MipsSEISelLowering.cpp revision aee7825762830536956b9e634fd7ffd59396984d
1//===-- MipsSEISelLowering.cpp - MipsSE DAG Lowering Interface --*- C++ -*-===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// Subclass of MipsTargetLowering specialized for mips32/64. 11// 12//===----------------------------------------------------------------------===// 13#define DEBUG_TYPE "mips-isel" 14#include "MipsSEISelLowering.h" 15#include "MipsRegisterInfo.h" 16#include "MipsTargetMachine.h" 17#include "llvm/CodeGen/MachineInstrBuilder.h" 18#include "llvm/CodeGen/MachineRegisterInfo.h" 19#include "llvm/IR/Intrinsics.h" 20#include "llvm/Support/CommandLine.h" 21#include "llvm/Support/Debug.h" 22#include "llvm/Support/raw_ostream.h" 23#include "llvm/Target/TargetInstrInfo.h" 24 25using namespace llvm; 26 27static cl::opt<bool> 28EnableMipsTailCalls("enable-mips-tail-calls", cl::Hidden, 29 cl::desc("MIPS: Enable tail calls."), cl::init(false)); 30 31static cl::opt<bool> NoDPLoadStore("mno-ldc1-sdc1", cl::init(false), 32 cl::desc("Expand double precision loads and " 33 "stores to their single precision " 34 "counterparts")); 35 36MipsSETargetLowering::MipsSETargetLowering(MipsTargetMachine &TM) 37 : MipsTargetLowering(TM) { 38 // Set up the register classes 39 addRegisterClass(MVT::i32, &Mips::GPR32RegClass); 40 41 if (HasMips64) 42 addRegisterClass(MVT::i64, &Mips::GPR64RegClass); 43 44 if (Subtarget->hasDSP() || Subtarget->hasMSA()) { 45 // Expand all truncating stores and extending loads. 46 unsigned FirstVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE; 47 unsigned LastVT = (unsigned)MVT::LAST_VECTOR_VALUETYPE; 48 49 for (unsigned VT0 = FirstVT; VT0 <= LastVT; ++VT0) { 50 for (unsigned VT1 = FirstVT; VT1 <= LastVT; ++VT1) 51 setTruncStoreAction((MVT::SimpleValueType)VT0, 52 (MVT::SimpleValueType)VT1, Expand); 53 54 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT0, Expand); 55 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT0, Expand); 56 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT0, Expand); 57 } 58 } 59 60 if (Subtarget->hasDSP()) { 61 MVT::SimpleValueType VecTys[2] = {MVT::v2i16, MVT::v4i8}; 62 63 for (unsigned i = 0; i < array_lengthof(VecTys); ++i) { 64 addRegisterClass(VecTys[i], &Mips::DSPRRegClass); 65 66 // Expand all builtin opcodes. 67 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc) 68 setOperationAction(Opc, VecTys[i], Expand); 69 70 setOperationAction(ISD::ADD, VecTys[i], Legal); 71 setOperationAction(ISD::SUB, VecTys[i], Legal); 72 setOperationAction(ISD::LOAD, VecTys[i], Legal); 73 setOperationAction(ISD::STORE, VecTys[i], Legal); 74 setOperationAction(ISD::BITCAST, VecTys[i], Legal); 75 } 76 77 setTargetDAGCombine(ISD::SHL); 78 setTargetDAGCombine(ISD::SRA); 79 setTargetDAGCombine(ISD::SRL); 80 setTargetDAGCombine(ISD::SETCC); 81 setTargetDAGCombine(ISD::VSELECT); 82 } 83 84 if (Subtarget->hasDSPR2()) 85 setOperationAction(ISD::MUL, MVT::v2i16, Legal); 86 87 if (Subtarget->hasMSA()) { 88 addMSAIntType(MVT::v16i8, &Mips::MSA128BRegClass); 89 addMSAIntType(MVT::v8i16, &Mips::MSA128HRegClass); 90 addMSAIntType(MVT::v4i32, &Mips::MSA128WRegClass); 91 addMSAIntType(MVT::v2i64, &Mips::MSA128DRegClass); 92 addMSAFloatType(MVT::v8f16, &Mips::MSA128HRegClass); 93 addMSAFloatType(MVT::v4f32, &Mips::MSA128WRegClass); 94 addMSAFloatType(MVT::v2f64, &Mips::MSA128DRegClass); 95 96 setTargetDAGCombine(ISD::AND); 97 setTargetDAGCombine(ISD::OR); 98 setTargetDAGCombine(ISD::SRA); 99 setTargetDAGCombine(ISD::VSELECT); 100 setTargetDAGCombine(ISD::XOR); 101 } 102 103 if (!Subtarget->mipsSEUsesSoftFloat()) { 104 addRegisterClass(MVT::f32, &Mips::FGR32RegClass); 105 106 // When dealing with single precision only, use libcalls 107 if (!Subtarget->isSingleFloat()) { 108 if (Subtarget->isFP64bit()) 109 addRegisterClass(MVT::f64, &Mips::FGR64RegClass); 110 else 111 addRegisterClass(MVT::f64, &Mips::AFGR64RegClass); 112 } 113 } 114 115 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Custom); 116 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Custom); 117 setOperationAction(ISD::MULHS, MVT::i32, Custom); 118 setOperationAction(ISD::MULHU, MVT::i32, Custom); 119 120 if (HasMips64) { 121 setOperationAction(ISD::MULHS, MVT::i64, Custom); 122 setOperationAction(ISD::MULHU, MVT::i64, Custom); 123 setOperationAction(ISD::MUL, MVT::i64, Custom); 124 } 125 126 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom); 127 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom); 128 129 setOperationAction(ISD::SDIVREM, MVT::i32, Custom); 130 setOperationAction(ISD::UDIVREM, MVT::i32, Custom); 131 setOperationAction(ISD::SDIVREM, MVT::i64, Custom); 132 setOperationAction(ISD::UDIVREM, MVT::i64, Custom); 133 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom); 134 setOperationAction(ISD::LOAD, MVT::i32, Custom); 135 setOperationAction(ISD::STORE, MVT::i32, Custom); 136 137 setTargetDAGCombine(ISD::ADDE); 138 setTargetDAGCombine(ISD::SUBE); 139 setTargetDAGCombine(ISD::MUL); 140 141 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); 142 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom); 143 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom); 144 145 if (NoDPLoadStore) { 146 setOperationAction(ISD::LOAD, MVT::f64, Custom); 147 setOperationAction(ISD::STORE, MVT::f64, Custom); 148 } 149 150 computeRegisterProperties(); 151} 152 153const MipsTargetLowering * 154llvm::createMipsSETargetLowering(MipsTargetMachine &TM) { 155 return new MipsSETargetLowering(TM); 156} 157 158// Enable MSA support for the given integer type and Register class. 159void MipsSETargetLowering:: 160addMSAIntType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) { 161 addRegisterClass(Ty, RC); 162 163 // Expand all builtin opcodes. 164 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc) 165 setOperationAction(Opc, Ty, Expand); 166 167 setOperationAction(ISD::BITCAST, Ty, Legal); 168 setOperationAction(ISD::LOAD, Ty, Legal); 169 setOperationAction(ISD::STORE, Ty, Legal); 170 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Custom); 171 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal); 172 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom); 173 174 setOperationAction(ISD::ADD, Ty, Legal); 175 setOperationAction(ISD::AND, Ty, Legal); 176 setOperationAction(ISD::CTLZ, Ty, Legal); 177 setOperationAction(ISD::CTPOP, Ty, Legal); 178 setOperationAction(ISD::MUL, Ty, Legal); 179 setOperationAction(ISD::OR, Ty, Legal); 180 setOperationAction(ISD::SDIV, Ty, Legal); 181 setOperationAction(ISD::SREM, Ty, Legal); 182 setOperationAction(ISD::SHL, Ty, Legal); 183 setOperationAction(ISD::SRA, Ty, Legal); 184 setOperationAction(ISD::SRL, Ty, Legal); 185 setOperationAction(ISD::SUB, Ty, Legal); 186 setOperationAction(ISD::UDIV, Ty, Legal); 187 setOperationAction(ISD::UREM, Ty, Legal); 188 setOperationAction(ISD::VECTOR_SHUFFLE, Ty, Custom); 189 setOperationAction(ISD::VSELECT, Ty, Legal); 190 setOperationAction(ISD::XOR, Ty, Legal); 191 192 if (Ty == MVT::v4i32 || Ty == MVT::v2i64) { 193 setOperationAction(ISD::FP_TO_SINT, Ty, Legal); 194 setOperationAction(ISD::FP_TO_UINT, Ty, Legal); 195 setOperationAction(ISD::SINT_TO_FP, Ty, Legal); 196 setOperationAction(ISD::UINT_TO_FP, Ty, Legal); 197 } 198 199 setOperationAction(ISD::SETCC, Ty, Legal); 200 setCondCodeAction(ISD::SETNE, Ty, Expand); 201 setCondCodeAction(ISD::SETGE, Ty, Expand); 202 setCondCodeAction(ISD::SETGT, Ty, Expand); 203 setCondCodeAction(ISD::SETUGE, Ty, Expand); 204 setCondCodeAction(ISD::SETUGT, Ty, Expand); 205} 206 207// Enable MSA support for the given floating-point type and Register class. 208void MipsSETargetLowering:: 209addMSAFloatType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) { 210 addRegisterClass(Ty, RC); 211 212 // Expand all builtin opcodes. 213 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc) 214 setOperationAction(Opc, Ty, Expand); 215 216 setOperationAction(ISD::LOAD, Ty, Legal); 217 setOperationAction(ISD::STORE, Ty, Legal); 218 setOperationAction(ISD::BITCAST, Ty, Legal); 219 setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Legal); 220 setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal); 221 setOperationAction(ISD::BUILD_VECTOR, Ty, Custom); 222 223 if (Ty != MVT::v8f16) { 224 setOperationAction(ISD::FABS, Ty, Legal); 225 setOperationAction(ISD::FADD, Ty, Legal); 226 setOperationAction(ISD::FDIV, Ty, Legal); 227 setOperationAction(ISD::FEXP2, Ty, Legal); 228 setOperationAction(ISD::FLOG2, Ty, Legal); 229 setOperationAction(ISD::FMA, Ty, Legal); 230 setOperationAction(ISD::FMUL, Ty, Legal); 231 setOperationAction(ISD::FRINT, Ty, Legal); 232 setOperationAction(ISD::FSQRT, Ty, Legal); 233 setOperationAction(ISD::FSUB, Ty, Legal); 234 setOperationAction(ISD::VSELECT, Ty, Legal); 235 236 setOperationAction(ISD::SETCC, Ty, Legal); 237 setCondCodeAction(ISD::SETOGE, Ty, Expand); 238 setCondCodeAction(ISD::SETOGT, Ty, Expand); 239 setCondCodeAction(ISD::SETUGE, Ty, Expand); 240 setCondCodeAction(ISD::SETUGT, Ty, Expand); 241 setCondCodeAction(ISD::SETGE, Ty, Expand); 242 setCondCodeAction(ISD::SETGT, Ty, Expand); 243 } 244} 245 246bool 247MipsSETargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const { 248 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy; 249 250 switch (SVT) { 251 case MVT::i64: 252 case MVT::i32: 253 if (Fast) 254 *Fast = true; 255 return true; 256 default: 257 return false; 258 } 259} 260 261SDValue MipsSETargetLowering::LowerOperation(SDValue Op, 262 SelectionDAG &DAG) const { 263 switch(Op.getOpcode()) { 264 case ISD::LOAD: return lowerLOAD(Op, DAG); 265 case ISD::STORE: return lowerSTORE(Op, DAG); 266 case ISD::SMUL_LOHI: return lowerMulDiv(Op, MipsISD::Mult, true, true, DAG); 267 case ISD::UMUL_LOHI: return lowerMulDiv(Op, MipsISD::Multu, true, true, DAG); 268 case ISD::MULHS: return lowerMulDiv(Op, MipsISD::Mult, false, true, DAG); 269 case ISD::MULHU: return lowerMulDiv(Op, MipsISD::Multu, false, true, DAG); 270 case ISD::MUL: return lowerMulDiv(Op, MipsISD::Mult, true, false, DAG); 271 case ISD::SDIVREM: return lowerMulDiv(Op, MipsISD::DivRem, true, true, DAG); 272 case ISD::UDIVREM: return lowerMulDiv(Op, MipsISD::DivRemU, true, true, 273 DAG); 274 case ISD::INTRINSIC_WO_CHAIN: return lowerINTRINSIC_WO_CHAIN(Op, DAG); 275 case ISD::INTRINSIC_W_CHAIN: return lowerINTRINSIC_W_CHAIN(Op, DAG); 276 case ISD::INTRINSIC_VOID: return lowerINTRINSIC_VOID(Op, DAG); 277 case ISD::EXTRACT_VECTOR_ELT: return lowerEXTRACT_VECTOR_ELT(Op, DAG); 278 case ISD::BUILD_VECTOR: return lowerBUILD_VECTOR(Op, DAG); 279 case ISD::VECTOR_SHUFFLE: return lowerVECTOR_SHUFFLE(Op, DAG); 280 } 281 282 return MipsTargetLowering::LowerOperation(Op, DAG); 283} 284 285// selectMADD - 286// Transforms a subgraph in CurDAG if the following pattern is found: 287// (addc multLo, Lo0), (adde multHi, Hi0), 288// where, 289// multHi/Lo: product of multiplication 290// Lo0: initial value of Lo register 291// Hi0: initial value of Hi register 292// Return true if pattern matching was successful. 293static bool selectMADD(SDNode *ADDENode, SelectionDAG *CurDAG) { 294 // ADDENode's second operand must be a flag output of an ADDC node in order 295 // for the matching to be successful. 296 SDNode *ADDCNode = ADDENode->getOperand(2).getNode(); 297 298 if (ADDCNode->getOpcode() != ISD::ADDC) 299 return false; 300 301 SDValue MultHi = ADDENode->getOperand(0); 302 SDValue MultLo = ADDCNode->getOperand(0); 303 SDNode *MultNode = MultHi.getNode(); 304 unsigned MultOpc = MultHi.getOpcode(); 305 306 // MultHi and MultLo must be generated by the same node, 307 if (MultLo.getNode() != MultNode) 308 return false; 309 310 // and it must be a multiplication. 311 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI) 312 return false; 313 314 // MultLo amd MultHi must be the first and second output of MultNode 315 // respectively. 316 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0) 317 return false; 318 319 // Transform this to a MADD only if ADDENode and ADDCNode are the only users 320 // of the values of MultNode, in which case MultNode will be removed in later 321 // phases. 322 // If there exist users other than ADDENode or ADDCNode, this function returns 323 // here, which will result in MultNode being mapped to a single MULT 324 // instruction node rather than a pair of MULT and MADD instructions being 325 // produced. 326 if (!MultHi.hasOneUse() || !MultLo.hasOneUse()) 327 return false; 328 329 SDLoc DL(ADDENode); 330 331 // Initialize accumulator. 332 SDValue ACCIn = CurDAG->getNode(MipsISD::MTLOHI, DL, MVT::Untyped, 333 ADDCNode->getOperand(1), 334 ADDENode->getOperand(1)); 335 336 // create MipsMAdd(u) node 337 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd; 338 339 SDValue MAdd = CurDAG->getNode(MultOpc, DL, MVT::Untyped, 340 MultNode->getOperand(0),// Factor 0 341 MultNode->getOperand(1),// Factor 1 342 ACCIn); 343 344 // replace uses of adde and addc here 345 if (!SDValue(ADDCNode, 0).use_empty()) { 346 SDValue LoOut = CurDAG->getNode(MipsISD::MFLO, DL, MVT::i32, MAdd); 347 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), LoOut); 348 } 349 if (!SDValue(ADDENode, 0).use_empty()) { 350 SDValue HiOut = CurDAG->getNode(MipsISD::MFHI, DL, MVT::i32, MAdd); 351 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), HiOut); 352 } 353 354 return true; 355} 356 357// selectMSUB - 358// Transforms a subgraph in CurDAG if the following pattern is found: 359// (addc Lo0, multLo), (sube Hi0, multHi), 360// where, 361// multHi/Lo: product of multiplication 362// Lo0: initial value of Lo register 363// Hi0: initial value of Hi register 364// Return true if pattern matching was successful. 365static bool selectMSUB(SDNode *SUBENode, SelectionDAG *CurDAG) { 366 // SUBENode's second operand must be a flag output of an SUBC node in order 367 // for the matching to be successful. 368 SDNode *SUBCNode = SUBENode->getOperand(2).getNode(); 369 370 if (SUBCNode->getOpcode() != ISD::SUBC) 371 return false; 372 373 SDValue MultHi = SUBENode->getOperand(1); 374 SDValue MultLo = SUBCNode->getOperand(1); 375 SDNode *MultNode = MultHi.getNode(); 376 unsigned MultOpc = MultHi.getOpcode(); 377 378 // MultHi and MultLo must be generated by the same node, 379 if (MultLo.getNode() != MultNode) 380 return false; 381 382 // and it must be a multiplication. 383 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI) 384 return false; 385 386 // MultLo amd MultHi must be the first and second output of MultNode 387 // respectively. 388 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0) 389 return false; 390 391 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users 392 // of the values of MultNode, in which case MultNode will be removed in later 393 // phases. 394 // If there exist users other than SUBENode or SUBCNode, this function returns 395 // here, which will result in MultNode being mapped to a single MULT 396 // instruction node rather than a pair of MULT and MSUB instructions being 397 // produced. 398 if (!MultHi.hasOneUse() || !MultLo.hasOneUse()) 399 return false; 400 401 SDLoc DL(SUBENode); 402 403 // Initialize accumulator. 404 SDValue ACCIn = CurDAG->getNode(MipsISD::MTLOHI, DL, MVT::Untyped, 405 SUBCNode->getOperand(0), 406 SUBENode->getOperand(0)); 407 408 // create MipsSub(u) node 409 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub; 410 411 SDValue MSub = CurDAG->getNode(MultOpc, DL, MVT::Glue, 412 MultNode->getOperand(0),// Factor 0 413 MultNode->getOperand(1),// Factor 1 414 ACCIn); 415 416 // replace uses of sube and subc here 417 if (!SDValue(SUBCNode, 0).use_empty()) { 418 SDValue LoOut = CurDAG->getNode(MipsISD::MFLO, DL, MVT::i32, MSub); 419 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), LoOut); 420 } 421 if (!SDValue(SUBENode, 0).use_empty()) { 422 SDValue HiOut = CurDAG->getNode(MipsISD::MFHI, DL, MVT::i32, MSub); 423 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), HiOut); 424 } 425 426 return true; 427} 428 429static SDValue performADDECombine(SDNode *N, SelectionDAG &DAG, 430 TargetLowering::DAGCombinerInfo &DCI, 431 const MipsSubtarget *Subtarget) { 432 if (DCI.isBeforeLegalize()) 433 return SDValue(); 434 435 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 && 436 selectMADD(N, &DAG)) 437 return SDValue(N, 0); 438 439 return SDValue(); 440} 441 442// Fold zero extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT 443// 444// Performs the following transformations: 445// - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to zero extension if its 446// sign/zero-extension is completely overwritten by the new one performed by 447// the ISD::AND. 448// - Removes redundant zero extensions performed by an ISD::AND. 449static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG, 450 TargetLowering::DAGCombinerInfo &DCI, 451 const MipsSubtarget *Subtarget) { 452 if (!Subtarget->hasMSA()) 453 return SDValue(); 454 455 SDValue Op0 = N->getOperand(0); 456 SDValue Op1 = N->getOperand(1); 457 unsigned Op0Opcode = Op0->getOpcode(); 458 459 // (and (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d) 460 // where $d + 1 == 2^n and n == 32 461 // or $d + 1 == 2^n and n <= 32 and ZExt 462 // -> (MipsVExtractZExt $a, $b, $c) 463 if (Op0Opcode == MipsISD::VEXTRACT_SEXT_ELT || 464 Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT) { 465 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(Op1); 466 467 if (!Mask) 468 return SDValue(); 469 470 int32_t Log2IfPositive = (Mask->getAPIntValue() + 1).exactLogBase2(); 471 472 if (Log2IfPositive <= 0) 473 return SDValue(); // Mask+1 is not a power of 2 474 475 SDValue Op0Op2 = Op0->getOperand(2); 476 EVT ExtendTy = cast<VTSDNode>(Op0Op2)->getVT(); 477 unsigned ExtendTySize = ExtendTy.getSizeInBits(); 478 unsigned Log2 = Log2IfPositive; 479 480 if ((Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT && Log2 >= ExtendTySize) || 481 Log2 == ExtendTySize) { 482 SDValue Ops[] = { Op0->getOperand(0), Op0->getOperand(1), Op0Op2 }; 483 DAG.MorphNodeTo(Op0.getNode(), MipsISD::VEXTRACT_ZEXT_ELT, 484 Op0->getVTList(), Ops, Op0->getNumOperands()); 485 return Op0; 486 } 487 } 488 489 return SDValue(); 490} 491 492// Determine if the specified node is a constant vector splat. 493// 494// Returns true and sets Imm if: 495// * N is a ISD::BUILD_VECTOR representing a constant splat 496// 497// This function is quite similar to MipsSEDAGToDAGISel::selectVSplat. The 498// differences are that it assumes the MSA has already been checked and the 499// arbitrary requirement for a maximum of 32-bit integers isn't applied (and 500// must not be in order for binsri.d to be selectable). 501static bool isVSplat(SDValue N, APInt &Imm, bool IsLittleEndian) { 502 BuildVectorSDNode *Node = dyn_cast<BuildVectorSDNode>(N.getNode()); 503 504 if (Node == NULL) 505 return false; 506 507 APInt SplatValue, SplatUndef; 508 unsigned SplatBitSize; 509 bool HasAnyUndefs; 510 511 if (!Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs, 512 8, !IsLittleEndian)) 513 return false; 514 515 Imm = SplatValue; 516 517 return true; 518} 519 520// Test whether the given node is an all-ones build_vector. 521static bool isVectorAllOnes(SDValue N) { 522 // Look through bitcasts. Endianness doesn't matter because we are looking 523 // for an all-ones value. 524 if (N->getOpcode() == ISD::BITCAST) 525 N = N->getOperand(0); 526 527 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N); 528 529 if (!BVN) 530 return false; 531 532 APInt SplatValue, SplatUndef; 533 unsigned SplatBitSize; 534 bool HasAnyUndefs; 535 536 // Endianness doesn't matter in this context because we are looking for 537 // an all-ones value. 538 if (BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs)) 539 return SplatValue.isAllOnesValue(); 540 541 return false; 542} 543 544// Test whether N is the bitwise inverse of OfNode. 545static bool isBitwiseInverse(SDValue N, SDValue OfNode) { 546 if (N->getOpcode() != ISD::XOR) 547 return false; 548 549 if (isVectorAllOnes(N->getOperand(0))) 550 return N->getOperand(1) == OfNode; 551 552 if (isVectorAllOnes(N->getOperand(1))) 553 return N->getOperand(0) == OfNode; 554 555 return false; 556} 557 558// Perform combines where ISD::OR is the root node. 559// 560// Performs the following transformations: 561// - (or (and $a, $mask), (and $b, $inv_mask)) => (vselect $mask, $a, $b) 562// where $inv_mask is the bitwise inverse of $mask and the 'or' has a 128-bit 563// vector type. 564static SDValue performORCombine(SDNode *N, SelectionDAG &DAG, 565 TargetLowering::DAGCombinerInfo &DCI, 566 const MipsSubtarget *Subtarget) { 567 if (!Subtarget->hasMSA()) 568 return SDValue(); 569 570 EVT Ty = N->getValueType(0); 571 572 if (!Ty.is128BitVector()) 573 return SDValue(); 574 575 SDValue Op0 = N->getOperand(0); 576 SDValue Op1 = N->getOperand(1); 577 578 if (Op0->getOpcode() == ISD::AND && Op1->getOpcode() == ISD::AND) { 579 SDValue Op0Op0 = Op0->getOperand(0); 580 SDValue Op0Op1 = Op0->getOperand(1); 581 SDValue Op1Op0 = Op1->getOperand(0); 582 SDValue Op1Op1 = Op1->getOperand(1); 583 bool IsLittleEndian = !Subtarget->isLittle(); 584 585 SDValue IfSet, IfClr, Cond; 586 bool IsConstantMask = false; 587 APInt Mask, InvMask; 588 589 // If Op0Op0 is an appropriate mask, try to find it's inverse in either 590 // Op1Op0, or Op1Op1. Keep track of the Cond, IfSet, and IfClr nodes, while 591 // looking. 592 // IfClr will be set if we find a valid match. 593 if (isVSplat(Op0Op0, Mask, IsLittleEndian)) { 594 Cond = Op0Op0; 595 IfSet = Op0Op1; 596 597 if (isVSplat(Op1Op0, InvMask, IsLittleEndian) && Mask == ~InvMask) 598 IfClr = Op1Op1; 599 else if (isVSplat(Op1Op1, InvMask, IsLittleEndian) && Mask == ~InvMask) 600 IfClr = Op1Op0; 601 602 IsConstantMask = true; 603 } 604 605 // If IfClr is not yet set, and Op0Op1 is an appropriate mask, try the same 606 // thing again using this mask. 607 // IfClr will be set if we find a valid match. 608 if (!IfClr.getNode() && isVSplat(Op0Op1, Mask, IsLittleEndian)) { 609 Cond = Op0Op1; 610 IfSet = Op0Op0; 611 612 if (isVSplat(Op1Op0, InvMask, IsLittleEndian) && Mask == ~InvMask) 613 IfClr = Op1Op1; 614 else if (isVSplat(Op1Op1, InvMask, IsLittleEndian) && Mask == ~InvMask) 615 IfClr = Op1Op0; 616 617 IsConstantMask = true; 618 } 619 620 // If IfClr is not yet set, try looking for a non-constant match. 621 // IfClr will be set if we find a valid match amongst the eight 622 // possibilities. 623 if (!IfClr.getNode()) { 624 if (isBitwiseInverse(Op0Op0, Op1Op0)) { 625 Cond = Op1Op0; 626 IfSet = Op1Op1; 627 IfClr = Op0Op1; 628 } else if (isBitwiseInverse(Op0Op1, Op1Op0)) { 629 Cond = Op1Op0; 630 IfSet = Op1Op1; 631 IfClr = Op0Op0; 632 } else if (isBitwiseInverse(Op0Op0, Op1Op1)) { 633 Cond = Op1Op1; 634 IfSet = Op1Op0; 635 IfClr = Op0Op1; 636 } else if (isBitwiseInverse(Op0Op1, Op1Op1)) { 637 Cond = Op1Op1; 638 IfSet = Op1Op0; 639 IfClr = Op0Op0; 640 } else if (isBitwiseInverse(Op1Op0, Op0Op0)) { 641 Cond = Op0Op0; 642 IfSet = Op0Op1; 643 IfClr = Op1Op1; 644 } else if (isBitwiseInverse(Op1Op1, Op0Op0)) { 645 Cond = Op0Op0; 646 IfSet = Op0Op1; 647 IfClr = Op1Op0; 648 } else if (isBitwiseInverse(Op1Op0, Op0Op1)) { 649 Cond = Op0Op1; 650 IfSet = Op0Op0; 651 IfClr = Op1Op1; 652 } else if (isBitwiseInverse(Op1Op1, Op0Op1)) { 653 Cond = Op0Op1; 654 IfSet = Op0Op0; 655 IfClr = Op1Op0; 656 } 657 } 658 659 // At this point, IfClr will be set if we have a valid match. 660 if (!IfClr.getNode()) 661 return SDValue(); 662 663 assert(Cond.getNode() && IfSet.getNode()); 664 665 // Fold degenerate cases. 666 if (IsConstantMask) { 667 if (Mask.isAllOnesValue()) 668 return IfSet; 669 else if (Mask == 0) 670 return IfClr; 671 } 672 673 // Transform the DAG into an equivalent VSELECT. 674 return DAG.getNode(ISD::VSELECT, SDLoc(N), Ty, Cond, IfClr, IfSet); 675 } 676 677 return SDValue(); 678} 679 680static SDValue performSUBECombine(SDNode *N, SelectionDAG &DAG, 681 TargetLowering::DAGCombinerInfo &DCI, 682 const MipsSubtarget *Subtarget) { 683 if (DCI.isBeforeLegalize()) 684 return SDValue(); 685 686 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 && 687 selectMSUB(N, &DAG)) 688 return SDValue(N, 0); 689 690 return SDValue(); 691} 692 693static SDValue genConstMult(SDValue X, uint64_t C, SDLoc DL, EVT VT, 694 EVT ShiftTy, SelectionDAG &DAG) { 695 // Clear the upper (64 - VT.sizeInBits) bits. 696 C &= ((uint64_t)-1) >> (64 - VT.getSizeInBits()); 697 698 // Return 0. 699 if (C == 0) 700 return DAG.getConstant(0, VT); 701 702 // Return x. 703 if (C == 1) 704 return X; 705 706 // If c is power of 2, return (shl x, log2(c)). 707 if (isPowerOf2_64(C)) 708 return DAG.getNode(ISD::SHL, DL, VT, X, 709 DAG.getConstant(Log2_64(C), ShiftTy)); 710 711 unsigned Log2Ceil = Log2_64_Ceil(C); 712 uint64_t Floor = 1LL << Log2_64(C); 713 uint64_t Ceil = Log2Ceil == 64 ? 0LL : 1LL << Log2Ceil; 714 715 // If |c - floor_c| <= |c - ceil_c|, 716 // where floor_c = pow(2, floor(log2(c))) and ceil_c = pow(2, ceil(log2(c))), 717 // return (add constMult(x, floor_c), constMult(x, c - floor_c)). 718 if (C - Floor <= Ceil - C) { 719 SDValue Op0 = genConstMult(X, Floor, DL, VT, ShiftTy, DAG); 720 SDValue Op1 = genConstMult(X, C - Floor, DL, VT, ShiftTy, DAG); 721 return DAG.getNode(ISD::ADD, DL, VT, Op0, Op1); 722 } 723 724 // If |c - floor_c| > |c - ceil_c|, 725 // return (sub constMult(x, ceil_c), constMult(x, ceil_c - c)). 726 SDValue Op0 = genConstMult(X, Ceil, DL, VT, ShiftTy, DAG); 727 SDValue Op1 = genConstMult(X, Ceil - C, DL, VT, ShiftTy, DAG); 728 return DAG.getNode(ISD::SUB, DL, VT, Op0, Op1); 729} 730 731static SDValue performMULCombine(SDNode *N, SelectionDAG &DAG, 732 const TargetLowering::DAGCombinerInfo &DCI, 733 const MipsSETargetLowering *TL) { 734 EVT VT = N->getValueType(0); 735 736 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) 737 if (!VT.isVector()) 738 return genConstMult(N->getOperand(0), C->getZExtValue(), SDLoc(N), 739 VT, TL->getScalarShiftAmountTy(VT), DAG); 740 741 return SDValue(N, 0); 742} 743 744static SDValue performDSPShiftCombine(unsigned Opc, SDNode *N, EVT Ty, 745 SelectionDAG &DAG, 746 const MipsSubtarget *Subtarget) { 747 // See if this is a vector splat immediate node. 748 APInt SplatValue, SplatUndef; 749 unsigned SplatBitSize; 750 bool HasAnyUndefs; 751 unsigned EltSize = Ty.getVectorElementType().getSizeInBits(); 752 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N->getOperand(1)); 753 754 if (!BV || 755 !BV->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs, 756 EltSize, !Subtarget->isLittle()) || 757 (SplatBitSize != EltSize) || 758 (SplatValue.getZExtValue() >= EltSize)) 759 return SDValue(); 760 761 return DAG.getNode(Opc, SDLoc(N), Ty, N->getOperand(0), 762 DAG.getConstant(SplatValue.getZExtValue(), MVT::i32)); 763} 764 765static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG, 766 TargetLowering::DAGCombinerInfo &DCI, 767 const MipsSubtarget *Subtarget) { 768 EVT Ty = N->getValueType(0); 769 770 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8)) 771 return SDValue(); 772 773 return performDSPShiftCombine(MipsISD::SHLL_DSP, N, Ty, DAG, Subtarget); 774} 775 776// Fold sign-extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT for MSA and fold 777// constant splats into MipsISD::SHRA_DSP for DSPr2. 778// 779// Performs the following transformations: 780// - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to sign extension if its 781// sign/zero-extension is completely overwritten by the new one performed by 782// the ISD::SRA and ISD::SHL nodes. 783// - Removes redundant sign extensions performed by an ISD::SRA and ISD::SHL 784// sequence. 785// 786// See performDSPShiftCombine for more information about the transformation 787// used for DSPr2. 788static SDValue performSRACombine(SDNode *N, SelectionDAG &DAG, 789 TargetLowering::DAGCombinerInfo &DCI, 790 const MipsSubtarget *Subtarget) { 791 EVT Ty = N->getValueType(0); 792 793 if (Subtarget->hasMSA()) { 794 SDValue Op0 = N->getOperand(0); 795 SDValue Op1 = N->getOperand(1); 796 797 // (sra (shl (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d), imm:$d) 798 // where $d + sizeof($c) == 32 799 // or $d + sizeof($c) <= 32 and SExt 800 // -> (MipsVExtractSExt $a, $b, $c) 801 if (Op0->getOpcode() == ISD::SHL && Op1 == Op0->getOperand(1)) { 802 SDValue Op0Op0 = Op0->getOperand(0); 803 ConstantSDNode *ShAmount = dyn_cast<ConstantSDNode>(Op1); 804 805 if (!ShAmount) 806 return SDValue(); 807 808 if (Op0Op0->getOpcode() != MipsISD::VEXTRACT_SEXT_ELT && 809 Op0Op0->getOpcode() != MipsISD::VEXTRACT_ZEXT_ELT) 810 return SDValue(); 811 812 EVT ExtendTy = cast<VTSDNode>(Op0Op0->getOperand(2))->getVT(); 813 unsigned TotalBits = ShAmount->getZExtValue() + ExtendTy.getSizeInBits(); 814 815 if (TotalBits == 32 || 816 (Op0Op0->getOpcode() == MipsISD::VEXTRACT_SEXT_ELT && 817 TotalBits <= 32)) { 818 SDValue Ops[] = { Op0Op0->getOperand(0), Op0Op0->getOperand(1), 819 Op0Op0->getOperand(2) }; 820 DAG.MorphNodeTo(Op0Op0.getNode(), MipsISD::VEXTRACT_SEXT_ELT, 821 Op0Op0->getVTList(), Ops, Op0Op0->getNumOperands()); 822 return Op0Op0; 823 } 824 } 825 } 826 827 if ((Ty != MVT::v2i16) && ((Ty != MVT::v4i8) || !Subtarget->hasDSPR2())) 828 return SDValue(); 829 830 return performDSPShiftCombine(MipsISD::SHRA_DSP, N, Ty, DAG, Subtarget); 831} 832 833 834static SDValue performSRLCombine(SDNode *N, SelectionDAG &DAG, 835 TargetLowering::DAGCombinerInfo &DCI, 836 const MipsSubtarget *Subtarget) { 837 EVT Ty = N->getValueType(0); 838 839 if (((Ty != MVT::v2i16) || !Subtarget->hasDSPR2()) && (Ty != MVT::v4i8)) 840 return SDValue(); 841 842 return performDSPShiftCombine(MipsISD::SHRL_DSP, N, Ty, DAG, Subtarget); 843} 844 845static bool isLegalDSPCondCode(EVT Ty, ISD::CondCode CC) { 846 bool IsV216 = (Ty == MVT::v2i16); 847 848 switch (CC) { 849 case ISD::SETEQ: 850 case ISD::SETNE: return true; 851 case ISD::SETLT: 852 case ISD::SETLE: 853 case ISD::SETGT: 854 case ISD::SETGE: return IsV216; 855 case ISD::SETULT: 856 case ISD::SETULE: 857 case ISD::SETUGT: 858 case ISD::SETUGE: return !IsV216; 859 default: return false; 860 } 861} 862 863static SDValue performSETCCCombine(SDNode *N, SelectionDAG &DAG) { 864 EVT Ty = N->getValueType(0); 865 866 if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8)) 867 return SDValue(); 868 869 if (!isLegalDSPCondCode(Ty, cast<CondCodeSDNode>(N->getOperand(2))->get())) 870 return SDValue(); 871 872 return DAG.getNode(MipsISD::SETCC_DSP, SDLoc(N), Ty, N->getOperand(0), 873 N->getOperand(1), N->getOperand(2)); 874} 875 876static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG) { 877 EVT Ty = N->getValueType(0); 878 879 if (Ty.is128BitVector() && Ty.isInteger()) { 880 // Try the following combines: 881 // (vselect (setcc $a, $b, SETLT), $b, $a)) -> (vsmax $a, $b) 882 // (vselect (setcc $a, $b, SETLE), $b, $a)) -> (vsmax $a, $b) 883 // (vselect (setcc $a, $b, SETLT), $a, $b)) -> (vsmin $a, $b) 884 // (vselect (setcc $a, $b, SETLE), $a, $b)) -> (vsmin $a, $b) 885 // (vselect (setcc $a, $b, SETULT), $b, $a)) -> (vumax $a, $b) 886 // (vselect (setcc $a, $b, SETULE), $b, $a)) -> (vumax $a, $b) 887 // (vselect (setcc $a, $b, SETULT), $a, $b)) -> (vumin $a, $b) 888 // (vselect (setcc $a, $b, SETULE), $a, $b)) -> (vumin $a, $b) 889 // SETGT/SETGE/SETUGT/SETUGE variants of these will show up initially but 890 // will be expanded to equivalent SETLT/SETLE/SETULT/SETULE versions by the 891 // legalizer. 892 SDValue Op0 = N->getOperand(0); 893 894 if (Op0->getOpcode() != ISD::SETCC) 895 return SDValue(); 896 897 ISD::CondCode CondCode = cast<CondCodeSDNode>(Op0->getOperand(2))->get(); 898 bool Signed; 899 900 if (CondCode == ISD::SETLT || CondCode == ISD::SETLE) 901 Signed = true; 902 else if (CondCode == ISD::SETULT || CondCode == ISD::SETULE) 903 Signed = false; 904 else 905 return SDValue(); 906 907 SDValue Op1 = N->getOperand(1); 908 SDValue Op2 = N->getOperand(2); 909 SDValue Op0Op0 = Op0->getOperand(0); 910 SDValue Op0Op1 = Op0->getOperand(1); 911 912 if (Op1 == Op0Op0 && Op2 == Op0Op1) 913 return DAG.getNode(Signed ? MipsISD::VSMIN : MipsISD::VUMIN, SDLoc(N), 914 Ty, Op1, Op2); 915 else if (Op1 == Op0Op1 && Op2 == Op0Op0) 916 return DAG.getNode(Signed ? MipsISD::VSMAX : MipsISD::VUMAX, SDLoc(N), 917 Ty, Op1, Op2); 918 } else if ((Ty == MVT::v2i16) || (Ty == MVT::v4i8)) { 919 SDValue SetCC = N->getOperand(0); 920 921 if (SetCC.getOpcode() != MipsISD::SETCC_DSP) 922 return SDValue(); 923 924 return DAG.getNode(MipsISD::SELECT_CC_DSP, SDLoc(N), Ty, 925 SetCC.getOperand(0), SetCC.getOperand(1), 926 N->getOperand(1), N->getOperand(2), SetCC.getOperand(2)); 927 } 928 929 return SDValue(); 930} 931 932static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG, 933 const MipsSubtarget *Subtarget) { 934 EVT Ty = N->getValueType(0); 935 936 if (Subtarget->hasMSA() && Ty.is128BitVector() && Ty.isInteger()) { 937 // Try the following combines: 938 // (xor (or $a, $b), (build_vector allones)) 939 // (xor (or $a, $b), (bitcast (build_vector allones))) 940 SDValue Op0 = N->getOperand(0); 941 SDValue Op1 = N->getOperand(1); 942 SDValue NotOp; 943 944 if (ISD::isBuildVectorAllOnes(Op0.getNode())) 945 NotOp = Op1; 946 else if (ISD::isBuildVectorAllOnes(Op1.getNode())) 947 NotOp = Op0; 948 else 949 return SDValue(); 950 951 if (NotOp->getOpcode() == ISD::OR) 952 return DAG.getNode(MipsISD::VNOR, SDLoc(N), Ty, NotOp->getOperand(0), 953 NotOp->getOperand(1)); 954 } 955 956 return SDValue(); 957} 958 959SDValue 960MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const { 961 SelectionDAG &DAG = DCI.DAG; 962 SDValue Val; 963 964 switch (N->getOpcode()) { 965 case ISD::ADDE: 966 return performADDECombine(N, DAG, DCI, Subtarget); 967 case ISD::AND: 968 Val = performANDCombine(N, DAG, DCI, Subtarget); 969 break; 970 case ISD::OR: 971 Val = performORCombine(N, DAG, DCI, Subtarget); 972 break; 973 case ISD::SUBE: 974 return performSUBECombine(N, DAG, DCI, Subtarget); 975 case ISD::MUL: 976 return performMULCombine(N, DAG, DCI, this); 977 case ISD::SHL: 978 return performSHLCombine(N, DAG, DCI, Subtarget); 979 case ISD::SRA: 980 return performSRACombine(N, DAG, DCI, Subtarget); 981 case ISD::SRL: 982 return performSRLCombine(N, DAG, DCI, Subtarget); 983 case ISD::VSELECT: 984 return performVSELECTCombine(N, DAG); 985 case ISD::XOR: 986 Val = performXORCombine(N, DAG, Subtarget); 987 break; 988 case ISD::SETCC: 989 Val = performSETCCCombine(N, DAG); 990 break; 991 } 992 993 if (Val.getNode()) { 994 DEBUG(dbgs() << "\nMipsSE DAG Combine:\n"; 995 N->printrWithDepth(dbgs(), &DAG); 996 dbgs() << "\n=> \n"; 997 Val.getNode()->printrWithDepth(dbgs(), &DAG); 998 dbgs() << "\n"); 999 return Val; 1000 } 1001 1002 return MipsTargetLowering::PerformDAGCombine(N, DCI); 1003} 1004 1005MachineBasicBlock * 1006MipsSETargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 1007 MachineBasicBlock *BB) const { 1008 switch (MI->getOpcode()) { 1009 default: 1010 return MipsTargetLowering::EmitInstrWithCustomInserter(MI, BB); 1011 case Mips::BPOSGE32_PSEUDO: 1012 return emitBPOSGE32(MI, BB); 1013 case Mips::SNZ_B_PSEUDO: 1014 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_B); 1015 case Mips::SNZ_H_PSEUDO: 1016 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_H); 1017 case Mips::SNZ_W_PSEUDO: 1018 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_W); 1019 case Mips::SNZ_D_PSEUDO: 1020 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_D); 1021 case Mips::SNZ_V_PSEUDO: 1022 return emitMSACBranchPseudo(MI, BB, Mips::BNZ_V); 1023 case Mips::SZ_B_PSEUDO: 1024 return emitMSACBranchPseudo(MI, BB, Mips::BZ_B); 1025 case Mips::SZ_H_PSEUDO: 1026 return emitMSACBranchPseudo(MI, BB, Mips::BZ_H); 1027 case Mips::SZ_W_PSEUDO: 1028 return emitMSACBranchPseudo(MI, BB, Mips::BZ_W); 1029 case Mips::SZ_D_PSEUDO: 1030 return emitMSACBranchPseudo(MI, BB, Mips::BZ_D); 1031 case Mips::SZ_V_PSEUDO: 1032 return emitMSACBranchPseudo(MI, BB, Mips::BZ_V); 1033 case Mips::COPY_FW_PSEUDO: 1034 return emitCOPY_FW(MI, BB); 1035 case Mips::COPY_FD_PSEUDO: 1036 return emitCOPY_FD(MI, BB); 1037 case Mips::INSERT_FW_PSEUDO: 1038 return emitINSERT_FW(MI, BB); 1039 case Mips::INSERT_FD_PSEUDO: 1040 return emitINSERT_FD(MI, BB); 1041 case Mips::FILL_FW_PSEUDO: 1042 return emitFILL_FW(MI, BB); 1043 case Mips::FILL_FD_PSEUDO: 1044 return emitFILL_FD(MI, BB); 1045 case Mips::FEXP2_W_1_PSEUDO: 1046 return emitFEXP2_W_1(MI, BB); 1047 case Mips::FEXP2_D_1_PSEUDO: 1048 return emitFEXP2_D_1(MI, BB); 1049 } 1050} 1051 1052bool MipsSETargetLowering:: 1053isEligibleForTailCallOptimization(const MipsCC &MipsCCInfo, 1054 unsigned NextStackOffset, 1055 const MipsFunctionInfo& FI) const { 1056 if (!EnableMipsTailCalls) 1057 return false; 1058 1059 // Return false if either the callee or caller has a byval argument. 1060 if (MipsCCInfo.hasByValArg() || FI.hasByvalArg()) 1061 return false; 1062 1063 // Return true if the callee's argument area is no larger than the 1064 // caller's. 1065 return NextStackOffset <= FI.getIncomingArgSize(); 1066} 1067 1068void MipsSETargetLowering:: 1069getOpndList(SmallVectorImpl<SDValue> &Ops, 1070 std::deque< std::pair<unsigned, SDValue> > &RegsToPass, 1071 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage, 1072 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const { 1073 // T9 should contain the address of the callee function if 1074 // -reloction-model=pic or it is an indirect call. 1075 if (IsPICCall || !GlobalOrExternal) { 1076 unsigned T9Reg = IsN64 ? Mips::T9_64 : Mips::T9; 1077 RegsToPass.push_front(std::make_pair(T9Reg, Callee)); 1078 } else 1079 Ops.push_back(Callee); 1080 1081 MipsTargetLowering::getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, 1082 InternalLinkage, CLI, Callee, Chain); 1083} 1084 1085SDValue MipsSETargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const { 1086 LoadSDNode &Nd = *cast<LoadSDNode>(Op); 1087 1088 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore) 1089 return MipsTargetLowering::lowerLOAD(Op, DAG); 1090 1091 // Replace a double precision load with two i32 loads and a buildpair64. 1092 SDLoc DL(Op); 1093 SDValue Ptr = Nd.getBasePtr(), Chain = Nd.getChain(); 1094 EVT PtrVT = Ptr.getValueType(); 1095 1096 // i32 load from lower address. 1097 SDValue Lo = DAG.getLoad(MVT::i32, DL, Chain, Ptr, 1098 MachinePointerInfo(), Nd.isVolatile(), 1099 Nd.isNonTemporal(), Nd.isInvariant(), 1100 Nd.getAlignment()); 1101 1102 // i32 load from higher address. 1103 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, PtrVT)); 1104 SDValue Hi = DAG.getLoad(MVT::i32, DL, Lo.getValue(1), Ptr, 1105 MachinePointerInfo(), Nd.isVolatile(), 1106 Nd.isNonTemporal(), Nd.isInvariant(), 1107 std::min(Nd.getAlignment(), 4U)); 1108 1109 if (!Subtarget->isLittle()) 1110 std::swap(Lo, Hi); 1111 1112 SDValue BP = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, Lo, Hi); 1113 SDValue Ops[2] = {BP, Hi.getValue(1)}; 1114 return DAG.getMergeValues(Ops, 2, DL); 1115} 1116 1117SDValue MipsSETargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const { 1118 StoreSDNode &Nd = *cast<StoreSDNode>(Op); 1119 1120 if (Nd.getMemoryVT() != MVT::f64 || !NoDPLoadStore) 1121 return MipsTargetLowering::lowerSTORE(Op, DAG); 1122 1123 // Replace a double precision store with two extractelement64s and i32 stores. 1124 SDLoc DL(Op); 1125 SDValue Val = Nd.getValue(), Ptr = Nd.getBasePtr(), Chain = Nd.getChain(); 1126 EVT PtrVT = Ptr.getValueType(); 1127 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 1128 Val, DAG.getConstant(0, MVT::i32)); 1129 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 1130 Val, DAG.getConstant(1, MVT::i32)); 1131 1132 if (!Subtarget->isLittle()) 1133 std::swap(Lo, Hi); 1134 1135 // i32 store to lower address. 1136 Chain = DAG.getStore(Chain, DL, Lo, Ptr, MachinePointerInfo(), 1137 Nd.isVolatile(), Nd.isNonTemporal(), Nd.getAlignment(), 1138 Nd.getTBAAInfo()); 1139 1140 // i32 store to higher address. 1141 Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Ptr, DAG.getConstant(4, PtrVT)); 1142 return DAG.getStore(Chain, DL, Hi, Ptr, MachinePointerInfo(), 1143 Nd.isVolatile(), Nd.isNonTemporal(), 1144 std::min(Nd.getAlignment(), 4U), Nd.getTBAAInfo()); 1145} 1146 1147SDValue MipsSETargetLowering::lowerMulDiv(SDValue Op, unsigned NewOpc, 1148 bool HasLo, bool HasHi, 1149 SelectionDAG &DAG) const { 1150 EVT Ty = Op.getOperand(0).getValueType(); 1151 SDLoc DL(Op); 1152 SDValue Mult = DAG.getNode(NewOpc, DL, MVT::Untyped, 1153 Op.getOperand(0), Op.getOperand(1)); 1154 SDValue Lo, Hi; 1155 1156 if (HasLo) 1157 Lo = DAG.getNode(MipsISD::MFLO, DL, Ty, Mult); 1158 if (HasHi) 1159 Hi = DAG.getNode(MipsISD::MFHI, DL, Ty, Mult); 1160 1161 if (!HasLo || !HasHi) 1162 return HasLo ? Lo : Hi; 1163 1164 SDValue Vals[] = { Lo, Hi }; 1165 return DAG.getMergeValues(Vals, 2, DL); 1166} 1167 1168 1169static SDValue initAccumulator(SDValue In, SDLoc DL, SelectionDAG &DAG) { 1170 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In, 1171 DAG.getConstant(0, MVT::i32)); 1172 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In, 1173 DAG.getConstant(1, MVT::i32)); 1174 return DAG.getNode(MipsISD::MTLOHI, DL, MVT::Untyped, InLo, InHi); 1175} 1176 1177static SDValue extractLOHI(SDValue Op, SDLoc DL, SelectionDAG &DAG) { 1178 SDValue Lo = DAG.getNode(MipsISD::MFLO, DL, MVT::i32, Op); 1179 SDValue Hi = DAG.getNode(MipsISD::MFHI, DL, MVT::i32, Op); 1180 return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi); 1181} 1182 1183// This function expands mips intrinsic nodes which have 64-bit input operands 1184// or output values. 1185// 1186// out64 = intrinsic-node in64 1187// => 1188// lo = copy (extract-element (in64, 0)) 1189// hi = copy (extract-element (in64, 1)) 1190// mips-specific-node 1191// v0 = copy lo 1192// v1 = copy hi 1193// out64 = merge-values (v0, v1) 1194// 1195static SDValue lowerDSPIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) { 1196 SDLoc DL(Op); 1197 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other; 1198 SmallVector<SDValue, 3> Ops; 1199 unsigned OpNo = 0; 1200 1201 // See if Op has a chain input. 1202 if (HasChainIn) 1203 Ops.push_back(Op->getOperand(OpNo++)); 1204 1205 // The next operand is the intrinsic opcode. 1206 assert(Op->getOperand(OpNo).getOpcode() == ISD::TargetConstant); 1207 1208 // See if the next operand has type i64. 1209 SDValue Opnd = Op->getOperand(++OpNo), In64; 1210 1211 if (Opnd.getValueType() == MVT::i64) 1212 In64 = initAccumulator(Opnd, DL, DAG); 1213 else 1214 Ops.push_back(Opnd); 1215 1216 // Push the remaining operands. 1217 for (++OpNo ; OpNo < Op->getNumOperands(); ++OpNo) 1218 Ops.push_back(Op->getOperand(OpNo)); 1219 1220 // Add In64 to the end of the list. 1221 if (In64.getNode()) 1222 Ops.push_back(In64); 1223 1224 // Scan output. 1225 SmallVector<EVT, 2> ResTys; 1226 1227 for (SDNode::value_iterator I = Op->value_begin(), E = Op->value_end(); 1228 I != E; ++I) 1229 ResTys.push_back((*I == MVT::i64) ? MVT::Untyped : *I); 1230 1231 // Create node. 1232 SDValue Val = DAG.getNode(Opc, DL, ResTys, &Ops[0], Ops.size()); 1233 SDValue Out = (ResTys[0] == MVT::Untyped) ? extractLOHI(Val, DL, DAG) : Val; 1234 1235 if (!HasChainIn) 1236 return Out; 1237 1238 assert(Val->getValueType(1) == MVT::Other); 1239 SDValue Vals[] = { Out, SDValue(Val.getNode(), 1) }; 1240 return DAG.getMergeValues(Vals, 2, DL); 1241} 1242 1243// Lower an MSA copy intrinsic into the specified SelectionDAG node 1244static SDValue lowerMSACopyIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) { 1245 SDLoc DL(Op); 1246 SDValue Vec = Op->getOperand(1); 1247 SDValue Idx = Op->getOperand(2); 1248 EVT ResTy = Op->getValueType(0); 1249 EVT EltTy = Vec->getValueType(0).getVectorElementType(); 1250 1251 SDValue Result = DAG.getNode(Opc, DL, ResTy, Vec, Idx, 1252 DAG.getValueType(EltTy)); 1253 1254 return Result; 1255} 1256 1257static SDValue 1258lowerMSASplatImm(SDLoc DL, EVT ResTy, SDValue ImmOp, SelectionDAG &DAG) { 1259 EVT ViaVecTy = ResTy; 1260 SmallVector<SDValue, 16> Ops; 1261 SDValue ImmHiOp; 1262 1263 if (ViaVecTy == MVT::v2i64) { 1264 ImmHiOp = DAG.getNode(ISD::SRA, DL, MVT::i32, ImmOp, 1265 DAG.getConstant(31, MVT::i32)); 1266 for (unsigned i = 0; i < ViaVecTy.getVectorNumElements(); ++i) { 1267 Ops.push_back(ImmHiOp); 1268 Ops.push_back(ImmOp); 1269 } 1270 ViaVecTy = MVT::v4i32; 1271 } else { 1272 for (unsigned i = 0; i < ResTy.getVectorNumElements(); ++i) 1273 Ops.push_back(ImmOp); 1274 } 1275 1276 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy, &Ops[0], 1277 Ops.size()); 1278 1279 if (ResTy != ViaVecTy) 1280 Result = DAG.getNode(ISD::BITCAST, DL, ResTy, Result); 1281 1282 return Result; 1283} 1284 1285static SDValue 1286lowerMSASplatImm(SDValue Op, unsigned ImmOp, SelectionDAG &DAG) { 1287 return lowerMSASplatImm(SDLoc(Op), Op->getValueType(0), 1288 Op->getOperand(ImmOp), DAG); 1289} 1290 1291static SDValue lowerMSABinaryBitImmIntr(SDValue Op, SelectionDAG &DAG, 1292 unsigned Opc, SDValue Imm, 1293 bool BigEndian) { 1294 EVT VecTy = Op->getValueType(0); 1295 SDValue Exp2Imm; 1296 SDLoc DL(Op); 1297 1298 // The DAG Combiner can't constant fold bitcasted vectors so we must do it 1299 // here. 1300 if (VecTy == MVT::v2i64) { 1301 if (ConstantSDNode *CImm = dyn_cast<ConstantSDNode>(Imm)) { 1302 APInt BitImm = APInt(64, 1) << CImm->getAPIntValue(); 1303 1304 SDValue BitImmHiOp = DAG.getConstant(BitImm.lshr(32).trunc(32), MVT::i32); 1305 SDValue BitImmOp = DAG.getConstant(BitImm.trunc(32), MVT::i32); 1306 Exp2Imm = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, 1307 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v4i32, 1308 BitImmHiOp, BitImmOp, 1309 BitImmHiOp, BitImmOp)); 1310 } 1311 } 1312 1313 if (Exp2Imm.getNode() == NULL) { 1314 // We couldnt constant fold, do a vector shift instead 1315 SDValue One = lowerMSASplatImm(DL, VecTy, DAG.getConstant(1, MVT::i32), 1316 DAG); 1317 Exp2Imm = lowerMSASplatImm(DL, VecTy, Imm, DAG); 1318 Exp2Imm = DAG.getNode(ISD::SHL, DL, VecTy, One, Exp2Imm); 1319 } 1320 1321 return DAG.getNode(Opc, DL, VecTy, Op->getOperand(1), Exp2Imm); 1322} 1323 1324SDValue MipsSETargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op, 1325 SelectionDAG &DAG) const { 1326 SDLoc DL(Op); 1327 1328 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) { 1329 default: 1330 return SDValue(); 1331 case Intrinsic::mips_shilo: 1332 return lowerDSPIntr(Op, DAG, MipsISD::SHILO); 1333 case Intrinsic::mips_dpau_h_qbl: 1334 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL); 1335 case Intrinsic::mips_dpau_h_qbr: 1336 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR); 1337 case Intrinsic::mips_dpsu_h_qbl: 1338 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL); 1339 case Intrinsic::mips_dpsu_h_qbr: 1340 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR); 1341 case Intrinsic::mips_dpa_w_ph: 1342 return lowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH); 1343 case Intrinsic::mips_dps_w_ph: 1344 return lowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH); 1345 case Intrinsic::mips_dpax_w_ph: 1346 return lowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH); 1347 case Intrinsic::mips_dpsx_w_ph: 1348 return lowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH); 1349 case Intrinsic::mips_mulsa_w_ph: 1350 return lowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH); 1351 case Intrinsic::mips_mult: 1352 return lowerDSPIntr(Op, DAG, MipsISD::Mult); 1353 case Intrinsic::mips_multu: 1354 return lowerDSPIntr(Op, DAG, MipsISD::Multu); 1355 case Intrinsic::mips_madd: 1356 return lowerDSPIntr(Op, DAG, MipsISD::MAdd); 1357 case Intrinsic::mips_maddu: 1358 return lowerDSPIntr(Op, DAG, MipsISD::MAddu); 1359 case Intrinsic::mips_msub: 1360 return lowerDSPIntr(Op, DAG, MipsISD::MSub); 1361 case Intrinsic::mips_msubu: 1362 return lowerDSPIntr(Op, DAG, MipsISD::MSubu); 1363 case Intrinsic::mips_addv_b: 1364 case Intrinsic::mips_addv_h: 1365 case Intrinsic::mips_addv_w: 1366 case Intrinsic::mips_addv_d: 1367 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1), 1368 Op->getOperand(2)); 1369 case Intrinsic::mips_addvi_b: 1370 case Intrinsic::mips_addvi_h: 1371 case Intrinsic::mips_addvi_w: 1372 case Intrinsic::mips_addvi_d: 1373 return DAG.getNode(ISD::ADD, DL, Op->getValueType(0), Op->getOperand(1), 1374 lowerMSASplatImm(Op, 2, DAG)); 1375 case Intrinsic::mips_and_v: 1376 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1), 1377 Op->getOperand(2)); 1378 case Intrinsic::mips_andi_b: 1379 return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1), 1380 lowerMSASplatImm(Op, 2, DAG)); 1381 case Intrinsic::mips_binsli_b: 1382 case Intrinsic::mips_binsli_h: 1383 case Intrinsic::mips_binsli_w: 1384 case Intrinsic::mips_binsli_d: { 1385 EVT VecTy = Op->getValueType(0); 1386 EVT EltTy = VecTy.getVectorElementType(); 1387 APInt Mask = APInt::getHighBitsSet(EltTy.getSizeInBits(), 1388 Op->getConstantOperandVal(3)); 1389 return DAG.getNode(ISD::VSELECT, DL, VecTy, 1390 DAG.getConstant(Mask, VecTy, true), Op->getOperand(1), 1391 Op->getOperand(2)); 1392 } 1393 case Intrinsic::mips_binsri_b: 1394 case Intrinsic::mips_binsri_h: 1395 case Intrinsic::mips_binsri_w: 1396 case Intrinsic::mips_binsri_d: { 1397 EVT VecTy = Op->getValueType(0); 1398 EVT EltTy = VecTy.getVectorElementType(); 1399 APInt Mask = APInt::getLowBitsSet(EltTy.getSizeInBits(), 1400 Op->getConstantOperandVal(3)); 1401 return DAG.getNode(ISD::VSELECT, DL, VecTy, 1402 DAG.getConstant(Mask, VecTy, true), Op->getOperand(1), 1403 Op->getOperand(2)); 1404 } 1405 case Intrinsic::mips_bmnz_v: 1406 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), Op->getOperand(3), 1407 Op->getOperand(2), Op->getOperand(1)); 1408 case Intrinsic::mips_bmnzi_b: 1409 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), 1410 lowerMSASplatImm(Op, 3, DAG), Op->getOperand(2), 1411 Op->getOperand(1)); 1412 case Intrinsic::mips_bmz_v: 1413 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), Op->getOperand(3), 1414 Op->getOperand(1), Op->getOperand(2)); 1415 case Intrinsic::mips_bmzi_b: 1416 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), 1417 lowerMSASplatImm(Op, 3, DAG), Op->getOperand(1), 1418 Op->getOperand(2)); 1419 case Intrinsic::mips_bneg_b: 1420 case Intrinsic::mips_bneg_h: 1421 case Intrinsic::mips_bneg_w: 1422 case Intrinsic::mips_bneg_d: { 1423 EVT VecTy = Op->getValueType(0); 1424 SDValue One = lowerMSASplatImm(DL, VecTy, DAG.getConstant(1, MVT::i32), 1425 DAG); 1426 1427 return DAG.getNode(ISD::XOR, DL, VecTy, Op->getOperand(1), 1428 DAG.getNode(ISD::SHL, DL, VecTy, One, 1429 Op->getOperand(2))); 1430 } 1431 case Intrinsic::mips_bnegi_b: 1432 case Intrinsic::mips_bnegi_h: 1433 case Intrinsic::mips_bnegi_w: 1434 case Intrinsic::mips_bnegi_d: 1435 return lowerMSABinaryBitImmIntr(Op, DAG, ISD::XOR, Op->getOperand(2), 1436 !Subtarget->isLittle()); 1437 case Intrinsic::mips_bnz_b: 1438 case Intrinsic::mips_bnz_h: 1439 case Intrinsic::mips_bnz_w: 1440 case Intrinsic::mips_bnz_d: 1441 return DAG.getNode(MipsISD::VALL_NONZERO, DL, Op->getValueType(0), 1442 Op->getOperand(1)); 1443 case Intrinsic::mips_bnz_v: 1444 return DAG.getNode(MipsISD::VANY_NONZERO, DL, Op->getValueType(0), 1445 Op->getOperand(1)); 1446 case Intrinsic::mips_bsel_v: 1447 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), 1448 Op->getOperand(1), Op->getOperand(2), 1449 Op->getOperand(3)); 1450 case Intrinsic::mips_bseli_b: 1451 return DAG.getNode(ISD::VSELECT, DL, Op->getValueType(0), 1452 Op->getOperand(1), Op->getOperand(2), 1453 lowerMSASplatImm(Op, 3, DAG)); 1454 case Intrinsic::mips_bset_b: 1455 case Intrinsic::mips_bset_h: 1456 case Intrinsic::mips_bset_w: 1457 case Intrinsic::mips_bset_d: { 1458 EVT VecTy = Op->getValueType(0); 1459 SDValue One = lowerMSASplatImm(DL, VecTy, DAG.getConstant(1, MVT::i32), 1460 DAG); 1461 1462 return DAG.getNode(ISD::OR, DL, VecTy, Op->getOperand(1), 1463 DAG.getNode(ISD::SHL, DL, VecTy, One, 1464 Op->getOperand(2))); 1465 } 1466 case Intrinsic::mips_bseti_b: 1467 case Intrinsic::mips_bseti_h: 1468 case Intrinsic::mips_bseti_w: 1469 case Intrinsic::mips_bseti_d: 1470 return lowerMSABinaryBitImmIntr(Op, DAG, ISD::OR, Op->getOperand(2), 1471 !Subtarget->isLittle()); 1472 case Intrinsic::mips_bz_b: 1473 case Intrinsic::mips_bz_h: 1474 case Intrinsic::mips_bz_w: 1475 case Intrinsic::mips_bz_d: 1476 return DAG.getNode(MipsISD::VALL_ZERO, DL, Op->getValueType(0), 1477 Op->getOperand(1)); 1478 case Intrinsic::mips_bz_v: 1479 return DAG.getNode(MipsISD::VANY_ZERO, DL, Op->getValueType(0), 1480 Op->getOperand(1)); 1481 case Intrinsic::mips_ceq_b: 1482 case Intrinsic::mips_ceq_h: 1483 case Intrinsic::mips_ceq_w: 1484 case Intrinsic::mips_ceq_d: 1485 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1486 Op->getOperand(2), ISD::SETEQ); 1487 case Intrinsic::mips_ceqi_b: 1488 case Intrinsic::mips_ceqi_h: 1489 case Intrinsic::mips_ceqi_w: 1490 case Intrinsic::mips_ceqi_d: 1491 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1492 lowerMSASplatImm(Op, 2, DAG), ISD::SETEQ); 1493 case Intrinsic::mips_cle_s_b: 1494 case Intrinsic::mips_cle_s_h: 1495 case Intrinsic::mips_cle_s_w: 1496 case Intrinsic::mips_cle_s_d: 1497 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1498 Op->getOperand(2), ISD::SETLE); 1499 case Intrinsic::mips_clei_s_b: 1500 case Intrinsic::mips_clei_s_h: 1501 case Intrinsic::mips_clei_s_w: 1502 case Intrinsic::mips_clei_s_d: 1503 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1504 lowerMSASplatImm(Op, 2, DAG), ISD::SETLE); 1505 case Intrinsic::mips_cle_u_b: 1506 case Intrinsic::mips_cle_u_h: 1507 case Intrinsic::mips_cle_u_w: 1508 case Intrinsic::mips_cle_u_d: 1509 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1510 Op->getOperand(2), ISD::SETULE); 1511 case Intrinsic::mips_clei_u_b: 1512 case Intrinsic::mips_clei_u_h: 1513 case Intrinsic::mips_clei_u_w: 1514 case Intrinsic::mips_clei_u_d: 1515 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1516 lowerMSASplatImm(Op, 2, DAG), ISD::SETULE); 1517 case Intrinsic::mips_clt_s_b: 1518 case Intrinsic::mips_clt_s_h: 1519 case Intrinsic::mips_clt_s_w: 1520 case Intrinsic::mips_clt_s_d: 1521 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1522 Op->getOperand(2), ISD::SETLT); 1523 case Intrinsic::mips_clti_s_b: 1524 case Intrinsic::mips_clti_s_h: 1525 case Intrinsic::mips_clti_s_w: 1526 case Intrinsic::mips_clti_s_d: 1527 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1528 lowerMSASplatImm(Op, 2, DAG), ISD::SETLT); 1529 case Intrinsic::mips_clt_u_b: 1530 case Intrinsic::mips_clt_u_h: 1531 case Intrinsic::mips_clt_u_w: 1532 case Intrinsic::mips_clt_u_d: 1533 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1534 Op->getOperand(2), ISD::SETULT); 1535 case Intrinsic::mips_clti_u_b: 1536 case Intrinsic::mips_clti_u_h: 1537 case Intrinsic::mips_clti_u_w: 1538 case Intrinsic::mips_clti_u_d: 1539 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1540 lowerMSASplatImm(Op, 2, DAG), ISD::SETULT); 1541 case Intrinsic::mips_copy_s_b: 1542 case Intrinsic::mips_copy_s_h: 1543 case Intrinsic::mips_copy_s_w: 1544 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT); 1545 case Intrinsic::mips_copy_s_d: 1546 // Don't lower directly into VEXTRACT_SEXT_ELT since i64 might be illegal. 1547 // Instead lower to the generic EXTRACT_VECTOR_ELT node and let the type 1548 // legalizer and EXTRACT_VECTOR_ELT lowering sort it out. 1549 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op), Op->getValueType(0), 1550 Op->getOperand(1), Op->getOperand(2)); 1551 case Intrinsic::mips_copy_u_b: 1552 case Intrinsic::mips_copy_u_h: 1553 case Intrinsic::mips_copy_u_w: 1554 return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT); 1555 case Intrinsic::mips_copy_u_d: 1556 // Don't lower directly into VEXTRACT_ZEXT_ELT since i64 might be illegal. 1557 // Instead lower to the generic EXTRACT_VECTOR_ELT node and let the type 1558 // legalizer and EXTRACT_VECTOR_ELT lowering sort it out. 1559 // 1560 // Note: When i64 is illegal, this results in copy_s.w instructions instead 1561 // of copy_u.w instructions. This makes no difference to the behaviour 1562 // since i64 is only illegal when the register file is 32-bit. 1563 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op), Op->getValueType(0), 1564 Op->getOperand(1), Op->getOperand(2)); 1565 case Intrinsic::mips_div_s_b: 1566 case Intrinsic::mips_div_s_h: 1567 case Intrinsic::mips_div_s_w: 1568 case Intrinsic::mips_div_s_d: 1569 return DAG.getNode(ISD::SDIV, DL, Op->getValueType(0), Op->getOperand(1), 1570 Op->getOperand(2)); 1571 case Intrinsic::mips_div_u_b: 1572 case Intrinsic::mips_div_u_h: 1573 case Intrinsic::mips_div_u_w: 1574 case Intrinsic::mips_div_u_d: 1575 return DAG.getNode(ISD::UDIV, DL, Op->getValueType(0), Op->getOperand(1), 1576 Op->getOperand(2)); 1577 case Intrinsic::mips_fadd_w: 1578 case Intrinsic::mips_fadd_d: 1579 return DAG.getNode(ISD::FADD, DL, Op->getValueType(0), Op->getOperand(1), 1580 Op->getOperand(2)); 1581 // Don't lower mips_fcaf_[wd] since LLVM folds SETFALSE condcodes away 1582 case Intrinsic::mips_fceq_w: 1583 case Intrinsic::mips_fceq_d: 1584 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1585 Op->getOperand(2), ISD::SETOEQ); 1586 case Intrinsic::mips_fcle_w: 1587 case Intrinsic::mips_fcle_d: 1588 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1589 Op->getOperand(2), ISD::SETOLE); 1590 case Intrinsic::mips_fclt_w: 1591 case Intrinsic::mips_fclt_d: 1592 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1593 Op->getOperand(2), ISD::SETOLT); 1594 case Intrinsic::mips_fcne_w: 1595 case Intrinsic::mips_fcne_d: 1596 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1597 Op->getOperand(2), ISD::SETONE); 1598 case Intrinsic::mips_fcor_w: 1599 case Intrinsic::mips_fcor_d: 1600 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1601 Op->getOperand(2), ISD::SETO); 1602 case Intrinsic::mips_fcueq_w: 1603 case Intrinsic::mips_fcueq_d: 1604 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1605 Op->getOperand(2), ISD::SETUEQ); 1606 case Intrinsic::mips_fcule_w: 1607 case Intrinsic::mips_fcule_d: 1608 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1609 Op->getOperand(2), ISD::SETULE); 1610 case Intrinsic::mips_fcult_w: 1611 case Intrinsic::mips_fcult_d: 1612 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1613 Op->getOperand(2), ISD::SETULT); 1614 case Intrinsic::mips_fcun_w: 1615 case Intrinsic::mips_fcun_d: 1616 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1617 Op->getOperand(2), ISD::SETUO); 1618 case Intrinsic::mips_fcune_w: 1619 case Intrinsic::mips_fcune_d: 1620 return DAG.getSetCC(DL, Op->getValueType(0), Op->getOperand(1), 1621 Op->getOperand(2), ISD::SETUNE); 1622 case Intrinsic::mips_fdiv_w: 1623 case Intrinsic::mips_fdiv_d: 1624 return DAG.getNode(ISD::FDIV, DL, Op->getValueType(0), Op->getOperand(1), 1625 Op->getOperand(2)); 1626 case Intrinsic::mips_ffint_u_w: 1627 case Intrinsic::mips_ffint_u_d: 1628 return DAG.getNode(ISD::UINT_TO_FP, DL, Op->getValueType(0), 1629 Op->getOperand(1)); 1630 case Intrinsic::mips_ffint_s_w: 1631 case Intrinsic::mips_ffint_s_d: 1632 return DAG.getNode(ISD::SINT_TO_FP, DL, Op->getValueType(0), 1633 Op->getOperand(1)); 1634 case Intrinsic::mips_fill_b: 1635 case Intrinsic::mips_fill_h: 1636 case Intrinsic::mips_fill_w: 1637 case Intrinsic::mips_fill_d: { 1638 SmallVector<SDValue, 16> Ops; 1639 EVT ResTy = Op->getValueType(0); 1640 1641 for (unsigned i = 0; i < ResTy.getVectorNumElements(); ++i) 1642 Ops.push_back(Op->getOperand(1)); 1643 1644 // If ResTy is v2i64 then the type legalizer will break this node down into 1645 // an equivalent v4i32. 1646 return DAG.getNode(ISD::BUILD_VECTOR, DL, ResTy, &Ops[0], Ops.size()); 1647 } 1648 case Intrinsic::mips_fexp2_w: 1649 case Intrinsic::mips_fexp2_d: { 1650 EVT ResTy = Op->getValueType(0); 1651 return DAG.getNode( 1652 ISD::FMUL, SDLoc(Op), ResTy, Op->getOperand(1), 1653 DAG.getNode(ISD::FEXP2, SDLoc(Op), ResTy, Op->getOperand(2))); 1654 } 1655 case Intrinsic::mips_flog2_w: 1656 case Intrinsic::mips_flog2_d: 1657 return DAG.getNode(ISD::FLOG2, DL, Op->getValueType(0), Op->getOperand(1)); 1658 case Intrinsic::mips_fmadd_w: 1659 case Intrinsic::mips_fmadd_d: 1660 return DAG.getNode(ISD::FMA, SDLoc(Op), Op->getValueType(0), 1661 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3)); 1662 case Intrinsic::mips_fmul_w: 1663 case Intrinsic::mips_fmul_d: 1664 return DAG.getNode(ISD::FMUL, DL, Op->getValueType(0), Op->getOperand(1), 1665 Op->getOperand(2)); 1666 case Intrinsic::mips_fmsub_w: 1667 case Intrinsic::mips_fmsub_d: { 1668 EVT ResTy = Op->getValueType(0); 1669 return DAG.getNode(ISD::FSUB, SDLoc(Op), ResTy, Op->getOperand(1), 1670 DAG.getNode(ISD::FMUL, SDLoc(Op), ResTy, 1671 Op->getOperand(2), Op->getOperand(3))); 1672 } 1673 case Intrinsic::mips_frint_w: 1674 case Intrinsic::mips_frint_d: 1675 return DAG.getNode(ISD::FRINT, DL, Op->getValueType(0), Op->getOperand(1)); 1676 case Intrinsic::mips_fsqrt_w: 1677 case Intrinsic::mips_fsqrt_d: 1678 return DAG.getNode(ISD::FSQRT, DL, Op->getValueType(0), Op->getOperand(1)); 1679 case Intrinsic::mips_fsub_w: 1680 case Intrinsic::mips_fsub_d: 1681 return DAG.getNode(ISD::FSUB, DL, Op->getValueType(0), Op->getOperand(1), 1682 Op->getOperand(2)); 1683 case Intrinsic::mips_ftrunc_u_w: 1684 case Intrinsic::mips_ftrunc_u_d: 1685 return DAG.getNode(ISD::FP_TO_UINT, DL, Op->getValueType(0), 1686 Op->getOperand(1)); 1687 case Intrinsic::mips_ftrunc_s_w: 1688 case Intrinsic::mips_ftrunc_s_d: 1689 return DAG.getNode(ISD::FP_TO_SINT, DL, Op->getValueType(0), 1690 Op->getOperand(1)); 1691 case Intrinsic::mips_ilvev_b: 1692 case Intrinsic::mips_ilvev_h: 1693 case Intrinsic::mips_ilvev_w: 1694 case Intrinsic::mips_ilvev_d: 1695 return DAG.getNode(MipsISD::ILVEV, DL, Op->getValueType(0), 1696 Op->getOperand(1), Op->getOperand(2)); 1697 case Intrinsic::mips_ilvl_b: 1698 case Intrinsic::mips_ilvl_h: 1699 case Intrinsic::mips_ilvl_w: 1700 case Intrinsic::mips_ilvl_d: 1701 return DAG.getNode(MipsISD::ILVL, DL, Op->getValueType(0), 1702 Op->getOperand(1), Op->getOperand(2)); 1703 case Intrinsic::mips_ilvod_b: 1704 case Intrinsic::mips_ilvod_h: 1705 case Intrinsic::mips_ilvod_w: 1706 case Intrinsic::mips_ilvod_d: 1707 return DAG.getNode(MipsISD::ILVOD, DL, Op->getValueType(0), 1708 Op->getOperand(1), Op->getOperand(2)); 1709 case Intrinsic::mips_ilvr_b: 1710 case Intrinsic::mips_ilvr_h: 1711 case Intrinsic::mips_ilvr_w: 1712 case Intrinsic::mips_ilvr_d: 1713 return DAG.getNode(MipsISD::ILVR, DL, Op->getValueType(0), 1714 Op->getOperand(1), Op->getOperand(2)); 1715 case Intrinsic::mips_insert_b: 1716 case Intrinsic::mips_insert_h: 1717 case Intrinsic::mips_insert_w: 1718 case Intrinsic::mips_insert_d: 1719 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(Op), Op->getValueType(0), 1720 Op->getOperand(1), Op->getOperand(3), Op->getOperand(2)); 1721 case Intrinsic::mips_ldi_b: 1722 case Intrinsic::mips_ldi_h: 1723 case Intrinsic::mips_ldi_w: 1724 case Intrinsic::mips_ldi_d: 1725 return lowerMSASplatImm(Op, 1, DAG); 1726 case Intrinsic::mips_lsa: { 1727 EVT ResTy = Op->getValueType(0); 1728 return DAG.getNode(ISD::ADD, SDLoc(Op), ResTy, Op->getOperand(1), 1729 DAG.getNode(ISD::SHL, SDLoc(Op), ResTy, 1730 Op->getOperand(2), Op->getOperand(3))); 1731 } 1732 case Intrinsic::mips_maddv_b: 1733 case Intrinsic::mips_maddv_h: 1734 case Intrinsic::mips_maddv_w: 1735 case Intrinsic::mips_maddv_d: { 1736 EVT ResTy = Op->getValueType(0); 1737 return DAG.getNode(ISD::ADD, SDLoc(Op), ResTy, Op->getOperand(1), 1738 DAG.getNode(ISD::MUL, SDLoc(Op), ResTy, 1739 Op->getOperand(2), Op->getOperand(3))); 1740 } 1741 case Intrinsic::mips_max_s_b: 1742 case Intrinsic::mips_max_s_h: 1743 case Intrinsic::mips_max_s_w: 1744 case Intrinsic::mips_max_s_d: 1745 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0), 1746 Op->getOperand(1), Op->getOperand(2)); 1747 case Intrinsic::mips_max_u_b: 1748 case Intrinsic::mips_max_u_h: 1749 case Intrinsic::mips_max_u_w: 1750 case Intrinsic::mips_max_u_d: 1751 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0), 1752 Op->getOperand(1), Op->getOperand(2)); 1753 case Intrinsic::mips_maxi_s_b: 1754 case Intrinsic::mips_maxi_s_h: 1755 case Intrinsic::mips_maxi_s_w: 1756 case Intrinsic::mips_maxi_s_d: 1757 return DAG.getNode(MipsISD::VSMAX, DL, Op->getValueType(0), 1758 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG)); 1759 case Intrinsic::mips_maxi_u_b: 1760 case Intrinsic::mips_maxi_u_h: 1761 case Intrinsic::mips_maxi_u_w: 1762 case Intrinsic::mips_maxi_u_d: 1763 return DAG.getNode(MipsISD::VUMAX, DL, Op->getValueType(0), 1764 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG)); 1765 case Intrinsic::mips_min_s_b: 1766 case Intrinsic::mips_min_s_h: 1767 case Intrinsic::mips_min_s_w: 1768 case Intrinsic::mips_min_s_d: 1769 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0), 1770 Op->getOperand(1), Op->getOperand(2)); 1771 case Intrinsic::mips_min_u_b: 1772 case Intrinsic::mips_min_u_h: 1773 case Intrinsic::mips_min_u_w: 1774 case Intrinsic::mips_min_u_d: 1775 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0), 1776 Op->getOperand(1), Op->getOperand(2)); 1777 case Intrinsic::mips_mini_s_b: 1778 case Intrinsic::mips_mini_s_h: 1779 case Intrinsic::mips_mini_s_w: 1780 case Intrinsic::mips_mini_s_d: 1781 return DAG.getNode(MipsISD::VSMIN, DL, Op->getValueType(0), 1782 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG)); 1783 case Intrinsic::mips_mini_u_b: 1784 case Intrinsic::mips_mini_u_h: 1785 case Intrinsic::mips_mini_u_w: 1786 case Intrinsic::mips_mini_u_d: 1787 return DAG.getNode(MipsISD::VUMIN, DL, Op->getValueType(0), 1788 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG)); 1789 case Intrinsic::mips_mod_s_b: 1790 case Intrinsic::mips_mod_s_h: 1791 case Intrinsic::mips_mod_s_w: 1792 case Intrinsic::mips_mod_s_d: 1793 return DAG.getNode(ISD::SREM, DL, Op->getValueType(0), Op->getOperand(1), 1794 Op->getOperand(2)); 1795 case Intrinsic::mips_mod_u_b: 1796 case Intrinsic::mips_mod_u_h: 1797 case Intrinsic::mips_mod_u_w: 1798 case Intrinsic::mips_mod_u_d: 1799 return DAG.getNode(ISD::UREM, DL, Op->getValueType(0), Op->getOperand(1), 1800 Op->getOperand(2)); 1801 case Intrinsic::mips_mulv_b: 1802 case Intrinsic::mips_mulv_h: 1803 case Intrinsic::mips_mulv_w: 1804 case Intrinsic::mips_mulv_d: 1805 return DAG.getNode(ISD::MUL, DL, Op->getValueType(0), Op->getOperand(1), 1806 Op->getOperand(2)); 1807 case Intrinsic::mips_msubv_b: 1808 case Intrinsic::mips_msubv_h: 1809 case Intrinsic::mips_msubv_w: 1810 case Intrinsic::mips_msubv_d: { 1811 EVT ResTy = Op->getValueType(0); 1812 return DAG.getNode(ISD::SUB, SDLoc(Op), ResTy, Op->getOperand(1), 1813 DAG.getNode(ISD::MUL, SDLoc(Op), ResTy, 1814 Op->getOperand(2), Op->getOperand(3))); 1815 } 1816 case Intrinsic::mips_nlzc_b: 1817 case Intrinsic::mips_nlzc_h: 1818 case Intrinsic::mips_nlzc_w: 1819 case Intrinsic::mips_nlzc_d: 1820 return DAG.getNode(ISD::CTLZ, DL, Op->getValueType(0), Op->getOperand(1)); 1821 case Intrinsic::mips_nor_v: { 1822 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0), 1823 Op->getOperand(1), Op->getOperand(2)); 1824 return DAG.getNOT(DL, Res, Res->getValueType(0)); 1825 } 1826 case Intrinsic::mips_nori_b: { 1827 SDValue Res = DAG.getNode(ISD::OR, DL, Op->getValueType(0), 1828 Op->getOperand(1), 1829 lowerMSASplatImm(Op, 2, DAG)); 1830 return DAG.getNOT(DL, Res, Res->getValueType(0)); 1831 } 1832 case Intrinsic::mips_or_v: 1833 return DAG.getNode(ISD::OR, DL, Op->getValueType(0), Op->getOperand(1), 1834 Op->getOperand(2)); 1835 case Intrinsic::mips_ori_b: 1836 return DAG.getNode(ISD::OR, DL, Op->getValueType(0), 1837 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG)); 1838 case Intrinsic::mips_pckev_b: 1839 case Intrinsic::mips_pckev_h: 1840 case Intrinsic::mips_pckev_w: 1841 case Intrinsic::mips_pckev_d: 1842 return DAG.getNode(MipsISD::PCKEV, DL, Op->getValueType(0), 1843 Op->getOperand(1), Op->getOperand(2)); 1844 case Intrinsic::mips_pckod_b: 1845 case Intrinsic::mips_pckod_h: 1846 case Intrinsic::mips_pckod_w: 1847 case Intrinsic::mips_pckod_d: 1848 return DAG.getNode(MipsISD::PCKOD, DL, Op->getValueType(0), 1849 Op->getOperand(1), Op->getOperand(2)); 1850 case Intrinsic::mips_pcnt_b: 1851 case Intrinsic::mips_pcnt_h: 1852 case Intrinsic::mips_pcnt_w: 1853 case Intrinsic::mips_pcnt_d: 1854 return DAG.getNode(ISD::CTPOP, DL, Op->getValueType(0), Op->getOperand(1)); 1855 case Intrinsic::mips_shf_b: 1856 case Intrinsic::mips_shf_h: 1857 case Intrinsic::mips_shf_w: 1858 return DAG.getNode(MipsISD::SHF, DL, Op->getValueType(0), 1859 Op->getOperand(2), Op->getOperand(1)); 1860 case Intrinsic::mips_sll_b: 1861 case Intrinsic::mips_sll_h: 1862 case Intrinsic::mips_sll_w: 1863 case Intrinsic::mips_sll_d: 1864 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0), Op->getOperand(1), 1865 Op->getOperand(2)); 1866 case Intrinsic::mips_slli_b: 1867 case Intrinsic::mips_slli_h: 1868 case Intrinsic::mips_slli_w: 1869 case Intrinsic::mips_slli_d: 1870 return DAG.getNode(ISD::SHL, DL, Op->getValueType(0), 1871 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG)); 1872 case Intrinsic::mips_splat_b: 1873 case Intrinsic::mips_splat_h: 1874 case Intrinsic::mips_splat_w: 1875 case Intrinsic::mips_splat_d: 1876 // We can't lower via VECTOR_SHUFFLE because it requires constant shuffle 1877 // masks, nor can we lower via BUILD_VECTOR & EXTRACT_VECTOR_ELT because 1878 // EXTRACT_VECTOR_ELT can't extract i64's on MIPS32. 1879 // Instead we lower to MipsISD::VSHF and match from there. 1880 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0), 1881 lowerMSASplatImm(Op, 2, DAG), Op->getOperand(1), 1882 Op->getOperand(1)); 1883 case Intrinsic::mips_splati_b: 1884 case Intrinsic::mips_splati_h: 1885 case Intrinsic::mips_splati_w: 1886 case Intrinsic::mips_splati_d: 1887 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0), 1888 lowerMSASplatImm(Op, 2, DAG), Op->getOperand(1), 1889 Op->getOperand(1)); 1890 case Intrinsic::mips_sra_b: 1891 case Intrinsic::mips_sra_h: 1892 case Intrinsic::mips_sra_w: 1893 case Intrinsic::mips_sra_d: 1894 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0), Op->getOperand(1), 1895 Op->getOperand(2)); 1896 case Intrinsic::mips_srai_b: 1897 case Intrinsic::mips_srai_h: 1898 case Intrinsic::mips_srai_w: 1899 case Intrinsic::mips_srai_d: 1900 return DAG.getNode(ISD::SRA, DL, Op->getValueType(0), 1901 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG)); 1902 case Intrinsic::mips_srl_b: 1903 case Intrinsic::mips_srl_h: 1904 case Intrinsic::mips_srl_w: 1905 case Intrinsic::mips_srl_d: 1906 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0), Op->getOperand(1), 1907 Op->getOperand(2)); 1908 case Intrinsic::mips_srli_b: 1909 case Intrinsic::mips_srli_h: 1910 case Intrinsic::mips_srli_w: 1911 case Intrinsic::mips_srli_d: 1912 return DAG.getNode(ISD::SRL, DL, Op->getValueType(0), 1913 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG)); 1914 case Intrinsic::mips_subv_b: 1915 case Intrinsic::mips_subv_h: 1916 case Intrinsic::mips_subv_w: 1917 case Intrinsic::mips_subv_d: 1918 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0), Op->getOperand(1), 1919 Op->getOperand(2)); 1920 case Intrinsic::mips_subvi_b: 1921 case Intrinsic::mips_subvi_h: 1922 case Intrinsic::mips_subvi_w: 1923 case Intrinsic::mips_subvi_d: 1924 return DAG.getNode(ISD::SUB, DL, Op->getValueType(0), 1925 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG)); 1926 case Intrinsic::mips_vshf_b: 1927 case Intrinsic::mips_vshf_h: 1928 case Intrinsic::mips_vshf_w: 1929 case Intrinsic::mips_vshf_d: 1930 return DAG.getNode(MipsISD::VSHF, DL, Op->getValueType(0), 1931 Op->getOperand(1), Op->getOperand(2), Op->getOperand(3)); 1932 case Intrinsic::mips_xor_v: 1933 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0), Op->getOperand(1), 1934 Op->getOperand(2)); 1935 case Intrinsic::mips_xori_b: 1936 return DAG.getNode(ISD::XOR, DL, Op->getValueType(0), 1937 Op->getOperand(1), lowerMSASplatImm(Op, 2, DAG)); 1938 } 1939} 1940 1941static SDValue lowerMSALoadIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) { 1942 SDLoc DL(Op); 1943 SDValue ChainIn = Op->getOperand(0); 1944 SDValue Address = Op->getOperand(2); 1945 SDValue Offset = Op->getOperand(3); 1946 EVT ResTy = Op->getValueType(0); 1947 EVT PtrTy = Address->getValueType(0); 1948 1949 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset); 1950 1951 return DAG.getLoad(ResTy, DL, ChainIn, Address, MachinePointerInfo(), false, 1952 false, false, 16); 1953} 1954 1955SDValue MipsSETargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op, 1956 SelectionDAG &DAG) const { 1957 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue(); 1958 switch (Intr) { 1959 default: 1960 return SDValue(); 1961 case Intrinsic::mips_extp: 1962 return lowerDSPIntr(Op, DAG, MipsISD::EXTP); 1963 case Intrinsic::mips_extpdp: 1964 return lowerDSPIntr(Op, DAG, MipsISD::EXTPDP); 1965 case Intrinsic::mips_extr_w: 1966 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_W); 1967 case Intrinsic::mips_extr_r_w: 1968 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W); 1969 case Intrinsic::mips_extr_rs_w: 1970 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W); 1971 case Intrinsic::mips_extr_s_h: 1972 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H); 1973 case Intrinsic::mips_mthlip: 1974 return lowerDSPIntr(Op, DAG, MipsISD::MTHLIP); 1975 case Intrinsic::mips_mulsaq_s_w_ph: 1976 return lowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH); 1977 case Intrinsic::mips_maq_s_w_phl: 1978 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL); 1979 case Intrinsic::mips_maq_s_w_phr: 1980 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR); 1981 case Intrinsic::mips_maq_sa_w_phl: 1982 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL); 1983 case Intrinsic::mips_maq_sa_w_phr: 1984 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR); 1985 case Intrinsic::mips_dpaq_s_w_ph: 1986 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH); 1987 case Intrinsic::mips_dpsq_s_w_ph: 1988 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH); 1989 case Intrinsic::mips_dpaq_sa_l_w: 1990 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W); 1991 case Intrinsic::mips_dpsq_sa_l_w: 1992 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W); 1993 case Intrinsic::mips_dpaqx_s_w_ph: 1994 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH); 1995 case Intrinsic::mips_dpaqx_sa_w_ph: 1996 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH); 1997 case Intrinsic::mips_dpsqx_s_w_ph: 1998 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH); 1999 case Intrinsic::mips_dpsqx_sa_w_ph: 2000 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH); 2001 case Intrinsic::mips_ld_b: 2002 case Intrinsic::mips_ld_h: 2003 case Intrinsic::mips_ld_w: 2004 case Intrinsic::mips_ld_d: 2005 return lowerMSALoadIntr(Op, DAG, Intr); 2006 } 2007} 2008 2009static SDValue lowerMSAStoreIntr(SDValue Op, SelectionDAG &DAG, unsigned Intr) { 2010 SDLoc DL(Op); 2011 SDValue ChainIn = Op->getOperand(0); 2012 SDValue Value = Op->getOperand(2); 2013 SDValue Address = Op->getOperand(3); 2014 SDValue Offset = Op->getOperand(4); 2015 EVT PtrTy = Address->getValueType(0); 2016 2017 Address = DAG.getNode(ISD::ADD, DL, PtrTy, Address, Offset); 2018 2019 return DAG.getStore(ChainIn, DL, Value, Address, MachinePointerInfo(), false, 2020 false, 16); 2021} 2022 2023SDValue MipsSETargetLowering::lowerINTRINSIC_VOID(SDValue Op, 2024 SelectionDAG &DAG) const { 2025 unsigned Intr = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue(); 2026 switch (Intr) { 2027 default: 2028 return SDValue(); 2029 case Intrinsic::mips_st_b: 2030 case Intrinsic::mips_st_h: 2031 case Intrinsic::mips_st_w: 2032 case Intrinsic::mips_st_d: 2033 return lowerMSAStoreIntr(Op, DAG, Intr); 2034 } 2035} 2036 2037/// \brief Check if the given BuildVectorSDNode is a splat. 2038/// This method currently relies on DAG nodes being reused when equivalent, 2039/// so it's possible for this to return false even when isConstantSplat returns 2040/// true. 2041static bool isSplatVector(const BuildVectorSDNode *N) { 2042 unsigned int nOps = N->getNumOperands(); 2043 assert(nOps > 1 && "isSplatVector has 0 or 1 sized build vector"); 2044 2045 SDValue Operand0 = N->getOperand(0); 2046 2047 for (unsigned int i = 1; i < nOps; ++i) { 2048 if (N->getOperand(i) != Operand0) 2049 return false; 2050 } 2051 2052 return true; 2053} 2054 2055// Lower ISD::EXTRACT_VECTOR_ELT into MipsISD::VEXTRACT_SEXT_ELT. 2056// 2057// The non-value bits resulting from ISD::EXTRACT_VECTOR_ELT are undefined. We 2058// choose to sign-extend but we could have equally chosen zero-extend. The 2059// DAGCombiner will fold any sign/zero extension of the ISD::EXTRACT_VECTOR_ELT 2060// result into this node later (possibly changing it to a zero-extend in the 2061// process). 2062SDValue MipsSETargetLowering:: 2063lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const { 2064 SDLoc DL(Op); 2065 EVT ResTy = Op->getValueType(0); 2066 SDValue Op0 = Op->getOperand(0); 2067 EVT VecTy = Op0->getValueType(0); 2068 2069 if (!VecTy.is128BitVector()) 2070 return SDValue(); 2071 2072 if (ResTy.isInteger()) { 2073 SDValue Op1 = Op->getOperand(1); 2074 EVT EltTy = VecTy.getVectorElementType(); 2075 return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, DL, ResTy, Op0, Op1, 2076 DAG.getValueType(EltTy)); 2077 } 2078 2079 return Op; 2080} 2081 2082static bool isConstantOrUndef(const SDValue Op) { 2083 if (Op->getOpcode() == ISD::UNDEF) 2084 return true; 2085 if (dyn_cast<ConstantSDNode>(Op)) 2086 return true; 2087 if (dyn_cast<ConstantFPSDNode>(Op)) 2088 return true; 2089 return false; 2090} 2091 2092static bool isConstantOrUndefBUILD_VECTOR(const BuildVectorSDNode *Op) { 2093 for (unsigned i = 0; i < Op->getNumOperands(); ++i) 2094 if (isConstantOrUndef(Op->getOperand(i))) 2095 return true; 2096 return false; 2097} 2098 2099// Lowers ISD::BUILD_VECTOR into appropriate SelectionDAG nodes for the 2100// backend. 2101// 2102// Lowers according to the following rules: 2103// - Constant splats are legal as-is as long as the SplatBitSize is a power of 2104// 2 less than or equal to 64 and the value fits into a signed 10-bit 2105// immediate 2106// - Constant splats are lowered to bitconverted BUILD_VECTORs if SplatBitSize 2107// is a power of 2 less than or equal to 64 and the value does not fit into a 2108// signed 10-bit immediate 2109// - Non-constant splats are legal as-is. 2110// - Non-constant non-splats are lowered to sequences of INSERT_VECTOR_ELT. 2111// - All others are illegal and must be expanded. 2112SDValue MipsSETargetLowering::lowerBUILD_VECTOR(SDValue Op, 2113 SelectionDAG &DAG) const { 2114 BuildVectorSDNode *Node = cast<BuildVectorSDNode>(Op); 2115 EVT ResTy = Op->getValueType(0); 2116 SDLoc DL(Op); 2117 APInt SplatValue, SplatUndef; 2118 unsigned SplatBitSize; 2119 bool HasAnyUndefs; 2120 2121 if (!Subtarget->hasMSA() || !ResTy.is128BitVector()) 2122 return SDValue(); 2123 2124 if (Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, 2125 HasAnyUndefs, 8, 2126 !Subtarget->isLittle()) && SplatBitSize <= 64) { 2127 // We can only cope with 8, 16, 32, or 64-bit elements 2128 if (SplatBitSize != 8 && SplatBitSize != 16 && SplatBitSize != 32 && 2129 SplatBitSize != 64) 2130 return SDValue(); 2131 2132 // If the value fits into a simm10 then we can use ldi.[bhwd] 2133 if (SplatValue.isSignedIntN(10)) 2134 return Op; 2135 2136 EVT ViaVecTy; 2137 2138 switch (SplatBitSize) { 2139 default: 2140 return SDValue(); 2141 case 8: 2142 ViaVecTy = MVT::v16i8; 2143 break; 2144 case 16: 2145 ViaVecTy = MVT::v8i16; 2146 break; 2147 case 32: 2148 ViaVecTy = MVT::v4i32; 2149 break; 2150 case 64: 2151 // There's no fill.d to fall back on for 64-bit values 2152 return SDValue(); 2153 } 2154 2155 SmallVector<SDValue, 16> Ops; 2156 SDValue Constant = DAG.getConstant(SplatValue.sextOrSelf(32), MVT::i32); 2157 2158 for (unsigned i = 0; i < ViaVecTy.getVectorNumElements(); ++i) 2159 Ops.push_back(Constant); 2160 2161 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Node), ViaVecTy, 2162 &Ops[0], Ops.size()); 2163 2164 if (ViaVecTy != ResTy) 2165 Result = DAG.getNode(ISD::BITCAST, SDLoc(Node), ResTy, Result); 2166 2167 return Result; 2168 } else if (isSplatVector(Node)) 2169 return Op; 2170 else if (!isConstantOrUndefBUILD_VECTOR(Node)) { 2171 // Use INSERT_VECTOR_ELT operations rather than expand to stores. 2172 // The resulting code is the same length as the expansion, but it doesn't 2173 // use memory operations 2174 EVT ResTy = Node->getValueType(0); 2175 2176 assert(ResTy.isVector()); 2177 2178 unsigned NumElts = ResTy.getVectorNumElements(); 2179 SDValue Vector = DAG.getUNDEF(ResTy); 2180 for (unsigned i = 0; i < NumElts; ++i) { 2181 Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Vector, 2182 Node->getOperand(i), 2183 DAG.getConstant(i, MVT::i32)); 2184 } 2185 return Vector; 2186 } 2187 2188 return SDValue(); 2189} 2190 2191// Lower VECTOR_SHUFFLE into SHF (if possible). 2192// 2193// SHF splits the vector into blocks of four elements, then shuffles these 2194// elements according to a <4 x i2> constant (encoded as an integer immediate). 2195// 2196// It is therefore possible to lower into SHF when the mask takes the form: 2197// <a, b, c, d, a+4, b+4, c+4, d+4, a+8, b+8, c+8, d+8, ...> 2198// When undef's appear they are treated as if they were whatever value is 2199// necessary in order to fit the above form. 2200// 2201// For example: 2202// %2 = shufflevector <8 x i16> %0, <8 x i16> undef, 2203// <8 x i32> <i32 3, i32 2, i32 1, i32 0, 2204// i32 7, i32 6, i32 5, i32 4> 2205// is lowered to: 2206// (SHF_H $w0, $w1, 27) 2207// where the 27 comes from: 2208// 3 + (2 << 2) + (1 << 4) + (0 << 6) 2209static SDValue lowerVECTOR_SHUFFLE_SHF(SDValue Op, EVT ResTy, 2210 SmallVector<int, 16> Indices, 2211 SelectionDAG &DAG) { 2212 int SHFIndices[4] = { -1, -1, -1, -1 }; 2213 2214 if (Indices.size() < 4) 2215 return SDValue(); 2216 2217 for (unsigned i = 0; i < 4; ++i) { 2218 for (unsigned j = i; j < Indices.size(); j += 4) { 2219 int Idx = Indices[j]; 2220 2221 // Convert from vector index to 4-element subvector index 2222 // If an index refers to an element outside of the subvector then give up 2223 if (Idx != -1) { 2224 Idx -= 4 * (j / 4); 2225 if (Idx < 0 || Idx >= 4) 2226 return SDValue(); 2227 } 2228 2229 // If the mask has an undef, replace it with the current index. 2230 // Note that it might still be undef if the current index is also undef 2231 if (SHFIndices[i] == -1) 2232 SHFIndices[i] = Idx; 2233 2234 // Check that non-undef values are the same as in the mask. If they 2235 // aren't then give up 2236 if (!(Idx == -1 || Idx == SHFIndices[i])) 2237 return SDValue(); 2238 } 2239 } 2240 2241 // Calculate the immediate. Replace any remaining undefs with zero 2242 APInt Imm(32, 0); 2243 for (int i = 3; i >= 0; --i) { 2244 int Idx = SHFIndices[i]; 2245 2246 if (Idx == -1) 2247 Idx = 0; 2248 2249 Imm <<= 2; 2250 Imm |= Idx & 0x3; 2251 } 2252 2253 return DAG.getNode(MipsISD::SHF, SDLoc(Op), ResTy, 2254 DAG.getConstant(Imm, MVT::i32), Op->getOperand(0)); 2255} 2256 2257// Lower VECTOR_SHUFFLE into ILVEV (if possible). 2258// 2259// ILVEV interleaves the even elements from each vector. 2260// 2261// It is possible to lower into ILVEV when the mask takes the form: 2262// <0, n, 2, n+2, 4, n+4, ...> 2263// where n is the number of elements in the vector. 2264// 2265// When undef's appear in the mask they are treated as if they were whatever 2266// value is necessary in order to fit the above form. 2267static SDValue lowerVECTOR_SHUFFLE_ILVEV(SDValue Op, EVT ResTy, 2268 SmallVector<int, 16> Indices, 2269 SelectionDAG &DAG) { 2270 assert ((Indices.size() % 2) == 0); 2271 int WsIdx = 0; 2272 int WtIdx = ResTy.getVectorNumElements(); 2273 2274 for (unsigned i = 0; i < Indices.size(); i += 2) { 2275 if (Indices[i] != -1 && Indices[i] != WsIdx) 2276 return SDValue(); 2277 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx) 2278 return SDValue(); 2279 WsIdx += 2; 2280 WtIdx += 2; 2281 } 2282 2283 return DAG.getNode(MipsISD::ILVEV, SDLoc(Op), ResTy, Op->getOperand(0), 2284 Op->getOperand(1)); 2285} 2286 2287// Lower VECTOR_SHUFFLE into ILVOD (if possible). 2288// 2289// ILVOD interleaves the odd elements from each vector. 2290// 2291// It is possible to lower into ILVOD when the mask takes the form: 2292// <1, n+1, 3, n+3, 5, n+5, ...> 2293// where n is the number of elements in the vector. 2294// 2295// When undef's appear in the mask they are treated as if they were whatever 2296// value is necessary in order to fit the above form. 2297static SDValue lowerVECTOR_SHUFFLE_ILVOD(SDValue Op, EVT ResTy, 2298 SmallVector<int, 16> Indices, 2299 SelectionDAG &DAG) { 2300 assert ((Indices.size() % 2) == 0); 2301 int WsIdx = 1; 2302 int WtIdx = ResTy.getVectorNumElements() + 1; 2303 2304 for (unsigned i = 0; i < Indices.size(); i += 2) { 2305 if (Indices[i] != -1 && Indices[i] != WsIdx) 2306 return SDValue(); 2307 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx) 2308 return SDValue(); 2309 WsIdx += 2; 2310 WtIdx += 2; 2311 } 2312 2313 return DAG.getNode(MipsISD::ILVOD, SDLoc(Op), ResTy, Op->getOperand(0), 2314 Op->getOperand(1)); 2315} 2316 2317// Lower VECTOR_SHUFFLE into ILVL (if possible). 2318// 2319// ILVL interleaves consecutive elements from the left half of each vector. 2320// 2321// It is possible to lower into ILVL when the mask takes the form: 2322// <0, n, 1, n+1, 2, n+2, ...> 2323// where n is the number of elements in the vector. 2324// 2325// When undef's appear in the mask they are treated as if they were whatever 2326// value is necessary in order to fit the above form. 2327static SDValue lowerVECTOR_SHUFFLE_ILVL(SDValue Op, EVT ResTy, 2328 SmallVector<int, 16> Indices, 2329 SelectionDAG &DAG) { 2330 assert ((Indices.size() % 2) == 0); 2331 int WsIdx = 0; 2332 int WtIdx = ResTy.getVectorNumElements(); 2333 2334 for (unsigned i = 0; i < Indices.size(); i += 2) { 2335 if (Indices[i] != -1 && Indices[i] != WsIdx) 2336 return SDValue(); 2337 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx) 2338 return SDValue(); 2339 WsIdx ++; 2340 WtIdx ++; 2341 } 2342 2343 return DAG.getNode(MipsISD::ILVL, SDLoc(Op), ResTy, Op->getOperand(0), 2344 Op->getOperand(1)); 2345} 2346 2347// Lower VECTOR_SHUFFLE into ILVR (if possible). 2348// 2349// ILVR interleaves consecutive elements from the right half of each vector. 2350// 2351// It is possible to lower into ILVR when the mask takes the form: 2352// <x, n+x, x+1, n+x+1, x+2, n+x+2, ...> 2353// where n is the number of elements in the vector and x is half n. 2354// 2355// When undef's appear in the mask they are treated as if they were whatever 2356// value is necessary in order to fit the above form. 2357static SDValue lowerVECTOR_SHUFFLE_ILVR(SDValue Op, EVT ResTy, 2358 SmallVector<int, 16> Indices, 2359 SelectionDAG &DAG) { 2360 assert ((Indices.size() % 2) == 0); 2361 unsigned NumElts = ResTy.getVectorNumElements(); 2362 int WsIdx = NumElts / 2; 2363 int WtIdx = NumElts + NumElts / 2; 2364 2365 for (unsigned i = 0; i < Indices.size(); i += 2) { 2366 if (Indices[i] != -1 && Indices[i] != WsIdx) 2367 return SDValue(); 2368 if (Indices[i+1] != -1 && Indices[i+1] != WtIdx) 2369 return SDValue(); 2370 WsIdx ++; 2371 WtIdx ++; 2372 } 2373 2374 return DAG.getNode(MipsISD::ILVR, SDLoc(Op), ResTy, Op->getOperand(0), 2375 Op->getOperand(1)); 2376} 2377 2378// Lower VECTOR_SHUFFLE into PCKEV (if possible). 2379// 2380// PCKEV copies the even elements of each vector into the result vector. 2381// 2382// It is possible to lower into PCKEV when the mask takes the form: 2383// <0, 2, 4, ..., n, n+2, n+4, ...> 2384// where n is the number of elements in the vector. 2385// 2386// When undef's appear in the mask they are treated as if they were whatever 2387// value is necessary in order to fit the above form. 2388static SDValue lowerVECTOR_SHUFFLE_PCKEV(SDValue Op, EVT ResTy, 2389 SmallVector<int, 16> Indices, 2390 SelectionDAG &DAG) { 2391 assert ((Indices.size() % 2) == 0); 2392 int Idx = 0; 2393 2394 for (unsigned i = 0; i < Indices.size(); ++i) { 2395 if (Indices[i] != -1 && Indices[i] != Idx) 2396 return SDValue(); 2397 Idx += 2; 2398 } 2399 2400 return DAG.getNode(MipsISD::PCKEV, SDLoc(Op), ResTy, Op->getOperand(0), 2401 Op->getOperand(1)); 2402} 2403 2404// Lower VECTOR_SHUFFLE into PCKOD (if possible). 2405// 2406// PCKOD copies the odd elements of each vector into the result vector. 2407// 2408// It is possible to lower into PCKOD when the mask takes the form: 2409// <1, 3, 5, ..., n+1, n+3, n+5, ...> 2410// where n is the number of elements in the vector. 2411// 2412// When undef's appear in the mask they are treated as if they were whatever 2413// value is necessary in order to fit the above form. 2414static SDValue lowerVECTOR_SHUFFLE_PCKOD(SDValue Op, EVT ResTy, 2415 SmallVector<int, 16> Indices, 2416 SelectionDAG &DAG) { 2417 assert ((Indices.size() % 2) == 0); 2418 int Idx = 1; 2419 2420 for (unsigned i = 0; i < Indices.size(); ++i) { 2421 if (Indices[i] != -1 && Indices[i] != Idx) 2422 return SDValue(); 2423 Idx += 2; 2424 } 2425 2426 return DAG.getNode(MipsISD::PCKOD, SDLoc(Op), ResTy, Op->getOperand(0), 2427 Op->getOperand(1)); 2428} 2429 2430// Lower VECTOR_SHUFFLE into VSHF. 2431// 2432// This mostly consists of converting the shuffle indices in Indices into a 2433// BUILD_VECTOR and adding it as an operand to the resulting VSHF. There is 2434// also code to eliminate unused operands of the VECTOR_SHUFFLE. For example, 2435// if the type is v8i16 and all the indices are less than 8 then the second 2436// operand is unused and can be replaced with anything. We choose to replace it 2437// with the used operand since this reduces the number of instructions overall. 2438static SDValue lowerVECTOR_SHUFFLE_VSHF(SDValue Op, EVT ResTy, 2439 SmallVector<int, 16> Indices, 2440 SelectionDAG &DAG) { 2441 SmallVector<SDValue, 16> Ops; 2442 SDValue Op0; 2443 SDValue Op1; 2444 EVT MaskVecTy = ResTy.changeVectorElementTypeToInteger(); 2445 EVT MaskEltTy = MaskVecTy.getVectorElementType(); 2446 bool Using1stVec = false; 2447 bool Using2ndVec = false; 2448 SDLoc DL(Op); 2449 int ResTyNumElts = ResTy.getVectorNumElements(); 2450 2451 for (int i = 0; i < ResTyNumElts; ++i) { 2452 // Idx == -1 means UNDEF 2453 int Idx = Indices[i]; 2454 2455 if (0 <= Idx && Idx < ResTyNumElts) 2456 Using1stVec = true; 2457 if (ResTyNumElts <= Idx && Idx < ResTyNumElts * 2) 2458 Using2ndVec = true; 2459 } 2460 2461 for (SmallVector<int, 16>::iterator I = Indices.begin(); I != Indices.end(); 2462 ++I) 2463 Ops.push_back(DAG.getTargetConstant(*I, MaskEltTy)); 2464 2465 SDValue MaskVec = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVecTy, &Ops[0], 2466 Ops.size()); 2467 2468 if (Using1stVec && Using2ndVec) { 2469 Op0 = Op->getOperand(0); 2470 Op1 = Op->getOperand(1); 2471 } else if (Using1stVec) 2472 Op0 = Op1 = Op->getOperand(0); 2473 else if (Using2ndVec) 2474 Op0 = Op1 = Op->getOperand(1); 2475 else 2476 llvm_unreachable("shuffle vector mask references neither vector operand?"); 2477 2478 return DAG.getNode(MipsISD::VSHF, DL, ResTy, MaskVec, Op0, Op1); 2479} 2480 2481// Lower VECTOR_SHUFFLE into one of a number of instructions depending on the 2482// indices in the shuffle. 2483SDValue MipsSETargetLowering::lowerVECTOR_SHUFFLE(SDValue Op, 2484 SelectionDAG &DAG) const { 2485 ShuffleVectorSDNode *Node = cast<ShuffleVectorSDNode>(Op); 2486 EVT ResTy = Op->getValueType(0); 2487 2488 if (!ResTy.is128BitVector()) 2489 return SDValue(); 2490 2491 int ResTyNumElts = ResTy.getVectorNumElements(); 2492 SmallVector<int, 16> Indices; 2493 2494 for (int i = 0; i < ResTyNumElts; ++i) 2495 Indices.push_back(Node->getMaskElt(i)); 2496 2497 SDValue Result = lowerVECTOR_SHUFFLE_SHF(Op, ResTy, Indices, DAG); 2498 if (Result.getNode()) 2499 return Result; 2500 Result = lowerVECTOR_SHUFFLE_ILVEV(Op, ResTy, Indices, DAG); 2501 if (Result.getNode()) 2502 return Result; 2503 Result = lowerVECTOR_SHUFFLE_ILVOD(Op, ResTy, Indices, DAG); 2504 if (Result.getNode()) 2505 return Result; 2506 Result = lowerVECTOR_SHUFFLE_ILVL(Op, ResTy, Indices, DAG); 2507 if (Result.getNode()) 2508 return Result; 2509 Result = lowerVECTOR_SHUFFLE_ILVR(Op, ResTy, Indices, DAG); 2510 if (Result.getNode()) 2511 return Result; 2512 Result = lowerVECTOR_SHUFFLE_PCKEV(Op, ResTy, Indices, DAG); 2513 if (Result.getNode()) 2514 return Result; 2515 Result = lowerVECTOR_SHUFFLE_PCKOD(Op, ResTy, Indices, DAG); 2516 if (Result.getNode()) 2517 return Result; 2518 return lowerVECTOR_SHUFFLE_VSHF(Op, ResTy, Indices, DAG); 2519} 2520 2521MachineBasicBlock * MipsSETargetLowering:: 2522emitBPOSGE32(MachineInstr *MI, MachineBasicBlock *BB) const{ 2523 // $bb: 2524 // bposge32_pseudo $vr0 2525 // => 2526 // $bb: 2527 // bposge32 $tbb 2528 // $fbb: 2529 // li $vr2, 0 2530 // b $sink 2531 // $tbb: 2532 // li $vr1, 1 2533 // $sink: 2534 // $vr0 = phi($vr2, $fbb, $vr1, $tbb) 2535 2536 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo(); 2537 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 2538 const TargetRegisterClass *RC = &Mips::GPR32RegClass; 2539 DebugLoc DL = MI->getDebugLoc(); 2540 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 2541 MachineFunction::iterator It = llvm::next(MachineFunction::iterator(BB)); 2542 MachineFunction *F = BB->getParent(); 2543 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB); 2544 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB); 2545 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB); 2546 F->insert(It, FBB); 2547 F->insert(It, TBB); 2548 F->insert(It, Sink); 2549 2550 // Transfer the remainder of BB and its successor edges to Sink. 2551 Sink->splice(Sink->begin(), BB, llvm::next(MachineBasicBlock::iterator(MI)), 2552 BB->end()); 2553 Sink->transferSuccessorsAndUpdatePHIs(BB); 2554 2555 // Add successors. 2556 BB->addSuccessor(FBB); 2557 BB->addSuccessor(TBB); 2558 FBB->addSuccessor(Sink); 2559 TBB->addSuccessor(Sink); 2560 2561 // Insert the real bposge32 instruction to $BB. 2562 BuildMI(BB, DL, TII->get(Mips::BPOSGE32)).addMBB(TBB); 2563 2564 // Fill $FBB. 2565 unsigned VR2 = RegInfo.createVirtualRegister(RC); 2566 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), VR2) 2567 .addReg(Mips::ZERO).addImm(0); 2568 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink); 2569 2570 // Fill $TBB. 2571 unsigned VR1 = RegInfo.createVirtualRegister(RC); 2572 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), VR1) 2573 .addReg(Mips::ZERO).addImm(1); 2574 2575 // Insert phi function to $Sink. 2576 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI), 2577 MI->getOperand(0).getReg()) 2578 .addReg(VR2).addMBB(FBB).addReg(VR1).addMBB(TBB); 2579 2580 MI->eraseFromParent(); // The pseudo instruction is gone now. 2581 return Sink; 2582} 2583 2584MachineBasicBlock * MipsSETargetLowering:: 2585emitMSACBranchPseudo(MachineInstr *MI, MachineBasicBlock *BB, 2586 unsigned BranchOp) const{ 2587 // $bb: 2588 // vany_nonzero $rd, $ws 2589 // => 2590 // $bb: 2591 // bnz.b $ws, $tbb 2592 // b $fbb 2593 // $fbb: 2594 // li $rd1, 0 2595 // b $sink 2596 // $tbb: 2597 // li $rd2, 1 2598 // $sink: 2599 // $rd = phi($rd1, $fbb, $rd2, $tbb) 2600 2601 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo(); 2602 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 2603 const TargetRegisterClass *RC = &Mips::GPR32RegClass; 2604 DebugLoc DL = MI->getDebugLoc(); 2605 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 2606 MachineFunction::iterator It = llvm::next(MachineFunction::iterator(BB)); 2607 MachineFunction *F = BB->getParent(); 2608 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB); 2609 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB); 2610 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB); 2611 F->insert(It, FBB); 2612 F->insert(It, TBB); 2613 F->insert(It, Sink); 2614 2615 // Transfer the remainder of BB and its successor edges to Sink. 2616 Sink->splice(Sink->begin(), BB, llvm::next(MachineBasicBlock::iterator(MI)), 2617 BB->end()); 2618 Sink->transferSuccessorsAndUpdatePHIs(BB); 2619 2620 // Add successors. 2621 BB->addSuccessor(FBB); 2622 BB->addSuccessor(TBB); 2623 FBB->addSuccessor(Sink); 2624 TBB->addSuccessor(Sink); 2625 2626 // Insert the real bnz.b instruction to $BB. 2627 BuildMI(BB, DL, TII->get(BranchOp)) 2628 .addReg(MI->getOperand(1).getReg()) 2629 .addMBB(TBB); 2630 2631 // Fill $FBB. 2632 unsigned RD1 = RegInfo.createVirtualRegister(RC); 2633 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), RD1) 2634 .addReg(Mips::ZERO).addImm(0); 2635 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink); 2636 2637 // Fill $TBB. 2638 unsigned RD2 = RegInfo.createVirtualRegister(RC); 2639 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), RD2) 2640 .addReg(Mips::ZERO).addImm(1); 2641 2642 // Insert phi function to $Sink. 2643 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI), 2644 MI->getOperand(0).getReg()) 2645 .addReg(RD1).addMBB(FBB).addReg(RD2).addMBB(TBB); 2646 2647 MI->eraseFromParent(); // The pseudo instruction is gone now. 2648 return Sink; 2649} 2650 2651// Emit the COPY_FW pseudo instruction. 2652// 2653// copy_fw_pseudo $fd, $ws, n 2654// => 2655// copy_u_w $rt, $ws, $n 2656// mtc1 $rt, $fd 2657// 2658// When n is zero, the equivalent operation can be performed with (potentially) 2659// zero instructions due to register overlaps. This optimization is never valid 2660// for lane 1 because it would require FR=0 mode which isn't supported by MSA. 2661MachineBasicBlock * MipsSETargetLowering:: 2662emitCOPY_FW(MachineInstr *MI, MachineBasicBlock *BB) const{ 2663 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 2664 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo(); 2665 DebugLoc DL = MI->getDebugLoc(); 2666 unsigned Fd = MI->getOperand(0).getReg(); 2667 unsigned Ws = MI->getOperand(1).getReg(); 2668 unsigned Lane = MI->getOperand(2).getImm(); 2669 2670 if (Lane == 0) 2671 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Ws, 0, Mips::sub_lo); 2672 else { 2673 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass); 2674 2675 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wt).addReg(Ws).addImm(1); 2676 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_lo); 2677 } 2678 2679 MI->eraseFromParent(); // The pseudo instruction is gone now. 2680 return BB; 2681} 2682 2683// Emit the COPY_FD pseudo instruction. 2684// 2685// copy_fd_pseudo $fd, $ws, n 2686// => 2687// splati.d $wt, $ws, $n 2688// copy $fd, $wt:sub_64 2689// 2690// When n is zero, the equivalent operation can be performed with (potentially) 2691// zero instructions due to register overlaps. This optimization is always 2692// valid because FR=1 mode which is the only supported mode in MSA. 2693MachineBasicBlock * MipsSETargetLowering:: 2694emitCOPY_FD(MachineInstr *MI, MachineBasicBlock *BB) const{ 2695 assert(Subtarget->isFP64bit()); 2696 2697 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 2698 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo(); 2699 unsigned Fd = MI->getOperand(0).getReg(); 2700 unsigned Ws = MI->getOperand(1).getReg(); 2701 unsigned Lane = MI->getOperand(2).getImm() * 2; 2702 DebugLoc DL = MI->getDebugLoc(); 2703 2704 if (Lane == 0) 2705 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Ws, 0, Mips::sub_64); 2706 else { 2707 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass); 2708 2709 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wt).addReg(Ws).addImm(1); 2710 BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_64); 2711 } 2712 2713 MI->eraseFromParent(); // The pseudo instruction is gone now. 2714 return BB; 2715} 2716 2717// Emit the INSERT_FW pseudo instruction. 2718// 2719// insert_fw_pseudo $wd, $wd_in, $n, $fs 2720// => 2721// subreg_to_reg $wt:sub_lo, $fs 2722// insve_w $wd[$n], $wd_in, $wt[0] 2723MachineBasicBlock * 2724MipsSETargetLowering::emitINSERT_FW(MachineInstr *MI, 2725 MachineBasicBlock *BB) const { 2726 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 2727 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo(); 2728 DebugLoc DL = MI->getDebugLoc(); 2729 unsigned Wd = MI->getOperand(0).getReg(); 2730 unsigned Wd_in = MI->getOperand(1).getReg(); 2731 unsigned Lane = MI->getOperand(2).getImm(); 2732 unsigned Fs = MI->getOperand(3).getReg(); 2733 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass); 2734 2735 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt) 2736 .addImm(0) 2737 .addReg(Fs) 2738 .addImm(Mips::sub_lo); 2739 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_W), Wd) 2740 .addReg(Wd_in) 2741 .addImm(Lane) 2742 .addReg(Wt); 2743 2744 MI->eraseFromParent(); // The pseudo instruction is gone now. 2745 return BB; 2746} 2747 2748// Emit the INSERT_FD pseudo instruction. 2749// 2750// insert_fd_pseudo $wd, $fs, n 2751// => 2752// subreg_to_reg $wt:sub_64, $fs 2753// insve_d $wd[$n], $wd_in, $wt[0] 2754MachineBasicBlock * 2755MipsSETargetLowering::emitINSERT_FD(MachineInstr *MI, 2756 MachineBasicBlock *BB) const { 2757 assert(Subtarget->isFP64bit()); 2758 2759 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 2760 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo(); 2761 DebugLoc DL = MI->getDebugLoc(); 2762 unsigned Wd = MI->getOperand(0).getReg(); 2763 unsigned Wd_in = MI->getOperand(1).getReg(); 2764 unsigned Lane = MI->getOperand(2).getImm(); 2765 unsigned Fs = MI->getOperand(3).getReg(); 2766 unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass); 2767 2768 BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt) 2769 .addImm(0) 2770 .addReg(Fs) 2771 .addImm(Mips::sub_64); 2772 BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_D), Wd) 2773 .addReg(Wd_in) 2774 .addImm(Lane) 2775 .addReg(Wt); 2776 2777 MI->eraseFromParent(); // The pseudo instruction is gone now. 2778 return BB; 2779} 2780 2781// Emit the FILL_FW pseudo instruction. 2782// 2783// fill_fw_pseudo $wd, $fs 2784// => 2785// implicit_def $wt1 2786// insert_subreg $wt2:subreg_lo, $wt1, $fs 2787// splati.w $wd, $wt2[0] 2788MachineBasicBlock * 2789MipsSETargetLowering::emitFILL_FW(MachineInstr *MI, 2790 MachineBasicBlock *BB) const { 2791 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 2792 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo(); 2793 DebugLoc DL = MI->getDebugLoc(); 2794 unsigned Wd = MI->getOperand(0).getReg(); 2795 unsigned Fs = MI->getOperand(1).getReg(); 2796 unsigned Wt1 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass); 2797 unsigned Wt2 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass); 2798 2799 BuildMI(*BB, MI, DL, TII->get(Mips::IMPLICIT_DEF), Wt1); 2800 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_SUBREG), Wt2) 2801 .addReg(Wt1) 2802 .addReg(Fs) 2803 .addImm(Mips::sub_lo); 2804 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wd).addReg(Wt2).addImm(0); 2805 2806 MI->eraseFromParent(); // The pseudo instruction is gone now. 2807 return BB; 2808} 2809 2810// Emit the FILL_FD pseudo instruction. 2811// 2812// fill_fd_pseudo $wd, $fs 2813// => 2814// implicit_def $wt1 2815// insert_subreg $wt2:subreg_64, $wt1, $fs 2816// splati.d $wd, $wt2[0] 2817MachineBasicBlock * 2818MipsSETargetLowering::emitFILL_FD(MachineInstr *MI, 2819 MachineBasicBlock *BB) const { 2820 assert(Subtarget->isFP64bit()); 2821 2822 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 2823 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo(); 2824 DebugLoc DL = MI->getDebugLoc(); 2825 unsigned Wd = MI->getOperand(0).getReg(); 2826 unsigned Fs = MI->getOperand(1).getReg(); 2827 unsigned Wt1 = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass); 2828 unsigned Wt2 = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass); 2829 2830 BuildMI(*BB, MI, DL, TII->get(Mips::IMPLICIT_DEF), Wt1); 2831 BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_SUBREG), Wt2) 2832 .addReg(Wt1) 2833 .addReg(Fs) 2834 .addImm(Mips::sub_64); 2835 BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wd).addReg(Wt2).addImm(0); 2836 2837 MI->eraseFromParent(); // The pseudo instruction is gone now. 2838 return BB; 2839} 2840 2841// Emit the FEXP2_W_1 pseudo instructions. 2842// 2843// fexp2_w_1_pseudo $wd, $wt 2844// => 2845// ldi.w $ws, 1 2846// fexp2.w $wd, $ws, $wt 2847MachineBasicBlock * 2848MipsSETargetLowering::emitFEXP2_W_1(MachineInstr *MI, 2849 MachineBasicBlock *BB) const { 2850 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 2851 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo(); 2852 const TargetRegisterClass *RC = &Mips::MSA128WRegClass; 2853 unsigned Ws1 = RegInfo.createVirtualRegister(RC); 2854 unsigned Ws2 = RegInfo.createVirtualRegister(RC); 2855 DebugLoc DL = MI->getDebugLoc(); 2856 2857 // Splat 1.0 into a vector 2858 BuildMI(*BB, MI, DL, TII->get(Mips::LDI_W), Ws1).addImm(1); 2859 BuildMI(*BB, MI, DL, TII->get(Mips::FFINT_U_W), Ws2).addReg(Ws1); 2860 2861 // Emit 1.0 * fexp2(Wt) 2862 BuildMI(*BB, MI, DL, TII->get(Mips::FEXP2_W), MI->getOperand(0).getReg()) 2863 .addReg(Ws2) 2864 .addReg(MI->getOperand(1).getReg()); 2865 2866 MI->eraseFromParent(); // The pseudo instruction is gone now. 2867 return BB; 2868} 2869 2870// Emit the FEXP2_D_1 pseudo instructions. 2871// 2872// fexp2_d_1_pseudo $wd, $wt 2873// => 2874// ldi.d $ws, 1 2875// fexp2.d $wd, $ws, $wt 2876MachineBasicBlock * 2877MipsSETargetLowering::emitFEXP2_D_1(MachineInstr *MI, 2878 MachineBasicBlock *BB) const { 2879 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 2880 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo(); 2881 const TargetRegisterClass *RC = &Mips::MSA128DRegClass; 2882 unsigned Ws1 = RegInfo.createVirtualRegister(RC); 2883 unsigned Ws2 = RegInfo.createVirtualRegister(RC); 2884 DebugLoc DL = MI->getDebugLoc(); 2885 2886 // Splat 1.0 into a vector 2887 BuildMI(*BB, MI, DL, TII->get(Mips::LDI_D), Ws1).addImm(1); 2888 BuildMI(*BB, MI, DL, TII->get(Mips::FFINT_U_D), Ws2).addReg(Ws1); 2889 2890 // Emit 1.0 * fexp2(Wt) 2891 BuildMI(*BB, MI, DL, TII->get(Mips::FEXP2_D), MI->getOperand(0).getReg()) 2892 .addReg(Ws2) 2893 .addReg(MI->getOperand(1).getReg()); 2894 2895 MI->eraseFromParent(); // The pseudo instruction is gone now. 2896 return BB; 2897} 2898