMM, 302 // Vector AND with NOT of immediate 303 VBICIMM, 304 305 // Pseudo vector bitwise select 306 VBSP, 307 308 // Pseudo-instruction representing a memory copy using ldm/stm 309 // instructions. 310 MEMCPY, 311 312 // Pseudo-instruction representing a memory copy using a tail predicated 313 // loop 314 MEMCPYLOOP, 315 // Pseudo-instruction representing a memset using a tail predicated 316 // loop 317 MEMSETLOOP, 318 319 // V8.1MMainline condition select 320 CSINV, // Conditional select invert. 321 CSNEG, // Conditional select negate. 322 CSINC, // Conditional select increment. 323 324 // Vector load N-element structure to all lanes: 325 VLD1DUP = ISD::FIRST_TARGET_MEMORY_OPCODE, 326 VLD2DUP, 327 VLD3DUP, 328 VLD4DUP, 329 330 // NEON loads with post-increment base updates: 331 VLD1_UPD, 332 VLD2_UPD, 333 VLD3_UPD, 334 VLD4_UPD, 335 VLD2LN_UPD, 336 VLD3LN_UPD, 337 VLD4LN_UPD, 338 VLD1DUP_UPD, 339 VLD2DUP_UPD, 340 VLD3DUP_UPD, 341 VLD4DUP_UPD, 342 VLD1x2_UPD, 343 VLD1x3_UPD, 344 VLD1x4_UPD, 345 346 // NEON stores with post-increment base updates: 347 VST1_UPD, 348 VST2_UPD, 349 VST3_UPD, 350 VST4_UPD, 351 VST2LN_UPD, 352 VST3LN_UPD, 353 VST4LN_UPD, 354 VST1x2_UPD, 355 VST1x3_UPD, 356 VST1x4_UPD, 357 358 // Load/Store of dual registers 359 LDRD, 360 STRD 361 }; 362 363 } // end namespace ARMISD 364 365 namespace ARM { 366 /// Possible values of current rounding mode, which is specified in bits 367 /// 23:22 of FPSCR. 368 enum Rounding { 369 RN = 0, // Round to Nearest 370 RP = 1, // Round towards Plus infinity 371 RM = 2, // Round towards Minus infinity 372 RZ = 3, // Round towards Zero 373 rmMask = 3 // Bit mask selecting rounding mode 374 }; 375 376 // Bit position of rounding mode bits in FPSCR. 377 const unsigned RoundingBitsPos = 22; 378 379 // Bits of floating-point status. These are NZCV flags, QC bit and cumulative 380 // FP exception bits. 381 const unsigned FPStatusBits = 0xf800009f; 382 383 // Some bits in the FPSCR are not yet defined. They must be preserved when 384 // modifying the contents. 385 const unsigned FPReservedBits = 0x00006060; 386 } // namespace ARM 387 388 /// Define some predicates that are used for node matching. 389 namespace ARM { 390 391 bool isBitFieldInvertedMask(unsigned v); 392 393 } // end namespace ARM 394 395 //===--------------------------------------------------------------------===// 396 // ARMTargetLowering - ARM Implementation of the TargetLowering interface 397 398 class ARMTargetLowering : public TargetLowering { 399 public: 400 explicit ARMTargetLowering(const TargetMachine &TM, 401 const ARMSubtarget &STI); 402 403 unsigned getJumpTableEncoding() const override; 404 bool useSoftFloat() const override; 405 406 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override; 407 408 /// ReplaceNodeResults - Replace the results of node with an illegal result 409 /// type with new values built out of custom code. 410 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results, 411 SelectionDAG &DAG) const override; 412 413 const char *getTargetNodeName(unsigned Opcode) const override; 414 isSelectSupported(SelectSupportKind Kind)415 bool isSelectSupported(SelectSupportKind Kind) const override { 416 // ARM does not support scalar condition selects on vectors. 417 return (Kind != ScalarCondVectorVal); 418 } 419 420 bool isReadOnly(const GlobalValue *GV) const; 421 422 /// getSetCCResultType - Return the value type to use for ISD::SETCC. 423 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context, 424 EVT VT) const override; 425 426 MachineBasicBlock * 427 EmitInstrWithCustomInserter(MachineInstr &MI, 428 MachineBasicBlock *MBB) const override; 429 430 void AdjustInstrPostInstrSelection(MachineInstr &MI, 431 SDNode *Node) const override; 432 433 SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const; 434 SDValue PerformBRCONDCombine(SDNode *N, SelectionDAG &DAG) const; 435 SDValue PerformCMOVToBFICombine(SDNode *N, SelectionDAG &DAG) const; 436 SDValue PerformIntrinsicCombine(SDNode *N, DAGCombinerInfo &DCI) const; 437 SDValue PerformMVEExtCombine(SDNode *N, DAGCombinerInfo &DCI) const; 438 SDValue PerformMVETruncCombine(SDNode *N, DAGCombinerInfo &DCI) const; 439 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override; 440 441 bool SimplifyDemandedBitsForTargetNode(SDValue Op, 442 const APInt &OriginalDemandedBits, 443 const APInt &OriginalDemandedElts, 444 KnownBits &Known, 445 TargetLoweringOpt &TLO, 446 unsigned Depth) const override; 447 448 bool isDesirableToTransformToIntegerOp(unsigned Opc, EVT VT) const override; 449 450 /// allowsMisalignedMemoryAccesses - Returns true if the target allows 451 /// unaligned memory accesses of the specified type. Returns whether it 452 /// is "fast" by reference in the second argument. 453 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AddrSpace, 454 Align Alignment, 455 MachineMemOperand::Flags Flags, 456 unsigned *Fast) const override; 457 458 EVT getOptimalMemOpType(const MemOp &Op, 459 const AttributeList &FuncAttributes) const override; 460 461 bool isTruncateFree(Type *SrcTy, Type *DstTy) const override; 462 bool isTruncateFree(EVT SrcVT, EVT DstVT) const override; 463 bool isZExtFree(SDValue Val, EVT VT2) const override; 464 bool shouldSinkOperands(Instruction *I, 465 SmallVectorImpl<Use *> &Ops) const override; 466 Type* shouldConvertSplatType(ShuffleVectorInst* SVI) const override; 467 468 bool isFNegFree(EVT VT) const override; 469 470 bool isVectorLoadExtDesirable(SDValue ExtVal) const override; 471 472 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override; 473 474 475 /// isLegalAddressingMode - Return true if the addressing mode represented 476 /// by AM is legal for this target, for a load/store of the specified type. 477 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, 478 Type *Ty, unsigned AS, 479 Instruction *I = nullptr) const override; 480 481 bool isLegalT2ScaledAddressingMode(const AddrMode &AM, EVT VT) const; 482 483 /// Returns true if the addressing mode representing by AM is legal 484 /// for the Thumb1 target, for a load/store of the specified type. 485 bool isLegalT1ScaledAddressingMode(const AddrMode &AM, EVT VT) const; 486 487 /// isLegalICmpImmediate - Return true if the specified immediate is legal 488 /// icmp immediate, that is the target has icmp instructions which can 489 /// compare a register against the immediate without having to materialize 490 /// the immediate into a register. 491 bool isLegalICmpImmediate(int64_t Imm) const override; 492 493 /// isLegalAddImmediate - Return true if the specified immediate is legal 494 /// add immediate, that is the target has add instructions which can 495 /// add a register and the immediate without having to materialize 496 /// the immediate into a register. 497 bool isLegalAddImmediate(int64_t Imm) const override; 498 499 /// getPreIndexedAddressParts - returns true by value, base pointer and 500 /// offset pointer and addressing mode by reference if the node's address 501 /// can be legally represented as pre-indexed load / store address. 502 bool getPreIndexedAddressParts(SDNode *N, SDValue &Base, SDValue &Offset, 503 ISD::MemIndexedMode &AM, 504 SelectionDAG &DAG) const override; 505 506 /// getPostIndexedAddressParts - returns true by value, base pointer and 507 /// offset pointer and addressing mode by reference if this node can be 508 /// combined with a load / store to form a post-indexed load / store. 509 bool getPostIndexedAddressParts(SDNode *N, SDNode *Op, SDValue &Base, 510 SDValue &Offset, ISD::MemIndexedMode &AM, 511 SelectionDAG &DAG) const override; 512 513 void computeKnownBitsForTargetNode(const SDValue Op, KnownBits &Known, 514 const APInt &DemandedElts, 515 const SelectionDAG &DAG, 516 unsigned Depth) const override; 517 518 bool targetShrinkDemandedConstant(SDValue Op, const APInt &DemandedBits, 519 const APInt &DemandedElts, 520 TargetLoweringOpt &TLO) const override; 521 522 bool ExpandInlineAsm(CallInst *CI) const override; 523 524 ConstraintType getConstraintType(StringRef Constraint) const override; 525 526 /// Examine constraint string and operand type and determine a weight value. 527 /// The operand object must already have been set up with the operand type. 528 ConstraintWeight getSingleConstraintMatchWeight( 529 AsmOperandInfo &info, const char *constraint) const override; 530 531 std::pair<unsigned, const TargetRegisterClass *> 532 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, 533 StringRef Constraint, MVT VT) const override; 534 535 const char *LowerXConstraint(EVT ConstraintVT) const override; 536 537 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 538 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is 539 /// true it means one of the asm constraint of the inline asm instruction 540 /// being processed is 'm'. 541 void LowerAsmOperandForConstraint(SDValue Op, StringRef Constraint, 542 std::vector<SDValue> &Ops, 543 SelectionDAG &DAG) const override; 544 545 InlineAsm::ConstraintCode getInlineAsmMemConstraint(StringRef ConstraintCode)546 getInlineAsmMemConstraint(StringRef ConstraintCode) const override { 547 if (ConstraintCode == "Q") 548 return InlineAsm::ConstraintCode::Q; 549 if (ConstraintCode.size() == 2) { 550 if (ConstraintCode[0] == 'U') { 551 switch(ConstraintCode[1]) { 552 default: 553 break; 554 case 'm': 555 return InlineAsm::ConstraintCode::Um; 556 case 'n': 557 return InlineAsm::ConstraintCode::Un; 558 case 'q': 559 return InlineAsm::ConstraintCode::Uq; 560 case 's': 561 return InlineAsm::ConstraintCode::Us; 562 case 't': 563 return InlineAsm::ConstraintCode::Ut; 564 case 'v': 565 return InlineAsm::ConstraintCode::Uv; 566 case 'y': 567 return InlineAsm::ConstraintCode::Uy; 568 } 569 } 570 } 571 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode); 572 } 573 getSubtarget()574 const ARMSubtarget* getSubtarget() const { 575 return Subtarget; 576 } 577 578 /// getRegClassFor - Return the register class that should be used for the 579 /// specified value type. 580 const TargetRegisterClass * 581 getRegClassFor(MVT VT, bool isDivergent = false) const override; 582 583 bool shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize, 584 Align &PrefAlign) const override; 585 586 /// createFastISel - This method returns a target specific FastISel object, 587 /// or null if the target does not support "fast" ISel. 588 FastISel *createFastISel(FunctionLoweringInfo &funcInfo, 589 const TargetLibraryInfo *libInfo) const override; 590 591 Sched::Preference getSchedulingPreference(SDNode *N) const override; 592 preferZeroCompareBranch()593 bool preferZeroCompareBranch() const override { return true; } 594 595 bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override; 596 597 bool 598 isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const override; 599 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override; 600 601 /// isFPImmLegal - Returns true if the target can instruction select the 602 /// specified FP immediate natively. If false, the legalizer will 603 /// materialize the FP immediate as a load from a constant pool. 604 bool isFPImmLegal(const APFloat &Imm, EVT VT, 605 bool ForCodeSize = false) const override; 606 607 bool getTgtMemIntrinsic(IntrinsicInfo &Info, 608 const CallInst &I, 609 MachineFunction &MF, 610 unsigned Intrinsic) const override; 611 612 /// Returns true if it is beneficial to convert a load of a constant 613 /// to just the constant itself. 614 bool shouldConvertConstantLoadToIntImm(const APInt &Imm, 615 Type *Ty) const override; 616 617 /// Return true if EXTRACT_SUBVECTOR is cheap for this result type 618 /// with this index. 619 bool isExtractSubvectorCheap(EVT ResVT, EVT SrcVT, 620 unsigned Index) const override; 621 shouldFormOverflowOp(unsigned Opcode,EVT VT,bool MathUsed)622 bool shouldFormOverflowOp(unsigned Opcode, EVT VT, 623 bool MathUsed) const override { 624 // Using overflow ops for overflow checks only should beneficial on ARM. 625 return TargetLowering::shouldFormOverflowOp(Opcode, VT, true); 626 } 627 shouldReassociateReduction(unsigned Opc,EVT VT)628 bool shouldReassociateReduction(unsigned Opc, EVT VT) const override { 629 return Opc != ISD::VECREDUCE_ADD; 630 } 631 632 /// Returns true if an argument of type Ty needs to be passed in a 633 /// contiguous block of registers in calling convention CallConv. 634 bool functionArgumentNeedsConsecutiveRegisters( 635 Type *Ty, CallingConv::ID CallConv, bool isVarArg, 636 const DataLayout &DL) const override; 637 638 /// If a physical register, this returns the register that receives the 639 /// exception address on entry to an EH pad. 640 Register 641 getExceptionPointerRegister(const Constant *PersonalityFn) const override; 642 643 /// If a physical register, this returns the register that receives the 644 /// exception typeid on entry to a landing pad. 645 Register 646 getExceptionSelectorRegister(const Constant *PersonalityFn) const override; 647 648 Instruction *makeDMB(IRBuilderBase &Builder, ARM_MB::MemBOpt Domain) const; 649 Value *emitLoadLinked(IRBuilderBase &Builder, Type *ValueTy, Value *Addr, 650 AtomicOrdering Ord) const override; 651 Value *emitStoreConditional(IRBuilderBase &Builder, Value *Val, Value *Addr, 652 AtomicOrdering Ord) const override; 653 654 void 655 emitAtomicCmpXchgNoStoreLLBalance(IRBuilderBase &Builder) const override; 656 657 Instruction *emitLeadingFence(IRBuilderBase &Builder, Instruction *Inst, 658 AtomicOrdering Ord) const override; 659 Instruction *emitTrailingFence(IRBuilderBase &Builder, Instruction *Inst, 660 AtomicOrdering Ord) const override; 661 662 unsigned getMaxSupportedInterleaveFactor() const override; 663 664 bool lowerInterleavedLoad(LoadInst *LI, 665 ArrayRef<ShuffleVectorInst *> Shuffles, 666 ArrayRef<unsigned> Indices, 667 unsigned Factor) const override; 668 bool lowerInterleavedStore(StoreInst *SI, ShuffleVectorInst *SVI, 669 unsigned Factor) const override; 670 671 bool shouldInsertFencesForAtomic(const Instruction *I) const override; 672 TargetLoweringBase::AtomicExpansionKind 673 shouldExpandAtomicLoadInIR(LoadInst *LI) const override; 674 TargetLoweringBase::AtomicExpansionKind 675 shouldExpandAtomicStoreInIR(StoreInst *SI) const override; 676 TargetLoweringBase::AtomicExpansionKind 677 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override; 678 TargetLoweringBase::AtomicExpansionKind 679 shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const override; 680 681 bool useLoadStackGuardNode() const override; 682 683 void insertSSPDeclarations(Module &M) const override; 684 Value *getSDagStackGuard(const Module &M) const override; 685 Function *getSSPStackGuardCheck(const Module &M) const override; 686 687 bool canCombineStoreAndExtract(Type *VectorTy, Value *Idx, 688 unsigned &Cost) const override; 689 canMergeStoresTo(unsigned AddressSpace,EVT MemVT,const MachineFunction & MF)690 bool canMergeStoresTo(unsigned AddressSpace, EVT MemVT, 691 const MachineFunction &MF) const override { 692 // Do not merge to larger than i32. 693 return (MemVT.getSizeInBits() <= 32); 694 } 695 696 bool isCheapToSpeculateCttz(Type *Ty) const override; 697 bool isCheapToSpeculateCtlz(Type *Ty) const override; 698 convertSetCCLogicToBitwiseLogic(EVT VT)699 bool convertSetCCLogicToBitwiseLogic(EVT VT) const override { 700 return VT.isScalarInteger(); 701 } 702 supportSwiftError()703 bool supportSwiftError() const override { 704 return true; 705 } 706 hasStandaloneRem(EVT VT)707 bool hasStandaloneRem(EVT VT) const override { 708 return HasStandaloneRem; 709 } 710 711 ShiftLegalizationStrategy 712 preferredShiftLegalizationStrategy(SelectionDAG &DAG, SDNode *N, 713 unsigned ExpansionFactor) const override; 714 715 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool isVarArg) const; 716 CCAssignFn *CCAssignFnForReturn(CallingConv::ID CC, bool isVarArg) const; 717 718 /// Returns true if \p VecTy is a legal interleaved access type. This 719 /// function checks the vector element type and the overall width of the 720 /// vector. 721 bool isLegalInterleavedAccessType(unsigned Factor, FixedVectorType *VecTy, 722 Align Alignment, 723 const DataLayout &DL) const; 724 725 bool isMulAddWithConstProfitable(SDValue AddNode, 726 SDValue ConstNode) const override; 727 728 bool alignLoopsWithOptSize() const override; 729 730 /// Returns the number of interleaved accesses that will be generated when 731 /// lowering accesses of the given type. 732 unsigned getNumInterleavedAccesses(VectorType *VecTy, 733 const DataLayout &DL) const; 734 735 void finalizeLowering(MachineFunction &MF) const override; 736 737 /// Return the correct alignment for the current calling convention. 738 Align getABIAlignmentForCallingConv(Type *ArgTy, 739 const DataLayout &DL) const override; 740 741 bool isDesirableToCommuteWithShift(const SDNode *N, 742 CombineLevel Level) const override; 743 744 bool isDesirableToCommuteXorWithShift(const SDNode *N) const override; 745 746 bool shouldFoldConstantShiftPairToMask(const SDNode *N, 747 CombineLevel Level) const override; 748 749 bool shouldFoldSelectWithIdentityConstant(unsigned BinOpcode, 750 EVT VT) const override; 751 752 bool preferIncOfAddToSubOfNot(EVT VT) const override; 753 754 bool shouldConvertFpToSat(unsigned Op, EVT FPVT, EVT VT) const override; 755 756 bool isComplexDeinterleavingSupported() const override; 757 bool isComplexDeinterleavingOperationSupported( 758 ComplexDeinterleavingOperation Operation, Type *Ty) const override; 759 760 Value *createComplexDeinterleavingIR( 761 IRBuilderBase &B, ComplexDeinterleavingOperation OperationType, 762 ComplexDeinterleavingRotation Rotation, Value *InputA, Value *InputB, 763 Value *Accumulator = nullptr) const override; 764 765 protected: 766 std::pair<const TargetRegisterClass *, uint8_t> 767 findRepresentativeClass(const TargetRegisterInfo *TRI, 768 MVT VT) const override; 769 770 private: 771 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can 772 /// make the right decision when generating code for different targets. 773 const ARMSubtarget *Subtarget; 774 775 const TargetRegisterInfo *RegInfo; 776 777 const InstrItineraryData *Itins; 778 779 // TODO: remove this, and have shouldInsertFencesForAtomic do the proper 780 // check. 781 bool InsertFencesForAtomic; 782 783 bool HasStandaloneRem = true; 784 785 void addTypeForNEON(MVT VT, MVT PromotedLdStVT); 786 void addDRTypeForNEON(MVT VT); 787 void addQRTypeForNEON(MVT VT); 788 std::pair<SDValue, SDValue> getARMXALUOOp(SDValue Op, SelectionDAG &DAG, SDValue &ARMcc) const; 789 790 using RegsToPassVector = SmallVector<std::pair<unsigned, SDValue>, 8>; 791 792 void PassF64ArgInRegs(const SDLoc &dl, SelectionDAG &DAG, SDValue Chain, 793 SDValue &Arg, RegsToPassVector &RegsToPass, 794 CCValAssign &VA, CCValAssign &NextVA, 795 SDValue &StackPtr, 796 SmallVectorImpl<SDValue> &MemOpChains, 797 bool IsTailCall, 798 int SPDiff) const; 799 SDValue GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA, 800 SDValue &Root, SelectionDAG &DAG, 801 const SDLoc &dl) const; 802 803 CallingConv::ID getEffectiveCallingConv(CallingConv::ID CC, 804 bool isVarArg) const; 805 CCAssignFn *CCAssignFnForNode(CallingConv::ID CC, bool Return, 806 bool isVarArg) const; 807 std::pair<SDValue, MachinePointerInfo> 808 computeAddrForCallArg(const SDLoc &dl, SelectionDAG &DAG, 809 const CCValAssign &VA, SDValue StackPtr, 810 bool IsTailCall, int SPDiff) const; 811 SDValue LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const; 812 SDValue LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const; 813 SDValue LowerEH_SJLJ_SETUP_DISPATCH(SDValue Op, SelectionDAG &DAG) const; 814 SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG, 815 const ARMSubtarget *Subtarget) const; 816 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG, 817 const ARMSubtarget *Subtarget) const; 818 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const; 819 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const; 820 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const; 821 SDValue LowerGlobalAddressDarwin(SDValue Op, SelectionDAG &DAG) const; 822 SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG) const; 823 SDValue LowerGlobalAddressWindows(SDValue Op, SelectionDAG &DAG) const; 824 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const; 825 SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, 826 SelectionDAG &DAG) const; 827 SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA, 828 SelectionDAG &DAG, 829 TLSModel::Model model) const; 830 SDValue LowerGlobalTLSAddressDarwin(SDValue Op, SelectionDAG &DAG) const; 831 SDValue LowerGlobalTLSAddressWindows(SDValue Op, SelectionDAG &DAG) const; 832 SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const; 833 SDValue LowerSignedALUO(SDValue Op, SelectionDAG &DAG) const; 834 SDValue LowerUnsignedALUO(SDValue Op, SelectionDAG &DAG) const; 835 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const; 836 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const; 837 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const; 838 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const; 839 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const; 840 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const; 841 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const; 842 SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const; 843 SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const; 844 SDValue LowerGET_ROUNDING(SDValue Op, SelectionDAG &DAG) const; 845 SDValue LowerSET_ROUNDING(SDValue Op, SelectionDAG &DAG) const; 846 SDValue LowerSET_FPMODE(SDValue Op, SelectionDAG &DAG) const; 847 SDValue LowerRESET_FPMODE(SDValue Op, SelectionDAG &DAG) const; 848 SDValue LowerConstantFP(SDValue Op, SelectionDAG &DAG, 849 const ARMSubtarget *ST) const; 850 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG, 851 const ARMSubtarget *ST) const; 852 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const; 853 SDValue LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const; 854 SDValue LowerDivRem(SDValue Op, SelectionDAG &DAG) const; 855 SDValue LowerDIV_Windows(SDValue Op, SelectionDAG &DAG, bool Signed) const; 856 void ExpandDIV_Windows(SDValue Op, SelectionDAG &DAG, bool Signed, 857 SmallVectorImpl<SDValue> &Results) const; 858 SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG, 859 const ARMSubtarget *Subtarget) const; 860 SDValue LowerWindowsDIVLibCall(SDValue Op, SelectionDAG &DAG, bool Signed, 861 SDValue &Chain) const; 862 SDValue LowerREM(SDNode *N, SelectionDAG &DAG) const; 863 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const; 864 SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const; 865 SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const; 866 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const; 867 SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const; 868 SDValue LowerFSETCC(SDValue Op, SelectionDAG &DAG) const; 869 SDValue LowerSPONENTRY(SDValue Op, SelectionDAG &DAG) const; 870 void LowerLOAD(SDNode *N, SmallVectorImpl<SDValue> &Results, 871 SelectionDAG &DAG) const; 872 873 Register getRegisterByName(const char* RegName, LLT VT, 874 const MachineFunction &MF) const override; 875 876 SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG, 877 SmallVectorImpl<SDNode *> &Created) const override; 878 879 bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF, 880 EVT VT) const override; 881 882 SDValue MoveToHPR(const SDLoc &dl, SelectionDAG &DAG, MVT LocVT, MVT ValVT, 883 SDValue Val) const; 884 SDValue MoveFromHPR(const SDLoc &dl, SelectionDAG &DAG, MVT LocVT, 885 MVT ValVT, SDValue Val) const; 886 887 SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const; 888 889 SDValue LowerCallResult(SDValue Chain, SDValue InGlue, 890 CallingConv::ID CallConv, bool isVarArg, 891 const SmallVectorImpl<ISD::InputArg> &Ins, 892 const SDLoc &dl, SelectionDAG &DAG, 893 SmallVectorImpl<SDValue> &InVals, bool isThisReturn, 894 SDValue ThisVal) const; 895 supportSplitCSR(MachineFunction * MF)896 bool supportSplitCSR(MachineFunction *MF) const override { 897 return MF->getFunction().getCallingConv() == CallingConv::CXX_FAST_TLS && 898 MF->getFunction().hasFnAttribute(Attribute::NoUnwind); 899 } 900 901 void initializeSplitCSR(MachineBasicBlock *Entry) const override; 902 void insertCopiesSplitCSR( 903 MachineBasicBlock *Entry, 904 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override; 905 906 bool splitValueIntoRegisterParts( 907 SelectionDAG & DAG, const SDLoc &DL, SDValue Val, SDValue *Parts, 908 unsigned NumParts, MVT PartVT, std::optional<CallingConv::ID> CC) 909 const override; 910 911 SDValue joinRegisterPartsIntoValue( 912 SelectionDAG & DAG, const SDLoc &DL, const SDValue *Parts, 913 unsigned NumParts, MVT PartVT, EVT ValueVT, 914 std::optional<CallingConv::ID> CC) const override; 915 916 SDValue 917 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, 918 const SmallVectorImpl<ISD::InputArg> &Ins, 919 const SDLoc &dl, SelectionDAG &DAG, 920 SmallVectorImpl<SDValue> &InVals) const override; 921 922 int StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG, const SDLoc &dl, 923 SDValue &Chain, const Value *OrigArg, 924 unsigned InRegsParamRecordIdx, int ArgOffset, 925 unsigned ArgSize) const; 926 927 void VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG, 928 const SDLoc &dl, SDValue &Chain, 929 unsigned ArgOffset, unsigned TotalArgRegsSaveSize, 930 bool ForceMutable = false) const; 931 932 SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI, 933 SmallVectorImpl<SDValue> &InVals) const override; 934 935 /// HandleByVal - Target-specific cleanup for ByVal support. 936 void HandleByVal(CCState *, unsigned &, Align) const override; 937 938 /// IsEligibleForTailCallOptimization - Check whether the call is eligible 939 /// for tail call optimization. Targets which want to do tail call 940 /// optimization should implement this function. 941 bool IsEligibleForTailCallOptimization( 942 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg, 943 bool isCalleeStructRet, bool isCallerStructRet, 944 const SmallVectorImpl<ISD::OutputArg> &Outs, 945 const SmallVectorImpl<SDValue> &OutVals, 946 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG, 947 const bool isIndirect) const; 948 949 bool CanLowerReturn(CallingConv::ID CallConv, 950 MachineFunction &MF, bool isVarArg, 951 const SmallVectorImpl<ISD::OutputArg> &Outs, 952 LLVMContext &Context) const override; 953 954 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, 955 const SmallVectorImpl<ISD::OutputArg> &Outs, 956 const SmallVectorImpl<SDValue> &OutVals, 957 const SDLoc &dl, SelectionDAG &DAG) const override; 958 959 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override; 960 961 bool mayBeEmittedAsTailCall(const CallInst *CI) const override; 962 shouldConsiderGEPOffsetSplit()963 bool shouldConsiderGEPOffsetSplit() const override { return true; } 964 965 bool isUnsupportedFloatingType(EVT VT) const; 966 967 SDValue getCMOV(const SDLoc &dl, EVT VT, SDValue FalseVal, SDValue TrueVal, 968 SDValue ARMcc, SDValue CCR, SDValue Cmp, 969 SelectionDAG &DAG) const; 970 SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC, 971 SDValue &ARMcc, SelectionDAG &DAG, const SDLoc &dl) const; 972 SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG, 973 const SDLoc &dl, bool Signaling = false) const; 974 SDValue duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const; 975 976 SDValue OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const; 977 978 void SetupEntryBlockForSjLj(MachineInstr &MI, MachineBasicBlock *MBB, 979 MachineBasicBlock *DispatchBB, int FI) const; 980 981 void EmitSjLjDispatchBlock(MachineInstr &MI, MachineBasicBlock *MBB) const; 982 983 MachineBasicBlock *EmitStructByval(MachineInstr &MI, 984 MachineBasicBlock *MBB) const; 985 986 MachineBasicBlock *EmitLowered__chkstk(MachineInstr &MI, 987 MachineBasicBlock *MBB) const; 988 MachineBasicBlock *EmitLowered__dbzchk(MachineInstr &MI, 989 MachineBasicBlock *MBB) const; 990 void addMVEVectorTypes(bool HasMVEFP); 991 void addAllExtLoads(const MVT From, const MVT To, LegalizeAction Action); 992 void setAllExpand(MVT VT); 993 }; 994 995 enum VMOVModImmType { 996 VMOVModImm, 997 VMVNModImm, 998 MVEVMVNModImm, 999 OtherModImm 1000 }; 1001 1002 namespace ARM { 1003 1004 FastISel *createFastISel(FunctionLoweringInfo &funcInfo, 1005 const TargetLibraryInfo *libInfo); 1006 1007 } // end namespace ARM 1008 1009 } // end namespace llvm 1010 1011 #endif // LLVM_LIB_TARGET_ARM_ARMISELLOWERING_H 1012