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