LegalizeVectorTypes.cpp revision da2d8e1032eb4c2fefb1f647d7877910b9483835
1//===------- LegalizeVectorTypes.cpp - Legalization of vector types -------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file performs vector type splitting and scalarization for LegalizeTypes. 11// Scalarization is the act of changing a computation in an illegal one-element 12// vector type to be a computation in its scalar element type. For example, 13// implementing <1 x f32> arithmetic in a scalar f32 register. This is needed 14// as a base case when scalarizing vector arithmetic like <4 x f32>, which 15// eventually decomposes to scalars if the target doesn't support v4f32 or v2f32 16// types. 17// Splitting is the act of changing a computation in an invalid vector type to 18// be a computation in two vectors of half the size. For example, implementing 19// <128 x f32> operations in terms of two <64 x f32> operations. 20// 21//===----------------------------------------------------------------------===// 22 23#include "LegalizeTypes.h" 24#include "llvm/CodeGen/PseudoSourceValue.h" 25#include "llvm/Target/TargetData.h" 26#include "llvm/Support/ErrorHandling.h" 27#include "llvm/Support/raw_ostream.h" 28using namespace llvm; 29 30//===----------------------------------------------------------------------===// 31// Result Vector Scalarization: <1 x ty> -> ty. 32//===----------------------------------------------------------------------===// 33 34void DAGTypeLegalizer::ScalarizeVectorResult(SDNode *N, unsigned ResNo) { 35 DEBUG(dbgs() << "Scalarize node result " << ResNo << ": "; 36 N->dump(&DAG); 37 dbgs() << "\n"); 38 SDValue R = SDValue(); 39 40 switch (N->getOpcode()) { 41 default: 42#ifndef NDEBUG 43 dbgs() << "ScalarizeVectorResult #" << ResNo << ": "; 44 N->dump(&DAG); 45 dbgs() << "\n"; 46#endif 47 llvm_unreachable("Do not know how to scalarize the result of this operator!"); 48 49 case ISD::BIT_CONVERT: R = ScalarizeVecRes_BIT_CONVERT(N); break; 50 case ISD::BUILD_VECTOR: R = N->getOperand(0); break; 51 case ISD::CONVERT_RNDSAT: R = ScalarizeVecRes_CONVERT_RNDSAT(N); break; 52 case ISD::EXTRACT_SUBVECTOR: R = ScalarizeVecRes_EXTRACT_SUBVECTOR(N); break; 53 case ISD::FP_ROUND_INREG: R = ScalarizeVecRes_InregOp(N); break; 54 case ISD::FPOWI: R = ScalarizeVecRes_FPOWI(N); break; 55 case ISD::INSERT_VECTOR_ELT: R = ScalarizeVecRes_INSERT_VECTOR_ELT(N); break; 56 case ISD::LOAD: R = ScalarizeVecRes_LOAD(cast<LoadSDNode>(N));break; 57 case ISD::SCALAR_TO_VECTOR: R = ScalarizeVecRes_SCALAR_TO_VECTOR(N); break; 58 case ISD::SIGN_EXTEND_INREG: R = ScalarizeVecRes_InregOp(N); break; 59 case ISD::SELECT: R = ScalarizeVecRes_SELECT(N); break; 60 case ISD::SELECT_CC: R = ScalarizeVecRes_SELECT_CC(N); break; 61 case ISD::SETCC: R = ScalarizeVecRes_SETCC(N); break; 62 case ISD::UNDEF: R = ScalarizeVecRes_UNDEF(N); break; 63 case ISD::VECTOR_SHUFFLE: R = ScalarizeVecRes_VECTOR_SHUFFLE(N); break; 64 case ISD::VSETCC: R = ScalarizeVecRes_VSETCC(N); break; 65 66 case ISD::CTLZ: 67 case ISD::CTPOP: 68 case ISD::CTTZ: 69 case ISD::FABS: 70 case ISD::FCOS: 71 case ISD::FNEG: 72 case ISD::FP_TO_SINT: 73 case ISD::FP_TO_UINT: 74 case ISD::FSIN: 75 case ISD::FSQRT: 76 case ISD::FTRUNC: 77 case ISD::FFLOOR: 78 case ISD::FCEIL: 79 case ISD::FRINT: 80 case ISD::FNEARBYINT: 81 case ISD::UINT_TO_FP: 82 case ISD::SINT_TO_FP: 83 case ISD::TRUNCATE: 84 case ISD::SIGN_EXTEND: 85 case ISD::ZERO_EXTEND: 86 case ISD::ANY_EXTEND: 87 R = ScalarizeVecRes_UnaryOp(N); 88 break; 89 90 case ISD::ADD: 91 case ISD::AND: 92 case ISD::FADD: 93 case ISD::FDIV: 94 case ISD::FMUL: 95 case ISD::FPOW: 96 case ISD::FREM: 97 case ISD::FSUB: 98 case ISD::MUL: 99 case ISD::OR: 100 case ISD::SDIV: 101 case ISD::SREM: 102 case ISD::SUB: 103 case ISD::UDIV: 104 case ISD::UREM: 105 case ISD::XOR: 106 case ISD::SHL: 107 case ISD::SRA: 108 case ISD::SRL: 109 R = ScalarizeVecRes_BinOp(N); 110 break; 111 } 112 113 // If R is null, the sub-method took care of registering the result. 114 if (R.getNode()) 115 SetScalarizedVector(SDValue(N, ResNo), R); 116} 117 118SDValue DAGTypeLegalizer::ScalarizeVecRes_BinOp(SDNode *N) { 119 SDValue LHS = GetScalarizedVector(N->getOperand(0)); 120 SDValue RHS = GetScalarizedVector(N->getOperand(1)); 121 return DAG.getNode(N->getOpcode(), N->getDebugLoc(), 122 LHS.getValueType(), LHS, RHS); 123} 124 125SDValue DAGTypeLegalizer::ScalarizeVecRes_BIT_CONVERT(SDNode *N) { 126 EVT NewVT = N->getValueType(0).getVectorElementType(); 127 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), 128 NewVT, N->getOperand(0)); 129} 130 131SDValue DAGTypeLegalizer::ScalarizeVecRes_CONVERT_RNDSAT(SDNode *N) { 132 EVT NewVT = N->getValueType(0).getVectorElementType(); 133 SDValue Op0 = GetScalarizedVector(N->getOperand(0)); 134 return DAG.getConvertRndSat(NewVT, N->getDebugLoc(), 135 Op0, DAG.getValueType(NewVT), 136 DAG.getValueType(Op0.getValueType()), 137 N->getOperand(3), 138 N->getOperand(4), 139 cast<CvtRndSatSDNode>(N)->getCvtCode()); 140} 141 142SDValue DAGTypeLegalizer::ScalarizeVecRes_EXTRACT_SUBVECTOR(SDNode *N) { 143 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, N->getDebugLoc(), 144 N->getValueType(0).getVectorElementType(), 145 N->getOperand(0), N->getOperand(1)); 146} 147 148SDValue DAGTypeLegalizer::ScalarizeVecRes_FPOWI(SDNode *N) { 149 SDValue Op = GetScalarizedVector(N->getOperand(0)); 150 return DAG.getNode(ISD::FPOWI, N->getDebugLoc(), 151 Op.getValueType(), Op, N->getOperand(1)); 152} 153 154SDValue DAGTypeLegalizer::ScalarizeVecRes_INSERT_VECTOR_ELT(SDNode *N) { 155 // The value to insert may have a wider type than the vector element type, 156 // so be sure to truncate it to the element type if necessary. 157 SDValue Op = N->getOperand(1); 158 EVT EltVT = N->getValueType(0).getVectorElementType(); 159 if (Op.getValueType() != EltVT) 160 // FIXME: Can this happen for floating point types? 161 Op = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), EltVT, Op); 162 return Op; 163} 164 165SDValue DAGTypeLegalizer::ScalarizeVecRes_LOAD(LoadSDNode *N) { 166 assert(N->isUnindexed() && "Indexed vector load?"); 167 168 SDValue Result = DAG.getLoad(ISD::UNINDEXED, 169 N->getExtensionType(), 170 N->getValueType(0).getVectorElementType(), 171 N->getDebugLoc(), 172 N->getChain(), N->getBasePtr(), 173 DAG.getUNDEF(N->getBasePtr().getValueType()), 174 N->getPointerInfo(), 175 N->getMemoryVT().getVectorElementType(), 176 N->isVolatile(), N->isNonTemporal(), 177 N->getOriginalAlignment()); 178 179 // Legalized the chain result - switch anything that used the old chain to 180 // use the new one. 181 ReplaceValueWith(SDValue(N, 1), Result.getValue(1)); 182 return Result; 183} 184 185SDValue DAGTypeLegalizer::ScalarizeVecRes_UnaryOp(SDNode *N) { 186 // Get the dest type - it doesn't always match the input type, e.g. int_to_fp. 187 EVT DestVT = N->getValueType(0).getVectorElementType(); 188 SDValue Op = GetScalarizedVector(N->getOperand(0)); 189 return DAG.getNode(N->getOpcode(), N->getDebugLoc(), DestVT, Op); 190} 191 192SDValue DAGTypeLegalizer::ScalarizeVecRes_InregOp(SDNode *N) { 193 EVT EltVT = N->getValueType(0).getVectorElementType(); 194 EVT ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT().getVectorElementType(); 195 SDValue LHS = GetScalarizedVector(N->getOperand(0)); 196 return DAG.getNode(N->getOpcode(), N->getDebugLoc(), EltVT, 197 LHS, DAG.getValueType(ExtVT)); 198} 199 200SDValue DAGTypeLegalizer::ScalarizeVecRes_SCALAR_TO_VECTOR(SDNode *N) { 201 // If the operand is wider than the vector element type then it is implicitly 202 // truncated. Make that explicit here. 203 EVT EltVT = N->getValueType(0).getVectorElementType(); 204 SDValue InOp = N->getOperand(0); 205 if (InOp.getValueType() != EltVT) 206 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), EltVT, InOp); 207 return InOp; 208} 209 210SDValue DAGTypeLegalizer::ScalarizeVecRes_SELECT(SDNode *N) { 211 SDValue LHS = GetScalarizedVector(N->getOperand(1)); 212 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), 213 LHS.getValueType(), N->getOperand(0), LHS, 214 GetScalarizedVector(N->getOperand(2))); 215} 216 217SDValue DAGTypeLegalizer::ScalarizeVecRes_SELECT_CC(SDNode *N) { 218 SDValue LHS = GetScalarizedVector(N->getOperand(2)); 219 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), LHS.getValueType(), 220 N->getOperand(0), N->getOperand(1), 221 LHS, GetScalarizedVector(N->getOperand(3)), 222 N->getOperand(4)); 223} 224 225SDValue DAGTypeLegalizer::ScalarizeVecRes_SETCC(SDNode *N) { 226 SDValue LHS = GetScalarizedVector(N->getOperand(0)); 227 SDValue RHS = GetScalarizedVector(N->getOperand(1)); 228 DebugLoc DL = N->getDebugLoc(); 229 230 // Turn it into a scalar SETCC. 231 return DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS, N->getOperand(2)); 232} 233 234SDValue DAGTypeLegalizer::ScalarizeVecRes_UNDEF(SDNode *N) { 235 return DAG.getUNDEF(N->getValueType(0).getVectorElementType()); 236} 237 238SDValue DAGTypeLegalizer::ScalarizeVecRes_VECTOR_SHUFFLE(SDNode *N) { 239 // Figure out if the scalar is the LHS or RHS and return it. 240 SDValue Arg = N->getOperand(2).getOperand(0); 241 if (Arg.getOpcode() == ISD::UNDEF) 242 return DAG.getUNDEF(N->getValueType(0).getVectorElementType()); 243 unsigned Op = !cast<ConstantSDNode>(Arg)->isNullValue(); 244 return GetScalarizedVector(N->getOperand(Op)); 245} 246 247SDValue DAGTypeLegalizer::ScalarizeVecRes_VSETCC(SDNode *N) { 248 SDValue LHS = GetScalarizedVector(N->getOperand(0)); 249 SDValue RHS = GetScalarizedVector(N->getOperand(1)); 250 EVT NVT = N->getValueType(0).getVectorElementType(); 251 EVT SVT = TLI.getSetCCResultType(LHS.getValueType()); 252 DebugLoc DL = N->getDebugLoc(); 253 254 // Turn it into a scalar SETCC. 255 SDValue Res = DAG.getNode(ISD::SETCC, DL, SVT, LHS, RHS, N->getOperand(2)); 256 257 // VSETCC always returns a sign-extended value, while SETCC may not. The 258 // SETCC result type may not match the vector element type. Correct these. 259 if (NVT.bitsLE(SVT)) { 260 // The SETCC result type is bigger than the vector element type. 261 // Ensure the SETCC result is sign-extended. 262 if (TLI.getBooleanContents() != 263 TargetLowering::ZeroOrNegativeOneBooleanContent) 264 Res = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, SVT, Res, 265 DAG.getValueType(MVT::i1)); 266 // Truncate to the final type. 267 return DAG.getNode(ISD::TRUNCATE, DL, NVT, Res); 268 } 269 270 // The SETCC result type is smaller than the vector element type. 271 // If the SetCC result is not sign-extended, chop it down to MVT::i1. 272 if (TLI.getBooleanContents() != 273 TargetLowering::ZeroOrNegativeOneBooleanContent) 274 Res = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Res); 275 // Sign extend to the final type. 276 return DAG.getNode(ISD::SIGN_EXTEND, DL, NVT, Res); 277} 278 279 280//===----------------------------------------------------------------------===// 281// Operand Vector Scalarization <1 x ty> -> ty. 282//===----------------------------------------------------------------------===// 283 284bool DAGTypeLegalizer::ScalarizeVectorOperand(SDNode *N, unsigned OpNo) { 285 DEBUG(dbgs() << "Scalarize node operand " << OpNo << ": "; 286 N->dump(&DAG); 287 dbgs() << "\n"); 288 SDValue Res = SDValue(); 289 290 if (Res.getNode() == 0) { 291 switch (N->getOpcode()) { 292 default: 293#ifndef NDEBUG 294 dbgs() << "ScalarizeVectorOperand Op #" << OpNo << ": "; 295 N->dump(&DAG); 296 dbgs() << "\n"; 297#endif 298 llvm_unreachable("Do not know how to scalarize this operator's operand!"); 299 case ISD::BIT_CONVERT: 300 Res = ScalarizeVecOp_BIT_CONVERT(N); 301 break; 302 case ISD::CONCAT_VECTORS: 303 Res = ScalarizeVecOp_CONCAT_VECTORS(N); 304 break; 305 case ISD::EXTRACT_VECTOR_ELT: 306 Res = ScalarizeVecOp_EXTRACT_VECTOR_ELT(N); 307 break; 308 case ISD::STORE: 309 Res = ScalarizeVecOp_STORE(cast<StoreSDNode>(N), OpNo); 310 break; 311 } 312 } 313 314 // If the result is null, the sub-method took care of registering results etc. 315 if (!Res.getNode()) return false; 316 317 // If the result is N, the sub-method updated N in place. Tell the legalizer 318 // core about this. 319 if (Res.getNode() == N) 320 return true; 321 322 assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 && 323 "Invalid operand expansion"); 324 325 ReplaceValueWith(SDValue(N, 0), Res); 326 return false; 327} 328 329/// ScalarizeVecOp_BIT_CONVERT - If the value to convert is a vector that needs 330/// to be scalarized, it must be <1 x ty>. Convert the element instead. 331SDValue DAGTypeLegalizer::ScalarizeVecOp_BIT_CONVERT(SDNode *N) { 332 SDValue Elt = GetScalarizedVector(N->getOperand(0)); 333 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), 334 N->getValueType(0), Elt); 335} 336 337/// ScalarizeVecOp_CONCAT_VECTORS - The vectors to concatenate have length one - 338/// use a BUILD_VECTOR instead. 339SDValue DAGTypeLegalizer::ScalarizeVecOp_CONCAT_VECTORS(SDNode *N) { 340 SmallVector<SDValue, 8> Ops(N->getNumOperands()); 341 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i) 342 Ops[i] = GetScalarizedVector(N->getOperand(i)); 343 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), N->getValueType(0), 344 &Ops[0], Ops.size()); 345} 346 347/// ScalarizeVecOp_EXTRACT_VECTOR_ELT - If the input is a vector that needs to 348/// be scalarized, it must be <1 x ty>, so just return the element, ignoring the 349/// index. 350SDValue DAGTypeLegalizer::ScalarizeVecOp_EXTRACT_VECTOR_ELT(SDNode *N) { 351 SDValue Res = GetScalarizedVector(N->getOperand(0)); 352 if (Res.getValueType() != N->getValueType(0)) 353 Res = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), N->getValueType(0), 354 Res); 355 return Res; 356} 357 358/// ScalarizeVecOp_STORE - If the value to store is a vector that needs to be 359/// scalarized, it must be <1 x ty>. Just store the element. 360SDValue DAGTypeLegalizer::ScalarizeVecOp_STORE(StoreSDNode *N, unsigned OpNo){ 361 assert(N->isUnindexed() && "Indexed store of one-element vector?"); 362 assert(OpNo == 1 && "Do not know how to scalarize this operand!"); 363 DebugLoc dl = N->getDebugLoc(); 364 365 if (N->isTruncatingStore()) 366 return DAG.getTruncStore(N->getChain(), dl, 367 GetScalarizedVector(N->getOperand(1)), 368 N->getBasePtr(), N->getPointerInfo(), 369 N->getMemoryVT().getVectorElementType(), 370 N->isVolatile(), N->isNonTemporal(), 371 N->getAlignment()); 372 373 return DAG.getStore(N->getChain(), dl, GetScalarizedVector(N->getOperand(1)), 374 N->getBasePtr(), N->getPointerInfo(), 375 N->isVolatile(), N->isNonTemporal(), 376 N->getOriginalAlignment()); 377} 378 379 380//===----------------------------------------------------------------------===// 381// Result Vector Splitting 382//===----------------------------------------------------------------------===// 383 384/// SplitVectorResult - This method is called when the specified result of the 385/// specified node is found to need vector splitting. At this point, the node 386/// may also have invalid operands or may have other results that need 387/// legalization, we just know that (at least) one result needs vector 388/// splitting. 389void DAGTypeLegalizer::SplitVectorResult(SDNode *N, unsigned ResNo) { 390 DEBUG(dbgs() << "Split node result: "; 391 N->dump(&DAG); 392 dbgs() << "\n"); 393 SDValue Lo, Hi; 394 395 switch (N->getOpcode()) { 396 default: 397#ifndef NDEBUG 398 dbgs() << "SplitVectorResult #" << ResNo << ": "; 399 N->dump(&DAG); 400 dbgs() << "\n"; 401#endif 402 llvm_unreachable("Do not know how to split the result of this operator!"); 403 404 case ISD::MERGE_VALUES: SplitRes_MERGE_VALUES(N, Lo, Hi); break; 405 case ISD::SELECT: SplitRes_SELECT(N, Lo, Hi); break; 406 case ISD::SELECT_CC: SplitRes_SELECT_CC(N, Lo, Hi); break; 407 case ISD::UNDEF: SplitRes_UNDEF(N, Lo, Hi); break; 408 409 case ISD::BIT_CONVERT: SplitVecRes_BIT_CONVERT(N, Lo, Hi); break; 410 case ISD::BUILD_VECTOR: SplitVecRes_BUILD_VECTOR(N, Lo, Hi); break; 411 case ISD::CONCAT_VECTORS: SplitVecRes_CONCAT_VECTORS(N, Lo, Hi); break; 412 case ISD::CONVERT_RNDSAT: SplitVecRes_CONVERT_RNDSAT(N, Lo, Hi); break; 413 case ISD::EXTRACT_SUBVECTOR: SplitVecRes_EXTRACT_SUBVECTOR(N, Lo, Hi); break; 414 case ISD::FP_ROUND_INREG: SplitVecRes_InregOp(N, Lo, Hi); break; 415 case ISD::FPOWI: SplitVecRes_FPOWI(N, Lo, Hi); break; 416 case ISD::INSERT_VECTOR_ELT: SplitVecRes_INSERT_VECTOR_ELT(N, Lo, Hi); break; 417 case ISD::SCALAR_TO_VECTOR: SplitVecRes_SCALAR_TO_VECTOR(N, Lo, Hi); break; 418 case ISD::SIGN_EXTEND_INREG: SplitVecRes_InregOp(N, Lo, Hi); break; 419 case ISD::LOAD: 420 SplitVecRes_LOAD(cast<LoadSDNode>(N), Lo, Hi); 421 break; 422 case ISD::SETCC: 423 case ISD::VSETCC: 424 SplitVecRes_SETCC(N, Lo, Hi); 425 break; 426 case ISD::VECTOR_SHUFFLE: 427 SplitVecRes_VECTOR_SHUFFLE(cast<ShuffleVectorSDNode>(N), Lo, Hi); 428 break; 429 430 case ISD::CTTZ: 431 case ISD::CTLZ: 432 case ISD::CTPOP: 433 case ISD::FNEG: 434 case ISD::FABS: 435 case ISD::FSQRT: 436 case ISD::FSIN: 437 case ISD::FCOS: 438 case ISD::FTRUNC: 439 case ISD::FFLOOR: 440 case ISD::FCEIL: 441 case ISD::FRINT: 442 case ISD::FNEARBYINT: 443 case ISD::FP_TO_SINT: 444 case ISD::FP_TO_UINT: 445 case ISD::SINT_TO_FP: 446 case ISD::UINT_TO_FP: 447 case ISD::TRUNCATE: 448 case ISD::SIGN_EXTEND: 449 case ISD::ZERO_EXTEND: 450 case ISD::ANY_EXTEND: 451 case ISD::FEXP: 452 case ISD::FEXP2: 453 case ISD::FLOG: 454 case ISD::FLOG2: 455 case ISD::FLOG10: 456 SplitVecRes_UnaryOp(N, Lo, Hi); 457 break; 458 459 case ISD::ADD: 460 case ISD::SUB: 461 case ISD::MUL: 462 case ISD::FADD: 463 case ISD::FSUB: 464 case ISD::FMUL: 465 case ISD::SDIV: 466 case ISD::UDIV: 467 case ISD::FDIV: 468 case ISD::FPOW: 469 case ISD::AND: 470 case ISD::OR: 471 case ISD::XOR: 472 case ISD::SHL: 473 case ISD::SRA: 474 case ISD::SRL: 475 case ISD::UREM: 476 case ISD::SREM: 477 case ISD::FREM: 478 SplitVecRes_BinOp(N, Lo, Hi); 479 break; 480 } 481 482 // If Lo/Hi is null, the sub-method took care of registering results etc. 483 if (Lo.getNode()) 484 SetSplitVector(SDValue(N, ResNo), Lo, Hi); 485} 486 487void DAGTypeLegalizer::SplitVecRes_BinOp(SDNode *N, SDValue &Lo, 488 SDValue &Hi) { 489 SDValue LHSLo, LHSHi; 490 GetSplitVector(N->getOperand(0), LHSLo, LHSHi); 491 SDValue RHSLo, RHSHi; 492 GetSplitVector(N->getOperand(1), RHSLo, RHSHi); 493 DebugLoc dl = N->getDebugLoc(); 494 495 Lo = DAG.getNode(N->getOpcode(), dl, LHSLo.getValueType(), LHSLo, RHSLo); 496 Hi = DAG.getNode(N->getOpcode(), dl, LHSHi.getValueType(), LHSHi, RHSHi); 497} 498 499void DAGTypeLegalizer::SplitVecRes_BIT_CONVERT(SDNode *N, SDValue &Lo, 500 SDValue &Hi) { 501 // We know the result is a vector. The input may be either a vector or a 502 // scalar value. 503 EVT LoVT, HiVT; 504 GetSplitDestVTs(N->getValueType(0), LoVT, HiVT); 505 DebugLoc dl = N->getDebugLoc(); 506 507 SDValue InOp = N->getOperand(0); 508 EVT InVT = InOp.getValueType(); 509 510 // Handle some special cases efficiently. 511 switch (getTypeAction(InVT)) { 512 default: 513 assert(false && "Unknown type action!"); 514 case Legal: 515 case PromoteInteger: 516 case SoftenFloat: 517 case ScalarizeVector: 518 break; 519 case ExpandInteger: 520 case ExpandFloat: 521 // A scalar to vector conversion, where the scalar needs expansion. 522 // If the vector is being split in two then we can just convert the 523 // expanded pieces. 524 if (LoVT == HiVT) { 525 GetExpandedOp(InOp, Lo, Hi); 526 if (TLI.isBigEndian()) 527 std::swap(Lo, Hi); 528 Lo = DAG.getNode(ISD::BIT_CONVERT, dl, LoVT, Lo); 529 Hi = DAG.getNode(ISD::BIT_CONVERT, dl, HiVT, Hi); 530 return; 531 } 532 break; 533 case SplitVector: 534 // If the input is a vector that needs to be split, convert each split 535 // piece of the input now. 536 GetSplitVector(InOp, Lo, Hi); 537 Lo = DAG.getNode(ISD::BIT_CONVERT, dl, LoVT, Lo); 538 Hi = DAG.getNode(ISD::BIT_CONVERT, dl, HiVT, Hi); 539 return; 540 } 541 542 // In the general case, convert the input to an integer and split it by hand. 543 EVT LoIntVT = EVT::getIntegerVT(*DAG.getContext(), LoVT.getSizeInBits()); 544 EVT HiIntVT = EVT::getIntegerVT(*DAG.getContext(), HiVT.getSizeInBits()); 545 if (TLI.isBigEndian()) 546 std::swap(LoIntVT, HiIntVT); 547 548 SplitInteger(BitConvertToInteger(InOp), LoIntVT, HiIntVT, Lo, Hi); 549 550 if (TLI.isBigEndian()) 551 std::swap(Lo, Hi); 552 Lo = DAG.getNode(ISD::BIT_CONVERT, dl, LoVT, Lo); 553 Hi = DAG.getNode(ISD::BIT_CONVERT, dl, HiVT, Hi); 554} 555 556void DAGTypeLegalizer::SplitVecRes_BUILD_VECTOR(SDNode *N, SDValue &Lo, 557 SDValue &Hi) { 558 EVT LoVT, HiVT; 559 DebugLoc dl = N->getDebugLoc(); 560 GetSplitDestVTs(N->getValueType(0), LoVT, HiVT); 561 unsigned LoNumElts = LoVT.getVectorNumElements(); 562 SmallVector<SDValue, 8> LoOps(N->op_begin(), N->op_begin()+LoNumElts); 563 Lo = DAG.getNode(ISD::BUILD_VECTOR, dl, LoVT, &LoOps[0], LoOps.size()); 564 565 SmallVector<SDValue, 8> HiOps(N->op_begin()+LoNumElts, N->op_end()); 566 Hi = DAG.getNode(ISD::BUILD_VECTOR, dl, HiVT, &HiOps[0], HiOps.size()); 567} 568 569void DAGTypeLegalizer::SplitVecRes_CONCAT_VECTORS(SDNode *N, SDValue &Lo, 570 SDValue &Hi) { 571 assert(!(N->getNumOperands() & 1) && "Unsupported CONCAT_VECTORS"); 572 DebugLoc dl = N->getDebugLoc(); 573 unsigned NumSubvectors = N->getNumOperands() / 2; 574 if (NumSubvectors == 1) { 575 Lo = N->getOperand(0); 576 Hi = N->getOperand(1); 577 return; 578 } 579 580 EVT LoVT, HiVT; 581 GetSplitDestVTs(N->getValueType(0), LoVT, HiVT); 582 583 SmallVector<SDValue, 8> LoOps(N->op_begin(), N->op_begin()+NumSubvectors); 584 Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, LoVT, &LoOps[0], LoOps.size()); 585 586 SmallVector<SDValue, 8> HiOps(N->op_begin()+NumSubvectors, N->op_end()); 587 Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HiVT, &HiOps[0], HiOps.size()); 588} 589 590void DAGTypeLegalizer::SplitVecRes_CONVERT_RNDSAT(SDNode *N, SDValue &Lo, 591 SDValue &Hi) { 592 EVT LoVT, HiVT; 593 DebugLoc dl = N->getDebugLoc(); 594 GetSplitDestVTs(N->getValueType(0), LoVT, HiVT); 595 596 SDValue DTyOpLo = DAG.getValueType(LoVT); 597 SDValue DTyOpHi = DAG.getValueType(HiVT); 598 599 SDValue RndOp = N->getOperand(3); 600 SDValue SatOp = N->getOperand(4); 601 ISD::CvtCode CvtCode = cast<CvtRndSatSDNode>(N)->getCvtCode(); 602 603 // Split the input. 604 SDValue VLo, VHi; 605 EVT InVT = N->getOperand(0).getValueType(); 606 switch (getTypeAction(InVT)) { 607 default: llvm_unreachable("Unexpected type action!"); 608 case Legal: { 609 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), InVT.getVectorElementType(), 610 LoVT.getVectorNumElements()); 611 VLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, N->getOperand(0), 612 DAG.getIntPtrConstant(0)); 613 VHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, N->getOperand(0), 614 DAG.getIntPtrConstant(InNVT.getVectorNumElements())); 615 break; 616 } 617 case SplitVector: 618 GetSplitVector(N->getOperand(0), VLo, VHi); 619 break; 620 case WidenVector: { 621 // If the result needs to be split and the input needs to be widened, 622 // the two types must have different lengths. Use the widened result 623 // and extract from it to do the split. 624 SDValue InOp = GetWidenedVector(N->getOperand(0)); 625 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), InVT.getVectorElementType(), 626 LoVT.getVectorNumElements()); 627 VLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, InOp, 628 DAG.getIntPtrConstant(0)); 629 VHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, InOp, 630 DAG.getIntPtrConstant(InNVT.getVectorNumElements())); 631 break; 632 } 633 } 634 635 SDValue STyOpLo = DAG.getValueType(VLo.getValueType()); 636 SDValue STyOpHi = DAG.getValueType(VHi.getValueType()); 637 638 Lo = DAG.getConvertRndSat(LoVT, dl, VLo, DTyOpLo, STyOpLo, RndOp, SatOp, 639 CvtCode); 640 Hi = DAG.getConvertRndSat(HiVT, dl, VHi, DTyOpHi, STyOpHi, RndOp, SatOp, 641 CvtCode); 642} 643 644void DAGTypeLegalizer::SplitVecRes_EXTRACT_SUBVECTOR(SDNode *N, SDValue &Lo, 645 SDValue &Hi) { 646 SDValue Vec = N->getOperand(0); 647 SDValue Idx = N->getOperand(1); 648 EVT IdxVT = Idx.getValueType(); 649 DebugLoc dl = N->getDebugLoc(); 650 651 EVT LoVT, HiVT; 652 GetSplitDestVTs(N->getValueType(0), LoVT, HiVT); 653 654 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, LoVT, Vec, Idx); 655 Idx = DAG.getNode(ISD::ADD, dl, IdxVT, Idx, 656 DAG.getConstant(LoVT.getVectorNumElements(), IdxVT)); 657 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, HiVT, Vec, Idx); 658} 659 660void DAGTypeLegalizer::SplitVecRes_FPOWI(SDNode *N, SDValue &Lo, 661 SDValue &Hi) { 662 DebugLoc dl = N->getDebugLoc(); 663 GetSplitVector(N->getOperand(0), Lo, Hi); 664 Lo = DAG.getNode(ISD::FPOWI, dl, Lo.getValueType(), Lo, N->getOperand(1)); 665 Hi = DAG.getNode(ISD::FPOWI, dl, Hi.getValueType(), Hi, N->getOperand(1)); 666} 667 668void DAGTypeLegalizer::SplitVecRes_InregOp(SDNode *N, SDValue &Lo, 669 SDValue &Hi) { 670 SDValue LHSLo, LHSHi; 671 GetSplitVector(N->getOperand(0), LHSLo, LHSHi); 672 DebugLoc dl = N->getDebugLoc(); 673 674 EVT LoVT, HiVT; 675 GetSplitDestVTs(cast<VTSDNode>(N->getOperand(1))->getVT(), LoVT, HiVT); 676 677 Lo = DAG.getNode(N->getOpcode(), dl, LHSLo.getValueType(), LHSLo, 678 DAG.getValueType(LoVT)); 679 Hi = DAG.getNode(N->getOpcode(), dl, LHSHi.getValueType(), LHSHi, 680 DAG.getValueType(HiVT)); 681} 682 683void DAGTypeLegalizer::SplitVecRes_INSERT_VECTOR_ELT(SDNode *N, SDValue &Lo, 684 SDValue &Hi) { 685 SDValue Vec = N->getOperand(0); 686 SDValue Elt = N->getOperand(1); 687 SDValue Idx = N->getOperand(2); 688 DebugLoc dl = N->getDebugLoc(); 689 GetSplitVector(Vec, Lo, Hi); 690 691 if (ConstantSDNode *CIdx = dyn_cast<ConstantSDNode>(Idx)) { 692 unsigned IdxVal = CIdx->getZExtValue(); 693 unsigned LoNumElts = Lo.getValueType().getVectorNumElements(); 694 if (IdxVal < LoNumElts) 695 Lo = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, 696 Lo.getValueType(), Lo, Elt, Idx); 697 else 698 Hi = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, Hi.getValueType(), Hi, Elt, 699 DAG.getIntPtrConstant(IdxVal - LoNumElts)); 700 return; 701 } 702 703 // Spill the vector to the stack. 704 EVT VecVT = Vec.getValueType(); 705 EVT EltVT = VecVT.getVectorElementType(); 706 SDValue StackPtr = DAG.CreateStackTemporary(VecVT); 707 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, 708 MachinePointerInfo(), false, false, 0); 709 710 // Store the new element. This may be larger than the vector element type, 711 // so use a truncating store. 712 SDValue EltPtr = GetVectorElementPointer(StackPtr, EltVT, Idx); 713 const Type *VecType = VecVT.getTypeForEVT(*DAG.getContext()); 714 unsigned Alignment = 715 TLI.getTargetData()->getPrefTypeAlignment(VecType); 716 Store = DAG.getTruncStore(Store, dl, Elt, EltPtr, MachinePointerInfo(), EltVT, 717 false, false, 0); 718 719 // Load the Lo part from the stack slot. 720 Lo = DAG.getLoad(Lo.getValueType(), dl, Store, StackPtr, MachinePointerInfo(), 721 false, false, 0); 722 723 // Increment the pointer to the other part. 724 unsigned IncrementSize = Lo.getValueType().getSizeInBits() / 8; 725 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr, 726 DAG.getIntPtrConstant(IncrementSize)); 727 728 // Load the Hi part from the stack slot. 729 Hi = DAG.getLoad(Hi.getValueType(), dl, Store, StackPtr, MachinePointerInfo(), 730 false, false, MinAlign(Alignment, IncrementSize)); 731} 732 733void DAGTypeLegalizer::SplitVecRes_SCALAR_TO_VECTOR(SDNode *N, SDValue &Lo, 734 SDValue &Hi) { 735 EVT LoVT, HiVT; 736 DebugLoc dl = N->getDebugLoc(); 737 GetSplitDestVTs(N->getValueType(0), LoVT, HiVT); 738 Lo = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoVT, N->getOperand(0)); 739 Hi = DAG.getUNDEF(HiVT); 740} 741 742void DAGTypeLegalizer::SplitVecRes_LOAD(LoadSDNode *LD, SDValue &Lo, 743 SDValue &Hi) { 744 assert(ISD::isUNINDEXEDLoad(LD) && "Indexed load during type legalization!"); 745 EVT LoVT, HiVT; 746 DebugLoc dl = LD->getDebugLoc(); 747 GetSplitDestVTs(LD->getValueType(0), LoVT, HiVT); 748 749 ISD::LoadExtType ExtType = LD->getExtensionType(); 750 SDValue Ch = LD->getChain(); 751 SDValue Ptr = LD->getBasePtr(); 752 SDValue Offset = DAG.getUNDEF(Ptr.getValueType()); 753 EVT MemoryVT = LD->getMemoryVT(); 754 unsigned Alignment = LD->getOriginalAlignment(); 755 bool isVolatile = LD->isVolatile(); 756 bool isNonTemporal = LD->isNonTemporal(); 757 758 EVT LoMemVT, HiMemVT; 759 GetSplitDestVTs(MemoryVT, LoMemVT, HiMemVT); 760 761 Lo = DAG.getLoad(ISD::UNINDEXED, ExtType, LoVT, dl, Ch, Ptr, Offset, 762 LD->getPointerInfo(), LoMemVT, isVolatile, isNonTemporal, 763 Alignment); 764 765 unsigned IncrementSize = LoMemVT.getSizeInBits()/8; 766 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, 767 DAG.getIntPtrConstant(IncrementSize)); 768 Hi = DAG.getLoad(ISD::UNINDEXED, ExtType, HiVT, dl, Ch, Ptr, Offset, 769 LD->getPointerInfo().getWithOffset(IncrementSize), 770 HiMemVT, isVolatile, isNonTemporal, Alignment); 771 772 // Build a factor node to remember that this load is independent of the 773 // other one. 774 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1), 775 Hi.getValue(1)); 776 777 // Legalized the chain result - switch anything that used the old chain to 778 // use the new one. 779 ReplaceValueWith(SDValue(LD, 1), Ch); 780} 781 782void DAGTypeLegalizer::SplitVecRes_SETCC(SDNode *N, SDValue &Lo, SDValue &Hi) { 783 EVT LoVT, HiVT; 784 DebugLoc DL = N->getDebugLoc(); 785 GetSplitDestVTs(N->getValueType(0), LoVT, HiVT); 786 787 // Split the input. 788 EVT InVT = N->getOperand(0).getValueType(); 789 SDValue LL, LH, RL, RH; 790 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), InVT.getVectorElementType(), 791 LoVT.getVectorNumElements()); 792 LL = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, N->getOperand(0), 793 DAG.getIntPtrConstant(0)); 794 LH = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, N->getOperand(0), 795 DAG.getIntPtrConstant(InNVT.getVectorNumElements())); 796 797 RL = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, N->getOperand(1), 798 DAG.getIntPtrConstant(0)); 799 RH = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, N->getOperand(1), 800 DAG.getIntPtrConstant(InNVT.getVectorNumElements())); 801 802 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2)); 803 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2)); 804} 805 806void DAGTypeLegalizer::SplitVecRes_UnaryOp(SDNode *N, SDValue &Lo, 807 SDValue &Hi) { 808 // Get the dest types - they may not match the input types, e.g. int_to_fp. 809 EVT LoVT, HiVT; 810 DebugLoc dl = N->getDebugLoc(); 811 GetSplitDestVTs(N->getValueType(0), LoVT, HiVT); 812 813 // Split the input. 814 EVT InVT = N->getOperand(0).getValueType(); 815 switch (getTypeAction(InVT)) { 816 default: llvm_unreachable("Unexpected type action!"); 817 case Legal: { 818 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), InVT.getVectorElementType(), 819 LoVT.getVectorNumElements()); 820 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, N->getOperand(0), 821 DAG.getIntPtrConstant(0)); 822 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, N->getOperand(0), 823 DAG.getIntPtrConstant(InNVT.getVectorNumElements())); 824 break; 825 } 826 case SplitVector: 827 GetSplitVector(N->getOperand(0), Lo, Hi); 828 break; 829 case WidenVector: { 830 // If the result needs to be split and the input needs to be widened, 831 // the two types must have different lengths. Use the widened result 832 // and extract from it to do the split. 833 SDValue InOp = GetWidenedVector(N->getOperand(0)); 834 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), InVT.getVectorElementType(), 835 LoVT.getVectorNumElements()); 836 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, InOp, 837 DAG.getIntPtrConstant(0)); 838 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, InOp, 839 DAG.getIntPtrConstant(InNVT.getVectorNumElements())); 840 break; 841 } 842 } 843 844 Lo = DAG.getNode(N->getOpcode(), dl, LoVT, Lo); 845 Hi = DAG.getNode(N->getOpcode(), dl, HiVT, Hi); 846} 847 848void DAGTypeLegalizer::SplitVecRes_VECTOR_SHUFFLE(ShuffleVectorSDNode *N, 849 SDValue &Lo, SDValue &Hi) { 850 // The low and high parts of the original input give four input vectors. 851 SDValue Inputs[4]; 852 DebugLoc dl = N->getDebugLoc(); 853 GetSplitVector(N->getOperand(0), Inputs[0], Inputs[1]); 854 GetSplitVector(N->getOperand(1), Inputs[2], Inputs[3]); 855 EVT NewVT = Inputs[0].getValueType(); 856 unsigned NewElts = NewVT.getVectorNumElements(); 857 858 // If Lo or Hi uses elements from at most two of the four input vectors, then 859 // express it as a vector shuffle of those two inputs. Otherwise extract the 860 // input elements by hand and construct the Lo/Hi output using a BUILD_VECTOR. 861 SmallVector<int, 16> Ops; 862 for (unsigned High = 0; High < 2; ++High) { 863 SDValue &Output = High ? Hi : Lo; 864 865 // Build a shuffle mask for the output, discovering on the fly which 866 // input vectors to use as shuffle operands (recorded in InputUsed). 867 // If building a suitable shuffle vector proves too hard, then bail 868 // out with useBuildVector set. 869 unsigned InputUsed[2] = { -1U, -1U }; // Not yet discovered. 870 unsigned FirstMaskIdx = High * NewElts; 871 bool useBuildVector = false; 872 for (unsigned MaskOffset = 0; MaskOffset < NewElts; ++MaskOffset) { 873 // The mask element. This indexes into the input. 874 int Idx = N->getMaskElt(FirstMaskIdx + MaskOffset); 875 876 // The input vector this mask element indexes into. 877 unsigned Input = (unsigned)Idx / NewElts; 878 879 if (Input >= array_lengthof(Inputs)) { 880 // The mask element does not index into any input vector. 881 Ops.push_back(-1); 882 continue; 883 } 884 885 // Turn the index into an offset from the start of the input vector. 886 Idx -= Input * NewElts; 887 888 // Find or create a shuffle vector operand to hold this input. 889 unsigned OpNo; 890 for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) { 891 if (InputUsed[OpNo] == Input) { 892 // This input vector is already an operand. 893 break; 894 } else if (InputUsed[OpNo] == -1U) { 895 // Create a new operand for this input vector. 896 InputUsed[OpNo] = Input; 897 break; 898 } 899 } 900 901 if (OpNo >= array_lengthof(InputUsed)) { 902 // More than two input vectors used! Give up on trying to create a 903 // shuffle vector. Insert all elements into a BUILD_VECTOR instead. 904 useBuildVector = true; 905 break; 906 } 907 908 // Add the mask index for the new shuffle vector. 909 Ops.push_back(Idx + OpNo * NewElts); 910 } 911 912 if (useBuildVector) { 913 EVT EltVT = NewVT.getVectorElementType(); 914 SmallVector<SDValue, 16> SVOps; 915 916 // Extract the input elements by hand. 917 for (unsigned MaskOffset = 0; MaskOffset < NewElts; ++MaskOffset) { 918 // The mask element. This indexes into the input. 919 int Idx = N->getMaskElt(FirstMaskIdx + MaskOffset); 920 921 // The input vector this mask element indexes into. 922 unsigned Input = (unsigned)Idx / NewElts; 923 924 if (Input >= array_lengthof(Inputs)) { 925 // The mask element is "undef" or indexes off the end of the input. 926 SVOps.push_back(DAG.getUNDEF(EltVT)); 927 continue; 928 } 929 930 // Turn the index into an offset from the start of the input vector. 931 Idx -= Input * NewElts; 932 933 // Extract the vector element by hand. 934 SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, 935 Inputs[Input], DAG.getIntPtrConstant(Idx))); 936 } 937 938 // Construct the Lo/Hi output using a BUILD_VECTOR. 939 Output = DAG.getNode(ISD::BUILD_VECTOR,dl,NewVT, &SVOps[0], SVOps.size()); 940 } else if (InputUsed[0] == -1U) { 941 // No input vectors were used! The result is undefined. 942 Output = DAG.getUNDEF(NewVT); 943 } else { 944 SDValue Op0 = Inputs[InputUsed[0]]; 945 // If only one input was used, use an undefined vector for the other. 946 SDValue Op1 = InputUsed[1] == -1U ? 947 DAG.getUNDEF(NewVT) : Inputs[InputUsed[1]]; 948 // At least one input vector was used. Create a new shuffle vector. 949 Output = DAG.getVectorShuffle(NewVT, dl, Op0, Op1, &Ops[0]); 950 } 951 952 Ops.clear(); 953 } 954} 955 956 957//===----------------------------------------------------------------------===// 958// Operand Vector Splitting 959//===----------------------------------------------------------------------===// 960 961/// SplitVectorOperand - This method is called when the specified operand of the 962/// specified node is found to need vector splitting. At this point, all of the 963/// result types of the node are known to be legal, but other operands of the 964/// node may need legalization as well as the specified one. 965bool DAGTypeLegalizer::SplitVectorOperand(SDNode *N, unsigned OpNo) { 966 DEBUG(dbgs() << "Split node operand: "; 967 N->dump(&DAG); 968 dbgs() << "\n"); 969 SDValue Res = SDValue(); 970 971 if (Res.getNode() == 0) { 972 switch (N->getOpcode()) { 973 default: 974#ifndef NDEBUG 975 dbgs() << "SplitVectorOperand Op #" << OpNo << ": "; 976 N->dump(&DAG); 977 dbgs() << "\n"; 978#endif 979 llvm_unreachable("Do not know how to split this operator's operand!"); 980 981 case ISD::BIT_CONVERT: Res = SplitVecOp_BIT_CONVERT(N); break; 982 case ISD::EXTRACT_SUBVECTOR: Res = SplitVecOp_EXTRACT_SUBVECTOR(N); break; 983 case ISD::EXTRACT_VECTOR_ELT:Res = SplitVecOp_EXTRACT_VECTOR_ELT(N); break; 984 case ISD::CONCAT_VECTORS: Res = SplitVecOp_CONCAT_VECTORS(N); break; 985 case ISD::STORE: 986 Res = SplitVecOp_STORE(cast<StoreSDNode>(N), OpNo); 987 break; 988 989 case ISD::CTTZ: 990 case ISD::CTLZ: 991 case ISD::CTPOP: 992 case ISD::FP_TO_SINT: 993 case ISD::FP_TO_UINT: 994 case ISD::SINT_TO_FP: 995 case ISD::UINT_TO_FP: 996 case ISD::TRUNCATE: 997 case ISD::SIGN_EXTEND: 998 case ISD::ZERO_EXTEND: 999 case ISD::ANY_EXTEND: 1000 Res = SplitVecOp_UnaryOp(N); 1001 break; 1002 } 1003 } 1004 1005 // If the result is null, the sub-method took care of registering results etc. 1006 if (!Res.getNode()) return false; 1007 1008 // If the result is N, the sub-method updated N in place. Tell the legalizer 1009 // core about this. 1010 if (Res.getNode() == N) 1011 return true; 1012 1013 assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 && 1014 "Invalid operand expansion"); 1015 1016 ReplaceValueWith(SDValue(N, 0), Res); 1017 return false; 1018} 1019 1020SDValue DAGTypeLegalizer::SplitVecOp_UnaryOp(SDNode *N) { 1021 // The result has a legal vector type, but the input needs splitting. 1022 EVT ResVT = N->getValueType(0); 1023 SDValue Lo, Hi; 1024 DebugLoc dl = N->getDebugLoc(); 1025 GetSplitVector(N->getOperand(0), Lo, Hi); 1026 EVT InVT = Lo.getValueType(); 1027 1028 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), ResVT.getVectorElementType(), 1029 InVT.getVectorNumElements()); 1030 1031 Lo = DAG.getNode(N->getOpcode(), dl, OutVT, Lo); 1032 Hi = DAG.getNode(N->getOpcode(), dl, OutVT, Hi); 1033 1034 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi); 1035} 1036 1037SDValue DAGTypeLegalizer::SplitVecOp_BIT_CONVERT(SDNode *N) { 1038 // For example, i64 = BIT_CONVERT v4i16 on alpha. Typically the vector will 1039 // end up being split all the way down to individual components. Convert the 1040 // split pieces into integers and reassemble. 1041 SDValue Lo, Hi; 1042 GetSplitVector(N->getOperand(0), Lo, Hi); 1043 Lo = BitConvertToInteger(Lo); 1044 Hi = BitConvertToInteger(Hi); 1045 1046 if (TLI.isBigEndian()) 1047 std::swap(Lo, Hi); 1048 1049 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), N->getValueType(0), 1050 JoinIntegers(Lo, Hi)); 1051} 1052 1053SDValue DAGTypeLegalizer::SplitVecOp_EXTRACT_SUBVECTOR(SDNode *N) { 1054 // We know that the extracted result type is legal. For now, assume the index 1055 // is a constant. 1056 EVT SubVT = N->getValueType(0); 1057 SDValue Idx = N->getOperand(1); 1058 DebugLoc dl = N->getDebugLoc(); 1059 SDValue Lo, Hi; 1060 GetSplitVector(N->getOperand(0), Lo, Hi); 1061 1062 uint64_t LoElts = Lo.getValueType().getVectorNumElements(); 1063 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); 1064 1065 if (IdxVal < LoElts) { 1066 assert(IdxVal + SubVT.getVectorNumElements() <= LoElts && 1067 "Extracted subvector crosses vector split!"); 1068 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, Lo, Idx); 1069 } else { 1070 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, Hi, 1071 DAG.getConstant(IdxVal - LoElts, Idx.getValueType())); 1072 } 1073} 1074 1075SDValue DAGTypeLegalizer::SplitVecOp_EXTRACT_VECTOR_ELT(SDNode *N) { 1076 SDValue Vec = N->getOperand(0); 1077 SDValue Idx = N->getOperand(1); 1078 EVT VecVT = Vec.getValueType(); 1079 1080 if (isa<ConstantSDNode>(Idx)) { 1081 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); 1082 assert(IdxVal < VecVT.getVectorNumElements() && "Invalid vector index!"); 1083 1084 SDValue Lo, Hi; 1085 GetSplitVector(Vec, Lo, Hi); 1086 1087 uint64_t LoElts = Lo.getValueType().getVectorNumElements(); 1088 1089 if (IdxVal < LoElts) 1090 return SDValue(DAG.UpdateNodeOperands(N, Lo, Idx), 0); 1091 return SDValue(DAG.UpdateNodeOperands(N, Hi, 1092 DAG.getConstant(IdxVal - LoElts, 1093 Idx.getValueType())), 0); 1094 } 1095 1096 // Store the vector to the stack. 1097 EVT EltVT = VecVT.getVectorElementType(); 1098 DebugLoc dl = N->getDebugLoc(); 1099 SDValue StackPtr = DAG.CreateStackTemporary(VecVT); 1100 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex(); 1101 const Value *SV = PseudoSourceValue::getFixedStack(SPFI); 1102 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, SV, 0, 1103 false, false, 0); 1104 1105 // Load back the required element. 1106 StackPtr = GetVectorElementPointer(StackPtr, EltVT, Idx); 1107 return DAG.getExtLoad(ISD::EXTLOAD, N->getValueType(0), dl, Store, StackPtr, 1108 MachinePointerInfo::getFixedStack(SPFI), 1109 EltVT, false, false, 0); 1110} 1111 1112SDValue DAGTypeLegalizer::SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo) { 1113 assert(N->isUnindexed() && "Indexed store of vector?"); 1114 assert(OpNo == 1 && "Can only split the stored value"); 1115 DebugLoc DL = N->getDebugLoc(); 1116 1117 bool isTruncating = N->isTruncatingStore(); 1118 SDValue Ch = N->getChain(); 1119 SDValue Ptr = N->getBasePtr(); 1120 EVT MemoryVT = N->getMemoryVT(); 1121 unsigned Alignment = N->getOriginalAlignment(); 1122 bool isVol = N->isVolatile(); 1123 bool isNT = N->isNonTemporal(); 1124 SDValue Lo, Hi; 1125 GetSplitVector(N->getOperand(1), Lo, Hi); 1126 1127 EVT LoMemVT, HiMemVT; 1128 GetSplitDestVTs(MemoryVT, LoMemVT, HiMemVT); 1129 1130 unsigned IncrementSize = LoMemVT.getSizeInBits()/8; 1131 1132 if (isTruncating) 1133 Lo = DAG.getTruncStore(Ch, DL, Lo, Ptr, N->getPointerInfo(), 1134 LoMemVT, isVol, isNT, Alignment); 1135 else 1136 Lo = DAG.getStore(Ch, DL, Lo, Ptr, N->getPointerInfo(), 1137 isVol, isNT, Alignment); 1138 1139 // Increment the pointer to the other half. 1140 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr, 1141 DAG.getIntPtrConstant(IncrementSize)); 1142 1143 if (isTruncating) 1144 Hi = DAG.getTruncStore(Ch, DL, Hi, Ptr, 1145 N->getPointerInfo().getWithOffset(IncrementSize), 1146 HiMemVT, isVol, isNT, Alignment); 1147 else 1148 Hi = DAG.getStore(Ch, DL, Hi, Ptr, 1149 N->getPointerInfo().getWithOffset(IncrementSize), 1150 isVol, isNT, Alignment); 1151 1152 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi); 1153} 1154 1155SDValue DAGTypeLegalizer::SplitVecOp_CONCAT_VECTORS(SDNode *N) { 1156 DebugLoc DL = N->getDebugLoc(); 1157 1158 // The input operands all must have the same type, and we know the result the 1159 // result type is valid. Convert this to a buildvector which extracts all the 1160 // input elements. 1161 // TODO: If the input elements are power-two vectors, we could convert this to 1162 // a new CONCAT_VECTORS node with elements that are half-wide. 1163 SmallVector<SDValue, 32> Elts; 1164 EVT EltVT = N->getValueType(0).getVectorElementType(); 1165 for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) { 1166 SDValue Op = N->getOperand(op); 1167 for (unsigned i = 0, e = Op.getValueType().getVectorNumElements(); 1168 i != e; ++i) { 1169 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, 1170 Op, DAG.getIntPtrConstant(i))); 1171 1172 } 1173 } 1174 1175 return DAG.getNode(ISD::BUILD_VECTOR, DL, N->getValueType(0), 1176 &Elts[0], Elts.size()); 1177} 1178 1179 1180//===----------------------------------------------------------------------===// 1181// Result Vector Widening 1182//===----------------------------------------------------------------------===// 1183 1184void DAGTypeLegalizer::WidenVectorResult(SDNode *N, unsigned ResNo) { 1185 DEBUG(dbgs() << "Widen node result " << ResNo << ": "; 1186 N->dump(&DAG); 1187 dbgs() << "\n"); 1188 1189 // See if the target wants to custom widen this node. 1190 if (CustomWidenLowerNode(N, N->getValueType(ResNo))) 1191 return; 1192 1193 SDValue Res = SDValue(); 1194 switch (N->getOpcode()) { 1195 default: 1196#ifndef NDEBUG 1197 dbgs() << "WidenVectorResult #" << ResNo << ": "; 1198 N->dump(&DAG); 1199 dbgs() << "\n"; 1200#endif 1201 llvm_unreachable("Do not know how to widen the result of this operator!"); 1202 1203 case ISD::BIT_CONVERT: Res = WidenVecRes_BIT_CONVERT(N); break; 1204 case ISD::BUILD_VECTOR: Res = WidenVecRes_BUILD_VECTOR(N); break; 1205 case ISD::CONCAT_VECTORS: Res = WidenVecRes_CONCAT_VECTORS(N); break; 1206 case ISD::CONVERT_RNDSAT: Res = WidenVecRes_CONVERT_RNDSAT(N); break; 1207 case ISD::EXTRACT_SUBVECTOR: Res = WidenVecRes_EXTRACT_SUBVECTOR(N); break; 1208 case ISD::FP_ROUND_INREG: Res = WidenVecRes_InregOp(N); break; 1209 case ISD::INSERT_VECTOR_ELT: Res = WidenVecRes_INSERT_VECTOR_ELT(N); break; 1210 case ISD::LOAD: Res = WidenVecRes_LOAD(N); break; 1211 case ISD::SCALAR_TO_VECTOR: Res = WidenVecRes_SCALAR_TO_VECTOR(N); break; 1212 case ISD::SIGN_EXTEND_INREG: Res = WidenVecRes_InregOp(N); break; 1213 case ISD::SELECT: Res = WidenVecRes_SELECT(N); break; 1214 case ISD::SELECT_CC: Res = WidenVecRes_SELECT_CC(N); break; 1215 case ISD::SETCC: Res = WidenVecRes_SETCC(N); break; 1216 case ISD::UNDEF: Res = WidenVecRes_UNDEF(N); break; 1217 case ISD::VECTOR_SHUFFLE: 1218 Res = WidenVecRes_VECTOR_SHUFFLE(cast<ShuffleVectorSDNode>(N)); 1219 break; 1220 case ISD::VSETCC: 1221 Res = WidenVecRes_VSETCC(N); 1222 break; 1223 1224 case ISD::ADD: 1225 case ISD::AND: 1226 case ISD::BSWAP: 1227 case ISD::FADD: 1228 case ISD::FCOPYSIGN: 1229 case ISD::FDIV: 1230 case ISD::FMUL: 1231 case ISD::FPOW: 1232 case ISD::FREM: 1233 case ISD::FSUB: 1234 case ISD::MUL: 1235 case ISD::MULHS: 1236 case ISD::MULHU: 1237 case ISD::OR: 1238 case ISD::SDIV: 1239 case ISD::SREM: 1240 case ISD::UDIV: 1241 case ISD::UREM: 1242 case ISD::SUB: 1243 case ISD::XOR: 1244 Res = WidenVecRes_Binary(N); 1245 break; 1246 1247 case ISD::FPOWI: 1248 Res = WidenVecRes_POWI(N); 1249 break; 1250 1251 case ISD::SHL: 1252 case ISD::SRA: 1253 case ISD::SRL: 1254 Res = WidenVecRes_Shift(N); 1255 break; 1256 1257 case ISD::FP_ROUND: 1258 case ISD::FP_TO_SINT: 1259 case ISD::FP_TO_UINT: 1260 case ISD::SINT_TO_FP: 1261 case ISD::UINT_TO_FP: 1262 case ISD::TRUNCATE: 1263 case ISD::SIGN_EXTEND: 1264 case ISD::ZERO_EXTEND: 1265 case ISD::ANY_EXTEND: 1266 Res = WidenVecRes_Convert(N); 1267 break; 1268 1269 case ISD::CTLZ: 1270 case ISD::CTPOP: 1271 case ISD::CTTZ: 1272 case ISD::FABS: 1273 case ISD::FCOS: 1274 case ISD::FNEG: 1275 case ISD::FSIN: 1276 case ISD::FSQRT: 1277 case ISD::FEXP: 1278 case ISD::FEXP2: 1279 case ISD::FLOG: 1280 case ISD::FLOG2: 1281 case ISD::FLOG10: 1282 Res = WidenVecRes_Unary(N); 1283 break; 1284 } 1285 1286 // If Res is null, the sub-method took care of registering the result. 1287 if (Res.getNode()) 1288 SetWidenedVector(SDValue(N, ResNo), Res); 1289} 1290 1291SDValue DAGTypeLegalizer::WidenVecRes_Binary(SDNode *N) { 1292 // Binary op widening. 1293 unsigned Opcode = N->getOpcode(); 1294 DebugLoc dl = N->getDebugLoc(); 1295 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); 1296 EVT WidenEltVT = WidenVT.getVectorElementType(); 1297 EVT VT = WidenVT; 1298 unsigned NumElts = VT.getVectorNumElements(); 1299 while (!TLI.isTypeSynthesizable(VT) && NumElts != 1) { 1300 NumElts = NumElts / 2; 1301 VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts); 1302 } 1303 1304 if (NumElts != 1 && !TLI.canOpTrap(N->getOpcode(), VT)) { 1305 // Operation doesn't trap so just widen as normal. 1306 SDValue InOp1 = GetWidenedVector(N->getOperand(0)); 1307 SDValue InOp2 = GetWidenedVector(N->getOperand(1)); 1308 return DAG.getNode(N->getOpcode(), dl, WidenVT, InOp1, InOp2); 1309 } 1310 1311 // No legal vector version so unroll the vector operation and then widen. 1312 if (NumElts == 1) 1313 return DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()); 1314 1315 // Since the operation can trap, apply operation on the original vector. 1316 EVT MaxVT = VT; 1317 SDValue InOp1 = GetWidenedVector(N->getOperand(0)); 1318 SDValue InOp2 = GetWidenedVector(N->getOperand(1)); 1319 unsigned CurNumElts = N->getValueType(0).getVectorNumElements(); 1320 1321 SmallVector<SDValue, 16> ConcatOps(CurNumElts); 1322 unsigned ConcatEnd = 0; // Current ConcatOps index. 1323 int Idx = 0; // Current Idx into input vectors. 1324 1325 // NumElts := greatest synthesizable vector size (at most WidenVT) 1326 // while (orig. vector has unhandled elements) { 1327 // take munches of size NumElts from the beginning and add to ConcatOps 1328 // NumElts := next smaller supported vector size or 1 1329 // } 1330 while (CurNumElts != 0) { 1331 while (CurNumElts >= NumElts) { 1332 SDValue EOp1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp1, 1333 DAG.getIntPtrConstant(Idx)); 1334 SDValue EOp2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp2, 1335 DAG.getIntPtrConstant(Idx)); 1336 ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, dl, VT, EOp1, EOp2); 1337 Idx += NumElts; 1338 CurNumElts -= NumElts; 1339 } 1340 do { 1341 NumElts = NumElts / 2; 1342 VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts); 1343 } while (!TLI.isTypeSynthesizable(VT) && NumElts != 1); 1344 1345 if (NumElts == 1) { 1346 for (unsigned i = 0; i != CurNumElts; ++i, ++Idx) { 1347 SDValue EOp1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT, 1348 InOp1, DAG.getIntPtrConstant(Idx)); 1349 SDValue EOp2 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT, 1350 InOp2, DAG.getIntPtrConstant(Idx)); 1351 ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, dl, WidenEltVT, 1352 EOp1, EOp2); 1353 } 1354 CurNumElts = 0; 1355 } 1356 } 1357 1358 // Check to see if we have a single operation with the widen type. 1359 if (ConcatEnd == 1) { 1360 VT = ConcatOps[0].getValueType(); 1361 if (VT == WidenVT) 1362 return ConcatOps[0]; 1363 } 1364 1365 // while (Some element of ConcatOps is not of type MaxVT) { 1366 // From the end of ConcatOps, collect elements of the same type and put 1367 // them into an op of the next larger supported type 1368 // } 1369 while (ConcatOps[ConcatEnd-1].getValueType() != MaxVT) { 1370 Idx = ConcatEnd - 1; 1371 VT = ConcatOps[Idx--].getValueType(); 1372 while (Idx >= 0 && ConcatOps[Idx].getValueType() == VT) 1373 Idx--; 1374 1375 int NextSize = VT.isVector() ? VT.getVectorNumElements() : 1; 1376 EVT NextVT; 1377 do { 1378 NextSize *= 2; 1379 NextVT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NextSize); 1380 } while (!TLI.isTypeSynthesizable(NextVT)); 1381 1382 if (!VT.isVector()) { 1383 // Scalar type, create an INSERT_VECTOR_ELEMENT of type NextVT 1384 SDValue VecOp = DAG.getUNDEF(NextVT); 1385 unsigned NumToInsert = ConcatEnd - Idx - 1; 1386 for (unsigned i = 0, OpIdx = Idx+1; i < NumToInsert; i++, OpIdx++) { 1387 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, NextVT, VecOp, 1388 ConcatOps[OpIdx], DAG.getIntPtrConstant(i)); 1389 } 1390 ConcatOps[Idx+1] = VecOp; 1391 ConcatEnd = Idx + 2; 1392 } else { 1393 // Vector type, create a CONCAT_VECTORS of type NextVT 1394 SDValue undefVec = DAG.getUNDEF(VT); 1395 unsigned OpsToConcat = NextSize/VT.getVectorNumElements(); 1396 SmallVector<SDValue, 16> SubConcatOps(OpsToConcat); 1397 unsigned RealVals = ConcatEnd - Idx - 1; 1398 unsigned SubConcatEnd = 0; 1399 unsigned SubConcatIdx = Idx + 1; 1400 while (SubConcatEnd < RealVals) 1401 SubConcatOps[SubConcatEnd++] = ConcatOps[++Idx]; 1402 while (SubConcatEnd < OpsToConcat) 1403 SubConcatOps[SubConcatEnd++] = undefVec; 1404 ConcatOps[SubConcatIdx] = DAG.getNode(ISD::CONCAT_VECTORS, dl, 1405 NextVT, &SubConcatOps[0], 1406 OpsToConcat); 1407 ConcatEnd = SubConcatIdx + 1; 1408 } 1409 } 1410 1411 // Check to see if we have a single operation with the widen type. 1412 if (ConcatEnd == 1) { 1413 VT = ConcatOps[0].getValueType(); 1414 if (VT == WidenVT) 1415 return ConcatOps[0]; 1416 } 1417 1418 // add undefs of size MaxVT until ConcatOps grows to length of WidenVT 1419 unsigned NumOps = WidenVT.getVectorNumElements()/MaxVT.getVectorNumElements(); 1420 if (NumOps != ConcatEnd ) { 1421 SDValue UndefVal = DAG.getUNDEF(MaxVT); 1422 for (unsigned j = ConcatEnd; j < NumOps; ++j) 1423 ConcatOps[j] = UndefVal; 1424 } 1425 return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &ConcatOps[0], NumOps); 1426} 1427 1428SDValue DAGTypeLegalizer::WidenVecRes_Convert(SDNode *N) { 1429 SDValue InOp = N->getOperand(0); 1430 DebugLoc dl = N->getDebugLoc(); 1431 1432 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); 1433 unsigned WidenNumElts = WidenVT.getVectorNumElements(); 1434 1435 EVT InVT = InOp.getValueType(); 1436 EVT InEltVT = InVT.getVectorElementType(); 1437 EVT InWidenVT = EVT::getVectorVT(*DAG.getContext(), InEltVT, WidenNumElts); 1438 1439 unsigned Opcode = N->getOpcode(); 1440 unsigned InVTNumElts = InVT.getVectorNumElements(); 1441 1442 if (getTypeAction(InVT) == WidenVector) { 1443 InOp = GetWidenedVector(N->getOperand(0)); 1444 InVT = InOp.getValueType(); 1445 InVTNumElts = InVT.getVectorNumElements(); 1446 if (InVTNumElts == WidenNumElts) 1447 return DAG.getNode(Opcode, dl, WidenVT, InOp); 1448 } 1449 1450 if (TLI.isTypeSynthesizable(InWidenVT)) { 1451 // Because the result and the input are different vector types, widening 1452 // the result could create a legal type but widening the input might make 1453 // it an illegal type that might lead to repeatedly splitting the input 1454 // and then widening it. To avoid this, we widen the input only if 1455 // it results in a legal type. 1456 if (WidenNumElts % InVTNumElts == 0) { 1457 // Widen the input and call convert on the widened input vector. 1458 unsigned NumConcat = WidenNumElts/InVTNumElts; 1459 SmallVector<SDValue, 16> Ops(NumConcat); 1460 Ops[0] = InOp; 1461 SDValue UndefVal = DAG.getUNDEF(InVT); 1462 for (unsigned i = 1; i != NumConcat; ++i) 1463 Ops[i] = UndefVal; 1464 return DAG.getNode(Opcode, dl, WidenVT, 1465 DAG.getNode(ISD::CONCAT_VECTORS, dl, InWidenVT, 1466 &Ops[0], NumConcat)); 1467 } 1468 1469 if (InVTNumElts % WidenNumElts == 0) { 1470 // Extract the input and convert the shorten input vector. 1471 return DAG.getNode(Opcode, dl, WidenVT, 1472 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InWidenVT, 1473 InOp, DAG.getIntPtrConstant(0))); 1474 } 1475 } 1476 1477 // Otherwise unroll into some nasty scalar code and rebuild the vector. 1478 SmallVector<SDValue, 16> Ops(WidenNumElts); 1479 EVT EltVT = WidenVT.getVectorElementType(); 1480 unsigned MinElts = std::min(InVTNumElts, WidenNumElts); 1481 unsigned i; 1482 for (i=0; i < MinElts; ++i) 1483 Ops[i] = DAG.getNode(Opcode, dl, EltVT, 1484 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, InEltVT, InOp, 1485 DAG.getIntPtrConstant(i))); 1486 1487 SDValue UndefVal = DAG.getUNDEF(EltVT); 1488 for (; i < WidenNumElts; ++i) 1489 Ops[i] = UndefVal; 1490 1491 return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, &Ops[0], WidenNumElts); 1492} 1493 1494SDValue DAGTypeLegalizer::WidenVecRes_POWI(SDNode *N) { 1495 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); 1496 SDValue InOp = GetWidenedVector(N->getOperand(0)); 1497 SDValue ShOp = N->getOperand(1); 1498 return DAG.getNode(N->getOpcode(), N->getDebugLoc(), WidenVT, InOp, ShOp); 1499} 1500 1501SDValue DAGTypeLegalizer::WidenVecRes_Shift(SDNode *N) { 1502 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); 1503 SDValue InOp = GetWidenedVector(N->getOperand(0)); 1504 SDValue ShOp = N->getOperand(1); 1505 1506 EVT ShVT = ShOp.getValueType(); 1507 if (getTypeAction(ShVT) == WidenVector) { 1508 ShOp = GetWidenedVector(ShOp); 1509 ShVT = ShOp.getValueType(); 1510 } 1511 EVT ShWidenVT = EVT::getVectorVT(*DAG.getContext(), 1512 ShVT.getVectorElementType(), 1513 WidenVT.getVectorNumElements()); 1514 if (ShVT != ShWidenVT) 1515 ShOp = ModifyToType(ShOp, ShWidenVT); 1516 1517 return DAG.getNode(N->getOpcode(), N->getDebugLoc(), WidenVT, InOp, ShOp); 1518} 1519 1520SDValue DAGTypeLegalizer::WidenVecRes_Unary(SDNode *N) { 1521 // Unary op widening. 1522 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); 1523 SDValue InOp = GetWidenedVector(N->getOperand(0)); 1524 return DAG.getNode(N->getOpcode(), N->getDebugLoc(), WidenVT, InOp); 1525} 1526 1527SDValue DAGTypeLegalizer::WidenVecRes_InregOp(SDNode *N) { 1528 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); 1529 EVT ExtVT = EVT::getVectorVT(*DAG.getContext(), 1530 cast<VTSDNode>(N->getOperand(1))->getVT() 1531 .getVectorElementType(), 1532 WidenVT.getVectorNumElements()); 1533 SDValue WidenLHS = GetWidenedVector(N->getOperand(0)); 1534 return DAG.getNode(N->getOpcode(), N->getDebugLoc(), 1535 WidenVT, WidenLHS, DAG.getValueType(ExtVT)); 1536} 1537 1538SDValue DAGTypeLegalizer::WidenVecRes_BIT_CONVERT(SDNode *N) { 1539 SDValue InOp = N->getOperand(0); 1540 EVT InVT = InOp.getValueType(); 1541 EVT VT = N->getValueType(0); 1542 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT); 1543 DebugLoc dl = N->getDebugLoc(); 1544 1545 switch (getTypeAction(InVT)) { 1546 default: 1547 assert(false && "Unknown type action!"); 1548 break; 1549 case Legal: 1550 break; 1551 case PromoteInteger: 1552 // If the InOp is promoted to the same size, convert it. Otherwise, 1553 // fall out of the switch and widen the promoted input. 1554 InOp = GetPromotedInteger(InOp); 1555 InVT = InOp.getValueType(); 1556 if (WidenVT.bitsEq(InVT)) 1557 return DAG.getNode(ISD::BIT_CONVERT, dl, WidenVT, InOp); 1558 break; 1559 case SoftenFloat: 1560 case ExpandInteger: 1561 case ExpandFloat: 1562 case ScalarizeVector: 1563 case SplitVector: 1564 break; 1565 case WidenVector: 1566 // If the InOp is widened to the same size, convert it. Otherwise, fall 1567 // out of the switch and widen the widened input. 1568 InOp = GetWidenedVector(InOp); 1569 InVT = InOp.getValueType(); 1570 if (WidenVT.bitsEq(InVT)) 1571 // The input widens to the same size. Convert to the widen value. 1572 return DAG.getNode(ISD::BIT_CONVERT, dl, WidenVT, InOp); 1573 break; 1574 } 1575 1576 unsigned WidenSize = WidenVT.getSizeInBits(); 1577 unsigned InSize = InVT.getSizeInBits(); 1578 if (WidenSize % InSize == 0) { 1579 // Determine new input vector type. The new input vector type will use 1580 // the same element type (if its a vector) or use the input type as a 1581 // vector. It is the same size as the type to widen to. 1582 EVT NewInVT; 1583 unsigned NewNumElts = WidenSize / InSize; 1584 if (InVT.isVector()) { 1585 EVT InEltVT = InVT.getVectorElementType(); 1586 NewInVT = EVT::getVectorVT(*DAG.getContext(), InEltVT, 1587 WidenSize / InEltVT.getSizeInBits()); 1588 } else { 1589 NewInVT = EVT::getVectorVT(*DAG.getContext(), InVT, NewNumElts); 1590 } 1591 1592 if (TLI.isTypeSynthesizable(NewInVT)) { 1593 // Because the result and the input are different vector types, widening 1594 // the result could create a legal type but widening the input might make 1595 // it an illegal type that might lead to repeatedly splitting the input 1596 // and then widening it. To avoid this, we widen the input only if 1597 // it results in a legal type. 1598 SmallVector<SDValue, 16> Ops(NewNumElts); 1599 SDValue UndefVal = DAG.getUNDEF(InVT); 1600 Ops[0] = InOp; 1601 for (unsigned i = 1; i < NewNumElts; ++i) 1602 Ops[i] = UndefVal; 1603 1604 SDValue NewVec; 1605 if (InVT.isVector()) 1606 NewVec = DAG.getNode(ISD::CONCAT_VECTORS, dl, 1607 NewInVT, &Ops[0], NewNumElts); 1608 else 1609 NewVec = DAG.getNode(ISD::BUILD_VECTOR, dl, 1610 NewInVT, &Ops[0], NewNumElts); 1611 return DAG.getNode(ISD::BIT_CONVERT, dl, WidenVT, NewVec); 1612 } 1613 } 1614 1615 return CreateStackStoreLoad(InOp, WidenVT); 1616} 1617 1618SDValue DAGTypeLegalizer::WidenVecRes_BUILD_VECTOR(SDNode *N) { 1619 DebugLoc dl = N->getDebugLoc(); 1620 // Build a vector with undefined for the new nodes. 1621 EVT VT = N->getValueType(0); 1622 EVT EltVT = VT.getVectorElementType(); 1623 unsigned NumElts = VT.getVectorNumElements(); 1624 1625 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT); 1626 unsigned WidenNumElts = WidenVT.getVectorNumElements(); 1627 1628 SmallVector<SDValue, 16> NewOps(N->op_begin(), N->op_end()); 1629 NewOps.reserve(WidenNumElts); 1630 for (unsigned i = NumElts; i < WidenNumElts; ++i) 1631 NewOps.push_back(DAG.getUNDEF(EltVT)); 1632 1633 return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, &NewOps[0], NewOps.size()); 1634} 1635 1636SDValue DAGTypeLegalizer::WidenVecRes_CONCAT_VECTORS(SDNode *N) { 1637 EVT InVT = N->getOperand(0).getValueType(); 1638 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); 1639 DebugLoc dl = N->getDebugLoc(); 1640 unsigned WidenNumElts = WidenVT.getVectorNumElements(); 1641 unsigned NumOperands = N->getNumOperands(); 1642 1643 bool InputWidened = false; // Indicates we need to widen the input. 1644 if (getTypeAction(InVT) != WidenVector) { 1645 if (WidenVT.getVectorNumElements() % InVT.getVectorNumElements() == 0) { 1646 // Add undef vectors to widen to correct length. 1647 unsigned NumConcat = WidenVT.getVectorNumElements() / 1648 InVT.getVectorNumElements(); 1649 SDValue UndefVal = DAG.getUNDEF(InVT); 1650 SmallVector<SDValue, 16> Ops(NumConcat); 1651 for (unsigned i=0; i < NumOperands; ++i) 1652 Ops[i] = N->getOperand(i); 1653 for (unsigned i = NumOperands; i != NumConcat; ++i) 1654 Ops[i] = UndefVal; 1655 return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &Ops[0], NumConcat); 1656 } 1657 } else { 1658 InputWidened = true; 1659 if (WidenVT == TLI.getTypeToTransformTo(*DAG.getContext(), InVT)) { 1660 // The inputs and the result are widen to the same value. 1661 unsigned i; 1662 for (i=1; i < NumOperands; ++i) 1663 if (N->getOperand(i).getOpcode() != ISD::UNDEF) 1664 break; 1665 1666 if (i > NumOperands) 1667 // Everything but the first operand is an UNDEF so just return the 1668 // widened first operand. 1669 return GetWidenedVector(N->getOperand(0)); 1670 1671 if (NumOperands == 2) { 1672 // Replace concat of two operands with a shuffle. 1673 SmallVector<int, 16> MaskOps(WidenNumElts); 1674 for (unsigned i=0; i < WidenNumElts/2; ++i) { 1675 MaskOps[i] = i; 1676 MaskOps[i+WidenNumElts/2] = i+WidenNumElts; 1677 } 1678 return DAG.getVectorShuffle(WidenVT, dl, 1679 GetWidenedVector(N->getOperand(0)), 1680 GetWidenedVector(N->getOperand(1)), 1681 &MaskOps[0]); 1682 } 1683 } 1684 } 1685 1686 // Fall back to use extracts and build vector. 1687 EVT EltVT = WidenVT.getVectorElementType(); 1688 unsigned NumInElts = InVT.getVectorNumElements(); 1689 SmallVector<SDValue, 16> Ops(WidenNumElts); 1690 unsigned Idx = 0; 1691 for (unsigned i=0; i < NumOperands; ++i) { 1692 SDValue InOp = N->getOperand(i); 1693 if (InputWidened) 1694 InOp = GetWidenedVector(InOp); 1695 for (unsigned j=0; j < NumInElts; ++j) 1696 Ops[Idx++] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp, 1697 DAG.getIntPtrConstant(j)); 1698 } 1699 SDValue UndefVal = DAG.getUNDEF(EltVT); 1700 for (; Idx < WidenNumElts; ++Idx) 1701 Ops[Idx] = UndefVal; 1702 return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, &Ops[0], WidenNumElts); 1703} 1704 1705SDValue DAGTypeLegalizer::WidenVecRes_CONVERT_RNDSAT(SDNode *N) { 1706 DebugLoc dl = N->getDebugLoc(); 1707 SDValue InOp = N->getOperand(0); 1708 SDValue RndOp = N->getOperand(3); 1709 SDValue SatOp = N->getOperand(4); 1710 1711 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); 1712 unsigned WidenNumElts = WidenVT.getVectorNumElements(); 1713 1714 EVT InVT = InOp.getValueType(); 1715 EVT InEltVT = InVT.getVectorElementType(); 1716 EVT InWidenVT = EVT::getVectorVT(*DAG.getContext(), InEltVT, WidenNumElts); 1717 1718 SDValue DTyOp = DAG.getValueType(WidenVT); 1719 SDValue STyOp = DAG.getValueType(InWidenVT); 1720 ISD::CvtCode CvtCode = cast<CvtRndSatSDNode>(N)->getCvtCode(); 1721 1722 unsigned InVTNumElts = InVT.getVectorNumElements(); 1723 if (getTypeAction(InVT) == WidenVector) { 1724 InOp = GetWidenedVector(InOp); 1725 InVT = InOp.getValueType(); 1726 InVTNumElts = InVT.getVectorNumElements(); 1727 if (InVTNumElts == WidenNumElts) 1728 return DAG.getConvertRndSat(WidenVT, dl, InOp, DTyOp, STyOp, RndOp, 1729 SatOp, CvtCode); 1730 } 1731 1732 if (TLI.isTypeSynthesizable(InWidenVT)) { 1733 // Because the result and the input are different vector types, widening 1734 // the result could create a legal type but widening the input might make 1735 // it an illegal type that might lead to repeatedly splitting the input 1736 // and then widening it. To avoid this, we widen the input only if 1737 // it results in a legal type. 1738 if (WidenNumElts % InVTNumElts == 0) { 1739 // Widen the input and call convert on the widened input vector. 1740 unsigned NumConcat = WidenNumElts/InVTNumElts; 1741 SmallVector<SDValue, 16> Ops(NumConcat); 1742 Ops[0] = InOp; 1743 SDValue UndefVal = DAG.getUNDEF(InVT); 1744 for (unsigned i = 1; i != NumConcat; ++i) 1745 Ops[i] = UndefVal; 1746 1747 InOp = DAG.getNode(ISD::CONCAT_VECTORS, dl, InWidenVT, &Ops[0],NumConcat); 1748 return DAG.getConvertRndSat(WidenVT, dl, InOp, DTyOp, STyOp, RndOp, 1749 SatOp, CvtCode); 1750 } 1751 1752 if (InVTNumElts % WidenNumElts == 0) { 1753 // Extract the input and convert the shorten input vector. 1754 InOp = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InWidenVT, InOp, 1755 DAG.getIntPtrConstant(0)); 1756 return DAG.getConvertRndSat(WidenVT, dl, InOp, DTyOp, STyOp, RndOp, 1757 SatOp, CvtCode); 1758 } 1759 } 1760 1761 // Otherwise unroll into some nasty scalar code and rebuild the vector. 1762 SmallVector<SDValue, 16> Ops(WidenNumElts); 1763 EVT EltVT = WidenVT.getVectorElementType(); 1764 DTyOp = DAG.getValueType(EltVT); 1765 STyOp = DAG.getValueType(InEltVT); 1766 1767 unsigned MinElts = std::min(InVTNumElts, WidenNumElts); 1768 unsigned i; 1769 for (i=0; i < MinElts; ++i) { 1770 SDValue ExtVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, InEltVT, InOp, 1771 DAG.getIntPtrConstant(i)); 1772 Ops[i] = DAG.getConvertRndSat(WidenVT, dl, ExtVal, DTyOp, STyOp, RndOp, 1773 SatOp, CvtCode); 1774 } 1775 1776 SDValue UndefVal = DAG.getUNDEF(EltVT); 1777 for (; i < WidenNumElts; ++i) 1778 Ops[i] = UndefVal; 1779 1780 return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, &Ops[0], WidenNumElts); 1781} 1782 1783SDValue DAGTypeLegalizer::WidenVecRes_EXTRACT_SUBVECTOR(SDNode *N) { 1784 EVT VT = N->getValueType(0); 1785 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT); 1786 unsigned WidenNumElts = WidenVT.getVectorNumElements(); 1787 SDValue InOp = N->getOperand(0); 1788 SDValue Idx = N->getOperand(1); 1789 DebugLoc dl = N->getDebugLoc(); 1790 1791 if (getTypeAction(InOp.getValueType()) == WidenVector) 1792 InOp = GetWidenedVector(InOp); 1793 1794 EVT InVT = InOp.getValueType(); 1795 1796 ConstantSDNode *CIdx = dyn_cast<ConstantSDNode>(Idx); 1797 if (CIdx) { 1798 unsigned IdxVal = CIdx->getZExtValue(); 1799 // Check if we can just return the input vector after widening. 1800 if (IdxVal == 0 && InVT == WidenVT) 1801 return InOp; 1802 1803 // Check if we can extract from the vector. 1804 unsigned InNumElts = InVT.getVectorNumElements(); 1805 if (IdxVal % WidenNumElts == 0 && IdxVal + WidenNumElts < InNumElts) 1806 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, WidenVT, InOp, Idx); 1807 } 1808 1809 // We could try widening the input to the right length but for now, extract 1810 // the original elements, fill the rest with undefs and build a vector. 1811 SmallVector<SDValue, 16> Ops(WidenNumElts); 1812 EVT EltVT = VT.getVectorElementType(); 1813 EVT IdxVT = Idx.getValueType(); 1814 unsigned NumElts = VT.getVectorNumElements(); 1815 unsigned i; 1816 if (CIdx) { 1817 unsigned IdxVal = CIdx->getZExtValue(); 1818 for (i=0; i < NumElts; ++i) 1819 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp, 1820 DAG.getConstant(IdxVal+i, IdxVT)); 1821 } else { 1822 Ops[0] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp, Idx); 1823 for (i=1; i < NumElts; ++i) { 1824 SDValue NewIdx = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, 1825 DAG.getConstant(i, IdxVT)); 1826 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp, NewIdx); 1827 } 1828 } 1829 1830 SDValue UndefVal = DAG.getUNDEF(EltVT); 1831 for (; i < WidenNumElts; ++i) 1832 Ops[i] = UndefVal; 1833 return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, &Ops[0], WidenNumElts); 1834} 1835 1836SDValue DAGTypeLegalizer::WidenVecRes_INSERT_VECTOR_ELT(SDNode *N) { 1837 SDValue InOp = GetWidenedVector(N->getOperand(0)); 1838 return DAG.getNode(ISD::INSERT_VECTOR_ELT, N->getDebugLoc(), 1839 InOp.getValueType(), InOp, 1840 N->getOperand(1), N->getOperand(2)); 1841} 1842 1843SDValue DAGTypeLegalizer::WidenVecRes_LOAD(SDNode *N) { 1844 LoadSDNode *LD = cast<LoadSDNode>(N); 1845 ISD::LoadExtType ExtType = LD->getExtensionType(); 1846 1847 SDValue Result; 1848 SmallVector<SDValue, 16> LdChain; // Chain for the series of load 1849 if (ExtType != ISD::NON_EXTLOAD) 1850 Result = GenWidenVectorExtLoads(LdChain, LD, ExtType); 1851 else 1852 Result = GenWidenVectorLoads(LdChain, LD); 1853 1854 // If we generate a single load, we can use that for the chain. Otherwise, 1855 // build a factor node to remember the multiple loads are independent and 1856 // chain to that. 1857 SDValue NewChain; 1858 if (LdChain.size() == 1) 1859 NewChain = LdChain[0]; 1860 else 1861 NewChain = DAG.getNode(ISD::TokenFactor, LD->getDebugLoc(), MVT::Other, 1862 &LdChain[0], LdChain.size()); 1863 1864 // Modified the chain - switch anything that used the old chain to use 1865 // the new one. 1866 ReplaceValueWith(SDValue(N, 1), NewChain); 1867 1868 return Result; 1869} 1870 1871SDValue DAGTypeLegalizer::WidenVecRes_SCALAR_TO_VECTOR(SDNode *N) { 1872 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); 1873 return DAG.getNode(ISD::SCALAR_TO_VECTOR, N->getDebugLoc(), 1874 WidenVT, N->getOperand(0)); 1875} 1876 1877SDValue DAGTypeLegalizer::WidenVecRes_SELECT(SDNode *N) { 1878 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); 1879 unsigned WidenNumElts = WidenVT.getVectorNumElements(); 1880 1881 SDValue Cond1 = N->getOperand(0); 1882 EVT CondVT = Cond1.getValueType(); 1883 if (CondVT.isVector()) { 1884 EVT CondEltVT = CondVT.getVectorElementType(); 1885 EVT CondWidenVT = EVT::getVectorVT(*DAG.getContext(), 1886 CondEltVT, WidenNumElts); 1887 if (getTypeAction(CondVT) == WidenVector) 1888 Cond1 = GetWidenedVector(Cond1); 1889 1890 if (Cond1.getValueType() != CondWidenVT) 1891 Cond1 = ModifyToType(Cond1, CondWidenVT); 1892 } 1893 1894 SDValue InOp1 = GetWidenedVector(N->getOperand(1)); 1895 SDValue InOp2 = GetWidenedVector(N->getOperand(2)); 1896 assert(InOp1.getValueType() == WidenVT && InOp2.getValueType() == WidenVT); 1897 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), 1898 WidenVT, Cond1, InOp1, InOp2); 1899} 1900 1901SDValue DAGTypeLegalizer::WidenVecRes_SELECT_CC(SDNode *N) { 1902 SDValue InOp1 = GetWidenedVector(N->getOperand(2)); 1903 SDValue InOp2 = GetWidenedVector(N->getOperand(3)); 1904 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), 1905 InOp1.getValueType(), N->getOperand(0), 1906 N->getOperand(1), InOp1, InOp2, N->getOperand(4)); 1907} 1908 1909SDValue DAGTypeLegalizer::WidenVecRes_SETCC(SDNode *N) { 1910 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); 1911 SDValue InOp1 = GetWidenedVector(N->getOperand(0)); 1912 SDValue InOp2 = GetWidenedVector(N->getOperand(1)); 1913 return DAG.getNode(ISD::SETCC, N->getDebugLoc(), WidenVT, 1914 InOp1, InOp2, N->getOperand(2)); 1915} 1916 1917SDValue DAGTypeLegalizer::WidenVecRes_UNDEF(SDNode *N) { 1918 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); 1919 return DAG.getUNDEF(WidenVT); 1920} 1921 1922SDValue DAGTypeLegalizer::WidenVecRes_VECTOR_SHUFFLE(ShuffleVectorSDNode *N) { 1923 EVT VT = N->getValueType(0); 1924 DebugLoc dl = N->getDebugLoc(); 1925 1926 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT); 1927 unsigned NumElts = VT.getVectorNumElements(); 1928 unsigned WidenNumElts = WidenVT.getVectorNumElements(); 1929 1930 SDValue InOp1 = GetWidenedVector(N->getOperand(0)); 1931 SDValue InOp2 = GetWidenedVector(N->getOperand(1)); 1932 1933 // Adjust mask based on new input vector length. 1934 SmallVector<int, 16> NewMask; 1935 for (unsigned i = 0; i != NumElts; ++i) { 1936 int Idx = N->getMaskElt(i); 1937 if (Idx < (int)NumElts) 1938 NewMask.push_back(Idx); 1939 else 1940 NewMask.push_back(Idx - NumElts + WidenNumElts); 1941 } 1942 for (unsigned i = NumElts; i != WidenNumElts; ++i) 1943 NewMask.push_back(-1); 1944 return DAG.getVectorShuffle(WidenVT, dl, InOp1, InOp2, &NewMask[0]); 1945} 1946 1947SDValue DAGTypeLegalizer::WidenVecRes_VSETCC(SDNode *N) { 1948 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); 1949 unsigned WidenNumElts = WidenVT.getVectorNumElements(); 1950 1951 SDValue InOp1 = N->getOperand(0); 1952 EVT InVT = InOp1.getValueType(); 1953 assert(InVT.isVector() && "can not widen non vector type"); 1954 EVT WidenInVT = EVT::getVectorVT(*DAG.getContext(), 1955 InVT.getVectorElementType(), WidenNumElts); 1956 InOp1 = GetWidenedVector(InOp1); 1957 SDValue InOp2 = GetWidenedVector(N->getOperand(1)); 1958 1959 // Assume that the input and output will be widen appropriately. If not, 1960 // we will have to unroll it at some point. 1961 assert(InOp1.getValueType() == WidenInVT && 1962 InOp2.getValueType() == WidenInVT && 1963 "Input not widened to expected type!"); 1964 return DAG.getNode(ISD::VSETCC, N->getDebugLoc(), 1965 WidenVT, InOp1, InOp2, N->getOperand(2)); 1966} 1967 1968 1969//===----------------------------------------------------------------------===// 1970// Widen Vector Operand 1971//===----------------------------------------------------------------------===// 1972bool DAGTypeLegalizer::WidenVectorOperand(SDNode *N, unsigned ResNo) { 1973 DEBUG(dbgs() << "Widen node operand " << ResNo << ": "; 1974 N->dump(&DAG); 1975 dbgs() << "\n"); 1976 SDValue Res = SDValue(); 1977 1978 switch (N->getOpcode()) { 1979 default: 1980#ifndef NDEBUG 1981 dbgs() << "WidenVectorOperand op #" << ResNo << ": "; 1982 N->dump(&DAG); 1983 dbgs() << "\n"; 1984#endif 1985 llvm_unreachable("Do not know how to widen this operator's operand!"); 1986 1987 case ISD::BIT_CONVERT: Res = WidenVecOp_BIT_CONVERT(N); break; 1988 case ISD::CONCAT_VECTORS: Res = WidenVecOp_CONCAT_VECTORS(N); break; 1989 case ISD::EXTRACT_SUBVECTOR: Res = WidenVecOp_EXTRACT_SUBVECTOR(N); break; 1990 case ISD::EXTRACT_VECTOR_ELT: Res = WidenVecOp_EXTRACT_VECTOR_ELT(N); break; 1991 case ISD::STORE: Res = WidenVecOp_STORE(N); break; 1992 1993 case ISD::FP_ROUND: 1994 case ISD::FP_TO_SINT: 1995 case ISD::FP_TO_UINT: 1996 case ISD::SINT_TO_FP: 1997 case ISD::UINT_TO_FP: 1998 case ISD::TRUNCATE: 1999 case ISD::SIGN_EXTEND: 2000 case ISD::ZERO_EXTEND: 2001 case ISD::ANY_EXTEND: 2002 Res = WidenVecOp_Convert(N); 2003 break; 2004 } 2005 2006 // If Res is null, the sub-method took care of registering the result. 2007 if (!Res.getNode()) return false; 2008 2009 // If the result is N, the sub-method updated N in place. Tell the legalizer 2010 // core about this. 2011 if (Res.getNode() == N) 2012 return true; 2013 2014 2015 assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 && 2016 "Invalid operand expansion"); 2017 2018 ReplaceValueWith(SDValue(N, 0), Res); 2019 return false; 2020} 2021 2022SDValue DAGTypeLegalizer::WidenVecOp_Convert(SDNode *N) { 2023 // Since the result is legal and the input is illegal, it is unlikely 2024 // that we can fix the input to a legal type so unroll the convert 2025 // into some scalar code and create a nasty build vector. 2026 EVT VT = N->getValueType(0); 2027 EVT EltVT = VT.getVectorElementType(); 2028 DebugLoc dl = N->getDebugLoc(); 2029 unsigned NumElts = VT.getVectorNumElements(); 2030 SDValue InOp = N->getOperand(0); 2031 if (getTypeAction(InOp.getValueType()) == WidenVector) 2032 InOp = GetWidenedVector(InOp); 2033 EVT InVT = InOp.getValueType(); 2034 EVT InEltVT = InVT.getVectorElementType(); 2035 2036 unsigned Opcode = N->getOpcode(); 2037 SmallVector<SDValue, 16> Ops(NumElts); 2038 for (unsigned i=0; i < NumElts; ++i) 2039 Ops[i] = DAG.getNode(Opcode, dl, EltVT, 2040 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, InEltVT, InOp, 2041 DAG.getIntPtrConstant(i))); 2042 2043 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], NumElts); 2044} 2045 2046SDValue DAGTypeLegalizer::WidenVecOp_BIT_CONVERT(SDNode *N) { 2047 EVT VT = N->getValueType(0); 2048 SDValue InOp = GetWidenedVector(N->getOperand(0)); 2049 EVT InWidenVT = InOp.getValueType(); 2050 DebugLoc dl = N->getDebugLoc(); 2051 2052 // Check if we can convert between two legal vector types and extract. 2053 unsigned InWidenSize = InWidenVT.getSizeInBits(); 2054 unsigned Size = VT.getSizeInBits(); 2055 if (InWidenSize % Size == 0 && !VT.isVector()) { 2056 unsigned NewNumElts = InWidenSize / Size; 2057 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), VT, NewNumElts); 2058 if (TLI.isTypeSynthesizable(NewVT)) { 2059 SDValue BitOp = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, InOp); 2060 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, BitOp, 2061 DAG.getIntPtrConstant(0)); 2062 } 2063 } 2064 2065 return CreateStackStoreLoad(InOp, VT); 2066} 2067 2068SDValue DAGTypeLegalizer::WidenVecOp_CONCAT_VECTORS(SDNode *N) { 2069 // If the input vector is not legal, it is likely that we will not find a 2070 // legal vector of the same size. Replace the concatenate vector with a 2071 // nasty build vector. 2072 EVT VT = N->getValueType(0); 2073 EVT EltVT = VT.getVectorElementType(); 2074 DebugLoc dl = N->getDebugLoc(); 2075 unsigned NumElts = VT.getVectorNumElements(); 2076 SmallVector<SDValue, 16> Ops(NumElts); 2077 2078 EVT InVT = N->getOperand(0).getValueType(); 2079 unsigned NumInElts = InVT.getVectorNumElements(); 2080 2081 unsigned Idx = 0; 2082 unsigned NumOperands = N->getNumOperands(); 2083 for (unsigned i=0; i < NumOperands; ++i) { 2084 SDValue InOp = N->getOperand(i); 2085 if (getTypeAction(InOp.getValueType()) == WidenVector) 2086 InOp = GetWidenedVector(InOp); 2087 for (unsigned j=0; j < NumInElts; ++j) 2088 Ops[Idx++] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp, 2089 DAG.getIntPtrConstant(j)); 2090 } 2091 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], NumElts); 2092} 2093 2094SDValue DAGTypeLegalizer::WidenVecOp_EXTRACT_SUBVECTOR(SDNode *N) { 2095 SDValue InOp = GetWidenedVector(N->getOperand(0)); 2096 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, N->getDebugLoc(), 2097 N->getValueType(0), InOp, N->getOperand(1)); 2098} 2099 2100SDValue DAGTypeLegalizer::WidenVecOp_EXTRACT_VECTOR_ELT(SDNode *N) { 2101 SDValue InOp = GetWidenedVector(N->getOperand(0)); 2102 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, N->getDebugLoc(), 2103 N->getValueType(0), InOp, N->getOperand(1)); 2104} 2105 2106SDValue DAGTypeLegalizer::WidenVecOp_STORE(SDNode *N) { 2107 // We have to widen the value but we want only to store the original 2108 // vector type. 2109 StoreSDNode *ST = cast<StoreSDNode>(N); 2110 2111 SmallVector<SDValue, 16> StChain; 2112 if (ST->isTruncatingStore()) 2113 GenWidenVectorTruncStores(StChain, ST); 2114 else 2115 GenWidenVectorStores(StChain, ST); 2116 2117 if (StChain.size() == 1) 2118 return StChain[0]; 2119 else 2120 return DAG.getNode(ISD::TokenFactor, ST->getDebugLoc(), 2121 MVT::Other,&StChain[0],StChain.size()); 2122} 2123 2124//===----------------------------------------------------------------------===// 2125// Vector Widening Utilities 2126//===----------------------------------------------------------------------===// 2127 2128// Utility function to find the type to chop up a widen vector for load/store 2129// TLI: Target lowering used to determine legal types. 2130// Width: Width left need to load/store. 2131// WidenVT: The widen vector type to load to/store from 2132// Align: If 0, don't allow use of a wider type 2133// WidenEx: If Align is not 0, the amount additional we can load/store from. 2134 2135static EVT FindMemType(SelectionDAG& DAG, const TargetLowering &TLI, 2136 unsigned Width, EVT WidenVT, 2137 unsigned Align = 0, unsigned WidenEx = 0) { 2138 EVT WidenEltVT = WidenVT.getVectorElementType(); 2139 unsigned WidenWidth = WidenVT.getSizeInBits(); 2140 unsigned WidenEltWidth = WidenEltVT.getSizeInBits(); 2141 unsigned AlignInBits = Align*8; 2142 2143 // If we have one element to load/store, return it. 2144 EVT RetVT = WidenEltVT; 2145 if (Width == WidenEltWidth) 2146 return RetVT; 2147 2148 // See if there is larger legal integer than the element type to load/store 2149 unsigned VT; 2150 for (VT = (unsigned)MVT::LAST_INTEGER_VALUETYPE; 2151 VT >= (unsigned)MVT::FIRST_INTEGER_VALUETYPE; --VT) { 2152 EVT MemVT((MVT::SimpleValueType) VT); 2153 unsigned MemVTWidth = MemVT.getSizeInBits(); 2154 if (MemVT.getSizeInBits() <= WidenEltWidth) 2155 break; 2156 if (TLI.isTypeSynthesizable(MemVT) && (WidenWidth % MemVTWidth) == 0 && 2157 (MemVTWidth <= Width || 2158 (Align!=0 && MemVTWidth<=AlignInBits && MemVTWidth<=Width+WidenEx))) { 2159 RetVT = MemVT; 2160 break; 2161 } 2162 } 2163 2164 // See if there is a larger vector type to load/store that has the same vector 2165 // element type and is evenly divisible with the WidenVT. 2166 for (VT = (unsigned)MVT::LAST_VECTOR_VALUETYPE; 2167 VT >= (unsigned)MVT::FIRST_VECTOR_VALUETYPE; --VT) { 2168 EVT MemVT = (MVT::SimpleValueType) VT; 2169 unsigned MemVTWidth = MemVT.getSizeInBits(); 2170 if (TLI.isTypeSynthesizable(MemVT) && WidenEltVT == MemVT.getVectorElementType() && 2171 (WidenWidth % MemVTWidth) == 0 && 2172 (MemVTWidth <= Width || 2173 (Align!=0 && MemVTWidth<=AlignInBits && MemVTWidth<=Width+WidenEx))) { 2174 if (RetVT.getSizeInBits() < MemVTWidth || MemVT == WidenVT) 2175 return MemVT; 2176 } 2177 } 2178 2179 return RetVT; 2180} 2181 2182// Builds a vector type from scalar loads 2183// VecTy: Resulting Vector type 2184// LDOps: Load operators to build a vector type 2185// [Start,End) the list of loads to use. 2186static SDValue BuildVectorFromScalar(SelectionDAG& DAG, EVT VecTy, 2187 SmallVector<SDValue, 16>& LdOps, 2188 unsigned Start, unsigned End) { 2189 DebugLoc dl = LdOps[Start].getDebugLoc(); 2190 EVT LdTy = LdOps[Start].getValueType(); 2191 unsigned Width = VecTy.getSizeInBits(); 2192 unsigned NumElts = Width / LdTy.getSizeInBits(); 2193 EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), LdTy, NumElts); 2194 2195 unsigned Idx = 1; 2196 SDValue VecOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NewVecVT,LdOps[Start]); 2197 2198 for (unsigned i = Start + 1; i != End; ++i) { 2199 EVT NewLdTy = LdOps[i].getValueType(); 2200 if (NewLdTy != LdTy) { 2201 NumElts = Width / NewLdTy.getSizeInBits(); 2202 NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewLdTy, NumElts); 2203 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, NewVecVT, VecOp); 2204 // Readjust position and vector position based on new load type 2205 Idx = Idx * LdTy.getSizeInBits() / NewLdTy.getSizeInBits(); 2206 LdTy = NewLdTy; 2207 } 2208 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, NewVecVT, VecOp, LdOps[i], 2209 DAG.getIntPtrConstant(Idx++)); 2210 } 2211 return DAG.getNode(ISD::BIT_CONVERT, dl, VecTy, VecOp); 2212} 2213 2214SDValue DAGTypeLegalizer::GenWidenVectorLoads(SmallVector<SDValue, 16> &LdChain, 2215 LoadSDNode *LD) { 2216 // The strategy assumes that we can efficiently load powers of two widths. 2217 // The routines chops the vector into the largest vector loads with the same 2218 // element type or scalar loads and then recombines it to the widen vector 2219 // type. 2220 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(),LD->getValueType(0)); 2221 unsigned WidenWidth = WidenVT.getSizeInBits(); 2222 EVT LdVT = LD->getMemoryVT(); 2223 DebugLoc dl = LD->getDebugLoc(); 2224 assert(LdVT.isVector() && WidenVT.isVector()); 2225 assert(LdVT.getVectorElementType() == WidenVT.getVectorElementType()); 2226 2227 // Load information 2228 SDValue Chain = LD->getChain(); 2229 SDValue BasePtr = LD->getBasePtr(); 2230 unsigned Align = LD->getAlignment(); 2231 bool isVolatile = LD->isVolatile(); 2232 bool isNonTemporal = LD->isNonTemporal(); 2233 2234 int LdWidth = LdVT.getSizeInBits(); 2235 int WidthDiff = WidenWidth - LdWidth; // Difference 2236 unsigned LdAlign = (isVolatile) ? 0 : Align; // Allow wider loads 2237 2238 // Find the vector type that can load from. 2239 EVT NewVT = FindMemType(DAG, TLI, LdWidth, WidenVT, LdAlign, WidthDiff); 2240 int NewVTWidth = NewVT.getSizeInBits(); 2241 SDValue LdOp = DAG.getLoad(NewVT, dl, Chain, BasePtr, LD->getPointerInfo(), 2242 isVolatile, isNonTemporal, Align); 2243 LdChain.push_back(LdOp.getValue(1)); 2244 2245 // Check if we can load the element with one instruction 2246 if (LdWidth <= NewVTWidth) { 2247 if (!NewVT.isVector()) { 2248 unsigned NumElts = WidenWidth / NewVTWidth; 2249 EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewVT, NumElts); 2250 SDValue VecOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NewVecVT, LdOp); 2251 return DAG.getNode(ISD::BIT_CONVERT, dl, WidenVT, VecOp); 2252 } 2253 if (NewVT == WidenVT) 2254 return LdOp; 2255 2256 assert(WidenWidth % NewVTWidth == 0); 2257 unsigned NumConcat = WidenWidth / NewVTWidth; 2258 SmallVector<SDValue, 16> ConcatOps(NumConcat); 2259 SDValue UndefVal = DAG.getUNDEF(NewVT); 2260 ConcatOps[0] = LdOp; 2261 for (unsigned i = 1; i != NumConcat; ++i) 2262 ConcatOps[i] = UndefVal; 2263 return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &ConcatOps[0], 2264 NumConcat); 2265 } 2266 2267 // Load vector by using multiple loads from largest vector to scalar 2268 SmallVector<SDValue, 16> LdOps; 2269 LdOps.push_back(LdOp); 2270 2271 LdWidth -= NewVTWidth; 2272 unsigned Offset = 0; 2273 2274 while (LdWidth > 0) { 2275 unsigned Increment = NewVTWidth / 8; 2276 Offset += Increment; 2277 BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, 2278 DAG.getIntPtrConstant(Increment)); 2279 2280 if (LdWidth < NewVTWidth) { 2281 // Our current type we are using is too large, find a better size 2282 NewVT = FindMemType(DAG, TLI, LdWidth, WidenVT, LdAlign, WidthDiff); 2283 NewVTWidth = NewVT.getSizeInBits(); 2284 } 2285 2286 SDValue LdOp = DAG.getLoad(NewVT, dl, Chain, BasePtr, 2287 LD->getPointerInfo().getWithOffset(Offset), 2288 isVolatile, 2289 isNonTemporal, MinAlign(Align, Increment)); 2290 LdChain.push_back(LdOp.getValue(1)); 2291 LdOps.push_back(LdOp); 2292 2293 LdWidth -= NewVTWidth; 2294 } 2295 2296 // Build the vector from the loads operations 2297 unsigned End = LdOps.size(); 2298 if (!LdOps[0].getValueType().isVector()) 2299 // All the loads are scalar loads. 2300 return BuildVectorFromScalar(DAG, WidenVT, LdOps, 0, End); 2301 2302 // If the load contains vectors, build the vector using concat vector. 2303 // All of the vectors used to loads are power of 2 and the scalars load 2304 // can be combined to make a power of 2 vector. 2305 SmallVector<SDValue, 16> ConcatOps(End); 2306 int i = End - 1; 2307 int Idx = End; 2308 EVT LdTy = LdOps[i].getValueType(); 2309 // First combine the scalar loads to a vector 2310 if (!LdTy.isVector()) { 2311 for (--i; i >= 0; --i) { 2312 LdTy = LdOps[i].getValueType(); 2313 if (LdTy.isVector()) 2314 break; 2315 } 2316 ConcatOps[--Idx] = BuildVectorFromScalar(DAG, LdTy, LdOps, i+1, End); 2317 } 2318 ConcatOps[--Idx] = LdOps[i]; 2319 for (--i; i >= 0; --i) { 2320 EVT NewLdTy = LdOps[i].getValueType(); 2321 if (NewLdTy != LdTy) { 2322 // Create a larger vector 2323 ConcatOps[End-1] = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewLdTy, 2324 &ConcatOps[Idx], End - Idx); 2325 Idx = End - 1; 2326 LdTy = NewLdTy; 2327 } 2328 ConcatOps[--Idx] = LdOps[i]; 2329 } 2330 2331 if (WidenWidth == LdTy.getSizeInBits()*(End - Idx)) 2332 return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, 2333 &ConcatOps[Idx], End - Idx); 2334 2335 // We need to fill the rest with undefs to build the vector 2336 unsigned NumOps = WidenWidth / LdTy.getSizeInBits(); 2337 SmallVector<SDValue, 16> WidenOps(NumOps); 2338 SDValue UndefVal = DAG.getUNDEF(LdTy); 2339 { 2340 unsigned i = 0; 2341 for (; i != End-Idx; ++i) 2342 WidenOps[i] = ConcatOps[Idx+i]; 2343 for (; i != NumOps; ++i) 2344 WidenOps[i] = UndefVal; 2345 } 2346 return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &WidenOps[0],NumOps); 2347} 2348 2349SDValue 2350DAGTypeLegalizer::GenWidenVectorExtLoads(SmallVector<SDValue, 16>& LdChain, 2351 LoadSDNode * LD, 2352 ISD::LoadExtType ExtType) { 2353 // For extension loads, it may not be more efficient to chop up the vector 2354 // and then extended it. Instead, we unroll the load and build a new vector. 2355 EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(),LD->getValueType(0)); 2356 EVT LdVT = LD->getMemoryVT(); 2357 DebugLoc dl = LD->getDebugLoc(); 2358 assert(LdVT.isVector() && WidenVT.isVector()); 2359 2360 // Load information 2361 SDValue Chain = LD->getChain(); 2362 SDValue BasePtr = LD->getBasePtr(); 2363 unsigned Align = LD->getAlignment(); 2364 bool isVolatile = LD->isVolatile(); 2365 bool isNonTemporal = LD->isNonTemporal(); 2366 2367 EVT EltVT = WidenVT.getVectorElementType(); 2368 EVT LdEltVT = LdVT.getVectorElementType(); 2369 unsigned NumElts = LdVT.getVectorNumElements(); 2370 2371 // Load each element and widen 2372 unsigned WidenNumElts = WidenVT.getVectorNumElements(); 2373 SmallVector<SDValue, 16> Ops(WidenNumElts); 2374 unsigned Increment = LdEltVT.getSizeInBits() / 8; 2375 Ops[0] = DAG.getExtLoad(ExtType, EltVT, dl, Chain, BasePtr, 2376 LD->getPointerInfo(), 2377 LdEltVT, isVolatile, isNonTemporal, Align); 2378 LdChain.push_back(Ops[0].getValue(1)); 2379 unsigned i = 0, Offset = Increment; 2380 for (i=1; i < NumElts; ++i, Offset += Increment) { 2381 SDValue NewBasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), 2382 BasePtr, DAG.getIntPtrConstant(Offset)); 2383 Ops[i] = DAG.getExtLoad(ExtType, EltVT, dl, Chain, NewBasePtr, 2384 LD->getPointerInfo().getWithOffset(Offset), LdEltVT, 2385 isVolatile, isNonTemporal, Align); 2386 LdChain.push_back(Ops[i].getValue(1)); 2387 } 2388 2389 // Fill the rest with undefs 2390 SDValue UndefVal = DAG.getUNDEF(EltVT); 2391 for (; i != WidenNumElts; ++i) 2392 Ops[i] = UndefVal; 2393 2394 return DAG.getNode(ISD::BUILD_VECTOR, dl, WidenVT, &Ops[0], Ops.size()); 2395} 2396 2397 2398void DAGTypeLegalizer::GenWidenVectorStores(SmallVector<SDValue, 16>& StChain, 2399 StoreSDNode *ST) { 2400 // The strategy assumes that we can efficiently store powers of two widths. 2401 // The routines chops the vector into the largest vector stores with the same 2402 // element type or scalar stores. 2403 SDValue Chain = ST->getChain(); 2404 SDValue BasePtr = ST->getBasePtr(); 2405 unsigned Align = ST->getAlignment(); 2406 bool isVolatile = ST->isVolatile(); 2407 bool isNonTemporal = ST->isNonTemporal(); 2408 SDValue ValOp = GetWidenedVector(ST->getValue()); 2409 DebugLoc dl = ST->getDebugLoc(); 2410 2411 EVT StVT = ST->getMemoryVT(); 2412 unsigned StWidth = StVT.getSizeInBits(); 2413 EVT ValVT = ValOp.getValueType(); 2414 unsigned ValWidth = ValVT.getSizeInBits(); 2415 EVT ValEltVT = ValVT.getVectorElementType(); 2416 unsigned ValEltWidth = ValEltVT.getSizeInBits(); 2417 assert(StVT.getVectorElementType() == ValEltVT); 2418 2419 int Idx = 0; // current index to store 2420 unsigned Offset = 0; // offset from base to store 2421 while (StWidth != 0) { 2422 // Find the largest vector type we can store with 2423 EVT NewVT = FindMemType(DAG, TLI, StWidth, ValVT); 2424 unsigned NewVTWidth = NewVT.getSizeInBits(); 2425 unsigned Increment = NewVTWidth / 8; 2426 if (NewVT.isVector()) { 2427 unsigned NumVTElts = NewVT.getVectorNumElements(); 2428 do { 2429 SDValue EOp = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NewVT, ValOp, 2430 DAG.getIntPtrConstant(Idx)); 2431 StChain.push_back(DAG.getStore(Chain, dl, EOp, BasePtr, 2432 ST->getPointerInfo().getWithOffset(Offset), 2433 isVolatile, isNonTemporal, 2434 MinAlign(Align, Offset))); 2435 StWidth -= NewVTWidth; 2436 Offset += Increment; 2437 Idx += NumVTElts; 2438 BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, 2439 DAG.getIntPtrConstant(Increment)); 2440 } while (StWidth != 0 && StWidth >= NewVTWidth); 2441 } else { 2442 // Cast the vector to the scalar type we can store 2443 unsigned NumElts = ValWidth / NewVTWidth; 2444 EVT NewVecVT = EVT::getVectorVT(*DAG.getContext(), NewVT, NumElts); 2445 SDValue VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, NewVecVT, ValOp); 2446 // Readjust index position based on new vector type 2447 Idx = Idx * ValEltWidth / NewVTWidth; 2448 do { 2449 SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NewVT, VecOp, 2450 DAG.getIntPtrConstant(Idx++)); 2451 StChain.push_back(DAG.getStore(Chain, dl, EOp, BasePtr, 2452 ST->getPointerInfo().getWithOffset(Offset), 2453 isVolatile, isNonTemporal, 2454 MinAlign(Align, Offset))); 2455 StWidth -= NewVTWidth; 2456 Offset += Increment; 2457 BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, 2458 DAG.getIntPtrConstant(Increment)); 2459 } while (StWidth != 0 && StWidth >= NewVTWidth); 2460 // Restore index back to be relative to the original widen element type 2461 Idx = Idx * NewVTWidth / ValEltWidth; 2462 } 2463 } 2464} 2465 2466void 2467DAGTypeLegalizer::GenWidenVectorTruncStores(SmallVector<SDValue, 16>& StChain, 2468 StoreSDNode *ST) { 2469 // For extension loads, it may not be more efficient to truncate the vector 2470 // and then store it. Instead, we extract each element and then store it. 2471 SDValue Chain = ST->getChain(); 2472 SDValue BasePtr = ST->getBasePtr(); 2473 unsigned Align = ST->getAlignment(); 2474 bool isVolatile = ST->isVolatile(); 2475 bool isNonTemporal = ST->isNonTemporal(); 2476 SDValue ValOp = GetWidenedVector(ST->getValue()); 2477 DebugLoc dl = ST->getDebugLoc(); 2478 2479 EVT StVT = ST->getMemoryVT(); 2480 EVT ValVT = ValOp.getValueType(); 2481 2482 // It must be true that we the widen vector type is bigger than where 2483 // we need to store. 2484 assert(StVT.isVector() && ValOp.getValueType().isVector()); 2485 assert(StVT.bitsLT(ValOp.getValueType())); 2486 2487 // For truncating stores, we can not play the tricks of chopping legal 2488 // vector types and bit cast it to the right type. Instead, we unroll 2489 // the store. 2490 EVT StEltVT = StVT.getVectorElementType(); 2491 EVT ValEltVT = ValVT.getVectorElementType(); 2492 unsigned Increment = ValEltVT.getSizeInBits() / 8; 2493 unsigned NumElts = StVT.getVectorNumElements(); 2494 SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ValEltVT, ValOp, 2495 DAG.getIntPtrConstant(0)); 2496 StChain.push_back(DAG.getTruncStore(Chain, dl, EOp, BasePtr, 2497 ST->getPointerInfo(), StEltVT, 2498 isVolatile, isNonTemporal, Align)); 2499 unsigned Offset = Increment; 2500 for (unsigned i=1; i < NumElts; ++i, Offset += Increment) { 2501 SDValue NewBasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), 2502 BasePtr, DAG.getIntPtrConstant(Offset)); 2503 SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ValEltVT, ValOp, 2504 DAG.getIntPtrConstant(0)); 2505 StChain.push_back(DAG.getTruncStore(Chain, dl, EOp, NewBasePtr, 2506 ST->getPointerInfo().getWithOffset(Offset), 2507 StEltVT, isVolatile, isNonTemporal, 2508 MinAlign(Align, Offset))); 2509 } 2510} 2511 2512/// Modifies a vector input (widen or narrows) to a vector of NVT. The 2513/// input vector must have the same element type as NVT. 2514SDValue DAGTypeLegalizer::ModifyToType(SDValue InOp, EVT NVT) { 2515 // Note that InOp might have been widened so it might already have 2516 // the right width or it might need be narrowed. 2517 EVT InVT = InOp.getValueType(); 2518 assert(InVT.getVectorElementType() == NVT.getVectorElementType() && 2519 "input and widen element type must match"); 2520 DebugLoc dl = InOp.getDebugLoc(); 2521 2522 // Check if InOp already has the right width. 2523 if (InVT == NVT) 2524 return InOp; 2525 2526 unsigned InNumElts = InVT.getVectorNumElements(); 2527 unsigned WidenNumElts = NVT.getVectorNumElements(); 2528 if (WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0) { 2529 unsigned NumConcat = WidenNumElts / InNumElts; 2530 SmallVector<SDValue, 16> Ops(NumConcat); 2531 SDValue UndefVal = DAG.getUNDEF(InVT); 2532 Ops[0] = InOp; 2533 for (unsigned i = 1; i != NumConcat; ++i) 2534 Ops[i] = UndefVal; 2535 2536 return DAG.getNode(ISD::CONCAT_VECTORS, dl, NVT, &Ops[0], NumConcat); 2537 } 2538 2539 if (WidenNumElts < InNumElts && InNumElts % WidenNumElts) 2540 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT, InOp, 2541 DAG.getIntPtrConstant(0)); 2542 2543 // Fall back to extract and build. 2544 SmallVector<SDValue, 16> Ops(WidenNumElts); 2545 EVT EltVT = NVT.getVectorElementType(); 2546 unsigned MinNumElts = std::min(WidenNumElts, InNumElts); 2547 unsigned Idx; 2548 for (Idx = 0; Idx < MinNumElts; ++Idx) 2549 Ops[Idx] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InOp, 2550 DAG.getIntPtrConstant(Idx)); 2551 2552 SDValue UndefVal = DAG.getUNDEF(EltVT); 2553 for ( ; Idx < WidenNumElts; ++Idx) 2554 Ops[Idx] = UndefVal; 2555 return DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, &Ops[0], WidenNumElts); 2556} 2557