| 1 | //===------- LegalizeVectorTypes.cpp - Legalization of vector types -------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file performs vector type splitting and scalarization for LegalizeTypes. |
| 10 | // Scalarization is the act of changing a computation in an illegal one-element |
| 11 | // vector type to be a computation in its scalar element type. For example, |
| 12 | // implementing <1 x f32> arithmetic in a scalar f32 register. This is needed |
| 13 | // as a base case when scalarizing vector arithmetic like <4 x f32>, which |
| 14 | // eventually decomposes to scalars if the target doesn't support v4f32 or v2f32 |
| 15 | // types. |
| 16 | // Splitting is the act of changing a computation in an invalid vector type to |
| 17 | // be a computation in two vectors of half the size. For example, implementing |
| 18 | // <128 x f32> operations in terms of two <64 x f32> operations. |
| 19 | // |
| 20 | //===----------------------------------------------------------------------===// |
| 21 | |
| 22 | #include "LegalizeTypes.h" |
| 23 | #include "llvm/ADT/SmallBitVector.h" |
| 24 | #include "llvm/Analysis/MemoryLocation.h" |
| 25 | #include "llvm/Analysis/VectorUtils.h" |
| 26 | #include "llvm/CodeGen/ISDOpcodes.h" |
| 27 | #include "llvm/IR/DataLayout.h" |
| 28 | #include "llvm/Support/ErrorHandling.h" |
| 29 | #include "llvm/Support/TypeSize.h" |
| 30 | #include "llvm/Support/raw_ostream.h" |
| 31 | #include <numeric> |
| 32 | |
| 33 | using namespace llvm; |
| 34 | |
| 35 | #define DEBUG_TYPE "legalize-types" |
| 36 | |
| 37 | //===----------------------------------------------------------------------===// |
| 38 | // Result Vector Scalarization: <1 x ty> -> ty. |
| 39 | //===----------------------------------------------------------------------===// |
| 40 | |
| 41 | void DAGTypeLegalizer::ScalarizeVectorResult(SDNode *N, unsigned ResNo) { |
| 42 | LLVM_DEBUG(dbgs() << "Scalarize node result " << ResNo << ": " ; |
| 43 | N->dump(&DAG)); |
| 44 | SDValue R = SDValue(); |
| 45 | |
| 46 | // See if the target wants to custom expand this node. |
| 47 | if (CustomLowerNode(N, VT: N->getValueType(ResNo), LegalizeResult: true)) |
| 48 | return; |
| 49 | |
| 50 | switch (N->getOpcode()) { |
| 51 | default: |
| 52 | #ifndef NDEBUG |
| 53 | dbgs() << "ScalarizeVectorResult #" << ResNo << ": " ; |
| 54 | N->dump(&DAG); |
| 55 | dbgs() << "\n" ; |
| 56 | #endif |
| 57 | report_fatal_error(reason: "Do not know how to scalarize the result of this " |
| 58 | "operator!\n" ); |
| 59 | |
| 60 | case ISD::LOOP_DEPENDENCE_WAR_MASK: |
| 61 | case ISD::LOOP_DEPENDENCE_RAW_MASK: |
| 62 | R = ScalarizeVecRes_LOOP_DEPENDENCE_MASK(N); |
| 63 | break; |
| 64 | case ISD::MERGE_VALUES: R = ScalarizeVecRes_MERGE_VALUES(N, ResNo);break; |
| 65 | case ISD::BITCAST: R = ScalarizeVecRes_BITCAST(N); break; |
| 66 | case ISD::BUILD_VECTOR: R = ScalarizeVecRes_BUILD_VECTOR(N); break; |
| 67 | case ISD::EXTRACT_SUBVECTOR: R = ScalarizeVecRes_EXTRACT_SUBVECTOR(N); break; |
| 68 | case ISD::FP_ROUND: R = ScalarizeVecRes_FP_ROUND(N); break; |
| 69 | case ISD::CONVERT_FROM_ARBITRARY_FP: |
| 70 | R = ScalarizeVecRes_CONVERT_FROM_ARBITRARY_FP(N); |
| 71 | break; |
| 72 | case ISD::CONVERT_TO_ARBITRARY_FP: |
| 73 | R = ScalarizeVecRes_CONVERT_TO_ARBITRARY_FP(N); |
| 74 | break; |
| 75 | case ISD::AssertZext: |
| 76 | case ISD::AssertSext: |
| 77 | case ISD::FPOWI: |
| 78 | case ISD::AssertNoFPClass: |
| 79 | R = ScalarizeVecRes_UnaryOpWithExtraInput(N); |
| 80 | break; |
| 81 | case ISD::INSERT_VECTOR_ELT: R = ScalarizeVecRes_INSERT_VECTOR_ELT(N); break; |
| 82 | case ISD::ATOMIC_LOAD: |
| 83 | R = ScalarizeVecRes_ATOMIC_LOAD(N: cast<AtomicSDNode>(Val: N)); |
| 84 | break; |
| 85 | case ISD::LOAD: R = ScalarizeVecRes_LOAD(N: cast<LoadSDNode>(Val: N));break; |
| 86 | case ISD::SCALAR_TO_VECTOR: R = ScalarizeVecRes_SCALAR_TO_VECTOR(N); break; |
| 87 | case ISD::SIGN_EXTEND_INREG: R = ScalarizeVecRes_InregOp(N); break; |
| 88 | case ISD::VSELECT: R = ScalarizeVecRes_VSELECT(N); break; |
| 89 | case ISD::SELECT: R = ScalarizeVecRes_SELECT(N); break; |
| 90 | case ISD::SELECT_CC: R = ScalarizeVecRes_SELECT_CC(N); break; |
| 91 | case ISD::SETCC: R = ScalarizeVecRes_SETCC(N); break; |
| 92 | case ISD::POISON: |
| 93 | case ISD::UNDEF: R = ScalarizeVecRes_UNDEF(N); break; |
| 94 | case ISD::VECTOR_SHUFFLE: R = ScalarizeVecRes_VECTOR_SHUFFLE(N); break; |
| 95 | case ISD::IS_FPCLASS: R = ScalarizeVecRes_IS_FPCLASS(N); break; |
| 96 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 97 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 98 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 99 | R = ScalarizeVecRes_VecInregOp(N); |
| 100 | break; |
| 101 | case ISD::ABS: |
| 102 | case ISD::ABS_MIN_POISON: |
| 103 | case ISD::ANY_EXTEND: |
| 104 | case ISD::BITREVERSE: |
| 105 | case ISD::BSWAP: |
| 106 | case ISD::CTLZ: |
| 107 | case ISD::CTLZ_ZERO_POISON: |
| 108 | case ISD::CTPOP: |
| 109 | case ISD::CTTZ: |
| 110 | case ISD::CTTZ_ZERO_POISON: |
| 111 | case ISD::FABS: |
| 112 | case ISD::FACOS: |
| 113 | case ISD::FASIN: |
| 114 | case ISD::FATAN: |
| 115 | case ISD::FCEIL: |
| 116 | case ISD::FCOS: |
| 117 | case ISD::FCOSH: |
| 118 | case ISD::FEXP: |
| 119 | case ISD::FEXP2: |
| 120 | case ISD::FEXP10: |
| 121 | case ISD::FFLOOR: |
| 122 | case ISD::FLOG: |
| 123 | case ISD::FLOG10: |
| 124 | case ISD::FLOG2: |
| 125 | case ISD::FNEARBYINT: |
| 126 | case ISD::FNEG: |
| 127 | case ISD::FREEZE: |
| 128 | case ISD::ARITH_FENCE: |
| 129 | case ISD::FP_EXTEND: |
| 130 | case ISD::FP_TO_SINT: |
| 131 | case ISD::FP_TO_UINT: |
| 132 | case ISD::FRINT: |
| 133 | case ISD::LRINT: |
| 134 | case ISD::LLRINT: |
| 135 | case ISD::FROUND: |
| 136 | case ISD::FROUNDEVEN: |
| 137 | case ISD::LROUND: |
| 138 | case ISD::LLROUND: |
| 139 | case ISD::FSIN: |
| 140 | case ISD::FSINH: |
| 141 | case ISD::FSQRT: |
| 142 | case ISD::FTAN: |
| 143 | case ISD::FTANH: |
| 144 | case ISD::FTRUNC: |
| 145 | case ISD::SIGN_EXTEND: |
| 146 | case ISD::SINT_TO_FP: |
| 147 | case ISD::TRUNCATE: |
| 148 | case ISD::UINT_TO_FP: |
| 149 | case ISD::ZERO_EXTEND: |
| 150 | case ISD::FCANONICALIZE: |
| 151 | R = ScalarizeVecRes_UnaryOp(N); |
| 152 | break; |
| 153 | case ISD::ADDRSPACECAST: |
| 154 | R = ScalarizeVecRes_ADDRSPACECAST(N); |
| 155 | break; |
| 156 | case ISD::FMODF: |
| 157 | case ISD::FFREXP: |
| 158 | case ISD::FSINCOS: |
| 159 | case ISD::FSINCOSPI: |
| 160 | R = ScalarizeVecRes_UnaryOpWithTwoResults(N, ResNo); |
| 161 | break; |
| 162 | case ISD::ADD: |
| 163 | case ISD::AND: |
| 164 | case ISD::AVGCEILS: |
| 165 | case ISD::AVGCEILU: |
| 166 | case ISD::AVGFLOORS: |
| 167 | case ISD::AVGFLOORU: |
| 168 | case ISD::FADD: |
| 169 | case ISD::FCOPYSIGN: |
| 170 | case ISD::FDIV: |
| 171 | case ISD::FMUL: |
| 172 | case ISD::FMINNUM: |
| 173 | case ISD::FMAXNUM: |
| 174 | case ISD::FMINNUM_IEEE: |
| 175 | case ISD::FMAXNUM_IEEE: |
| 176 | case ISD::FMINIMUM: |
| 177 | case ISD::FMAXIMUM: |
| 178 | case ISD::FMINIMUMNUM: |
| 179 | case ISD::FMAXIMUMNUM: |
| 180 | case ISD::FLDEXP: |
| 181 | case ISD::ABDS: |
| 182 | case ISD::ABDU: |
| 183 | case ISD::SMIN: |
| 184 | case ISD::SMAX: |
| 185 | case ISD::UMIN: |
| 186 | case ISD::UMAX: |
| 187 | |
| 188 | case ISD::SADDSAT: |
| 189 | case ISD::UADDSAT: |
| 190 | case ISD::SSUBSAT: |
| 191 | case ISD::USUBSAT: |
| 192 | case ISD::SSHLSAT: |
| 193 | case ISD::USHLSAT: |
| 194 | |
| 195 | case ISD::FPOW: |
| 196 | case ISD::FATAN2: |
| 197 | case ISD::FREM: |
| 198 | case ISD::FSUB: |
| 199 | case ISD::MUL: |
| 200 | case ISD::MULHS: |
| 201 | case ISD::MULHU: |
| 202 | case ISD::OR: |
| 203 | case ISD::SDIV: |
| 204 | case ISD::SREM: |
| 205 | case ISD::SUB: |
| 206 | case ISD::UDIV: |
| 207 | case ISD::UREM: |
| 208 | case ISD::XOR: |
| 209 | case ISD::SHL: |
| 210 | case ISD::SRA: |
| 211 | case ISD::SRL: |
| 212 | case ISD::ROTL: |
| 213 | case ISD::ROTR: |
| 214 | case ISD::CLMUL: |
| 215 | case ISD::CLMULR: |
| 216 | case ISD::CLMULH: |
| 217 | case ISD::PEXT: |
| 218 | case ISD::PDEP: |
| 219 | R = ScalarizeVecRes_BinOp(N); |
| 220 | break; |
| 221 | |
| 222 | case ISD::MASKED_UDIV: |
| 223 | case ISD::MASKED_SDIV: |
| 224 | case ISD::MASKED_UREM: |
| 225 | case ISD::MASKED_SREM: |
| 226 | R = ScalarizeVecRes_MaskedBinOp(N); |
| 227 | break; |
| 228 | |
| 229 | case ISD::SCMP: |
| 230 | case ISD::UCMP: |
| 231 | R = ScalarizeVecRes_CMP(N); |
| 232 | break; |
| 233 | |
| 234 | case ISD::FMA: |
| 235 | case ISD::FSHL: |
| 236 | case ISD::FSHR: |
| 237 | R = ScalarizeVecRes_TernaryOp(N); |
| 238 | break; |
| 239 | |
| 240 | #define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \ |
| 241 | case ISD::STRICT_##DAGN: |
| 242 | #include "llvm/IR/ConstrainedOps.def" |
| 243 | R = ScalarizeVecRes_StrictFPOp(N); |
| 244 | break; |
| 245 | |
| 246 | case ISD::FP_TO_UINT_SAT: |
| 247 | case ISD::FP_TO_SINT_SAT: |
| 248 | R = ScalarizeVecRes_FP_TO_XINT_SAT(N); |
| 249 | break; |
| 250 | |
| 251 | case ISD::UADDO: |
| 252 | case ISD::SADDO: |
| 253 | case ISD::USUBO: |
| 254 | case ISD::SSUBO: |
| 255 | case ISD::UMULO: |
| 256 | case ISD::SMULO: |
| 257 | R = ScalarizeVecRes_OverflowOp(N, ResNo); |
| 258 | break; |
| 259 | case ISD::SMULFIX: |
| 260 | case ISD::SMULFIXSAT: |
| 261 | case ISD::UMULFIX: |
| 262 | case ISD::UMULFIXSAT: |
| 263 | case ISD::SDIVFIX: |
| 264 | case ISD::SDIVFIXSAT: |
| 265 | case ISD::UDIVFIX: |
| 266 | case ISD::UDIVFIXSAT: |
| 267 | R = ScalarizeVecRes_FIX(N); |
| 268 | break; |
| 269 | } |
| 270 | |
| 271 | // If R is null, the sub-method took care of registering the result. |
| 272 | if (R.getNode()) |
| 273 | SetScalarizedVector(Op: SDValue(N, ResNo), Result: R); |
| 274 | } |
| 275 | |
| 276 | SDValue DAGTypeLegalizer::ScalarizeVecRes_BinOp(SDNode *N) { |
| 277 | SDValue LHS = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 278 | SDValue RHS = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 279 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), |
| 280 | VT: LHS.getValueType(), N1: LHS, N2: RHS, Flags: N->getFlags()); |
| 281 | } |
| 282 | |
| 283 | SDValue DAGTypeLegalizer::ScalarizeVecRes_MaskedBinOp(SDNode *N) { |
| 284 | SDLoc DL(N); |
| 285 | SDValue LHS = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 286 | SDValue RHS = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 287 | SDValue Mask = N->getOperand(Num: 2); |
| 288 | EVT MaskVT = Mask.getValueType(); |
| 289 | // The vselect result and input vectors need scalarizing, but it's |
| 290 | // not a given that the mask does. For instance, in AVX512 v1i1 is legal. |
| 291 | // See the similar logic in ScalarizeVecRes_SETCC. |
| 292 | if (getTypeAction(VT: MaskVT) == TargetLowering::TypeScalarizeVector) |
| 293 | Mask = GetScalarizedVector(Op: Mask); |
| 294 | else |
| 295 | Mask = DAG.getExtractVectorElt(DL, VT: MaskVT.getVectorElementType(), Vec: Mask, Idx: 0); |
| 296 | // Vectors may have a different boolean contents to scalars, so truncate to i1 |
| 297 | // and let type legalization promote appropriately. |
| 298 | Mask = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: MVT::i1, Operand: Mask); |
| 299 | // Masked binary ops don't have UB on disabled lanes but produce poison, so |
| 300 | // use 1 as the divisor to avoid division by zero and overflow. |
| 301 | SDValue Divisor = DAG.getSelect(DL, VT: LHS.getValueType(), Cond: Mask, LHS: RHS, |
| 302 | RHS: DAG.getConstant(Val: 1, DL, VT: LHS.getValueType())); |
| 303 | return DAG.getNode(Opcode: ISD::getUnmaskedBinOpOpcode(MaskedOpc: N->getOpcode()), DL, |
| 304 | VT: LHS.getValueType(), N1: LHS, N2: Divisor); |
| 305 | } |
| 306 | |
| 307 | SDValue DAGTypeLegalizer::ScalarizeVecRes_CMP(SDNode *N) { |
| 308 | SDLoc DL(N); |
| 309 | |
| 310 | SDValue LHS = N->getOperand(Num: 0); |
| 311 | SDValue RHS = N->getOperand(Num: 1); |
| 312 | if (getTypeAction(VT: LHS.getValueType()) == |
| 313 | TargetLowering::TypeScalarizeVector) { |
| 314 | LHS = GetScalarizedVector(Op: LHS); |
| 315 | RHS = GetScalarizedVector(Op: RHS); |
| 316 | } else { |
| 317 | EVT VT = LHS.getValueType().getVectorElementType(); |
| 318 | LHS = DAG.getExtractVectorElt(DL, VT, Vec: LHS, Idx: 0); |
| 319 | RHS = DAG.getExtractVectorElt(DL, VT, Vec: RHS, Idx: 0); |
| 320 | } |
| 321 | |
| 322 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), |
| 323 | VT: N->getValueType(ResNo: 0).getVectorElementType(), N1: LHS, N2: RHS); |
| 324 | } |
| 325 | |
| 326 | SDValue DAGTypeLegalizer::ScalarizeVecRes_TernaryOp(SDNode *N) { |
| 327 | SDValue Op0 = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 328 | SDValue Op1 = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 329 | SDValue Op2 = GetScalarizedVector(Op: N->getOperand(Num: 2)); |
| 330 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: Op0.getValueType(), N1: Op0, N2: Op1, |
| 331 | N3: Op2, Flags: N->getFlags()); |
| 332 | } |
| 333 | |
| 334 | SDValue DAGTypeLegalizer::ScalarizeVecRes_FIX(SDNode *N) { |
| 335 | SDValue Op0 = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 336 | SDValue Op1 = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 337 | SDValue Op2 = N->getOperand(Num: 2); |
| 338 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: Op0.getValueType(), N1: Op0, N2: Op1, |
| 339 | N3: Op2, Flags: N->getFlags()); |
| 340 | } |
| 341 | |
| 342 | SDValue |
| 343 | DAGTypeLegalizer::ScalarizeVecRes_UnaryOpWithTwoResults(SDNode *N, |
| 344 | unsigned ResNo) { |
| 345 | assert(N->getValueType(0).getVectorNumElements() == 1 && |
| 346 | "Unexpected vector type!" ); |
| 347 | SDValue Elt = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 348 | |
| 349 | EVT VT0 = N->getValueType(ResNo: 0); |
| 350 | EVT VT1 = N->getValueType(ResNo: 1); |
| 351 | SDLoc dl(N); |
| 352 | |
| 353 | SDNode *ScalarNode = |
| 354 | DAG.getNode(Opcode: N->getOpcode(), DL: dl, |
| 355 | ResultTys: {VT0.getScalarType(), VT1.getScalarType()}, Ops: Elt) |
| 356 | .getNode(); |
| 357 | |
| 358 | // Replace the other vector result not being explicitly scalarized here. |
| 359 | unsigned OtherNo = 1 - ResNo; |
| 360 | EVT OtherVT = N->getValueType(ResNo: OtherNo); |
| 361 | if (getTypeAction(VT: OtherVT) == TargetLowering::TypeScalarizeVector) { |
| 362 | SetScalarizedVector(Op: SDValue(N, OtherNo), Result: SDValue(ScalarNode, OtherNo)); |
| 363 | } else { |
| 364 | SDValue OtherVal = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: dl, VT: OtherVT, |
| 365 | Operand: SDValue(ScalarNode, OtherNo)); |
| 366 | ReplaceValueWith(From: SDValue(N, OtherNo), To: OtherVal); |
| 367 | } |
| 368 | |
| 369 | return SDValue(ScalarNode, ResNo); |
| 370 | } |
| 371 | |
| 372 | SDValue DAGTypeLegalizer::ScalarizeVecRes_StrictFPOp(SDNode *N) { |
| 373 | EVT VT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 374 | unsigned NumOpers = N->getNumOperands(); |
| 375 | SDValue Chain = N->getOperand(Num: 0); |
| 376 | EVT ValueVTs[] = {VT, MVT::Other}; |
| 377 | SDLoc dl(N); |
| 378 | |
| 379 | SmallVector<SDValue, 4> Opers(NumOpers); |
| 380 | |
| 381 | // The Chain is the first operand. |
| 382 | Opers[0] = Chain; |
| 383 | |
| 384 | // Now process the remaining operands. |
| 385 | for (unsigned i = 1; i < NumOpers; ++i) { |
| 386 | SDValue Oper = N->getOperand(Num: i); |
| 387 | EVT OperVT = Oper.getValueType(); |
| 388 | |
| 389 | if (OperVT.isVector()) { |
| 390 | if (getTypeAction(VT: OperVT) == TargetLowering::TypeScalarizeVector) |
| 391 | Oper = GetScalarizedVector(Op: Oper); |
| 392 | else |
| 393 | Oper = |
| 394 | DAG.getExtractVectorElt(DL: dl, VT: OperVT.getVectorElementType(), Vec: Oper, Idx: 0); |
| 395 | } |
| 396 | |
| 397 | Opers[i] = Oper; |
| 398 | } |
| 399 | |
| 400 | SDValue Result = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VTList: DAG.getVTList(VTs: ValueVTs), |
| 401 | Ops: Opers, Flags: N->getFlags()); |
| 402 | |
| 403 | // Legalize the chain result - switch anything that used the old chain to |
| 404 | // use the new one. |
| 405 | ReplaceValueWith(From: SDValue(N, 1), To: Result.getValue(R: 1)); |
| 406 | return Result; |
| 407 | } |
| 408 | |
| 409 | SDValue DAGTypeLegalizer::ScalarizeVecRes_OverflowOp(SDNode *N, |
| 410 | unsigned ResNo) { |
| 411 | SDLoc DL(N); |
| 412 | EVT ResVT = N->getValueType(ResNo: 0); |
| 413 | EVT OvVT = N->getValueType(ResNo: 1); |
| 414 | |
| 415 | SDValue ScalarLHS, ScalarRHS; |
| 416 | if (getTypeAction(VT: ResVT) == TargetLowering::TypeScalarizeVector) { |
| 417 | ScalarLHS = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 418 | ScalarRHS = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 419 | } else { |
| 420 | SmallVector<SDValue, 1> ElemsLHS, ElemsRHS; |
| 421 | DAG.ExtractVectorElements(Op: N->getOperand(Num: 0), Args&: ElemsLHS); |
| 422 | DAG.ExtractVectorElements(Op: N->getOperand(Num: 1), Args&: ElemsRHS); |
| 423 | ScalarLHS = ElemsLHS[0]; |
| 424 | ScalarRHS = ElemsRHS[0]; |
| 425 | } |
| 426 | |
| 427 | SDVTList ScalarVTs = DAG.getVTList( |
| 428 | VT1: ResVT.getVectorElementType(), VT2: OvVT.getVectorElementType()); |
| 429 | SDNode *ScalarNode = DAG.getNode(Opcode: N->getOpcode(), DL, VTList: ScalarVTs, |
| 430 | Ops: {ScalarLHS, ScalarRHS}, Flags: N->getFlags()) |
| 431 | .getNode(); |
| 432 | |
| 433 | // Replace the other vector result not being explicitly scalarized here. |
| 434 | unsigned OtherNo = 1 - ResNo; |
| 435 | EVT OtherVT = N->getValueType(ResNo: OtherNo); |
| 436 | if (getTypeAction(VT: OtherVT) == TargetLowering::TypeScalarizeVector) { |
| 437 | SetScalarizedVector(Op: SDValue(N, OtherNo), Result: SDValue(ScalarNode, OtherNo)); |
| 438 | } else { |
| 439 | SDValue OtherVal = DAG.getNode( |
| 440 | Opcode: ISD::SCALAR_TO_VECTOR, DL, VT: OtherVT, Operand: SDValue(ScalarNode, OtherNo)); |
| 441 | ReplaceValueWith(From: SDValue(N, OtherNo), To: OtherVal); |
| 442 | } |
| 443 | |
| 444 | return SDValue(ScalarNode, ResNo); |
| 445 | } |
| 446 | |
| 447 | SDValue DAGTypeLegalizer::ScalarizeVecRes_MERGE_VALUES(SDNode *N, |
| 448 | unsigned ResNo) { |
| 449 | SDValue Op = DisintegrateMERGE_VALUES(N, ResNo); |
| 450 | return GetScalarizedVector(Op); |
| 451 | } |
| 452 | |
| 453 | SDValue DAGTypeLegalizer::ScalarizeVecRes_LOOP_DEPENDENCE_MASK(SDNode *N) { |
| 454 | SDLoc DL(N); |
| 455 | // Reuse the expansion (which should scalarize). |
| 456 | SDValue Mask = TLI.expandLoopDependenceMask(N, DAG); |
| 457 | return DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: SDLoc(N), |
| 458 | VT: N->getValueType(ResNo: 0).getScalarType(), N1: Mask, |
| 459 | N2: DAG.getVectorIdxConstant(Val: 0, DL)); |
| 460 | } |
| 461 | |
| 462 | SDValue DAGTypeLegalizer::ScalarizeVecRes_BITCAST(SDNode *N) { |
| 463 | SDValue Op = N->getOperand(Num: 0); |
| 464 | if (getTypeAction(VT: Op.getValueType()) == TargetLowering::TypeScalarizeVector) |
| 465 | Op = GetScalarizedVector(Op); |
| 466 | EVT NewVT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 467 | return DAG.getNode(Opcode: ISD::BITCAST, DL: SDLoc(N), |
| 468 | VT: NewVT, Operand: Op); |
| 469 | } |
| 470 | |
| 471 | SDValue DAGTypeLegalizer::ScalarizeVecRes_BUILD_VECTOR(SDNode *N) { |
| 472 | EVT EltVT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 473 | SDValue InOp = N->getOperand(Num: 0); |
| 474 | // The BUILD_VECTOR operands may be of wider element types and |
| 475 | // we may need to truncate them back to the requested return type. |
| 476 | if (EltVT.isInteger()) |
| 477 | return DAG.getNode(Opcode: ISD::TRUNCATE, DL: SDLoc(N), VT: EltVT, Operand: InOp); |
| 478 | return InOp; |
| 479 | } |
| 480 | |
| 481 | SDValue DAGTypeLegalizer::(SDNode *N) { |
| 482 | return DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: SDLoc(N), |
| 483 | VT: N->getValueType(ResNo: 0).getVectorElementType(), |
| 484 | N1: N->getOperand(Num: 0), N2: N->getOperand(Num: 1)); |
| 485 | } |
| 486 | |
| 487 | SDValue DAGTypeLegalizer::ScalarizeVecRes_FP_ROUND(SDNode *N) { |
| 488 | SDLoc DL(N); |
| 489 | SDValue Op = N->getOperand(Num: 0); |
| 490 | EVT OpVT = Op.getValueType(); |
| 491 | // The result needs scalarizing, but it's not a given that the source does. |
| 492 | // See similar logic in ScalarizeVecRes_UnaryOp. |
| 493 | if (getTypeAction(VT: OpVT) == TargetLowering::TypeScalarizeVector) { |
| 494 | Op = GetScalarizedVector(Op); |
| 495 | } else { |
| 496 | EVT VT = OpVT.getVectorElementType(); |
| 497 | Op = DAG.getExtractVectorElt(DL, VT, Vec: Op, Idx: 0); |
| 498 | } |
| 499 | return DAG.getNode(Opcode: ISD::FP_ROUND, DL, |
| 500 | VT: N->getValueType(ResNo: 0).getVectorElementType(), N1: Op, |
| 501 | N2: N->getOperand(Num: 1)); |
| 502 | } |
| 503 | |
| 504 | SDValue DAGTypeLegalizer::ScalarizeVecRes_CONVERT_FROM_ARBITRARY_FP(SDNode *N) { |
| 505 | SDLoc DL(N); |
| 506 | SDValue Op = N->getOperand(Num: 0); |
| 507 | EVT OpVT = Op.getValueType(); |
| 508 | // The result needs scalarizing, but it's not a given that the source does. |
| 509 | // See similar logic in ScalarizeVecRes_UnaryOp. |
| 510 | if (getTypeAction(VT: OpVT) == TargetLowering::TypeScalarizeVector) { |
| 511 | Op = GetScalarizedVector(Op); |
| 512 | } else { |
| 513 | EVT VT = OpVT.getVectorElementType(); |
| 514 | Op = DAG.getExtractVectorElt(DL, VT, Vec: Op, Idx: 0); |
| 515 | } |
| 516 | return DAG.getNode(Opcode: ISD::CONVERT_FROM_ARBITRARY_FP, DL, |
| 517 | VT: N->getValueType(ResNo: 0).getVectorElementType(), N1: Op, |
| 518 | N2: N->getOperand(Num: 1)); |
| 519 | } |
| 520 | |
| 521 | SDValue DAGTypeLegalizer::ScalarizeVecRes_CONVERT_TO_ARBITRARY_FP(SDNode *N) { |
| 522 | SDLoc DL(N); |
| 523 | SDValue Op = N->getOperand(Num: 0); |
| 524 | EVT OpVT = Op.getValueType(); |
| 525 | // The result needs scalarizing, but it's not a given that the source does. |
| 526 | if (getTypeAction(VT: OpVT) == TargetLowering::TypeScalarizeVector) { |
| 527 | Op = GetScalarizedVector(Op); |
| 528 | } else { |
| 529 | EVT VT = OpVT.getVectorElementType(); |
| 530 | Op = DAG.getExtractVectorElt(DL, VT, Vec: Op, Idx: 0); |
| 531 | } |
| 532 | return DAG.getNode(Opcode: ISD::CONVERT_TO_ARBITRARY_FP, DL, |
| 533 | VT: N->getValueType(ResNo: 0).getVectorElementType(), N1: Op, |
| 534 | N2: N->getOperand(Num: 1), N3: N->getOperand(Num: 2), N4: N->getOperand(Num: 3)); |
| 535 | } |
| 536 | |
| 537 | SDValue DAGTypeLegalizer::(SDNode *N) { |
| 538 | SDValue Op = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 539 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: Op.getValueType(), N1: Op, |
| 540 | N2: N->getOperand(Num: 1)); |
| 541 | } |
| 542 | |
| 543 | SDValue DAGTypeLegalizer::ScalarizeVecRes_INSERT_VECTOR_ELT(SDNode *N) { |
| 544 | // The value to insert may have a wider type than the vector element type, |
| 545 | // so be sure to truncate it to the element type if necessary. |
| 546 | SDValue Op = N->getOperand(Num: 1); |
| 547 | EVT EltVT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 548 | if (Op.getValueType() != EltVT) |
| 549 | // FIXME: Can this happen for floating point types? |
| 550 | Op = DAG.getNode(Opcode: ISD::TRUNCATE, DL: SDLoc(N), VT: EltVT, Operand: Op); |
| 551 | return Op; |
| 552 | } |
| 553 | |
| 554 | SDValue DAGTypeLegalizer::ScalarizeVecRes_ATOMIC_LOAD(AtomicSDNode *N) { |
| 555 | SDValue Result = DAG.getAtomicLoad( |
| 556 | ExtType: N->getExtensionType(), dl: SDLoc(N), MemVT: N->getMemoryVT().getVectorElementType(), |
| 557 | VT: N->getValueType(ResNo: 0).getVectorElementType(), Chain: N->getChain(), Ptr: N->getBasePtr(), |
| 558 | MMO: N->getMemOperand()); |
| 559 | |
| 560 | // Legalize the chain result - switch anything that used the old chain to |
| 561 | // use the new one. |
| 562 | ReplaceValueWith(From: SDValue(N, 1), To: Result.getValue(R: 1)); |
| 563 | return Result; |
| 564 | } |
| 565 | |
| 566 | SDValue DAGTypeLegalizer::ScalarizeVecRes_LOAD(LoadSDNode *N) { |
| 567 | assert(N->isUnindexed() && "Indexed vector load?" ); |
| 568 | |
| 569 | SDValue Result = DAG.getLoad( |
| 570 | AM: ISD::UNINDEXED, ExtType: N->getExtensionType(), |
| 571 | VT: N->getValueType(ResNo: 0).getVectorElementType(), dl: SDLoc(N), Chain: N->getChain(), |
| 572 | Ptr: N->getBasePtr(), Offset: DAG.getUNDEF(VT: N->getBasePtr().getValueType()), |
| 573 | PtrInfo: N->getPointerInfo(), MemVT: N->getMemoryVT().getVectorElementType(), |
| 574 | Alignment: N->getBaseAlign(), MMOFlags: N->getMemOperand()->getFlags(), AAInfo: N->getAAInfo()); |
| 575 | |
| 576 | // Legalize the chain result - switch anything that used the old chain to |
| 577 | // use the new one. |
| 578 | ReplaceValueWith(From: SDValue(N, 1), To: Result.getValue(R: 1)); |
| 579 | return Result; |
| 580 | } |
| 581 | |
| 582 | SDValue DAGTypeLegalizer::ScalarizeVecRes_UnaryOp(SDNode *N) { |
| 583 | // Get the dest type - it doesn't always match the input type, e.g. int_to_fp. |
| 584 | EVT DestVT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 585 | SDValue Op = N->getOperand(Num: 0); |
| 586 | EVT OpVT = Op.getValueType(); |
| 587 | SDLoc DL(N); |
| 588 | // The result needs scalarizing, but it's not a given that the source does. |
| 589 | // This is a workaround for targets where it's impossible to scalarize the |
| 590 | // result of a conversion, because the source type is legal. |
| 591 | // For instance, this happens on AArch64: v1i1 is illegal but v1i{8,16,32} |
| 592 | // are widened to v8i8, v4i16, and v2i32, which is legal, because v1i64 is |
| 593 | // legal and was not scalarized. |
| 594 | // See the similar logic in ScalarizeVecRes_SETCC |
| 595 | if (getTypeAction(VT: OpVT) == TargetLowering::TypeScalarizeVector) { |
| 596 | Op = GetScalarizedVector(Op); |
| 597 | } else { |
| 598 | EVT VT = OpVT.getVectorElementType(); |
| 599 | Op = DAG.getExtractVectorElt(DL, VT, Vec: Op, Idx: 0); |
| 600 | } |
| 601 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: DestVT, Operand: Op, Flags: N->getFlags()); |
| 602 | } |
| 603 | |
| 604 | SDValue DAGTypeLegalizer::ScalarizeVecRes_InregOp(SDNode *N) { |
| 605 | EVT EltVT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 606 | EVT ExtVT = cast<VTSDNode>(Val: N->getOperand(Num: 1))->getVT().getVectorElementType(); |
| 607 | SDValue LHS = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 608 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: EltVT, |
| 609 | N1: LHS, N2: DAG.getValueType(ExtVT)); |
| 610 | } |
| 611 | |
| 612 | SDValue DAGTypeLegalizer::ScalarizeVecRes_VecInregOp(SDNode *N) { |
| 613 | SDLoc DL(N); |
| 614 | SDValue Op = N->getOperand(Num: 0); |
| 615 | |
| 616 | EVT OpVT = Op.getValueType(); |
| 617 | EVT OpEltVT = OpVT.getVectorElementType(); |
| 618 | EVT EltVT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 619 | |
| 620 | if (getTypeAction(VT: OpVT) == TargetLowering::TypeScalarizeVector) { |
| 621 | Op = GetScalarizedVector(Op); |
| 622 | } else { |
| 623 | Op = DAG.getExtractVectorElt(DL, VT: OpEltVT, Vec: Op, Idx: 0); |
| 624 | } |
| 625 | |
| 626 | switch (N->getOpcode()) { |
| 627 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 628 | return DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT: EltVT, Operand: Op); |
| 629 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 630 | return DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL, VT: EltVT, Operand: Op); |
| 631 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 632 | return DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: EltVT, Operand: Op); |
| 633 | } |
| 634 | |
| 635 | llvm_unreachable("Illegal extend_vector_inreg opcode" ); |
| 636 | } |
| 637 | |
| 638 | SDValue DAGTypeLegalizer::ScalarizeVecRes_ADDRSPACECAST(SDNode *N) { |
| 639 | EVT DestVT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 640 | SDValue Op = N->getOperand(Num: 0); |
| 641 | EVT OpVT = Op.getValueType(); |
| 642 | SDLoc DL(N); |
| 643 | // The result needs scalarizing, but it's not a given that the source does. |
| 644 | // This is a workaround for targets where it's impossible to scalarize the |
| 645 | // result of a conversion, because the source type is legal. |
| 646 | // For instance, this happens on AArch64: v1i1 is illegal but v1i{8,16,32} |
| 647 | // are widened to v8i8, v4i16, and v2i32, which is legal, because v1i64 is |
| 648 | // legal and was not scalarized. |
| 649 | // See the similar logic in ScalarizeVecRes_SETCC |
| 650 | if (getTypeAction(VT: OpVT) == TargetLowering::TypeScalarizeVector) { |
| 651 | Op = GetScalarizedVector(Op); |
| 652 | } else { |
| 653 | EVT VT = OpVT.getVectorElementType(); |
| 654 | Op = DAG.getExtractVectorElt(DL, VT, Vec: Op, Idx: 0); |
| 655 | } |
| 656 | auto *AddrSpaceCastN = cast<AddrSpaceCastSDNode>(Val: N); |
| 657 | unsigned SrcAS = AddrSpaceCastN->getSrcAddressSpace(); |
| 658 | unsigned DestAS = AddrSpaceCastN->getDestAddressSpace(); |
| 659 | return DAG.getAddrSpaceCast(dl: DL, VT: DestVT, Ptr: Op, SrcAS, DestAS); |
| 660 | } |
| 661 | |
| 662 | SDValue DAGTypeLegalizer::ScalarizeVecRes_SCALAR_TO_VECTOR(SDNode *N) { |
| 663 | // If the operand is wider than the vector element type then it is implicitly |
| 664 | // truncated. Make that explicit here. |
| 665 | EVT EltVT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 666 | SDValue InOp = N->getOperand(Num: 0); |
| 667 | if (InOp.getValueType() != EltVT) |
| 668 | return DAG.getNode(Opcode: ISD::TRUNCATE, DL: SDLoc(N), VT: EltVT, Operand: InOp); |
| 669 | return InOp; |
| 670 | } |
| 671 | |
| 672 | SDValue DAGTypeLegalizer::ScalarizeVecRes_VSELECT(SDNode *N) { |
| 673 | SDValue Cond = N->getOperand(Num: 0); |
| 674 | EVT OpVT = Cond.getValueType(); |
| 675 | SDLoc DL(N); |
| 676 | // The vselect result and true/value operands needs scalarizing, but it's |
| 677 | // not a given that the Cond does. For instance, in AVX512 v1i1 is legal. |
| 678 | // See the similar logic in ScalarizeVecRes_SETCC |
| 679 | if (getTypeAction(VT: OpVT) == TargetLowering::TypeScalarizeVector) { |
| 680 | Cond = GetScalarizedVector(Op: Cond); |
| 681 | } else { |
| 682 | EVT VT = OpVT.getVectorElementType(); |
| 683 | Cond = DAG.getExtractVectorElt(DL, VT, Vec: Cond, Idx: 0); |
| 684 | } |
| 685 | |
| 686 | SDValue LHS = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 687 | TargetLowering::BooleanContent ScalarBool = |
| 688 | TLI.getBooleanContents(isVec: false, isFloat: false); |
| 689 | TargetLowering::BooleanContent VecBool = TLI.getBooleanContents(isVec: true, isFloat: false); |
| 690 | |
| 691 | // If integer and float booleans have different contents then we can't |
| 692 | // reliably optimize in all cases. There is a full explanation for this in |
| 693 | // DAGCombiner::visitSELECT() where the same issue affects folding |
| 694 | // (select C, 0, 1) to (xor C, 1). |
| 695 | if (TLI.getBooleanContents(isVec: false, isFloat: false) != |
| 696 | TLI.getBooleanContents(isVec: false, isFloat: true)) { |
| 697 | // At least try the common case where the boolean is generated by a |
| 698 | // comparison. |
| 699 | if (Cond->getOpcode() == ISD::SETCC) { |
| 700 | EVT OpVT = Cond->getOperand(Num: 0).getValueType(); |
| 701 | ScalarBool = TLI.getBooleanContents(Type: OpVT.getScalarType()); |
| 702 | VecBool = TLI.getBooleanContents(Type: OpVT); |
| 703 | } else |
| 704 | ScalarBool = TargetLowering::UndefinedBooleanContent; |
| 705 | } |
| 706 | |
| 707 | EVT CondVT = Cond.getValueType(); |
| 708 | if (ScalarBool != VecBool) { |
| 709 | switch (ScalarBool) { |
| 710 | case TargetLowering::UndefinedBooleanContent: |
| 711 | break; |
| 712 | case TargetLowering::ZeroOrOneBooleanContent: |
| 713 | assert(VecBool == TargetLowering::UndefinedBooleanContent || |
| 714 | VecBool == TargetLowering::ZeroOrNegativeOneBooleanContent); |
| 715 | // Vector read from all ones, scalar expects a single 1 so mask. |
| 716 | Cond = DAG.getNode(Opcode: ISD::AND, DL: SDLoc(N), VT: CondVT, |
| 717 | N1: Cond, N2: DAG.getConstant(Val: 1, DL: SDLoc(N), VT: CondVT)); |
| 718 | break; |
| 719 | case TargetLowering::ZeroOrNegativeOneBooleanContent: |
| 720 | assert(VecBool == TargetLowering::UndefinedBooleanContent || |
| 721 | VecBool == TargetLowering::ZeroOrOneBooleanContent); |
| 722 | // Vector reads from a one, scalar from all ones so sign extend. |
| 723 | Cond = DAG.getNode(Opcode: ISD::SIGN_EXTEND_INREG, DL: SDLoc(N), VT: CondVT, |
| 724 | N1: Cond, N2: DAG.getValueType(MVT::i1)); |
| 725 | break; |
| 726 | } |
| 727 | } |
| 728 | |
| 729 | // Truncate the condition if needed |
| 730 | auto BoolVT = getSetCCResultType(VT: CondVT); |
| 731 | if (BoolVT.bitsLT(VT: CondVT)) |
| 732 | Cond = DAG.getNode(Opcode: ISD::TRUNCATE, DL: SDLoc(N), VT: BoolVT, Operand: Cond); |
| 733 | |
| 734 | return DAG.getSelect(DL: SDLoc(N), |
| 735 | VT: LHS.getValueType(), Cond, LHS, |
| 736 | RHS: GetScalarizedVector(Op: N->getOperand(Num: 2))); |
| 737 | } |
| 738 | |
| 739 | SDValue DAGTypeLegalizer::ScalarizeVecRes_SELECT(SDNode *N) { |
| 740 | SDValue LHS = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 741 | return DAG.getSelect(DL: SDLoc(N), |
| 742 | VT: LHS.getValueType(), Cond: N->getOperand(Num: 0), LHS, |
| 743 | RHS: GetScalarizedVector(Op: N->getOperand(Num: 2))); |
| 744 | } |
| 745 | |
| 746 | SDValue DAGTypeLegalizer::ScalarizeVecRes_SELECT_CC(SDNode *N) { |
| 747 | SDValue LHS = GetScalarizedVector(Op: N->getOperand(Num: 2)); |
| 748 | return DAG.getNode(Opcode: ISD::SELECT_CC, DL: SDLoc(N), VT: LHS.getValueType(), |
| 749 | N1: N->getOperand(Num: 0), N2: N->getOperand(Num: 1), |
| 750 | N3: LHS, N4: GetScalarizedVector(Op: N->getOperand(Num: 3)), |
| 751 | N5: N->getOperand(Num: 4)); |
| 752 | } |
| 753 | |
| 754 | SDValue DAGTypeLegalizer::ScalarizeVecRes_UNDEF(SDNode *N) { |
| 755 | return DAG.getUNDEF(VT: N->getValueType(ResNo: 0).getVectorElementType()); |
| 756 | } |
| 757 | |
| 758 | SDValue DAGTypeLegalizer::ScalarizeVecRes_VECTOR_SHUFFLE(SDNode *N) { |
| 759 | // Figure out if the scalar is the LHS or RHS and return it. |
| 760 | SDValue Arg = N->getOperand(Num: 2).getOperand(i: 0); |
| 761 | if (Arg.isUndef()) |
| 762 | return DAG.getUNDEF(VT: N->getValueType(ResNo: 0).getVectorElementType()); |
| 763 | unsigned Op = !cast<ConstantSDNode>(Val&: Arg)->isZero(); |
| 764 | return GetScalarizedVector(Op: N->getOperand(Num: Op)); |
| 765 | } |
| 766 | |
| 767 | SDValue DAGTypeLegalizer::ScalarizeVecRes_FP_TO_XINT_SAT(SDNode *N) { |
| 768 | SDValue Src = N->getOperand(Num: 0); |
| 769 | EVT SrcVT = Src.getValueType(); |
| 770 | SDLoc dl(N); |
| 771 | |
| 772 | // Handle case where result is scalarized but operand is not |
| 773 | if (getTypeAction(VT: SrcVT) == TargetLowering::TypeScalarizeVector) |
| 774 | Src = GetScalarizedVector(Op: Src); |
| 775 | else |
| 776 | Src = DAG.getNode( |
| 777 | Opcode: ISD::EXTRACT_VECTOR_ELT, DL: dl, VT: SrcVT.getVectorElementType(), N1: Src, |
| 778 | N2: DAG.getConstant(Val: 0, DL: dl, VT: TLI.getVectorIdxTy(DL: DAG.getDataLayout()))); |
| 779 | |
| 780 | EVT DstVT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 781 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: DstVT, N1: Src, N2: N->getOperand(Num: 1)); |
| 782 | } |
| 783 | |
| 784 | SDValue DAGTypeLegalizer::ScalarizeVecRes_SETCC(SDNode *N) { |
| 785 | assert(N->getValueType(0).isVector() && |
| 786 | N->getOperand(0).getValueType().isVector() && |
| 787 | "Operand types must be vectors" ); |
| 788 | SDValue LHS = N->getOperand(Num: 0); |
| 789 | SDValue RHS = N->getOperand(Num: 1); |
| 790 | EVT OpVT = LHS.getValueType(); |
| 791 | EVT NVT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 792 | SDLoc DL(N); |
| 793 | |
| 794 | // The result needs scalarizing, but it's not a given that the source does. |
| 795 | if (getTypeAction(VT: OpVT) == TargetLowering::TypeScalarizeVector) { |
| 796 | LHS = GetScalarizedVector(Op: LHS); |
| 797 | RHS = GetScalarizedVector(Op: RHS); |
| 798 | } else { |
| 799 | EVT VT = OpVT.getVectorElementType(); |
| 800 | LHS = DAG.getExtractVectorElt(DL, VT, Vec: LHS, Idx: 0); |
| 801 | RHS = DAG.getExtractVectorElt(DL, VT, Vec: RHS, Idx: 0); |
| 802 | } |
| 803 | |
| 804 | // Turn it into a scalar SETCC. |
| 805 | SDValue Res = DAG.getNode(Opcode: ISD::SETCC, DL, VT: MVT::i1, N1: LHS, N2: RHS, |
| 806 | N3: N->getOperand(Num: 2)); |
| 807 | // Vectors may have a different boolean contents to scalars. Promote the |
| 808 | // value appropriately. |
| 809 | ISD::NodeType ExtendCode = |
| 810 | TargetLowering::getExtendForContent(Content: TLI.getBooleanContents(Type: OpVT)); |
| 811 | return DAG.getNode(Opcode: ExtendCode, DL, VT: NVT, Operand: Res); |
| 812 | } |
| 813 | |
| 814 | SDValue DAGTypeLegalizer::ScalarizeVecRes_IS_FPCLASS(SDNode *N) { |
| 815 | SDLoc DL(N); |
| 816 | SDValue Arg = N->getOperand(Num: 0); |
| 817 | SDValue Test = N->getOperand(Num: 1); |
| 818 | EVT ArgVT = Arg.getValueType(); |
| 819 | EVT ResultVT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 820 | |
| 821 | if (getTypeAction(VT: ArgVT) == TargetLowering::TypeScalarizeVector) { |
| 822 | Arg = GetScalarizedVector(Op: Arg); |
| 823 | } else { |
| 824 | EVT VT = ArgVT.getVectorElementType(); |
| 825 | Arg = DAG.getExtractVectorElt(DL, VT, Vec: Arg, Idx: 0); |
| 826 | } |
| 827 | |
| 828 | SDValue Res = |
| 829 | DAG.getNode(Opcode: ISD::IS_FPCLASS, DL, VT: MVT::i1, Ops: {Arg, Test}, Flags: N->getFlags()); |
| 830 | // Vectors may have a different boolean contents to scalars. Promote the |
| 831 | // value appropriately. |
| 832 | ISD::NodeType ExtendCode = |
| 833 | TargetLowering::getExtendForContent(Content: TLI.getBooleanContents(Type: ArgVT)); |
| 834 | return DAG.getNode(Opcode: ExtendCode, DL, VT: ResultVT, Operand: Res); |
| 835 | } |
| 836 | |
| 837 | //===----------------------------------------------------------------------===// |
| 838 | // Operand Vector Scalarization <1 x ty> -> ty. |
| 839 | //===----------------------------------------------------------------------===// |
| 840 | |
| 841 | bool DAGTypeLegalizer::ScalarizeVectorOperand(SDNode *N, unsigned OpNo) { |
| 842 | LLVM_DEBUG(dbgs() << "Scalarize node operand " << OpNo << ": " ; |
| 843 | N->dump(&DAG)); |
| 844 | SDValue Res = SDValue(); |
| 845 | |
| 846 | // See if the target wants to custom scalarize this node. |
| 847 | if (CustomLowerNode(N, VT: N->getOperand(Num: OpNo).getValueType(), LegalizeResult: false)) |
| 848 | return false; |
| 849 | |
| 850 | switch (N->getOpcode()) { |
| 851 | default: |
| 852 | #ifndef NDEBUG |
| 853 | dbgs() << "ScalarizeVectorOperand Op #" << OpNo << ": " ; |
| 854 | N->dump(&DAG); |
| 855 | dbgs() << "\n" ; |
| 856 | #endif |
| 857 | report_fatal_error(reason: "Do not know how to scalarize this operator's " |
| 858 | "operand!\n" ); |
| 859 | case ISD::BITCAST: |
| 860 | Res = ScalarizeVecOp_BITCAST(N); |
| 861 | break; |
| 862 | case ISD::FAKE_USE: |
| 863 | Res = ScalarizeVecOp_FAKE_USE(N); |
| 864 | break; |
| 865 | case ISD::ANY_EXTEND: |
| 866 | case ISD::ZERO_EXTEND: |
| 867 | case ISD::SIGN_EXTEND: |
| 868 | case ISD::TRUNCATE: |
| 869 | case ISD::FP_TO_SINT: |
| 870 | case ISD::FP_TO_UINT: |
| 871 | case ISD::SINT_TO_FP: |
| 872 | case ISD::UINT_TO_FP: |
| 873 | case ISD::LROUND: |
| 874 | case ISD::LLROUND: |
| 875 | case ISD::LRINT: |
| 876 | case ISD::LLRINT: |
| 877 | Res = ScalarizeVecOp_UnaryOp(N); |
| 878 | break; |
| 879 | case ISD::FP_TO_SINT_SAT: |
| 880 | case ISD::FP_TO_UINT_SAT: |
| 881 | case ISD::CONVERT_FROM_ARBITRARY_FP: |
| 882 | Res = ScalarizeVecOp_UnaryOpWithExtraInput(N); |
| 883 | break; |
| 884 | case ISD::CONVERT_TO_ARBITRARY_FP: { |
| 885 | assert(N->getValueType(0).getVectorNumElements() == 1 && |
| 886 | "Unexpected vector type!" ); |
| 887 | SDValue Elt = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 888 | SDValue Op = DAG.getNode( |
| 889 | Opcode: N->getOpcode(), DL: SDLoc(N), VT: N->getValueType(ResNo: 0).getScalarType(), N1: Elt, |
| 890 | N2: N->getOperand(Num: 1), N3: N->getOperand(Num: 2), N4: N->getOperand(Num: 3)); |
| 891 | Res = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: SDLoc(N), VT: N->getValueType(ResNo: 0), Operand: Op); |
| 892 | break; |
| 893 | } |
| 894 | case ISD::STRICT_SINT_TO_FP: |
| 895 | case ISD::STRICT_UINT_TO_FP: |
| 896 | case ISD::STRICT_FP_TO_SINT: |
| 897 | case ISD::STRICT_FP_TO_UINT: |
| 898 | Res = ScalarizeVecOp_UnaryOp_StrictFP(N); |
| 899 | break; |
| 900 | case ISD::CONCAT_VECTORS: |
| 901 | Res = ScalarizeVecOp_CONCAT_VECTORS(N); |
| 902 | break; |
| 903 | case ISD::INSERT_SUBVECTOR: |
| 904 | Res = ScalarizeVecOp_INSERT_SUBVECTOR(N, OpNo); |
| 905 | break; |
| 906 | case ISD::EXTRACT_VECTOR_ELT: |
| 907 | Res = ScalarizeVecOp_EXTRACT_VECTOR_ELT(N); |
| 908 | break; |
| 909 | case ISD::VSELECT: |
| 910 | Res = ScalarizeVecOp_VSELECT(N); |
| 911 | break; |
| 912 | case ISD::SETCC: |
| 913 | Res = ScalarizeVecOp_VSETCC(N); |
| 914 | break; |
| 915 | case ISD::STRICT_FSETCC: |
| 916 | case ISD::STRICT_FSETCCS: |
| 917 | Res = ScalarizeVecOp_VSTRICT_FSETCC(N, OpNo); |
| 918 | break; |
| 919 | case ISD::STORE: |
| 920 | Res = ScalarizeVecOp_STORE(N: cast<StoreSDNode>(Val: N), OpNo); |
| 921 | break; |
| 922 | case ISD::ATOMIC_STORE: |
| 923 | Res = ScalarizeVecOp_ATOMIC_STORE(N: cast<AtomicSDNode>(Val: N)); |
| 924 | break; |
| 925 | case ISD::STRICT_FP_ROUND: |
| 926 | Res = ScalarizeVecOp_STRICT_FP_ROUND(N, OpNo); |
| 927 | break; |
| 928 | case ISD::FP_ROUND: |
| 929 | Res = ScalarizeVecOp_FP_ROUND(N, OpNo); |
| 930 | break; |
| 931 | case ISD::STRICT_FP_EXTEND: |
| 932 | Res = ScalarizeVecOp_STRICT_FP_EXTEND(N); |
| 933 | break; |
| 934 | case ISD::FP_EXTEND: |
| 935 | Res = ScalarizeVecOp_FP_EXTEND(N); |
| 936 | break; |
| 937 | case ISD::VECREDUCE_FADD: |
| 938 | case ISD::VECREDUCE_FMUL: |
| 939 | case ISD::VECREDUCE_ADD: |
| 940 | case ISD::VECREDUCE_MUL: |
| 941 | case ISD::VECREDUCE_AND: |
| 942 | case ISD::VECREDUCE_OR: |
| 943 | case ISD::VECREDUCE_XOR: |
| 944 | case ISD::VECREDUCE_SMAX: |
| 945 | case ISD::VECREDUCE_SMIN: |
| 946 | case ISD::VECREDUCE_UMAX: |
| 947 | case ISD::VECREDUCE_UMIN: |
| 948 | case ISD::VECREDUCE_FMAX: |
| 949 | case ISD::VECREDUCE_FMIN: |
| 950 | case ISD::VECREDUCE_FMAXIMUM: |
| 951 | case ISD::VECREDUCE_FMINIMUM: |
| 952 | Res = ScalarizeVecOp_VECREDUCE(N); |
| 953 | break; |
| 954 | case ISD::VECREDUCE_SEQ_FADD: |
| 955 | case ISD::VECREDUCE_SEQ_FMUL: |
| 956 | Res = ScalarizeVecOp_VECREDUCE_SEQ(N); |
| 957 | break; |
| 958 | case ISD::SCMP: |
| 959 | case ISD::UCMP: |
| 960 | Res = ScalarizeVecOp_CMP(N); |
| 961 | break; |
| 962 | case ISD::VECTOR_FIND_LAST_ACTIVE: |
| 963 | Res = ScalarizeVecOp_VECTOR_FIND_LAST_ACTIVE(N); |
| 964 | break; |
| 965 | case ISD::CTTZ_ELTS: |
| 966 | case ISD::CTTZ_ELTS_ZERO_POISON: |
| 967 | Res = ScalarizeVecOp_CTTZ_ELTS(N); |
| 968 | break; |
| 969 | case ISD::MASKED_UDIV: |
| 970 | case ISD::MASKED_SDIV: |
| 971 | case ISD::MASKED_UREM: |
| 972 | case ISD::MASKED_SREM: |
| 973 | Res = ScalarizeVecOp_MaskedBinOp(N, OpNo); |
| 974 | break; |
| 975 | } |
| 976 | |
| 977 | // If the result is null, the sub-method took care of registering results etc. |
| 978 | if (!Res.getNode()) return false; |
| 979 | |
| 980 | // If the result is N, the sub-method updated N in place. Tell the legalizer |
| 981 | // core about this. |
| 982 | if (Res.getNode() == N) |
| 983 | return true; |
| 984 | |
| 985 | assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 && |
| 986 | "Invalid operand expansion" ); |
| 987 | |
| 988 | ReplaceValueWith(From: SDValue(N, 0), To: Res); |
| 989 | return false; |
| 990 | } |
| 991 | |
| 992 | /// If the value to convert is a vector that needs to be scalarized, it must be |
| 993 | /// <1 x ty>. Convert the element instead. |
| 994 | SDValue DAGTypeLegalizer::ScalarizeVecOp_BITCAST(SDNode *N) { |
| 995 | SDValue Elt = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 996 | return DAG.getNode(Opcode: ISD::BITCAST, DL: SDLoc(N), |
| 997 | VT: N->getValueType(ResNo: 0), Operand: Elt); |
| 998 | } |
| 999 | |
| 1000 | // Need to legalize vector operands of fake uses. Must be <1 x ty>. |
| 1001 | SDValue DAGTypeLegalizer::ScalarizeVecOp_FAKE_USE(SDNode *N) { |
| 1002 | assert(N->getOperand(1).getValueType().getVectorNumElements() == 1 && |
| 1003 | "Fake Use: Unexpected vector type!" ); |
| 1004 | SDValue Elt = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 1005 | return DAG.getNode(Opcode: ISD::FAKE_USE, DL: SDLoc(), VT: MVT::Other, N1: N->getOperand(Num: 0), N2: Elt); |
| 1006 | } |
| 1007 | |
| 1008 | /// If the input is a vector that needs to be scalarized, it must be <1 x ty>. |
| 1009 | /// Do the operation on the element instead. |
| 1010 | SDValue DAGTypeLegalizer::ScalarizeVecOp_UnaryOp(SDNode *N) { |
| 1011 | assert(N->getValueType(0).getVectorNumElements() == 1 && |
| 1012 | "Unexpected vector type!" ); |
| 1013 | SDValue Elt = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 1014 | SDValue Op = DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), |
| 1015 | VT: N->getValueType(ResNo: 0).getScalarType(), Operand: Elt); |
| 1016 | // Revectorize the result so the types line up with what the uses of this |
| 1017 | // expression expect. |
| 1018 | return DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: SDLoc(N), VT: N->getValueType(ResNo: 0), Operand: Op); |
| 1019 | } |
| 1020 | |
| 1021 | /// Same as ScalarizeVecOp_UnaryOp with an extra operand (for example a |
| 1022 | /// typesize). |
| 1023 | SDValue DAGTypeLegalizer::(SDNode *N) { |
| 1024 | assert(N->getValueType(0).getVectorNumElements() == 1 && |
| 1025 | "Unexpected vector type!" ); |
| 1026 | SDValue Elt = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 1027 | SDValue Op = |
| 1028 | DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: N->getValueType(ResNo: 0).getScalarType(), |
| 1029 | N1: Elt, N2: N->getOperand(Num: 1)); |
| 1030 | // Revectorize the result so the types line up with what the uses of this |
| 1031 | // expression expect. |
| 1032 | return DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: SDLoc(N), VT: N->getValueType(ResNo: 0), Operand: Op); |
| 1033 | } |
| 1034 | |
| 1035 | /// If the input is a vector that needs to be scalarized, it must be <1 x ty>. |
| 1036 | /// Do the strict FP operation on the element instead. |
| 1037 | SDValue DAGTypeLegalizer::ScalarizeVecOp_UnaryOp_StrictFP(SDNode *N) { |
| 1038 | assert(N->getValueType(0).getVectorNumElements() == 1 && |
| 1039 | "Unexpected vector type!" ); |
| 1040 | SDValue Elt = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 1041 | SDValue Res = DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), |
| 1042 | ResultTys: { N->getValueType(ResNo: 0).getScalarType(), MVT::Other }, |
| 1043 | Ops: { N->getOperand(Num: 0), Elt }); |
| 1044 | // Legalize the chain result - switch anything that used the old chain to |
| 1045 | // use the new one. |
| 1046 | ReplaceValueWith(From: SDValue(N, 1), To: Res.getValue(R: 1)); |
| 1047 | // Revectorize the result so the types line up with what the uses of this |
| 1048 | // expression expect. |
| 1049 | Res = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: SDLoc(N), VT: N->getValueType(ResNo: 0), Operand: Res); |
| 1050 | |
| 1051 | // Do our own replacement and return SDValue() to tell the caller that we |
| 1052 | // handled all replacements since caller can only handle a single result. |
| 1053 | ReplaceValueWith(From: SDValue(N, 0), To: Res); |
| 1054 | return SDValue(); |
| 1055 | } |
| 1056 | |
| 1057 | /// The vectors to concatenate have length one - use a BUILD_VECTOR instead. |
| 1058 | SDValue DAGTypeLegalizer::ScalarizeVecOp_CONCAT_VECTORS(SDNode *N) { |
| 1059 | SmallVector<SDValue, 8> Ops(N->getNumOperands()); |
| 1060 | for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i) |
| 1061 | Ops[i] = GetScalarizedVector(Op: N->getOperand(Num: i)); |
| 1062 | return DAG.getBuildVector(VT: N->getValueType(ResNo: 0), DL: SDLoc(N), Ops); |
| 1063 | } |
| 1064 | |
| 1065 | /// The inserted subvector is to be scalarized - use insert vector element |
| 1066 | /// instead. |
| 1067 | SDValue DAGTypeLegalizer::ScalarizeVecOp_INSERT_SUBVECTOR(SDNode *N, |
| 1068 | unsigned OpNo) { |
| 1069 | // We should not be attempting to scalarize the containing vector |
| 1070 | assert(OpNo == 1); |
| 1071 | SDValue Elt = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 1072 | SDValue ContainingVec = N->getOperand(Num: 0); |
| 1073 | return DAG.getNode(Opcode: ISD::INSERT_VECTOR_ELT, DL: SDLoc(N), |
| 1074 | VT: ContainingVec.getValueType(), N1: ContainingVec, N2: Elt, |
| 1075 | N3: N->getOperand(Num: 2)); |
| 1076 | } |
| 1077 | |
| 1078 | /// If the input is a vector that needs to be scalarized, it must be <1 x ty>, |
| 1079 | /// so just return the element, ignoring the index. |
| 1080 | SDValue DAGTypeLegalizer::(SDNode *N) { |
| 1081 | EVT VT = N->getValueType(ResNo: 0); |
| 1082 | SDValue Res = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 1083 | if (Res.getValueType() != VT) |
| 1084 | Res = VT.isFloatingPoint() |
| 1085 | ? DAG.getNode(Opcode: ISD::FP_EXTEND, DL: SDLoc(N), VT, Operand: Res) |
| 1086 | : DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: SDLoc(N), VT, Operand: Res); |
| 1087 | return Res; |
| 1088 | } |
| 1089 | |
| 1090 | /// If the input condition is a vector that needs to be scalarized, it must be |
| 1091 | /// <1 x i1>, so just convert to a normal ISD::SELECT |
| 1092 | /// (still with vector output type since that was acceptable if we got here). |
| 1093 | SDValue DAGTypeLegalizer::ScalarizeVecOp_VSELECT(SDNode *N) { |
| 1094 | SDValue ScalarCond = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 1095 | EVT VT = N->getValueType(ResNo: 0); |
| 1096 | |
| 1097 | return DAG.getNode(Opcode: ISD::SELECT, DL: SDLoc(N), VT, N1: ScalarCond, N2: N->getOperand(Num: 1), |
| 1098 | N3: N->getOperand(Num: 2)); |
| 1099 | } |
| 1100 | |
| 1101 | /// If the operand is a vector that needs to be scalarized then the |
| 1102 | /// result must be v1i1, so just convert to a scalar SETCC and wrap |
| 1103 | /// with a scalar_to_vector since the res type is legal if we got here |
| 1104 | SDValue DAGTypeLegalizer::ScalarizeVecOp_VSETCC(SDNode *N) { |
| 1105 | assert(N->getValueType(0).isVector() && |
| 1106 | N->getOperand(0).getValueType().isVector() && |
| 1107 | "Operand types must be vectors" ); |
| 1108 | assert(N->getValueType(0) == MVT::v1i1 && "Expected v1i1 type" ); |
| 1109 | |
| 1110 | EVT VT = N->getValueType(ResNo: 0); |
| 1111 | SDValue LHS = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 1112 | SDValue RHS = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 1113 | |
| 1114 | EVT OpVT = N->getOperand(Num: 0).getValueType(); |
| 1115 | EVT NVT = VT.getVectorElementType(); |
| 1116 | SDLoc DL(N); |
| 1117 | // Turn it into a scalar SETCC. |
| 1118 | SDValue Res = DAG.getNode(Opcode: ISD::SETCC, DL, VT: MVT::i1, N1: LHS, N2: RHS, |
| 1119 | N3: N->getOperand(Num: 2)); |
| 1120 | |
| 1121 | // Vectors may have a different boolean contents to scalars. Promote the |
| 1122 | // value appropriately. |
| 1123 | ISD::NodeType ExtendCode = |
| 1124 | TargetLowering::getExtendForContent(Content: TLI.getBooleanContents(Type: OpVT)); |
| 1125 | |
| 1126 | Res = DAG.getNode(Opcode: ExtendCode, DL, VT: NVT, Operand: Res); |
| 1127 | |
| 1128 | return DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL, VT, Operand: Res); |
| 1129 | } |
| 1130 | |
| 1131 | // Similiar to ScalarizeVecOp_VSETCC, with added logic to update chains. |
| 1132 | SDValue DAGTypeLegalizer::ScalarizeVecOp_VSTRICT_FSETCC(SDNode *N, |
| 1133 | unsigned OpNo) { |
| 1134 | assert(OpNo == 1 && "Wrong operand for scalarization!" ); |
| 1135 | assert(N->getValueType(0).isVector() && |
| 1136 | N->getOperand(1).getValueType().isVector() && |
| 1137 | "Operand types must be vectors" ); |
| 1138 | assert(N->getValueType(0) == MVT::v1i1 && "Expected v1i1 type" ); |
| 1139 | |
| 1140 | EVT VT = N->getValueType(ResNo: 0); |
| 1141 | SDValue Ch = N->getOperand(Num: 0); |
| 1142 | SDValue LHS = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 1143 | SDValue RHS = GetScalarizedVector(Op: N->getOperand(Num: 2)); |
| 1144 | SDValue CC = N->getOperand(Num: 3); |
| 1145 | |
| 1146 | EVT OpVT = N->getOperand(Num: 1).getValueType(); |
| 1147 | EVT NVT = VT.getVectorElementType(); |
| 1148 | SDLoc DL(N); |
| 1149 | SDValue Res = DAG.getNode(Opcode: N->getOpcode(), DL, ResultTys: {MVT::i1, MVT::Other}, |
| 1150 | Ops: {Ch, LHS, RHS, CC}); |
| 1151 | |
| 1152 | // Legalize the chain result - switch anything that used the old chain to |
| 1153 | // use the new one. |
| 1154 | ReplaceValueWith(From: SDValue(N, 1), To: Res.getValue(R: 1)); |
| 1155 | |
| 1156 | ISD::NodeType ExtendCode = |
| 1157 | TargetLowering::getExtendForContent(Content: TLI.getBooleanContents(Type: OpVT)); |
| 1158 | |
| 1159 | Res = DAG.getNode(Opcode: ExtendCode, DL, VT: NVT, Operand: Res); |
| 1160 | Res = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL, VT, Operand: Res); |
| 1161 | |
| 1162 | // Do our own replacement and return SDValue() to tell the caller that we |
| 1163 | // handled all replacements since caller can only handle a single result. |
| 1164 | ReplaceValueWith(From: SDValue(N, 0), To: Res); |
| 1165 | return SDValue(); |
| 1166 | } |
| 1167 | |
| 1168 | /// If the value to store is a vector that needs to be scalarized, it must be |
| 1169 | /// <1 x ty>. Just store the element. |
| 1170 | SDValue DAGTypeLegalizer::ScalarizeVecOp_STORE(StoreSDNode *N, unsigned OpNo){ |
| 1171 | assert(N->isUnindexed() && "Indexed store of one-element vector?" ); |
| 1172 | assert(OpNo == 1 && "Do not know how to scalarize this operand!" ); |
| 1173 | SDLoc dl(N); |
| 1174 | |
| 1175 | if (N->isTruncatingStore()) |
| 1176 | return DAG.getTruncStore( |
| 1177 | Chain: N->getChain(), dl, Val: GetScalarizedVector(Op: N->getOperand(Num: 1)), |
| 1178 | Ptr: N->getBasePtr(), PtrInfo: N->getPointerInfo(), |
| 1179 | SVT: N->getMemoryVT().getVectorElementType(), Alignment: N->getBaseAlign(), |
| 1180 | MMOFlags: N->getMemOperand()->getFlags(), AAInfo: N->getAAInfo()); |
| 1181 | |
| 1182 | return DAG.getStore(Chain: N->getChain(), dl, Val: GetScalarizedVector(Op: N->getOperand(Num: 1)), |
| 1183 | Ptr: N->getBasePtr(), PtrInfo: N->getPointerInfo(), Alignment: N->getBaseAlign(), |
| 1184 | MMOFlags: N->getMemOperand()->getFlags(), AAInfo: N->getAAInfo()); |
| 1185 | } |
| 1186 | |
| 1187 | /// If the value to store is a vector that needs to be scalarized, it must be |
| 1188 | /// <1 x ty>. Just store the element. |
| 1189 | SDValue DAGTypeLegalizer::ScalarizeVecOp_ATOMIC_STORE(AtomicSDNode *N) { |
| 1190 | SDValue ScalarVal = GetScalarizedVector(Op: N->getVal()); |
| 1191 | return DAG.getAtomic(Opcode: ISD::ATOMIC_STORE, dl: SDLoc(N), |
| 1192 | MemVT: N->getMemoryVT().getVectorElementType(), Chain: N->getChain(), |
| 1193 | Ptr: ScalarVal, Val: N->getBasePtr(), MMO: N->getMemOperand()); |
| 1194 | } |
| 1195 | |
| 1196 | /// If the value to round is a vector that needs to be scalarized, it must be |
| 1197 | /// <1 x ty>. Convert the element instead. |
| 1198 | SDValue DAGTypeLegalizer::ScalarizeVecOp_FP_ROUND(SDNode *N, unsigned OpNo) { |
| 1199 | assert(OpNo == 0 && "Wrong operand for scalarization!" ); |
| 1200 | SDValue Elt = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 1201 | SDValue Res = DAG.getNode(Opcode: ISD::FP_ROUND, DL: SDLoc(N), |
| 1202 | VT: N->getValueType(ResNo: 0).getVectorElementType(), N1: Elt, |
| 1203 | N2: N->getOperand(Num: 1)); |
| 1204 | return DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: SDLoc(N), VT: N->getValueType(ResNo: 0), Operand: Res); |
| 1205 | } |
| 1206 | |
| 1207 | SDValue DAGTypeLegalizer::ScalarizeVecOp_STRICT_FP_ROUND(SDNode *N, |
| 1208 | unsigned OpNo) { |
| 1209 | assert(OpNo == 1 && "Wrong operand for scalarization!" ); |
| 1210 | SDValue Elt = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 1211 | SDValue Res = |
| 1212 | DAG.getNode(Opcode: ISD::STRICT_FP_ROUND, DL: SDLoc(N), |
| 1213 | ResultTys: {N->getValueType(ResNo: 0).getVectorElementType(), MVT::Other}, |
| 1214 | Ops: {N->getOperand(Num: 0), Elt, N->getOperand(Num: 2)}); |
| 1215 | // Legalize the chain result - switch anything that used the old chain to |
| 1216 | // use the new one. |
| 1217 | ReplaceValueWith(From: SDValue(N, 1), To: Res.getValue(R: 1)); |
| 1218 | |
| 1219 | Res = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: SDLoc(N), VT: N->getValueType(ResNo: 0), Operand: Res); |
| 1220 | |
| 1221 | // Do our own replacement and return SDValue() to tell the caller that we |
| 1222 | // handled all replacements since caller can only handle a single result. |
| 1223 | ReplaceValueWith(From: SDValue(N, 0), To: Res); |
| 1224 | return SDValue(); |
| 1225 | } |
| 1226 | |
| 1227 | /// If the value to extend is a vector that needs to be scalarized, it must be |
| 1228 | /// <1 x ty>. Convert the element instead. |
| 1229 | SDValue DAGTypeLegalizer::ScalarizeVecOp_FP_EXTEND(SDNode *N) { |
| 1230 | SDValue Elt = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 1231 | SDValue Res = DAG.getNode(Opcode: ISD::FP_EXTEND, DL: SDLoc(N), |
| 1232 | VT: N->getValueType(ResNo: 0).getVectorElementType(), Operand: Elt); |
| 1233 | return DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: SDLoc(N), VT: N->getValueType(ResNo: 0), Operand: Res); |
| 1234 | } |
| 1235 | |
| 1236 | /// If the value to extend is a vector that needs to be scalarized, it must be |
| 1237 | /// <1 x ty>. Convert the element instead. |
| 1238 | SDValue DAGTypeLegalizer::ScalarizeVecOp_STRICT_FP_EXTEND(SDNode *N) { |
| 1239 | SDValue Elt = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 1240 | SDValue Res = |
| 1241 | DAG.getNode(Opcode: ISD::STRICT_FP_EXTEND, DL: SDLoc(N), |
| 1242 | ResultTys: {N->getValueType(ResNo: 0).getVectorElementType(), MVT::Other}, |
| 1243 | Ops: {N->getOperand(Num: 0), Elt}); |
| 1244 | // Legalize the chain result - switch anything that used the old chain to |
| 1245 | // use the new one. |
| 1246 | ReplaceValueWith(From: SDValue(N, 1), To: Res.getValue(R: 1)); |
| 1247 | |
| 1248 | Res = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: SDLoc(N), VT: N->getValueType(ResNo: 0), Operand: Res); |
| 1249 | |
| 1250 | // Do our own replacement and return SDValue() to tell the caller that we |
| 1251 | // handled all replacements since caller can only handle a single result. |
| 1252 | ReplaceValueWith(From: SDValue(N, 0), To: Res); |
| 1253 | return SDValue(); |
| 1254 | } |
| 1255 | |
| 1256 | SDValue DAGTypeLegalizer::ScalarizeVecOp_VECREDUCE(SDNode *N) { |
| 1257 | SDValue Res = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 1258 | // Result type may be wider than element type. |
| 1259 | if (Res.getValueType() != N->getValueType(ResNo: 0)) |
| 1260 | Res = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: SDLoc(N), VT: N->getValueType(ResNo: 0), Operand: Res); |
| 1261 | return Res; |
| 1262 | } |
| 1263 | |
| 1264 | SDValue DAGTypeLegalizer::ScalarizeVecOp_VECREDUCE_SEQ(SDNode *N) { |
| 1265 | SDValue AccOp = N->getOperand(Num: 0); |
| 1266 | SDValue VecOp = N->getOperand(Num: 1); |
| 1267 | |
| 1268 | unsigned BaseOpc = ISD::getVecReduceBaseOpcode(VecReduceOpcode: N->getOpcode()); |
| 1269 | |
| 1270 | SDValue Op = GetScalarizedVector(Op: VecOp); |
| 1271 | return DAG.getNode(Opcode: BaseOpc, DL: SDLoc(N), VT: N->getValueType(ResNo: 0), |
| 1272 | N1: AccOp, N2: Op, Flags: N->getFlags()); |
| 1273 | } |
| 1274 | |
| 1275 | SDValue DAGTypeLegalizer::ScalarizeVecOp_CMP(SDNode *N) { |
| 1276 | SDValue LHS = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 1277 | SDValue RHS = GetScalarizedVector(Op: N->getOperand(Num: 1)); |
| 1278 | |
| 1279 | EVT ResVT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 1280 | SDValue Cmp = DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: ResVT, N1: LHS, N2: RHS); |
| 1281 | return DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: SDLoc(N), VT: N->getValueType(ResNo: 0), Operand: Cmp); |
| 1282 | } |
| 1283 | |
| 1284 | SDValue DAGTypeLegalizer::ScalarizeVecOp_VECTOR_FIND_LAST_ACTIVE(SDNode *N) { |
| 1285 | // Since there is no "none-active" result, the only valid return for <1 x ty> |
| 1286 | // is 0. Note: Since we check the high mask during splitting this is safe. |
| 1287 | // As e.g., a <2 x ty> operation would split to: |
| 1288 | // any_active(%hi_mask) ? (1 + last_active(%hi_mask)) |
| 1289 | // : `last_active(%lo_mask)` |
| 1290 | // Which then scalarizes to: |
| 1291 | // %mask[1] ? 1 : 0 |
| 1292 | EVT VT = N->getValueType(ResNo: 0); |
| 1293 | return DAG.getConstant(Val: 0, DL: SDLoc(N), VT); |
| 1294 | } |
| 1295 | |
| 1296 | SDValue DAGTypeLegalizer::ScalarizeVecOp_CTTZ_ELTS(SDNode *N) { |
| 1297 | // The number of trailing zero elements is 1 if the element is 0, and 0 |
| 1298 | // otherwise. |
| 1299 | if (N->getOpcode() == ISD::CTTZ_ELTS_ZERO_POISON) |
| 1300 | return DAG.getConstant(Val: 0, DL: SDLoc(N), VT: N->getValueType(ResNo: 0)); |
| 1301 | SDValue Op = GetScalarizedVector(Op: N->getOperand(Num: 0)); |
| 1302 | SDValue SetCC = |
| 1303 | DAG.getSetCC(DL: SDLoc(N), VT: MVT::i1, LHS: Op, |
| 1304 | RHS: DAG.getConstant(Val: 0, DL: SDLoc(N), VT: Op.getValueType()), Cond: ISD::SETEQ); |
| 1305 | return DAG.getZExtOrTrunc(Op: SetCC, DL: SDLoc(N), VT: N->getValueType(ResNo: 0)); |
| 1306 | } |
| 1307 | |
| 1308 | SDValue DAGTypeLegalizer::ScalarizeVecOp_MaskedBinOp(SDNode *N, unsigned OpNo) { |
| 1309 | assert(OpNo == 2 && "Can only scalarize mask operand" ); |
| 1310 | SDLoc DL(N); |
| 1311 | EVT VT = N->getOperand(Num: 0).getValueType().getVectorElementType(); |
| 1312 | SDValue LHS = DAG.getExtractVectorElt(DL, VT, Vec: N->getOperand(Num: 0), Idx: 0); |
| 1313 | SDValue RHS = DAG.getExtractVectorElt(DL, VT, Vec: N->getOperand(Num: 1), Idx: 0); |
| 1314 | SDValue Mask = GetScalarizedVector(Op: N->getOperand(Num: 2)); |
| 1315 | // Vectors may have a different boolean contents to scalars, so truncate to i1 |
| 1316 | // and let type legalization promote appropriately. |
| 1317 | Mask = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: MVT::i1, Operand: Mask); |
| 1318 | // Masked binary ops don't have UB on disabled lanes but produce poison, so |
| 1319 | // use 1 as the divisor to avoid division by zero and overflow. |
| 1320 | SDValue BinOp = |
| 1321 | DAG.getNode(Opcode: ISD::getUnmaskedBinOpOpcode(MaskedOpc: N->getOpcode()), DL, VT, N1: LHS, |
| 1322 | N2: DAG.getSelect(DL, VT, Cond: Mask, LHS: RHS, RHS: DAG.getConstant(Val: 1, DL, VT))); |
| 1323 | return DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL, VT: N->getValueType(ResNo: 0), Operand: BinOp); |
| 1324 | } |
| 1325 | |
| 1326 | //===----------------------------------------------------------------------===// |
| 1327 | // Result Vector Splitting |
| 1328 | //===----------------------------------------------------------------------===// |
| 1329 | |
| 1330 | /// This method is called when the specified result of the specified node is |
| 1331 | /// found to need vector splitting. At this point, the node may also have |
| 1332 | /// invalid operands or may have other results that need legalization, we just |
| 1333 | /// know that (at least) one result needs vector splitting. |
| 1334 | void DAGTypeLegalizer::SplitVectorResult(SDNode *N, unsigned ResNo) { |
| 1335 | LLVM_DEBUG(dbgs() << "Split node result: " ; N->dump(&DAG)); |
| 1336 | SDValue Lo, Hi; |
| 1337 | |
| 1338 | // See if the target wants to custom expand this node. |
| 1339 | if (CustomLowerNode(N, VT: N->getValueType(ResNo), LegalizeResult: true)) |
| 1340 | return; |
| 1341 | |
| 1342 | switch (N->getOpcode()) { |
| 1343 | default: |
| 1344 | #ifndef NDEBUG |
| 1345 | dbgs() << "SplitVectorResult #" << ResNo << ": " ; |
| 1346 | N->dump(&DAG); |
| 1347 | dbgs() << "\n" ; |
| 1348 | #endif |
| 1349 | report_fatal_error(reason: "Do not know how to split the result of this " |
| 1350 | "operator!\n" ); |
| 1351 | |
| 1352 | case ISD::LOOP_DEPENDENCE_RAW_MASK: |
| 1353 | case ISD::LOOP_DEPENDENCE_WAR_MASK: |
| 1354 | SplitVecRes_LOOP_DEPENDENCE_MASK(N, Lo, Hi); |
| 1355 | break; |
| 1356 | case ISD::MERGE_VALUES: SplitRes_MERGE_VALUES(N, ResNo, Lo, Hi); break; |
| 1357 | case ISD::AssertZext: SplitVecRes_AssertZext(N, Lo, Hi); break; |
| 1358 | case ISD::AssertSext: SplitVecRes_AssertSext(N, Lo, Hi); break; |
| 1359 | case ISD::VSELECT: |
| 1360 | case ISD::SELECT: |
| 1361 | case ISD::VP_MERGE: |
| 1362 | case ISD::VP_SELECT: SplitRes_Select(N, Lo, Hi); break; |
| 1363 | case ISD::SELECT_CC: SplitRes_SELECT_CC(N, Lo, Hi); break; |
| 1364 | case ISD::POISON: |
| 1365 | case ISD::UNDEF: SplitRes_UNDEF(N, Lo, Hi); break; |
| 1366 | case ISD::BITCAST: SplitVecRes_BITCAST(N, Lo, Hi); break; |
| 1367 | case ISD::BUILD_VECTOR: SplitVecRes_BUILD_VECTOR(N, Lo, Hi); break; |
| 1368 | case ISD::CONCAT_VECTORS: SplitVecRes_CONCAT_VECTORS(N, Lo, Hi); break; |
| 1369 | case ISD::EXTRACT_SUBVECTOR: SplitVecRes_EXTRACT_SUBVECTOR(N, Lo, Hi); break; |
| 1370 | case ISD::INSERT_SUBVECTOR: SplitVecRes_INSERT_SUBVECTOR(N, Lo, Hi); break; |
| 1371 | case ISD::FPOWI: |
| 1372 | case ISD::FLDEXP: |
| 1373 | case ISD::FCOPYSIGN: SplitVecRes_FPOp_MultiType(N, Lo, Hi); break; |
| 1374 | case ISD::IS_FPCLASS: SplitVecRes_IS_FPCLASS(N, Lo, Hi); break; |
| 1375 | case ISD::INSERT_VECTOR_ELT: SplitVecRes_INSERT_VECTOR_ELT(N, Lo, Hi); break; |
| 1376 | case ISD::SPLAT_VECTOR: |
| 1377 | case ISD::SCALAR_TO_VECTOR: |
| 1378 | SplitVecRes_ScalarOp(N, Lo, Hi); |
| 1379 | break; |
| 1380 | case ISD::STEP_VECTOR: |
| 1381 | SplitVecRes_STEP_VECTOR(N, Lo, Hi); |
| 1382 | break; |
| 1383 | case ISD::SIGN_EXTEND_INREG: SplitVecRes_InregOp(N, Lo, Hi); break; |
| 1384 | case ISD::ATOMIC_LOAD: |
| 1385 | SplitVecRes_ATOMIC_LOAD(LD: cast<AtomicSDNode>(Val: N), Lo, Hi); |
| 1386 | break; |
| 1387 | case ISD::LOAD: |
| 1388 | SplitVecRes_LOAD(LD: cast<LoadSDNode>(Val: N), Lo, Hi); |
| 1389 | break; |
| 1390 | case ISD::VP_LOAD: |
| 1391 | SplitVecRes_VP_LOAD(LD: cast<VPLoadSDNode>(Val: N), Lo, Hi); |
| 1392 | break; |
| 1393 | case ISD::VP_LOAD_FF: |
| 1394 | SplitVecRes_VP_LOAD_FF(LD: cast<VPLoadFFSDNode>(Val: N), Lo, Hi); |
| 1395 | break; |
| 1396 | case ISD::EXPERIMENTAL_VP_STRIDED_LOAD: |
| 1397 | SplitVecRes_VP_STRIDED_LOAD(SLD: cast<VPStridedLoadSDNode>(Val: N), Lo, Hi); |
| 1398 | break; |
| 1399 | case ISD::MLOAD: |
| 1400 | SplitVecRes_MLOAD(MLD: cast<MaskedLoadSDNode>(Val: N), Lo, Hi); |
| 1401 | break; |
| 1402 | case ISD::MGATHER: |
| 1403 | case ISD::VP_GATHER: |
| 1404 | SplitVecRes_Gather(VPGT: cast<MemSDNode>(Val: N), Lo, Hi, /*SplitSETCC*/ true); |
| 1405 | break; |
| 1406 | case ISD::VECTOR_COMPRESS: |
| 1407 | SplitVecRes_VECTOR_COMPRESS(N, Lo, Hi); |
| 1408 | break; |
| 1409 | case ISD::SETCC: |
| 1410 | case ISD::VP_SETCC: |
| 1411 | SplitVecRes_SETCC(N, Lo, Hi); |
| 1412 | break; |
| 1413 | case ISD::VECTOR_REVERSE: |
| 1414 | SplitVecRes_VECTOR_REVERSE(N, Lo, Hi); |
| 1415 | break; |
| 1416 | case ISD::VECTOR_SHUFFLE: |
| 1417 | SplitVecRes_VECTOR_SHUFFLE(N: cast<ShuffleVectorSDNode>(Val: N), Lo, Hi); |
| 1418 | break; |
| 1419 | case ISD::VECTOR_SPLICE_LEFT: |
| 1420 | case ISD::VECTOR_SPLICE_RIGHT: |
| 1421 | SplitVecRes_VECTOR_SPLICE(N, Lo, Hi); |
| 1422 | break; |
| 1423 | case ISD::VECTOR_DEINTERLEAVE: |
| 1424 | SplitVecRes_VECTOR_DEINTERLEAVE(N); |
| 1425 | return; |
| 1426 | case ISD::VECTOR_INTERLEAVE: |
| 1427 | SplitVecRes_VECTOR_INTERLEAVE(N); |
| 1428 | return; |
| 1429 | case ISD::VAARG: |
| 1430 | SplitVecRes_VAARG(N, Lo, Hi); |
| 1431 | break; |
| 1432 | |
| 1433 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 1434 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 1435 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 1436 | SplitVecRes_ExtVecInRegOp(N, Lo, Hi); |
| 1437 | break; |
| 1438 | |
| 1439 | case ISD::ABS: |
| 1440 | case ISD::ABS_MIN_POISON: |
| 1441 | case ISD::VP_ABS: |
| 1442 | case ISD::BITREVERSE: |
| 1443 | case ISD::VP_BITREVERSE: |
| 1444 | case ISD::BSWAP: |
| 1445 | case ISD::VP_BSWAP: |
| 1446 | case ISD::CTLZ: |
| 1447 | case ISD::VP_CTLZ: |
| 1448 | case ISD::CTTZ: |
| 1449 | case ISD::VP_CTTZ: |
| 1450 | case ISD::CTLZ_ZERO_POISON: |
| 1451 | case ISD::VP_CTLZ_ZERO_POISON: |
| 1452 | case ISD::CTTZ_ZERO_POISON: |
| 1453 | case ISD::VP_CTTZ_ZERO_POISON: |
| 1454 | case ISD::CTPOP: |
| 1455 | case ISD::VP_CTPOP: |
| 1456 | case ISD::FABS: case ISD::VP_FABS: |
| 1457 | case ISD::FACOS: |
| 1458 | case ISD::FASIN: |
| 1459 | case ISD::FATAN: |
| 1460 | case ISD::FCEIL: |
| 1461 | case ISD::VP_FCEIL: |
| 1462 | case ISD::FCOS: |
| 1463 | case ISD::FCOSH: |
| 1464 | case ISD::FEXP: |
| 1465 | case ISD::FEXP2: |
| 1466 | case ISD::FEXP10: |
| 1467 | case ISD::FFLOOR: |
| 1468 | case ISD::VP_FFLOOR: |
| 1469 | case ISD::FLOG: |
| 1470 | case ISD::FLOG10: |
| 1471 | case ISD::FLOG2: |
| 1472 | case ISD::FNEARBYINT: |
| 1473 | case ISD::VP_FNEARBYINT: |
| 1474 | case ISD::FNEG: case ISD::VP_FNEG: |
| 1475 | case ISD::FREEZE: |
| 1476 | case ISD::ARITH_FENCE: |
| 1477 | case ISD::FP_EXTEND: |
| 1478 | case ISD::VP_FP_EXTEND: |
| 1479 | case ISD::FP_ROUND: |
| 1480 | case ISD::VP_FP_ROUND: |
| 1481 | case ISD::FP_TO_SINT: |
| 1482 | case ISD::VP_FP_TO_SINT: |
| 1483 | case ISD::FP_TO_UINT: |
| 1484 | case ISD::VP_FP_TO_UINT: |
| 1485 | case ISD::FRINT: |
| 1486 | case ISD::VP_FRINT: |
| 1487 | case ISD::LRINT: |
| 1488 | case ISD::VP_LRINT: |
| 1489 | case ISD::LLRINT: |
| 1490 | case ISD::VP_LLRINT: |
| 1491 | case ISD::FROUND: |
| 1492 | case ISD::VP_FROUND: |
| 1493 | case ISD::FROUNDEVEN: |
| 1494 | case ISD::VP_FROUNDEVEN: |
| 1495 | case ISD::LROUND: |
| 1496 | case ISD::LLROUND: |
| 1497 | case ISD::FSIN: |
| 1498 | case ISD::FSINH: |
| 1499 | case ISD::FSQRT: case ISD::VP_SQRT: |
| 1500 | case ISD::FTAN: |
| 1501 | case ISD::FTANH: |
| 1502 | case ISD::FTRUNC: |
| 1503 | case ISD::VP_FROUNDTOZERO: |
| 1504 | case ISD::SINT_TO_FP: |
| 1505 | case ISD::VP_SINT_TO_FP: |
| 1506 | case ISD::TRUNCATE: |
| 1507 | case ISD::VP_TRUNCATE: |
| 1508 | case ISD::UINT_TO_FP: |
| 1509 | case ISD::VP_UINT_TO_FP: |
| 1510 | case ISD::FCANONICALIZE: |
| 1511 | case ISD::AssertNoFPClass: |
| 1512 | case ISD::CONVERT_FROM_ARBITRARY_FP: |
| 1513 | case ISD::CONVERT_TO_ARBITRARY_FP: |
| 1514 | SplitVecRes_UnaryOp(N, Lo, Hi); |
| 1515 | break; |
| 1516 | case ISD::ADDRSPACECAST: |
| 1517 | SplitVecRes_ADDRSPACECAST(N, Lo, Hi); |
| 1518 | break; |
| 1519 | case ISD::FMODF: |
| 1520 | case ISD::FFREXP: |
| 1521 | case ISD::FSINCOS: |
| 1522 | case ISD::FSINCOSPI: |
| 1523 | SplitVecRes_UnaryOpWithTwoResults(N, ResNo, Lo, Hi); |
| 1524 | break; |
| 1525 | |
| 1526 | case ISD::ANY_EXTEND: |
| 1527 | case ISD::SIGN_EXTEND: |
| 1528 | case ISD::ZERO_EXTEND: |
| 1529 | case ISD::VP_SIGN_EXTEND: |
| 1530 | case ISD::VP_ZERO_EXTEND: |
| 1531 | SplitVecRes_ExtendOp(N, Lo, Hi); |
| 1532 | break; |
| 1533 | |
| 1534 | case ISD::ADD: case ISD::VP_ADD: |
| 1535 | case ISD::SUB: case ISD::VP_SUB: |
| 1536 | case ISD::MUL: case ISD::VP_MUL: |
| 1537 | case ISD::CLMUL: |
| 1538 | case ISD::CLMULR: |
| 1539 | case ISD::CLMULH: |
| 1540 | case ISD::PEXT: |
| 1541 | case ISD::PDEP: |
| 1542 | case ISD::MULHS: |
| 1543 | case ISD::MULHU: |
| 1544 | case ISD::ABDS: |
| 1545 | case ISD::ABDU: |
| 1546 | case ISD::AVGCEILS: |
| 1547 | case ISD::AVGCEILU: |
| 1548 | case ISD::AVGFLOORS: |
| 1549 | case ISD::AVGFLOORU: |
| 1550 | case ISD::FADD: case ISD::VP_FADD: |
| 1551 | case ISD::FSUB: case ISD::VP_FSUB: |
| 1552 | case ISD::FMUL: case ISD::VP_FMUL: |
| 1553 | case ISD::FMINNUM: |
| 1554 | case ISD::FMINNUM_IEEE: |
| 1555 | case ISD::VP_FMINNUM: |
| 1556 | case ISD::FMAXNUM: |
| 1557 | case ISD::FMAXNUM_IEEE: |
| 1558 | case ISD::VP_FMAXNUM: |
| 1559 | case ISD::FMINIMUM: |
| 1560 | case ISD::VP_FMINIMUM: |
| 1561 | case ISD::FMAXIMUM: |
| 1562 | case ISD::VP_FMAXIMUM: |
| 1563 | case ISD::FMINIMUMNUM: |
| 1564 | case ISD::FMAXIMUMNUM: |
| 1565 | case ISD::SDIV: case ISD::VP_SDIV: |
| 1566 | case ISD::UDIV: case ISD::VP_UDIV: |
| 1567 | case ISD::FDIV: case ISD::VP_FDIV: |
| 1568 | case ISD::FPOW: |
| 1569 | case ISD::FATAN2: |
| 1570 | case ISD::AND: case ISD::VP_AND: |
| 1571 | case ISD::OR: case ISD::VP_OR: |
| 1572 | case ISD::XOR: case ISD::VP_XOR: |
| 1573 | case ISD::SHL: case ISD::VP_SHL: |
| 1574 | case ISD::SRA: case ISD::VP_SRA: |
| 1575 | case ISD::SRL: case ISD::VP_SRL: |
| 1576 | case ISD::UREM: case ISD::VP_UREM: |
| 1577 | case ISD::SREM: case ISD::VP_SREM: |
| 1578 | case ISD::FREM: case ISD::VP_FREM: |
| 1579 | case ISD::SMIN: case ISD::VP_SMIN: |
| 1580 | case ISD::SMAX: case ISD::VP_SMAX: |
| 1581 | case ISD::UMIN: case ISD::VP_UMIN: |
| 1582 | case ISD::UMAX: case ISD::VP_UMAX: |
| 1583 | case ISD::SADDSAT: case ISD::VP_SADDSAT: |
| 1584 | case ISD::UADDSAT: case ISD::VP_UADDSAT: |
| 1585 | case ISD::SSUBSAT: case ISD::VP_SSUBSAT: |
| 1586 | case ISD::USUBSAT: case ISD::VP_USUBSAT: |
| 1587 | case ISD::SSHLSAT: |
| 1588 | case ISD::USHLSAT: |
| 1589 | case ISD::ROTL: |
| 1590 | case ISD::ROTR: |
| 1591 | case ISD::VP_FCOPYSIGN: |
| 1592 | SplitVecRes_BinOp(N, Lo, Hi); |
| 1593 | break; |
| 1594 | case ISD::MASKED_UDIV: |
| 1595 | case ISD::MASKED_SDIV: |
| 1596 | case ISD::MASKED_UREM: |
| 1597 | case ISD::MASKED_SREM: |
| 1598 | SplitVecRes_MaskedBinOp(N, Lo, Hi); |
| 1599 | break; |
| 1600 | case ISD::FMA: case ISD::VP_FMA: |
| 1601 | case ISD::FSHL: |
| 1602 | case ISD::VP_FSHL: |
| 1603 | case ISD::FSHR: |
| 1604 | case ISD::VP_FSHR: |
| 1605 | SplitVecRes_TernaryOp(N, Lo, Hi); |
| 1606 | break; |
| 1607 | |
| 1608 | case ISD::SCMP: case ISD::UCMP: |
| 1609 | SplitVecRes_CMP(N, Lo, Hi); |
| 1610 | break; |
| 1611 | |
| 1612 | #define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \ |
| 1613 | case ISD::STRICT_##DAGN: |
| 1614 | #include "llvm/IR/ConstrainedOps.def" |
| 1615 | SplitVecRes_StrictFPOp(N, Lo, Hi); |
| 1616 | break; |
| 1617 | |
| 1618 | case ISD::FP_TO_UINT_SAT: |
| 1619 | case ISD::FP_TO_SINT_SAT: |
| 1620 | SplitVecRes_FP_TO_XINT_SAT(N, Lo, Hi); |
| 1621 | break; |
| 1622 | |
| 1623 | case ISD::UADDO: |
| 1624 | case ISD::SADDO: |
| 1625 | case ISD::USUBO: |
| 1626 | case ISD::SSUBO: |
| 1627 | case ISD::UMULO: |
| 1628 | case ISD::SMULO: |
| 1629 | SplitVecRes_OverflowOp(N, ResNo, Lo, Hi); |
| 1630 | break; |
| 1631 | case ISD::SMULFIX: |
| 1632 | case ISD::SMULFIXSAT: |
| 1633 | case ISD::UMULFIX: |
| 1634 | case ISD::UMULFIXSAT: |
| 1635 | case ISD::SDIVFIX: |
| 1636 | case ISD::SDIVFIXSAT: |
| 1637 | case ISD::UDIVFIX: |
| 1638 | case ISD::UDIVFIXSAT: |
| 1639 | SplitVecRes_FIX(N, Lo, Hi); |
| 1640 | break; |
| 1641 | case ISD::EXPERIMENTAL_VP_SPLICE: |
| 1642 | SplitVecRes_VP_SPLICE(N, Lo, Hi); |
| 1643 | break; |
| 1644 | case ISD::EXPERIMENTAL_VP_REVERSE: |
| 1645 | SplitVecRes_VP_REVERSE(N, Lo, Hi); |
| 1646 | break; |
| 1647 | case ISD::PARTIAL_REDUCE_UMLA: |
| 1648 | case ISD::PARTIAL_REDUCE_SMLA: |
| 1649 | case ISD::PARTIAL_REDUCE_SUMLA: |
| 1650 | case ISD::PARTIAL_REDUCE_FMLA: |
| 1651 | SplitVecRes_PARTIAL_REDUCE_MLA(N, Lo, Hi); |
| 1652 | break; |
| 1653 | case ISD::GET_ACTIVE_LANE_MASK: |
| 1654 | SplitVecRes_GET_ACTIVE_LANE_MASK(N, Lo, Hi); |
| 1655 | break; |
| 1656 | } |
| 1657 | |
| 1658 | // If Lo/Hi is null, the sub-method took care of registering results etc. |
| 1659 | if (Lo.getNode()) |
| 1660 | SetSplitVector(Op: SDValue(N, ResNo), Lo, Hi); |
| 1661 | } |
| 1662 | |
| 1663 | void DAGTypeLegalizer::IncrementPointer(MemSDNode *N, EVT MemVT, |
| 1664 | MachinePointerInfo &MPI, SDValue &Ptr, |
| 1665 | uint64_t *ScaledOffset) { |
| 1666 | SDLoc DL(N); |
| 1667 | unsigned IncrementSize = MemVT.getSizeInBits().getKnownMinValue() / 8; |
| 1668 | |
| 1669 | if (MemVT.isScalableVector()) { |
| 1670 | SDValue BytesIncrement = DAG.getVScale( |
| 1671 | DL, VT: Ptr.getValueType(), |
| 1672 | MulImm: APInt(Ptr.getValueSizeInBits().getFixedValue(), IncrementSize)); |
| 1673 | MPI = MachinePointerInfo(N->getPointerInfo().getAddrSpace()); |
| 1674 | if (ScaledOffset) |
| 1675 | *ScaledOffset += IncrementSize; |
| 1676 | Ptr = DAG.getNode(Opcode: ISD::ADD, DL, VT: Ptr.getValueType(), N1: Ptr, N2: BytesIncrement, |
| 1677 | Flags: SDNodeFlags::NoUnsignedWrap); |
| 1678 | } else { |
| 1679 | MPI = N->getPointerInfo().getWithOffset(O: IncrementSize); |
| 1680 | // Increment the pointer to the other half. |
| 1681 | Ptr = DAG.getObjectPtrOffset(SL: DL, Ptr, Offset: TypeSize::getFixed(ExactSize: IncrementSize)); |
| 1682 | } |
| 1683 | } |
| 1684 | |
| 1685 | std::pair<SDValue, SDValue> DAGTypeLegalizer::SplitMask(SDValue Mask) { |
| 1686 | return SplitMask(Mask, DL: SDLoc(Mask)); |
| 1687 | } |
| 1688 | |
| 1689 | std::pair<SDValue, SDValue> DAGTypeLegalizer::SplitMask(SDValue Mask, |
| 1690 | const SDLoc &DL) { |
| 1691 | SDValue MaskLo, MaskHi; |
| 1692 | EVT MaskVT = Mask.getValueType(); |
| 1693 | if (getTypeAction(VT: MaskVT) == TargetLowering::TypeSplitVector) |
| 1694 | GetSplitVector(Op: Mask, Lo&: MaskLo, Hi&: MaskHi); |
| 1695 | else |
| 1696 | std::tie(args&: MaskLo, args&: MaskHi) = DAG.SplitVector(N: Mask, DL); |
| 1697 | return std::make_pair(x&: MaskLo, y&: MaskHi); |
| 1698 | } |
| 1699 | |
| 1700 | void DAGTypeLegalizer::SplitVecRes_BinOp(SDNode *N, SDValue &Lo, SDValue &Hi) { |
| 1701 | SDValue LHSLo, LHSHi; |
| 1702 | GetSplitVector(Op: N->getOperand(Num: 0), Lo&: LHSLo, Hi&: LHSHi); |
| 1703 | SDValue RHSLo, RHSHi; |
| 1704 | GetSplitVector(Op: N->getOperand(Num: 1), Lo&: RHSLo, Hi&: RHSHi); |
| 1705 | SDLoc dl(N); |
| 1706 | |
| 1707 | const SDNodeFlags Flags = N->getFlags(); |
| 1708 | unsigned Opcode = N->getOpcode(); |
| 1709 | if (N->getNumOperands() == 2) { |
| 1710 | Lo = DAG.getNode(Opcode, DL: dl, VT: LHSLo.getValueType(), N1: LHSLo, N2: RHSLo, Flags); |
| 1711 | Hi = DAG.getNode(Opcode, DL: dl, VT: LHSHi.getValueType(), N1: LHSHi, N2: RHSHi, Flags); |
| 1712 | return; |
| 1713 | } |
| 1714 | |
| 1715 | assert(N->getNumOperands() == 4 && "Unexpected number of operands!" ); |
| 1716 | assert(N->isVPOpcode() && "Expected VP opcode" ); |
| 1717 | |
| 1718 | SDValue MaskLo, MaskHi; |
| 1719 | std::tie(args&: MaskLo, args&: MaskHi) = SplitMask(Mask: N->getOperand(Num: 2)); |
| 1720 | |
| 1721 | SDValue EVLLo, EVLHi; |
| 1722 | std::tie(args&: EVLLo, args&: EVLHi) = |
| 1723 | DAG.SplitEVL(N: N->getOperand(Num: 3), VecVT: N->getValueType(ResNo: 0), DL: dl); |
| 1724 | |
| 1725 | Lo = DAG.getNode(Opcode, DL: dl, VT: LHSLo.getValueType(), |
| 1726 | Ops: {LHSLo, RHSLo, MaskLo, EVLLo}, Flags); |
| 1727 | Hi = DAG.getNode(Opcode, DL: dl, VT: LHSHi.getValueType(), |
| 1728 | Ops: {LHSHi, RHSHi, MaskHi, EVLHi}, Flags); |
| 1729 | } |
| 1730 | |
| 1731 | void DAGTypeLegalizer::SplitVecRes_MaskedBinOp(SDNode *N, SDValue &Lo, |
| 1732 | SDValue &Hi) { |
| 1733 | SDValue LHSLo, LHSHi; |
| 1734 | GetSplitVector(Op: N->getOperand(Num: 0), Lo&: LHSLo, Hi&: LHSHi); |
| 1735 | SDValue RHSLo, RHSHi; |
| 1736 | GetSplitVector(Op: N->getOperand(Num: 1), Lo&: RHSLo, Hi&: RHSHi); |
| 1737 | auto [MaskLo, MaskHi] = SplitMask(Mask: N->getOperand(Num: 2)); |
| 1738 | SDLoc dl(N); |
| 1739 | |
| 1740 | const SDNodeFlags Flags = N->getFlags(); |
| 1741 | unsigned Opcode = N->getOpcode(); |
| 1742 | Lo = DAG.getNode(Opcode, DL: dl, VT: LHSLo.getValueType(), N1: LHSLo, N2: RHSLo, N3: MaskLo, |
| 1743 | Flags); |
| 1744 | Hi = DAG.getNode(Opcode, DL: dl, VT: LHSHi.getValueType(), N1: LHSHi, N2: RHSHi, N3: MaskHi, |
| 1745 | Flags); |
| 1746 | } |
| 1747 | |
| 1748 | void DAGTypeLegalizer::SplitVecRes_TernaryOp(SDNode *N, SDValue &Lo, |
| 1749 | SDValue &Hi) { |
| 1750 | SDValue Op0Lo, Op0Hi; |
| 1751 | GetSplitVector(Op: N->getOperand(Num: 0), Lo&: Op0Lo, Hi&: Op0Hi); |
| 1752 | SDValue Op1Lo, Op1Hi; |
| 1753 | GetSplitVector(Op: N->getOperand(Num: 1), Lo&: Op1Lo, Hi&: Op1Hi); |
| 1754 | SDValue Op2Lo, Op2Hi; |
| 1755 | GetSplitVector(Op: N->getOperand(Num: 2), Lo&: Op2Lo, Hi&: Op2Hi); |
| 1756 | SDLoc dl(N); |
| 1757 | |
| 1758 | const SDNodeFlags Flags = N->getFlags(); |
| 1759 | unsigned Opcode = N->getOpcode(); |
| 1760 | if (N->getNumOperands() == 3) { |
| 1761 | Lo = DAG.getNode(Opcode, DL: dl, VT: Op0Lo.getValueType(), N1: Op0Lo, N2: Op1Lo, N3: Op2Lo, Flags); |
| 1762 | Hi = DAG.getNode(Opcode, DL: dl, VT: Op0Hi.getValueType(), N1: Op0Hi, N2: Op1Hi, N3: Op2Hi, Flags); |
| 1763 | return; |
| 1764 | } |
| 1765 | |
| 1766 | assert(N->getNumOperands() == 5 && "Unexpected number of operands!" ); |
| 1767 | assert(N->isVPOpcode() && "Expected VP opcode" ); |
| 1768 | |
| 1769 | SDValue MaskLo, MaskHi; |
| 1770 | std::tie(args&: MaskLo, args&: MaskHi) = SplitMask(Mask: N->getOperand(Num: 3)); |
| 1771 | |
| 1772 | SDValue EVLLo, EVLHi; |
| 1773 | std::tie(args&: EVLLo, args&: EVLHi) = |
| 1774 | DAG.SplitEVL(N: N->getOperand(Num: 4), VecVT: N->getValueType(ResNo: 0), DL: dl); |
| 1775 | |
| 1776 | Lo = DAG.getNode(Opcode, DL: dl, VT: Op0Lo.getValueType(), |
| 1777 | Ops: {Op0Lo, Op1Lo, Op2Lo, MaskLo, EVLLo}, Flags); |
| 1778 | Hi = DAG.getNode(Opcode, DL: dl, VT: Op0Hi.getValueType(), |
| 1779 | Ops: {Op0Hi, Op1Hi, Op2Hi, MaskHi, EVLHi}, Flags); |
| 1780 | } |
| 1781 | |
| 1782 | void DAGTypeLegalizer::SplitVecRes_CMP(SDNode *N, SDValue &Lo, SDValue &Hi) { |
| 1783 | LLVMContext &Ctxt = *DAG.getContext(); |
| 1784 | SDLoc dl(N); |
| 1785 | |
| 1786 | SDValue LHS = N->getOperand(Num: 0); |
| 1787 | SDValue RHS = N->getOperand(Num: 1); |
| 1788 | |
| 1789 | SDValue LHSLo, LHSHi, RHSLo, RHSHi; |
| 1790 | if (getTypeAction(VT: LHS.getValueType()) == TargetLowering::TypeSplitVector) { |
| 1791 | GetSplitVector(Op: LHS, Lo&: LHSLo, Hi&: LHSHi); |
| 1792 | GetSplitVector(Op: RHS, Lo&: RHSLo, Hi&: RHSHi); |
| 1793 | } else { |
| 1794 | std::tie(args&: LHSLo, args&: LHSHi) = DAG.SplitVector(N: LHS, DL: dl); |
| 1795 | std::tie(args&: RHSLo, args&: RHSHi) = DAG.SplitVector(N: RHS, DL: dl); |
| 1796 | } |
| 1797 | |
| 1798 | EVT SplitResVT = N->getValueType(ResNo: 0).getHalfNumVectorElementsVT(Context&: Ctxt); |
| 1799 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: SplitResVT, N1: LHSLo, N2: RHSLo); |
| 1800 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: SplitResVT, N1: LHSHi, N2: RHSHi); |
| 1801 | } |
| 1802 | |
| 1803 | void DAGTypeLegalizer::SplitVecRes_FIX(SDNode *N, SDValue &Lo, SDValue &Hi) { |
| 1804 | SDValue LHSLo, LHSHi; |
| 1805 | GetSplitVector(Op: N->getOperand(Num: 0), Lo&: LHSLo, Hi&: LHSHi); |
| 1806 | SDValue RHSLo, RHSHi; |
| 1807 | GetSplitVector(Op: N->getOperand(Num: 1), Lo&: RHSLo, Hi&: RHSHi); |
| 1808 | SDLoc dl(N); |
| 1809 | SDValue Op2 = N->getOperand(Num: 2); |
| 1810 | |
| 1811 | unsigned Opcode = N->getOpcode(); |
| 1812 | Lo = DAG.getNode(Opcode, DL: dl, VT: LHSLo.getValueType(), N1: LHSLo, N2: RHSLo, N3: Op2, |
| 1813 | Flags: N->getFlags()); |
| 1814 | Hi = DAG.getNode(Opcode, DL: dl, VT: LHSHi.getValueType(), N1: LHSHi, N2: RHSHi, N3: Op2, |
| 1815 | Flags: N->getFlags()); |
| 1816 | } |
| 1817 | |
| 1818 | void DAGTypeLegalizer::SplitVecRes_BITCAST(SDNode *N, SDValue &Lo, |
| 1819 | SDValue &Hi) { |
| 1820 | // We know the result is a vector. The input may be either a vector or a |
| 1821 | // scalar value. |
| 1822 | EVT LoVT, HiVT; |
| 1823 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 1824 | SDLoc dl(N); |
| 1825 | |
| 1826 | SDValue InOp = N->getOperand(Num: 0); |
| 1827 | EVT InVT = InOp.getValueType(); |
| 1828 | |
| 1829 | // Handle some special cases efficiently. |
| 1830 | switch (getTypeAction(VT: InVT)) { |
| 1831 | case TargetLowering::TypeLegal: |
| 1832 | case TargetLowering::TypePromoteInteger: |
| 1833 | case TargetLowering::TypeSoftPromoteHalf: |
| 1834 | case TargetLowering::TypeSoftenFloat: |
| 1835 | case TargetLowering::TypeScalarizeVector: |
| 1836 | case TargetLowering::TypeWidenVector: |
| 1837 | break; |
| 1838 | case TargetLowering::TypeExpandInteger: |
| 1839 | case TargetLowering::TypeExpandFloat: |
| 1840 | // A scalar to vector conversion, where the scalar needs expansion. |
| 1841 | // If the vector is being split in two then we can just convert the |
| 1842 | // expanded pieces. |
| 1843 | if (LoVT == HiVT) { |
| 1844 | GetExpandedOp(Op: InOp, Lo, Hi); |
| 1845 | if (DAG.getDataLayout().isBigEndian()) |
| 1846 | std::swap(a&: Lo, b&: Hi); |
| 1847 | Lo = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: LoVT, Operand: Lo); |
| 1848 | Hi = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: HiVT, Operand: Hi); |
| 1849 | return; |
| 1850 | } |
| 1851 | break; |
| 1852 | case TargetLowering::TypeSplitVector: |
| 1853 | // If the input is a vector that needs to be split, convert each split |
| 1854 | // piece of the input now. |
| 1855 | GetSplitVector(Op: InOp, Lo, Hi); |
| 1856 | Lo = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: LoVT, Operand: Lo); |
| 1857 | Hi = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: HiVT, Operand: Hi); |
| 1858 | return; |
| 1859 | case TargetLowering::TypeScalarizeScalableVector: |
| 1860 | report_fatal_error(reason: "Scalarization of scalable vectors is not supported." ); |
| 1861 | } |
| 1862 | |
| 1863 | if (LoVT.isScalableVector()) { |
| 1864 | auto [InLo, InHi] = DAG.SplitVectorOperand(N, OpNo: 0); |
| 1865 | Lo = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: LoVT, Operand: InLo); |
| 1866 | Hi = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: HiVT, Operand: InHi); |
| 1867 | return; |
| 1868 | } |
| 1869 | |
| 1870 | // In the general case, convert the input to an integer and split it by hand. |
| 1871 | EVT LoIntVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: LoVT.getSizeInBits()); |
| 1872 | EVT HiIntVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: HiVT.getSizeInBits()); |
| 1873 | if (DAG.getDataLayout().isBigEndian()) |
| 1874 | std::swap(a&: LoIntVT, b&: HiIntVT); |
| 1875 | |
| 1876 | SplitInteger(Op: BitConvertToInteger(Op: InOp), LoVT: LoIntVT, HiVT: HiIntVT, Lo, Hi); |
| 1877 | |
| 1878 | if (DAG.getDataLayout().isBigEndian()) |
| 1879 | std::swap(a&: Lo, b&: Hi); |
| 1880 | Lo = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: LoVT, Operand: Lo); |
| 1881 | Hi = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: HiVT, Operand: Hi); |
| 1882 | } |
| 1883 | |
| 1884 | void DAGTypeLegalizer::SplitVecRes_LOOP_DEPENDENCE_MASK(SDNode *N, SDValue &Lo, |
| 1885 | SDValue &Hi) { |
| 1886 | SDLoc DL(N); |
| 1887 | EVT LoVT, HiVT; |
| 1888 | SDValue PtrA = N->getOperand(Num: 0); |
| 1889 | SDValue PtrB = N->getOperand(Num: 1); |
| 1890 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 1891 | |
| 1892 | // The lane offset for the "Lo" half of the mask is unchanged. |
| 1893 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL, VT: LoVT, N1: PtrA, N2: PtrB, |
| 1894 | /*ElementSizeInBytes=*/N3: N->getOperand(Num: 2), |
| 1895 | /*LaneOffset=*/N4: N->getOperand(Num: 3)); |
| 1896 | // The lane offset for the "Hi" half of the mask is incremented by the number |
| 1897 | // of elements in the "Lo" half. |
| 1898 | unsigned LaneOffset = |
| 1899 | N->getConstantOperandVal(Num: 3) + LoVT.getVectorMinNumElements(); |
| 1900 | // Note: The lane offset is implicitly scalable for scalable masks. |
| 1901 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL, VT: HiVT, N1: PtrA, N2: PtrB, |
| 1902 | /*ElementSizeInBytes=*/N3: N->getOperand(Num: 2), |
| 1903 | /*LaneOffset=*/N4: DAG.getConstant(Val: LaneOffset, DL, VT: MVT::i64)); |
| 1904 | } |
| 1905 | |
| 1906 | void DAGTypeLegalizer::SplitVecRes_BUILD_VECTOR(SDNode *N, SDValue &Lo, |
| 1907 | SDValue &Hi) { |
| 1908 | EVT LoVT, HiVT; |
| 1909 | SDLoc dl(N); |
| 1910 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 1911 | unsigned LoNumElts = LoVT.getVectorNumElements(); |
| 1912 | SmallVector<SDValue, 8> LoOps(N->op_begin(), N->op_begin()+LoNumElts); |
| 1913 | Lo = DAG.getBuildVector(VT: LoVT, DL: dl, Ops: LoOps); |
| 1914 | |
| 1915 | SmallVector<SDValue, 8> HiOps(N->op_begin()+LoNumElts, N->op_end()); |
| 1916 | Hi = DAG.getBuildVector(VT: HiVT, DL: dl, Ops: HiOps); |
| 1917 | } |
| 1918 | |
| 1919 | void DAGTypeLegalizer::SplitVecRes_CONCAT_VECTORS(SDNode *N, SDValue &Lo, |
| 1920 | SDValue &Hi) { |
| 1921 | assert(!(N->getNumOperands() & 1) && "Unsupported CONCAT_VECTORS" ); |
| 1922 | SDLoc dl(N); |
| 1923 | unsigned NumSubvectors = N->getNumOperands() / 2; |
| 1924 | if (NumSubvectors == 1) { |
| 1925 | Lo = N->getOperand(Num: 0); |
| 1926 | Hi = N->getOperand(Num: 1); |
| 1927 | return; |
| 1928 | } |
| 1929 | |
| 1930 | EVT LoVT, HiVT; |
| 1931 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 1932 | |
| 1933 | SmallVector<SDValue, 8> LoOps(N->op_begin(), N->op_begin()+NumSubvectors); |
| 1934 | Lo = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: LoVT, Ops: LoOps); |
| 1935 | |
| 1936 | SmallVector<SDValue, 8> HiOps(N->op_begin()+NumSubvectors, N->op_end()); |
| 1937 | Hi = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: HiVT, Ops: HiOps); |
| 1938 | } |
| 1939 | |
| 1940 | void DAGTypeLegalizer::(SDNode *N, SDValue &Lo, |
| 1941 | SDValue &Hi) { |
| 1942 | SDValue Vec = N->getOperand(Num: 0); |
| 1943 | SDValue Idx = N->getOperand(Num: 1); |
| 1944 | SDLoc dl(N); |
| 1945 | |
| 1946 | EVT LoVT, HiVT; |
| 1947 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 1948 | |
| 1949 | Lo = DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL: dl, VT: LoVT, N1: Vec, N2: Idx); |
| 1950 | uint64_t IdxVal = Idx->getAsZExtVal(); |
| 1951 | Hi = DAG.getNode( |
| 1952 | Opcode: ISD::EXTRACT_SUBVECTOR, DL: dl, VT: HiVT, N1: Vec, |
| 1953 | N2: DAG.getVectorIdxConstant(Val: IdxVal + LoVT.getVectorMinNumElements(), DL: dl)); |
| 1954 | } |
| 1955 | |
| 1956 | void DAGTypeLegalizer::SplitVecRes_INSERT_SUBVECTOR(SDNode *N, SDValue &Lo, |
| 1957 | SDValue &Hi) { |
| 1958 | SDValue Vec = N->getOperand(Num: 0); |
| 1959 | SDValue SubVec = N->getOperand(Num: 1); |
| 1960 | SDValue Idx = N->getOperand(Num: 2); |
| 1961 | SDLoc dl(N); |
| 1962 | GetSplitVector(Op: Vec, Lo, Hi); |
| 1963 | |
| 1964 | EVT VecVT = Vec.getValueType(); |
| 1965 | EVT LoVT = Lo.getValueType(); |
| 1966 | EVT SubVecVT = SubVec.getValueType(); |
| 1967 | unsigned VecElems = VecVT.getVectorMinNumElements(); |
| 1968 | unsigned SubElems = SubVecVT.getVectorMinNumElements(); |
| 1969 | unsigned LoElems = LoVT.getVectorMinNumElements(); |
| 1970 | |
| 1971 | // If we know the index is in the first half, and we know the subvector |
| 1972 | // doesn't cross the boundary between the halves, we can avoid spilling the |
| 1973 | // vector, and insert into the lower half of the split vector directly. |
| 1974 | unsigned IdxVal = Idx->getAsZExtVal(); |
| 1975 | if (IdxVal + SubElems <= LoElems) { |
| 1976 | Lo = DAG.getNode(Opcode: ISD::INSERT_SUBVECTOR, DL: dl, VT: LoVT, N1: Lo, N2: SubVec, N3: Idx); |
| 1977 | return; |
| 1978 | } |
| 1979 | // Similarly if the subvector is fully in the high half, but mind that we |
| 1980 | // can't tell whether a fixed-length subvector is fully within the high half |
| 1981 | // of a scalable vector. |
| 1982 | if (VecVT.isScalableVector() == SubVecVT.isScalableVector() && |
| 1983 | IdxVal >= LoElems && IdxVal + SubElems <= VecElems) { |
| 1984 | Hi = DAG.getNode(Opcode: ISD::INSERT_SUBVECTOR, DL: dl, VT: Hi.getValueType(), N1: Hi, N2: SubVec, |
| 1985 | N3: DAG.getVectorIdxConstant(Val: IdxVal - LoElems, DL: dl)); |
| 1986 | return; |
| 1987 | } |
| 1988 | |
| 1989 | if (getTypeAction(VT: SubVecVT) == TargetLowering::TypeWidenVector && |
| 1990 | Vec.isUndef() && SubVecVT.getVectorElementType() == MVT::i1) { |
| 1991 | SDValue WideSubVec = GetWidenedVector(Op: SubVec); |
| 1992 | if (WideSubVec.getValueType() == VecVT) { |
| 1993 | std::tie(args&: Lo, args&: Hi) = DAG.SplitVector(N: WideSubVec, DL: SDLoc(WideSubVec)); |
| 1994 | return; |
| 1995 | } |
| 1996 | } |
| 1997 | |
| 1998 | // Spill the vector to the stack. |
| 1999 | // In cases where the vector is illegal it will be broken down into parts |
| 2000 | // and stored in parts - we should use the alignment for the smallest part. |
| 2001 | Align SmallestAlign = DAG.getReducedAlign(VT: VecVT, /*UseABI=*/false); |
| 2002 | SDValue StackPtr = |
| 2003 | DAG.CreateStackTemporary(Bytes: VecVT.getStoreSize(), Alignment: SmallestAlign); |
| 2004 | auto &MF = DAG.getMachineFunction(); |
| 2005 | auto FrameIndex = cast<FrameIndexSDNode>(Val: StackPtr.getNode())->getIndex(); |
| 2006 | auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: FrameIndex); |
| 2007 | |
| 2008 | SDValue Store = DAG.getStore(Chain: DAG.getEntryNode(), dl, Val: Vec, Ptr: StackPtr, PtrInfo, |
| 2009 | Alignment: SmallestAlign); |
| 2010 | |
| 2011 | // Store the new subvector into the specified index. |
| 2012 | SDValue SubVecPtr = |
| 2013 | TLI.getVectorSubVecPointer(DAG, VecPtr: StackPtr, VecVT, SubVecVT, Index: Idx); |
| 2014 | Store = DAG.getStore(Chain: Store, dl, Val: SubVec, Ptr: SubVecPtr, |
| 2015 | PtrInfo: MachinePointerInfo::getUnknownStack(MF)); |
| 2016 | |
| 2017 | // Load the Lo part from the stack slot. |
| 2018 | Lo = DAG.getLoad(VT: Lo.getValueType(), dl, Chain: Store, Ptr: StackPtr, PtrInfo, |
| 2019 | Alignment: SmallestAlign); |
| 2020 | |
| 2021 | // Increment the pointer to the other part. |
| 2022 | auto *Load = cast<LoadSDNode>(Val&: Lo); |
| 2023 | MachinePointerInfo MPI = Load->getPointerInfo(); |
| 2024 | IncrementPointer(N: Load, MemVT: LoVT, MPI, Ptr&: StackPtr); |
| 2025 | |
| 2026 | // Load the Hi part from the stack slot. |
| 2027 | Hi = DAG.getLoad(VT: Hi.getValueType(), dl, Chain: Store, Ptr: StackPtr, PtrInfo: MPI, Alignment: SmallestAlign); |
| 2028 | } |
| 2029 | |
| 2030 | // Handle splitting an FP where the second operand does not match the first |
| 2031 | // type. The second operand may be a scalar, or a vector that has exactly as |
| 2032 | // many elements as the first |
| 2033 | void DAGTypeLegalizer::SplitVecRes_FPOp_MultiType(SDNode *N, SDValue &Lo, |
| 2034 | SDValue &Hi) { |
| 2035 | SDValue LHSLo, LHSHi; |
| 2036 | GetSplitVector(Op: N->getOperand(Num: 0), Lo&: LHSLo, Hi&: LHSHi); |
| 2037 | SDLoc DL(N); |
| 2038 | |
| 2039 | SDValue RHSLo, RHSHi; |
| 2040 | SDValue RHS = N->getOperand(Num: 1); |
| 2041 | EVT RHSVT = RHS.getValueType(); |
| 2042 | if (RHSVT.isVector()) { |
| 2043 | if (getTypeAction(VT: RHSVT) == TargetLowering::TypeSplitVector) |
| 2044 | GetSplitVector(Op: RHS, Lo&: RHSLo, Hi&: RHSHi); |
| 2045 | else |
| 2046 | std::tie(args&: RHSLo, args&: RHSHi) = DAG.SplitVector(N: RHS, DL: SDLoc(RHS)); |
| 2047 | |
| 2048 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL, VT: LHSLo.getValueType(), N1: LHSLo, N2: RHSLo); |
| 2049 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL, VT: LHSHi.getValueType(), N1: LHSHi, N2: RHSHi); |
| 2050 | } else { |
| 2051 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL, VT: LHSLo.getValueType(), N1: LHSLo, N2: RHS); |
| 2052 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL, VT: LHSHi.getValueType(), N1: LHSHi, N2: RHS); |
| 2053 | } |
| 2054 | } |
| 2055 | |
| 2056 | void DAGTypeLegalizer::SplitVecRes_IS_FPCLASS(SDNode *N, SDValue &Lo, |
| 2057 | SDValue &Hi) { |
| 2058 | SDLoc DL(N); |
| 2059 | SDValue ArgLo, ArgHi; |
| 2060 | SDValue Test = N->getOperand(Num: 1); |
| 2061 | SDValue FpValue = N->getOperand(Num: 0); |
| 2062 | if (getTypeAction(VT: FpValue.getValueType()) == TargetLowering::TypeSplitVector) |
| 2063 | GetSplitVector(Op: FpValue, Lo&: ArgLo, Hi&: ArgHi); |
| 2064 | else |
| 2065 | std::tie(args&: ArgLo, args&: ArgHi) = DAG.SplitVector(N: FpValue, DL: SDLoc(FpValue)); |
| 2066 | EVT LoVT, HiVT; |
| 2067 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 2068 | |
| 2069 | Lo = DAG.getNode(Opcode: ISD::IS_FPCLASS, DL, VT: LoVT, N1: ArgLo, N2: Test, Flags: N->getFlags()); |
| 2070 | Hi = DAG.getNode(Opcode: ISD::IS_FPCLASS, DL, VT: HiVT, N1: ArgHi, N2: Test, Flags: N->getFlags()); |
| 2071 | } |
| 2072 | |
| 2073 | void DAGTypeLegalizer::SplitVecRes_InregOp(SDNode *N, SDValue &Lo, |
| 2074 | SDValue &Hi) { |
| 2075 | SDValue LHSLo, LHSHi; |
| 2076 | GetSplitVector(Op: N->getOperand(Num: 0), Lo&: LHSLo, Hi&: LHSHi); |
| 2077 | SDLoc dl(N); |
| 2078 | |
| 2079 | EVT LoVT, HiVT; |
| 2080 | std::tie(args&: LoVT, args&: HiVT) = |
| 2081 | DAG.GetSplitDestVTs(VT: cast<VTSDNode>(Val: N->getOperand(Num: 1))->getVT()); |
| 2082 | |
| 2083 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: LHSLo.getValueType(), N1: LHSLo, |
| 2084 | N2: DAG.getValueType(LoVT)); |
| 2085 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: LHSHi.getValueType(), N1: LHSHi, |
| 2086 | N2: DAG.getValueType(HiVT)); |
| 2087 | } |
| 2088 | |
| 2089 | void DAGTypeLegalizer::SplitVecRes_ExtVecInRegOp(SDNode *N, SDValue &Lo, |
| 2090 | SDValue &Hi) { |
| 2091 | unsigned Opcode = N->getOpcode(); |
| 2092 | SDValue N0 = N->getOperand(Num: 0); |
| 2093 | |
| 2094 | SDLoc dl(N); |
| 2095 | SDValue InLo, InHi; |
| 2096 | |
| 2097 | if (getTypeAction(VT: N0.getValueType()) == TargetLowering::TypeSplitVector) |
| 2098 | GetSplitVector(Op: N0, Lo&: InLo, Hi&: InHi); |
| 2099 | else |
| 2100 | std::tie(args&: InLo, args&: InHi) = DAG.SplitVectorOperand(N, OpNo: 0); |
| 2101 | |
| 2102 | EVT InLoVT = InLo.getValueType(); |
| 2103 | unsigned InNumElements = InLoVT.getVectorNumElements(); |
| 2104 | |
| 2105 | EVT OutLoVT, OutHiVT; |
| 2106 | std::tie(args&: OutLoVT, args&: OutHiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 2107 | unsigned OutNumElements = OutLoVT.getVectorNumElements(); |
| 2108 | assert((2 * OutNumElements) <= InNumElements && |
| 2109 | "Illegal extend vector in reg split" ); |
| 2110 | |
| 2111 | // *_EXTEND_VECTOR_INREG instructions extend the lowest elements of the |
| 2112 | // input vector (i.e. we only use InLo): |
| 2113 | // OutLo will extend the first OutNumElements from InLo. |
| 2114 | // OutHi will extend the next OutNumElements from InLo. |
| 2115 | |
| 2116 | // Shuffle the elements from InLo for OutHi into the bottom elements to |
| 2117 | // create a 'fake' InHi. |
| 2118 | SmallVector<int, 8> SplitHi(InNumElements, -1); |
| 2119 | for (unsigned i = 0; i != OutNumElements; ++i) |
| 2120 | SplitHi[i] = i + OutNumElements; |
| 2121 | InHi = DAG.getVectorShuffle(VT: InLoVT, dl, N1: InLo, N2: DAG.getPOISON(VT: InLoVT), Mask: SplitHi); |
| 2122 | |
| 2123 | Lo = DAG.getNode(Opcode, DL: dl, VT: OutLoVT, Operand: InLo); |
| 2124 | Hi = DAG.getNode(Opcode, DL: dl, VT: OutHiVT, Operand: InHi); |
| 2125 | } |
| 2126 | |
| 2127 | void DAGTypeLegalizer::SplitVecRes_StrictFPOp(SDNode *N, SDValue &Lo, |
| 2128 | SDValue &Hi) { |
| 2129 | unsigned NumOps = N->getNumOperands(); |
| 2130 | SDValue Chain = N->getOperand(Num: 0); |
| 2131 | EVT LoVT, HiVT; |
| 2132 | SDLoc dl(N); |
| 2133 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 2134 | |
| 2135 | SmallVector<SDValue, 4> OpsLo(NumOps); |
| 2136 | SmallVector<SDValue, 4> OpsHi(NumOps); |
| 2137 | |
| 2138 | // The Chain is the first operand. |
| 2139 | OpsLo[0] = Chain; |
| 2140 | OpsHi[0] = Chain; |
| 2141 | |
| 2142 | // Now process the remaining operands. |
| 2143 | for (unsigned i = 1; i < NumOps; ++i) { |
| 2144 | SDValue Op = N->getOperand(Num: i); |
| 2145 | SDValue OpLo = Op; |
| 2146 | SDValue OpHi = Op; |
| 2147 | |
| 2148 | EVT InVT = Op.getValueType(); |
| 2149 | if (InVT.isVector()) { |
| 2150 | // If the input also splits, handle it directly for a |
| 2151 | // compile time speedup. Otherwise split it by hand. |
| 2152 | if (getTypeAction(VT: InVT) == TargetLowering::TypeSplitVector) |
| 2153 | GetSplitVector(Op, Lo&: OpLo, Hi&: OpHi); |
| 2154 | else |
| 2155 | std::tie(args&: OpLo, args&: OpHi) = DAG.SplitVectorOperand(N, OpNo: i); |
| 2156 | } |
| 2157 | |
| 2158 | OpsLo[i] = OpLo; |
| 2159 | OpsHi[i] = OpHi; |
| 2160 | } |
| 2161 | |
| 2162 | EVT LoValueVTs[] = {LoVT, MVT::Other}; |
| 2163 | EVT HiValueVTs[] = {HiVT, MVT::Other}; |
| 2164 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VTList: DAG.getVTList(VTs: LoValueVTs), Ops: OpsLo, |
| 2165 | Flags: N->getFlags()); |
| 2166 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VTList: DAG.getVTList(VTs: HiValueVTs), Ops: OpsHi, |
| 2167 | Flags: N->getFlags()); |
| 2168 | |
| 2169 | // Build a factor node to remember that this Op is independent of the |
| 2170 | // other one. |
| 2171 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, |
| 2172 | N1: Lo.getValue(R: 1), N2: Hi.getValue(R: 1)); |
| 2173 | |
| 2174 | // Legalize the chain result - switch anything that used the old chain to |
| 2175 | // use the new one. |
| 2176 | ReplaceValueWith(From: SDValue(N, 1), To: Chain); |
| 2177 | } |
| 2178 | |
| 2179 | SDValue DAGTypeLegalizer::UnrollVectorOp_StrictFP(SDNode *N, unsigned ResNE) { |
| 2180 | SDValue Chain = N->getOperand(Num: 0); |
| 2181 | EVT VT = N->getValueType(ResNo: 0); |
| 2182 | unsigned NE = VT.getVectorNumElements(); |
| 2183 | EVT EltVT = VT.getVectorElementType(); |
| 2184 | SDLoc dl(N); |
| 2185 | |
| 2186 | SmallVector<SDValue, 8> Scalars; |
| 2187 | SmallVector<SDValue, 4> Operands(N->getNumOperands()); |
| 2188 | |
| 2189 | // If ResNE is 0, fully unroll the vector op. |
| 2190 | if (ResNE == 0) |
| 2191 | ResNE = NE; |
| 2192 | else if (NE > ResNE) |
| 2193 | NE = ResNE; |
| 2194 | |
| 2195 | //The results of each unrolled operation, including the chain. |
| 2196 | SDVTList ChainVTs = DAG.getVTList(VT1: EltVT, VT2: MVT::Other); |
| 2197 | SmallVector<SDValue, 8> Chains; |
| 2198 | |
| 2199 | unsigned i; |
| 2200 | for (i = 0; i != NE; ++i) { |
| 2201 | Operands[0] = Chain; |
| 2202 | for (unsigned j = 1, e = N->getNumOperands(); j != e; ++j) { |
| 2203 | SDValue Operand = N->getOperand(Num: j); |
| 2204 | EVT OperandVT = Operand.getValueType(); |
| 2205 | if (OperandVT.isVector()) { |
| 2206 | EVT OperandEltVT = OperandVT.getVectorElementType(); |
| 2207 | Operands[j] = DAG.getExtractVectorElt(DL: dl, VT: OperandEltVT, Vec: Operand, Idx: i); |
| 2208 | } else { |
| 2209 | Operands[j] = Operand; |
| 2210 | } |
| 2211 | } |
| 2212 | SDValue Scalar = |
| 2213 | DAG.getNode(Opcode: N->getOpcode(), DL: dl, VTList: ChainVTs, Ops: Operands, Flags: N->getFlags()); |
| 2214 | |
| 2215 | //Add in the scalar as well as its chain value to the |
| 2216 | //result vectors. |
| 2217 | Scalars.push_back(Elt: Scalar); |
| 2218 | Chains.push_back(Elt: Scalar.getValue(R: 1)); |
| 2219 | } |
| 2220 | |
| 2221 | for (; i < ResNE; ++i) |
| 2222 | Scalars.push_back(Elt: DAG.getPOISON(VT: EltVT)); |
| 2223 | |
| 2224 | // Build a new factor node to connect the chain back together. |
| 2225 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: Chains); |
| 2226 | ReplaceValueWith(From: SDValue(N, 1), To: Chain); |
| 2227 | |
| 2228 | // Create a new BUILD_VECTOR node |
| 2229 | EVT VecVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: EltVT, NumElements: ResNE); |
| 2230 | return DAG.getBuildVector(VT: VecVT, DL: dl, Ops: Scalars); |
| 2231 | } |
| 2232 | |
| 2233 | void DAGTypeLegalizer::SplitVecRes_OverflowOp(SDNode *N, unsigned ResNo, |
| 2234 | SDValue &Lo, SDValue &Hi) { |
| 2235 | SDLoc dl(N); |
| 2236 | EVT ResVT = N->getValueType(ResNo: 0); |
| 2237 | EVT OvVT = N->getValueType(ResNo: 1); |
| 2238 | EVT LoResVT, HiResVT, LoOvVT, HiOvVT; |
| 2239 | std::tie(args&: LoResVT, args&: HiResVT) = DAG.GetSplitDestVTs(VT: ResVT); |
| 2240 | std::tie(args&: LoOvVT, args&: HiOvVT) = DAG.GetSplitDestVTs(VT: OvVT); |
| 2241 | |
| 2242 | SDValue LoLHS, HiLHS, LoRHS, HiRHS; |
| 2243 | if (getTypeAction(VT: ResVT) == TargetLowering::TypeSplitVector) { |
| 2244 | GetSplitVector(Op: N->getOperand(Num: 0), Lo&: LoLHS, Hi&: HiLHS); |
| 2245 | GetSplitVector(Op: N->getOperand(Num: 1), Lo&: LoRHS, Hi&: HiRHS); |
| 2246 | } else { |
| 2247 | std::tie(args&: LoLHS, args&: HiLHS) = DAG.SplitVectorOperand(N, OpNo: 0); |
| 2248 | std::tie(args&: LoRHS, args&: HiRHS) = DAG.SplitVectorOperand(N, OpNo: 1); |
| 2249 | } |
| 2250 | |
| 2251 | unsigned Opcode = N->getOpcode(); |
| 2252 | SDVTList LoVTs = DAG.getVTList(VT1: LoResVT, VT2: LoOvVT); |
| 2253 | SDVTList HiVTs = DAG.getVTList(VT1: HiResVT, VT2: HiOvVT); |
| 2254 | SDNode *LoNode = |
| 2255 | DAG.getNode(Opcode, DL: dl, VTList: LoVTs, Ops: {LoLHS, LoRHS}, Flags: N->getFlags()).getNode(); |
| 2256 | SDNode *HiNode = |
| 2257 | DAG.getNode(Opcode, DL: dl, VTList: HiVTs, Ops: {HiLHS, HiRHS}, Flags: N->getFlags()).getNode(); |
| 2258 | |
| 2259 | Lo = SDValue(LoNode, ResNo); |
| 2260 | Hi = SDValue(HiNode, ResNo); |
| 2261 | |
| 2262 | // Replace the other vector result not being explicitly split here. |
| 2263 | unsigned OtherNo = 1 - ResNo; |
| 2264 | EVT OtherVT = N->getValueType(ResNo: OtherNo); |
| 2265 | if (getTypeAction(VT: OtherVT) == TargetLowering::TypeSplitVector) { |
| 2266 | SetSplitVector(Op: SDValue(N, OtherNo), |
| 2267 | Lo: SDValue(LoNode, OtherNo), Hi: SDValue(HiNode, OtherNo)); |
| 2268 | } else { |
| 2269 | SDValue OtherVal = DAG.getNode( |
| 2270 | Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: OtherVT, |
| 2271 | N1: SDValue(LoNode, OtherNo), N2: SDValue(HiNode, OtherNo)); |
| 2272 | ReplaceValueWith(From: SDValue(N, OtherNo), To: OtherVal); |
| 2273 | } |
| 2274 | } |
| 2275 | |
| 2276 | void DAGTypeLegalizer::SplitVecRes_INSERT_VECTOR_ELT(SDNode *N, SDValue &Lo, |
| 2277 | SDValue &Hi) { |
| 2278 | SDValue Vec = N->getOperand(Num: 0); |
| 2279 | SDValue Elt = N->getOperand(Num: 1); |
| 2280 | SDValue Idx = N->getOperand(Num: 2); |
| 2281 | SDLoc dl(N); |
| 2282 | GetSplitVector(Op: Vec, Lo, Hi); |
| 2283 | |
| 2284 | if (ConstantSDNode *CIdx = dyn_cast<ConstantSDNode>(Val&: Idx)) { |
| 2285 | unsigned IdxVal = CIdx->getZExtValue(); |
| 2286 | unsigned LoNumElts = Lo.getValueType().getVectorMinNumElements(); |
| 2287 | if (IdxVal < LoNumElts) { |
| 2288 | Lo = DAG.getNode(Opcode: ISD::INSERT_VECTOR_ELT, DL: dl, |
| 2289 | VT: Lo.getValueType(), N1: Lo, N2: Elt, N3: Idx); |
| 2290 | return; |
| 2291 | } else if (!Vec.getValueType().isScalableVector()) { |
| 2292 | Hi = DAG.getInsertVectorElt(DL: dl, Vec: Hi, Elt, Idx: IdxVal - LoNumElts); |
| 2293 | return; |
| 2294 | } |
| 2295 | } |
| 2296 | |
| 2297 | // Make the vector elements byte-addressable if they aren't already. |
| 2298 | EVT VecVT = Vec.getValueType(); |
| 2299 | EVT EltVT = VecVT.getVectorElementType(); |
| 2300 | if (!EltVT.isByteSized()) { |
| 2301 | EltVT = EltVT.changeTypeToInteger().getRoundIntegerType(Context&: *DAG.getContext()); |
| 2302 | VecVT = VecVT.changeElementType(Context&: *DAG.getContext(), EltVT); |
| 2303 | Vec = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT: VecVT, Operand: Vec); |
| 2304 | // Extend the element type to match if needed. |
| 2305 | if (EltVT.bitsGT(VT: Elt.getValueType())) |
| 2306 | Elt = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT: EltVT, Operand: Elt); |
| 2307 | } |
| 2308 | |
| 2309 | // Spill the vector to the stack. |
| 2310 | // In cases where the vector is illegal it will be broken down into parts |
| 2311 | // and stored in parts - we should use the alignment for the smallest part. |
| 2312 | Align SmallestAlign = DAG.getReducedAlign(VT: VecVT, /*UseABI=*/false); |
| 2313 | SDValue StackPtr = |
| 2314 | DAG.CreateStackTemporary(Bytes: VecVT.getStoreSize(), Alignment: SmallestAlign); |
| 2315 | auto &MF = DAG.getMachineFunction(); |
| 2316 | auto FrameIndex = cast<FrameIndexSDNode>(Val: StackPtr.getNode())->getIndex(); |
| 2317 | auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: FrameIndex); |
| 2318 | |
| 2319 | SDValue Store = DAG.getStore(Chain: DAG.getEntryNode(), dl, Val: Vec, Ptr: StackPtr, PtrInfo, |
| 2320 | Alignment: SmallestAlign); |
| 2321 | |
| 2322 | // Store the new element. This may be larger than the vector element type, |
| 2323 | // so use a truncating store. |
| 2324 | SDValue EltPtr = TLI.getVectorElementPointer(DAG, VecPtr: StackPtr, VecVT, Index: Idx); |
| 2325 | Store = DAG.getTruncStore( |
| 2326 | Chain: Store, dl, Val: Elt, Ptr: EltPtr, PtrInfo: MachinePointerInfo::getUnknownStack(MF), SVT: EltVT, |
| 2327 | Alignment: commonAlignment(A: SmallestAlign, |
| 2328 | Offset: EltVT.getFixedSizeInBits() / 8)); |
| 2329 | |
| 2330 | EVT LoVT, HiVT; |
| 2331 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: VecVT); |
| 2332 | |
| 2333 | // Load the Lo part from the stack slot. |
| 2334 | Lo = DAG.getLoad(VT: LoVT, dl, Chain: Store, Ptr: StackPtr, PtrInfo, Alignment: SmallestAlign); |
| 2335 | |
| 2336 | // Increment the pointer to the other part. |
| 2337 | auto Load = cast<LoadSDNode>(Val&: Lo); |
| 2338 | MachinePointerInfo MPI = Load->getPointerInfo(); |
| 2339 | IncrementPointer(N: Load, MemVT: LoVT, MPI, Ptr&: StackPtr); |
| 2340 | |
| 2341 | Hi = DAG.getLoad(VT: HiVT, dl, Chain: Store, Ptr: StackPtr, PtrInfo: MPI, Alignment: SmallestAlign); |
| 2342 | |
| 2343 | // If we adjusted the original type, we need to truncate the results. |
| 2344 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 2345 | if (LoVT != Lo.getValueType()) |
| 2346 | Lo = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: LoVT, Operand: Lo); |
| 2347 | if (HiVT != Hi.getValueType()) |
| 2348 | Hi = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: HiVT, Operand: Hi); |
| 2349 | } |
| 2350 | |
| 2351 | void DAGTypeLegalizer::SplitVecRes_STEP_VECTOR(SDNode *N, SDValue &Lo, |
| 2352 | SDValue &Hi) { |
| 2353 | EVT LoVT, HiVT; |
| 2354 | SDLoc dl(N); |
| 2355 | assert(N->getValueType(0).isScalableVector() && |
| 2356 | "Only scalable vectors are supported for STEP_VECTOR" ); |
| 2357 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 2358 | SDValue Step = N->getOperand(Num: 0); |
| 2359 | |
| 2360 | Lo = DAG.getNode(Opcode: ISD::STEP_VECTOR, DL: dl, VT: LoVT, Operand: Step); |
| 2361 | |
| 2362 | // Hi = Lo + (EltCnt * Step) |
| 2363 | EVT EltVT = Step.getValueType(); |
| 2364 | APInt StepVal = Step->getAsAPIntVal(); |
| 2365 | SDValue StartOfHi = |
| 2366 | DAG.getVScale(DL: dl, VT: EltVT, MulImm: StepVal * LoVT.getVectorMinNumElements()); |
| 2367 | StartOfHi = DAG.getSExtOrTrunc(Op: StartOfHi, DL: dl, VT: HiVT.getVectorElementType()); |
| 2368 | StartOfHi = DAG.getNode(Opcode: ISD::SPLAT_VECTOR, DL: dl, VT: HiVT, Operand: StartOfHi); |
| 2369 | |
| 2370 | Hi = DAG.getNode(Opcode: ISD::STEP_VECTOR, DL: dl, VT: HiVT, Operand: Step); |
| 2371 | Hi = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: HiVT, N1: Hi, N2: StartOfHi); |
| 2372 | } |
| 2373 | |
| 2374 | void DAGTypeLegalizer::SplitVecRes_ScalarOp(SDNode *N, SDValue &Lo, |
| 2375 | SDValue &Hi) { |
| 2376 | EVT LoVT, HiVT; |
| 2377 | SDLoc dl(N); |
| 2378 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 2379 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: LoVT, Operand: N->getOperand(Num: 0)); |
| 2380 | if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) { |
| 2381 | Hi = DAG.getPOISON(VT: HiVT); |
| 2382 | } else { |
| 2383 | assert(N->getOpcode() == ISD::SPLAT_VECTOR && "Unexpected opcode" ); |
| 2384 | Hi = Lo; |
| 2385 | } |
| 2386 | } |
| 2387 | |
| 2388 | void DAGTypeLegalizer::SplitVecRes_ATOMIC_LOAD(AtomicSDNode *LD, SDValue &Lo, |
| 2389 | SDValue &Hi) { |
| 2390 | assert(LD->getExtensionType() == ISD::NON_EXTLOAD && |
| 2391 | "Extended load during type legalization!" ); |
| 2392 | SDLoc dl(LD); |
| 2393 | EVT VT = LD->getValueType(ResNo: 0); |
| 2394 | EVT LoVT, HiVT; |
| 2395 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT); |
| 2396 | |
| 2397 | SDValue Ch = LD->getChain(); |
| 2398 | SDValue Ptr = LD->getBasePtr(); |
| 2399 | |
| 2400 | EVT IntVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: VT.getSizeInBits()); |
| 2401 | EVT MemIntVT = |
| 2402 | EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: LD->getMemoryVT().getSizeInBits()); |
| 2403 | SDValue ALD = DAG.getAtomicLoad(ExtType: LD->getExtensionType(), dl, MemVT: MemIntVT, VT: IntVT, |
| 2404 | Chain: Ch, Ptr, MMO: LD->getMemOperand()); |
| 2405 | |
| 2406 | EVT LoIntVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: LoVT.getSizeInBits()); |
| 2407 | EVT HiIntVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: HiVT.getSizeInBits()); |
| 2408 | SDValue , ; |
| 2409 | SplitInteger(Op: ALD, LoVT: LoIntVT, HiVT: HiIntVT, Lo&: ExtractLo, Hi&: ExtractHi); |
| 2410 | |
| 2411 | Lo = DAG.getBitcast(VT: LoVT, V: ExtractLo); |
| 2412 | Hi = DAG.getBitcast(VT: HiVT, V: ExtractHi); |
| 2413 | |
| 2414 | // Legalize the chain result - switch anything that used the old chain to |
| 2415 | // use the new one. |
| 2416 | ReplaceValueWith(From: SDValue(LD, 1), To: ALD.getValue(R: 1)); |
| 2417 | } |
| 2418 | |
| 2419 | void DAGTypeLegalizer::SplitVecRes_LOAD(LoadSDNode *LD, SDValue &Lo, |
| 2420 | SDValue &Hi) { |
| 2421 | assert(ISD::isUNINDEXEDLoad(LD) && "Indexed load during type legalization!" ); |
| 2422 | EVT LoVT, HiVT; |
| 2423 | SDLoc dl(LD); |
| 2424 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: LD->getValueType(ResNo: 0)); |
| 2425 | |
| 2426 | ISD::LoadExtType ExtType = LD->getExtensionType(); |
| 2427 | SDValue Ch = LD->getChain(); |
| 2428 | SDValue Ptr = LD->getBasePtr(); |
| 2429 | SDValue Offset = DAG.getUNDEF(VT: Ptr.getValueType()); |
| 2430 | EVT MemoryVT = LD->getMemoryVT(); |
| 2431 | MachineMemOperand::Flags MMOFlags = LD->getMemOperand()->getFlags(); |
| 2432 | AAMDNodes AAInfo = LD->getAAInfo(); |
| 2433 | |
| 2434 | EVT LoMemVT, HiMemVT; |
| 2435 | std::tie(args&: LoMemVT, args&: HiMemVT) = DAG.GetSplitDestVTs(VT: MemoryVT); |
| 2436 | |
| 2437 | if (!LoMemVT.isByteSized() || !HiMemVT.isByteSized()) { |
| 2438 | SDValue Value, NewChain; |
| 2439 | std::tie(args&: Value, args&: NewChain) = TLI.scalarizeVectorLoad(LD, DAG); |
| 2440 | std::tie(args&: Lo, args&: Hi) = DAG.SplitVector(N: Value, DL: dl); |
| 2441 | ReplaceValueWith(From: SDValue(LD, 1), To: NewChain); |
| 2442 | return; |
| 2443 | } |
| 2444 | |
| 2445 | Lo = DAG.getLoad(AM: ISD::UNINDEXED, ExtType, VT: LoVT, dl, Chain: Ch, Ptr, Offset, |
| 2446 | PtrInfo: LD->getPointerInfo(), MemVT: LoMemVT, Alignment: LD->getBaseAlign(), MMOFlags, |
| 2447 | AAInfo); |
| 2448 | |
| 2449 | MachinePointerInfo MPI; |
| 2450 | IncrementPointer(N: LD, MemVT: LoMemVT, MPI, Ptr); |
| 2451 | |
| 2452 | Hi = DAG.getLoad(AM: ISD::UNINDEXED, ExtType, VT: HiVT, dl, Chain: Ch, Ptr, Offset, PtrInfo: MPI, |
| 2453 | MemVT: HiMemVT, Alignment: LD->getBaseAlign(), MMOFlags, AAInfo); |
| 2454 | |
| 2455 | // Build a factor node to remember that this load is independent of the |
| 2456 | // other one. |
| 2457 | Ch = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Lo.getValue(R: 1), |
| 2458 | N2: Hi.getValue(R: 1)); |
| 2459 | |
| 2460 | // Legalize the chain result - switch anything that used the old chain to |
| 2461 | // use the new one. |
| 2462 | ReplaceValueWith(From: SDValue(LD, 1), To: Ch); |
| 2463 | } |
| 2464 | |
| 2465 | void DAGTypeLegalizer::SplitVecRes_VP_LOAD(VPLoadSDNode *LD, SDValue &Lo, |
| 2466 | SDValue &Hi) { |
| 2467 | assert(LD->isUnindexed() && "Indexed VP load during type legalization!" ); |
| 2468 | EVT LoVT, HiVT; |
| 2469 | SDLoc dl(LD); |
| 2470 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: LD->getValueType(ResNo: 0)); |
| 2471 | |
| 2472 | ISD::LoadExtType ExtType = LD->getExtensionType(); |
| 2473 | SDValue Ch = LD->getChain(); |
| 2474 | SDValue Ptr = LD->getBasePtr(); |
| 2475 | SDValue Offset = LD->getOffset(); |
| 2476 | assert(Offset.isUndef() && "Unexpected indexed variable-length load offset" ); |
| 2477 | Align Alignment = LD->getBaseAlign(); |
| 2478 | SDValue Mask = LD->getMask(); |
| 2479 | SDValue EVL = LD->getVectorLength(); |
| 2480 | EVT MemoryVT = LD->getMemoryVT(); |
| 2481 | |
| 2482 | EVT LoMemVT, HiMemVT; |
| 2483 | bool HiIsEmpty = false; |
| 2484 | std::tie(args&: LoMemVT, args&: HiMemVT) = |
| 2485 | DAG.GetDependentSplitDestVTs(VT: MemoryVT, EnvVT: LoVT, HiIsEmpty: &HiIsEmpty); |
| 2486 | |
| 2487 | // Split Mask operand |
| 2488 | SDValue MaskLo, MaskHi; |
| 2489 | if (Mask.getOpcode() == ISD::SETCC) { |
| 2490 | SplitVecRes_SETCC(N: Mask.getNode(), Lo&: MaskLo, Hi&: MaskHi); |
| 2491 | } else { |
| 2492 | if (getTypeAction(VT: Mask.getValueType()) == TargetLowering::TypeSplitVector) |
| 2493 | GetSplitVector(Op: Mask, Lo&: MaskLo, Hi&: MaskHi); |
| 2494 | else |
| 2495 | std::tie(args&: MaskLo, args&: MaskHi) = DAG.SplitVector(N: Mask, DL: dl); |
| 2496 | } |
| 2497 | |
| 2498 | // Split EVL operand |
| 2499 | SDValue EVLLo, EVLHi; |
| 2500 | std::tie(args&: EVLLo, args&: EVLHi) = DAG.SplitEVL(N: EVL, VecVT: LD->getValueType(ResNo: 0), DL: dl); |
| 2501 | |
| 2502 | MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand( |
| 2503 | PtrInfo: LD->getPointerInfo(), F: MachineMemOperand::MOLoad, |
| 2504 | Size: LocationSize::beforeOrAfterPointer(), BaseAlignment: Alignment, AAInfo: LD->getAAInfo(), |
| 2505 | Ranges: LD->getRanges()); |
| 2506 | |
| 2507 | Lo = |
| 2508 | DAG.getLoadVP(AM: LD->getAddressingMode(), ExtType, VT: LoVT, dl, Chain: Ch, Ptr, Offset, |
| 2509 | Mask: MaskLo, EVL: EVLLo, MemVT: LoMemVT, MMO, IsExpanding: LD->isExpandingLoad()); |
| 2510 | |
| 2511 | if (HiIsEmpty) { |
| 2512 | // The hi vp_load has zero storage size. We therefore simply set it to |
| 2513 | // the low vp_load and rely on subsequent removal from the chain. |
| 2514 | Hi = Lo; |
| 2515 | } else { |
| 2516 | // Generate hi vp_load. |
| 2517 | Ptr = TLI.IncrementMemoryAddress(Addr: Ptr, Mask: MaskLo, DL: dl, DataVT: LoMemVT, DAG, |
| 2518 | IsCompressedMemory: LD->isExpandingLoad()); |
| 2519 | |
| 2520 | MachinePointerInfo MPI; |
| 2521 | if (LoMemVT.isScalableVector()) |
| 2522 | MPI = MachinePointerInfo(LD->getPointerInfo().getAddrSpace()); |
| 2523 | else |
| 2524 | MPI = LD->getPointerInfo().getWithOffset( |
| 2525 | O: LoMemVT.getStoreSize().getFixedValue()); |
| 2526 | |
| 2527 | MMO = DAG.getMachineFunction().getMachineMemOperand( |
| 2528 | PtrInfo: MPI, F: MachineMemOperand::MOLoad, Size: LocationSize::beforeOrAfterPointer(), |
| 2529 | BaseAlignment: Alignment, AAInfo: LD->getAAInfo(), Ranges: LD->getRanges()); |
| 2530 | |
| 2531 | Hi = DAG.getLoadVP(AM: LD->getAddressingMode(), ExtType, VT: HiVT, dl, Chain: Ch, Ptr, |
| 2532 | Offset, Mask: MaskHi, EVL: EVLHi, MemVT: HiMemVT, MMO, |
| 2533 | IsExpanding: LD->isExpandingLoad()); |
| 2534 | } |
| 2535 | |
| 2536 | // Build a factor node to remember that this load is independent of the |
| 2537 | // other one. |
| 2538 | Ch = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Lo.getValue(R: 1), |
| 2539 | N2: Hi.getValue(R: 1)); |
| 2540 | |
| 2541 | // Legalize the chain result - switch anything that used the old chain to |
| 2542 | // use the new one. |
| 2543 | ReplaceValueWith(From: SDValue(LD, 1), To: Ch); |
| 2544 | } |
| 2545 | |
| 2546 | void DAGTypeLegalizer::SplitVecRes_VP_LOAD_FF(VPLoadFFSDNode *LD, SDValue &Lo, |
| 2547 | SDValue &Hi) { |
| 2548 | SDLoc dl(LD); |
| 2549 | auto [LoVT, HiVT] = DAG.GetSplitDestVTs(VT: LD->getValueType(ResNo: 0)); |
| 2550 | |
| 2551 | SDValue Ch = LD->getChain(); |
| 2552 | SDValue Ptr = LD->getBasePtr(); |
| 2553 | Align Alignment = LD->getBaseAlign(); |
| 2554 | SDValue Mask = LD->getMask(); |
| 2555 | SDValue EVL = LD->getVectorLength(); |
| 2556 | |
| 2557 | // Split Mask operand |
| 2558 | SDValue MaskLo, MaskHi; |
| 2559 | if (Mask.getOpcode() == ISD::SETCC) { |
| 2560 | SplitVecRes_SETCC(N: Mask.getNode(), Lo&: MaskLo, Hi&: MaskHi); |
| 2561 | } else { |
| 2562 | if (getTypeAction(VT: Mask.getValueType()) == TargetLowering::TypeSplitVector) |
| 2563 | GetSplitVector(Op: Mask, Lo&: MaskLo, Hi&: MaskHi); |
| 2564 | else |
| 2565 | std::tie(args&: MaskLo, args&: MaskHi) = DAG.SplitVector(N: Mask, DL: dl); |
| 2566 | } |
| 2567 | |
| 2568 | // Split EVL operand |
| 2569 | auto [EVLLo, EVLHi] = DAG.SplitEVL(N: EVL, VecVT: LD->getValueType(ResNo: 0), DL: dl); |
| 2570 | |
| 2571 | MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand( |
| 2572 | PtrInfo: LD->getPointerInfo(), F: MachineMemOperand::MOLoad, |
| 2573 | Size: LocationSize::beforeOrAfterPointer(), BaseAlignment: Alignment, AAInfo: LD->getAAInfo(), |
| 2574 | Ranges: LD->getRanges()); |
| 2575 | |
| 2576 | Lo = DAG.getLoadFFVP(VT: LoVT, DL: dl, Chain: Ch, Ptr, Mask: MaskLo, EVL: EVLLo, MMO); |
| 2577 | |
| 2578 | // Fill the upper half with poison. |
| 2579 | Hi = DAG.getPOISON(VT: HiVT); |
| 2580 | |
| 2581 | ReplaceValueWith(From: SDValue(LD, 1), To: Lo.getValue(R: 1)); |
| 2582 | ReplaceValueWith(From: SDValue(LD, 2), To: Lo.getValue(R: 2)); |
| 2583 | } |
| 2584 | |
| 2585 | void DAGTypeLegalizer::SplitVecRes_VP_STRIDED_LOAD(VPStridedLoadSDNode *SLD, |
| 2586 | SDValue &Lo, SDValue &Hi) { |
| 2587 | assert(SLD->isUnindexed() && |
| 2588 | "Indexed VP strided load during type legalization!" ); |
| 2589 | assert(SLD->getOffset().isUndef() && |
| 2590 | "Unexpected indexed variable-length load offset" ); |
| 2591 | |
| 2592 | SDLoc DL(SLD); |
| 2593 | |
| 2594 | EVT LoVT, HiVT; |
| 2595 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: SLD->getValueType(ResNo: 0)); |
| 2596 | |
| 2597 | EVT LoMemVT, HiMemVT; |
| 2598 | bool HiIsEmpty = false; |
| 2599 | std::tie(args&: LoMemVT, args&: HiMemVT) = |
| 2600 | DAG.GetDependentSplitDestVTs(VT: SLD->getMemoryVT(), EnvVT: LoVT, HiIsEmpty: &HiIsEmpty); |
| 2601 | |
| 2602 | SDValue Mask = SLD->getMask(); |
| 2603 | SDValue LoMask, HiMask; |
| 2604 | if (Mask.getOpcode() == ISD::SETCC) { |
| 2605 | SplitVecRes_SETCC(N: Mask.getNode(), Lo&: LoMask, Hi&: HiMask); |
| 2606 | } else { |
| 2607 | if (getTypeAction(VT: Mask.getValueType()) == TargetLowering::TypeSplitVector) |
| 2608 | GetSplitVector(Op: Mask, Lo&: LoMask, Hi&: HiMask); |
| 2609 | else |
| 2610 | std::tie(args&: LoMask, args&: HiMask) = DAG.SplitVector(N: Mask, DL); |
| 2611 | } |
| 2612 | |
| 2613 | SDValue LoEVL, HiEVL; |
| 2614 | std::tie(args&: LoEVL, args&: HiEVL) = |
| 2615 | DAG.SplitEVL(N: SLD->getVectorLength(), VecVT: SLD->getValueType(ResNo: 0), DL); |
| 2616 | |
| 2617 | // Generate the low vp_strided_load |
| 2618 | Lo = DAG.getStridedLoadVP( |
| 2619 | AM: SLD->getAddressingMode(), ExtType: SLD->getExtensionType(), VT: LoVT, DL, |
| 2620 | Chain: SLD->getChain(), Ptr: SLD->getBasePtr(), Offset: SLD->getOffset(), Stride: SLD->getStride(), |
| 2621 | Mask: LoMask, EVL: LoEVL, MemVT: LoMemVT, MMO: SLD->getMemOperand(), IsExpanding: SLD->isExpandingLoad()); |
| 2622 | |
| 2623 | if (HiIsEmpty) { |
| 2624 | // The high vp_strided_load has zero storage size. We therefore simply set |
| 2625 | // it to the low vp_strided_load and rely on subsequent removal from the |
| 2626 | // chain. |
| 2627 | Hi = Lo; |
| 2628 | } else { |
| 2629 | // Generate the high vp_strided_load. |
| 2630 | // To calculate the high base address, we need to sum to the low base |
| 2631 | // address stride number of bytes for each element already loaded by low, |
| 2632 | // that is: Ptr = Ptr + (LoEVL * Stride) |
| 2633 | EVT PtrVT = SLD->getBasePtr().getValueType(); |
| 2634 | SDValue Increment = |
| 2635 | DAG.getNode(Opcode: ISD::MUL, DL, VT: PtrVT, N1: LoEVL, |
| 2636 | N2: DAG.getSExtOrTrunc(Op: SLD->getStride(), DL, VT: PtrVT)); |
| 2637 | SDValue Ptr = |
| 2638 | DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: SLD->getBasePtr(), N2: Increment); |
| 2639 | |
| 2640 | Align Alignment = SLD->getBaseAlign(); |
| 2641 | if (LoMemVT.isScalableVector()) |
| 2642 | Alignment = commonAlignment( |
| 2643 | A: Alignment, Offset: LoMemVT.getSizeInBits().getKnownMinValue() / 8); |
| 2644 | |
| 2645 | MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand( |
| 2646 | PtrInfo: MachinePointerInfo(SLD->getPointerInfo().getAddrSpace()), |
| 2647 | F: MachineMemOperand::MOLoad, Size: LocationSize::beforeOrAfterPointer(), |
| 2648 | BaseAlignment: Alignment, AAInfo: SLD->getAAInfo(), Ranges: SLD->getRanges()); |
| 2649 | |
| 2650 | Hi = DAG.getStridedLoadVP(AM: SLD->getAddressingMode(), ExtType: SLD->getExtensionType(), |
| 2651 | VT: HiVT, DL, Chain: SLD->getChain(), Ptr, Offset: SLD->getOffset(), |
| 2652 | Stride: SLD->getStride(), Mask: HiMask, EVL: HiEVL, MemVT: HiMemVT, MMO, |
| 2653 | IsExpanding: SLD->isExpandingLoad()); |
| 2654 | } |
| 2655 | |
| 2656 | // Build a factor node to remember that this load is independent of the |
| 2657 | // other one. |
| 2658 | SDValue Ch = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, N1: Lo.getValue(R: 1), |
| 2659 | N2: Hi.getValue(R: 1)); |
| 2660 | |
| 2661 | // Legalize the chain result - switch anything that used the old chain to |
| 2662 | // use the new one. |
| 2663 | ReplaceValueWith(From: SDValue(SLD, 1), To: Ch); |
| 2664 | } |
| 2665 | |
| 2666 | void DAGTypeLegalizer::SplitVecRes_MLOAD(MaskedLoadSDNode *MLD, |
| 2667 | SDValue &Lo, SDValue &Hi) { |
| 2668 | assert(MLD->isUnindexed() && "Indexed masked load during type legalization!" ); |
| 2669 | EVT LoVT, HiVT; |
| 2670 | SDLoc dl(MLD); |
| 2671 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: MLD->getValueType(ResNo: 0)); |
| 2672 | |
| 2673 | SDValue Ch = MLD->getChain(); |
| 2674 | SDValue Ptr = MLD->getBasePtr(); |
| 2675 | SDValue Offset = MLD->getOffset(); |
| 2676 | assert(Offset.isUndef() && "Unexpected indexed masked load offset" ); |
| 2677 | SDValue Mask = MLD->getMask(); |
| 2678 | SDValue PassThru = MLD->getPassThru(); |
| 2679 | Align Alignment = MLD->getBaseAlign(); |
| 2680 | ISD::LoadExtType ExtType = MLD->getExtensionType(); |
| 2681 | MachineMemOperand::Flags MMOFlags = MLD->getMemOperand()->getFlags(); |
| 2682 | |
| 2683 | // Split Mask operand |
| 2684 | SDValue MaskLo, MaskHi; |
| 2685 | if (Mask.getOpcode() == ISD::SETCC) { |
| 2686 | SplitVecRes_SETCC(N: Mask.getNode(), Lo&: MaskLo, Hi&: MaskHi); |
| 2687 | } else { |
| 2688 | if (getTypeAction(VT: Mask.getValueType()) == TargetLowering::TypeSplitVector) |
| 2689 | GetSplitVector(Op: Mask, Lo&: MaskLo, Hi&: MaskHi); |
| 2690 | else |
| 2691 | std::tie(args&: MaskLo, args&: MaskHi) = DAG.SplitVector(N: Mask, DL: dl); |
| 2692 | } |
| 2693 | |
| 2694 | EVT MemoryVT = MLD->getMemoryVT(); |
| 2695 | EVT LoMemVT, HiMemVT; |
| 2696 | bool HiIsEmpty = false; |
| 2697 | std::tie(args&: LoMemVT, args&: HiMemVT) = |
| 2698 | DAG.GetDependentSplitDestVTs(VT: MemoryVT, EnvVT: LoVT, HiIsEmpty: &HiIsEmpty); |
| 2699 | |
| 2700 | SDValue PassThruLo, PassThruHi; |
| 2701 | if (getTypeAction(VT: PassThru.getValueType()) == TargetLowering::TypeSplitVector) |
| 2702 | GetSplitVector(Op: PassThru, Lo&: PassThruLo, Hi&: PassThruHi); |
| 2703 | else |
| 2704 | std::tie(args&: PassThruLo, args&: PassThruHi) = DAG.SplitVector(N: PassThru, DL: dl); |
| 2705 | |
| 2706 | MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand( |
| 2707 | PtrInfo: MLD->getPointerInfo(), F: MMOFlags, Size: LocationSize::beforeOrAfterPointer(), |
| 2708 | BaseAlignment: Alignment, AAInfo: MLD->getAAInfo(), Ranges: MLD->getRanges()); |
| 2709 | |
| 2710 | Lo = DAG.getMaskedLoad(VT: LoVT, dl, Chain: Ch, Base: Ptr, Offset, Mask: MaskLo, Src0: PassThruLo, MemVT: LoMemVT, |
| 2711 | MMO, AM: MLD->getAddressingMode(), ExtType, |
| 2712 | IsExpanding: MLD->isExpandingLoad()); |
| 2713 | |
| 2714 | if (HiIsEmpty) { |
| 2715 | // The hi masked load has zero storage size. We therefore simply set it to |
| 2716 | // the low masked load and rely on subsequent removal from the chain. |
| 2717 | Hi = Lo; |
| 2718 | } else { |
| 2719 | // Generate hi masked load. |
| 2720 | Ptr = TLI.IncrementMemoryAddress(Addr: Ptr, Mask: MaskLo, DL: dl, DataVT: LoMemVT, DAG, |
| 2721 | IsCompressedMemory: MLD->isExpandingLoad()); |
| 2722 | |
| 2723 | MachinePointerInfo MPI; |
| 2724 | if (LoMemVT.isScalableVector()) |
| 2725 | MPI = MachinePointerInfo(MLD->getPointerInfo().getAddrSpace()); |
| 2726 | else |
| 2727 | MPI = MLD->getPointerInfo().getWithOffset( |
| 2728 | O: LoMemVT.getStoreSize().getFixedValue()); |
| 2729 | |
| 2730 | MMO = DAG.getMachineFunction().getMachineMemOperand( |
| 2731 | PtrInfo: MPI, F: MMOFlags, Size: LocationSize::beforeOrAfterPointer(), BaseAlignment: Alignment, |
| 2732 | AAInfo: MLD->getAAInfo(), Ranges: MLD->getRanges()); |
| 2733 | |
| 2734 | Hi = DAG.getMaskedLoad(VT: HiVT, dl, Chain: Ch, Base: Ptr, Offset, Mask: MaskHi, Src0: PassThruHi, |
| 2735 | MemVT: HiMemVT, MMO, AM: MLD->getAddressingMode(), ExtType, |
| 2736 | IsExpanding: MLD->isExpandingLoad()); |
| 2737 | } |
| 2738 | |
| 2739 | // Build a factor node to remember that this load is independent of the |
| 2740 | // other one. |
| 2741 | Ch = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Lo.getValue(R: 1), |
| 2742 | N2: Hi.getValue(R: 1)); |
| 2743 | |
| 2744 | // Legalize the chain result - switch anything that used the old chain to |
| 2745 | // use the new one. |
| 2746 | ReplaceValueWith(From: SDValue(MLD, 1), To: Ch); |
| 2747 | |
| 2748 | } |
| 2749 | |
| 2750 | void DAGTypeLegalizer::SplitVecRes_Gather(MemSDNode *N, SDValue &Lo, |
| 2751 | SDValue &Hi, bool SplitSETCC) { |
| 2752 | EVT LoVT, HiVT; |
| 2753 | SDLoc dl(N); |
| 2754 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 2755 | |
| 2756 | SDValue Ch = N->getChain(); |
| 2757 | SDValue Ptr = N->getBasePtr(); |
| 2758 | struct Operands { |
| 2759 | SDValue Mask; |
| 2760 | SDValue Index; |
| 2761 | SDValue Scale; |
| 2762 | } Ops = [&]() -> Operands { |
| 2763 | if (auto *MSC = dyn_cast<MaskedGatherSDNode>(Val: N)) { |
| 2764 | return {.Mask: MSC->getMask(), .Index: MSC->getIndex(), .Scale: MSC->getScale()}; |
| 2765 | } |
| 2766 | auto *VPSC = cast<VPGatherSDNode>(Val: N); |
| 2767 | return {.Mask: VPSC->getMask(), .Index: VPSC->getIndex(), .Scale: VPSC->getScale()}; |
| 2768 | }(); |
| 2769 | |
| 2770 | EVT MemoryVT = N->getMemoryVT(); |
| 2771 | Align Alignment = N->getBaseAlign(); |
| 2772 | |
| 2773 | // Split Mask operand |
| 2774 | SDValue MaskLo, MaskHi; |
| 2775 | if (SplitSETCC && Ops.Mask.getOpcode() == ISD::SETCC) { |
| 2776 | SplitVecRes_SETCC(N: Ops.Mask.getNode(), Lo&: MaskLo, Hi&: MaskHi); |
| 2777 | } else { |
| 2778 | std::tie(args&: MaskLo, args&: MaskHi) = SplitMask(Mask: Ops.Mask, DL: dl); |
| 2779 | } |
| 2780 | |
| 2781 | EVT LoMemVT, HiMemVT; |
| 2782 | // Split MemoryVT |
| 2783 | std::tie(args&: LoMemVT, args&: HiMemVT) = DAG.GetSplitDestVTs(VT: MemoryVT); |
| 2784 | |
| 2785 | SDValue IndexHi, IndexLo; |
| 2786 | if (getTypeAction(VT: Ops.Index.getValueType()) == |
| 2787 | TargetLowering::TypeSplitVector) |
| 2788 | GetSplitVector(Op: Ops.Index, Lo&: IndexLo, Hi&: IndexHi); |
| 2789 | else |
| 2790 | std::tie(args&: IndexLo, args&: IndexHi) = DAG.SplitVector(N: Ops.Index, DL: dl); |
| 2791 | |
| 2792 | MachineMemOperand::Flags MMOFlags = N->getMemOperand()->getFlags(); |
| 2793 | MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand( |
| 2794 | PtrInfo: N->getPointerInfo(), F: MMOFlags, Size: LocationSize::beforeOrAfterPointer(), |
| 2795 | BaseAlignment: Alignment, AAInfo: N->getAAInfo(), Ranges: N->getRanges()); |
| 2796 | |
| 2797 | if (auto *MGT = dyn_cast<MaskedGatherSDNode>(Val: N)) { |
| 2798 | SDValue PassThru = MGT->getPassThru(); |
| 2799 | SDValue PassThruLo, PassThruHi; |
| 2800 | if (getTypeAction(VT: PassThru.getValueType()) == |
| 2801 | TargetLowering::TypeSplitVector) |
| 2802 | GetSplitVector(Op: PassThru, Lo&: PassThruLo, Hi&: PassThruHi); |
| 2803 | else |
| 2804 | std::tie(args&: PassThruLo, args&: PassThruHi) = DAG.SplitVector(N: PassThru, DL: dl); |
| 2805 | |
| 2806 | ISD::LoadExtType ExtType = MGT->getExtensionType(); |
| 2807 | ISD::MemIndexType IndexTy = MGT->getIndexType(); |
| 2808 | |
| 2809 | SDValue OpsLo[] = {Ch, PassThruLo, MaskLo, Ptr, IndexLo, Ops.Scale}; |
| 2810 | Lo = DAG.getMaskedGather(VTs: DAG.getVTList(VT1: LoVT, VT2: MVT::Other), MemVT: LoMemVT, dl, |
| 2811 | Ops: OpsLo, MMO, IndexType: IndexTy, ExtTy: ExtType); |
| 2812 | |
| 2813 | SDValue OpsHi[] = {Ch, PassThruHi, MaskHi, Ptr, IndexHi, Ops.Scale}; |
| 2814 | Hi = DAG.getMaskedGather(VTs: DAG.getVTList(VT1: HiVT, VT2: MVT::Other), MemVT: HiMemVT, dl, |
| 2815 | Ops: OpsHi, MMO, IndexType: IndexTy, ExtTy: ExtType); |
| 2816 | } else { |
| 2817 | auto *VPGT = cast<VPGatherSDNode>(Val: N); |
| 2818 | SDValue EVLLo, EVLHi; |
| 2819 | std::tie(args&: EVLLo, args&: EVLHi) = |
| 2820 | DAG.SplitEVL(N: VPGT->getVectorLength(), VecVT: MemoryVT, DL: dl); |
| 2821 | |
| 2822 | SDValue OpsLo[] = {Ch, Ptr, IndexLo, Ops.Scale, MaskLo, EVLLo}; |
| 2823 | Lo = DAG.getGatherVP(VTs: DAG.getVTList(VT1: LoVT, VT2: MVT::Other), VT: LoMemVT, dl, Ops: OpsLo, |
| 2824 | MMO, IndexType: VPGT->getIndexType()); |
| 2825 | |
| 2826 | SDValue OpsHi[] = {Ch, Ptr, IndexHi, Ops.Scale, MaskHi, EVLHi}; |
| 2827 | Hi = DAG.getGatherVP(VTs: DAG.getVTList(VT1: HiVT, VT2: MVT::Other), VT: HiMemVT, dl, Ops: OpsHi, |
| 2828 | MMO, IndexType: VPGT->getIndexType()); |
| 2829 | } |
| 2830 | |
| 2831 | // Build a factor node to remember that this load is independent of the |
| 2832 | // other one. |
| 2833 | Ch = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Lo.getValue(R: 1), |
| 2834 | N2: Hi.getValue(R: 1)); |
| 2835 | |
| 2836 | // Legalize the chain result - switch anything that used the old chain to |
| 2837 | // use the new one. |
| 2838 | ReplaceValueWith(From: SDValue(N, 1), To: Ch); |
| 2839 | } |
| 2840 | |
| 2841 | void DAGTypeLegalizer::SplitVecRes_VECTOR_COMPRESS(SDNode *N, SDValue &Lo, |
| 2842 | SDValue &Hi) { |
| 2843 | // This is not "trivial", as there is a dependency between the two subvectors. |
| 2844 | // Depending on the number of 1s in the mask, the elements from the Hi vector |
| 2845 | // need to be moved to the Lo vector. Passthru values make this even harder. |
| 2846 | // We try to use VECTOR_COMPRESS if the target has custom lowering with |
| 2847 | // smaller types and passthru is undef, as it is most likely faster than the |
| 2848 | // fully expand path. Otherwise, just do the full expansion as one "big" |
| 2849 | // operation and then extract the Lo and Hi vectors from that. This gets |
| 2850 | // rid of VECTOR_COMPRESS and all other operands can be legalized later. |
| 2851 | SDLoc DL(N); |
| 2852 | EVT VecVT = N->getValueType(ResNo: 0); |
| 2853 | |
| 2854 | auto [LoVT, HiVT] = DAG.GetSplitDestVTs(VT: VecVT); |
| 2855 | bool HasCustomLowering = false; |
| 2856 | EVT CheckVT = LoVT; |
| 2857 | while (CheckVT.getVectorMinNumElements() > 1) { |
| 2858 | // TLI.isOperationLegalOrCustom requires a legal type, but we could have a |
| 2859 | // custom lowering for illegal types. So we do the checks separately. |
| 2860 | if (TLI.isOperationLegal(Op: ISD::VECTOR_COMPRESS, VT: CheckVT) || |
| 2861 | TLI.isOperationCustom(Op: ISD::VECTOR_COMPRESS, VT: CheckVT)) { |
| 2862 | HasCustomLowering = true; |
| 2863 | break; |
| 2864 | } |
| 2865 | CheckVT = CheckVT.getHalfNumVectorElementsVT(Context&: *DAG.getContext()); |
| 2866 | } |
| 2867 | |
| 2868 | SDValue Passthru = N->getOperand(Num: 2); |
| 2869 | if (!HasCustomLowering) { |
| 2870 | SDValue Compressed = TLI.expandVECTOR_COMPRESS(Node: N, DAG); |
| 2871 | std::tie(args&: Lo, args&: Hi) = DAG.SplitVector(N: Compressed, DL, LoVT, HiVT); |
| 2872 | return; |
| 2873 | } |
| 2874 | |
| 2875 | // Try to VECTOR_COMPRESS smaller vectors and combine via a stack store+load. |
| 2876 | SDValue Mask = N->getOperand(Num: 1); |
| 2877 | SDValue LoMask, HiMask; |
| 2878 | std::tie(args&: Lo, args&: Hi) = DAG.SplitVectorOperand(N, OpNo: 0); |
| 2879 | std::tie(args&: LoMask, args&: HiMask) = SplitMask(Mask); |
| 2880 | |
| 2881 | SDValue UndefPassthru = DAG.getPOISON(VT: LoVT); |
| 2882 | Lo = DAG.getNode(Opcode: ISD::VECTOR_COMPRESS, DL, VT: LoVT, N1: Lo, N2: LoMask, N3: UndefPassthru); |
| 2883 | Hi = DAG.getNode(Opcode: ISD::VECTOR_COMPRESS, DL, VT: HiVT, N1: Hi, N2: HiMask, N3: UndefPassthru); |
| 2884 | |
| 2885 | SDValue StackPtr = DAG.CreateStackTemporary( |
| 2886 | Bytes: VecVT.getStoreSize(), Alignment: DAG.getReducedAlign(VT: VecVT, /*UseABI=*/false)); |
| 2887 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2888 | MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack( |
| 2889 | MF, FI: cast<FrameIndexSDNode>(Val: StackPtr.getNode())->getIndex()); |
| 2890 | |
| 2891 | EVT MaskVT = LoMask.getValueType(); |
| 2892 | assert(MaskVT.getScalarType() == MVT::i1 && "Expected vector of i1s" ); |
| 2893 | |
| 2894 | // We store LoVec and then insert HiVec starting at offset=|1s| in LoMask. |
| 2895 | EVT WideMaskVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: MVT::i32, |
| 2896 | EC: MaskVT.getVectorElementCount()); |
| 2897 | SDValue WideMask = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: WideMaskVT, Operand: LoMask); |
| 2898 | SDValue Offset = DAG.getNode(Opcode: ISD::VECREDUCE_ADD, DL, VT: MVT::i32, Operand: WideMask); |
| 2899 | Offset = TLI.getVectorElementPointer(DAG, VecPtr: StackPtr, VecVT, Index: Offset); |
| 2900 | |
| 2901 | SDValue Chain = DAG.getEntryNode(); |
| 2902 | Chain = DAG.getStore(Chain, dl: DL, Val: Lo, Ptr: StackPtr, PtrInfo); |
| 2903 | Chain = DAG.getStore(Chain, dl: DL, Val: Hi, Ptr: Offset, |
| 2904 | PtrInfo: MachinePointerInfo::getUnknownStack(MF)); |
| 2905 | |
| 2906 | SDValue Compressed = DAG.getLoad(VT: VecVT, dl: DL, Chain, Ptr: StackPtr, PtrInfo); |
| 2907 | if (!Passthru.isUndef()) { |
| 2908 | Compressed = |
| 2909 | DAG.getNode(Opcode: ISD::VSELECT, DL, VT: VecVT, N1: Mask, N2: Compressed, N3: Passthru); |
| 2910 | } |
| 2911 | std::tie(args&: Lo, args&: Hi) = DAG.SplitVector(N: Compressed, DL); |
| 2912 | } |
| 2913 | |
| 2914 | void DAGTypeLegalizer::SplitVecRes_SETCC(SDNode *N, SDValue &Lo, SDValue &Hi) { |
| 2915 | assert(N->getValueType(0).isVector() && |
| 2916 | N->getOperand(0).getValueType().isVector() && |
| 2917 | "Operand types must be vectors" ); |
| 2918 | |
| 2919 | EVT LoVT, HiVT; |
| 2920 | SDLoc DL(N); |
| 2921 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 2922 | |
| 2923 | // If the input also splits, handle it directly. Otherwise split it by hand. |
| 2924 | SDValue LL, LH, RL, RH; |
| 2925 | if (getTypeAction(VT: N->getOperand(Num: 0).getValueType()) == |
| 2926 | TargetLowering::TypeSplitVector) |
| 2927 | GetSplitVector(Op: N->getOperand(Num: 0), Lo&: LL, Hi&: LH); |
| 2928 | else |
| 2929 | std::tie(args&: LL, args&: LH) = DAG.SplitVectorOperand(N, OpNo: 0); |
| 2930 | |
| 2931 | if (getTypeAction(VT: N->getOperand(Num: 1).getValueType()) == |
| 2932 | TargetLowering::TypeSplitVector) |
| 2933 | GetSplitVector(Op: N->getOperand(Num: 1), Lo&: RL, Hi&: RH); |
| 2934 | else |
| 2935 | std::tie(args&: RL, args&: RH) = DAG.SplitVectorOperand(N, OpNo: 1); |
| 2936 | |
| 2937 | if (N->getOpcode() == ISD::SETCC) { |
| 2938 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL, VT: LoVT, N1: LL, N2: RL, N3: N->getOperand(Num: 2)); |
| 2939 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL, VT: HiVT, N1: LH, N2: RH, N3: N->getOperand(Num: 2)); |
| 2940 | } else { |
| 2941 | assert(N->getOpcode() == ISD::VP_SETCC && "Expected VP_SETCC opcode" ); |
| 2942 | SDValue MaskLo, MaskHi, EVLLo, EVLHi; |
| 2943 | std::tie(args&: MaskLo, args&: MaskHi) = SplitMask(Mask: N->getOperand(Num: 3)); |
| 2944 | std::tie(args&: EVLLo, args&: EVLHi) = |
| 2945 | DAG.SplitEVL(N: N->getOperand(Num: 4), VecVT: N->getValueType(ResNo: 0), DL); |
| 2946 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL, VT: LoVT, N1: LL, N2: RL, N3: N->getOperand(Num: 2), N4: MaskLo, |
| 2947 | N5: EVLLo); |
| 2948 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL, VT: HiVT, N1: LH, N2: RH, N3: N->getOperand(Num: 2), N4: MaskHi, |
| 2949 | N5: EVLHi); |
| 2950 | } |
| 2951 | } |
| 2952 | |
| 2953 | void DAGTypeLegalizer::SplitVecRes_UnaryOp(SDNode *N, SDValue &Lo, |
| 2954 | SDValue &Hi) { |
| 2955 | // Get the dest types - they may not match the input types, e.g. int_to_fp. |
| 2956 | EVT LoVT, HiVT; |
| 2957 | SDLoc dl(N); |
| 2958 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 2959 | |
| 2960 | // If the input also splits, handle it directly for a compile time speedup. |
| 2961 | // Otherwise split it by hand. |
| 2962 | EVT InVT = N->getOperand(Num: 0).getValueType(); |
| 2963 | if (getTypeAction(VT: InVT) == TargetLowering::TypeSplitVector) |
| 2964 | GetSplitVector(Op: N->getOperand(Num: 0), Lo, Hi); |
| 2965 | else |
| 2966 | std::tie(args&: Lo, args&: Hi) = DAG.SplitVectorOperand(N, OpNo: 0); |
| 2967 | |
| 2968 | const SDNodeFlags Flags = N->getFlags(); |
| 2969 | unsigned Opcode = N->getOpcode(); |
| 2970 | if (Opcode == ISD::CONVERT_TO_ARBITRARY_FP) { |
| 2971 | Lo = DAG.getNode(Opcode, DL: dl, VT: LoVT, N1: Lo, N2: N->getOperand(Num: 1), N3: N->getOperand(Num: 2), |
| 2972 | N4: N->getOperand(Num: 3), Flags); |
| 2973 | Hi = DAG.getNode(Opcode, DL: dl, VT: HiVT, N1: Hi, N2: N->getOperand(Num: 1), N3: N->getOperand(Num: 2), |
| 2974 | N4: N->getOperand(Num: 3), Flags); |
| 2975 | return; |
| 2976 | } |
| 2977 | if (N->getNumOperands() <= 2) { |
| 2978 | if (Opcode == ISD::FP_ROUND || Opcode == ISD::AssertNoFPClass || |
| 2979 | Opcode == ISD::CONVERT_FROM_ARBITRARY_FP) { |
| 2980 | Lo = DAG.getNode(Opcode, DL: dl, VT: LoVT, N1: Lo, N2: N->getOperand(Num: 1), Flags); |
| 2981 | Hi = DAG.getNode(Opcode, DL: dl, VT: HiVT, N1: Hi, N2: N->getOperand(Num: 1), Flags); |
| 2982 | } else { |
| 2983 | Lo = DAG.getNode(Opcode, DL: dl, VT: LoVT, Operand: Lo, Flags); |
| 2984 | Hi = DAG.getNode(Opcode, DL: dl, VT: HiVT, Operand: Hi, Flags); |
| 2985 | } |
| 2986 | return; |
| 2987 | } |
| 2988 | |
| 2989 | assert(N->getNumOperands() == 3 && "Unexpected number of operands!" ); |
| 2990 | assert(N->isVPOpcode() && "Expected VP opcode" ); |
| 2991 | |
| 2992 | SDValue MaskLo, MaskHi; |
| 2993 | std::tie(args&: MaskLo, args&: MaskHi) = SplitMask(Mask: N->getOperand(Num: 1)); |
| 2994 | |
| 2995 | SDValue EVLLo, EVLHi; |
| 2996 | std::tie(args&: EVLLo, args&: EVLHi) = |
| 2997 | DAG.SplitEVL(N: N->getOperand(Num: 2), VecVT: N->getValueType(ResNo: 0), DL: dl); |
| 2998 | |
| 2999 | Lo = DAG.getNode(Opcode, DL: dl, VT: LoVT, Ops: {Lo, MaskLo, EVLLo}, Flags); |
| 3000 | Hi = DAG.getNode(Opcode, DL: dl, VT: HiVT, Ops: {Hi, MaskHi, EVLHi}, Flags); |
| 3001 | } |
| 3002 | |
| 3003 | void DAGTypeLegalizer::SplitVecRes_ADDRSPACECAST(SDNode *N, SDValue &Lo, |
| 3004 | SDValue &Hi) { |
| 3005 | SDLoc dl(N); |
| 3006 | auto [LoVT, HiVT] = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 3007 | |
| 3008 | // If the input also splits, handle it directly for a compile time speedup. |
| 3009 | // Otherwise split it by hand. |
| 3010 | EVT InVT = N->getOperand(Num: 0).getValueType(); |
| 3011 | if (getTypeAction(VT: InVT) == TargetLowering::TypeSplitVector) |
| 3012 | GetSplitVector(Op: N->getOperand(Num: 0), Lo, Hi); |
| 3013 | else |
| 3014 | std::tie(args&: Lo, args&: Hi) = DAG.SplitVectorOperand(N, OpNo: 0); |
| 3015 | |
| 3016 | auto *AddrSpaceCastN = cast<AddrSpaceCastSDNode>(Val: N); |
| 3017 | unsigned SrcAS = AddrSpaceCastN->getSrcAddressSpace(); |
| 3018 | unsigned DestAS = AddrSpaceCastN->getDestAddressSpace(); |
| 3019 | Lo = DAG.getAddrSpaceCast(dl, VT: LoVT, Ptr: Lo, SrcAS, DestAS); |
| 3020 | Hi = DAG.getAddrSpaceCast(dl, VT: HiVT, Ptr: Hi, SrcAS, DestAS); |
| 3021 | } |
| 3022 | |
| 3023 | void DAGTypeLegalizer::SplitVecRes_UnaryOpWithTwoResults(SDNode *N, |
| 3024 | unsigned ResNo, |
| 3025 | SDValue &Lo, |
| 3026 | SDValue &Hi) { |
| 3027 | SDLoc dl(N); |
| 3028 | auto [LoVT, HiVT] = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 3029 | auto [LoVT1, HiVT1] = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 1)); |
| 3030 | |
| 3031 | // If the input also splits, handle it directly for a compile time speedup. |
| 3032 | // Otherwise split it by hand. |
| 3033 | EVT InVT = N->getOperand(Num: 0).getValueType(); |
| 3034 | if (getTypeAction(VT: InVT) == TargetLowering::TypeSplitVector) |
| 3035 | GetSplitVector(Op: N->getOperand(Num: 0), Lo, Hi); |
| 3036 | else |
| 3037 | std::tie(args&: Lo, args&: Hi) = DAG.SplitVectorOperand(N, OpNo: 0); |
| 3038 | |
| 3039 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL: dl, ResultTys: {LoVT, LoVT1}, Ops: Lo, Flags: N->getFlags()); |
| 3040 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL: dl, ResultTys: {HiVT, HiVT1}, Ops: Hi, Flags: N->getFlags()); |
| 3041 | |
| 3042 | SDNode *HiNode = Hi.getNode(); |
| 3043 | SDNode *LoNode = Lo.getNode(); |
| 3044 | |
| 3045 | // Replace the other vector result not being explicitly split here. |
| 3046 | unsigned OtherNo = 1 - ResNo; |
| 3047 | EVT OtherVT = N->getValueType(ResNo: OtherNo); |
| 3048 | if (getTypeAction(VT: OtherVT) == TargetLowering::TypeSplitVector) { |
| 3049 | SetSplitVector(Op: SDValue(N, OtherNo), Lo: SDValue(LoNode, OtherNo), |
| 3050 | Hi: SDValue(HiNode, OtherNo)); |
| 3051 | } else { |
| 3052 | SDValue OtherVal = |
| 3053 | DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: OtherVT, N1: SDValue(LoNode, OtherNo), |
| 3054 | N2: SDValue(HiNode, OtherNo)); |
| 3055 | ReplaceValueWith(From: SDValue(N, OtherNo), To: OtherVal); |
| 3056 | } |
| 3057 | } |
| 3058 | |
| 3059 | void DAGTypeLegalizer::SplitVecRes_ExtendOp(SDNode *N, SDValue &Lo, |
| 3060 | SDValue &Hi) { |
| 3061 | SDLoc dl(N); |
| 3062 | EVT SrcVT = N->getOperand(Num: 0).getValueType(); |
| 3063 | EVT DestVT = N->getValueType(ResNo: 0); |
| 3064 | EVT LoVT, HiVT; |
| 3065 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: DestVT); |
| 3066 | |
| 3067 | // We can do better than a generic split operation if the extend is doing |
| 3068 | // more than just doubling the width of the elements and the following are |
| 3069 | // true: |
| 3070 | // - The number of vector elements is even, |
| 3071 | // - the source type is legal, |
| 3072 | // - the type of a split source is illegal, |
| 3073 | // - the type of an extended (by doubling element size) source is legal, and |
| 3074 | // - the type of that extended source when split is legal. |
| 3075 | // |
| 3076 | // This won't necessarily completely legalize the operation, but it will |
| 3077 | // more effectively move in the right direction and prevent falling down |
| 3078 | // to scalarization in many cases due to the input vector being split too |
| 3079 | // far. |
| 3080 | if (SrcVT.getVectorElementCount().isKnownEven() && |
| 3081 | SrcVT.getScalarSizeInBits() * 2 < DestVT.getScalarSizeInBits()) { |
| 3082 | LLVMContext &Ctx = *DAG.getContext(); |
| 3083 | EVT NewSrcVT = SrcVT.widenIntegerVectorElementType(Context&: Ctx); |
| 3084 | EVT SplitSrcVT = SrcVT.getHalfNumVectorElementsVT(Context&: Ctx); |
| 3085 | |
| 3086 | EVT SplitLoVT, SplitHiVT; |
| 3087 | std::tie(args&: SplitLoVT, args&: SplitHiVT) = DAG.GetSplitDestVTs(VT: NewSrcVT); |
| 3088 | if (TLI.isTypeLegal(VT: SrcVT) && !TLI.isTypeLegal(VT: SplitSrcVT) && |
| 3089 | TLI.isTypeLegal(VT: NewSrcVT) && TLI.isTypeLegal(VT: SplitLoVT)) { |
| 3090 | LLVM_DEBUG(dbgs() << "Split vector extend via incremental extend:" ; |
| 3091 | N->dump(&DAG); dbgs() << "\n" ); |
| 3092 | if (!N->isVPOpcode()) { |
| 3093 | // Extend the source vector by one step. |
| 3094 | SDValue NewSrc = |
| 3095 | DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: NewSrcVT, Operand: N->getOperand(Num: 0)); |
| 3096 | // Get the low and high halves of the new, extended one step, vector. |
| 3097 | std::tie(args&: Lo, args&: Hi) = DAG.SplitVector(N: NewSrc, DL: dl); |
| 3098 | // Extend those vector halves the rest of the way. |
| 3099 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: LoVT, Operand: Lo); |
| 3100 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: HiVT, Operand: Hi); |
| 3101 | return; |
| 3102 | } |
| 3103 | |
| 3104 | // Extend the source vector by one step. |
| 3105 | SDValue NewSrc = |
| 3106 | DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: NewSrcVT, N1: N->getOperand(Num: 0), |
| 3107 | N2: N->getOperand(Num: 1), N3: N->getOperand(Num: 2)); |
| 3108 | // Get the low and high halves of the new, extended one step, vector. |
| 3109 | std::tie(args&: Lo, args&: Hi) = DAG.SplitVector(N: NewSrc, DL: dl); |
| 3110 | |
| 3111 | SDValue MaskLo, MaskHi; |
| 3112 | std::tie(args&: MaskLo, args&: MaskHi) = SplitMask(Mask: N->getOperand(Num: 1)); |
| 3113 | |
| 3114 | SDValue EVLLo, EVLHi; |
| 3115 | std::tie(args&: EVLLo, args&: EVLHi) = |
| 3116 | DAG.SplitEVL(N: N->getOperand(Num: 2), VecVT: N->getValueType(ResNo: 0), DL: dl); |
| 3117 | // Extend those vector halves the rest of the way. |
| 3118 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: LoVT, Ops: {Lo, MaskLo, EVLLo}); |
| 3119 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: HiVT, Ops: {Hi, MaskHi, EVLHi}); |
| 3120 | return; |
| 3121 | } |
| 3122 | } |
| 3123 | // Fall back to the generic unary operator splitting otherwise. |
| 3124 | SplitVecRes_UnaryOp(N, Lo, Hi); |
| 3125 | } |
| 3126 | |
| 3127 | void DAGTypeLegalizer::SplitVecRes_VECTOR_SHUFFLE(ShuffleVectorSDNode *N, |
| 3128 | SDValue &Lo, SDValue &Hi) { |
| 3129 | // The low and high parts of the original input give four input vectors. |
| 3130 | SDValue Inputs[4]; |
| 3131 | SDLoc DL(N); |
| 3132 | GetSplitVector(Op: N->getOperand(Num: 0), Lo&: Inputs[0], Hi&: Inputs[1]); |
| 3133 | GetSplitVector(Op: N->getOperand(Num: 1), Lo&: Inputs[2], Hi&: Inputs[3]); |
| 3134 | EVT NewVT = Inputs[0].getValueType(); |
| 3135 | unsigned NewElts = NewVT.getVectorNumElements(); |
| 3136 | |
| 3137 | auto &&IsConstant = [](const SDValue &N) { |
| 3138 | APInt SplatValue; |
| 3139 | return N.getResNo() == 0 && |
| 3140 | (ISD::isConstantSplatVector(N: N.getNode(), SplatValue) || |
| 3141 | ISD::isBuildVectorOfConstantSDNodes(N: N.getNode())); |
| 3142 | }; |
| 3143 | auto &&BuildVector = [NewElts, &DAG = DAG, NewVT, &DL](SDValue &Input1, |
| 3144 | SDValue &Input2, |
| 3145 | ArrayRef<int> Mask) { |
| 3146 | assert(Input1->getOpcode() == ISD::BUILD_VECTOR && |
| 3147 | Input2->getOpcode() == ISD::BUILD_VECTOR && |
| 3148 | "Expected build vector node." ); |
| 3149 | EVT EltVT = NewVT.getVectorElementType(); |
| 3150 | SmallVector<SDValue> Ops(NewElts, DAG.getPOISON(VT: EltVT)); |
| 3151 | for (unsigned I = 0; I < NewElts; ++I) { |
| 3152 | if (Mask[I] == PoisonMaskElem) |
| 3153 | continue; |
| 3154 | unsigned Idx = Mask[I]; |
| 3155 | if (Idx >= NewElts) |
| 3156 | Ops[I] = Input2.getOperand(i: Idx - NewElts); |
| 3157 | else |
| 3158 | Ops[I] = Input1.getOperand(i: Idx); |
| 3159 | // Make the type of all elements the same as the element type. |
| 3160 | if (Ops[I].getValueType().bitsGT(VT: EltVT)) |
| 3161 | Ops[I] = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: EltVT, Operand: Ops[I]); |
| 3162 | } |
| 3163 | return DAG.getBuildVector(VT: NewVT, DL, Ops); |
| 3164 | }; |
| 3165 | |
| 3166 | // If Lo or Hi uses elements from at most two of the four input vectors, then |
| 3167 | // express it as a vector shuffle of those two inputs. Otherwise extract the |
| 3168 | // input elements by hand and construct the Lo/Hi output using a BUILD_VECTOR. |
| 3169 | SmallVector<int> OrigMask(N->getMask()); |
| 3170 | // Try to pack incoming shuffles/inputs. |
| 3171 | auto &&TryPeekThroughShufflesInputs = [&Inputs, &NewVT, this, NewElts, |
| 3172 | &DL](SmallVectorImpl<int> &Mask) { |
| 3173 | // Check if all inputs are shuffles of the same operands or non-shuffles. |
| 3174 | MapVector<std::pair<SDValue, SDValue>, SmallVector<unsigned>> ShufflesIdxs; |
| 3175 | for (unsigned Idx = 0; Idx < std::size(Inputs); ++Idx) { |
| 3176 | SDValue Input = Inputs[Idx]; |
| 3177 | auto *Shuffle = dyn_cast<ShuffleVectorSDNode>(Val: Input.getNode()); |
| 3178 | if (!Shuffle || |
| 3179 | Input.getOperand(i: 0).getValueType() != Input.getValueType()) |
| 3180 | continue; |
| 3181 | ShufflesIdxs[std::make_pair(x: Input.getOperand(i: 0), y: Input.getOperand(i: 1))] |
| 3182 | .push_back(Elt: Idx); |
| 3183 | ShufflesIdxs[std::make_pair(x: Input.getOperand(i: 1), y: Input.getOperand(i: 0))] |
| 3184 | .push_back(Elt: Idx); |
| 3185 | } |
| 3186 | for (auto &P : ShufflesIdxs) { |
| 3187 | if (P.second.size() < 2) |
| 3188 | continue; |
| 3189 | // Use shuffles operands instead of shuffles themselves. |
| 3190 | // 1. Adjust mask. |
| 3191 | for (int &Idx : Mask) { |
| 3192 | if (Idx == PoisonMaskElem) |
| 3193 | continue; |
| 3194 | unsigned SrcRegIdx = Idx / NewElts; |
| 3195 | if (Inputs[SrcRegIdx].isUndef()) { |
| 3196 | Idx = PoisonMaskElem; |
| 3197 | continue; |
| 3198 | } |
| 3199 | auto *Shuffle = |
| 3200 | dyn_cast<ShuffleVectorSDNode>(Val: Inputs[SrcRegIdx].getNode()); |
| 3201 | if (!Shuffle || !is_contained(Range&: P.second, Element: SrcRegIdx)) |
| 3202 | continue; |
| 3203 | int MaskElt = Shuffle->getMaskElt(Idx: Idx % NewElts); |
| 3204 | if (MaskElt == PoisonMaskElem) { |
| 3205 | Idx = PoisonMaskElem; |
| 3206 | continue; |
| 3207 | } |
| 3208 | Idx = MaskElt % NewElts + |
| 3209 | P.second[Shuffle->getOperand(Num: MaskElt / NewElts) == P.first.first |
| 3210 | ? 0 |
| 3211 | : 1] * |
| 3212 | NewElts; |
| 3213 | } |
| 3214 | // 2. Update inputs. |
| 3215 | Inputs[P.second[0]] = P.first.first; |
| 3216 | Inputs[P.second[1]] = P.first.second; |
| 3217 | // Clear the pair data. |
| 3218 | P.second.clear(); |
| 3219 | ShufflesIdxs[std::make_pair(x&: P.first.second, y&: P.first.first)].clear(); |
| 3220 | } |
| 3221 | // Check if any concat_vectors can be simplified. |
| 3222 | SmallBitVector UsedSubVector(2 * std::size(Inputs)); |
| 3223 | for (int &Idx : Mask) { |
| 3224 | if (Idx == PoisonMaskElem) |
| 3225 | continue; |
| 3226 | unsigned SrcRegIdx = Idx / NewElts; |
| 3227 | if (Inputs[SrcRegIdx].isUndef()) { |
| 3228 | Idx = PoisonMaskElem; |
| 3229 | continue; |
| 3230 | } |
| 3231 | TargetLowering::LegalizeTypeAction TypeAction = |
| 3232 | getTypeAction(VT: Inputs[SrcRegIdx].getValueType()); |
| 3233 | if (Inputs[SrcRegIdx].getOpcode() == ISD::CONCAT_VECTORS && |
| 3234 | Inputs[SrcRegIdx].getNumOperands() == 2 && |
| 3235 | !Inputs[SrcRegIdx].getOperand(i: 1).isUndef() && |
| 3236 | (TypeAction == TargetLowering::TypeLegal || |
| 3237 | TypeAction == TargetLowering::TypeWidenVector)) |
| 3238 | UsedSubVector.set(2 * SrcRegIdx + (Idx % NewElts) / (NewElts / 2)); |
| 3239 | } |
| 3240 | if (UsedSubVector.count() > 1) { |
| 3241 | SmallVector<SmallVector<std::pair<unsigned, int>, 2>> Pairs; |
| 3242 | for (unsigned I = 0; I < std::size(Inputs); ++I) { |
| 3243 | if (UsedSubVector.test(Idx: 2 * I) == UsedSubVector.test(Idx: 2 * I + 1)) |
| 3244 | continue; |
| 3245 | if (Pairs.empty() || Pairs.back().size() == 2) |
| 3246 | Pairs.emplace_back(); |
| 3247 | if (UsedSubVector.test(Idx: 2 * I)) { |
| 3248 | Pairs.back().emplace_back(Args&: I, Args: 0); |
| 3249 | } else { |
| 3250 | assert(UsedSubVector.test(2 * I + 1) && |
| 3251 | "Expected to be used one of the subvectors." ); |
| 3252 | Pairs.back().emplace_back(Args&: I, Args: 1); |
| 3253 | } |
| 3254 | } |
| 3255 | if (!Pairs.empty() && Pairs.front().size() > 1) { |
| 3256 | // Adjust mask. |
| 3257 | for (int &Idx : Mask) { |
| 3258 | if (Idx == PoisonMaskElem) |
| 3259 | continue; |
| 3260 | unsigned SrcRegIdx = Idx / NewElts; |
| 3261 | auto *It = find_if( |
| 3262 | Range&: Pairs, P: [SrcRegIdx](ArrayRef<std::pair<unsigned, int>> Idxs) { |
| 3263 | return Idxs.front().first == SrcRegIdx || |
| 3264 | Idxs.back().first == SrcRegIdx; |
| 3265 | }); |
| 3266 | if (It == Pairs.end()) |
| 3267 | continue; |
| 3268 | Idx = It->front().first * NewElts + (Idx % NewElts) % (NewElts / 2) + |
| 3269 | (SrcRegIdx == It->front().first ? 0 : (NewElts / 2)); |
| 3270 | } |
| 3271 | // Adjust inputs. |
| 3272 | for (ArrayRef<std::pair<unsigned, int>> Idxs : Pairs) { |
| 3273 | Inputs[Idxs.front().first] = DAG.getNode( |
| 3274 | Opcode: ISD::CONCAT_VECTORS, DL, |
| 3275 | VT: Inputs[Idxs.front().first].getValueType(), |
| 3276 | N1: Inputs[Idxs.front().first].getOperand(i: Idxs.front().second), |
| 3277 | N2: Inputs[Idxs.back().first].getOperand(i: Idxs.back().second)); |
| 3278 | } |
| 3279 | } |
| 3280 | } |
| 3281 | bool Changed; |
| 3282 | do { |
| 3283 | // Try to remove extra shuffles (except broadcasts) and shuffles with the |
| 3284 | // reused operands. |
| 3285 | Changed = false; |
| 3286 | for (unsigned I = 0; I < std::size(Inputs); ++I) { |
| 3287 | auto *Shuffle = dyn_cast<ShuffleVectorSDNode>(Val: Inputs[I].getNode()); |
| 3288 | if (!Shuffle) |
| 3289 | continue; |
| 3290 | if (Shuffle->getOperand(Num: 0).getValueType() != NewVT) |
| 3291 | continue; |
| 3292 | int Op = -1; |
| 3293 | if (!Inputs[I].hasOneUse() && Shuffle->getOperand(Num: 1).isUndef() && |
| 3294 | !Shuffle->isSplat()) { |
| 3295 | Op = 0; |
| 3296 | } else if (!Inputs[I].hasOneUse() && |
| 3297 | !Shuffle->getOperand(Num: 1).isUndef()) { |
| 3298 | // Find the only used operand, if possible. |
| 3299 | for (int &Idx : Mask) { |
| 3300 | if (Idx == PoisonMaskElem) |
| 3301 | continue; |
| 3302 | unsigned SrcRegIdx = Idx / NewElts; |
| 3303 | if (SrcRegIdx != I) |
| 3304 | continue; |
| 3305 | int MaskElt = Shuffle->getMaskElt(Idx: Idx % NewElts); |
| 3306 | if (MaskElt == PoisonMaskElem) { |
| 3307 | Idx = PoisonMaskElem; |
| 3308 | continue; |
| 3309 | } |
| 3310 | int OpIdx = MaskElt / NewElts; |
| 3311 | if (Op == -1) { |
| 3312 | Op = OpIdx; |
| 3313 | continue; |
| 3314 | } |
| 3315 | if (Op != OpIdx) { |
| 3316 | Op = -1; |
| 3317 | break; |
| 3318 | } |
| 3319 | } |
| 3320 | } |
| 3321 | if (Op < 0) { |
| 3322 | // Try to check if one of the shuffle operands is used already. |
| 3323 | for (int OpIdx = 0; OpIdx < 2; ++OpIdx) { |
| 3324 | if (Shuffle->getOperand(Num: OpIdx).isUndef()) |
| 3325 | continue; |
| 3326 | auto *It = find(Range&: Inputs, Val: Shuffle->getOperand(Num: OpIdx)); |
| 3327 | if (It == std::end(arr&: Inputs)) |
| 3328 | continue; |
| 3329 | int FoundOp = std::distance(first: std::begin(arr&: Inputs), last: It); |
| 3330 | // Found that operand is used already. |
| 3331 | // 1. Fix the mask for the reused operand. |
| 3332 | for (int &Idx : Mask) { |
| 3333 | if (Idx == PoisonMaskElem) |
| 3334 | continue; |
| 3335 | unsigned SrcRegIdx = Idx / NewElts; |
| 3336 | if (SrcRegIdx != I) |
| 3337 | continue; |
| 3338 | int MaskElt = Shuffle->getMaskElt(Idx: Idx % NewElts); |
| 3339 | if (MaskElt == PoisonMaskElem) { |
| 3340 | Idx = PoisonMaskElem; |
| 3341 | continue; |
| 3342 | } |
| 3343 | int MaskIdx = MaskElt / NewElts; |
| 3344 | if (OpIdx == MaskIdx) |
| 3345 | Idx = MaskElt % NewElts + FoundOp * NewElts; |
| 3346 | } |
| 3347 | // 2. Set Op to the unused OpIdx. |
| 3348 | Op = (OpIdx + 1) % 2; |
| 3349 | break; |
| 3350 | } |
| 3351 | } |
| 3352 | if (Op >= 0) { |
| 3353 | Changed = true; |
| 3354 | Inputs[I] = Shuffle->getOperand(Num: Op); |
| 3355 | // Adjust mask. |
| 3356 | for (int &Idx : Mask) { |
| 3357 | if (Idx == PoisonMaskElem) |
| 3358 | continue; |
| 3359 | unsigned SrcRegIdx = Idx / NewElts; |
| 3360 | if (SrcRegIdx != I) |
| 3361 | continue; |
| 3362 | int MaskElt = Shuffle->getMaskElt(Idx: Idx % NewElts); |
| 3363 | int OpIdx = MaskElt / NewElts; |
| 3364 | if (OpIdx != Op) |
| 3365 | continue; |
| 3366 | Idx = MaskElt % NewElts + SrcRegIdx * NewElts; |
| 3367 | } |
| 3368 | } |
| 3369 | } |
| 3370 | } while (Changed); |
| 3371 | }; |
| 3372 | TryPeekThroughShufflesInputs(OrigMask); |
| 3373 | // Proces unique inputs. |
| 3374 | auto &&MakeUniqueInputs = [&Inputs, &IsConstant, |
| 3375 | NewElts](SmallVectorImpl<int> &Mask) { |
| 3376 | SetVector<SDValue> UniqueInputs; |
| 3377 | SetVector<SDValue> UniqueConstantInputs; |
| 3378 | for (const auto &I : Inputs) { |
| 3379 | if (IsConstant(I)) |
| 3380 | UniqueConstantInputs.insert(X: I); |
| 3381 | else if (!I.isUndef()) |
| 3382 | UniqueInputs.insert(X: I); |
| 3383 | } |
| 3384 | // Adjust mask in case of reused inputs. Also, need to insert constant |
| 3385 | // inputs at first, otherwise it affects the final outcome. |
| 3386 | if (UniqueInputs.size() != std::size(Inputs)) { |
| 3387 | auto &&UniqueVec = UniqueInputs.takeVector(); |
| 3388 | auto &&UniqueConstantVec = UniqueConstantInputs.takeVector(); |
| 3389 | unsigned ConstNum = UniqueConstantVec.size(); |
| 3390 | for (int &Idx : Mask) { |
| 3391 | if (Idx == PoisonMaskElem) |
| 3392 | continue; |
| 3393 | unsigned SrcRegIdx = Idx / NewElts; |
| 3394 | if (Inputs[SrcRegIdx].isUndef()) { |
| 3395 | Idx = PoisonMaskElem; |
| 3396 | continue; |
| 3397 | } |
| 3398 | const auto It = find(Range&: UniqueConstantVec, Val: Inputs[SrcRegIdx]); |
| 3399 | if (It != UniqueConstantVec.end()) { |
| 3400 | Idx = (Idx % NewElts) + |
| 3401 | NewElts * std::distance(first: UniqueConstantVec.begin(), last: It); |
| 3402 | assert(Idx >= 0 && "Expected defined mask idx." ); |
| 3403 | continue; |
| 3404 | } |
| 3405 | const auto RegIt = find(Range&: UniqueVec, Val: Inputs[SrcRegIdx]); |
| 3406 | assert(RegIt != UniqueVec.end() && "Cannot find non-const value." ); |
| 3407 | Idx = (Idx % NewElts) + |
| 3408 | NewElts * (std::distance(first: UniqueVec.begin(), last: RegIt) + ConstNum); |
| 3409 | assert(Idx >= 0 && "Expected defined mask idx." ); |
| 3410 | } |
| 3411 | copy(Range&: UniqueConstantVec, Out: std::begin(arr&: Inputs)); |
| 3412 | copy(Range&: UniqueVec, Out: std::next(x: std::begin(arr&: Inputs), n: ConstNum)); |
| 3413 | } |
| 3414 | }; |
| 3415 | MakeUniqueInputs(OrigMask); |
| 3416 | SDValue OrigInputs[4]; |
| 3417 | copy(Range&: Inputs, Out: std::begin(arr&: OrigInputs)); |
| 3418 | for (unsigned High = 0; High < 2; ++High) { |
| 3419 | SDValue &Output = High ? Hi : Lo; |
| 3420 | |
| 3421 | // Build a shuffle mask for the output, discovering on the fly which |
| 3422 | // input vectors to use as shuffle operands. |
| 3423 | unsigned FirstMaskIdx = High * NewElts; |
| 3424 | SmallVector<int> Mask(NewElts * std::size(Inputs), PoisonMaskElem); |
| 3425 | copy(Range: ArrayRef(OrigMask).slice(N: FirstMaskIdx, M: NewElts), Out: Mask.begin()); |
| 3426 | assert(!Output && "Expected default initialized initial value." ); |
| 3427 | TryPeekThroughShufflesInputs(Mask); |
| 3428 | MakeUniqueInputs(Mask); |
| 3429 | SDValue TmpInputs[4]; |
| 3430 | copy(Range&: Inputs, Out: std::begin(arr&: TmpInputs)); |
| 3431 | // Track changes in the output registers. |
| 3432 | int UsedIdx = -1; |
| 3433 | bool SecondIteration = false; |
| 3434 | auto &&AccumulateResults = [&UsedIdx, &SecondIteration](unsigned Idx) { |
| 3435 | if (UsedIdx < 0) { |
| 3436 | UsedIdx = Idx; |
| 3437 | return false; |
| 3438 | } |
| 3439 | if (UsedIdx >= 0 && static_cast<unsigned>(UsedIdx) == Idx) |
| 3440 | SecondIteration = true; |
| 3441 | return SecondIteration; |
| 3442 | }; |
| 3443 | processShuffleMasks( |
| 3444 | Mask, NumOfSrcRegs: std::size(Inputs), NumOfDestRegs: std::size(Inputs), |
| 3445 | /*NumOfUsedRegs=*/1, |
| 3446 | NoInputAction: [&Output, &DAG = DAG, NewVT]() { Output = DAG.getPOISON(VT: NewVT); }, |
| 3447 | SingleInputAction: [&Output, &DAG = DAG, NewVT, &DL, &Inputs, |
| 3448 | &BuildVector](ArrayRef<int> Mask, unsigned Idx, unsigned /*Unused*/) { |
| 3449 | if (Inputs[Idx]->getOpcode() == ISD::BUILD_VECTOR) |
| 3450 | Output = BuildVector(Inputs[Idx], Inputs[Idx], Mask); |
| 3451 | else |
| 3452 | Output = DAG.getVectorShuffle(VT: NewVT, dl: DL, N1: Inputs[Idx], |
| 3453 | N2: DAG.getPOISON(VT: NewVT), Mask); |
| 3454 | Inputs[Idx] = Output; |
| 3455 | }, |
| 3456 | ManyInputsAction: [&AccumulateResults, &Output, &DAG = DAG, NewVT, &DL, &Inputs, |
| 3457 | &TmpInputs, &BuildVector](ArrayRef<int> Mask, unsigned Idx1, |
| 3458 | unsigned Idx2, bool /*Unused*/) { |
| 3459 | if (AccumulateResults(Idx1)) { |
| 3460 | if (Inputs[Idx1]->getOpcode() == ISD::BUILD_VECTOR && |
| 3461 | Inputs[Idx2]->getOpcode() == ISD::BUILD_VECTOR) |
| 3462 | Output = BuildVector(Inputs[Idx1], Inputs[Idx2], Mask); |
| 3463 | else |
| 3464 | Output = DAG.getVectorShuffle(VT: NewVT, dl: DL, N1: Inputs[Idx1], |
| 3465 | N2: Inputs[Idx2], Mask); |
| 3466 | } else { |
| 3467 | if (TmpInputs[Idx1]->getOpcode() == ISD::BUILD_VECTOR && |
| 3468 | TmpInputs[Idx2]->getOpcode() == ISD::BUILD_VECTOR) |
| 3469 | Output = BuildVector(TmpInputs[Idx1], TmpInputs[Idx2], Mask); |
| 3470 | else |
| 3471 | Output = DAG.getVectorShuffle(VT: NewVT, dl: DL, N1: TmpInputs[Idx1], |
| 3472 | N2: TmpInputs[Idx2], Mask); |
| 3473 | } |
| 3474 | Inputs[Idx1] = Output; |
| 3475 | }); |
| 3476 | copy(Range&: OrigInputs, Out: std::begin(arr&: Inputs)); |
| 3477 | } |
| 3478 | } |
| 3479 | |
| 3480 | void DAGTypeLegalizer::SplitVecRes_VAARG(SDNode *N, SDValue &Lo, SDValue &Hi) { |
| 3481 | EVT OVT = N->getValueType(ResNo: 0); |
| 3482 | EVT NVT = OVT.getHalfNumVectorElementsVT(Context&: *DAG.getContext()); |
| 3483 | SDValue Chain = N->getOperand(Num: 0); |
| 3484 | SDValue Ptr = N->getOperand(Num: 1); |
| 3485 | SDValue SV = N->getOperand(Num: 2); |
| 3486 | SDLoc dl(N); |
| 3487 | |
| 3488 | const Align Alignment = |
| 3489 | DAG.getDataLayout().getABITypeAlign(Ty: NVT.getTypeForEVT(Context&: *DAG.getContext())); |
| 3490 | |
| 3491 | Lo = DAG.getVAArg(VT: NVT, dl, Chain, Ptr, SV, Align: Alignment.value()); |
| 3492 | Hi = DAG.getVAArg(VT: NVT, dl, Chain: Lo.getValue(R: 1), Ptr, SV, Align: Alignment.value()); |
| 3493 | Chain = Hi.getValue(R: 1); |
| 3494 | |
| 3495 | // Modified the chain - switch anything that used the old chain to use |
| 3496 | // the new one. |
| 3497 | ReplaceValueWith(From: SDValue(N, 1), To: Chain); |
| 3498 | } |
| 3499 | |
| 3500 | void DAGTypeLegalizer::SplitVecRes_FP_TO_XINT_SAT(SDNode *N, SDValue &Lo, |
| 3501 | SDValue &Hi) { |
| 3502 | EVT DstVTLo, DstVTHi; |
| 3503 | std::tie(args&: DstVTLo, args&: DstVTHi) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 3504 | SDLoc dl(N); |
| 3505 | |
| 3506 | SDValue SrcLo, SrcHi; |
| 3507 | EVT SrcVT = N->getOperand(Num: 0).getValueType(); |
| 3508 | if (getTypeAction(VT: SrcVT) == TargetLowering::TypeSplitVector) |
| 3509 | GetSplitVector(Op: N->getOperand(Num: 0), Lo&: SrcLo, Hi&: SrcHi); |
| 3510 | else |
| 3511 | std::tie(args&: SrcLo, args&: SrcHi) = DAG.SplitVectorOperand(N, OpNo: 0); |
| 3512 | |
| 3513 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: DstVTLo, N1: SrcLo, N2: N->getOperand(Num: 1)); |
| 3514 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: DstVTHi, N1: SrcHi, N2: N->getOperand(Num: 1)); |
| 3515 | } |
| 3516 | |
| 3517 | void DAGTypeLegalizer::SplitVecRes_VECTOR_REVERSE(SDNode *N, SDValue &Lo, |
| 3518 | SDValue &Hi) { |
| 3519 | SDValue InLo, InHi; |
| 3520 | GetSplitVector(Op: N->getOperand(Num: 0), Lo&: InLo, Hi&: InHi); |
| 3521 | SDLoc DL(N); |
| 3522 | |
| 3523 | Lo = DAG.getNode(Opcode: ISD::VECTOR_REVERSE, DL, VT: InHi.getValueType(), Operand: InHi); |
| 3524 | Hi = DAG.getNode(Opcode: ISD::VECTOR_REVERSE, DL, VT: InLo.getValueType(), Operand: InLo); |
| 3525 | } |
| 3526 | |
| 3527 | void DAGTypeLegalizer::SplitVecRes_VECTOR_SPLICE(SDNode *N, SDValue &Lo, |
| 3528 | SDValue &Hi) { |
| 3529 | SDLoc DL(N); |
| 3530 | |
| 3531 | SDValue Expanded = TLI.expandVectorSplice(Node: N, DAG); |
| 3532 | std::tie(args&: Lo, args&: Hi) = DAG.SplitVector(N: Expanded, DL); |
| 3533 | } |
| 3534 | |
| 3535 | void DAGTypeLegalizer::SplitVecRes_VP_REVERSE(SDNode *N, SDValue &Lo, |
| 3536 | SDValue &Hi) { |
| 3537 | EVT VT = N->getValueType(ResNo: 0); |
| 3538 | SDValue Val = N->getOperand(Num: 0); |
| 3539 | SDValue Mask = N->getOperand(Num: 1); |
| 3540 | SDValue EVL = N->getOperand(Num: 2); |
| 3541 | SDLoc DL(N); |
| 3542 | |
| 3543 | // The stack round-trip uses a byte stride, so a sub-byte element (e.g. i1) |
| 3544 | // would get stride 0 and alias every lane. Widen to a byte integer, reverse, |
| 3545 | // then truncate back. |
| 3546 | EVT OrigVT = VT; |
| 3547 | if (!VT.getVectorElementType().isByteSized()) { |
| 3548 | EVT WideEltVT = VT.getVectorElementType().changeTypeToInteger(); |
| 3549 | WideEltVT = WideEltVT.getRoundIntegerType(Context&: *DAG.getContext()); |
| 3550 | VT = VT.changeVectorElementType(Context&: *DAG.getContext(), EltVT: WideEltVT); |
| 3551 | Val = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT, Operand: Val); |
| 3552 | } |
| 3553 | |
| 3554 | // Fallback to VP_STRIDED_STORE to stack followed by VP_LOAD. |
| 3555 | Align Alignment = DAG.getReducedAlign(VT, /*UseABI=*/false); |
| 3556 | |
| 3557 | EVT MemVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: VT.getVectorElementType(), |
| 3558 | EC: VT.getVectorElementCount()); |
| 3559 | SDValue StackPtr = DAG.CreateStackTemporary(Bytes: MemVT.getStoreSize(), Alignment); |
| 3560 | EVT PtrVT = StackPtr.getValueType(); |
| 3561 | auto &MF = DAG.getMachineFunction(); |
| 3562 | auto FrameIndex = cast<FrameIndexSDNode>(Val: StackPtr.getNode())->getIndex(); |
| 3563 | auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: FrameIndex); |
| 3564 | |
| 3565 | MachineMemOperand *StoreMMO = DAG.getMachineFunction().getMachineMemOperand( |
| 3566 | PtrInfo, F: MachineMemOperand::MOStore, Size: LocationSize::beforeOrAfterPointer(), |
| 3567 | BaseAlignment: Alignment); |
| 3568 | MachineMemOperand *LoadMMO = DAG.getMachineFunction().getMachineMemOperand( |
| 3569 | PtrInfo, F: MachineMemOperand::MOLoad, Size: LocationSize::beforeOrAfterPointer(), |
| 3570 | BaseAlignment: Alignment); |
| 3571 | |
| 3572 | unsigned EltWidth = VT.getScalarSizeInBits() / 8; |
| 3573 | SDValue NumElemMinus1 = |
| 3574 | DAG.getNode(Opcode: ISD::SUB, DL, VT: PtrVT, N1: DAG.getZExtOrTrunc(Op: EVL, DL, VT: PtrVT), |
| 3575 | N2: DAG.getConstant(Val: 1, DL, VT: PtrVT)); |
| 3576 | SDValue StartOffset = DAG.getNode(Opcode: ISD::MUL, DL, VT: PtrVT, N1: NumElemMinus1, |
| 3577 | N2: DAG.getConstant(Val: EltWidth, DL, VT: PtrVT)); |
| 3578 | SDValue StorePtr = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: StackPtr, N2: StartOffset); |
| 3579 | SDValue Stride = DAG.getConstant(Val: -(int64_t)EltWidth, DL, VT: PtrVT); |
| 3580 | |
| 3581 | SDValue TrueMask = DAG.getBoolConstant(V: true, DL, VT: Mask.getValueType(), OpVT: VT); |
| 3582 | SDValue Store = DAG.getStridedStoreVP(Chain: DAG.getEntryNode(), DL, Val, Ptr: StorePtr, |
| 3583 | Offset: DAG.getPOISON(VT: PtrVT), Stride, Mask: TrueMask, |
| 3584 | EVL, MemVT, MMO: StoreMMO, AM: ISD::UNINDEXED); |
| 3585 | |
| 3586 | SDValue Load = DAG.getLoadVP(VT, dl: DL, Chain: Store, Ptr: StackPtr, Mask, EVL, MMO: LoadMMO); |
| 3587 | |
| 3588 | // Truncate back if we widened above. |
| 3589 | if (OrigVT != VT) |
| 3590 | Load = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: OrigVT, Operand: Load); |
| 3591 | |
| 3592 | std::tie(args&: Lo, args&: Hi) = DAG.SplitVector(N: Load, DL); |
| 3593 | } |
| 3594 | |
| 3595 | void DAGTypeLegalizer::SplitVecRes_VP_SPLICE(SDNode *N, SDValue &Lo, |
| 3596 | SDValue &Hi) { |
| 3597 | EVT VT = N->getValueType(ResNo: 0); |
| 3598 | SDValue V1 = N->getOperand(Num: 0); |
| 3599 | SDValue V2 = N->getOperand(Num: 1); |
| 3600 | int64_t Imm = cast<ConstantSDNode>(Val: N->getOperand(Num: 2))->getSExtValue(); |
| 3601 | SDValue Mask = N->getOperand(Num: 3); |
| 3602 | SDValue EVL1 = N->getOperand(Num: 4); |
| 3603 | SDValue EVL2 = N->getOperand(Num: 5); |
| 3604 | SDLoc DL(N); |
| 3605 | |
| 3606 | // Since EVL2 is considered the real VL it gets promoted during |
| 3607 | // SelectionDAGBuilder. Promote EVL1 here if needed. |
| 3608 | if (getTypeAction(VT: EVL1.getValueType()) == TargetLowering::TypePromoteInteger) |
| 3609 | EVL1 = ZExtPromotedInteger(Op: EVL1); |
| 3610 | |
| 3611 | // The stack splice addresses elements by byte offset/stride, which breaks for |
| 3612 | // a sub-byte element (e.g. i1): getVectorElementPointer asserts and the |
| 3613 | // stride is 0. Widen to a byte integer, splice, then truncate back. |
| 3614 | EVT OrigVT = VT; |
| 3615 | if (!VT.getVectorElementType().isByteSized()) { |
| 3616 | EVT WideEltVT = VT.getVectorElementType().changeTypeToInteger(); |
| 3617 | WideEltVT = WideEltVT.getRoundIntegerType(Context&: *DAG.getContext()); |
| 3618 | VT = VT.changeVectorElementType(Context&: *DAG.getContext(), EltVT: WideEltVT); |
| 3619 | V1 = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT, Operand: V1); |
| 3620 | V2 = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT, Operand: V2); |
| 3621 | } |
| 3622 | |
| 3623 | Align Alignment = DAG.getReducedAlign(VT, /*UseABI=*/false); |
| 3624 | |
| 3625 | EVT MemVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: VT.getVectorElementType(), |
| 3626 | EC: VT.getVectorElementCount() * 2); |
| 3627 | SDValue StackPtr = DAG.CreateStackTemporary(Bytes: MemVT.getStoreSize(), Alignment); |
| 3628 | EVT PtrVT = StackPtr.getValueType(); |
| 3629 | auto &MF = DAG.getMachineFunction(); |
| 3630 | auto FrameIndex = cast<FrameIndexSDNode>(Val: StackPtr.getNode())->getIndex(); |
| 3631 | auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: FrameIndex); |
| 3632 | |
| 3633 | MachineMemOperand *StoreMMO = DAG.getMachineFunction().getMachineMemOperand( |
| 3634 | PtrInfo, F: MachineMemOperand::MOStore, Size: LocationSize::beforeOrAfterPointer(), |
| 3635 | BaseAlignment: Alignment); |
| 3636 | MachineMemOperand *LoadMMO = DAG.getMachineFunction().getMachineMemOperand( |
| 3637 | PtrInfo, F: MachineMemOperand::MOLoad, Size: LocationSize::beforeOrAfterPointer(), |
| 3638 | BaseAlignment: Alignment); |
| 3639 | |
| 3640 | SDValue EltByteSize = |
| 3641 | DAG.getTypeSize(DL, VT: PtrVT, TS: VT.getVectorElementType().getStoreSize()); |
| 3642 | SDValue EVL1Ptr = DAG.getZExtOrTrunc(Op: EVL1, DL, VT: PtrVT); |
| 3643 | SDValue EVL1Bytes = DAG.getNode(Opcode: ISD::MUL, DL, VT: PtrVT, N1: EVL1Ptr, N2: EltByteSize); |
| 3644 | // Clip EVL1Bytes to make sure we stay within the stack object. |
| 3645 | SDValue VTBytes = DAG.getTypeSize(DL, VT: PtrVT, TS: VT.getStoreSize()); |
| 3646 | EVL1Bytes = DAG.getNode(Opcode: ISD::UMIN, DL, VT: PtrVT, N1: EVL1Bytes, N2: VTBytes); |
| 3647 | SDValue StackPtr2 = DAG.getMemBasePlusOffset(Base: StackPtr, Offset: EVL1Bytes, DL); |
| 3648 | SDValue PoisonPtr = DAG.getPOISON(VT: PtrVT); |
| 3649 | |
| 3650 | SDValue TrueMask = DAG.getBoolConstant(V: true, DL, VT: Mask.getValueType(), OpVT: VT); |
| 3651 | SDValue StoreV1 = |
| 3652 | DAG.getStoreVP(Chain: DAG.getEntryNode(), dl: DL, Val: V1, Ptr: StackPtr, Offset: PoisonPtr, Mask: TrueMask, |
| 3653 | EVL: EVL1, MemVT: V1.getValueType(), MMO: StoreMMO, AM: ISD::UNINDEXED); |
| 3654 | |
| 3655 | SDValue StoreV2 = |
| 3656 | DAG.getStoreVP(Chain: StoreV1, dl: DL, Val: V2, Ptr: StackPtr2, Offset: PoisonPtr, Mask: TrueMask, EVL: EVL2, |
| 3657 | MemVT: V2.getValueType(), MMO: StoreMMO, AM: ISD::UNINDEXED); |
| 3658 | |
| 3659 | SDValue Load; |
| 3660 | if (Imm >= 0) { |
| 3661 | StackPtr = TLI.getVectorElementPointer(DAG, VecPtr: StackPtr, VecVT: VT, Index: N->getOperand(Num: 2)); |
| 3662 | Load = DAG.getLoadVP(VT, dl: DL, Chain: StoreV2, Ptr: StackPtr, Mask, EVL: EVL2, MMO: LoadMMO); |
| 3663 | } else { |
| 3664 | uint64_t TrailingElts = -Imm; |
| 3665 | unsigned EltWidth = VT.getScalarSizeInBits() / 8; |
| 3666 | SDValue TrailingBytes = DAG.getConstant(Val: TrailingElts * EltWidth, DL, VT: PtrVT); |
| 3667 | |
| 3668 | // Make sure TrailingBytes doesn't exceed the size of vec1. |
| 3669 | SDValue OffsetToV2 = DAG.getNode(Opcode: ISD::SUB, DL, VT: PtrVT, N1: StackPtr2, N2: StackPtr); |
| 3670 | TrailingBytes = |
| 3671 | DAG.getNode(Opcode: ISD::UMIN, DL, VT: PtrVT, N1: TrailingBytes, N2: OffsetToV2); |
| 3672 | |
| 3673 | // Calculate the start address of the spliced result. |
| 3674 | StackPtr2 = DAG.getNode(Opcode: ISD::SUB, DL, VT: PtrVT, N1: StackPtr2, N2: TrailingBytes); |
| 3675 | Load = DAG.getLoadVP(VT, dl: DL, Chain: StoreV2, Ptr: StackPtr2, Mask, EVL: EVL2, MMO: LoadMMO); |
| 3676 | } |
| 3677 | |
| 3678 | // Truncate back if we widened above. |
| 3679 | if (OrigVT != VT) |
| 3680 | Load = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: OrigVT, Operand: Load); |
| 3681 | |
| 3682 | EVT LoVT, HiVT; |
| 3683 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: OrigVT); |
| 3684 | Lo = DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL, VT: LoVT, N1: Load, |
| 3685 | N2: DAG.getVectorIdxConstant(Val: 0, DL)); |
| 3686 | Hi = |
| 3687 | DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL, VT: HiVT, N1: Load, |
| 3688 | N2: DAG.getVectorIdxConstant(Val: LoVT.getVectorMinNumElements(), DL)); |
| 3689 | } |
| 3690 | |
| 3691 | void DAGTypeLegalizer::SplitVecRes_PARTIAL_REDUCE_MLA(SDNode *N, SDValue &Lo, |
| 3692 | SDValue &Hi) { |
| 3693 | SDLoc DL(N); |
| 3694 | SDValue Acc = N->getOperand(Num: 0); |
| 3695 | SDValue Input1 = N->getOperand(Num: 1); |
| 3696 | SDValue Input2 = N->getOperand(Num: 2); |
| 3697 | |
| 3698 | SDValue AccLo, AccHi; |
| 3699 | GetSplitVector(Op: Acc, Lo&: AccLo, Hi&: AccHi); |
| 3700 | unsigned Opcode = N->getOpcode(); |
| 3701 | |
| 3702 | // If the input types don't need splitting, just accumulate into the |
| 3703 | // low part of the accumulator. |
| 3704 | if (getTypeAction(VT: Input1.getValueType()) != TargetLowering::TypeSplitVector) { |
| 3705 | Lo = DAG.getNode(Opcode, DL, VT: AccLo.getValueType(), N1: AccLo, N2: Input1, N3: Input2); |
| 3706 | Hi = AccHi; |
| 3707 | return; |
| 3708 | } |
| 3709 | |
| 3710 | SDValue Input1Lo, Input1Hi; |
| 3711 | SDValue Input2Lo, Input2Hi; |
| 3712 | GetSplitVector(Op: Input1, Lo&: Input1Lo, Hi&: Input1Hi); |
| 3713 | GetSplitVector(Op: Input2, Lo&: Input2Lo, Hi&: Input2Hi); |
| 3714 | EVT ResultVT = AccLo.getValueType(); |
| 3715 | |
| 3716 | Lo = DAG.getNode(Opcode, DL, VT: ResultVT, N1: AccLo, N2: Input1Lo, N3: Input2Lo); |
| 3717 | Hi = DAG.getNode(Opcode, DL, VT: ResultVT, N1: AccHi, N2: Input1Hi, N3: Input2Hi); |
| 3718 | } |
| 3719 | |
| 3720 | void DAGTypeLegalizer::SplitVecRes_GET_ACTIVE_LANE_MASK(SDNode *N, SDValue &Lo, |
| 3721 | SDValue &Hi) { |
| 3722 | SDLoc DL(N); |
| 3723 | SDValue Op0 = N->getOperand(Num: 0); |
| 3724 | SDValue Op1 = N->getOperand(Num: 1); |
| 3725 | EVT OpVT = Op0.getValueType(); |
| 3726 | |
| 3727 | EVT LoVT, HiVT; |
| 3728 | std::tie(args&: LoVT, args&: HiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 3729 | |
| 3730 | Lo = DAG.getNode(Opcode: ISD::GET_ACTIVE_LANE_MASK, DL, VT: LoVT, N1: Op0, N2: Op1); |
| 3731 | SDValue LoElts = DAG.getElementCount(DL, VT: OpVT, EC: LoVT.getVectorElementCount()); |
| 3732 | SDValue HiStartVal = DAG.getNode(Opcode: ISD::UADDSAT, DL, VT: OpVT, N1: Op0, N2: LoElts); |
| 3733 | Hi = DAG.getNode(Opcode: ISD::GET_ACTIVE_LANE_MASK, DL, VT: HiVT, N1: HiStartVal, N2: Op1); |
| 3734 | } |
| 3735 | |
| 3736 | void DAGTypeLegalizer::SplitVecRes_VECTOR_DEINTERLEAVE(SDNode *N) { |
| 3737 | unsigned Factor = N->getNumOperands(); |
| 3738 | |
| 3739 | SmallVector<SDValue, 8> Ops(Factor * 2); |
| 3740 | for (unsigned i = 0; i != Factor; ++i) { |
| 3741 | SDValue OpLo, OpHi; |
| 3742 | GetSplitVector(Op: N->getOperand(Num: i), Lo&: OpLo, Hi&: OpHi); |
| 3743 | Ops[i * 2] = OpLo; |
| 3744 | Ops[i * 2 + 1] = OpHi; |
| 3745 | } |
| 3746 | |
| 3747 | SmallVector<EVT, 8> VTs(Factor, Ops[0].getValueType()); |
| 3748 | |
| 3749 | SDLoc DL(N); |
| 3750 | SDValue ResLo = DAG.getNode(Opcode: ISD::VECTOR_DEINTERLEAVE, DL, ResultTys: VTs, |
| 3751 | Ops: ArrayRef(Ops).slice(N: 0, M: Factor)); |
| 3752 | SDValue ResHi = DAG.getNode(Opcode: ISD::VECTOR_DEINTERLEAVE, DL, ResultTys: VTs, |
| 3753 | Ops: ArrayRef(Ops).slice(N: Factor, M: Factor)); |
| 3754 | |
| 3755 | for (unsigned i = 0; i != Factor; ++i) |
| 3756 | SetSplitVector(Op: SDValue(N, i), Lo: ResLo.getValue(R: i), Hi: ResHi.getValue(R: i)); |
| 3757 | } |
| 3758 | |
| 3759 | void DAGTypeLegalizer::SplitVecRes_VECTOR_INTERLEAVE(SDNode *N) { |
| 3760 | unsigned Factor = N->getNumOperands(); |
| 3761 | |
| 3762 | SmallVector<SDValue, 8> Ops(Factor * 2); |
| 3763 | for (unsigned i = 0; i != Factor; ++i) { |
| 3764 | SDValue OpLo, OpHi; |
| 3765 | GetSplitVector(Op: N->getOperand(Num: i), Lo&: OpLo, Hi&: OpHi); |
| 3766 | Ops[i] = OpLo; |
| 3767 | Ops[i + Factor] = OpHi; |
| 3768 | } |
| 3769 | |
| 3770 | SmallVector<EVT, 8> VTs(Factor, Ops[0].getValueType()); |
| 3771 | |
| 3772 | SDLoc DL(N); |
| 3773 | SDValue Res[] = {DAG.getNode(Opcode: ISD::VECTOR_INTERLEAVE, DL, ResultTys: VTs, |
| 3774 | Ops: ArrayRef(Ops).slice(N: 0, M: Factor)), |
| 3775 | DAG.getNode(Opcode: ISD::VECTOR_INTERLEAVE, DL, ResultTys: VTs, |
| 3776 | Ops: ArrayRef(Ops).slice(N: Factor, M: Factor))}; |
| 3777 | |
| 3778 | for (unsigned i = 0; i != Factor; ++i) { |
| 3779 | unsigned IdxLo = 2 * i; |
| 3780 | unsigned IdxHi = 2 * i + 1; |
| 3781 | SetSplitVector(Op: SDValue(N, i), Lo: Res[IdxLo / Factor].getValue(R: IdxLo % Factor), |
| 3782 | Hi: Res[IdxHi / Factor].getValue(R: IdxHi % Factor)); |
| 3783 | } |
| 3784 | } |
| 3785 | |
| 3786 | //===----------------------------------------------------------------------===// |
| 3787 | // Operand Vector Splitting |
| 3788 | //===----------------------------------------------------------------------===// |
| 3789 | |
| 3790 | /// This method is called when the specified operand of the specified node is |
| 3791 | /// found to need vector splitting. At this point, all of the result types of |
| 3792 | /// the node are known to be legal, but other operands of the node may need |
| 3793 | /// legalization as well as the specified one. |
| 3794 | bool DAGTypeLegalizer::SplitVectorOperand(SDNode *N, unsigned OpNo) { |
| 3795 | LLVM_DEBUG(dbgs() << "Split node operand: " ; N->dump(&DAG)); |
| 3796 | SDValue Res = SDValue(); |
| 3797 | |
| 3798 | // See if the target wants to custom split this node. |
| 3799 | if (CustomLowerNode(N, VT: N->getOperand(Num: OpNo).getValueType(), LegalizeResult: false)) |
| 3800 | return false; |
| 3801 | |
| 3802 | switch (N->getOpcode()) { |
| 3803 | default: |
| 3804 | #ifndef NDEBUG |
| 3805 | dbgs() << "SplitVectorOperand Op #" << OpNo << ": " ; |
| 3806 | N->dump(&DAG); |
| 3807 | dbgs() << "\n" ; |
| 3808 | #endif |
| 3809 | report_fatal_error(reason: "Do not know how to split this operator's " |
| 3810 | "operand!\n" ); |
| 3811 | |
| 3812 | case ISD::VP_SETCC: |
| 3813 | case ISD::STRICT_FSETCC: |
| 3814 | case ISD::STRICT_FSETCCS: |
| 3815 | case ISD::SETCC: Res = SplitVecOp_VSETCC(N); break; |
| 3816 | case ISD::BITCAST: Res = SplitVecOp_BITCAST(N); break; |
| 3817 | case ISD::EXTRACT_SUBVECTOR: Res = SplitVecOp_EXTRACT_SUBVECTOR(N); break; |
| 3818 | case ISD::INSERT_SUBVECTOR: Res = SplitVecOp_INSERT_SUBVECTOR(N, OpNo); break; |
| 3819 | case ISD::EXTRACT_VECTOR_ELT:Res = SplitVecOp_EXTRACT_VECTOR_ELT(N); break; |
| 3820 | case ISD::CONCAT_VECTORS: Res = SplitVecOp_CONCAT_VECTORS(N); break; |
| 3821 | case ISD::VECTOR_FIND_LAST_ACTIVE: |
| 3822 | Res = SplitVecOp_VECTOR_FIND_LAST_ACTIVE(N); |
| 3823 | break; |
| 3824 | case ISD::VP_TRUNCATE: |
| 3825 | case ISD::TRUNCATE: |
| 3826 | Res = SplitVecOp_TruncateHelper(N); |
| 3827 | break; |
| 3828 | case ISD::STRICT_FP_ROUND: |
| 3829 | case ISD::VP_FP_ROUND: |
| 3830 | case ISD::FP_ROUND: |
| 3831 | case ISD::CONVERT_FROM_ARBITRARY_FP: |
| 3832 | case ISD::CONVERT_TO_ARBITRARY_FP: |
| 3833 | Res = SplitVecOp_FP_ROUND(N); |
| 3834 | break; |
| 3835 | case ISD::FCOPYSIGN: Res = SplitVecOp_FPOpDifferentTypes(N); break; |
| 3836 | case ISD::STORE: |
| 3837 | Res = SplitVecOp_STORE(N: cast<StoreSDNode>(Val: N), OpNo); |
| 3838 | break; |
| 3839 | case ISD::ATOMIC_STORE: |
| 3840 | Res = SplitVecOp_ATOMIC_STORE(N: cast<AtomicSDNode>(Val: N)); |
| 3841 | break; |
| 3842 | case ISD::VP_STORE: |
| 3843 | Res = SplitVecOp_VP_STORE(N: cast<VPStoreSDNode>(Val: N), OpNo); |
| 3844 | break; |
| 3845 | case ISD::EXPERIMENTAL_VP_STRIDED_STORE: |
| 3846 | Res = SplitVecOp_VP_STRIDED_STORE(N: cast<VPStridedStoreSDNode>(Val: N), OpNo); |
| 3847 | break; |
| 3848 | case ISD::MSTORE: |
| 3849 | Res = SplitVecOp_MSTORE(N: cast<MaskedStoreSDNode>(Val: N), OpNo); |
| 3850 | break; |
| 3851 | case ISD::MSCATTER: |
| 3852 | case ISD::VP_SCATTER: |
| 3853 | Res = SplitVecOp_Scatter(N: cast<MemSDNode>(Val: N), OpNo); |
| 3854 | break; |
| 3855 | case ISD::MGATHER: |
| 3856 | case ISD::VP_GATHER: |
| 3857 | Res = SplitVecOp_Gather(MGT: cast<MemSDNode>(Val: N), OpNo); |
| 3858 | break; |
| 3859 | case ISD::VSELECT: |
| 3860 | Res = SplitVecOp_VSELECT(N, OpNo); |
| 3861 | break; |
| 3862 | case ISD::VECTOR_COMPRESS: |
| 3863 | Res = SplitVecOp_VECTOR_COMPRESS(N, OpNo); |
| 3864 | break; |
| 3865 | case ISD::STRICT_SINT_TO_FP: |
| 3866 | case ISD::STRICT_UINT_TO_FP: |
| 3867 | case ISD::SINT_TO_FP: |
| 3868 | case ISD::UINT_TO_FP: |
| 3869 | case ISD::VP_SINT_TO_FP: |
| 3870 | case ISD::VP_UINT_TO_FP: |
| 3871 | if (N->getValueType(ResNo: 0).bitsLT( |
| 3872 | VT: N->getOperand(Num: N->isStrictFPOpcode() ? 1 : 0).getValueType())) |
| 3873 | Res = SplitVecOp_TruncateHelper(N); |
| 3874 | else |
| 3875 | Res = SplitVecOp_UnaryOp(N); |
| 3876 | break; |
| 3877 | case ISD::FP_TO_SINT_SAT: |
| 3878 | case ISD::FP_TO_UINT_SAT: |
| 3879 | Res = SplitVecOp_FP_TO_XINT_SAT(N); |
| 3880 | break; |
| 3881 | case ISD::FP_TO_SINT: |
| 3882 | case ISD::FP_TO_UINT: |
| 3883 | case ISD::VP_FP_TO_SINT: |
| 3884 | case ISD::VP_FP_TO_UINT: |
| 3885 | case ISD::STRICT_FP_TO_SINT: |
| 3886 | case ISD::STRICT_FP_TO_UINT: |
| 3887 | case ISD::STRICT_FP_EXTEND: |
| 3888 | case ISD::FP_EXTEND: |
| 3889 | case ISD::SIGN_EXTEND: |
| 3890 | case ISD::ZERO_EXTEND: |
| 3891 | case ISD::ANY_EXTEND: |
| 3892 | case ISD::FTRUNC: |
| 3893 | case ISD::LROUND: |
| 3894 | case ISD::LLROUND: |
| 3895 | case ISD::LRINT: |
| 3896 | case ISD::LLRINT: |
| 3897 | Res = SplitVecOp_UnaryOp(N); |
| 3898 | break; |
| 3899 | case ISD::FLDEXP: |
| 3900 | Res = SplitVecOp_FPOpDifferentTypes(N); |
| 3901 | break; |
| 3902 | |
| 3903 | case ISD::SCMP: |
| 3904 | case ISD::UCMP: |
| 3905 | Res = SplitVecOp_CMP(N); |
| 3906 | break; |
| 3907 | |
| 3908 | case ISD::FAKE_USE: |
| 3909 | Res = SplitVecOp_FAKE_USE(N); |
| 3910 | break; |
| 3911 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 3912 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 3913 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 3914 | Res = SplitVecOp_ExtVecInRegOp(N); |
| 3915 | break; |
| 3916 | |
| 3917 | case ISD::VECREDUCE_FADD: |
| 3918 | case ISD::VECREDUCE_FMUL: |
| 3919 | case ISD::VECREDUCE_ADD: |
| 3920 | case ISD::VECREDUCE_MUL: |
| 3921 | case ISD::VECREDUCE_AND: |
| 3922 | case ISD::VECREDUCE_OR: |
| 3923 | case ISD::VECREDUCE_XOR: |
| 3924 | case ISD::VECREDUCE_SMAX: |
| 3925 | case ISD::VECREDUCE_SMIN: |
| 3926 | case ISD::VECREDUCE_UMAX: |
| 3927 | case ISD::VECREDUCE_UMIN: |
| 3928 | case ISD::VECREDUCE_FMAX: |
| 3929 | case ISD::VECREDUCE_FMIN: |
| 3930 | case ISD::VECREDUCE_FMAXIMUM: |
| 3931 | case ISD::VECREDUCE_FMINIMUM: |
| 3932 | Res = SplitVecOp_VECREDUCE(N, OpNo); |
| 3933 | break; |
| 3934 | case ISD::VECREDUCE_SEQ_FADD: |
| 3935 | case ISD::VECREDUCE_SEQ_FMUL: |
| 3936 | Res = SplitVecOp_VECREDUCE_SEQ(N); |
| 3937 | break; |
| 3938 | case ISD::VP_REDUCE_FADD: |
| 3939 | case ISD::VP_REDUCE_SEQ_FADD: |
| 3940 | case ISD::VP_REDUCE_FMUL: |
| 3941 | case ISD::VP_REDUCE_SEQ_FMUL: |
| 3942 | case ISD::VP_REDUCE_ADD: |
| 3943 | case ISD::VP_REDUCE_MUL: |
| 3944 | case ISD::VP_REDUCE_AND: |
| 3945 | case ISD::VP_REDUCE_OR: |
| 3946 | case ISD::VP_REDUCE_XOR: |
| 3947 | case ISD::VP_REDUCE_SMAX: |
| 3948 | case ISD::VP_REDUCE_SMIN: |
| 3949 | case ISD::VP_REDUCE_UMAX: |
| 3950 | case ISD::VP_REDUCE_UMIN: |
| 3951 | case ISD::VP_REDUCE_FMAX: |
| 3952 | case ISD::VP_REDUCE_FMIN: |
| 3953 | case ISD::VP_REDUCE_FMAXIMUM: |
| 3954 | case ISD::VP_REDUCE_FMINIMUM: |
| 3955 | Res = SplitVecOp_VP_REDUCE(N, OpNo); |
| 3956 | break; |
| 3957 | case ISD::CTTZ_ELTS: |
| 3958 | case ISD::CTTZ_ELTS_ZERO_POISON: |
| 3959 | Res = SplitVecOp_CttzElts(N); |
| 3960 | break; |
| 3961 | case ISD::VP_CTTZ_ELTS: |
| 3962 | case ISD::VP_CTTZ_ELTS_ZERO_POISON: |
| 3963 | Res = SplitVecOp_VP_CttzElements(N); |
| 3964 | break; |
| 3965 | case ISD::EXPERIMENTAL_VECTOR_HISTOGRAM: |
| 3966 | Res = SplitVecOp_VECTOR_HISTOGRAM(N); |
| 3967 | break; |
| 3968 | case ISD::PARTIAL_REDUCE_UMLA: |
| 3969 | case ISD::PARTIAL_REDUCE_SMLA: |
| 3970 | case ISD::PARTIAL_REDUCE_SUMLA: |
| 3971 | case ISD::PARTIAL_REDUCE_FMLA: |
| 3972 | Res = SplitVecOp_PARTIAL_REDUCE_MLA(N); |
| 3973 | break; |
| 3974 | } |
| 3975 | |
| 3976 | // If the result is null, the sub-method took care of registering results etc. |
| 3977 | if (!Res.getNode()) return false; |
| 3978 | |
| 3979 | // If the result is N, the sub-method updated N in place. Tell the legalizer |
| 3980 | // core about this. |
| 3981 | if (Res.getNode() == N) |
| 3982 | return true; |
| 3983 | |
| 3984 | if (N->isStrictFPOpcode()) |
| 3985 | assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 2 && |
| 3986 | "Invalid operand expansion" ); |
| 3987 | else |
| 3988 | assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 && |
| 3989 | "Invalid operand expansion" ); |
| 3990 | |
| 3991 | ReplaceValueWith(From: SDValue(N, 0), To: Res); |
| 3992 | return false; |
| 3993 | } |
| 3994 | |
| 3995 | SDValue DAGTypeLegalizer::SplitVecOp_VECTOR_FIND_LAST_ACTIVE(SDNode *N) { |
| 3996 | SDLoc DL(N); |
| 3997 | |
| 3998 | SDValue LoMask, HiMask; |
| 3999 | GetSplitVector(Op: N->getOperand(Num: 0), Lo&: LoMask, Hi&: HiMask); |
| 4000 | |
| 4001 | EVT VT = N->getValueType(ResNo: 0); |
| 4002 | EVT SplitVT = LoMask.getValueType(); |
| 4003 | ElementCount SplitEC = SplitVT.getVectorElementCount(); |
| 4004 | |
| 4005 | // Find the last active in both the low and the high masks. |
| 4006 | SDValue LoFind = DAG.getNode(Opcode: ISD::VECTOR_FIND_LAST_ACTIVE, DL, VT, Operand: LoMask); |
| 4007 | SDValue HiFind = DAG.getNode(Opcode: ISD::VECTOR_FIND_LAST_ACTIVE, DL, VT, Operand: HiMask); |
| 4008 | |
| 4009 | // Check if any lane is active in the high mask. |
| 4010 | // FIXME: This would not be necessary if VECTOR_FIND_LAST_ACTIVE returned a |
| 4011 | // sentinel value for "none active". |
| 4012 | SDValue AnyHiActive = DAG.getNode(Opcode: ISD::VECREDUCE_OR, DL, VT: MVT::i1, Operand: HiMask); |
| 4013 | SDValue Cond = DAG.getBoolExtOrTrunc(Op: AnyHiActive, SL: DL, |
| 4014 | VT: getSetCCResultType(VT: MVT::i1), OpVT: MVT::i1); |
| 4015 | |
| 4016 | // Return: AnyHiActive ? (HiFind + SplitEC) : LoFind; |
| 4017 | return DAG.getNode(Opcode: ISD::SELECT, DL, VT, N1: Cond, |
| 4018 | N2: DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: HiFind, |
| 4019 | N2: DAG.getElementCount(DL, VT, EC: SplitEC)), |
| 4020 | N3: LoFind); |
| 4021 | } |
| 4022 | |
| 4023 | SDValue DAGTypeLegalizer::SplitVecOp_VSELECT(SDNode *N, unsigned OpNo) { |
| 4024 | // The only possibility for an illegal operand is the mask, since result type |
| 4025 | // legalization would have handled this node already otherwise. |
| 4026 | assert(OpNo == 0 && "Illegal operand must be mask" ); |
| 4027 | |
| 4028 | SDValue Mask = N->getOperand(Num: 0); |
| 4029 | SDValue Src0 = N->getOperand(Num: 1); |
| 4030 | SDValue Src1 = N->getOperand(Num: 2); |
| 4031 | EVT Src0VT = Src0.getValueType(); |
| 4032 | SDLoc DL(N); |
| 4033 | assert(Mask.getValueType().isVector() && "VSELECT without a vector mask?" ); |
| 4034 | |
| 4035 | SDValue Lo, Hi; |
| 4036 | GetSplitVector(Op: N->getOperand(Num: 0), Lo, Hi); |
| 4037 | assert(Lo.getValueType() == Hi.getValueType() && |
| 4038 | "Lo and Hi have differing types" ); |
| 4039 | |
| 4040 | EVT LoOpVT, HiOpVT; |
| 4041 | std::tie(args&: LoOpVT, args&: HiOpVT) = DAG.GetSplitDestVTs(VT: Src0VT); |
| 4042 | assert(LoOpVT == HiOpVT && "Asymmetric vector split?" ); |
| 4043 | |
| 4044 | SDValue LoOp0, HiOp0, LoOp1, HiOp1, LoMask, HiMask; |
| 4045 | std::tie(args&: LoOp0, args&: HiOp0) = DAG.SplitVector(N: Src0, DL); |
| 4046 | std::tie(args&: LoOp1, args&: HiOp1) = DAG.SplitVector(N: Src1, DL); |
| 4047 | std::tie(args&: LoMask, args&: HiMask) = DAG.SplitVector(N: Mask, DL); |
| 4048 | |
| 4049 | SDValue LoSelect = |
| 4050 | DAG.getNode(Opcode: ISD::VSELECT, DL, VT: LoOpVT, N1: LoMask, N2: LoOp0, N3: LoOp1); |
| 4051 | SDValue HiSelect = |
| 4052 | DAG.getNode(Opcode: ISD::VSELECT, DL, VT: HiOpVT, N1: HiMask, N2: HiOp0, N3: HiOp1); |
| 4053 | |
| 4054 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: Src0VT, N1: LoSelect, N2: HiSelect); |
| 4055 | } |
| 4056 | |
| 4057 | SDValue DAGTypeLegalizer::SplitVecOp_VECTOR_COMPRESS(SDNode *N, unsigned OpNo) { |
| 4058 | // The only possibility for an illegal operand is the mask, since result type |
| 4059 | // legalization would have handled this node already otherwise. |
| 4060 | assert(OpNo == 1 && "Illegal operand must be mask" ); |
| 4061 | |
| 4062 | // To split the mask, we need to split the result type too, so we can just |
| 4063 | // reuse that logic here. |
| 4064 | SDValue Lo, Hi; |
| 4065 | SplitVecRes_VECTOR_COMPRESS(N, Lo, Hi); |
| 4066 | |
| 4067 | EVT VecVT = N->getValueType(ResNo: 0); |
| 4068 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: SDLoc(N), VT: VecVT, N1: Lo, N2: Hi); |
| 4069 | } |
| 4070 | |
| 4071 | SDValue DAGTypeLegalizer::SplitVecOp_VECREDUCE(SDNode *N, unsigned OpNo) { |
| 4072 | EVT ResVT = N->getValueType(ResNo: 0); |
| 4073 | SDValue Lo, Hi; |
| 4074 | SDLoc dl(N); |
| 4075 | |
| 4076 | SDValue VecOp = N->getOperand(Num: OpNo); |
| 4077 | EVT VecVT = VecOp.getValueType(); |
| 4078 | assert(VecVT.isVector() && "Can only split reduce vector operand" ); |
| 4079 | GetSplitVector(Op: VecOp, Lo, Hi); |
| 4080 | EVT LoOpVT, HiOpVT; |
| 4081 | std::tie(args&: LoOpVT, args&: HiOpVT) = DAG.GetSplitDestVTs(VT: VecVT); |
| 4082 | |
| 4083 | // Use the appropriate scalar instruction on the split subvectors before |
| 4084 | // reducing the now partially reduced smaller vector. |
| 4085 | unsigned CombineOpc = ISD::getVecReduceBaseOpcode(VecReduceOpcode: N->getOpcode()); |
| 4086 | SDValue Partial = DAG.getNode(Opcode: CombineOpc, DL: dl, VT: LoOpVT, N1: Lo, N2: Hi, Flags: N->getFlags()); |
| 4087 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: ResVT, Operand: Partial, Flags: N->getFlags()); |
| 4088 | } |
| 4089 | |
| 4090 | SDValue DAGTypeLegalizer::SplitVecOp_VECREDUCE_SEQ(SDNode *N) { |
| 4091 | EVT ResVT = N->getValueType(ResNo: 0); |
| 4092 | SDValue Lo, Hi; |
| 4093 | SDLoc dl(N); |
| 4094 | |
| 4095 | SDValue AccOp = N->getOperand(Num: 0); |
| 4096 | SDValue VecOp = N->getOperand(Num: 1); |
| 4097 | SDNodeFlags Flags = N->getFlags(); |
| 4098 | |
| 4099 | EVT VecVT = VecOp.getValueType(); |
| 4100 | assert(VecVT.isVector() && "Can only split reduce vector operand" ); |
| 4101 | GetSplitVector(Op: VecOp, Lo, Hi); |
| 4102 | EVT LoOpVT, HiOpVT; |
| 4103 | std::tie(args&: LoOpVT, args&: HiOpVT) = DAG.GetSplitDestVTs(VT: VecVT); |
| 4104 | |
| 4105 | // Reduce low half. |
| 4106 | SDValue Partial = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: ResVT, N1: AccOp, N2: Lo, Flags); |
| 4107 | |
| 4108 | // Reduce high half, using low half result as initial value. |
| 4109 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: ResVT, N1: Partial, N2: Hi, Flags); |
| 4110 | } |
| 4111 | |
| 4112 | SDValue DAGTypeLegalizer::SplitVecOp_VP_REDUCE(SDNode *N, unsigned OpNo) { |
| 4113 | assert(N->isVPOpcode() && "Expected VP opcode" ); |
| 4114 | assert(OpNo == 1 && "Can only split reduce vector operand" ); |
| 4115 | |
| 4116 | unsigned Opc = N->getOpcode(); |
| 4117 | EVT ResVT = N->getValueType(ResNo: 0); |
| 4118 | SDValue Lo, Hi; |
| 4119 | SDLoc dl(N); |
| 4120 | |
| 4121 | SDValue VecOp = N->getOperand(Num: OpNo); |
| 4122 | EVT VecVT = VecOp.getValueType(); |
| 4123 | assert(VecVT.isVector() && "Can only split reduce vector operand" ); |
| 4124 | GetSplitVector(Op: VecOp, Lo, Hi); |
| 4125 | |
| 4126 | SDValue MaskLo, MaskHi; |
| 4127 | std::tie(args&: MaskLo, args&: MaskHi) = SplitMask(Mask: N->getOperand(Num: 2)); |
| 4128 | |
| 4129 | SDValue EVLLo, EVLHi; |
| 4130 | std::tie(args&: EVLLo, args&: EVLHi) = DAG.SplitEVL(N: N->getOperand(Num: 3), VecVT, DL: dl); |
| 4131 | |
| 4132 | const SDNodeFlags Flags = N->getFlags(); |
| 4133 | |
| 4134 | SDValue ResLo = |
| 4135 | DAG.getNode(Opcode: Opc, DL: dl, VT: ResVT, Ops: {N->getOperand(Num: 0), Lo, MaskLo, EVLLo}, Flags); |
| 4136 | return DAG.getNode(Opcode: Opc, DL: dl, VT: ResVT, Ops: {ResLo, Hi, MaskHi, EVLHi}, Flags); |
| 4137 | } |
| 4138 | |
| 4139 | SDValue DAGTypeLegalizer::SplitVecOp_UnaryOp(SDNode *N) { |
| 4140 | // The result has a legal vector type, but the input needs splitting. |
| 4141 | EVT ResVT = N->getValueType(ResNo: 0); |
| 4142 | SDValue Lo, Hi; |
| 4143 | SDLoc dl(N); |
| 4144 | GetSplitVector(Op: N->getOperand(Num: N->isStrictFPOpcode() ? 1 : 0), Lo, Hi); |
| 4145 | EVT InVT = Lo.getValueType(); |
| 4146 | |
| 4147 | EVT OutVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: ResVT.getVectorElementType(), |
| 4148 | EC: InVT.getVectorElementCount()); |
| 4149 | |
| 4150 | if (N->isStrictFPOpcode()) { |
| 4151 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL: dl, ResultTys: {OutVT, MVT::Other}, |
| 4152 | Ops: {N->getOperand(Num: 0), Lo}); |
| 4153 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL: dl, ResultTys: {OutVT, MVT::Other}, |
| 4154 | Ops: {N->getOperand(Num: 0), Hi}); |
| 4155 | |
| 4156 | // Build a factor node to remember that this operation is independent |
| 4157 | // of the other one. |
| 4158 | SDValue Ch = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Lo.getValue(R: 1), |
| 4159 | N2: Hi.getValue(R: 1)); |
| 4160 | |
| 4161 | // Legalize the chain result - switch anything that used the old chain to |
| 4162 | // use the new one. |
| 4163 | ReplaceValueWith(From: SDValue(N, 1), To: Ch); |
| 4164 | } else if (N->getNumOperands() == 3) { |
| 4165 | assert(N->isVPOpcode() && "Expected VP opcode" ); |
| 4166 | SDValue MaskLo, MaskHi, EVLLo, EVLHi; |
| 4167 | std::tie(args&: MaskLo, args&: MaskHi) = SplitMask(Mask: N->getOperand(Num: 1)); |
| 4168 | std::tie(args&: EVLLo, args&: EVLHi) = |
| 4169 | DAG.SplitEVL(N: N->getOperand(Num: 2), VecVT: N->getValueType(ResNo: 0), DL: dl); |
| 4170 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: OutVT, N1: Lo, N2: MaskLo, N3: EVLLo); |
| 4171 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: OutVT, N1: Hi, N2: MaskHi, N3: EVLHi); |
| 4172 | } else { |
| 4173 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: OutVT, Operand: Lo); |
| 4174 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: OutVT, Operand: Hi); |
| 4175 | } |
| 4176 | |
| 4177 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: ResVT, N1: Lo, N2: Hi); |
| 4178 | } |
| 4179 | |
| 4180 | // Split a FAKE_USE use of a vector into FAKE_USEs of hi and lo part. |
| 4181 | SDValue DAGTypeLegalizer::SplitVecOp_FAKE_USE(SDNode *N) { |
| 4182 | SDValue Lo, Hi; |
| 4183 | GetSplitVector(Op: N->getOperand(Num: 1), Lo, Hi); |
| 4184 | SDValue Chain = |
| 4185 | DAG.getNode(Opcode: ISD::FAKE_USE, DL: SDLoc(), VT: MVT::Other, N1: N->getOperand(Num: 0), N2: Lo); |
| 4186 | return DAG.getNode(Opcode: ISD::FAKE_USE, DL: SDLoc(), VT: MVT::Other, N1: Chain, N2: Hi); |
| 4187 | } |
| 4188 | |
| 4189 | SDValue DAGTypeLegalizer::SplitVecOp_BITCAST(SDNode *N) { |
| 4190 | // For example, i64 = BITCAST v4i16 on alpha. Typically the vector will |
| 4191 | // end up being split all the way down to individual components. Convert the |
| 4192 | // split pieces into integers and reassemble. |
| 4193 | EVT ResVT = N->getValueType(ResNo: 0); |
| 4194 | SDValue Lo, Hi; |
| 4195 | GetSplitVector(Op: N->getOperand(Num: 0), Lo, Hi); |
| 4196 | SDLoc dl(N); |
| 4197 | |
| 4198 | if (ResVT.isScalableVector()) { |
| 4199 | auto [LoVT, HiVT] = DAG.GetSplitDestVTs(VT: ResVT); |
| 4200 | Lo = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: LoVT, Operand: Lo); |
| 4201 | Hi = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: HiVT, Operand: Hi); |
| 4202 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: ResVT, N1: Lo, N2: Hi); |
| 4203 | } |
| 4204 | |
| 4205 | Lo = BitConvertToInteger(Op: Lo); |
| 4206 | Hi = BitConvertToInteger(Op: Hi); |
| 4207 | |
| 4208 | if (DAG.getDataLayout().isBigEndian()) |
| 4209 | std::swap(a&: Lo, b&: Hi); |
| 4210 | |
| 4211 | return DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: ResVT, Operand: JoinIntegers(Lo, Hi)); |
| 4212 | } |
| 4213 | |
| 4214 | SDValue DAGTypeLegalizer::SplitVecOp_INSERT_SUBVECTOR(SDNode *N, |
| 4215 | unsigned OpNo) { |
| 4216 | assert(OpNo == 1 && "Invalid OpNo; can only split SubVec." ); |
| 4217 | // We know that the result type is legal. |
| 4218 | EVT ResVT = N->getValueType(ResNo: 0); |
| 4219 | |
| 4220 | SDValue Vec = N->getOperand(Num: 0); |
| 4221 | SDValue SubVec = N->getOperand(Num: 1); |
| 4222 | SDValue Idx = N->getOperand(Num: 2); |
| 4223 | SDLoc dl(N); |
| 4224 | |
| 4225 | SDValue Lo, Hi; |
| 4226 | GetSplitVector(Op: SubVec, Lo, Hi); |
| 4227 | |
| 4228 | uint64_t IdxVal = Idx->getAsZExtVal(); |
| 4229 | uint64_t LoElts = Lo.getValueType().getVectorMinNumElements(); |
| 4230 | |
| 4231 | SDValue FirstInsertion = |
| 4232 | DAG.getNode(Opcode: ISD::INSERT_SUBVECTOR, DL: dl, VT: ResVT, N1: Vec, N2: Lo, N3: Idx); |
| 4233 | SDValue SecondInsertion = |
| 4234 | DAG.getNode(Opcode: ISD::INSERT_SUBVECTOR, DL: dl, VT: ResVT, N1: FirstInsertion, N2: Hi, |
| 4235 | N3: DAG.getVectorIdxConstant(Val: IdxVal + LoElts, DL: dl)); |
| 4236 | |
| 4237 | return SecondInsertion; |
| 4238 | } |
| 4239 | |
| 4240 | SDValue DAGTypeLegalizer::(SDNode *N) { |
| 4241 | // We know that the extracted result type is legal. |
| 4242 | EVT SubVT = N->getValueType(ResNo: 0); |
| 4243 | SDValue Idx = N->getOperand(Num: 1); |
| 4244 | SDLoc dl(N); |
| 4245 | SDValue Lo, Hi; |
| 4246 | |
| 4247 | GetSplitVector(Op: N->getOperand(Num: 0), Lo, Hi); |
| 4248 | |
| 4249 | ElementCount LoElts = Lo.getValueType().getVectorElementCount(); |
| 4250 | // Note: For scalable vectors, the index is scaled by vscale. |
| 4251 | ElementCount IdxVal = |
| 4252 | ElementCount::get(MinVal: Idx->getAsZExtVal(), Scalable: SubVT.isScalableVector()); |
| 4253 | uint64_t IdxValMin = IdxVal.getKnownMinValue(); |
| 4254 | |
| 4255 | EVT SrcVT = N->getOperand(Num: 0).getValueType(); |
| 4256 | ElementCount NumResultElts = SubVT.getVectorElementCount(); |
| 4257 | |
| 4258 | // If the extracted elements are all in the low half, do a simple extract. |
| 4259 | if (ElementCount::isKnownLE(LHS: IdxVal + NumResultElts, RHS: LoElts)) |
| 4260 | return DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL: dl, VT: SubVT, N1: Lo, N2: Idx); |
| 4261 | |
| 4262 | unsigned LoEltsMin = LoElts.getKnownMinValue(); |
| 4263 | if (IdxValMin < LoEltsMin && SubVT.isFixedLengthVector() && |
| 4264 | SrcVT.isFixedLengthVector()) { |
| 4265 | // Extracted subvector crosses vector split, so we need to blend the two |
| 4266 | // halves. |
| 4267 | // TODO: May be able to emit partial extract_subvector. |
| 4268 | SmallVector<SDValue, 8> Elts; |
| 4269 | Elts.reserve(N: NumResultElts.getFixedValue()); |
| 4270 | |
| 4271 | // This is not valid for scalable vectors. If SubVT is scalable, this is the |
| 4272 | // same as unrolling a scalable dimension (invalid). If ScrVT is scalable, |
| 4273 | // `Lo[LoEltsMin]` may not be the last element of `Lo`. |
| 4274 | DAG.ExtractVectorElements(Op: Lo, Args&: Elts, /*Start=*/IdxValMin, |
| 4275 | /*Count=*/LoEltsMin - IdxValMin); |
| 4276 | DAG.ExtractVectorElements(Op: Hi, Args&: Elts, /*Start=*/0, |
| 4277 | /*Count=*/SubVT.getVectorNumElements() - |
| 4278 | Elts.size()); |
| 4279 | return DAG.getBuildVector(VT: SubVT, DL: dl, Ops: Elts); |
| 4280 | } |
| 4281 | |
| 4282 | if (SubVT.isScalableVector() == SrcVT.isScalableVector()) { |
| 4283 | ElementCount = IdxVal - LoElts; |
| 4284 | if (ExtractIdx.isKnownMultipleOf(RHS: NumResultElts)) |
| 4285 | return DAG.getExtractSubvector(DL: dl, VT: SubVT, Vec: Hi, |
| 4286 | Idx: ExtractIdx.getKnownMinValue()); |
| 4287 | |
| 4288 | EVT HiVT = Hi.getValueType(); |
| 4289 | assert(HiVT.isFixedLengthVector() && |
| 4290 | "Only fixed-vector extracts are supported in this case" ); |
| 4291 | |
| 4292 | // We cannot create an extract_subvector that isn't a multiple of the |
| 4293 | // result size, which may go out of bounds for the last elements. Shuffle |
| 4294 | // the desired elements down to 0 and do a simple 0 extract. |
| 4295 | SmallVector<int, 8> Mask(HiVT.getVectorNumElements(), -1); |
| 4296 | for (int I = 0; I != int(NumResultElts.getFixedValue()); ++I) |
| 4297 | Mask[I] = int(ExtractIdx.getFixedValue()) + I; |
| 4298 | |
| 4299 | SDValue Shuffle = |
| 4300 | DAG.getVectorShuffle(VT: HiVT, dl, N1: Hi, N2: DAG.getPOISON(VT: HiVT), Mask); |
| 4301 | return DAG.getExtractSubvector(DL: dl, VT: SubVT, Vec: Shuffle, Idx: 0); |
| 4302 | } |
| 4303 | |
| 4304 | // After this point the DAG node only permits extracting fixed-width |
| 4305 | // subvectors from scalable vectors. |
| 4306 | assert(SubVT.isFixedLengthVector() && |
| 4307 | "Extracting scalable subvector from fixed-width unsupported" ); |
| 4308 | |
| 4309 | // If the element type is i1 and we're not promoting the result, then we may |
| 4310 | // end up loading the wrong data since the bits are packed tightly into |
| 4311 | // bytes. For example, if we extract a v4i1 (legal) from a nxv4i1 (legal) |
| 4312 | // type at index 4, then we will load a byte starting at index 0. |
| 4313 | if (SubVT.getScalarType() == MVT::i1) |
| 4314 | report_fatal_error(reason: "Don't know how to extract fixed-width predicate " |
| 4315 | "subvector from a scalable predicate vector" ); |
| 4316 | |
| 4317 | // Spill the vector to the stack. We should use the alignment for |
| 4318 | // the smallest part. |
| 4319 | SDValue Vec = N->getOperand(Num: 0); |
| 4320 | EVT VecVT = Vec.getValueType(); |
| 4321 | Align SmallestAlign = DAG.getReducedAlign(VT: VecVT, /*UseABI=*/false); |
| 4322 | SDValue StackPtr = |
| 4323 | DAG.CreateStackTemporary(Bytes: VecVT.getStoreSize(), Alignment: SmallestAlign); |
| 4324 | auto &MF = DAG.getMachineFunction(); |
| 4325 | auto FrameIndex = cast<FrameIndexSDNode>(Val: StackPtr.getNode())->getIndex(); |
| 4326 | auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: FrameIndex); |
| 4327 | |
| 4328 | SDValue Store = DAG.getStore(Chain: DAG.getEntryNode(), dl, Val: Vec, Ptr: StackPtr, PtrInfo, |
| 4329 | Alignment: SmallestAlign); |
| 4330 | |
| 4331 | // Extract the subvector by loading the correct part. |
| 4332 | StackPtr = TLI.getVectorSubVecPointer(DAG, VecPtr: StackPtr, VecVT, SubVecVT: SubVT, Index: Idx); |
| 4333 | |
| 4334 | return DAG.getLoad( |
| 4335 | VT: SubVT, dl, Chain: Store, Ptr: StackPtr, |
| 4336 | PtrInfo: MachinePointerInfo::getUnknownStack(MF&: DAG.getMachineFunction())); |
| 4337 | } |
| 4338 | |
| 4339 | SDValue DAGTypeLegalizer::(SDNode *N) { |
| 4340 | SDValue Vec = N->getOperand(Num: 0); |
| 4341 | SDValue Idx = N->getOperand(Num: 1); |
| 4342 | EVT VecVT = Vec.getValueType(); |
| 4343 | |
| 4344 | if (const ConstantSDNode *Index = dyn_cast<ConstantSDNode>(Val&: Idx)) { |
| 4345 | uint64_t IdxVal = Index->getZExtValue(); |
| 4346 | |
| 4347 | SDValue Lo, Hi; |
| 4348 | GetSplitVector(Op: Vec, Lo, Hi); |
| 4349 | |
| 4350 | uint64_t LoElts = Lo.getValueType().getVectorMinNumElements(); |
| 4351 | |
| 4352 | if (IdxVal < LoElts) |
| 4353 | return SDValue(DAG.UpdateNodeOperands(N, Op1: Lo, Op2: Idx), 0); |
| 4354 | else if (!Vec.getValueType().isScalableVector()) |
| 4355 | return SDValue(DAG.UpdateNodeOperands(N, Op1: Hi, |
| 4356 | Op2: DAG.getConstant(Val: IdxVal - LoElts, DL: SDLoc(N), |
| 4357 | VT: Idx.getValueType())), 0); |
| 4358 | } |
| 4359 | |
| 4360 | // See if the target wants to custom expand this node. |
| 4361 | if (CustomLowerNode(N, VT: N->getValueType(ResNo: 0), LegalizeResult: true)) |
| 4362 | return SDValue(); |
| 4363 | |
| 4364 | // Make the vector elements byte-addressable if they aren't already. |
| 4365 | SDLoc dl(N); |
| 4366 | EVT EltVT = VecVT.getVectorElementType(); |
| 4367 | if (!EltVT.isByteSized()) { |
| 4368 | EltVT = EltVT.changeTypeToInteger().getRoundIntegerType(Context&: *DAG.getContext()); |
| 4369 | VecVT = VecVT.changeElementType(Context&: *DAG.getContext(), EltVT); |
| 4370 | Vec = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT: VecVT, Operand: Vec); |
| 4371 | SDValue = |
| 4372 | DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: dl, VT: EltVT, N1: Vec, N2: Idx); |
| 4373 | return DAG.getAnyExtOrTrunc(Op: NewExtract, DL: dl, VT: N->getValueType(ResNo: 0)); |
| 4374 | } |
| 4375 | |
| 4376 | // Store the vector to the stack. |
| 4377 | // In cases where the vector is illegal it will be broken down into parts |
| 4378 | // and stored in parts - we should use the alignment for the smallest part. |
| 4379 | Align SmallestAlign = DAG.getReducedAlign(VT: VecVT, /*UseABI=*/false); |
| 4380 | SDValue StackPtr = |
| 4381 | DAG.CreateStackTemporary(Bytes: VecVT.getStoreSize(), Alignment: SmallestAlign); |
| 4382 | auto &MF = DAG.getMachineFunction(); |
| 4383 | auto FrameIndex = cast<FrameIndexSDNode>(Val: StackPtr.getNode())->getIndex(); |
| 4384 | auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: FrameIndex); |
| 4385 | SDValue Store = DAG.getStore(Chain: DAG.getEntryNode(), dl, Val: Vec, Ptr: StackPtr, PtrInfo, |
| 4386 | Alignment: SmallestAlign); |
| 4387 | |
| 4388 | // Load back the required element. |
| 4389 | StackPtr = TLI.getVectorElementPointer(DAG, VecPtr: StackPtr, VecVT, Index: Idx); |
| 4390 | |
| 4391 | // EXTRACT_VECTOR_ELT can extend the element type to the width of the return |
| 4392 | // type, leaving the high bits undefined. But it can't truncate. |
| 4393 | assert(N->getValueType(0).bitsGE(EltVT) && "Illegal EXTRACT_VECTOR_ELT." ); |
| 4394 | |
| 4395 | return DAG.getExtLoad( |
| 4396 | ExtType: ISD::EXTLOAD, dl, VT: N->getValueType(ResNo: 0), Chain: Store, Ptr: StackPtr, |
| 4397 | PtrInfo: MachinePointerInfo::getUnknownStack(MF&: DAG.getMachineFunction()), MemVT: EltVT, |
| 4398 | Alignment: commonAlignment(A: SmallestAlign, Offset: EltVT.getFixedSizeInBits() / 8)); |
| 4399 | } |
| 4400 | |
| 4401 | SDValue DAGTypeLegalizer::SplitVecOp_ExtVecInRegOp(SDNode *N) { |
| 4402 | SDValue Lo, Hi; |
| 4403 | |
| 4404 | // *_EXTEND_VECTOR_INREG only reference the lower half of the input, so |
| 4405 | // splitting the result has the same effect as splitting the input operand. |
| 4406 | SplitVecRes_ExtVecInRegOp(N, Lo, Hi); |
| 4407 | |
| 4408 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: SDLoc(N), VT: N->getValueType(ResNo: 0), N1: Lo, N2: Hi); |
| 4409 | } |
| 4410 | |
| 4411 | SDValue DAGTypeLegalizer::SplitVecOp_Gather(MemSDNode *N, unsigned OpNo) { |
| 4412 | (void)OpNo; |
| 4413 | SDValue Lo, Hi; |
| 4414 | SplitVecRes_Gather(N, Lo, Hi); |
| 4415 | |
| 4416 | SDValue Res = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: N, VT: N->getValueType(ResNo: 0), N1: Lo, N2: Hi); |
| 4417 | ReplaceValueWith(From: SDValue(N, 0), To: Res); |
| 4418 | return SDValue(); |
| 4419 | } |
| 4420 | |
| 4421 | SDValue DAGTypeLegalizer::SplitVecOp_VP_STORE(VPStoreSDNode *N, unsigned OpNo) { |
| 4422 | assert(N->isUnindexed() && "Indexed vp_store of vector?" ); |
| 4423 | SDValue Ch = N->getChain(); |
| 4424 | SDValue Ptr = N->getBasePtr(); |
| 4425 | SDValue Offset = N->getOffset(); |
| 4426 | assert(Offset.isUndef() && "Unexpected VP store offset" ); |
| 4427 | SDValue Mask = N->getMask(); |
| 4428 | SDValue EVL = N->getVectorLength(); |
| 4429 | SDValue Data = N->getValue(); |
| 4430 | Align Alignment = N->getBaseAlign(); |
| 4431 | SDLoc DL(N); |
| 4432 | |
| 4433 | SDValue DataLo, DataHi; |
| 4434 | if (getTypeAction(VT: Data.getValueType()) == TargetLowering::TypeSplitVector) |
| 4435 | // Split Data operand |
| 4436 | GetSplitVector(Op: Data, Lo&: DataLo, Hi&: DataHi); |
| 4437 | else |
| 4438 | std::tie(args&: DataLo, args&: DataHi) = DAG.SplitVector(N: Data, DL); |
| 4439 | |
| 4440 | // Split Mask operand |
| 4441 | SDValue MaskLo, MaskHi; |
| 4442 | if (OpNo == 1 && Mask.getOpcode() == ISD::SETCC) { |
| 4443 | SplitVecRes_SETCC(N: Mask.getNode(), Lo&: MaskLo, Hi&: MaskHi); |
| 4444 | } else { |
| 4445 | if (getTypeAction(VT: Mask.getValueType()) == TargetLowering::TypeSplitVector) |
| 4446 | GetSplitVector(Op: Mask, Lo&: MaskLo, Hi&: MaskHi); |
| 4447 | else |
| 4448 | std::tie(args&: MaskLo, args&: MaskHi) = DAG.SplitVector(N: Mask, DL); |
| 4449 | } |
| 4450 | |
| 4451 | EVT MemoryVT = N->getMemoryVT(); |
| 4452 | EVT LoMemVT, HiMemVT; |
| 4453 | bool HiIsEmpty = false; |
| 4454 | std::tie(args&: LoMemVT, args&: HiMemVT) = |
| 4455 | DAG.GetDependentSplitDestVTs(VT: MemoryVT, EnvVT: DataLo.getValueType(), HiIsEmpty: &HiIsEmpty); |
| 4456 | |
| 4457 | // Split EVL |
| 4458 | SDValue EVLLo, EVLHi; |
| 4459 | std::tie(args&: EVLLo, args&: EVLHi) = DAG.SplitEVL(N: EVL, VecVT: Data.getValueType(), DL); |
| 4460 | |
| 4461 | SDValue Lo, Hi; |
| 4462 | MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand( |
| 4463 | PtrInfo: N->getPointerInfo(), F: MachineMemOperand::MOStore, |
| 4464 | Size: LocationSize::beforeOrAfterPointer(), BaseAlignment: Alignment, AAInfo: N->getAAInfo(), |
| 4465 | Ranges: N->getRanges()); |
| 4466 | |
| 4467 | Lo = DAG.getStoreVP(Chain: Ch, dl: DL, Val: DataLo, Ptr, Offset, Mask: MaskLo, EVL: EVLLo, MemVT: LoMemVT, MMO, |
| 4468 | AM: N->getAddressingMode(), IsTruncating: N->isTruncatingStore(), |
| 4469 | IsCompressing: N->isCompressingStore()); |
| 4470 | |
| 4471 | // If the hi vp_store has zero storage size, only the lo vp_store is needed. |
| 4472 | if (HiIsEmpty) |
| 4473 | return Lo; |
| 4474 | |
| 4475 | Ptr = TLI.IncrementMemoryAddress(Addr: Ptr, Mask: MaskLo, DL, DataVT: LoMemVT, DAG, |
| 4476 | IsCompressedMemory: N->isCompressingStore()); |
| 4477 | |
| 4478 | MachinePointerInfo MPI; |
| 4479 | if (LoMemVT.isScalableVector()) { |
| 4480 | Alignment = commonAlignment(A: Alignment, |
| 4481 | Offset: LoMemVT.getSizeInBits().getKnownMinValue() / 8); |
| 4482 | MPI = MachinePointerInfo(N->getPointerInfo().getAddrSpace()); |
| 4483 | } else |
| 4484 | MPI = N->getPointerInfo().getWithOffset( |
| 4485 | O: LoMemVT.getStoreSize().getFixedValue()); |
| 4486 | |
| 4487 | MMO = DAG.getMachineFunction().getMachineMemOperand( |
| 4488 | PtrInfo: MPI, F: MachineMemOperand::MOStore, Size: LocationSize::beforeOrAfterPointer(), |
| 4489 | BaseAlignment: Alignment, AAInfo: N->getAAInfo(), Ranges: N->getRanges()); |
| 4490 | |
| 4491 | Hi = DAG.getStoreVP(Chain: Ch, dl: DL, Val: DataHi, Ptr, Offset, Mask: MaskHi, EVL: EVLHi, MemVT: HiMemVT, MMO, |
| 4492 | AM: N->getAddressingMode(), IsTruncating: N->isTruncatingStore(), |
| 4493 | IsCompressing: N->isCompressingStore()); |
| 4494 | |
| 4495 | // Build a factor node to remember that this store is independent of the |
| 4496 | // other one. |
| 4497 | return DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, N1: Lo, N2: Hi); |
| 4498 | } |
| 4499 | |
| 4500 | SDValue DAGTypeLegalizer::SplitVecOp_VP_STRIDED_STORE(VPStridedStoreSDNode *N, |
| 4501 | unsigned OpNo) { |
| 4502 | assert(N->isUnindexed() && "Indexed vp_strided_store of a vector?" ); |
| 4503 | assert(N->getOffset().isUndef() && "Unexpected VP strided store offset" ); |
| 4504 | |
| 4505 | SDLoc DL(N); |
| 4506 | |
| 4507 | SDValue Data = N->getValue(); |
| 4508 | SDValue LoData, HiData; |
| 4509 | if (getTypeAction(VT: Data.getValueType()) == TargetLowering::TypeSplitVector) |
| 4510 | GetSplitVector(Op: Data, Lo&: LoData, Hi&: HiData); |
| 4511 | else |
| 4512 | std::tie(args&: LoData, args&: HiData) = DAG.SplitVector(N: Data, DL); |
| 4513 | |
| 4514 | EVT LoMemVT, HiMemVT; |
| 4515 | bool HiIsEmpty = false; |
| 4516 | std::tie(args&: LoMemVT, args&: HiMemVT) = DAG.GetDependentSplitDestVTs( |
| 4517 | VT: N->getMemoryVT(), EnvVT: LoData.getValueType(), HiIsEmpty: &HiIsEmpty); |
| 4518 | |
| 4519 | SDValue Mask = N->getMask(); |
| 4520 | SDValue LoMask, HiMask; |
| 4521 | if (OpNo == 1 && Mask.getOpcode() == ISD::SETCC) |
| 4522 | SplitVecRes_SETCC(N: Mask.getNode(), Lo&: LoMask, Hi&: HiMask); |
| 4523 | else if (getTypeAction(VT: Mask.getValueType()) == |
| 4524 | TargetLowering::TypeSplitVector) |
| 4525 | GetSplitVector(Op: Mask, Lo&: LoMask, Hi&: HiMask); |
| 4526 | else |
| 4527 | std::tie(args&: LoMask, args&: HiMask) = DAG.SplitVector(N: Mask, DL); |
| 4528 | |
| 4529 | SDValue LoEVL, HiEVL; |
| 4530 | std::tie(args&: LoEVL, args&: HiEVL) = |
| 4531 | DAG.SplitEVL(N: N->getVectorLength(), VecVT: Data.getValueType(), DL); |
| 4532 | |
| 4533 | // Generate the low vp_strided_store |
| 4534 | SDValue Lo = DAG.getStridedStoreVP( |
| 4535 | Chain: N->getChain(), DL, Val: LoData, Ptr: N->getBasePtr(), Offset: N->getOffset(), |
| 4536 | Stride: N->getStride(), Mask: LoMask, EVL: LoEVL, MemVT: LoMemVT, MMO: N->getMemOperand(), |
| 4537 | AM: N->getAddressingMode(), IsTruncating: N->isTruncatingStore(), IsCompressing: N->isCompressingStore()); |
| 4538 | |
| 4539 | // If the high vp_strided_store has zero storage size, only the low |
| 4540 | // vp_strided_store is needed. |
| 4541 | if (HiIsEmpty) |
| 4542 | return Lo; |
| 4543 | |
| 4544 | // Generate the high vp_strided_store. |
| 4545 | // To calculate the high base address, we need to sum to the low base |
| 4546 | // address stride number of bytes for each element already stored by low, |
| 4547 | // that is: Ptr = Ptr + (LoEVL * Stride) |
| 4548 | EVT PtrVT = N->getBasePtr().getValueType(); |
| 4549 | SDValue Increment = |
| 4550 | DAG.getNode(Opcode: ISD::MUL, DL, VT: PtrVT, N1: LoEVL, |
| 4551 | N2: DAG.getSExtOrTrunc(Op: N->getStride(), DL, VT: PtrVT)); |
| 4552 | SDValue Ptr = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: N->getBasePtr(), N2: Increment); |
| 4553 | |
| 4554 | Align Alignment = N->getBaseAlign(); |
| 4555 | if (LoMemVT.isScalableVector()) |
| 4556 | Alignment = commonAlignment(A: Alignment, |
| 4557 | Offset: LoMemVT.getSizeInBits().getKnownMinValue() / 8); |
| 4558 | |
| 4559 | MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand( |
| 4560 | PtrInfo: MachinePointerInfo(N->getPointerInfo().getAddrSpace()), |
| 4561 | F: MachineMemOperand::MOStore, Size: LocationSize::beforeOrAfterPointer(), |
| 4562 | BaseAlignment: Alignment, AAInfo: N->getAAInfo(), Ranges: N->getRanges()); |
| 4563 | |
| 4564 | SDValue Hi = DAG.getStridedStoreVP( |
| 4565 | Chain: N->getChain(), DL, Val: HiData, Ptr, Offset: N->getOffset(), Stride: N->getStride(), Mask: HiMask, |
| 4566 | EVL: HiEVL, MemVT: HiMemVT, MMO, AM: N->getAddressingMode(), IsTruncating: N->isTruncatingStore(), |
| 4567 | IsCompressing: N->isCompressingStore()); |
| 4568 | |
| 4569 | // Build a factor node to remember that this store is independent of the |
| 4570 | // other one. |
| 4571 | return DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, N1: Lo, N2: Hi); |
| 4572 | } |
| 4573 | |
| 4574 | SDValue DAGTypeLegalizer::SplitVecOp_MSTORE(MaskedStoreSDNode *N, |
| 4575 | unsigned OpNo) { |
| 4576 | assert(N->isUnindexed() && "Indexed masked store of vector?" ); |
| 4577 | SDValue Ch = N->getChain(); |
| 4578 | SDValue Ptr = N->getBasePtr(); |
| 4579 | SDValue Offset = N->getOffset(); |
| 4580 | assert(Offset.isUndef() && "Unexpected indexed masked store offset" ); |
| 4581 | SDValue Mask = N->getMask(); |
| 4582 | SDValue Data = N->getValue(); |
| 4583 | Align Alignment = N->getBaseAlign(); |
| 4584 | SDLoc DL(N); |
| 4585 | |
| 4586 | SDValue DataLo, DataHi; |
| 4587 | if (getTypeAction(VT: Data.getValueType()) == TargetLowering::TypeSplitVector) |
| 4588 | // Split Data operand |
| 4589 | GetSplitVector(Op: Data, Lo&: DataLo, Hi&: DataHi); |
| 4590 | else |
| 4591 | std::tie(args&: DataLo, args&: DataHi) = DAG.SplitVector(N: Data, DL); |
| 4592 | |
| 4593 | // Split Mask operand |
| 4594 | SDValue MaskLo, MaskHi; |
| 4595 | if (OpNo == 1 && Mask.getOpcode() == ISD::SETCC) { |
| 4596 | SplitVecRes_SETCC(N: Mask.getNode(), Lo&: MaskLo, Hi&: MaskHi); |
| 4597 | } else { |
| 4598 | if (getTypeAction(VT: Mask.getValueType()) == TargetLowering::TypeSplitVector) |
| 4599 | GetSplitVector(Op: Mask, Lo&: MaskLo, Hi&: MaskHi); |
| 4600 | else |
| 4601 | std::tie(args&: MaskLo, args&: MaskHi) = DAG.SplitVector(N: Mask, DL); |
| 4602 | } |
| 4603 | |
| 4604 | EVT MemoryVT = N->getMemoryVT(); |
| 4605 | EVT LoMemVT, HiMemVT; |
| 4606 | bool HiIsEmpty = false; |
| 4607 | std::tie(args&: LoMemVT, args&: HiMemVT) = |
| 4608 | DAG.GetDependentSplitDestVTs(VT: MemoryVT, EnvVT: DataLo.getValueType(), HiIsEmpty: &HiIsEmpty); |
| 4609 | |
| 4610 | SDValue Lo, Hi, Res; |
| 4611 | MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand( |
| 4612 | PtrInfo: N->getPointerInfo(), F: MachineMemOperand::MOStore, |
| 4613 | Size: LocationSize::beforeOrAfterPointer(), BaseAlignment: Alignment, AAInfo: N->getAAInfo(), |
| 4614 | Ranges: N->getRanges()); |
| 4615 | |
| 4616 | Lo = DAG.getMaskedStore(Chain: Ch, dl: DL, Val: DataLo, Base: Ptr, Offset, Mask: MaskLo, MemVT: LoMemVT, MMO, |
| 4617 | AM: N->getAddressingMode(), IsTruncating: N->isTruncatingStore(), |
| 4618 | IsCompressing: N->isCompressingStore()); |
| 4619 | |
| 4620 | if (HiIsEmpty) { |
| 4621 | // The hi masked store has zero storage size. |
| 4622 | // Only the lo masked store is needed. |
| 4623 | Res = Lo; |
| 4624 | } else { |
| 4625 | |
| 4626 | Ptr = TLI.IncrementMemoryAddress(Addr: Ptr, Mask: MaskLo, DL, DataVT: LoMemVT, DAG, |
| 4627 | IsCompressedMemory: N->isCompressingStore()); |
| 4628 | |
| 4629 | MachinePointerInfo MPI; |
| 4630 | if (LoMemVT.isScalableVector()) { |
| 4631 | Alignment = commonAlignment( |
| 4632 | A: Alignment, Offset: LoMemVT.getSizeInBits().getKnownMinValue() / 8); |
| 4633 | MPI = MachinePointerInfo(N->getPointerInfo().getAddrSpace()); |
| 4634 | } else |
| 4635 | MPI = N->getPointerInfo().getWithOffset( |
| 4636 | O: LoMemVT.getStoreSize().getFixedValue()); |
| 4637 | |
| 4638 | MMO = DAG.getMachineFunction().getMachineMemOperand( |
| 4639 | PtrInfo: MPI, F: MachineMemOperand::MOStore, Size: LocationSize::beforeOrAfterPointer(), |
| 4640 | BaseAlignment: Alignment, AAInfo: N->getAAInfo(), Ranges: N->getRanges()); |
| 4641 | |
| 4642 | Hi = DAG.getMaskedStore(Chain: Ch, dl: DL, Val: DataHi, Base: Ptr, Offset, Mask: MaskHi, MemVT: HiMemVT, MMO, |
| 4643 | AM: N->getAddressingMode(), IsTruncating: N->isTruncatingStore(), |
| 4644 | IsCompressing: N->isCompressingStore()); |
| 4645 | |
| 4646 | // Build a factor node to remember that this store is independent of the |
| 4647 | // other one. |
| 4648 | Res = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, N1: Lo, N2: Hi); |
| 4649 | } |
| 4650 | |
| 4651 | return Res; |
| 4652 | } |
| 4653 | |
| 4654 | SDValue DAGTypeLegalizer::SplitVecOp_Scatter(MemSDNode *N, unsigned OpNo) { |
| 4655 | SDValue Ch = N->getChain(); |
| 4656 | SDValue Ptr = N->getBasePtr(); |
| 4657 | EVT MemoryVT = N->getMemoryVT(); |
| 4658 | Align Alignment = N->getBaseAlign(); |
| 4659 | SDLoc DL(N); |
| 4660 | struct Operands { |
| 4661 | SDValue Mask; |
| 4662 | SDValue Index; |
| 4663 | SDValue Scale; |
| 4664 | SDValue Data; |
| 4665 | } Ops = [&]() -> Operands { |
| 4666 | if (auto *MSC = dyn_cast<MaskedScatterSDNode>(Val: N)) { |
| 4667 | return {.Mask: MSC->getMask(), .Index: MSC->getIndex(), .Scale: MSC->getScale(), |
| 4668 | .Data: MSC->getValue()}; |
| 4669 | } |
| 4670 | auto *VPSC = cast<VPScatterSDNode>(Val: N); |
| 4671 | return {.Mask: VPSC->getMask(), .Index: VPSC->getIndex(), .Scale: VPSC->getScale(), |
| 4672 | .Data: VPSC->getValue()}; |
| 4673 | }(); |
| 4674 | // Split all operands |
| 4675 | |
| 4676 | EVT LoMemVT, HiMemVT; |
| 4677 | std::tie(args&: LoMemVT, args&: HiMemVT) = DAG.GetSplitDestVTs(VT: MemoryVT); |
| 4678 | |
| 4679 | SDValue DataLo, DataHi; |
| 4680 | if (getTypeAction(VT: Ops.Data.getValueType()) == TargetLowering::TypeSplitVector) |
| 4681 | // Split Data operand |
| 4682 | GetSplitVector(Op: Ops.Data, Lo&: DataLo, Hi&: DataHi); |
| 4683 | else |
| 4684 | std::tie(args&: DataLo, args&: DataHi) = DAG.SplitVector(N: Ops.Data, DL); |
| 4685 | |
| 4686 | // Split Mask operand |
| 4687 | SDValue MaskLo, MaskHi; |
| 4688 | if (OpNo == 1 && Ops.Mask.getOpcode() == ISD::SETCC) { |
| 4689 | SplitVecRes_SETCC(N: Ops.Mask.getNode(), Lo&: MaskLo, Hi&: MaskHi); |
| 4690 | } else { |
| 4691 | std::tie(args&: MaskLo, args&: MaskHi) = SplitMask(Mask: Ops.Mask, DL); |
| 4692 | } |
| 4693 | |
| 4694 | SDValue IndexHi, IndexLo; |
| 4695 | if (getTypeAction(VT: Ops.Index.getValueType()) == |
| 4696 | TargetLowering::TypeSplitVector) |
| 4697 | GetSplitVector(Op: Ops.Index, Lo&: IndexLo, Hi&: IndexHi); |
| 4698 | else |
| 4699 | std::tie(args&: IndexLo, args&: IndexHi) = DAG.SplitVector(N: Ops.Index, DL); |
| 4700 | |
| 4701 | SDValue Lo; |
| 4702 | MachineMemOperand::Flags MMOFlags = N->getMemOperand()->getFlags(); |
| 4703 | MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand( |
| 4704 | PtrInfo: N->getPointerInfo(), F: MMOFlags, Size: LocationSize::beforeOrAfterPointer(), |
| 4705 | BaseAlignment: Alignment, AAInfo: N->getAAInfo(), Ranges: N->getRanges()); |
| 4706 | |
| 4707 | if (auto *MSC = dyn_cast<MaskedScatterSDNode>(Val: N)) { |
| 4708 | SDValue OpsLo[] = {Ch, DataLo, MaskLo, Ptr, IndexLo, Ops.Scale}; |
| 4709 | Lo = |
| 4710 | DAG.getMaskedScatter(VTs: DAG.getVTList(VT: MVT::Other), MemVT: LoMemVT, dl: DL, Ops: OpsLo, MMO, |
| 4711 | IndexType: MSC->getIndexType(), IsTruncating: MSC->isTruncatingStore()); |
| 4712 | |
| 4713 | // The order of the Scatter operation after split is well defined. The "Hi" |
| 4714 | // part comes after the "Lo". So these two operations should be chained one |
| 4715 | // after another. |
| 4716 | SDValue OpsHi[] = {Lo, DataHi, MaskHi, Ptr, IndexHi, Ops.Scale}; |
| 4717 | return DAG.getMaskedScatter(VTs: DAG.getVTList(VT: MVT::Other), MemVT: HiMemVT, dl: DL, Ops: OpsHi, |
| 4718 | MMO, IndexType: MSC->getIndexType(), |
| 4719 | IsTruncating: MSC->isTruncatingStore()); |
| 4720 | } |
| 4721 | auto *VPSC = cast<VPScatterSDNode>(Val: N); |
| 4722 | SDValue EVLLo, EVLHi; |
| 4723 | std::tie(args&: EVLLo, args&: EVLHi) = |
| 4724 | DAG.SplitEVL(N: VPSC->getVectorLength(), VecVT: Ops.Data.getValueType(), DL); |
| 4725 | |
| 4726 | SDValue OpsLo[] = {Ch, DataLo, Ptr, IndexLo, Ops.Scale, MaskLo, EVLLo}; |
| 4727 | Lo = DAG.getScatterVP(VTs: DAG.getVTList(VT: MVT::Other), VT: LoMemVT, dl: DL, Ops: OpsLo, MMO, |
| 4728 | IndexType: VPSC->getIndexType()); |
| 4729 | |
| 4730 | // The order of the Scatter operation after split is well defined. The "Hi" |
| 4731 | // part comes after the "Lo". So these two operations should be chained one |
| 4732 | // after another. |
| 4733 | SDValue OpsHi[] = {Lo, DataHi, Ptr, IndexHi, Ops.Scale, MaskHi, EVLHi}; |
| 4734 | return DAG.getScatterVP(VTs: DAG.getVTList(VT: MVT::Other), VT: HiMemVT, dl: DL, Ops: OpsHi, MMO, |
| 4735 | IndexType: VPSC->getIndexType()); |
| 4736 | } |
| 4737 | |
| 4738 | SDValue DAGTypeLegalizer::SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo) { |
| 4739 | assert(N->isUnindexed() && "Indexed store of vector?" ); |
| 4740 | assert(OpNo == 1 && "Can only split the stored value" ); |
| 4741 | SDLoc DL(N); |
| 4742 | |
| 4743 | bool isTruncating = N->isTruncatingStore(); |
| 4744 | SDValue Ch = N->getChain(); |
| 4745 | SDValue Ptr = N->getBasePtr(); |
| 4746 | EVT MemoryVT = N->getMemoryVT(); |
| 4747 | Align Alignment = N->getBaseAlign(); |
| 4748 | MachineMemOperand::Flags MMOFlags = N->getMemOperand()->getFlags(); |
| 4749 | AAMDNodes AAInfo = N->getAAInfo(); |
| 4750 | SDValue Lo, Hi; |
| 4751 | GetSplitVector(Op: N->getOperand(Num: 1), Lo, Hi); |
| 4752 | |
| 4753 | EVT LoMemVT, HiMemVT; |
| 4754 | std::tie(args&: LoMemVT, args&: HiMemVT) = DAG.GetSplitDestVTs(VT: MemoryVT); |
| 4755 | |
| 4756 | // Scalarize if the split halves are not byte-sized. |
| 4757 | if (!LoMemVT.isByteSized() || !HiMemVT.isByteSized()) |
| 4758 | return TLI.scalarizeVectorStore(ST: N, DAG); |
| 4759 | |
| 4760 | if (isTruncating) |
| 4761 | Lo = DAG.getTruncStore(Chain: Ch, dl: DL, Val: Lo, Ptr, PtrInfo: N->getPointerInfo(), SVT: LoMemVT, |
| 4762 | Alignment, MMOFlags, AAInfo); |
| 4763 | else |
| 4764 | Lo = DAG.getStore(Chain: Ch, dl: DL, Val: Lo, Ptr, PtrInfo: N->getPointerInfo(), Alignment, MMOFlags, |
| 4765 | AAInfo); |
| 4766 | |
| 4767 | MachinePointerInfo MPI; |
| 4768 | IncrementPointer(N, MemVT: LoMemVT, MPI, Ptr); |
| 4769 | |
| 4770 | if (isTruncating) |
| 4771 | Hi = DAG.getTruncStore(Chain: Ch, dl: DL, Val: Hi, Ptr, PtrInfo: MPI, |
| 4772 | SVT: HiMemVT, Alignment, MMOFlags, AAInfo); |
| 4773 | else |
| 4774 | Hi = DAG.getStore(Chain: Ch, dl: DL, Val: Hi, Ptr, PtrInfo: MPI, Alignment, MMOFlags, AAInfo); |
| 4775 | |
| 4776 | return DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, N1: Lo, N2: Hi); |
| 4777 | } |
| 4778 | |
| 4779 | SDValue DAGTypeLegalizer::SplitVecOp_ATOMIC_STORE(AtomicSDNode *N) { |
| 4780 | SDLoc DL(N); |
| 4781 | LLVMContext &Ctx = *DAG.getContext(); |
| 4782 | SDValue StVal = N->getVal(); |
| 4783 | EVT VT = StVal.getValueType(); |
| 4784 | EVT MemIntVT = EVT::getIntegerVT(Context&: Ctx, BitWidth: N->getMemoryVT().getSizeInBits()); |
| 4785 | |
| 4786 | // The store needs a single value spanning the full memory width. If the |
| 4787 | // value can be held in a legal vector register, keep it there and extract |
| 4788 | // the low integer element of the memory width. This lets the store be issued |
| 4789 | // directly from a vector register (e.g. a single MOVQ/MOVD) instead of |
| 4790 | // bitcasting the split vector straight to a scalar integer, which would |
| 4791 | // reassemble the value element by element in GPRs. |
| 4792 | // |
| 4793 | // Reinterpret the value as a same-shaped integer vector first: an FP element |
| 4794 | // type may not have a legal vector form (e.g. bfloat on SSE2) while the |
| 4795 | // integer-of-element-size form does. Ask the target which legal vector type |
| 4796 | // it widens to. |
| 4797 | EVT IntVecVT = VT.changeVectorElementTypeToInteger(); |
| 4798 | EVT IntEltVT = IntVecVT.getVectorElementType(); |
| 4799 | EVT WideVT = TLI.getLegalTypeToTransformTo(Context&: Ctx, VT: IntVecVT); |
| 4800 | if (DAG.getDataLayout().isLittleEndian() && TLI.isTypeLegal(VT: MemIntVT) && |
| 4801 | WideVT.isVector() && WideVT.getVectorElementType() == IntEltVT && |
| 4802 | IntEltVT.getSizeInBits() <= MemIntVT.getSizeInBits() && |
| 4803 | WideVT.getSizeInBits() % MemIntVT.getSizeInBits() == 0) { |
| 4804 | SDValue Wide = ModifyToType(InOp: DAG.getBitcast(VT: IntVecVT, V: StVal), NVT: WideVT); |
| 4805 | unsigned NumMemElts = WideVT.getSizeInBits() / MemIntVT.getSizeInBits(); |
| 4806 | EVT MemVecVT = EVT::getVectorVT(Context&: Ctx, VT: MemIntVT, NumElements: NumMemElts); |
| 4807 | SDValue Elt = DAG.getExtractVectorElt(DL, VT: MemIntVT, |
| 4808 | Vec: DAG.getBitcast(VT: MemVecVT, V: Wide), Idx: 0); |
| 4809 | return DAG.getAtomic(Opcode: ISD::ATOMIC_STORE, dl: DL, MemVT: MemIntVT, Chain: N->getChain(), Ptr: Elt, |
| 4810 | Val: N->getBasePtr(), MMO: N->getMemOperand()); |
| 4811 | } |
| 4812 | |
| 4813 | // Otherwise issue a single atomic store of an integer that spans the full |
| 4814 | // memory width. Bitcasting the (illegal) vector value to that integer lets |
| 4815 | // the type legalizer further legalize the BITCAST input as needed, while the |
| 4816 | // ATOMIC_STORE itself uses only the legal integer type. |
| 4817 | EVT IntVT = EVT::getIntegerVT(Context&: Ctx, BitWidth: VT.getSizeInBits()); |
| 4818 | SDValue AsInt = DAG.getBitcast(VT: IntVT, V: StVal); |
| 4819 | return DAG.getAtomic(Opcode: ISD::ATOMIC_STORE, dl: DL, MemVT: MemIntVT, Chain: N->getChain(), Ptr: AsInt, |
| 4820 | Val: N->getBasePtr(), MMO: N->getMemOperand()); |
| 4821 | } |
| 4822 | |
| 4823 | SDValue DAGTypeLegalizer::SplitVecOp_CONCAT_VECTORS(SDNode *N) { |
| 4824 | SDLoc DL(N); |
| 4825 | |
| 4826 | // The input operands all must have the same type, and we know the result |
| 4827 | // type is valid. Convert this to a buildvector which extracts all the |
| 4828 | // input elements. |
| 4829 | // TODO: If the input elements are power-two vectors, we could convert this to |
| 4830 | // a new CONCAT_VECTORS node with elements that are half-wide. |
| 4831 | SmallVector<SDValue, 32> Elts; |
| 4832 | EVT EltVT = N->getValueType(ResNo: 0).getVectorElementType(); |
| 4833 | for (const SDValue &Op : N->op_values()) { |
| 4834 | for (unsigned i = 0, e = Op.getValueType().getVectorNumElements(); |
| 4835 | i != e; ++i) { |
| 4836 | Elts.push_back(Elt: DAG.getExtractVectorElt(DL, VT: EltVT, Vec: Op, Idx: i)); |
| 4837 | } |
| 4838 | } |
| 4839 | |
| 4840 | return DAG.getBuildVector(VT: N->getValueType(ResNo: 0), DL, Ops: Elts); |
| 4841 | } |
| 4842 | |
| 4843 | SDValue DAGTypeLegalizer::SplitVecOp_TruncateHelper(SDNode *N) { |
| 4844 | // The result type is legal, but the input type is illegal. If splitting |
| 4845 | // ends up with the result type of each half still being legal, just |
| 4846 | // do that. If, however, that would result in an illegal result type, |
| 4847 | // we can try to get more clever with power-two vectors. Specifically, |
| 4848 | // split the input type, but also widen the result element size, then |
| 4849 | // concatenate the halves and truncate again. For example, consider a target |
| 4850 | // where v8i8 is legal and v8i32 is not (ARM, which doesn't have 256-bit |
| 4851 | // vectors). To perform a "%res = v8i8 trunc v8i32 %in" we do: |
| 4852 | // %inlo = v4i32 extract_subvector %in, 0 |
| 4853 | // %inhi = v4i32 extract_subvector %in, 4 |
| 4854 | // %lo16 = v4i16 trunc v4i32 %inlo |
| 4855 | // %hi16 = v4i16 trunc v4i32 %inhi |
| 4856 | // %in16 = v8i16 concat_vectors v4i16 %lo16, v4i16 %hi16 |
| 4857 | // %res = v8i8 trunc v8i16 %in16 |
| 4858 | // |
| 4859 | // Without this transform, the original truncate would end up being |
| 4860 | // scalarized, which is pretty much always a last resort. |
| 4861 | unsigned OpNo = N->isStrictFPOpcode() ? 1 : 0; |
| 4862 | SDValue InVec = N->getOperand(Num: OpNo); |
| 4863 | EVT InVT = InVec->getValueType(ResNo: 0); |
| 4864 | EVT OutVT = N->getValueType(ResNo: 0); |
| 4865 | ElementCount NumElements = OutVT.getVectorElementCount(); |
| 4866 | bool IsFloat = OutVT.isFloatingPoint(); |
| 4867 | |
| 4868 | unsigned InElementSize = InVT.getScalarSizeInBits(); |
| 4869 | unsigned OutElementSize = OutVT.getScalarSizeInBits(); |
| 4870 | |
| 4871 | // Determine the split output VT. If its legal we can just split dirctly. |
| 4872 | EVT LoOutVT, HiOutVT; |
| 4873 | std::tie(args&: LoOutVT, args&: HiOutVT) = DAG.GetSplitDestVTs(VT: OutVT); |
| 4874 | assert(LoOutVT == HiOutVT && "Unequal split?" ); |
| 4875 | |
| 4876 | // If the input elements are only 1/2 the width of the result elements, |
| 4877 | // just use the normal splitting. Our trick only work if there's room |
| 4878 | // to split more than once. |
| 4879 | if (isTypeLegal(VT: LoOutVT) || InElementSize <= OutElementSize * 2 || |
| 4880 | (IsFloat && !isPowerOf2_32(Value: InElementSize))) |
| 4881 | return SplitVecOp_UnaryOp(N); |
| 4882 | SDLoc DL(N); |
| 4883 | |
| 4884 | // Don't touch if this will be scalarized. |
| 4885 | EVT FinalVT = InVT; |
| 4886 | while (getTypeAction(VT: FinalVT) == TargetLowering::TypeSplitVector) |
| 4887 | FinalVT = FinalVT.getHalfNumVectorElementsVT(Context&: *DAG.getContext()); |
| 4888 | |
| 4889 | if (getTypeAction(VT: FinalVT) == TargetLowering::TypeScalarizeVector) |
| 4890 | return SplitVecOp_UnaryOp(N); |
| 4891 | |
| 4892 | // Get the split input vector. |
| 4893 | SDValue InLoVec, InHiVec; |
| 4894 | GetSplitVector(Op: InVec, Lo&: InLoVec, Hi&: InHiVec); |
| 4895 | |
| 4896 | // Truncate them to 1/2 the element size. |
| 4897 | // |
| 4898 | // This assumes the number of elements is a power of two; any vector that |
| 4899 | // isn't should be widened, not split. |
| 4900 | EVT HalfElementVT = IsFloat ? |
| 4901 | EVT::getFloatingPointVT(BitWidth: InElementSize/2) : |
| 4902 | EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: InElementSize/2); |
| 4903 | EVT HalfVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: HalfElementVT, |
| 4904 | EC: NumElements.divideCoefficientBy(RHS: 2)); |
| 4905 | |
| 4906 | SDValue HalfLo; |
| 4907 | SDValue HalfHi; |
| 4908 | SDValue Chain; |
| 4909 | if (N->isStrictFPOpcode()) { |
| 4910 | HalfLo = DAG.getNode(Opcode: N->getOpcode(), DL, ResultTys: {HalfVT, MVT::Other}, |
| 4911 | Ops: {N->getOperand(Num: 0), InLoVec}); |
| 4912 | HalfHi = DAG.getNode(Opcode: N->getOpcode(), DL, ResultTys: {HalfVT, MVT::Other}, |
| 4913 | Ops: {N->getOperand(Num: 0), InHiVec}); |
| 4914 | // Legalize the chain result - switch anything that used the old chain to |
| 4915 | // use the new one. |
| 4916 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, N1: HalfLo.getValue(R: 1), |
| 4917 | N2: HalfHi.getValue(R: 1)); |
| 4918 | } else { |
| 4919 | HalfLo = DAG.getNode(Opcode: N->getOpcode(), DL, VT: HalfVT, Operand: InLoVec); |
| 4920 | HalfHi = DAG.getNode(Opcode: N->getOpcode(), DL, VT: HalfVT, Operand: InHiVec); |
| 4921 | } |
| 4922 | |
| 4923 | // Concatenate them to get the full intermediate truncation result. |
| 4924 | EVT InterVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: HalfElementVT, EC: NumElements); |
| 4925 | SDValue InterVec = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: InterVT, N1: HalfLo, |
| 4926 | N2: HalfHi); |
| 4927 | // Now finish up by truncating all the way down to the original result |
| 4928 | // type. This should normally be something that ends up being legal directly, |
| 4929 | // but in theory if a target has very wide vectors and an annoyingly |
| 4930 | // restricted set of legal types, this split can chain to build things up. |
| 4931 | |
| 4932 | if (N->isStrictFPOpcode()) { |
| 4933 | SDValue Res = DAG.getNode( |
| 4934 | Opcode: ISD::STRICT_FP_ROUND, DL, ResultTys: {OutVT, MVT::Other}, |
| 4935 | Ops: {Chain, InterVec, |
| 4936 | DAG.getTargetConstant(Val: 0, DL, VT: TLI.getPointerTy(DL: DAG.getDataLayout()))}); |
| 4937 | // Relink the chain |
| 4938 | ReplaceValueWith(From: SDValue(N, 1), To: SDValue(Res.getNode(), 1)); |
| 4939 | return Res; |
| 4940 | } |
| 4941 | |
| 4942 | return IsFloat |
| 4943 | ? DAG.getNode(Opcode: ISD::FP_ROUND, DL, VT: OutVT, N1: InterVec, |
| 4944 | N2: DAG.getTargetConstant( |
| 4945 | Val: 0, DL, VT: TLI.getPointerTy(DL: DAG.getDataLayout()))) |
| 4946 | : DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: OutVT, Operand: InterVec); |
| 4947 | } |
| 4948 | |
| 4949 | SDValue DAGTypeLegalizer::SplitVecOp_VSETCC(SDNode *N) { |
| 4950 | unsigned Opc = N->getOpcode(); |
| 4951 | bool isStrict = Opc == ISD::STRICT_FSETCC || Opc == ISD::STRICT_FSETCCS; |
| 4952 | assert(N->getValueType(0).isVector() && |
| 4953 | N->getOperand(isStrict ? 1 : 0).getValueType().isVector() && |
| 4954 | "Operand types must be vectors" ); |
| 4955 | // The result has a legal vector type, but the input needs splitting. |
| 4956 | SDValue Lo0, Hi0, Lo1, Hi1, LoRes, HiRes; |
| 4957 | SDLoc DL(N); |
| 4958 | GetSplitVector(Op: N->getOperand(Num: isStrict ? 1 : 0), Lo&: Lo0, Hi&: Hi0); |
| 4959 | GetSplitVector(Op: N->getOperand(Num: isStrict ? 2 : 1), Lo&: Lo1, Hi&: Hi1); |
| 4960 | |
| 4961 | EVT VT = N->getValueType(ResNo: 0); |
| 4962 | EVT PartResVT = getSetCCResultType(VT: Lo0.getValueType()); |
| 4963 | |
| 4964 | if (Opc == ISD::SETCC) { |
| 4965 | LoRes = DAG.getNode(Opcode: ISD::SETCC, DL, VT: PartResVT, N1: Lo0, N2: Lo1, N3: N->getOperand(Num: 2)); |
| 4966 | HiRes = DAG.getNode(Opcode: ISD::SETCC, DL, VT: PartResVT, N1: Hi0, N2: Hi1, N3: N->getOperand(Num: 2)); |
| 4967 | } else if (isStrict) { |
| 4968 | LoRes = DAG.getNode(Opcode: Opc, DL, VTList: DAG.getVTList(VT1: PartResVT, VT2: N->getValueType(ResNo: 1)), |
| 4969 | N1: N->getOperand(Num: 0), N2: Lo0, N3: Lo1, N4: N->getOperand(Num: 3)); |
| 4970 | HiRes = DAG.getNode(Opcode: Opc, DL, VTList: DAG.getVTList(VT1: PartResVT, VT2: N->getValueType(ResNo: 1)), |
| 4971 | N1: N->getOperand(Num: 0), N2: Hi0, N3: Hi1, N4: N->getOperand(Num: 3)); |
| 4972 | SDValue NewChain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, |
| 4973 | N1: LoRes.getValue(R: 1), N2: HiRes.getValue(R: 1)); |
| 4974 | ReplaceValueWith(From: SDValue(N, 1), To: NewChain); |
| 4975 | } else { |
| 4976 | assert(Opc == ISD::VP_SETCC && "Expected VP_SETCC opcode" ); |
| 4977 | SDValue MaskLo, MaskHi, EVLLo, EVLHi; |
| 4978 | std::tie(args&: MaskLo, args&: MaskHi) = SplitMask(Mask: N->getOperand(Num: 3)); |
| 4979 | std::tie(args&: EVLLo, args&: EVLHi) = |
| 4980 | DAG.SplitEVL(N: N->getOperand(Num: 4), VecVT: N->getValueType(ResNo: 0), DL); |
| 4981 | LoRes = DAG.getNode(Opcode: ISD::VP_SETCC, DL, VT: PartResVT, N1: Lo0, N2: Lo1, |
| 4982 | N3: N->getOperand(Num: 2), N4: MaskLo, N5: EVLLo); |
| 4983 | HiRes = DAG.getNode(Opcode: ISD::VP_SETCC, DL, VT: PartResVT, N1: Hi0, N2: Hi1, |
| 4984 | N3: N->getOperand(Num: 2), N4: MaskHi, N5: EVLHi); |
| 4985 | } |
| 4986 | |
| 4987 | EVT ConcatVT = PartResVT.getDoubleNumVectorElementsVT(Context&: *DAG.getContext()); |
| 4988 | SDValue Con = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: ConcatVT, N1: LoRes, N2: HiRes); |
| 4989 | if (VT == ConcatVT) |
| 4990 | return Con; |
| 4991 | |
| 4992 | EVT OpVT = N->getOperand(Num: 0).getValueType(); |
| 4993 | ISD::NodeType ExtendCode = |
| 4994 | TargetLowering::getExtendForContent(Content: TLI.getBooleanContents(Type: OpVT)); |
| 4995 | return DAG.getExtOrTrunc(Op: Con, DL, VT, Opcode: ExtendCode); |
| 4996 | } |
| 4997 | |
| 4998 | |
| 4999 | SDValue DAGTypeLegalizer::SplitVecOp_FP_ROUND(SDNode *N) { |
| 5000 | // The result has a legal vector type, but the input needs splitting. |
| 5001 | EVT ResVT = N->getValueType(ResNo: 0); |
| 5002 | SDValue Lo, Hi; |
| 5003 | SDLoc DL(N); |
| 5004 | GetSplitVector(Op: N->getOperand(Num: N->isStrictFPOpcode() ? 1 : 0), Lo, Hi); |
| 5005 | EVT InVT = Lo.getValueType(); |
| 5006 | |
| 5007 | EVT OutVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: ResVT.getVectorElementType(), |
| 5008 | EC: InVT.getVectorElementCount()); |
| 5009 | |
| 5010 | if (N->isStrictFPOpcode()) { |
| 5011 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL, ResultTys: {OutVT, MVT::Other}, |
| 5012 | Ops: {N->getOperand(Num: 0), Lo, N->getOperand(Num: 2)}); |
| 5013 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL, ResultTys: {OutVT, MVT::Other}, |
| 5014 | Ops: {N->getOperand(Num: 0), Hi, N->getOperand(Num: 2)}); |
| 5015 | // Legalize the chain result - switch anything that used the old chain to |
| 5016 | // use the new one. |
| 5017 | SDValue NewChain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, |
| 5018 | N1: Lo.getValue(R: 1), N2: Hi.getValue(R: 1)); |
| 5019 | ReplaceValueWith(From: SDValue(N, 1), To: NewChain); |
| 5020 | } else if (N->getOpcode() == ISD::VP_FP_ROUND) { |
| 5021 | SDValue MaskLo, MaskHi, EVLLo, EVLHi; |
| 5022 | std::tie(args&: MaskLo, args&: MaskHi) = SplitMask(Mask: N->getOperand(Num: 1)); |
| 5023 | std::tie(args&: EVLLo, args&: EVLHi) = |
| 5024 | DAG.SplitEVL(N: N->getOperand(Num: 2), VecVT: N->getValueType(ResNo: 0), DL); |
| 5025 | Lo = DAG.getNode(Opcode: ISD::VP_FP_ROUND, DL, VT: OutVT, N1: Lo, N2: MaskLo, N3: EVLLo); |
| 5026 | Hi = DAG.getNode(Opcode: ISD::VP_FP_ROUND, DL, VT: OutVT, N1: Hi, N2: MaskHi, N3: EVLHi); |
| 5027 | } else if (N->getOpcode() == ISD::CONVERT_TO_ARBITRARY_FP) { |
| 5028 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL, VT: OutVT, N1: Lo, N2: N->getOperand(Num: 1), |
| 5029 | N3: N->getOperand(Num: 2), N4: N->getOperand(Num: 3)); |
| 5030 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL, VT: OutVT, N1: Hi, N2: N->getOperand(Num: 1), |
| 5031 | N3: N->getOperand(Num: 2), N4: N->getOperand(Num: 3)); |
| 5032 | } else { |
| 5033 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL, VT: OutVT, N1: Lo, N2: N->getOperand(Num: 1)); |
| 5034 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL, VT: OutVT, N1: Hi, N2: N->getOperand(Num: 1)); |
| 5035 | } |
| 5036 | |
| 5037 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: ResVT, N1: Lo, N2: Hi); |
| 5038 | } |
| 5039 | |
| 5040 | // Split a vector type in an FP binary operation where the second operand has a |
| 5041 | // different type from the first. |
| 5042 | // |
| 5043 | // The result (and the first input) has a legal vector type, but the second |
| 5044 | // input needs splitting. |
| 5045 | SDValue DAGTypeLegalizer::SplitVecOp_FPOpDifferentTypes(SDNode *N) { |
| 5046 | SDLoc DL(N); |
| 5047 | |
| 5048 | EVT LHSLoVT, LHSHiVT; |
| 5049 | std::tie(args&: LHSLoVT, args&: LHSHiVT) = DAG.GetSplitDestVTs(VT: N->getValueType(ResNo: 0)); |
| 5050 | |
| 5051 | if (!isTypeLegal(VT: LHSLoVT) || !isTypeLegal(VT: LHSHiVT)) |
| 5052 | return DAG.UnrollVectorOp(N, ResNE: N->getValueType(ResNo: 0).getVectorNumElements()); |
| 5053 | |
| 5054 | SDValue LHSLo, LHSHi; |
| 5055 | std::tie(args&: LHSLo, args&: LHSHi) = |
| 5056 | DAG.SplitVector(N: N->getOperand(Num: 0), DL, LoVT: LHSLoVT, HiVT: LHSHiVT); |
| 5057 | |
| 5058 | SDValue RHSLo, RHSHi; |
| 5059 | std::tie(args&: RHSLo, args&: RHSHi) = DAG.SplitVector(N: N->getOperand(Num: 1), DL); |
| 5060 | |
| 5061 | SDValue Lo = DAG.getNode(Opcode: N->getOpcode(), DL, VT: LHSLoVT, N1: LHSLo, N2: RHSLo); |
| 5062 | SDValue Hi = DAG.getNode(Opcode: N->getOpcode(), DL, VT: LHSHiVT, N1: LHSHi, N2: RHSHi); |
| 5063 | |
| 5064 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: N->getValueType(ResNo: 0), N1: Lo, N2: Hi); |
| 5065 | } |
| 5066 | |
| 5067 | SDValue DAGTypeLegalizer::SplitVecOp_CMP(SDNode *N) { |
| 5068 | LLVMContext &Ctxt = *DAG.getContext(); |
| 5069 | SDLoc dl(N); |
| 5070 | |
| 5071 | SDValue LHSLo, LHSHi, RHSLo, RHSHi; |
| 5072 | GetSplitVector(Op: N->getOperand(Num: 0), Lo&: LHSLo, Hi&: LHSHi); |
| 5073 | GetSplitVector(Op: N->getOperand(Num: 1), Lo&: RHSLo, Hi&: RHSHi); |
| 5074 | |
| 5075 | EVT ResVT = N->getValueType(ResNo: 0); |
| 5076 | ElementCount SplitOpEC = LHSLo.getValueType().getVectorElementCount(); |
| 5077 | EVT NewResVT = |
| 5078 | EVT::getVectorVT(Context&: Ctxt, VT: ResVT.getVectorElementType(), EC: SplitOpEC); |
| 5079 | |
| 5080 | SDValue Lo = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: NewResVT, N1: LHSLo, N2: RHSLo); |
| 5081 | SDValue Hi = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: NewResVT, N1: LHSHi, N2: RHSHi); |
| 5082 | |
| 5083 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: ResVT, N1: Lo, N2: Hi); |
| 5084 | } |
| 5085 | |
| 5086 | SDValue DAGTypeLegalizer::SplitVecOp_FP_TO_XINT_SAT(SDNode *N) { |
| 5087 | EVT ResVT = N->getValueType(ResNo: 0); |
| 5088 | SDValue Lo, Hi; |
| 5089 | SDLoc dl(N); |
| 5090 | GetSplitVector(Op: N->getOperand(Num: 0), Lo, Hi); |
| 5091 | EVT InVT = Lo.getValueType(); |
| 5092 | |
| 5093 | EVT NewResVT = |
| 5094 | EVT::getVectorVT(Context&: *DAG.getContext(), VT: ResVT.getVectorElementType(), |
| 5095 | EC: InVT.getVectorElementCount()); |
| 5096 | |
| 5097 | Lo = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: NewResVT, N1: Lo, N2: N->getOperand(Num: 1)); |
| 5098 | Hi = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: NewResVT, N1: Hi, N2: N->getOperand(Num: 1)); |
| 5099 | |
| 5100 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: ResVT, N1: Lo, N2: Hi); |
| 5101 | } |
| 5102 | |
| 5103 | SDValue DAGTypeLegalizer::SplitVecOp_CttzElts(SDNode *N) { |
| 5104 | SDLoc DL(N); |
| 5105 | EVT ResVT = N->getValueType(ResNo: 0); |
| 5106 | |
| 5107 | SDValue Lo, Hi; |
| 5108 | SDValue VecOp = N->getOperand(Num: 0); |
| 5109 | GetSplitVector(Op: VecOp, Lo, Hi); |
| 5110 | |
| 5111 | // if CTTZ_ELTS(Lo) != VL => CTTZ_ELTS(Lo). |
| 5112 | // else => VL + (CTTZ_ELTS(Hi) or CTTZ_ELTS_ZERO_POISON(Hi)). |
| 5113 | SDValue ResLo = DAG.getNode(Opcode: ISD::CTTZ_ELTS, DL, VT: ResVT, Operand: Lo); |
| 5114 | SDValue VL = |
| 5115 | DAG.getElementCount(DL, VT: ResVT, EC: Lo.getValueType().getVectorElementCount()); |
| 5116 | SDValue ResLoNotVL = |
| 5117 | DAG.getSetCC(DL, VT: getSetCCResultType(VT: ResVT), LHS: ResLo, RHS: VL, Cond: ISD::SETNE); |
| 5118 | SDValue ResHi = DAG.getNode(Opcode: N->getOpcode(), DL, VT: ResVT, Operand: Hi); |
| 5119 | return DAG.getSelect(DL, VT: ResVT, Cond: ResLoNotVL, LHS: ResLo, |
| 5120 | RHS: DAG.getNode(Opcode: ISD::ADD, DL, VT: ResVT, N1: VL, N2: ResHi)); |
| 5121 | } |
| 5122 | |
| 5123 | SDValue DAGTypeLegalizer::SplitVecOp_VP_CttzElements(SDNode *N) { |
| 5124 | SDLoc DL(N); |
| 5125 | EVT ResVT = N->getValueType(ResNo: 0); |
| 5126 | |
| 5127 | SDValue Lo, Hi; |
| 5128 | SDValue VecOp = N->getOperand(Num: 0); |
| 5129 | GetSplitVector(Op: VecOp, Lo, Hi); |
| 5130 | |
| 5131 | auto [MaskLo, MaskHi] = SplitMask(Mask: N->getOperand(Num: 1)); |
| 5132 | auto [EVLLo, EVLHi] = |
| 5133 | DAG.SplitEVL(N: N->getOperand(Num: 2), VecVT: VecOp.getValueType(), DL); |
| 5134 | SDValue VLo = DAG.getZExtOrTrunc(Op: EVLLo, DL, VT: ResVT); |
| 5135 | |
| 5136 | // if VP_CTTZ_ELTS(Lo) != EVLLo => VP_CTTZ_ELTS(Lo). |
| 5137 | // else => EVLLo + (VP_CTTZ_ELTS(Hi) or VP_CTTZ_ELTS_ZERO_POISON(Hi)). |
| 5138 | SDValue ResLo = DAG.getNode(Opcode: ISD::VP_CTTZ_ELTS, DL, VT: ResVT, N1: Lo, N2: MaskLo, N3: EVLLo); |
| 5139 | SDValue ResLoNotEVL = |
| 5140 | DAG.getSetCC(DL, VT: getSetCCResultType(VT: ResVT), LHS: ResLo, RHS: VLo, Cond: ISD::SETNE); |
| 5141 | SDValue ResHi = DAG.getNode(Opcode: N->getOpcode(), DL, VT: ResVT, N1: Hi, N2: MaskHi, N3: EVLHi); |
| 5142 | return DAG.getSelect(DL, VT: ResVT, Cond: ResLoNotEVL, LHS: ResLo, |
| 5143 | RHS: DAG.getNode(Opcode: ISD::ADD, DL, VT: ResVT, N1: VLo, N2: ResHi)); |
| 5144 | } |
| 5145 | |
| 5146 | SDValue DAGTypeLegalizer::SplitVecOp_VECTOR_HISTOGRAM(SDNode *N) { |
| 5147 | MaskedHistogramSDNode *HG = cast<MaskedHistogramSDNode>(Val: N); |
| 5148 | SDLoc DL(HG); |
| 5149 | SDValue Inc = HG->getInc(); |
| 5150 | SDValue Ptr = HG->getBasePtr(); |
| 5151 | SDValue Scale = HG->getScale(); |
| 5152 | SDValue IntID = HG->getIntID(); |
| 5153 | EVT MemVT = HG->getMemoryVT(); |
| 5154 | MachineMemOperand *MMO = HG->getMemOperand(); |
| 5155 | ISD::MemIndexType IndexType = HG->getIndexType(); |
| 5156 | |
| 5157 | SDValue IndexLo, IndexHi, MaskLo, MaskHi; |
| 5158 | std::tie(args&: IndexLo, args&: IndexHi) = DAG.SplitVector(N: HG->getIndex(), DL); |
| 5159 | std::tie(args&: MaskLo, args&: MaskHi) = DAG.SplitVector(N: HG->getMask(), DL); |
| 5160 | SDValue OpsLo[] = {HG->getChain(), Inc, MaskLo, Ptr, IndexLo, Scale, IntID}; |
| 5161 | SDValue Lo = DAG.getMaskedHistogram(VTs: DAG.getVTList(VT: MVT::Other), MemVT, dl: DL, |
| 5162 | Ops: OpsLo, MMO, IndexType); |
| 5163 | SDValue OpsHi[] = {Lo, Inc, MaskHi, Ptr, IndexHi, Scale, IntID}; |
| 5164 | return DAG.getMaskedHistogram(VTs: DAG.getVTList(VT: MVT::Other), MemVT, dl: DL, Ops: OpsHi, |
| 5165 | MMO, IndexType); |
| 5166 | } |
| 5167 | |
| 5168 | SDValue DAGTypeLegalizer::SplitVecOp_PARTIAL_REDUCE_MLA(SDNode *N) { |
| 5169 | SDValue Acc = N->getOperand(Num: 0); |
| 5170 | assert(getTypeAction(Acc.getValueType()) != TargetLowering::TypeSplitVector && |
| 5171 | "Accumulator should already be a legal type, and shouldn't need " |
| 5172 | "further splitting" ); |
| 5173 | |
| 5174 | SDLoc DL(N); |
| 5175 | SDValue Input1Lo, Input1Hi, Input2Lo, Input2Hi; |
| 5176 | GetSplitVector(Op: N->getOperand(Num: 1), Lo&: Input1Lo, Hi&: Input1Hi); |
| 5177 | GetSplitVector(Op: N->getOperand(Num: 2), Lo&: Input2Lo, Hi&: Input2Hi); |
| 5178 | unsigned Opcode = N->getOpcode(); |
| 5179 | EVT ResultVT = Acc.getValueType(); |
| 5180 | |
| 5181 | SDValue Lo = DAG.getNode(Opcode, DL, VT: ResultVT, N1: Acc, N2: Input1Lo, N3: Input2Lo); |
| 5182 | return DAG.getNode(Opcode, DL, VT: ResultVT, N1: Lo, N2: Input1Hi, N3: Input2Hi); |
| 5183 | } |
| 5184 | |
| 5185 | //===----------------------------------------------------------------------===// |
| 5186 | // Result Vector Widening |
| 5187 | //===----------------------------------------------------------------------===// |
| 5188 | |
| 5189 | void DAGTypeLegalizer::ReplaceOtherWidenResults(SDNode *N, SDNode *WidenNode, |
| 5190 | unsigned WidenResNo) { |
| 5191 | unsigned NumResults = N->getNumValues(); |
| 5192 | for (unsigned ResNo = 0; ResNo < NumResults; ResNo++) { |
| 5193 | if (ResNo == WidenResNo) |
| 5194 | continue; |
| 5195 | EVT ResVT = N->getValueType(ResNo); |
| 5196 | if (getTypeAction(VT: ResVT) == TargetLowering::TypeWidenVector) { |
| 5197 | SetWidenedVector(Op: SDValue(N, ResNo), Result: SDValue(WidenNode, ResNo)); |
| 5198 | } else { |
| 5199 | SDLoc DL(N); |
| 5200 | SDValue ResVal = |
| 5201 | DAG.getExtractSubvector(DL, VT: ResVT, Vec: SDValue(WidenNode, ResNo), Idx: 0); |
| 5202 | ReplaceValueWith(From: SDValue(N, ResNo), To: ResVal); |
| 5203 | } |
| 5204 | } |
| 5205 | } |
| 5206 | |
| 5207 | void DAGTypeLegalizer::WidenVectorResult(SDNode *N, unsigned ResNo) { |
| 5208 | LLVM_DEBUG(dbgs() << "Widen node result " << ResNo << ": " ; N->dump(&DAG)); |
| 5209 | |
| 5210 | // See if the target wants to custom widen this node. |
| 5211 | if (CustomWidenLowerNode(N, VT: N->getValueType(ResNo))) |
| 5212 | return; |
| 5213 | |
| 5214 | SDValue Res = SDValue(); |
| 5215 | |
| 5216 | auto unrollExpandedOp = [&]() { |
| 5217 | // We're going to widen this vector op to a legal type by padding with undef |
| 5218 | // elements. If the wide vector op is eventually going to be expanded to |
| 5219 | // scalar libcalls, then unroll into scalar ops now to avoid unnecessary |
| 5220 | // libcalls on the undef elements. |
| 5221 | EVT VT = N->getValueType(ResNo: 0); |
| 5222 | EVT WideVecVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT); |
| 5223 | if (!TLI.isOperationLegalOrCustomOrPromote(Op: N->getOpcode(), VT: WideVecVT) && |
| 5224 | TLI.isOperationExpandOrLibCall(Op: N->getOpcode(), VT: VT.getScalarType())) { |
| 5225 | Res = DAG.UnrollVectorOp(N, ResNE: WideVecVT.getVectorNumElements()); |
| 5226 | if (N->getNumValues() > 1) |
| 5227 | ReplaceOtherWidenResults(N, WidenNode: Res.getNode(), WidenResNo: ResNo); |
| 5228 | return true; |
| 5229 | } |
| 5230 | return false; |
| 5231 | }; |
| 5232 | |
| 5233 | switch (N->getOpcode()) { |
| 5234 | default: |
| 5235 | #ifndef NDEBUG |
| 5236 | dbgs() << "WidenVectorResult #" << ResNo << ": " ; |
| 5237 | N->dump(&DAG); |
| 5238 | dbgs() << "\n" ; |
| 5239 | #endif |
| 5240 | report_fatal_error(reason: "Do not know how to widen the result of this operator!" ); |
| 5241 | |
| 5242 | case ISD::LOOP_DEPENDENCE_RAW_MASK: |
| 5243 | case ISD::LOOP_DEPENDENCE_WAR_MASK: |
| 5244 | Res = WidenVecRes_LOOP_DEPENDENCE_MASK(N); |
| 5245 | break; |
| 5246 | case ISD::MERGE_VALUES: Res = WidenVecRes_MERGE_VALUES(N, ResNo); break; |
| 5247 | case ISD::ADDRSPACECAST: |
| 5248 | Res = WidenVecRes_ADDRSPACECAST(N); |
| 5249 | break; |
| 5250 | case ISD::AssertZext: Res = WidenVecRes_AssertZext(N); break; |
| 5251 | case ISD::BITCAST: Res = WidenVecRes_BITCAST(N); break; |
| 5252 | case ISD::BUILD_VECTOR: Res = WidenVecRes_BUILD_VECTOR(N); break; |
| 5253 | case ISD::CONCAT_VECTORS: Res = WidenVecRes_CONCAT_VECTORS(N); break; |
| 5254 | case ISD::INSERT_SUBVECTOR: |
| 5255 | Res = WidenVecRes_INSERT_SUBVECTOR(N); |
| 5256 | break; |
| 5257 | case ISD::EXTRACT_SUBVECTOR: Res = WidenVecRes_EXTRACT_SUBVECTOR(N); break; |
| 5258 | case ISD::INSERT_VECTOR_ELT: Res = WidenVecRes_INSERT_VECTOR_ELT(N); break; |
| 5259 | case ISD::ATOMIC_LOAD: |
| 5260 | Res = WidenVecRes_ATOMIC_LOAD(N: cast<AtomicSDNode>(Val: N)); |
| 5261 | break; |
| 5262 | case ISD::LOAD: Res = WidenVecRes_LOAD(N); break; |
| 5263 | case ISD::STEP_VECTOR: |
| 5264 | case ISD::SPLAT_VECTOR: |
| 5265 | case ISD::SCALAR_TO_VECTOR: |
| 5266 | Res = WidenVecRes_ScalarOp(N); |
| 5267 | break; |
| 5268 | case ISD::SIGN_EXTEND_INREG: Res = WidenVecRes_InregOp(N); break; |
| 5269 | case ISD::VSELECT: |
| 5270 | case ISD::SELECT: |
| 5271 | case ISD::VP_SELECT: |
| 5272 | case ISD::VP_MERGE: |
| 5273 | Res = WidenVecRes_Select(N); |
| 5274 | break; |
| 5275 | case ISD::SELECT_CC: Res = WidenVecRes_SELECT_CC(N); break; |
| 5276 | case ISD::VP_SETCC: |
| 5277 | case ISD::SETCC: Res = WidenVecRes_SETCC(N); break; |
| 5278 | case ISD::POISON: |
| 5279 | case ISD::UNDEF: Res = WidenVecRes_UNDEF(N); break; |
| 5280 | case ISD::VECTOR_SHUFFLE: |
| 5281 | Res = WidenVecRes_VECTOR_SHUFFLE(N: cast<ShuffleVectorSDNode>(Val: N)); |
| 5282 | break; |
| 5283 | case ISD::VP_LOAD: |
| 5284 | Res = WidenVecRes_VP_LOAD(N: cast<VPLoadSDNode>(Val: N)); |
| 5285 | break; |
| 5286 | case ISD::VP_LOAD_FF: |
| 5287 | Res = WidenVecRes_VP_LOAD_FF(N: cast<VPLoadFFSDNode>(Val: N)); |
| 5288 | break; |
| 5289 | case ISD::EXPERIMENTAL_VP_STRIDED_LOAD: |
| 5290 | Res = WidenVecRes_VP_STRIDED_LOAD(N: cast<VPStridedLoadSDNode>(Val: N)); |
| 5291 | break; |
| 5292 | case ISD::VECTOR_COMPRESS: |
| 5293 | Res = WidenVecRes_VECTOR_COMPRESS(N); |
| 5294 | break; |
| 5295 | case ISD::MLOAD: |
| 5296 | Res = WidenVecRes_MLOAD(N: cast<MaskedLoadSDNode>(Val: N)); |
| 5297 | break; |
| 5298 | case ISD::MGATHER: |
| 5299 | Res = WidenVecRes_MGATHER(N: cast<MaskedGatherSDNode>(Val: N)); |
| 5300 | break; |
| 5301 | case ISD::VP_GATHER: |
| 5302 | Res = WidenVecRes_VP_GATHER(N: cast<VPGatherSDNode>(Val: N)); |
| 5303 | break; |
| 5304 | case ISD::VECTOR_REVERSE: |
| 5305 | Res = WidenVecRes_VECTOR_REVERSE(N); |
| 5306 | break; |
| 5307 | case ISD::GET_ACTIVE_LANE_MASK: |
| 5308 | Res = WidenVecRes_GET_ACTIVE_LANE_MASK(N); |
| 5309 | break; |
| 5310 | case ISD::VECTOR_DEINTERLEAVE: |
| 5311 | WidenVecRes_VECTOR_DEINTERLEAVE(N); |
| 5312 | break; |
| 5313 | |
| 5314 | case ISD::ADD: case ISD::VP_ADD: |
| 5315 | case ISD::AND: case ISD::VP_AND: |
| 5316 | case ISD::MUL: case ISD::VP_MUL: |
| 5317 | case ISD::MULHS: |
| 5318 | case ISD::MULHU: |
| 5319 | case ISD::ABDS: |
| 5320 | case ISD::ABDU: |
| 5321 | case ISD::OR: case ISD::VP_OR: |
| 5322 | case ISD::SUB: case ISD::VP_SUB: |
| 5323 | case ISD::XOR: case ISD::VP_XOR: |
| 5324 | case ISD::SHL: case ISD::VP_SHL: |
| 5325 | case ISD::SRA: case ISD::VP_SRA: |
| 5326 | case ISD::SRL: case ISD::VP_SRL: |
| 5327 | case ISD::CLMUL: |
| 5328 | case ISD::CLMULR: |
| 5329 | case ISD::CLMULH: |
| 5330 | case ISD::PEXT: |
| 5331 | case ISD::PDEP: |
| 5332 | case ISD::FMINNUM: |
| 5333 | case ISD::FMINNUM_IEEE: |
| 5334 | case ISD::VP_FMINNUM: |
| 5335 | case ISD::FMAXNUM: |
| 5336 | case ISD::FMAXNUM_IEEE: |
| 5337 | case ISD::VP_FMAXNUM: |
| 5338 | case ISD::FMINIMUM: |
| 5339 | case ISD::VP_FMINIMUM: |
| 5340 | case ISD::FMAXIMUM: |
| 5341 | case ISD::VP_FMAXIMUM: |
| 5342 | case ISD::FMINIMUMNUM: |
| 5343 | case ISD::FMAXIMUMNUM: |
| 5344 | case ISD::SMIN: case ISD::VP_SMIN: |
| 5345 | case ISD::SMAX: case ISD::VP_SMAX: |
| 5346 | case ISD::UMIN: case ISD::VP_UMIN: |
| 5347 | case ISD::UMAX: case ISD::VP_UMAX: |
| 5348 | case ISD::UADDSAT: case ISD::VP_UADDSAT: |
| 5349 | case ISD::SADDSAT: case ISD::VP_SADDSAT: |
| 5350 | case ISD::USUBSAT: case ISD::VP_USUBSAT: |
| 5351 | case ISD::SSUBSAT: case ISD::VP_SSUBSAT: |
| 5352 | case ISD::SSHLSAT: |
| 5353 | case ISD::USHLSAT: |
| 5354 | case ISD::ROTL: |
| 5355 | case ISD::ROTR: |
| 5356 | case ISD::AVGFLOORS: |
| 5357 | case ISD::AVGFLOORU: |
| 5358 | case ISD::AVGCEILS: |
| 5359 | case ISD::AVGCEILU: |
| 5360 | // Vector-predicated binary op widening. Note that -- unlike the |
| 5361 | // unpredicated versions -- we don't have to worry about trapping on |
| 5362 | // operations like UDIV, FADD, etc., as we pass on the original vector |
| 5363 | // length parameter. This means the widened elements containing garbage |
| 5364 | // aren't active. |
| 5365 | case ISD::VP_SDIV: |
| 5366 | case ISD::VP_UDIV: |
| 5367 | case ISD::VP_SREM: |
| 5368 | case ISD::VP_UREM: |
| 5369 | case ISD::VP_FADD: |
| 5370 | case ISD::VP_FSUB: |
| 5371 | case ISD::VP_FMUL: |
| 5372 | case ISD::VP_FDIV: |
| 5373 | case ISD::VP_FREM: |
| 5374 | case ISD::VP_FCOPYSIGN: |
| 5375 | Res = WidenVecRes_Binary(N); |
| 5376 | break; |
| 5377 | |
| 5378 | case ISD::MASKED_UDIV: |
| 5379 | case ISD::MASKED_SDIV: |
| 5380 | case ISD::MASKED_UREM: |
| 5381 | case ISD::MASKED_SREM: |
| 5382 | Res = WidenVecRes_MaskedBinary(N); |
| 5383 | break; |
| 5384 | |
| 5385 | case ISD::SCMP: |
| 5386 | case ISD::UCMP: |
| 5387 | Res = WidenVecRes_CMP(N); |
| 5388 | break; |
| 5389 | |
| 5390 | case ISD::FPOW: |
| 5391 | case ISD::FATAN2: |
| 5392 | case ISD::FREM: |
| 5393 | if (unrollExpandedOp()) |
| 5394 | break; |
| 5395 | // If the target has custom/legal support for the scalar FP intrinsic ops |
| 5396 | // (they are probably not destined to become libcalls), then widen those |
| 5397 | // like any other binary ops. |
| 5398 | [[fallthrough]]; |
| 5399 | |
| 5400 | case ISD::FADD: |
| 5401 | case ISD::FMUL: |
| 5402 | case ISD::FSUB: |
| 5403 | case ISD::FDIV: |
| 5404 | case ISD::SDIV: |
| 5405 | case ISD::UDIV: |
| 5406 | case ISD::SREM: |
| 5407 | case ISD::UREM: |
| 5408 | Res = WidenVecRes_BinaryCanTrap(N); |
| 5409 | break; |
| 5410 | |
| 5411 | case ISD::SMULFIX: |
| 5412 | case ISD::SMULFIXSAT: |
| 5413 | case ISD::UMULFIX: |
| 5414 | case ISD::UMULFIXSAT: |
| 5415 | // These are binary operations, but with an extra operand that shouldn't |
| 5416 | // be widened (the scale). |
| 5417 | Res = WidenVecRes_BinaryWithExtraScalarOp(N); |
| 5418 | break; |
| 5419 | |
| 5420 | #define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \ |
| 5421 | case ISD::STRICT_##DAGN: |
| 5422 | #include "llvm/IR/ConstrainedOps.def" |
| 5423 | Res = WidenVecRes_StrictFP(N); |
| 5424 | break; |
| 5425 | |
| 5426 | case ISD::UADDO: |
| 5427 | case ISD::SADDO: |
| 5428 | case ISD::USUBO: |
| 5429 | case ISD::SSUBO: |
| 5430 | case ISD::UMULO: |
| 5431 | case ISD::SMULO: |
| 5432 | Res = WidenVecRes_OverflowOp(N, ResNo); |
| 5433 | break; |
| 5434 | |
| 5435 | case ISD::FCOPYSIGN: |
| 5436 | Res = WidenVecRes_FCOPYSIGN(N); |
| 5437 | break; |
| 5438 | |
| 5439 | case ISD::IS_FPCLASS: |
| 5440 | case ISD::FPTRUNC_ROUND: |
| 5441 | Res = WidenVecRes_UnarySameEltsWithScalarArg(N); |
| 5442 | break; |
| 5443 | |
| 5444 | case ISD::FLDEXP: |
| 5445 | case ISD::FPOWI: |
| 5446 | if (!unrollExpandedOp()) |
| 5447 | Res = WidenVecRes_ExpOp(N); |
| 5448 | break; |
| 5449 | |
| 5450 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 5451 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 5452 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 5453 | Res = WidenVecRes_EXTEND_VECTOR_INREG(N); |
| 5454 | break; |
| 5455 | |
| 5456 | case ISD::ANY_EXTEND: |
| 5457 | case ISD::FP_EXTEND: |
| 5458 | case ISD::VP_FP_EXTEND: |
| 5459 | case ISD::FP_ROUND: |
| 5460 | case ISD::VP_FP_ROUND: |
| 5461 | case ISD::FP_TO_SINT: |
| 5462 | case ISD::VP_FP_TO_SINT: |
| 5463 | case ISD::FP_TO_UINT: |
| 5464 | case ISD::VP_FP_TO_UINT: |
| 5465 | case ISD::SIGN_EXTEND: |
| 5466 | case ISD::VP_SIGN_EXTEND: |
| 5467 | case ISD::SINT_TO_FP: |
| 5468 | case ISD::VP_SINT_TO_FP: |
| 5469 | case ISD::VP_TRUNCATE: |
| 5470 | case ISD::TRUNCATE: |
| 5471 | case ISD::UINT_TO_FP: |
| 5472 | case ISD::VP_UINT_TO_FP: |
| 5473 | case ISD::ZERO_EXTEND: |
| 5474 | case ISD::VP_ZERO_EXTEND: |
| 5475 | case ISD::CONVERT_FROM_ARBITRARY_FP: |
| 5476 | case ISD::CONVERT_TO_ARBITRARY_FP: |
| 5477 | Res = WidenVecRes_Convert(N); |
| 5478 | break; |
| 5479 | |
| 5480 | case ISD::FP_TO_SINT_SAT: |
| 5481 | case ISD::FP_TO_UINT_SAT: |
| 5482 | Res = WidenVecRes_FP_TO_XINT_SAT(N); |
| 5483 | break; |
| 5484 | |
| 5485 | case ISD::LRINT: |
| 5486 | case ISD::LLRINT: |
| 5487 | case ISD::VP_LRINT: |
| 5488 | case ISD::VP_LLRINT: |
| 5489 | case ISD::LROUND: |
| 5490 | case ISD::LLROUND: |
| 5491 | Res = WidenVecRes_XROUND(N); |
| 5492 | break; |
| 5493 | |
| 5494 | case ISD::FACOS: |
| 5495 | case ISD::FASIN: |
| 5496 | case ISD::FATAN: |
| 5497 | case ISD::FCEIL: |
| 5498 | case ISD::FCOS: |
| 5499 | case ISD::FCOSH: |
| 5500 | case ISD::FEXP: |
| 5501 | case ISD::FEXP2: |
| 5502 | case ISD::FEXP10: |
| 5503 | case ISD::FFLOOR: |
| 5504 | case ISD::FLOG: |
| 5505 | case ISD::FLOG10: |
| 5506 | case ISD::FLOG2: |
| 5507 | case ISD::FNEARBYINT: |
| 5508 | case ISD::FRINT: |
| 5509 | case ISD::FROUND: |
| 5510 | case ISD::FROUNDEVEN: |
| 5511 | case ISD::FSIN: |
| 5512 | case ISD::FSINH: |
| 5513 | case ISD::FSQRT: |
| 5514 | case ISD::FTAN: |
| 5515 | case ISD::FTANH: |
| 5516 | case ISD::FTRUNC: |
| 5517 | if (unrollExpandedOp()) |
| 5518 | break; |
| 5519 | // If the target has custom/legal support for the scalar FP intrinsic ops |
| 5520 | // (they are probably not destined to become libcalls), then widen those |
| 5521 | // like any other unary ops. |
| 5522 | [[fallthrough]]; |
| 5523 | |
| 5524 | case ISD::ABS: |
| 5525 | case ISD::ABS_MIN_POISON: |
| 5526 | case ISD::VP_ABS: |
| 5527 | case ISD::BITREVERSE: |
| 5528 | case ISD::VP_BITREVERSE: |
| 5529 | case ISD::BSWAP: |
| 5530 | case ISD::VP_BSWAP: |
| 5531 | case ISD::CTLZ: |
| 5532 | case ISD::VP_CTLZ: |
| 5533 | case ISD::CTLZ_ZERO_POISON: |
| 5534 | case ISD::VP_CTLZ_ZERO_POISON: |
| 5535 | case ISD::CTPOP: |
| 5536 | case ISD::VP_CTPOP: |
| 5537 | case ISD::CTTZ: |
| 5538 | case ISD::VP_CTTZ: |
| 5539 | case ISD::CTTZ_ZERO_POISON: |
| 5540 | case ISD::VP_CTTZ_ZERO_POISON: |
| 5541 | case ISD::FNEG: case ISD::VP_FNEG: |
| 5542 | case ISD::FABS: case ISD::VP_FABS: |
| 5543 | case ISD::VP_SQRT: |
| 5544 | case ISD::VP_FCEIL: |
| 5545 | case ISD::VP_FFLOOR: |
| 5546 | case ISD::VP_FRINT: |
| 5547 | case ISD::VP_FNEARBYINT: |
| 5548 | case ISD::VP_FROUND: |
| 5549 | case ISD::VP_FROUNDEVEN: |
| 5550 | case ISD::VP_FROUNDTOZERO: |
| 5551 | case ISD::FREEZE: |
| 5552 | case ISD::ARITH_FENCE: |
| 5553 | case ISD::FCANONICALIZE: |
| 5554 | case ISD::AssertNoFPClass: |
| 5555 | Res = WidenVecRes_Unary(N); |
| 5556 | break; |
| 5557 | case ISD::FMA: case ISD::VP_FMA: |
| 5558 | case ISD::FSHL: |
| 5559 | case ISD::VP_FSHL: |
| 5560 | case ISD::FSHR: |
| 5561 | case ISD::VP_FSHR: |
| 5562 | Res = WidenVecRes_Ternary(N); |
| 5563 | break; |
| 5564 | case ISD::FMODF: |
| 5565 | case ISD::FFREXP: |
| 5566 | case ISD::FSINCOS: |
| 5567 | case ISD::FSINCOSPI: { |
| 5568 | if (!unrollExpandedOp()) |
| 5569 | Res = WidenVecRes_UnaryOpWithTwoResults(N, ResNo); |
| 5570 | break; |
| 5571 | } |
| 5572 | } |
| 5573 | |
| 5574 | // If Res is null, the sub-method took care of registering the result. |
| 5575 | if (Res.getNode()) |
| 5576 | SetWidenedVector(Op: SDValue(N, ResNo), Result: Res); |
| 5577 | } |
| 5578 | |
| 5579 | SDValue DAGTypeLegalizer::WidenVecRes_Ternary(SDNode *N) { |
| 5580 | // Ternary op widening. |
| 5581 | SDLoc dl(N); |
| 5582 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 5583 | SDValue InOp1 = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 5584 | SDValue InOp2 = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 5585 | SDValue InOp3 = GetWidenedVector(Op: N->getOperand(Num: 2)); |
| 5586 | if (N->getNumOperands() == 3) |
| 5587 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: WidenVT, N1: InOp1, N2: InOp2, N3: InOp3); |
| 5588 | |
| 5589 | assert(N->getNumOperands() == 5 && "Unexpected number of operands!" ); |
| 5590 | assert(N->isVPOpcode() && "Expected VP opcode" ); |
| 5591 | |
| 5592 | SDValue Mask = |
| 5593 | GetWidenedMask(Mask: N->getOperand(Num: 3), EC: WidenVT.getVectorElementCount()); |
| 5594 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: WidenVT, |
| 5595 | Ops: {InOp1, InOp2, InOp3, Mask, N->getOperand(Num: 4)}); |
| 5596 | } |
| 5597 | |
| 5598 | SDValue DAGTypeLegalizer::WidenVecRes_Binary(SDNode *N) { |
| 5599 | // Binary op widening. |
| 5600 | SDLoc dl(N); |
| 5601 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 5602 | SDValue InOp1 = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 5603 | SDValue InOp2 = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 5604 | if (N->getNumOperands() == 2) |
| 5605 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: WidenVT, N1: InOp1, N2: InOp2, |
| 5606 | Flags: N->getFlags()); |
| 5607 | |
| 5608 | assert(N->getNumOperands() == 4 && "Unexpected number of operands!" ); |
| 5609 | assert(N->isVPOpcode() && "Expected VP opcode" ); |
| 5610 | |
| 5611 | SDValue Mask = |
| 5612 | GetWidenedMask(Mask: N->getOperand(Num: 2), EC: WidenVT.getVectorElementCount()); |
| 5613 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: WidenVT, |
| 5614 | Ops: {InOp1, InOp2, Mask, N->getOperand(Num: 3)}, Flags: N->getFlags()); |
| 5615 | } |
| 5616 | |
| 5617 | SDValue DAGTypeLegalizer::WidenVecRes_MaskedBinary(SDNode *N) { |
| 5618 | SDLoc dl(N); |
| 5619 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 5620 | SDValue InOp1 = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 5621 | SDValue InOp2 = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 5622 | SDValue Mask = N->getOperand(Num: 2); |
| 5623 | EVT WideMaskVT = WidenVT.changeVectorElementType( |
| 5624 | Context&: *DAG.getContext(), EltVT: Mask.getValueType().getVectorElementType()); |
| 5625 | Mask = ModifyToType(InOp: Mask, NVT: WideMaskVT, /*FillWithZeros=*/FillWithZeroes: true); |
| 5626 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: WidenVT, N1: InOp1, N2: InOp2, N3: Mask, |
| 5627 | Flags: N->getFlags()); |
| 5628 | } |
| 5629 | |
| 5630 | SDValue DAGTypeLegalizer::WidenVecRes_CMP(SDNode *N) { |
| 5631 | LLVMContext &Ctxt = *DAG.getContext(); |
| 5632 | SDLoc dl(N); |
| 5633 | |
| 5634 | SDValue LHS = N->getOperand(Num: 0); |
| 5635 | SDValue RHS = N->getOperand(Num: 1); |
| 5636 | EVT OpVT = LHS.getValueType(); |
| 5637 | if (getTypeAction(VT: OpVT) == TargetLowering::TypeWidenVector) { |
| 5638 | LHS = GetWidenedVector(Op: LHS); |
| 5639 | RHS = GetWidenedVector(Op: RHS); |
| 5640 | OpVT = LHS.getValueType(); |
| 5641 | } |
| 5642 | |
| 5643 | EVT WidenResVT = TLI.getTypeToTransformTo(Context&: Ctxt, VT: N->getValueType(ResNo: 0)); |
| 5644 | ElementCount WidenResEC = WidenResVT.getVectorElementCount(); |
| 5645 | if (WidenResEC == OpVT.getVectorElementCount()) { |
| 5646 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: WidenResVT, N1: LHS, N2: RHS); |
| 5647 | } |
| 5648 | |
| 5649 | return DAG.UnrollVectorOp(N, ResNE: WidenResVT.getVectorNumElements()); |
| 5650 | } |
| 5651 | |
| 5652 | SDValue DAGTypeLegalizer::(SDNode *N) { |
| 5653 | // Binary op widening, but with an extra operand that shouldn't be widened. |
| 5654 | SDLoc dl(N); |
| 5655 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 5656 | SDValue InOp1 = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 5657 | SDValue InOp2 = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 5658 | SDValue InOp3 = N->getOperand(Num: 2); |
| 5659 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: WidenVT, N1: InOp1, N2: InOp2, N3: InOp3, |
| 5660 | Flags: N->getFlags()); |
| 5661 | } |
| 5662 | |
| 5663 | // Given a vector of operations that have been broken up to widen, see |
| 5664 | // if we can collect them together into the next widest legal VT. This |
| 5665 | // implementation is trap-safe. |
| 5666 | static SDValue CollectOpsToWiden(SelectionDAG &DAG, const TargetLowering &TLI, |
| 5667 | SmallVectorImpl<SDValue> &ConcatOps, |
| 5668 | unsigned ConcatEnd, EVT VT, EVT MaxVT, |
| 5669 | EVT WidenVT) { |
| 5670 | // Check to see if we have a single operation with the widen type. |
| 5671 | if (ConcatEnd == 1) { |
| 5672 | VT = ConcatOps[0].getValueType(); |
| 5673 | if (VT == WidenVT) |
| 5674 | return ConcatOps[0]; |
| 5675 | } |
| 5676 | |
| 5677 | SDLoc dl(ConcatOps[0]); |
| 5678 | EVT WidenEltVT = WidenVT.getVectorElementType(); |
| 5679 | |
| 5680 | // while (Some element of ConcatOps is not of type MaxVT) { |
| 5681 | // From the end of ConcatOps, collect elements of the same type and put |
| 5682 | // them into an op of the next larger supported type |
| 5683 | // } |
| 5684 | while (ConcatOps[ConcatEnd-1].getValueType() != MaxVT) { |
| 5685 | int Idx = ConcatEnd - 1; |
| 5686 | VT = ConcatOps[Idx--].getValueType(); |
| 5687 | while (Idx >= 0 && ConcatOps[Idx].getValueType() == VT) |
| 5688 | Idx--; |
| 5689 | |
| 5690 | int NextSize = VT.isVector() ? VT.getVectorNumElements() : 1; |
| 5691 | EVT NextVT; |
| 5692 | do { |
| 5693 | NextSize *= 2; |
| 5694 | NextVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: WidenEltVT, NumElements: NextSize); |
| 5695 | } while (!TLI.isTypeLegal(VT: NextVT)); |
| 5696 | |
| 5697 | if (!VT.isVector()) { |
| 5698 | // Scalar type, create an INSERT_VECTOR_ELEMENT of type NextVT |
| 5699 | SDValue VecOp = DAG.getPOISON(VT: NextVT); |
| 5700 | unsigned NumToInsert = ConcatEnd - Idx - 1; |
| 5701 | for (unsigned i = 0, OpIdx = Idx + 1; i < NumToInsert; i++, OpIdx++) |
| 5702 | VecOp = DAG.getInsertVectorElt(DL: dl, Vec: VecOp, Elt: ConcatOps[OpIdx], Idx: i); |
| 5703 | ConcatOps[Idx+1] = VecOp; |
| 5704 | ConcatEnd = Idx + 2; |
| 5705 | } else { |
| 5706 | // Vector type, create a CONCAT_VECTORS of type NextVT |
| 5707 | SDValue undefVec = DAG.getPOISON(VT); |
| 5708 | unsigned OpsToConcat = NextSize/VT.getVectorNumElements(); |
| 5709 | SmallVector<SDValue, 16> SubConcatOps(OpsToConcat); |
| 5710 | unsigned RealVals = ConcatEnd - Idx - 1; |
| 5711 | unsigned SubConcatEnd = 0; |
| 5712 | unsigned SubConcatIdx = Idx + 1; |
| 5713 | while (SubConcatEnd < RealVals) |
| 5714 | SubConcatOps[SubConcatEnd++] = ConcatOps[++Idx]; |
| 5715 | while (SubConcatEnd < OpsToConcat) |
| 5716 | SubConcatOps[SubConcatEnd++] = undefVec; |
| 5717 | ConcatOps[SubConcatIdx] = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, |
| 5718 | VT: NextVT, Ops: SubConcatOps); |
| 5719 | ConcatEnd = SubConcatIdx + 1; |
| 5720 | } |
| 5721 | } |
| 5722 | |
| 5723 | // Check to see if we have a single operation with the widen type. |
| 5724 | if (ConcatEnd == 1) { |
| 5725 | VT = ConcatOps[0].getValueType(); |
| 5726 | if (VT == WidenVT) |
| 5727 | return ConcatOps[0]; |
| 5728 | } |
| 5729 | |
| 5730 | // add undefs of size MaxVT until ConcatOps grows to length of WidenVT |
| 5731 | unsigned NumOps = WidenVT.getVectorNumElements()/MaxVT.getVectorNumElements(); |
| 5732 | if (NumOps != ConcatEnd ) { |
| 5733 | SDValue UndefVal = DAG.getPOISON(VT: MaxVT); |
| 5734 | for (unsigned j = ConcatEnd; j < NumOps; ++j) |
| 5735 | ConcatOps[j] = UndefVal; |
| 5736 | } |
| 5737 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: WidenVT, |
| 5738 | Ops: ArrayRef(ConcatOps.data(), NumOps)); |
| 5739 | } |
| 5740 | |
| 5741 | SDValue DAGTypeLegalizer::WidenVecRes_BinaryCanTrap(SDNode *N) { |
| 5742 | // Binary op widening for operations that can trap. |
| 5743 | unsigned Opcode = N->getOpcode(); |
| 5744 | SDLoc dl(N); |
| 5745 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 5746 | EVT WidenEltVT = WidenVT.getVectorElementType(); |
| 5747 | EVT VT = WidenVT; |
| 5748 | unsigned NumElts = VT.getVectorMinNumElements(); |
| 5749 | const SDNodeFlags Flags = N->getFlags(); |
| 5750 | while (!TLI.isTypeLegal(VT) && NumElts != 1) { |
| 5751 | NumElts = NumElts / 2; |
| 5752 | VT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: WidenEltVT, NumElements: NumElts); |
| 5753 | } |
| 5754 | |
| 5755 | if (NumElts != 1 && !TLI.canOpTrap(Op: N->getOpcode(), VT)) { |
| 5756 | // Operation doesn't trap so just widen as normal. |
| 5757 | SDValue InOp1 = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 5758 | SDValue InOp2 = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 5759 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: WidenVT, N1: InOp1, N2: InOp2, Flags); |
| 5760 | } |
| 5761 | |
| 5762 | // Generate a vp.op if it is custom/legal for the target. This avoids need |
| 5763 | // to split and tile the subvectors (below), because the inactive lanes can |
| 5764 | // simply be disabled. To avoid possible recursion, only do this if the |
| 5765 | // widened mask type is legal. |
| 5766 | if (auto VPOpcode = ISD::getVPForBaseOpcode(Opcode); |
| 5767 | VPOpcode && TLI.isOperationLegalOrCustom(Op: *VPOpcode, VT: WidenVT)) { |
| 5768 | if (EVT WideMaskVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: MVT::i1, |
| 5769 | EC: WidenVT.getVectorElementCount()); |
| 5770 | TLI.isTypeLegal(VT: WideMaskVT)) { |
| 5771 | SDValue InOp1 = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 5772 | SDValue InOp2 = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 5773 | SDValue Mask = DAG.getAllOnesConstant(DL: dl, VT: WideMaskVT); |
| 5774 | SDValue EVL = |
| 5775 | DAG.getElementCount(DL: dl, VT: TLI.getVPExplicitVectorLengthTy(), |
| 5776 | EC: N->getValueType(ResNo: 0).getVectorElementCount()); |
| 5777 | return DAG.getNode(Opcode: *VPOpcode, DL: dl, VT: WidenVT, N1: InOp1, N2: InOp2, N3: Mask, N4: EVL, |
| 5778 | Flags); |
| 5779 | } |
| 5780 | } |
| 5781 | |
| 5782 | // FIXME: Improve support for scalable vectors. |
| 5783 | assert(!VT.isScalableVector() && "Scalable vectors not handled yet." ); |
| 5784 | |
| 5785 | // No legal vector version so unroll the vector operation and then widen. |
| 5786 | if (NumElts == 1) |
| 5787 | return DAG.UnrollVectorOp(N, ResNE: WidenVT.getVectorNumElements()); |
| 5788 | |
| 5789 | // Since the operation can trap, apply operation on the original vector. |
| 5790 | EVT MaxVT = VT; |
| 5791 | SDValue InOp1 = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 5792 | SDValue InOp2 = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 5793 | unsigned CurNumElts = N->getValueType(ResNo: 0).getVectorNumElements(); |
| 5794 | |
| 5795 | SmallVector<SDValue, 16> ConcatOps(CurNumElts); |
| 5796 | unsigned ConcatEnd = 0; // Current ConcatOps index. |
| 5797 | int Idx = 0; // Current Idx into input vectors. |
| 5798 | |
| 5799 | // NumElts := greatest legal vector size (at most WidenVT) |
| 5800 | // while (orig. vector has unhandled elements) { |
| 5801 | // take munches of size NumElts from the beginning and add to ConcatOps |
| 5802 | // NumElts := next smaller supported vector size or 1 |
| 5803 | // } |
| 5804 | while (CurNumElts != 0) { |
| 5805 | while (CurNumElts >= NumElts) { |
| 5806 | SDValue EOp1 = DAG.getExtractSubvector(DL: dl, VT, Vec: InOp1, Idx); |
| 5807 | SDValue EOp2 = DAG.getExtractSubvector(DL: dl, VT, Vec: InOp2, Idx); |
| 5808 | ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, DL: dl, VT, N1: EOp1, N2: EOp2, Flags); |
| 5809 | Idx += NumElts; |
| 5810 | CurNumElts -= NumElts; |
| 5811 | } |
| 5812 | do { |
| 5813 | NumElts = NumElts / 2; |
| 5814 | VT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: WidenEltVT, NumElements: NumElts); |
| 5815 | } while (!TLI.isTypeLegal(VT) && NumElts != 1); |
| 5816 | |
| 5817 | if (NumElts == 1) { |
| 5818 | for (unsigned i = 0; i != CurNumElts; ++i, ++Idx) { |
| 5819 | SDValue EOp1 = DAG.getExtractVectorElt(DL: dl, VT: WidenEltVT, Vec: InOp1, Idx); |
| 5820 | SDValue EOp2 = DAG.getExtractVectorElt(DL: dl, VT: WidenEltVT, Vec: InOp2, Idx); |
| 5821 | ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, DL: dl, VT: WidenEltVT, |
| 5822 | N1: EOp1, N2: EOp2, Flags); |
| 5823 | } |
| 5824 | CurNumElts = 0; |
| 5825 | } |
| 5826 | } |
| 5827 | |
| 5828 | return CollectOpsToWiden(DAG, TLI, ConcatOps, ConcatEnd, VT, MaxVT, WidenVT); |
| 5829 | } |
| 5830 | |
| 5831 | SDValue DAGTypeLegalizer::WidenVecRes_StrictFP(SDNode *N) { |
| 5832 | switch (N->getOpcode()) { |
| 5833 | case ISD::STRICT_FSETCC: |
| 5834 | case ISD::STRICT_FSETCCS: |
| 5835 | return WidenVecRes_STRICT_FSETCC(N); |
| 5836 | case ISD::STRICT_FP_EXTEND: |
| 5837 | case ISD::STRICT_FP_ROUND: |
| 5838 | case ISD::STRICT_FP_TO_SINT: |
| 5839 | case ISD::STRICT_FP_TO_UINT: |
| 5840 | case ISD::STRICT_SINT_TO_FP: |
| 5841 | case ISD::STRICT_UINT_TO_FP: |
| 5842 | return WidenVecRes_Convert_StrictFP(N); |
| 5843 | default: |
| 5844 | break; |
| 5845 | } |
| 5846 | |
| 5847 | // StrictFP op widening for operations that can trap. |
| 5848 | unsigned NumOpers = N->getNumOperands(); |
| 5849 | unsigned Opcode = N->getOpcode(); |
| 5850 | SDLoc dl(N); |
| 5851 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 5852 | EVT WidenEltVT = WidenVT.getVectorElementType(); |
| 5853 | EVT VT = WidenVT; |
| 5854 | unsigned NumElts = VT.getVectorNumElements(); |
| 5855 | while (!TLI.isTypeLegal(VT) && NumElts != 1) { |
| 5856 | NumElts = NumElts / 2; |
| 5857 | VT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: WidenEltVT, NumElements: NumElts); |
| 5858 | } |
| 5859 | |
| 5860 | // No legal vector version so unroll the vector operation and then widen. |
| 5861 | if (NumElts == 1) |
| 5862 | return UnrollVectorOp_StrictFP(N, ResNE: WidenVT.getVectorNumElements()); |
| 5863 | |
| 5864 | // Since the operation can trap, apply operation on the original vector. |
| 5865 | EVT MaxVT = VT; |
| 5866 | SmallVector<SDValue, 4> InOps; |
| 5867 | unsigned CurNumElts = N->getValueType(ResNo: 0).getVectorNumElements(); |
| 5868 | |
| 5869 | SmallVector<SDValue, 16> ConcatOps(CurNumElts); |
| 5870 | SmallVector<SDValue, 16> Chains; |
| 5871 | unsigned ConcatEnd = 0; // Current ConcatOps index. |
| 5872 | int Idx = 0; // Current Idx into input vectors. |
| 5873 | |
| 5874 | // The Chain is the first operand. |
| 5875 | InOps.push_back(Elt: N->getOperand(Num: 0)); |
| 5876 | |
| 5877 | // Now process the remaining operands. |
| 5878 | for (unsigned i = 1; i < NumOpers; ++i) { |
| 5879 | SDValue Oper = N->getOperand(Num: i); |
| 5880 | |
| 5881 | EVT OpVT = Oper.getValueType(); |
| 5882 | if (OpVT.isVector()) { |
| 5883 | if (getTypeAction(VT: OpVT) == TargetLowering::TypeWidenVector) |
| 5884 | Oper = GetWidenedVector(Op: Oper); |
| 5885 | else { |
| 5886 | EVT WideOpVT = |
| 5887 | EVT::getVectorVT(Context&: *DAG.getContext(), VT: OpVT.getVectorElementType(), |
| 5888 | EC: WidenVT.getVectorElementCount()); |
| 5889 | Oper = DAG.getNode(Opcode: ISD::INSERT_SUBVECTOR, DL: dl, VT: WideOpVT, |
| 5890 | N1: DAG.getPOISON(VT: WideOpVT), N2: Oper, |
| 5891 | N3: DAG.getVectorIdxConstant(Val: 0, DL: dl)); |
| 5892 | } |
| 5893 | } |
| 5894 | |
| 5895 | InOps.push_back(Elt: Oper); |
| 5896 | } |
| 5897 | |
| 5898 | // NumElts := greatest legal vector size (at most WidenVT) |
| 5899 | // while (orig. vector has unhandled elements) { |
| 5900 | // take munches of size NumElts from the beginning and add to ConcatOps |
| 5901 | // NumElts := next smaller supported vector size or 1 |
| 5902 | // } |
| 5903 | while (CurNumElts != 0) { |
| 5904 | while (CurNumElts >= NumElts) { |
| 5905 | SmallVector<SDValue, 4> EOps; |
| 5906 | |
| 5907 | for (unsigned i = 0; i < NumOpers; ++i) { |
| 5908 | SDValue Op = InOps[i]; |
| 5909 | |
| 5910 | EVT OpVT = Op.getValueType(); |
| 5911 | if (OpVT.isVector()) { |
| 5912 | EVT = |
| 5913 | EVT::getVectorVT(Context&: *DAG.getContext(), VT: OpVT.getVectorElementType(), |
| 5914 | EC: VT.getVectorElementCount()); |
| 5915 | Op = DAG.getExtractSubvector(DL: dl, VT: OpExtractVT, Vec: Op, Idx); |
| 5916 | } |
| 5917 | |
| 5918 | EOps.push_back(Elt: Op); |
| 5919 | } |
| 5920 | |
| 5921 | EVT OperVT[] = {VT, MVT::Other}; |
| 5922 | SDValue Oper = DAG.getNode(Opcode, DL: dl, ResultTys: OperVT, Ops: EOps); |
| 5923 | ConcatOps[ConcatEnd++] = Oper; |
| 5924 | Chains.push_back(Elt: Oper.getValue(R: 1)); |
| 5925 | Idx += NumElts; |
| 5926 | CurNumElts -= NumElts; |
| 5927 | } |
| 5928 | do { |
| 5929 | NumElts = NumElts / 2; |
| 5930 | VT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: WidenEltVT, NumElements: NumElts); |
| 5931 | } while (!TLI.isTypeLegal(VT) && NumElts != 1); |
| 5932 | |
| 5933 | if (NumElts == 1) { |
| 5934 | for (unsigned i = 0; i != CurNumElts; ++i, ++Idx) { |
| 5935 | SmallVector<SDValue, 4> EOps; |
| 5936 | |
| 5937 | for (unsigned i = 0; i < NumOpers; ++i) { |
| 5938 | SDValue Op = InOps[i]; |
| 5939 | |
| 5940 | EVT OpVT = Op.getValueType(); |
| 5941 | if (OpVT.isVector()) |
| 5942 | Op = DAG.getExtractVectorElt(DL: dl, VT: OpVT.getVectorElementType(), Vec: Op, |
| 5943 | Idx); |
| 5944 | |
| 5945 | EOps.push_back(Elt: Op); |
| 5946 | } |
| 5947 | |
| 5948 | EVT WidenVT[] = {WidenEltVT, MVT::Other}; |
| 5949 | SDValue Oper = DAG.getNode(Opcode, DL: dl, ResultTys: WidenVT, Ops: EOps); |
| 5950 | ConcatOps[ConcatEnd++] = Oper; |
| 5951 | Chains.push_back(Elt: Oper.getValue(R: 1)); |
| 5952 | } |
| 5953 | CurNumElts = 0; |
| 5954 | } |
| 5955 | } |
| 5956 | |
| 5957 | // Build a factor node to remember all the Ops that have been created. |
| 5958 | SDValue NewChain; |
| 5959 | if (Chains.size() == 1) |
| 5960 | NewChain = Chains[0]; |
| 5961 | else |
| 5962 | NewChain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: Chains); |
| 5963 | ReplaceValueWith(From: SDValue(N, 1), To: NewChain); |
| 5964 | |
| 5965 | return CollectOpsToWiden(DAG, TLI, ConcatOps, ConcatEnd, VT, MaxVT, WidenVT); |
| 5966 | } |
| 5967 | |
| 5968 | SDValue DAGTypeLegalizer::WidenVecRes_OverflowOp(SDNode *N, unsigned ResNo) { |
| 5969 | SDLoc DL(N); |
| 5970 | EVT ResVT = N->getValueType(ResNo: 0); |
| 5971 | EVT OvVT = N->getValueType(ResNo: 1); |
| 5972 | EVT WideResVT, WideOvVT; |
| 5973 | SDValue WideLHS, WideRHS; |
| 5974 | |
| 5975 | // TODO: This might result in a widen/split loop. |
| 5976 | if (ResNo == 0) { |
| 5977 | WideResVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: ResVT); |
| 5978 | WideOvVT = EVT::getVectorVT( |
| 5979 | Context&: *DAG.getContext(), VT: OvVT.getVectorElementType(), |
| 5980 | NumElements: WideResVT.getVectorNumElements()); |
| 5981 | |
| 5982 | WideLHS = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 5983 | WideRHS = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 5984 | } else { |
| 5985 | WideOvVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: OvVT); |
| 5986 | WideResVT = EVT::getVectorVT( |
| 5987 | Context&: *DAG.getContext(), VT: ResVT.getVectorElementType(), |
| 5988 | NumElements: WideOvVT.getVectorNumElements()); |
| 5989 | |
| 5990 | SDValue Zero = DAG.getVectorIdxConstant(Val: 0, DL); |
| 5991 | SDValue Poison = DAG.getPOISON(VT: WideResVT); |
| 5992 | |
| 5993 | WideLHS = DAG.getNode(Opcode: ISD::INSERT_SUBVECTOR, DL, VT: WideResVT, N1: Poison, |
| 5994 | N2: N->getOperand(Num: 0), N3: Zero); |
| 5995 | WideRHS = DAG.getNode(Opcode: ISD::INSERT_SUBVECTOR, DL, VT: WideResVT, N1: Poison, |
| 5996 | N2: N->getOperand(Num: 1), N3: Zero); |
| 5997 | } |
| 5998 | |
| 5999 | SDVTList WideVTs = DAG.getVTList(VT1: WideResVT, VT2: WideOvVT); |
| 6000 | SDNode *WideNode = DAG.getNode( |
| 6001 | Opcode: N->getOpcode(), DL, VTList: WideVTs, N1: WideLHS, N2: WideRHS).getNode(); |
| 6002 | |
| 6003 | // Replace the other vector result not being explicitly widened here. |
| 6004 | unsigned OtherNo = 1 - ResNo; |
| 6005 | EVT OtherVT = N->getValueType(ResNo: OtherNo); |
| 6006 | if (getTypeAction(VT: OtherVT) == TargetLowering::TypeWidenVector) { |
| 6007 | SetWidenedVector(Op: SDValue(N, OtherNo), Result: SDValue(WideNode, OtherNo)); |
| 6008 | } else { |
| 6009 | SDValue Zero = DAG.getVectorIdxConstant(Val: 0, DL); |
| 6010 | SDValue OtherVal = DAG.getNode( |
| 6011 | Opcode: ISD::EXTRACT_SUBVECTOR, DL, VT: OtherVT, N1: SDValue(WideNode, OtherNo), N2: Zero); |
| 6012 | ReplaceValueWith(From: SDValue(N, OtherNo), To: OtherVal); |
| 6013 | } |
| 6014 | |
| 6015 | return SDValue(WideNode, ResNo); |
| 6016 | } |
| 6017 | |
| 6018 | SDValue DAGTypeLegalizer::WidenVecRes_Convert(SDNode *N) { |
| 6019 | LLVMContext &Ctx = *DAG.getContext(); |
| 6020 | SDValue InOp = N->getOperand(Num: 0); |
| 6021 | SDLoc DL(N); |
| 6022 | |
| 6023 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: Ctx, VT: N->getValueType(ResNo: 0)); |
| 6024 | ElementCount WidenEC = WidenVT.getVectorElementCount(); |
| 6025 | |
| 6026 | EVT InVT = InOp.getValueType(); |
| 6027 | |
| 6028 | unsigned Opcode = N->getOpcode(); |
| 6029 | const SDNodeFlags Flags = N->getFlags(); |
| 6030 | |
| 6031 | // Handle the case of ZERO_EXTEND where the promoted InVT element size does |
| 6032 | // not equal that of WidenVT. |
| 6033 | if (N->getOpcode() == ISD::ZERO_EXTEND && |
| 6034 | getTypeAction(VT: InVT) == TargetLowering::TypePromoteInteger && |
| 6035 | TLI.getTypeToTransformTo(Context&: Ctx, VT: InVT).getScalarSizeInBits() != |
| 6036 | WidenVT.getScalarSizeInBits()) { |
| 6037 | InOp = ZExtPromotedInteger(Op: InOp); |
| 6038 | InVT = InOp.getValueType(); |
| 6039 | if (WidenVT.getScalarSizeInBits() < InVT.getScalarSizeInBits()) |
| 6040 | Opcode = ISD::TRUNCATE; |
| 6041 | } |
| 6042 | |
| 6043 | EVT InEltVT = InVT.getVectorElementType(); |
| 6044 | EVT InWidenVT = EVT::getVectorVT(Context&: Ctx, VT: InEltVT, EC: WidenEC); |
| 6045 | ElementCount InVTEC = InVT.getVectorElementCount(); |
| 6046 | |
| 6047 | // Helper to build node with all scalar trailing operands. |
| 6048 | auto MakeConvertNode = [&](EVT VT, SDValue Op) -> SDValue { |
| 6049 | if (N->getNumOperands() == 1) |
| 6050 | return DAG.getNode(Opcode, DL, VT, Operand: Op, Flags); |
| 6051 | if (Opcode == ISD::CONVERT_TO_ARBITRARY_FP) |
| 6052 | return DAG.getNode(Opcode, DL, VT, N1: Op, N2: N->getOperand(Num: 1), N3: N->getOperand(Num: 2), |
| 6053 | N4: N->getOperand(Num: 3), Flags); |
| 6054 | return DAG.getNode(Opcode, DL, VT, N1: Op, N2: N->getOperand(Num: 1), Flags); |
| 6055 | }; |
| 6056 | |
| 6057 | if (getTypeAction(VT: InVT) == TargetLowering::TypeWidenVector) { |
| 6058 | InOp = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 6059 | InVT = InOp.getValueType(); |
| 6060 | InVTEC = InVT.getVectorElementCount(); |
| 6061 | if (InVTEC == WidenEC) { |
| 6062 | if (N->getNumOperands() == 3 && N->isVPOpcode()) { |
| 6063 | SDValue Mask = |
| 6064 | GetWidenedMask(Mask: N->getOperand(Num: 1), EC: WidenVT.getVectorElementCount()); |
| 6065 | return DAG.getNode(Opcode, DL, VT: WidenVT, N1: InOp, N2: Mask, N3: N->getOperand(Num: 2)); |
| 6066 | } |
| 6067 | return MakeConvertNode(WidenVT, InOp); |
| 6068 | } |
| 6069 | if (WidenVT.getSizeInBits() == InVT.getSizeInBits()) { |
| 6070 | // If both input and result vector types are of same width, extend |
| 6071 | // operations should be done with SIGN/ZERO_EXTEND_VECTOR_INREG, which |
| 6072 | // accepts fewer elements in the result than in the input. |
| 6073 | if (Opcode == ISD::ANY_EXTEND) |
| 6074 | return DAG.getNode(Opcode: ISD::ANY_EXTEND_VECTOR_INREG, DL, VT: WidenVT, Operand: InOp); |
| 6075 | if (Opcode == ISD::SIGN_EXTEND) |
| 6076 | return DAG.getNode(Opcode: ISD::SIGN_EXTEND_VECTOR_INREG, DL, VT: WidenVT, Operand: InOp); |
| 6077 | if (Opcode == ISD::ZERO_EXTEND) |
| 6078 | return DAG.getNode(Opcode: ISD::ZERO_EXTEND_VECTOR_INREG, DL, VT: WidenVT, Operand: InOp); |
| 6079 | } |
| 6080 | |
| 6081 | // For TRUNCATE, try to widen using the legal EC of the input type instead |
| 6082 | // if the legalisation action for that intermediate type is not widening. |
| 6083 | // E.g. for trunc nxv1i64 -> nxv1i8 where |
| 6084 | // - nxv1i64 input gets widened to nxv2i64 |
| 6085 | // - nxv1i8 output gets widened to nxv16i8 |
| 6086 | // Then one can try widening the result to nxv2i8 (instead of going all the |
| 6087 | // way to nxv16i8) if this later allows type promotion. |
| 6088 | EVT MidResVT = |
| 6089 | EVT::getVectorVT(Context&: Ctx, VT: WidenVT.getVectorElementType(), EC: InVTEC); |
| 6090 | if (N->getOpcode() == ISD::TRUNCATE && |
| 6091 | getTypeAction(VT: MidResVT) == TargetLowering::TypePromoteInteger) { |
| 6092 | SDValue MidRes = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: MidResVT, Operand: InOp, Flags); |
| 6093 | return DAG.getInsertSubvector(DL, Vec: DAG.getPOISON(VT: WidenVT), SubVec: MidRes, Idx: 0); |
| 6094 | } |
| 6095 | } |
| 6096 | |
| 6097 | if (TLI.isTypeLegal(VT: InWidenVT)) { |
| 6098 | // Because the result and the input are different vector types, widening |
| 6099 | // the result could create a legal type but widening the input might make |
| 6100 | // it an illegal type that might lead to repeatedly splitting the input |
| 6101 | // and then widening it. To avoid this, we widen the input only if |
| 6102 | // it results in a legal type. |
| 6103 | if (WidenEC.isKnownMultipleOf(RHS: InVTEC.getKnownMinValue())) { |
| 6104 | // Widen the input and call convert on the widened input vector. |
| 6105 | unsigned NumConcat = |
| 6106 | WidenEC.getKnownMinValue() / InVTEC.getKnownMinValue(); |
| 6107 | SmallVector<SDValue, 16> Ops(NumConcat, DAG.getPOISON(VT: InVT)); |
| 6108 | Ops[0] = InOp; |
| 6109 | SDValue InVec = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: InWidenVT, Ops); |
| 6110 | return MakeConvertNode(WidenVT, InVec); |
| 6111 | } |
| 6112 | |
| 6113 | if (InVTEC.isKnownMultipleOf(RHS: WidenEC.getKnownMinValue())) { |
| 6114 | SDValue InVal = DAG.getExtractSubvector(DL, VT: InWidenVT, Vec: InOp, Idx: 0); |
| 6115 | // Extract the input and convert the shorten input vector. |
| 6116 | return MakeConvertNode(WidenVT, InVal); |
| 6117 | } |
| 6118 | } |
| 6119 | |
| 6120 | // Otherwise unroll into some nasty scalar code and rebuild the vector. |
| 6121 | EVT EltVT = WidenVT.getVectorElementType(); |
| 6122 | SmallVector<SDValue, 16> Ops(WidenEC.getFixedValue(), DAG.getPOISON(VT: EltVT)); |
| 6123 | // Use the original element count so we don't do more scalar opts than |
| 6124 | // necessary. |
| 6125 | unsigned MinElts = N->getValueType(ResNo: 0).getVectorNumElements(); |
| 6126 | for (unsigned i=0; i < MinElts; ++i) { |
| 6127 | SDValue Val = DAG.getExtractVectorElt(DL, VT: InEltVT, Vec: InOp, Idx: i); |
| 6128 | Ops[i] = MakeConvertNode(EltVT, Val); |
| 6129 | } |
| 6130 | |
| 6131 | return DAG.getBuildVector(VT: WidenVT, DL, Ops); |
| 6132 | } |
| 6133 | |
| 6134 | SDValue DAGTypeLegalizer::WidenVecRes_FP_TO_XINT_SAT(SDNode *N) { |
| 6135 | SDLoc dl(N); |
| 6136 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6137 | ElementCount WidenNumElts = WidenVT.getVectorElementCount(); |
| 6138 | |
| 6139 | SDValue Src = N->getOperand(Num: 0); |
| 6140 | EVT SrcVT = Src.getValueType(); |
| 6141 | |
| 6142 | // Also widen the input. |
| 6143 | if (getTypeAction(VT: SrcVT) == TargetLowering::TypeWidenVector) { |
| 6144 | Src = GetWidenedVector(Op: Src); |
| 6145 | SrcVT = Src.getValueType(); |
| 6146 | } |
| 6147 | |
| 6148 | // Input and output not widened to the same size, give up. |
| 6149 | if (WidenNumElts != SrcVT.getVectorElementCount()) |
| 6150 | return DAG.UnrollVectorOp(N, ResNE: WidenNumElts.getKnownMinValue()); |
| 6151 | |
| 6152 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: WidenVT, N1: Src, N2: N->getOperand(Num: 1)); |
| 6153 | } |
| 6154 | |
| 6155 | SDValue DAGTypeLegalizer::WidenVecRes_XROUND(SDNode *N) { |
| 6156 | SDLoc dl(N); |
| 6157 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6158 | ElementCount WidenNumElts = WidenVT.getVectorElementCount(); |
| 6159 | |
| 6160 | SDValue Src = N->getOperand(Num: 0); |
| 6161 | EVT SrcVT = Src.getValueType(); |
| 6162 | |
| 6163 | // Also widen the input. |
| 6164 | if (getTypeAction(VT: SrcVT) == TargetLowering::TypeWidenVector) { |
| 6165 | Src = GetWidenedVector(Op: Src); |
| 6166 | SrcVT = Src.getValueType(); |
| 6167 | } |
| 6168 | |
| 6169 | // Input and output not widened to the same size, give up. |
| 6170 | if (WidenNumElts != SrcVT.getVectorElementCount()) |
| 6171 | return DAG.UnrollVectorOp(N, ResNE: WidenNumElts.getKnownMinValue()); |
| 6172 | |
| 6173 | if (N->getNumOperands() == 1) |
| 6174 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: WidenVT, Operand: Src); |
| 6175 | |
| 6176 | assert(N->getNumOperands() == 3 && "Unexpected number of operands!" ); |
| 6177 | assert(N->isVPOpcode() && "Expected VP opcode" ); |
| 6178 | |
| 6179 | SDValue Mask = |
| 6180 | GetWidenedMask(Mask: N->getOperand(Num: 1), EC: WidenVT.getVectorElementCount()); |
| 6181 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: WidenVT, N1: Src, N2: Mask, N3: N->getOperand(Num: 2)); |
| 6182 | } |
| 6183 | |
| 6184 | SDValue DAGTypeLegalizer::WidenVecRes_Convert_StrictFP(SDNode *N) { |
| 6185 | SDValue InOp = N->getOperand(Num: 1); |
| 6186 | SDLoc DL(N); |
| 6187 | SmallVector<SDValue, 4> NewOps(N->ops()); |
| 6188 | |
| 6189 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6190 | unsigned WidenNumElts = WidenVT.getVectorNumElements(); |
| 6191 | |
| 6192 | EVT InVT = InOp.getValueType(); |
| 6193 | EVT InEltVT = InVT.getVectorElementType(); |
| 6194 | |
| 6195 | unsigned Opcode = N->getOpcode(); |
| 6196 | |
| 6197 | // FIXME: Optimizations need to be implemented here. |
| 6198 | |
| 6199 | // Otherwise unroll into some nasty scalar code and rebuild the vector. |
| 6200 | EVT EltVT = WidenVT.getVectorElementType(); |
| 6201 | std::array<EVT, 2> EltVTs = {._M_elems: {EltVT, MVT::Other}}; |
| 6202 | SmallVector<SDValue, 16> Ops(WidenNumElts, DAG.getPOISON(VT: EltVT)); |
| 6203 | SmallVector<SDValue, 32> OpChains; |
| 6204 | // Use the original element count so we don't do more scalar opts than |
| 6205 | // necessary. |
| 6206 | unsigned MinElts = N->getValueType(ResNo: 0).getVectorNumElements(); |
| 6207 | for (unsigned i=0; i < MinElts; ++i) { |
| 6208 | NewOps[1] = DAG.getExtractVectorElt(DL, VT: InEltVT, Vec: InOp, Idx: i); |
| 6209 | Ops[i] = DAG.getNode(Opcode, DL, ResultTys: EltVTs, Ops: NewOps); |
| 6210 | OpChains.push_back(Elt: Ops[i].getValue(R: 1)); |
| 6211 | } |
| 6212 | SDValue NewChain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, Ops: OpChains); |
| 6213 | ReplaceValueWith(From: SDValue(N, 1), To: NewChain); |
| 6214 | |
| 6215 | return DAG.getBuildVector(VT: WidenVT, DL, Ops); |
| 6216 | } |
| 6217 | |
| 6218 | SDValue DAGTypeLegalizer::WidenVecRes_EXTEND_VECTOR_INREG(SDNode *N) { |
| 6219 | unsigned Opcode = N->getOpcode(); |
| 6220 | SDValue InOp = N->getOperand(Num: 0); |
| 6221 | SDLoc DL(N); |
| 6222 | |
| 6223 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6224 | EVT WidenSVT = WidenVT.getVectorElementType(); |
| 6225 | unsigned WidenNumElts = WidenVT.getVectorNumElements(); |
| 6226 | |
| 6227 | EVT InVT = InOp.getValueType(); |
| 6228 | EVT InSVT = InVT.getVectorElementType(); |
| 6229 | unsigned InVTNumElts = InVT.getVectorNumElements(); |
| 6230 | |
| 6231 | if (getTypeAction(VT: InVT) == TargetLowering::TypeWidenVector) { |
| 6232 | InOp = GetWidenedVector(Op: InOp); |
| 6233 | InVT = InOp.getValueType(); |
| 6234 | if (InVT.getSizeInBits() == WidenVT.getSizeInBits()) { |
| 6235 | switch (Opcode) { |
| 6236 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 6237 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 6238 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 6239 | return DAG.getNode(Opcode, DL, VT: WidenVT, Operand: InOp); |
| 6240 | } |
| 6241 | } |
| 6242 | } |
| 6243 | |
| 6244 | // Unroll, extend the scalars and rebuild the vector. |
| 6245 | SmallVector<SDValue, 16> Ops; |
| 6246 | for (unsigned i = 0, e = std::min(a: InVTNumElts, b: WidenNumElts); i != e; ++i) { |
| 6247 | SDValue Val = DAG.getExtractVectorElt(DL, VT: InSVT, Vec: InOp, Idx: i); |
| 6248 | switch (Opcode) { |
| 6249 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 6250 | Val = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT: WidenSVT, Operand: Val); |
| 6251 | break; |
| 6252 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 6253 | Val = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL, VT: WidenSVT, Operand: Val); |
| 6254 | break; |
| 6255 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 6256 | Val = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: WidenSVT, Operand: Val); |
| 6257 | break; |
| 6258 | default: |
| 6259 | llvm_unreachable("A *_EXTEND_VECTOR_INREG node was expected" ); |
| 6260 | } |
| 6261 | Ops.push_back(Elt: Val); |
| 6262 | } |
| 6263 | |
| 6264 | while (Ops.size() != WidenNumElts) |
| 6265 | Ops.push_back(Elt: DAG.getPOISON(VT: WidenSVT)); |
| 6266 | |
| 6267 | return DAG.getBuildVector(VT: WidenVT, DL, Ops); |
| 6268 | } |
| 6269 | |
| 6270 | SDValue DAGTypeLegalizer::WidenVecRes_FCOPYSIGN(SDNode *N) { |
| 6271 | // If this is an FCOPYSIGN with same input types, we can treat it as a |
| 6272 | // normal (can trap) binary op. |
| 6273 | if (N->getOperand(Num: 0).getValueType() == N->getOperand(Num: 1).getValueType()) |
| 6274 | return WidenVecRes_BinaryCanTrap(N); |
| 6275 | |
| 6276 | // If the types are different, fall back to unrolling. |
| 6277 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6278 | return DAG.UnrollVectorOp(N, ResNE: WidenVT.getVectorNumElements()); |
| 6279 | } |
| 6280 | |
| 6281 | /// Result and first source operand are different scalar types, but must have |
| 6282 | /// the same number of elements. There is an additional control argument which |
| 6283 | /// should be passed through unchanged. |
| 6284 | SDValue DAGTypeLegalizer::WidenVecRes_UnarySameEltsWithScalarArg(SDNode *N) { |
| 6285 | SDValue FpValue = N->getOperand(Num: 0); |
| 6286 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6287 | if (getTypeAction(VT: FpValue.getValueType()) != TargetLowering::TypeWidenVector) |
| 6288 | return DAG.UnrollVectorOp(N, ResNE: WidenVT.getVectorNumElements()); |
| 6289 | SDValue Arg = GetWidenedVector(Op: FpValue); |
| 6290 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: WidenVT, Ops: {Arg, N->getOperand(Num: 1)}, |
| 6291 | Flags: N->getFlags()); |
| 6292 | } |
| 6293 | |
| 6294 | SDValue DAGTypeLegalizer::WidenVecRes_ExpOp(SDNode *N) { |
| 6295 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6296 | SDValue InOp = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 6297 | SDValue RHS = N->getOperand(Num: 1); |
| 6298 | EVT ExpVT = RHS.getValueType(); |
| 6299 | SDValue ExpOp = RHS; |
| 6300 | if (ExpVT.isVector()) { |
| 6301 | EVT WideExpVT = WidenVT.changeVectorElementType( |
| 6302 | Context&: *DAG.getContext(), EltVT: ExpVT.getVectorElementType()); |
| 6303 | ExpOp = ModifyToType(InOp: RHS, NVT: WideExpVT); |
| 6304 | } |
| 6305 | |
| 6306 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: WidenVT, N1: InOp, N2: ExpOp); |
| 6307 | } |
| 6308 | |
| 6309 | SDValue DAGTypeLegalizer::WidenVecRes_Unary(SDNode *N) { |
| 6310 | // Unary op widening. |
| 6311 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6312 | SDValue InOp = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 6313 | if (N->getNumOperands() == 1) |
| 6314 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: WidenVT, Operand: InOp, Flags: N->getFlags()); |
| 6315 | if (N->getOpcode() == ISD::AssertNoFPClass) |
| 6316 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: WidenVT, N1: InOp, |
| 6317 | N2: N->getOperand(Num: 1), Flags: N->getFlags()); |
| 6318 | |
| 6319 | assert(N->getNumOperands() == 3 && "Unexpected number of operands!" ); |
| 6320 | assert(N->isVPOpcode() && "Expected VP opcode" ); |
| 6321 | |
| 6322 | SDValue Mask = |
| 6323 | GetWidenedMask(Mask: N->getOperand(Num: 1), EC: WidenVT.getVectorElementCount()); |
| 6324 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: WidenVT, |
| 6325 | Ops: {InOp, Mask, N->getOperand(Num: 2)}); |
| 6326 | } |
| 6327 | |
| 6328 | SDValue DAGTypeLegalizer::WidenVecRes_InregOp(SDNode *N) { |
| 6329 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6330 | EVT ExtVT = EVT::getVectorVT( |
| 6331 | Context&: *DAG.getContext(), |
| 6332 | VT: cast<VTSDNode>(Val: N->getOperand(Num: 1))->getVT().getVectorElementType(), |
| 6333 | EC: WidenVT.getVectorElementCount()); |
| 6334 | SDValue WidenLHS = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 6335 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), |
| 6336 | VT: WidenVT, N1: WidenLHS, N2: DAG.getValueType(ExtVT)); |
| 6337 | } |
| 6338 | |
| 6339 | SDValue DAGTypeLegalizer::WidenVecRes_UnaryOpWithTwoResults(SDNode *N, |
| 6340 | unsigned ResNo) { |
| 6341 | EVT VT0 = N->getValueType(ResNo: 0); |
| 6342 | EVT VT1 = N->getValueType(ResNo: 1); |
| 6343 | |
| 6344 | assert(VT0.isVector() && VT1.isVector() && |
| 6345 | VT0.getVectorElementCount() == VT1.getVectorElementCount() && |
| 6346 | "expected both results to be vectors of matching element count" ); |
| 6347 | |
| 6348 | LLVMContext &Ctx = *DAG.getContext(); |
| 6349 | SDValue InOp = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 6350 | |
| 6351 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: Ctx, VT: N->getValueType(ResNo)); |
| 6352 | ElementCount WidenEC = WidenVT.getVectorElementCount(); |
| 6353 | |
| 6354 | EVT WidenVT0 = EVT::getVectorVT(Context&: Ctx, VT: VT0.getVectorElementType(), EC: WidenEC); |
| 6355 | EVT WidenVT1 = EVT::getVectorVT(Context&: Ctx, VT: VT1.getVectorElementType(), EC: WidenEC); |
| 6356 | |
| 6357 | SDNode *WidenNode = |
| 6358 | DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), ResultTys: {WidenVT0, WidenVT1}, Ops: InOp) |
| 6359 | .getNode(); |
| 6360 | |
| 6361 | ReplaceOtherWidenResults(N, WidenNode, WidenResNo: ResNo); |
| 6362 | return SDValue(WidenNode, ResNo); |
| 6363 | } |
| 6364 | |
| 6365 | SDValue DAGTypeLegalizer::WidenVecRes_MERGE_VALUES(SDNode *N, unsigned ResNo) { |
| 6366 | SDValue WidenVec = DisintegrateMERGE_VALUES(N, ResNo); |
| 6367 | return GetWidenedVector(Op: WidenVec); |
| 6368 | } |
| 6369 | |
| 6370 | SDValue DAGTypeLegalizer::WidenVecRes_ADDRSPACECAST(SDNode *N) { |
| 6371 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6372 | SDValue InOp = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 6373 | auto *AddrSpaceCastN = cast<AddrSpaceCastSDNode>(Val: N); |
| 6374 | |
| 6375 | return DAG.getAddrSpaceCast(dl: SDLoc(N), VT: WidenVT, Ptr: InOp, |
| 6376 | SrcAS: AddrSpaceCastN->getSrcAddressSpace(), |
| 6377 | DestAS: AddrSpaceCastN->getDestAddressSpace()); |
| 6378 | } |
| 6379 | |
| 6380 | SDValue DAGTypeLegalizer::WidenVecRes_BITCAST(SDNode *N) { |
| 6381 | SDValue InOp = N->getOperand(Num: 0); |
| 6382 | EVT InVT = InOp.getValueType(); |
| 6383 | EVT VT = N->getValueType(ResNo: 0); |
| 6384 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT); |
| 6385 | SDLoc dl(N); |
| 6386 | |
| 6387 | switch (getTypeAction(VT: InVT)) { |
| 6388 | case TargetLowering::TypeLegal: |
| 6389 | break; |
| 6390 | case TargetLowering::TypeScalarizeScalableVector: |
| 6391 | report_fatal_error(reason: "Scalarization of scalable vectors is not supported." ); |
| 6392 | case TargetLowering::TypePromoteInteger: { |
| 6393 | // If the incoming type is a vector that is being promoted, then |
| 6394 | // we know that the elements are arranged differently and that we |
| 6395 | // must perform the conversion using a stack slot. |
| 6396 | if (InVT.isVector()) |
| 6397 | break; |
| 6398 | |
| 6399 | // If the InOp is promoted to the same size, convert it. Otherwise, |
| 6400 | // fall out of the switch and widen the promoted input. |
| 6401 | SDValue NInOp = GetPromotedInteger(Op: InOp); |
| 6402 | EVT NInVT = NInOp.getValueType(); |
| 6403 | if (WidenVT.bitsEq(VT: NInVT)) { |
| 6404 | // For big endian targets we need to shift the input integer or the |
| 6405 | // interesting bits will end up at the wrong place. |
| 6406 | if (DAG.getDataLayout().isBigEndian()) { |
| 6407 | unsigned ShiftAmt = NInVT.getSizeInBits() - InVT.getSizeInBits(); |
| 6408 | NInOp = DAG.getNode(Opcode: ISD::SHL, DL: dl, VT: NInVT, N1: NInOp, |
| 6409 | N2: DAG.getShiftAmountConstant(Val: ShiftAmt, VT: NInVT, DL: dl)); |
| 6410 | } |
| 6411 | return DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: WidenVT, Operand: NInOp); |
| 6412 | } |
| 6413 | InOp = NInOp; |
| 6414 | InVT = NInVT; |
| 6415 | break; |
| 6416 | } |
| 6417 | case TargetLowering::TypeSoftenFloat: |
| 6418 | case TargetLowering::TypeSoftPromoteHalf: |
| 6419 | case TargetLowering::TypeExpandInteger: |
| 6420 | case TargetLowering::TypeExpandFloat: |
| 6421 | case TargetLowering::TypeScalarizeVector: |
| 6422 | case TargetLowering::TypeSplitVector: |
| 6423 | break; |
| 6424 | case TargetLowering::TypeWidenVector: |
| 6425 | // If the InOp is widened to the same size, convert it. Otherwise, fall |
| 6426 | // out of the switch and widen the widened input. |
| 6427 | InOp = GetWidenedVector(Op: InOp); |
| 6428 | InVT = InOp.getValueType(); |
| 6429 | if (WidenVT.bitsEq(VT: InVT)) |
| 6430 | // The input widens to the same size. Convert to the widen value. |
| 6431 | return DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: WidenVT, Operand: InOp); |
| 6432 | break; |
| 6433 | } |
| 6434 | |
| 6435 | unsigned WidenSize = WidenVT.getSizeInBits(); |
| 6436 | unsigned InSize = InVT.getSizeInBits(); |
| 6437 | unsigned InScalarSize = InVT.getScalarSizeInBits(); |
| 6438 | // x86mmx is not an acceptable vector element type, so don't try. |
| 6439 | if (WidenSize % InScalarSize == 0 && InVT != MVT::x86mmx) { |
| 6440 | // Determine new input vector type. The new input vector type will use |
| 6441 | // the same element type (if its a vector) or use the input type as a |
| 6442 | // vector. It is the same size as the type to widen to. |
| 6443 | EVT NewInVT; |
| 6444 | unsigned NewNumParts = WidenSize / InSize; |
| 6445 | if (InVT.isVector()) { |
| 6446 | EVT InEltVT = InVT.getVectorElementType(); |
| 6447 | NewInVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: InEltVT, |
| 6448 | NumElements: WidenSize / InEltVT.getSizeInBits()); |
| 6449 | } else { |
| 6450 | // For big endian systems, using the promoted input scalar type |
| 6451 | // to produce the scalar_to_vector would put the desired bits into |
| 6452 | // the least significant byte(s) of the wider element zero. This |
| 6453 | // will mean that the users of the result vector are using incorrect |
| 6454 | // bits. Use the original input type instead. Although either input |
| 6455 | // type can be used on little endian systems, for consistency we |
| 6456 | // use the original type there as well. |
| 6457 | EVT OrigInVT = N->getOperand(Num: 0).getValueType(); |
| 6458 | NewNumParts = WidenSize / OrigInVT.getSizeInBits(); |
| 6459 | NewInVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: OrigInVT, NumElements: NewNumParts); |
| 6460 | } |
| 6461 | |
| 6462 | if (TLI.isTypeLegal(VT: NewInVT)) { |
| 6463 | SDValue NewVec; |
| 6464 | if (InVT.isVector()) { |
| 6465 | // Because the result and the input are different vector types, widening |
| 6466 | // the result could create a legal type but widening the input might |
| 6467 | // make it an illegal type that might lead to repeatedly splitting the |
| 6468 | // input and then widening it. To avoid this, we widen the input only if |
| 6469 | // it results in a legal type. |
| 6470 | if (WidenSize % InSize == 0) { |
| 6471 | SmallVector<SDValue, 16> Ops(NewNumParts, DAG.getPOISON(VT: InVT)); |
| 6472 | Ops[0] = InOp; |
| 6473 | |
| 6474 | NewVec = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: NewInVT, Ops); |
| 6475 | } else { |
| 6476 | SmallVector<SDValue, 16> Ops; |
| 6477 | DAG.ExtractVectorElements(Op: InOp, Args&: Ops); |
| 6478 | Ops.append(NumInputs: WidenSize / InScalarSize - Ops.size(), |
| 6479 | Elt: DAG.getPOISON(VT: InVT.getVectorElementType())); |
| 6480 | |
| 6481 | NewVec = DAG.getNode(Opcode: ISD::BUILD_VECTOR, DL: dl, VT: NewInVT, Ops); |
| 6482 | } |
| 6483 | } else { |
| 6484 | NewVec = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: dl, VT: NewInVT, Operand: InOp); |
| 6485 | } |
| 6486 | return DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: WidenVT, Operand: NewVec); |
| 6487 | } |
| 6488 | } |
| 6489 | |
| 6490 | return CreateStackStoreLoad(Op: InOp, DestVT: WidenVT); |
| 6491 | } |
| 6492 | |
| 6493 | SDValue DAGTypeLegalizer::WidenVecRes_LOOP_DEPENDENCE_MASK(SDNode *N) { |
| 6494 | return DAG.getNode( |
| 6495 | Opcode: N->getOpcode(), DL: SDLoc(N), |
| 6496 | VT: TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)), |
| 6497 | N1: N->getOperand(Num: 0), N2: N->getOperand(Num: 1), N3: N->getOperand(Num: 2), N4: N->getOperand(Num: 3)); |
| 6498 | } |
| 6499 | |
| 6500 | SDValue DAGTypeLegalizer::WidenVecRes_BUILD_VECTOR(SDNode *N) { |
| 6501 | SDLoc dl(N); |
| 6502 | // Build a vector with poison for the new nodes. |
| 6503 | EVT VT = N->getValueType(ResNo: 0); |
| 6504 | |
| 6505 | // Integer BUILD_VECTOR operands may be larger than the node's vector element |
| 6506 | // type. The POISONs need to have the same type as the existing operands. |
| 6507 | EVT EltVT = N->getOperand(Num: 0).getValueType(); |
| 6508 | unsigned NumElts = VT.getVectorNumElements(); |
| 6509 | |
| 6510 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT); |
| 6511 | unsigned WidenNumElts = WidenVT.getVectorNumElements(); |
| 6512 | |
| 6513 | SmallVector<SDValue, 16> NewOps(N->ops()); |
| 6514 | assert(WidenNumElts >= NumElts && "Shrinking vector instead of widening!" ); |
| 6515 | NewOps.append(NumInputs: WidenNumElts - NumElts, Elt: DAG.getPOISON(VT: EltVT)); |
| 6516 | |
| 6517 | return DAG.getBuildVector(VT: WidenVT, DL: dl, Ops: NewOps); |
| 6518 | } |
| 6519 | |
| 6520 | SDValue DAGTypeLegalizer::WidenVecRes_CONCAT_VECTORS(SDNode *N) { |
| 6521 | EVT InVT = N->getOperand(Num: 0).getValueType(); |
| 6522 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6523 | SDLoc dl(N); |
| 6524 | unsigned NumOperands = N->getNumOperands(); |
| 6525 | |
| 6526 | bool InputWidened = false; // Indicates we need to widen the input. |
| 6527 | if (getTypeAction(VT: InVT) != TargetLowering::TypeWidenVector) { |
| 6528 | unsigned WidenNumElts = WidenVT.getVectorMinNumElements(); |
| 6529 | unsigned NumInElts = InVT.getVectorMinNumElements(); |
| 6530 | if (WidenNumElts % NumInElts == 0) { |
| 6531 | // Add undef vectors to widen to correct length. |
| 6532 | unsigned NumConcat = WidenNumElts / NumInElts; |
| 6533 | SDValue UndefVal = DAG.getPOISON(VT: InVT); |
| 6534 | SmallVector<SDValue, 16> Ops(NumConcat); |
| 6535 | for (unsigned i=0; i < NumOperands; ++i) |
| 6536 | Ops[i] = N->getOperand(Num: i); |
| 6537 | for (unsigned i = NumOperands; i != NumConcat; ++i) |
| 6538 | Ops[i] = UndefVal; |
| 6539 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: WidenVT, Ops); |
| 6540 | } |
| 6541 | } else { |
| 6542 | InputWidened = true; |
| 6543 | if (WidenVT == TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: InVT)) { |
| 6544 | // The inputs and the result are widen to the same value. |
| 6545 | unsigned i; |
| 6546 | for (i=1; i < NumOperands; ++i) |
| 6547 | if (!N->getOperand(Num: i).isUndef()) |
| 6548 | break; |
| 6549 | |
| 6550 | if (i == NumOperands) |
| 6551 | // Everything but the first operand is an UNDEF so just return the |
| 6552 | // widened first operand. |
| 6553 | return GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 6554 | |
| 6555 | if (NumOperands == 2) { |
| 6556 | assert(!WidenVT.isScalableVector() && |
| 6557 | "Cannot use vector shuffles to widen CONCAT_VECTOR result" ); |
| 6558 | unsigned WidenNumElts = WidenVT.getVectorNumElements(); |
| 6559 | unsigned NumInElts = InVT.getVectorNumElements(); |
| 6560 | |
| 6561 | // Replace concat of two operands with a shuffle. |
| 6562 | SmallVector<int, 16> MaskOps(WidenNumElts, -1); |
| 6563 | for (unsigned i = 0; i < NumInElts; ++i) { |
| 6564 | MaskOps[i] = i; |
| 6565 | MaskOps[i + NumInElts] = i + WidenNumElts; |
| 6566 | } |
| 6567 | return DAG.getVectorShuffle(VT: WidenVT, dl, |
| 6568 | N1: GetWidenedVector(Op: N->getOperand(Num: 0)), |
| 6569 | N2: GetWidenedVector(Op: N->getOperand(Num: 1)), |
| 6570 | Mask: MaskOps); |
| 6571 | } |
| 6572 | } |
| 6573 | } |
| 6574 | |
| 6575 | assert(!WidenVT.isScalableVector() && |
| 6576 | "Cannot use build vectors to widen CONCAT_VECTOR result" ); |
| 6577 | unsigned WidenNumElts = WidenVT.getVectorNumElements(); |
| 6578 | unsigned NumInElts = InVT.getVectorNumElements(); |
| 6579 | |
| 6580 | // Fall back to use extracts and build vector. |
| 6581 | EVT EltVT = WidenVT.getVectorElementType(); |
| 6582 | SmallVector<SDValue, 16> Ops(WidenNumElts); |
| 6583 | unsigned Idx = 0; |
| 6584 | for (unsigned i=0; i < NumOperands; ++i) { |
| 6585 | SDValue InOp = N->getOperand(Num: i); |
| 6586 | if (InputWidened) |
| 6587 | InOp = GetWidenedVector(Op: InOp); |
| 6588 | for (unsigned j = 0; j < NumInElts; ++j) |
| 6589 | Ops[Idx++] = DAG.getExtractVectorElt(DL: dl, VT: EltVT, Vec: InOp, Idx: j); |
| 6590 | } |
| 6591 | SDValue UndefVal = DAG.getPOISON(VT: EltVT); |
| 6592 | for (; Idx < WidenNumElts; ++Idx) |
| 6593 | Ops[Idx] = UndefVal; |
| 6594 | return DAG.getBuildVector(VT: WidenVT, DL: dl, Ops); |
| 6595 | } |
| 6596 | |
| 6597 | SDValue DAGTypeLegalizer::WidenVecRes_INSERT_SUBVECTOR(SDNode *N) { |
| 6598 | EVT VT = N->getValueType(ResNo: 0); |
| 6599 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT); |
| 6600 | SDValue InOp1 = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 6601 | SDValue InOp2 = N->getOperand(Num: 1); |
| 6602 | SDValue Idx = N->getOperand(Num: 2); |
| 6603 | SDLoc dl(N); |
| 6604 | return DAG.getNode(Opcode: ISD::INSERT_SUBVECTOR, DL: dl, VT: WidenVT, N1: InOp1, N2: InOp2, N3: Idx); |
| 6605 | } |
| 6606 | |
| 6607 | SDValue DAGTypeLegalizer::(SDNode *N) { |
| 6608 | EVT VT = N->getValueType(ResNo: 0); |
| 6609 | EVT EltVT = VT.getVectorElementType(); |
| 6610 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT); |
| 6611 | SDValue InOp = N->getOperand(Num: 0); |
| 6612 | SDValue Idx = N->getOperand(Num: 1); |
| 6613 | SDLoc dl(N); |
| 6614 | |
| 6615 | auto InOpTypeAction = getTypeAction(VT: InOp.getValueType()); |
| 6616 | if (InOpTypeAction == TargetLowering::TypeWidenVector) |
| 6617 | InOp = GetWidenedVector(Op: InOp); |
| 6618 | |
| 6619 | EVT InVT = InOp.getValueType(); |
| 6620 | |
| 6621 | // Check if we can just return the input vector after widening. |
| 6622 | uint64_t IdxVal = Idx->getAsZExtVal(); |
| 6623 | if (IdxVal == 0 && InVT == WidenVT) |
| 6624 | return InOp; |
| 6625 | |
| 6626 | // Check if we can extract from the vector. |
| 6627 | unsigned WidenNumElts = WidenVT.getVectorMinNumElements(); |
| 6628 | unsigned InNumElts = InVT.getVectorMinNumElements(); |
| 6629 | unsigned VTNumElts = VT.getVectorMinNumElements(); |
| 6630 | assert(IdxVal % VTNumElts == 0 && |
| 6631 | "Expected Idx to be a multiple of subvector minimum vector length" ); |
| 6632 | if (IdxVal % WidenNumElts == 0 && IdxVal + WidenNumElts < InNumElts) |
| 6633 | return DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL: dl, VT: WidenVT, N1: InOp, N2: Idx); |
| 6634 | |
| 6635 | if (VT.isScalableVector()) { |
| 6636 | // Try to split the operation up into smaller extracts and concat the |
| 6637 | // results together, e.g. |
| 6638 | // nxv6i64 extract_subvector(nxv12i64, 6) |
| 6639 | // <-> |
| 6640 | // nxv8i64 concat( |
| 6641 | // nxv2i64 extract_subvector(nxv16i64, 6) |
| 6642 | // nxv2i64 extract_subvector(nxv16i64, 8) |
| 6643 | // nxv2i64 extract_subvector(nxv16i64, 10) |
| 6644 | // undef) |
| 6645 | unsigned GCD = std::gcd(m: VTNumElts, n: WidenNumElts); |
| 6646 | assert((IdxVal % GCD) == 0 && "Expected Idx to be a multiple of the broken " |
| 6647 | "down type's element count" ); |
| 6648 | EVT PartVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: EltVT, |
| 6649 | EC: ElementCount::getScalable(MinVal: GCD)); |
| 6650 | // Avoid recursion around e.g. nxv1i8. |
| 6651 | if (getTypeAction(VT: PartVT) != TargetLowering::TypeWidenVector) { |
| 6652 | SmallVector<SDValue> Parts; |
| 6653 | unsigned I = 0; |
| 6654 | for (; I < VTNumElts / GCD; ++I) |
| 6655 | Parts.push_back( |
| 6656 | Elt: DAG.getExtractSubvector(DL: dl, VT: PartVT, Vec: InOp, Idx: IdxVal + I * GCD)); |
| 6657 | for (; I < WidenNumElts / GCD; ++I) |
| 6658 | Parts.push_back(Elt: DAG.getPOISON(VT: PartVT)); |
| 6659 | |
| 6660 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: WidenVT, Ops: Parts); |
| 6661 | } |
| 6662 | |
| 6663 | // Fallback to extracting through memory. |
| 6664 | |
| 6665 | Align Alignment = DAG.getReducedAlign(VT: InVT, /*UseABI=*/false); |
| 6666 | SDValue StackPtr = DAG.CreateStackTemporary(Bytes: InVT.getStoreSize(), Alignment); |
| 6667 | MachineFunction &MF = DAG.getMachineFunction(); |
| 6668 | int FrameIndex = cast<FrameIndexSDNode>(Val: StackPtr.getNode())->getIndex(); |
| 6669 | auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: FrameIndex); |
| 6670 | |
| 6671 | MachineMemOperand *StoreMMO = MF.getMachineMemOperand( |
| 6672 | PtrInfo, F: MachineMemOperand::MOStore, |
| 6673 | Size: LocationSize::beforeOrAfterPointer(), BaseAlignment: Alignment); |
| 6674 | MachineMemOperand *LoadMMO = MF.getMachineMemOperand( |
| 6675 | PtrInfo, F: MachineMemOperand::MOLoad, |
| 6676 | Size: LocationSize::beforeOrAfterPointer(), BaseAlignment: Alignment); |
| 6677 | |
| 6678 | // Write out the input vector. |
| 6679 | SDValue Ch = DAG.getStore(Chain: DAG.getEntryNode(), dl, Val: InOp, Ptr: StackPtr, MMO: StoreMMO); |
| 6680 | |
| 6681 | // Build a mask to match the length of the non-widened result. |
| 6682 | SDValue Mask = |
| 6683 | DAG.getMaskFromElementCount(DL: dl, VT: WidenVT, Len: VT.getVectorElementCount()); |
| 6684 | |
| 6685 | // Read back the sub-vector setting the remaining lanes to poison. |
| 6686 | StackPtr = TLI.getVectorSubVecPointer(DAG, VecPtr: StackPtr, VecVT: InVT, SubVecVT: VT, Index: Idx); |
| 6687 | return DAG.getMaskedLoad( |
| 6688 | VT: WidenVT, dl, Chain: Ch, Base: StackPtr, Offset: DAG.getPOISON(VT: StackPtr.getValueType()), Mask, |
| 6689 | Src0: DAG.getPOISON(VT: WidenVT), MemVT: VT, MMO: LoadMMO, AM: ISD::UNINDEXED, ISD::NON_EXTLOAD); |
| 6690 | } |
| 6691 | |
| 6692 | // We could try widening the input to the right length but for now, extract |
| 6693 | // the original elements, fill the rest with undefs and build a vector. |
| 6694 | SmallVector<SDValue, 16> Ops(WidenNumElts); |
| 6695 | unsigned i; |
| 6696 | for (i = 0; i < VTNumElts; ++i) |
| 6697 | Ops[i] = DAG.getExtractVectorElt(DL: dl, VT: EltVT, Vec: InOp, Idx: IdxVal + i); |
| 6698 | |
| 6699 | SDValue UndefVal = DAG.getPOISON(VT: EltVT); |
| 6700 | for (; i < WidenNumElts; ++i) |
| 6701 | Ops[i] = UndefVal; |
| 6702 | return DAG.getBuildVector(VT: WidenVT, DL: dl, Ops); |
| 6703 | } |
| 6704 | |
| 6705 | SDValue DAGTypeLegalizer::WidenVecRes_AssertZext(SDNode *N) { |
| 6706 | SDValue InOp = ModifyToType( |
| 6707 | InOp: N->getOperand(Num: 0), |
| 6708 | NVT: TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)), FillWithZeroes: true); |
| 6709 | return DAG.getNode(Opcode: ISD::AssertZext, DL: SDLoc(N), VT: InOp.getValueType(), N1: InOp, |
| 6710 | N2: N->getOperand(Num: 1)); |
| 6711 | } |
| 6712 | |
| 6713 | SDValue DAGTypeLegalizer::WidenVecRes_INSERT_VECTOR_ELT(SDNode *N) { |
| 6714 | SDValue InOp = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 6715 | return DAG.getNode(Opcode: ISD::INSERT_VECTOR_ELT, DL: SDLoc(N), |
| 6716 | VT: InOp.getValueType(), N1: InOp, |
| 6717 | N2: N->getOperand(Num: 1), N3: N->getOperand(Num: 2)); |
| 6718 | } |
| 6719 | |
| 6720 | /// Either return the same load or provide appropriate casts |
| 6721 | /// from the load and return that. |
| 6722 | static SDValue coerceLoadedValue(SDValue LdOp, EVT FirstVT, EVT WidenVT, |
| 6723 | TypeSize LdWidth, TypeSize FirstVTWidth, |
| 6724 | SDLoc dl, SelectionDAG &DAG) { |
| 6725 | assert(TypeSize::isKnownLE(LdWidth, FirstVTWidth) && |
| 6726 | "Load width must be less than or equal to first value type width" ); |
| 6727 | TypeSize WidenWidth = WidenVT.getSizeInBits(); |
| 6728 | if (!FirstVT.isVector()) { |
| 6729 | unsigned NumElts = |
| 6730 | WidenWidth.getFixedValue() / FirstVTWidth.getFixedValue(); |
| 6731 | EVT NewVecVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: FirstVT, NumElements: NumElts); |
| 6732 | SDValue VecOp = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: dl, VT: NewVecVT, Operand: LdOp); |
| 6733 | return DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: WidenVT, Operand: VecOp); |
| 6734 | } |
| 6735 | assert(FirstVT == WidenVT && "First value type must equal widen value type" ); |
| 6736 | return LdOp; |
| 6737 | } |
| 6738 | |
| 6739 | /// Inverse of coerceLoadedValue: pull a FirstVT-sized scalar/vector out of the |
| 6740 | /// widened value so it can be issued in a single atomic store. |
| 6741 | static SDValue coerceStoredValue(SDValue StVal, EVT FirstVT, EVT WidenVT, |
| 6742 | TypeSize FirstVTWidth, const SDLoc &dl, |
| 6743 | SelectionDAG &DAG) { |
| 6744 | TypeSize WidenWidth = WidenVT.getSizeInBits(); |
| 6745 | if (!FirstVT.isVector()) { |
| 6746 | unsigned NumElts = |
| 6747 | WidenWidth.getFixedValue() / FirstVTWidth.getFixedValue(); |
| 6748 | EVT NewVecVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: FirstVT, NumElements: NumElts); |
| 6749 | SDValue VecOp = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: NewVecVT, Operand: StVal); |
| 6750 | return DAG.getExtractVectorElt(DL: dl, VT: FirstVT, Vec: VecOp, Idx: 0); |
| 6751 | } |
| 6752 | assert(FirstVT == WidenVT && "First value type must equal widen value type" ); |
| 6753 | return StVal; |
| 6754 | } |
| 6755 | |
| 6756 | static std::optional<EVT> findMemType(SelectionDAG &DAG, |
| 6757 | const TargetLowering &TLI, unsigned Width, |
| 6758 | EVT WidenVT, unsigned Align, |
| 6759 | unsigned WidenEx); |
| 6760 | |
| 6761 | SDValue DAGTypeLegalizer::WidenVecRes_ATOMIC_LOAD(AtomicSDNode *LD) { |
| 6762 | EVT WidenVT = |
| 6763 | TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: LD->getValueType(ResNo: 0)); |
| 6764 | EVT LdVT = LD->getMemoryVT(); |
| 6765 | SDLoc dl(LD); |
| 6766 | |
| 6767 | // Load information |
| 6768 | SDValue Chain = LD->getChain(); |
| 6769 | SDValue BasePtr = LD->getBasePtr(); |
| 6770 | |
| 6771 | TypeSize LdWidth = LdVT.getSizeInBits(); |
| 6772 | TypeSize WidenWidth = WidenVT.getSizeInBits(); |
| 6773 | TypeSize WidthDiff = WidenWidth - LdWidth; |
| 6774 | |
| 6775 | // Find the vector type that can load from. |
| 6776 | std::optional<EVT> FirstVT = |
| 6777 | findMemType(DAG, TLI, Width: LdWidth.getKnownMinValue(), WidenVT, /*LdAlign=*/Align: 0, |
| 6778 | WidenEx: WidthDiff.getKnownMinValue()); |
| 6779 | |
| 6780 | if (!FirstVT) |
| 6781 | return SDValue(); |
| 6782 | |
| 6783 | SmallVector<EVT, 8> MemVTs; |
| 6784 | TypeSize FirstVTWidth = FirstVT->getSizeInBits(); |
| 6785 | |
| 6786 | SDValue LdOp = DAG.getAtomicLoad(ExtType: ISD::NON_EXTLOAD, dl, MemVT: *FirstVT, VT: *FirstVT, |
| 6787 | Chain, Ptr: BasePtr, MMO: LD->getMemOperand()); |
| 6788 | |
| 6789 | // Load the element with one instruction. |
| 6790 | SDValue Result = coerceLoadedValue(LdOp, FirstVT: *FirstVT, WidenVT, LdWidth, |
| 6791 | FirstVTWidth, dl, DAG); |
| 6792 | |
| 6793 | // Modified the chain - switch anything that used the old chain to use |
| 6794 | // the new one. |
| 6795 | ReplaceValueWith(From: SDValue(LD, 1), To: LdOp.getValue(R: 1)); |
| 6796 | return Result; |
| 6797 | } |
| 6798 | |
| 6799 | SDValue DAGTypeLegalizer::WidenVecRes_LOAD(SDNode *N) { |
| 6800 | LoadSDNode *LD = cast<LoadSDNode>(Val: N); |
| 6801 | ISD::LoadExtType ExtType = LD->getExtensionType(); |
| 6802 | |
| 6803 | // A vector must always be stored in memory as-is, i.e. without any padding |
| 6804 | // between the elements, since various code depend on it, e.g. in the |
| 6805 | // handling of a bitcast of a vector type to int, which may be done with a |
| 6806 | // vector store followed by an integer load. A vector that does not have |
| 6807 | // elements that are byte-sized must therefore be stored as an integer |
| 6808 | // built out of the extracted vector elements. |
| 6809 | if (!LD->getMemoryVT().isByteSized()) { |
| 6810 | SDValue Value, NewChain; |
| 6811 | std::tie(args&: Value, args&: NewChain) = TLI.scalarizeVectorLoad(LD, DAG); |
| 6812 | ReplaceValueWith(From: SDValue(LD, 0), To: Value); |
| 6813 | ReplaceValueWith(From: SDValue(LD, 1), To: NewChain); |
| 6814 | return SDValue(); |
| 6815 | } |
| 6816 | |
| 6817 | // Generate a vector-predicated load if it is custom/legal on the target. To |
| 6818 | // avoid possible recursion, only do this if the widened mask type is legal. |
| 6819 | // FIXME: Not all targets may support EVL in VP_LOAD. These will have been |
| 6820 | // removed from the IR by the ExpandVectorPredication pass but we're |
| 6821 | // reintroducing them here. |
| 6822 | EVT VT = LD->getValueType(ResNo: 0); |
| 6823 | EVT WideVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT); |
| 6824 | EVT WideMaskVT = getSetCCResultType(VT: WideVT); |
| 6825 | |
| 6826 | if (ExtType == ISD::NON_EXTLOAD && |
| 6827 | TLI.isOperationLegalOrCustom(Op: ISD::VP_LOAD, VT: WideVT) && |
| 6828 | TLI.isTypeLegal(VT: WideMaskVT)) { |
| 6829 | SDLoc DL(N); |
| 6830 | SDValue Mask = DAG.getAllOnesConstant(DL, VT: WideMaskVT); |
| 6831 | SDValue EVL = DAG.getElementCount(DL, VT: TLI.getVPExplicitVectorLengthTy(), |
| 6832 | EC: VT.getVectorElementCount()); |
| 6833 | SDValue NewLoad = |
| 6834 | DAG.getLoadVP(AM: LD->getAddressingMode(), ExtType: ISD::NON_EXTLOAD, VT: WideVT, dl: DL, |
| 6835 | Chain: LD->getChain(), Ptr: LD->getBasePtr(), Offset: LD->getOffset(), Mask, |
| 6836 | EVL, MemVT: LD->getMemoryVT(), MMO: LD->getMemOperand()); |
| 6837 | |
| 6838 | // Modified the chain - switch anything that used the old chain to use |
| 6839 | // the new one. |
| 6840 | ReplaceValueWith(From: SDValue(N, 1), To: NewLoad.getValue(R: 1)); |
| 6841 | |
| 6842 | return NewLoad; |
| 6843 | } |
| 6844 | |
| 6845 | SDValue Result; |
| 6846 | SmallVector<SDValue, 16> LdChain; // Chain for the series of load |
| 6847 | if (ExtType != ISD::NON_EXTLOAD) |
| 6848 | Result = GenWidenVectorExtLoads(LdChain, LD, ExtType); |
| 6849 | else |
| 6850 | Result = GenWidenVectorLoads(LdChain, LD); |
| 6851 | |
| 6852 | if (Result) { |
| 6853 | // If we generate a single load, we can use that for the chain. Otherwise, |
| 6854 | // build a factor node to remember the multiple loads are independent and |
| 6855 | // chain to that. |
| 6856 | SDValue NewChain; |
| 6857 | if (LdChain.size() == 1) |
| 6858 | NewChain = LdChain[0]; |
| 6859 | else |
| 6860 | NewChain = DAG.getNode(Opcode: ISD::TokenFactor, DL: SDLoc(LD), VT: MVT::Other, Ops: LdChain); |
| 6861 | |
| 6862 | // Modified the chain - switch anything that used the old chain to use |
| 6863 | // the new one. |
| 6864 | ReplaceValueWith(From: SDValue(N, 1), To: NewChain); |
| 6865 | |
| 6866 | return Result; |
| 6867 | } |
| 6868 | |
| 6869 | if (VT.isVector()) { |
| 6870 | // If all else fails replace the load with a wide masked load. |
| 6871 | SDLoc DL(N); |
| 6872 | SDValue Mask = |
| 6873 | DAG.getMaskFromElementCount(DL, VT: WideVT, Len: VT.getVectorElementCount()); |
| 6874 | |
| 6875 | SDValue NewLoad = DAG.getMaskedLoad( |
| 6876 | VT: WideVT, dl: DL, Chain: LD->getChain(), Base: LD->getBasePtr(), Offset: LD->getOffset(), Mask, |
| 6877 | Src0: DAG.getPOISON(VT: WideVT), MemVT: LD->getMemoryVT(), MMO: LD->getMemOperand(), |
| 6878 | AM: LD->getAddressingMode(), LD->getExtensionType()); |
| 6879 | |
| 6880 | ReplaceValueWith(From: SDValue(N, 1), To: NewLoad.getValue(R: 1)); |
| 6881 | return NewLoad; |
| 6882 | } |
| 6883 | |
| 6884 | report_fatal_error(reason: "Unable to widen vector load" ); |
| 6885 | } |
| 6886 | |
| 6887 | SDValue DAGTypeLegalizer::WidenVecRes_VP_LOAD(VPLoadSDNode *N) { |
| 6888 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6889 | SDValue Mask = N->getMask(); |
| 6890 | SDValue EVL = N->getVectorLength(); |
| 6891 | ISD::LoadExtType ExtType = N->getExtensionType(); |
| 6892 | SDLoc dl(N); |
| 6893 | |
| 6894 | // The mask should be widened as well |
| 6895 | assert(getTypeAction(Mask.getValueType()) == |
| 6896 | TargetLowering::TypeWidenVector && |
| 6897 | "Unable to widen binary VP op" ); |
| 6898 | Mask = GetWidenedVector(Op: Mask); |
| 6899 | assert(Mask.getValueType().getVectorElementCount() == |
| 6900 | TLI.getTypeToTransformTo(*DAG.getContext(), Mask.getValueType()) |
| 6901 | .getVectorElementCount() && |
| 6902 | "Unable to widen vector load" ); |
| 6903 | |
| 6904 | SDValue Res = |
| 6905 | DAG.getLoadVP(AM: N->getAddressingMode(), ExtType, VT: WidenVT, dl, Chain: N->getChain(), |
| 6906 | Ptr: N->getBasePtr(), Offset: N->getOffset(), Mask, EVL, |
| 6907 | MemVT: N->getMemoryVT(), MMO: N->getMemOperand(), IsExpanding: N->isExpandingLoad()); |
| 6908 | // Legalize the chain result - switch anything that used the old chain to |
| 6909 | // use the new one. |
| 6910 | ReplaceValueWith(From: SDValue(N, 1), To: Res.getValue(R: 1)); |
| 6911 | return Res; |
| 6912 | } |
| 6913 | |
| 6914 | SDValue DAGTypeLegalizer::WidenVecRes_VP_LOAD_FF(VPLoadFFSDNode *N) { |
| 6915 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6916 | SDValue Mask = N->getMask(); |
| 6917 | SDValue EVL = N->getVectorLength(); |
| 6918 | SDLoc dl(N); |
| 6919 | |
| 6920 | // The mask should be widened as well |
| 6921 | assert(getTypeAction(Mask.getValueType()) == |
| 6922 | TargetLowering::TypeWidenVector && |
| 6923 | "Unable to widen binary VP op" ); |
| 6924 | Mask = GetWidenedVector(Op: Mask); |
| 6925 | assert(Mask.getValueType().getVectorElementCount() == |
| 6926 | TLI.getTypeToTransformTo(*DAG.getContext(), Mask.getValueType()) |
| 6927 | .getVectorElementCount() && |
| 6928 | "Unable to widen vector load" ); |
| 6929 | |
| 6930 | SDValue Res = DAG.getLoadFFVP(VT: WidenVT, DL: dl, Chain: N->getChain(), Ptr: N->getBasePtr(), |
| 6931 | Mask, EVL, MMO: N->getMemOperand()); |
| 6932 | ReplaceValueWith(From: SDValue(N, 1), To: Res.getValue(R: 1)); |
| 6933 | ReplaceValueWith(From: SDValue(N, 2), To: Res.getValue(R: 2)); |
| 6934 | return Res; |
| 6935 | } |
| 6936 | |
| 6937 | SDValue DAGTypeLegalizer::WidenVecRes_VP_STRIDED_LOAD(VPStridedLoadSDNode *N) { |
| 6938 | SDLoc DL(N); |
| 6939 | |
| 6940 | // The mask should be widened as well |
| 6941 | SDValue Mask = N->getMask(); |
| 6942 | assert(getTypeAction(Mask.getValueType()) == |
| 6943 | TargetLowering::TypeWidenVector && |
| 6944 | "Unable to widen VP strided load" ); |
| 6945 | Mask = GetWidenedVector(Op: Mask); |
| 6946 | |
| 6947 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 6948 | assert(Mask.getValueType().getVectorElementCount() == |
| 6949 | WidenVT.getVectorElementCount() && |
| 6950 | "Data and mask vectors should have the same number of elements" ); |
| 6951 | |
| 6952 | SDValue Res = DAG.getStridedLoadVP( |
| 6953 | AM: N->getAddressingMode(), ExtType: N->getExtensionType(), VT: WidenVT, DL, Chain: N->getChain(), |
| 6954 | Ptr: N->getBasePtr(), Offset: N->getOffset(), Stride: N->getStride(), Mask, |
| 6955 | EVL: N->getVectorLength(), MemVT: N->getMemoryVT(), MMO: N->getMemOperand(), |
| 6956 | IsExpanding: N->isExpandingLoad()); |
| 6957 | |
| 6958 | // Legalize the chain result - switch anything that used the old chain to |
| 6959 | // use the new one. |
| 6960 | ReplaceValueWith(From: SDValue(N, 1), To: Res.getValue(R: 1)); |
| 6961 | return Res; |
| 6962 | } |
| 6963 | |
| 6964 | SDValue DAGTypeLegalizer::WidenVecRes_VECTOR_COMPRESS(SDNode *N) { |
| 6965 | SDValue Vec = N->getOperand(Num: 0); |
| 6966 | SDValue Mask = N->getOperand(Num: 1); |
| 6967 | SDValue Passthru = N->getOperand(Num: 2); |
| 6968 | EVT WideVecVT = |
| 6969 | TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: Vec.getValueType()); |
| 6970 | EVT WideMaskVT = EVT::getVectorVT(Context&: *DAG.getContext(), |
| 6971 | VT: Mask.getValueType().getVectorElementType(), |
| 6972 | EC: WideVecVT.getVectorElementCount()); |
| 6973 | |
| 6974 | SDValue WideVec = ModifyToType(InOp: Vec, NVT: WideVecVT); |
| 6975 | SDValue WideMask = ModifyToType(InOp: Mask, NVT: WideMaskVT, /*FillWithZeroes=*/true); |
| 6976 | SDValue WidePassthru = ModifyToType(InOp: Passthru, NVT: WideVecVT); |
| 6977 | return DAG.getNode(Opcode: ISD::VECTOR_COMPRESS, DL: SDLoc(N), VT: WideVecVT, N1: WideVec, |
| 6978 | N2: WideMask, N3: WidePassthru); |
| 6979 | } |
| 6980 | |
| 6981 | SDValue DAGTypeLegalizer::WidenVecRes_MLOAD(MaskedLoadSDNode *N) { |
| 6982 | EVT VT = N->getValueType(ResNo: 0); |
| 6983 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT); |
| 6984 | SDValue Mask = N->getMask(); |
| 6985 | EVT MaskVT = Mask.getValueType(); |
| 6986 | SDValue PassThru = GetWidenedVector(Op: N->getPassThru()); |
| 6987 | ISD::LoadExtType ExtType = N->getExtensionType(); |
| 6988 | SDLoc dl(N); |
| 6989 | |
| 6990 | EVT WideMaskVT = |
| 6991 | EVT::getVectorVT(Context&: *DAG.getContext(), VT: MaskVT.getVectorElementType(), |
| 6992 | EC: WidenVT.getVectorElementCount()); |
| 6993 | |
| 6994 | if (ExtType == ISD::NON_EXTLOAD && !N->isExpandingLoad() && |
| 6995 | TLI.isOperationLegalOrCustom(Op: ISD::VP_LOAD, VT: WidenVT) && |
| 6996 | TLI.isTypeLegal(VT: WideMaskVT) && |
| 6997 | // If there is a passthru, we shouldn't use vp.load. However, |
| 6998 | // type legalizer will struggle on masked.load with |
| 6999 | // scalable vectors, so for scalable vectors, we still use vp.load |
| 7000 | // but manually merge the load result with the passthru using vp.select. |
| 7001 | (N->getPassThru()->isUndef() || VT.isScalableVector())) { |
| 7002 | Mask = DAG.getInsertSubvector(DL: dl, Vec: DAG.getPOISON(VT: WideMaskVT), SubVec: Mask, Idx: 0); |
| 7003 | SDValue EVL = DAG.getElementCount(DL: dl, VT: TLI.getVPExplicitVectorLengthTy(), |
| 7004 | EC: VT.getVectorElementCount()); |
| 7005 | SDValue NewLoad = |
| 7006 | DAG.getLoadVP(AM: N->getAddressingMode(), ExtType: ISD::NON_EXTLOAD, VT: WidenVT, dl, |
| 7007 | Chain: N->getChain(), Ptr: N->getBasePtr(), Offset: N->getOffset(), Mask, EVL, |
| 7008 | MemVT: N->getMemoryVT(), MMO: N->getMemOperand()); |
| 7009 | SDValue NewVal = NewLoad; |
| 7010 | |
| 7011 | // Manually merge with vselect |
| 7012 | if (!N->getPassThru()->isUndef()) { |
| 7013 | assert(WidenVT.isScalableVector()); |
| 7014 | NewVal = DAG.getNode(Opcode: ISD::VSELECT, DL: dl, VT: WidenVT, N1: Mask, N2: NewVal, N3: PassThru); |
| 7015 | // The lanes past EVL are poison. |
| 7016 | NewVal = DAG.getNode(Opcode: ISD::VP_MERGE, DL: dl, VT: WidenVT, |
| 7017 | N1: DAG.getAllOnesConstant(DL: dl, VT: WideMaskVT), N2: NewVal, |
| 7018 | N3: DAG.getPOISON(VT: WidenVT), N4: EVL); |
| 7019 | } |
| 7020 | |
| 7021 | // Modified the chain - switch anything that used the old chain to use |
| 7022 | // the new one. |
| 7023 | ReplaceValueWith(From: SDValue(N, 1), To: NewLoad.getValue(R: 1)); |
| 7024 | |
| 7025 | return NewVal; |
| 7026 | } |
| 7027 | |
| 7028 | // The mask should be widened as well |
| 7029 | Mask = ModifyToType(InOp: Mask, NVT: WideMaskVT, FillWithZeroes: true); |
| 7030 | |
| 7031 | SDValue Res = DAG.getMaskedLoad( |
| 7032 | VT: WidenVT, dl, Chain: N->getChain(), Base: N->getBasePtr(), Offset: N->getOffset(), Mask, |
| 7033 | Src0: PassThru, MemVT: N->getMemoryVT(), MMO: N->getMemOperand(), AM: N->getAddressingMode(), |
| 7034 | ExtType, IsExpanding: N->isExpandingLoad()); |
| 7035 | // Legalize the chain result - switch anything that used the old chain to |
| 7036 | // use the new one. |
| 7037 | ReplaceValueWith(From: SDValue(N, 1), To: Res.getValue(R: 1)); |
| 7038 | return Res; |
| 7039 | } |
| 7040 | |
| 7041 | SDValue DAGTypeLegalizer::WidenVecRes_MGATHER(MaskedGatherSDNode *N) { |
| 7042 | |
| 7043 | EVT WideVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 7044 | SDValue Mask = N->getMask(); |
| 7045 | EVT MaskVT = Mask.getValueType(); |
| 7046 | SDValue PassThru = GetWidenedVector(Op: N->getPassThru()); |
| 7047 | SDValue Scale = N->getScale(); |
| 7048 | ElementCount WideEC = WideVT.getVectorElementCount(); |
| 7049 | SDLoc dl(N); |
| 7050 | |
| 7051 | // The mask should be widened as well |
| 7052 | EVT WideMaskVT = EVT::getVectorVT(Context&: *DAG.getContext(), |
| 7053 | VT: MaskVT.getVectorElementType(), EC: WideEC); |
| 7054 | Mask = ModifyToType(InOp: Mask, NVT: WideMaskVT, FillWithZeroes: true); |
| 7055 | |
| 7056 | // Widen the Index operand |
| 7057 | SDValue Index = N->getIndex(); |
| 7058 | EVT WideIndexVT = EVT::getVectorVT( |
| 7059 | Context&: *DAG.getContext(), VT: Index.getValueType().getScalarType(), EC: WideEC); |
| 7060 | Index = ModifyToType(InOp: Index, NVT: WideIndexVT); |
| 7061 | SDValue Ops[] = { N->getChain(), PassThru, Mask, N->getBasePtr(), Index, |
| 7062 | Scale }; |
| 7063 | |
| 7064 | // Widen the MemoryType |
| 7065 | EVT WideMemVT = EVT::getVectorVT(Context&: *DAG.getContext(), |
| 7066 | VT: N->getMemoryVT().getScalarType(), EC: WideEC); |
| 7067 | SDValue Res = DAG.getMaskedGather(VTs: DAG.getVTList(VT1: WideVT, VT2: MVT::Other), |
| 7068 | MemVT: WideMemVT, dl, Ops, MMO: N->getMemOperand(), |
| 7069 | IndexType: N->getIndexType(), ExtTy: N->getExtensionType()); |
| 7070 | |
| 7071 | // Legalize the chain result - switch anything that used the old chain to |
| 7072 | // use the new one. |
| 7073 | ReplaceValueWith(From: SDValue(N, 1), To: Res.getValue(R: 1)); |
| 7074 | return Res; |
| 7075 | } |
| 7076 | |
| 7077 | SDValue DAGTypeLegalizer::WidenVecRes_VP_GATHER(VPGatherSDNode *N) { |
| 7078 | EVT WideVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 7079 | SDValue Mask = N->getMask(); |
| 7080 | SDValue Scale = N->getScale(); |
| 7081 | ElementCount WideEC = WideVT.getVectorElementCount(); |
| 7082 | SDLoc dl(N); |
| 7083 | |
| 7084 | SDValue Index = GetWidenedVector(Op: N->getIndex()); |
| 7085 | EVT WideMemVT = EVT::getVectorVT(Context&: *DAG.getContext(), |
| 7086 | VT: N->getMemoryVT().getScalarType(), EC: WideEC); |
| 7087 | Mask = GetWidenedMask(Mask, EC: WideEC); |
| 7088 | |
| 7089 | SDValue Ops[] = {N->getChain(), N->getBasePtr(), Index, Scale, |
| 7090 | Mask, N->getVectorLength()}; |
| 7091 | SDValue Res = DAG.getGatherVP(VTs: DAG.getVTList(VT1: WideVT, VT2: MVT::Other), VT: WideMemVT, |
| 7092 | dl, Ops, MMO: N->getMemOperand(), IndexType: N->getIndexType()); |
| 7093 | |
| 7094 | // Legalize the chain result - switch anything that used the old chain to |
| 7095 | // use the new one. |
| 7096 | ReplaceValueWith(From: SDValue(N, 1), To: Res.getValue(R: 1)); |
| 7097 | return Res; |
| 7098 | } |
| 7099 | |
| 7100 | SDValue DAGTypeLegalizer::WidenVecRes_ScalarOp(SDNode *N) { |
| 7101 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 7102 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: WidenVT, Operand: N->getOperand(Num: 0)); |
| 7103 | } |
| 7104 | |
| 7105 | // Return true is this is a SETCC node or a strict version of it. |
| 7106 | static inline bool isSETCCOp(unsigned Opcode) { |
| 7107 | switch (Opcode) { |
| 7108 | case ISD::SETCC: |
| 7109 | case ISD::STRICT_FSETCC: |
| 7110 | case ISD::STRICT_FSETCCS: |
| 7111 | return true; |
| 7112 | } |
| 7113 | return false; |
| 7114 | } |
| 7115 | |
| 7116 | // Return true if this is a node that could have two SETCCs as operands. |
| 7117 | static inline bool isLogicalMaskOp(unsigned Opcode) { |
| 7118 | switch (Opcode) { |
| 7119 | case ISD::AND: |
| 7120 | case ISD::OR: |
| 7121 | case ISD::XOR: |
| 7122 | return true; |
| 7123 | } |
| 7124 | return false; |
| 7125 | } |
| 7126 | |
| 7127 | // If N is a SETCC or a strict variant of it, return the type |
| 7128 | // of the compare operands. |
| 7129 | static inline EVT getSETCCOperandType(SDValue N) { |
| 7130 | unsigned OpNo = N->isStrictFPOpcode() ? 1 : 0; |
| 7131 | return N->getOperand(Num: OpNo).getValueType(); |
| 7132 | } |
| 7133 | |
| 7134 | // This is used just for the assert in convertMask(). Check that this either |
| 7135 | // a SETCC or a previously handled SETCC by convertMask(). |
| 7136 | #ifndef NDEBUG |
| 7137 | static inline bool isSETCCorConvertedSETCC(SDValue N) { |
| 7138 | if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR) |
| 7139 | N = N.getOperand(0); |
| 7140 | else if (N.getOpcode() == ISD::CONCAT_VECTORS) { |
| 7141 | for (unsigned i = 1; i < N->getNumOperands(); ++i) |
| 7142 | if (!N->getOperand(i)->isUndef()) |
| 7143 | return false; |
| 7144 | N = N.getOperand(0); |
| 7145 | } |
| 7146 | |
| 7147 | if (N.getOpcode() == ISD::TRUNCATE) |
| 7148 | N = N.getOperand(0); |
| 7149 | else if (N.getOpcode() == ISD::SIGN_EXTEND) |
| 7150 | N = N.getOperand(0); |
| 7151 | |
| 7152 | if (isLogicalMaskOp(N.getOpcode())) |
| 7153 | return isSETCCorConvertedSETCC(N.getOperand(0)) && |
| 7154 | isSETCCorConvertedSETCC(N.getOperand(1)); |
| 7155 | |
| 7156 | return (isSETCCOp(N.getOpcode()) || |
| 7157 | ISD::isBuildVectorOfConstantSDNodes(N.getNode())); |
| 7158 | } |
| 7159 | #endif |
| 7160 | |
| 7161 | // Return a mask of vector type MaskVT to replace InMask. Also adjust MaskVT |
| 7162 | // to ToMaskVT if needed with vector extension or truncation. |
| 7163 | SDValue DAGTypeLegalizer::convertMask(SDValue InMask, EVT MaskVT, |
| 7164 | EVT ToMaskVT) { |
| 7165 | // Currently a SETCC or a AND/OR/XOR with two SETCCs are handled. |
| 7166 | // FIXME: This code seems to be too restrictive, we might consider |
| 7167 | // generalizing it or dropping it. |
| 7168 | assert(isSETCCorConvertedSETCC(InMask) && "Unexpected mask argument." ); |
| 7169 | |
| 7170 | // Make a new Mask node, with a legal result VT. |
| 7171 | SDValue Mask; |
| 7172 | SmallVector<SDValue, 4> Ops; |
| 7173 | for (unsigned i = 0, e = InMask->getNumOperands(); i < e; ++i) |
| 7174 | Ops.push_back(Elt: InMask->getOperand(Num: i)); |
| 7175 | if (InMask->isStrictFPOpcode()) { |
| 7176 | Mask = DAG.getNode(Opcode: InMask->getOpcode(), DL: SDLoc(InMask), |
| 7177 | ResultTys: { MaskVT, MVT::Other }, Ops); |
| 7178 | ReplaceValueWith(From: InMask.getValue(R: 1), To: Mask.getValue(R: 1)); |
| 7179 | } |
| 7180 | else |
| 7181 | Mask = DAG.getNode(Opcode: InMask->getOpcode(), DL: SDLoc(InMask), VT: MaskVT, Ops, |
| 7182 | Flags: InMask->getFlags()); |
| 7183 | |
| 7184 | // If MaskVT has smaller or bigger elements than ToMaskVT, a vector sign |
| 7185 | // extend or truncate is needed. |
| 7186 | LLVMContext &Ctx = *DAG.getContext(); |
| 7187 | unsigned MaskScalarBits = MaskVT.getScalarSizeInBits(); |
| 7188 | unsigned ToMaskScalBits = ToMaskVT.getScalarSizeInBits(); |
| 7189 | if (MaskScalarBits < ToMaskScalBits) { |
| 7190 | EVT ExtVT = EVT::getVectorVT(Context&: Ctx, VT: ToMaskVT.getVectorElementType(), |
| 7191 | NumElements: MaskVT.getVectorNumElements()); |
| 7192 | Mask = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL: SDLoc(Mask), VT: ExtVT, Operand: Mask); |
| 7193 | } else if (MaskScalarBits > ToMaskScalBits) { |
| 7194 | EVT TruncVT = EVT::getVectorVT(Context&: Ctx, VT: ToMaskVT.getVectorElementType(), |
| 7195 | NumElements: MaskVT.getVectorNumElements()); |
| 7196 | Mask = DAG.getNode(Opcode: ISD::TRUNCATE, DL: SDLoc(Mask), VT: TruncVT, Operand: Mask); |
| 7197 | } |
| 7198 | |
| 7199 | assert(Mask->getValueType(0).getScalarSizeInBits() == |
| 7200 | ToMaskVT.getScalarSizeInBits() && |
| 7201 | "Mask should have the right element size by now." ); |
| 7202 | |
| 7203 | // Adjust Mask to the right number of elements. |
| 7204 | unsigned CurrMaskNumEls = Mask->getValueType(ResNo: 0).getVectorNumElements(); |
| 7205 | if (CurrMaskNumEls > ToMaskVT.getVectorNumElements()) { |
| 7206 | Mask = DAG.getExtractSubvector(DL: SDLoc(Mask), VT: ToMaskVT, Vec: Mask, Idx: 0); |
| 7207 | } else if (CurrMaskNumEls < ToMaskVT.getVectorNumElements()) { |
| 7208 | unsigned NumSubVecs = (ToMaskVT.getVectorNumElements() / CurrMaskNumEls); |
| 7209 | EVT SubVT = Mask->getValueType(ResNo: 0); |
| 7210 | SmallVector<SDValue, 16> SubOps(NumSubVecs, DAG.getPOISON(VT: SubVT)); |
| 7211 | SubOps[0] = Mask; |
| 7212 | Mask = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: SDLoc(Mask), VT: ToMaskVT, Ops: SubOps); |
| 7213 | } |
| 7214 | |
| 7215 | assert((Mask->getValueType(0) == ToMaskVT) && |
| 7216 | "A mask of ToMaskVT should have been produced by now." ); |
| 7217 | |
| 7218 | return Mask; |
| 7219 | } |
| 7220 | |
| 7221 | // This method tries to handle some special cases for the vselect mask |
| 7222 | // and if needed adjusting the mask vector type to match that of the VSELECT. |
| 7223 | // Without it, many cases end up with scalarization of the SETCC, with many |
| 7224 | // unnecessary instructions. |
| 7225 | SDValue DAGTypeLegalizer::WidenVSELECTMask(SDNode *N) { |
| 7226 | LLVMContext &Ctx = *DAG.getContext(); |
| 7227 | SDValue Cond = N->getOperand(Num: 0); |
| 7228 | |
| 7229 | if (N->getOpcode() != ISD::VSELECT) |
| 7230 | return SDValue(); |
| 7231 | |
| 7232 | if (!isSETCCOp(Opcode: Cond->getOpcode()) && !isLogicalMaskOp(Opcode: Cond->getOpcode())) |
| 7233 | return SDValue(); |
| 7234 | |
| 7235 | // If this is a splitted VSELECT that was previously already handled, do |
| 7236 | // nothing. |
| 7237 | EVT CondVT = Cond->getValueType(ResNo: 0); |
| 7238 | if (CondVT.getScalarSizeInBits() != 1) |
| 7239 | return SDValue(); |
| 7240 | |
| 7241 | EVT VSelVT = N->getValueType(ResNo: 0); |
| 7242 | |
| 7243 | // This method can't handle scalable vector types. |
| 7244 | // FIXME: This support could be added in the future. |
| 7245 | if (VSelVT.isScalableVector()) |
| 7246 | return SDValue(); |
| 7247 | |
| 7248 | // Only handle vector types which are a power of 2. |
| 7249 | if (!isPowerOf2_64(Value: VSelVT.getSizeInBits())) |
| 7250 | return SDValue(); |
| 7251 | |
| 7252 | // Don't touch if this will be scalarized. |
| 7253 | EVT FinalVT = VSelVT; |
| 7254 | while (getTypeAction(VT: FinalVT) == TargetLowering::TypeSplitVector) |
| 7255 | FinalVT = FinalVT.getHalfNumVectorElementsVT(Context&: Ctx); |
| 7256 | |
| 7257 | if (FinalVT.getVectorNumElements() == 1) |
| 7258 | return SDValue(); |
| 7259 | |
| 7260 | // If there is support for an i1 vector mask, don't touch. |
| 7261 | if (isSETCCOp(Opcode: Cond.getOpcode())) { |
| 7262 | EVT SetCCOpVT = getSETCCOperandType(N: Cond); |
| 7263 | while (TLI.getTypeAction(Context&: Ctx, VT: SetCCOpVT) != TargetLowering::TypeLegal) |
| 7264 | SetCCOpVT = TLI.getTypeToTransformTo(Context&: Ctx, VT: SetCCOpVT); |
| 7265 | EVT SetCCResVT = getSetCCResultType(VT: SetCCOpVT); |
| 7266 | if (SetCCResVT.getScalarSizeInBits() == 1) |
| 7267 | return SDValue(); |
| 7268 | } else if (CondVT.getScalarType() == MVT::i1) { |
| 7269 | // If there is support for an i1 vector mask (or only scalar i1 conditions), |
| 7270 | // don't touch. |
| 7271 | while (TLI.getTypeAction(Context&: Ctx, VT: CondVT) != TargetLowering::TypeLegal) |
| 7272 | CondVT = TLI.getTypeToTransformTo(Context&: Ctx, VT: CondVT); |
| 7273 | |
| 7274 | if (CondVT.getScalarType() == MVT::i1) |
| 7275 | return SDValue(); |
| 7276 | } |
| 7277 | |
| 7278 | // Widen the vselect result type if needed. |
| 7279 | if (getTypeAction(VT: VSelVT) == TargetLowering::TypeWidenVector) |
| 7280 | VSelVT = TLI.getTypeToTransformTo(Context&: Ctx, VT: VSelVT); |
| 7281 | |
| 7282 | // The mask of the VSELECT should have integer elements. |
| 7283 | EVT ToMaskVT = VSelVT; |
| 7284 | if (!ToMaskVT.getScalarType().isInteger()) |
| 7285 | ToMaskVT = ToMaskVT.changeVectorElementTypeToInteger(); |
| 7286 | |
| 7287 | SDValue Mask; |
| 7288 | if (isSETCCOp(Opcode: Cond->getOpcode())) { |
| 7289 | EVT MaskVT = getSetCCResultType(VT: getSETCCOperandType(N: Cond)); |
| 7290 | Mask = convertMask(InMask: Cond, MaskVT, ToMaskVT); |
| 7291 | } else if (isLogicalMaskOp(Opcode: Cond->getOpcode()) && |
| 7292 | isSETCCOp(Opcode: Cond->getOperand(Num: 0).getOpcode()) && |
| 7293 | isSETCCOp(Opcode: Cond->getOperand(Num: 1).getOpcode())) { |
| 7294 | // Cond is (AND/OR/XOR (SETCC, SETCC)) |
| 7295 | SDValue SETCC0 = Cond->getOperand(Num: 0); |
| 7296 | SDValue SETCC1 = Cond->getOperand(Num: 1); |
| 7297 | EVT VT0 = getSetCCResultType(VT: getSETCCOperandType(N: SETCC0)); |
| 7298 | EVT VT1 = getSetCCResultType(VT: getSETCCOperandType(N: SETCC1)); |
| 7299 | unsigned ScalarBits0 = VT0.getScalarSizeInBits(); |
| 7300 | unsigned ScalarBits1 = VT1.getScalarSizeInBits(); |
| 7301 | unsigned ScalarBits_ToMask = ToMaskVT.getScalarSizeInBits(); |
| 7302 | EVT MaskVT; |
| 7303 | // If the two SETCCs have different VTs, either extend/truncate one of |
| 7304 | // them to the other "towards" ToMaskVT, or truncate one and extend the |
| 7305 | // other to ToMaskVT. |
| 7306 | if (ScalarBits0 != ScalarBits1) { |
| 7307 | EVT NarrowVT = ((ScalarBits0 < ScalarBits1) ? VT0 : VT1); |
| 7308 | EVT WideVT = ((NarrowVT == VT0) ? VT1 : VT0); |
| 7309 | if (ScalarBits_ToMask >= WideVT.getScalarSizeInBits()) |
| 7310 | MaskVT = WideVT; |
| 7311 | else if (ScalarBits_ToMask <= NarrowVT.getScalarSizeInBits()) |
| 7312 | MaskVT = NarrowVT; |
| 7313 | else |
| 7314 | MaskVT = ToMaskVT; |
| 7315 | } else |
| 7316 | // If the two SETCCs have the same VT, don't change it. |
| 7317 | MaskVT = VT0; |
| 7318 | |
| 7319 | // Make new SETCCs and logical nodes. |
| 7320 | SETCC0 = convertMask(InMask: SETCC0, MaskVT: VT0, ToMaskVT: MaskVT); |
| 7321 | SETCC1 = convertMask(InMask: SETCC1, MaskVT: VT1, ToMaskVT: MaskVT); |
| 7322 | Cond = DAG.getNode(Opcode: Cond->getOpcode(), DL: SDLoc(Cond), VT: MaskVT, N1: SETCC0, N2: SETCC1); |
| 7323 | |
| 7324 | // Convert the logical op for VSELECT if needed. |
| 7325 | Mask = convertMask(InMask: Cond, MaskVT, ToMaskVT); |
| 7326 | } else |
| 7327 | return SDValue(); |
| 7328 | |
| 7329 | return Mask; |
| 7330 | } |
| 7331 | |
| 7332 | SDValue DAGTypeLegalizer::WidenVecRes_Select(SDNode *N) { |
| 7333 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 7334 | ElementCount WidenEC = WidenVT.getVectorElementCount(); |
| 7335 | |
| 7336 | SDValue Cond1 = N->getOperand(Num: 0); |
| 7337 | EVT CondVT = Cond1.getValueType(); |
| 7338 | unsigned Opcode = N->getOpcode(); |
| 7339 | if (CondVT.isVector()) { |
| 7340 | if (SDValue WideCond = WidenVSELECTMask(N)) { |
| 7341 | SDValue InOp1 = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 7342 | SDValue InOp2 = GetWidenedVector(Op: N->getOperand(Num: 2)); |
| 7343 | assert(InOp1.getValueType() == WidenVT && InOp2.getValueType() == WidenVT); |
| 7344 | return DAG.getNode(Opcode, DL: SDLoc(N), VT: WidenVT, N1: WideCond, N2: InOp1, N3: InOp2); |
| 7345 | } |
| 7346 | |
| 7347 | EVT CondEltVT = CondVT.getVectorElementType(); |
| 7348 | EVT CondWidenVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: CondEltVT, EC: WidenEC); |
| 7349 | if (getTypeAction(VT: CondVT) == TargetLowering::TypeWidenVector) |
| 7350 | Cond1 = GetWidenedVector(Op: Cond1); |
| 7351 | |
| 7352 | // If we have to split the condition there is no point in widening the |
| 7353 | // select. This would result in an cycle of widening the select -> |
| 7354 | // widening the condition operand -> splitting the condition operand -> |
| 7355 | // splitting the select -> widening the select. Instead split this select |
| 7356 | // further and widen the resulting type. |
| 7357 | if (getTypeAction(VT: CondVT) == TargetLowering::TypeSplitVector) { |
| 7358 | SDValue SplitSelect = SplitVecOp_VSELECT(N, OpNo: 0); |
| 7359 | SDValue Res = ModifyToType(InOp: SplitSelect, NVT: WidenVT); |
| 7360 | return Res; |
| 7361 | } |
| 7362 | |
| 7363 | if (Cond1.getValueType() != CondWidenVT) |
| 7364 | Cond1 = ModifyToType(InOp: Cond1, NVT: CondWidenVT); |
| 7365 | } |
| 7366 | |
| 7367 | SDValue InOp1 = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 7368 | SDValue InOp2 = GetWidenedVector(Op: N->getOperand(Num: 2)); |
| 7369 | assert(InOp1.getValueType() == WidenVT && InOp2.getValueType() == WidenVT); |
| 7370 | if (Opcode == ISD::VP_SELECT || Opcode == ISD::VP_MERGE) |
| 7371 | return DAG.getNode(Opcode, DL: SDLoc(N), VT: WidenVT, N1: Cond1, N2: InOp1, N3: InOp2, |
| 7372 | N4: N->getOperand(Num: 3)); |
| 7373 | return DAG.getNode(Opcode, DL: SDLoc(N), VT: WidenVT, N1: Cond1, N2: InOp1, N3: InOp2); |
| 7374 | } |
| 7375 | |
| 7376 | SDValue DAGTypeLegalizer::WidenVecRes_SELECT_CC(SDNode *N) { |
| 7377 | SDValue InOp1 = GetWidenedVector(Op: N->getOperand(Num: 2)); |
| 7378 | SDValue InOp2 = GetWidenedVector(Op: N->getOperand(Num: 3)); |
| 7379 | return DAG.getNode(Opcode: ISD::SELECT_CC, DL: SDLoc(N), |
| 7380 | VT: InOp1.getValueType(), N1: N->getOperand(Num: 0), |
| 7381 | N2: N->getOperand(Num: 1), N3: InOp1, N4: InOp2, N5: N->getOperand(Num: 4)); |
| 7382 | } |
| 7383 | |
| 7384 | SDValue DAGTypeLegalizer::WidenVecRes_UNDEF(SDNode *N) { |
| 7385 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 7386 | return DAG.getUNDEF(VT: WidenVT); |
| 7387 | } |
| 7388 | |
| 7389 | SDValue DAGTypeLegalizer::WidenVecRes_VECTOR_SHUFFLE(ShuffleVectorSDNode *N) { |
| 7390 | EVT VT = N->getValueType(ResNo: 0); |
| 7391 | SDLoc dl(N); |
| 7392 | |
| 7393 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT); |
| 7394 | unsigned NumElts = VT.getVectorNumElements(); |
| 7395 | unsigned WidenNumElts = WidenVT.getVectorNumElements(); |
| 7396 | |
| 7397 | SDValue InOp1 = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 7398 | SDValue InOp2 = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 7399 | |
| 7400 | // Adjust mask based on new input vector length. |
| 7401 | SmallVector<int, 16> NewMask(WidenNumElts, -1); |
| 7402 | for (unsigned i = 0; i != NumElts; ++i) { |
| 7403 | int Idx = N->getMaskElt(Idx: i); |
| 7404 | if (Idx < (int)NumElts) |
| 7405 | NewMask[i] = Idx; |
| 7406 | else |
| 7407 | NewMask[i] = Idx - NumElts + WidenNumElts; |
| 7408 | } |
| 7409 | return DAG.getVectorShuffle(VT: WidenVT, dl, N1: InOp1, N2: InOp2, Mask: NewMask); |
| 7410 | } |
| 7411 | |
| 7412 | SDValue DAGTypeLegalizer::WidenVecRes_VECTOR_REVERSE(SDNode *N) { |
| 7413 | EVT VT = N->getValueType(ResNo: 0); |
| 7414 | EVT EltVT = VT.getVectorElementType(); |
| 7415 | SDLoc dl(N); |
| 7416 | |
| 7417 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT); |
| 7418 | SDValue OpValue = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 7419 | assert(WidenVT == OpValue.getValueType() && "Unexpected widened vector type" ); |
| 7420 | |
| 7421 | SDValue ReverseVal = DAG.getNode(Opcode: ISD::VECTOR_REVERSE, DL: dl, VT: WidenVT, Operand: OpValue); |
| 7422 | unsigned WidenNumElts = WidenVT.getVectorMinNumElements(); |
| 7423 | unsigned VTNumElts = VT.getVectorMinNumElements(); |
| 7424 | unsigned IdxVal = WidenNumElts - VTNumElts; |
| 7425 | |
| 7426 | if (VT.isScalableVector()) { |
| 7427 | // Try to split the 'Widen ReverseVal' into smaller extracts and concat the |
| 7428 | // results together, e.g.(nxv6i64 -> nxv8i64) |
| 7429 | // nxv8i64 vector_reverse |
| 7430 | // <-> |
| 7431 | // nxv8i64 concat( |
| 7432 | // nxv2i64 extract_subvector(nxv8i64, 2) |
| 7433 | // nxv2i64 extract_subvector(nxv8i64, 4) |
| 7434 | // nxv2i64 extract_subvector(nxv8i64, 6) |
| 7435 | // nxv2i64 undef) |
| 7436 | |
| 7437 | unsigned GCD = std::gcd(m: VTNumElts, n: WidenNumElts); |
| 7438 | EVT PartVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: EltVT, |
| 7439 | EC: ElementCount::getScalable(MinVal: GCD)); |
| 7440 | assert((IdxVal % GCD) == 0 && "Expected Idx to be a multiple of the broken " |
| 7441 | "down type's element count" ); |
| 7442 | SmallVector<SDValue> Parts; |
| 7443 | unsigned i = 0; |
| 7444 | for (; i < VTNumElts / GCD; ++i) |
| 7445 | Parts.push_back( |
| 7446 | Elt: DAG.getExtractSubvector(DL: dl, VT: PartVT, Vec: ReverseVal, Idx: IdxVal + i * GCD)); |
| 7447 | for (; i < WidenNumElts / GCD; ++i) |
| 7448 | Parts.push_back(Elt: DAG.getPOISON(VT: PartVT)); |
| 7449 | |
| 7450 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: WidenVT, Ops: Parts); |
| 7451 | } |
| 7452 | |
| 7453 | // Use VECTOR_SHUFFLE to combine new vector from 'ReverseVal' for |
| 7454 | // fixed-vectors. |
| 7455 | SmallVector<int, 16> Mask(WidenNumElts, -1); |
| 7456 | std::iota(first: Mask.begin(), last: Mask.begin() + VTNumElts, value: IdxVal); |
| 7457 | |
| 7458 | return DAG.getVectorShuffle(VT: WidenVT, dl, N1: ReverseVal, N2: DAG.getPOISON(VT: WidenVT), |
| 7459 | Mask); |
| 7460 | } |
| 7461 | |
| 7462 | SDValue DAGTypeLegalizer::WidenVecRes_GET_ACTIVE_LANE_MASK(SDNode *N) { |
| 7463 | EVT NVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 7464 | return DAG.getNode(Opcode: ISD::GET_ACTIVE_LANE_MASK, DL: SDLoc(N), VT: NVT, Ops: N->ops()); |
| 7465 | } |
| 7466 | |
| 7467 | void DAGTypeLegalizer::WidenVecRes_VECTOR_DEINTERLEAVE(SDNode *N) { |
| 7468 | EVT VT = N->getValueType(ResNo: 0); |
| 7469 | EVT EltVT = VT.getVectorElementType(); |
| 7470 | ElementCount OrigEC = VT.getVectorElementCount(); |
| 7471 | unsigned Factor = N->getNumOperands(); |
| 7472 | SDLoc DL(N); |
| 7473 | |
| 7474 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT); |
| 7475 | ElementCount WidenEC = WidenVT.getVectorElementCount(); |
| 7476 | // We cannot just use the widened operands directly: since they might be |
| 7477 | // individually widened, using them directly will result in de-interleaving |
| 7478 | // the "padded" lanes that sit in the middle of the vector. Instead, we should |
| 7479 | // not concat the widened operands but the original ones to effectively |
| 7480 | // generate a "packed" concated and widened vector, before extracting new |
| 7481 | // operand vectors with the widened type. |
| 7482 | EVT PackedWidenVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: EltVT, |
| 7483 | EC: WidenEC.multiplyCoefficientBy(RHS: Factor)); |
| 7484 | EVT ConcatVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: EltVT, |
| 7485 | EC: OrigEC.multiplyCoefficientBy(RHS: Factor)); |
| 7486 | SDValue ConcatOp = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: ConcatVT, Ops: N->ops()); |
| 7487 | SDValue PackedWidenVec = DAG.getInsertSubvector( |
| 7488 | DL, Vec: DAG.getUNDEF(VT: PackedWidenVT), SubVec: ConcatOp, /*Idx=*/0U); |
| 7489 | |
| 7490 | // Extract the new widened operand vectors. |
| 7491 | SmallVector<SDValue, 8> NewOps(Factor, SDValue()); |
| 7492 | for (unsigned Idx = 0U; Idx < Factor; ++Idx) { |
| 7493 | NewOps[Idx] = DAG.getExtractSubvector( |
| 7494 | DL, VT: WidenVT, Vec: PackedWidenVec, |
| 7495 | Idx: WidenEC.multiplyCoefficientBy(RHS: Idx).getKnownMinValue()); |
| 7496 | } |
| 7497 | |
| 7498 | SmallVector<EVT, 8> NewVTs(Factor, WidenVT); |
| 7499 | SDValue NewRes = DAG.getNode(Opcode: ISD::VECTOR_DEINTERLEAVE, DL, ResultTys: NewVTs, Ops: NewOps); |
| 7500 | // Set the widened results manually. |
| 7501 | for (unsigned Idx = 0U; Idx < Factor; ++Idx) |
| 7502 | SetWidenedVector(Op: SDValue(N, Idx), Result: NewRes.getValue(R: Idx)); |
| 7503 | } |
| 7504 | |
| 7505 | SDValue DAGTypeLegalizer::WidenVecRes_SETCC(SDNode *N) { |
| 7506 | assert(N->getValueType(0).isVector() && |
| 7507 | N->getOperand(0).getValueType().isVector() && |
| 7508 | "Operands must be vectors" ); |
| 7509 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: N->getValueType(ResNo: 0)); |
| 7510 | ElementCount WidenEC = WidenVT.getVectorElementCount(); |
| 7511 | |
| 7512 | SDValue InOp1 = N->getOperand(Num: 0); |
| 7513 | EVT InVT = InOp1.getValueType(); |
| 7514 | assert(InVT.isVector() && "can not widen non-vector type" ); |
| 7515 | EVT WidenInVT = |
| 7516 | EVT::getVectorVT(Context&: *DAG.getContext(), VT: InVT.getVectorElementType(), EC: WidenEC); |
| 7517 | |
| 7518 | // The input and output types often differ here, and it could be that while |
| 7519 | // we'd prefer to widen the result type, the input operands have been split. |
| 7520 | // In this case, we also need to split the result of this node as well. |
| 7521 | if (getTypeAction(VT: InVT) == TargetLowering::TypeSplitVector) { |
| 7522 | SDValue SplitVSetCC = SplitVecOp_VSETCC(N); |
| 7523 | SDValue Res = ModifyToType(InOp: SplitVSetCC, NVT: WidenVT); |
| 7524 | return Res; |
| 7525 | } |
| 7526 | |
| 7527 | // If the inputs also widen, handle them directly. Otherwise widen by hand. |
| 7528 | SDValue InOp2 = N->getOperand(Num: 1); |
| 7529 | if (getTypeAction(VT: InVT) == TargetLowering::TypeWidenVector) { |
| 7530 | InOp1 = GetWidenedVector(Op: InOp1); |
| 7531 | InOp2 = GetWidenedVector(Op: InOp2); |
| 7532 | } else { |
| 7533 | SDValue Poison = DAG.getPOISON(VT: WidenInVT); |
| 7534 | SDValue ZeroIdx = DAG.getVectorIdxConstant(Val: 0, DL: SDLoc(N)); |
| 7535 | InOp1 = DAG.getNode(Opcode: ISD::INSERT_SUBVECTOR, DL: SDLoc(N), VT: WidenInVT, N1: Poison, |
| 7536 | N2: InOp1, N3: ZeroIdx); |
| 7537 | InOp2 = DAG.getNode(Opcode: ISD::INSERT_SUBVECTOR, DL: SDLoc(N), VT: WidenInVT, N1: Poison, |
| 7538 | N2: InOp2, N3: ZeroIdx); |
| 7539 | } |
| 7540 | |
| 7541 | // Assume that the input and output will be widen appropriately. If not, |
| 7542 | // we will have to unroll it at some point. |
| 7543 | assert(InOp1.getValueType() == WidenInVT && |
| 7544 | InOp2.getValueType() == WidenInVT && |
| 7545 | "Input not widened to expected type!" ); |
| 7546 | (void)WidenInVT; |
| 7547 | if (N->getOpcode() == ISD::VP_SETCC) { |
| 7548 | SDValue Mask = |
| 7549 | GetWidenedMask(Mask: N->getOperand(Num: 3), EC: WidenVT.getVectorElementCount()); |
| 7550 | return DAG.getNode(Opcode: ISD::VP_SETCC, DL: SDLoc(N), VT: WidenVT, N1: InOp1, N2: InOp2, |
| 7551 | N3: N->getOperand(Num: 2), N4: Mask, N5: N->getOperand(Num: 4)); |
| 7552 | } |
| 7553 | return DAG.getNode(Opcode: ISD::SETCC, DL: SDLoc(N), VT: WidenVT, N1: InOp1, N2: InOp2, |
| 7554 | N3: N->getOperand(Num: 2)); |
| 7555 | } |
| 7556 | |
| 7557 | SDValue DAGTypeLegalizer::WidenVecRes_STRICT_FSETCC(SDNode *N) { |
| 7558 | assert(N->getValueType(0).isVector() && |
| 7559 | N->getOperand(1).getValueType().isVector() && |
| 7560 | "Operands must be vectors" ); |
| 7561 | EVT VT = N->getValueType(ResNo: 0); |
| 7562 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT); |
| 7563 | unsigned WidenNumElts = WidenVT.getVectorNumElements(); |
| 7564 | unsigned NumElts = VT.getVectorNumElements(); |
| 7565 | EVT EltVT = VT.getVectorElementType(); |
| 7566 | |
| 7567 | SDLoc dl(N); |
| 7568 | SDValue Chain = N->getOperand(Num: 0); |
| 7569 | SDValue LHS = N->getOperand(Num: 1); |
| 7570 | SDValue RHS = N->getOperand(Num: 2); |
| 7571 | SDValue CC = N->getOperand(Num: 3); |
| 7572 | EVT TmpEltVT = LHS.getValueType().getVectorElementType(); |
| 7573 | |
| 7574 | // Fully unroll and reassemble. |
| 7575 | SmallVector<SDValue, 8> Scalars(WidenNumElts, DAG.getPOISON(VT: EltVT)); |
| 7576 | SmallVector<SDValue, 8> Chains(NumElts); |
| 7577 | for (unsigned i = 0; i != NumElts; ++i) { |
| 7578 | SDValue LHSElem = DAG.getExtractVectorElt(DL: dl, VT: TmpEltVT, Vec: LHS, Idx: i); |
| 7579 | SDValue RHSElem = DAG.getExtractVectorElt(DL: dl, VT: TmpEltVT, Vec: RHS, Idx: i); |
| 7580 | |
| 7581 | Scalars[i] = DAG.getNode(Opcode: N->getOpcode(), DL: dl, ResultTys: {MVT::i1, MVT::Other}, |
| 7582 | Ops: {Chain, LHSElem, RHSElem, CC}); |
| 7583 | Chains[i] = Scalars[i].getValue(R: 1); |
| 7584 | Scalars[i] = DAG.getSelect(DL: dl, VT: EltVT, Cond: Scalars[i], |
| 7585 | LHS: DAG.getBoolConstant(V: true, DL: dl, VT: EltVT, OpVT: VT), |
| 7586 | RHS: DAG.getBoolConstant(V: false, DL: dl, VT: EltVT, OpVT: VT)); |
| 7587 | } |
| 7588 | |
| 7589 | SDValue NewChain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: Chains); |
| 7590 | ReplaceValueWith(From: SDValue(N, 1), To: NewChain); |
| 7591 | |
| 7592 | return DAG.getBuildVector(VT: WidenVT, DL: dl, Ops: Scalars); |
| 7593 | } |
| 7594 | |
| 7595 | //===----------------------------------------------------------------------===// |
| 7596 | // Widen Vector Operand |
| 7597 | //===----------------------------------------------------------------------===// |
| 7598 | bool DAGTypeLegalizer::WidenVectorOperand(SDNode *N, unsigned OpNo) { |
| 7599 | LLVM_DEBUG(dbgs() << "Widen node operand " << OpNo << ": " ; N->dump(&DAG)); |
| 7600 | SDValue Res = SDValue(); |
| 7601 | |
| 7602 | // See if the target wants to custom widen this node. |
| 7603 | if (CustomLowerNode(N, VT: N->getOperand(Num: OpNo).getValueType(), LegalizeResult: false)) |
| 7604 | return false; |
| 7605 | |
| 7606 | switch (N->getOpcode()) { |
| 7607 | default: |
| 7608 | #ifndef NDEBUG |
| 7609 | dbgs() << "WidenVectorOperand op #" << OpNo << ": " ; |
| 7610 | N->dump(&DAG); |
| 7611 | dbgs() << "\n" ; |
| 7612 | #endif |
| 7613 | report_fatal_error(reason: "Do not know how to widen this operator's operand!" ); |
| 7614 | |
| 7615 | case ISD::BITCAST: Res = WidenVecOp_BITCAST(N); break; |
| 7616 | case ISD::FAKE_USE: |
| 7617 | Res = WidenVecOp_FAKE_USE(N); |
| 7618 | break; |
| 7619 | case ISD::CONCAT_VECTORS: Res = WidenVecOp_CONCAT_VECTORS(N); break; |
| 7620 | case ISD::INSERT_SUBVECTOR: Res = WidenVecOp_INSERT_SUBVECTOR(N); break; |
| 7621 | case ISD::EXTRACT_SUBVECTOR: Res = WidenVecOp_EXTRACT_SUBVECTOR(N); break; |
| 7622 | case ISD::EXTRACT_VECTOR_ELT: Res = WidenVecOp_EXTRACT_VECTOR_ELT(N); break; |
| 7623 | case ISD::STORE: Res = WidenVecOp_STORE(N); break; |
| 7624 | case ISD::ATOMIC_STORE: |
| 7625 | Res = WidenVecOp_ATOMIC_STORE(ST: cast<AtomicSDNode>(Val: N)); |
| 7626 | break; |
| 7627 | case ISD::VP_STORE: Res = WidenVecOp_VP_STORE(N, OpNo); break; |
| 7628 | case ISD::EXPERIMENTAL_VP_STRIDED_STORE: |
| 7629 | Res = WidenVecOp_VP_STRIDED_STORE(N, OpNo); |
| 7630 | break; |
| 7631 | case ISD::ANY_EXTEND_VECTOR_INREG: |
| 7632 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
| 7633 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
| 7634 | Res = WidenVecOp_EXTEND_VECTOR_INREG(N); |
| 7635 | break; |
| 7636 | case ISD::MSTORE: Res = WidenVecOp_MSTORE(N, OpNo); break; |
| 7637 | case ISD::MGATHER: Res = WidenVecOp_MGATHER(N, OpNo); break; |
| 7638 | case ISD::MSCATTER: Res = WidenVecOp_MSCATTER(N, OpNo); break; |
| 7639 | case ISD::VP_SCATTER: Res = WidenVecOp_VP_SCATTER(N, OpNo); break; |
| 7640 | case ISD::SETCC: Res = WidenVecOp_SETCC(N); break; |
| 7641 | case ISD::STRICT_FSETCC: |
| 7642 | case ISD::STRICT_FSETCCS: Res = WidenVecOp_STRICT_FSETCC(N); break; |
| 7643 | case ISD::VSELECT: Res = WidenVecOp_VSELECT(N); break; |
| 7644 | case ISD::FLDEXP: |
| 7645 | case ISD::FCOPYSIGN: |
| 7646 | case ISD::LROUND: |
| 7647 | case ISD::LLROUND: |
| 7648 | case ISD::LRINT: |
| 7649 | case ISD::LLRINT: |
| 7650 | Res = WidenVecOp_UnrollVectorOp(N); |
| 7651 | break; |
| 7652 | case ISD::IS_FPCLASS: Res = WidenVecOp_IS_FPCLASS(N); break; |
| 7653 | |
| 7654 | case ISD::ANY_EXTEND: |
| 7655 | case ISD::SIGN_EXTEND: |
| 7656 | case ISD::ZERO_EXTEND: |
| 7657 | Res = WidenVecOp_EXTEND(N); |
| 7658 | break; |
| 7659 | |
| 7660 | case ISD::SCMP: |
| 7661 | case ISD::UCMP: |
| 7662 | Res = WidenVecOp_CMP(N); |
| 7663 | break; |
| 7664 | |
| 7665 | case ISD::FP_EXTEND: |
| 7666 | case ISD::STRICT_FP_EXTEND: |
| 7667 | case ISD::FP_ROUND: |
| 7668 | case ISD::STRICT_FP_ROUND: |
| 7669 | case ISD::FP_TO_SINT: |
| 7670 | case ISD::STRICT_FP_TO_SINT: |
| 7671 | case ISD::FP_TO_UINT: |
| 7672 | case ISD::STRICT_FP_TO_UINT: |
| 7673 | case ISD::SINT_TO_FP: |
| 7674 | case ISD::STRICT_SINT_TO_FP: |
| 7675 | case ISD::UINT_TO_FP: |
| 7676 | case ISD::STRICT_UINT_TO_FP: |
| 7677 | case ISD::TRUNCATE: |
| 7678 | case ISD::CONVERT_FROM_ARBITRARY_FP: |
| 7679 | case ISD::CONVERT_TO_ARBITRARY_FP: |
| 7680 | Res = WidenVecOp_Convert(N); |
| 7681 | break; |
| 7682 | |
| 7683 | case ISD::FP_TO_SINT_SAT: |
| 7684 | case ISD::FP_TO_UINT_SAT: |
| 7685 | Res = WidenVecOp_FP_TO_XINT_SAT(N); |
| 7686 | break; |
| 7687 | |
| 7688 | case ISD::VECREDUCE_FADD: |
| 7689 | case ISD::VECREDUCE_FMUL: |
| 7690 | case ISD::VECREDUCE_ADD: |
| 7691 | case ISD::VECREDUCE_MUL: |
| 7692 | case ISD::VECREDUCE_AND: |
| 7693 | case ISD::VECREDUCE_OR: |
| 7694 | case ISD::VECREDUCE_XOR: |
| 7695 | case ISD::VECREDUCE_SMAX: |
| 7696 | case ISD::VECREDUCE_SMIN: |
| 7697 | case ISD::VECREDUCE_UMAX: |
| 7698 | case ISD::VECREDUCE_UMIN: |
| 7699 | case ISD::VECREDUCE_FMAX: |
| 7700 | case ISD::VECREDUCE_FMIN: |
| 7701 | case ISD::VECREDUCE_FMAXIMUM: |
| 7702 | case ISD::VECREDUCE_FMINIMUM: |
| 7703 | Res = WidenVecOp_VECREDUCE(N); |
| 7704 | break; |
| 7705 | case ISD::VECREDUCE_SEQ_FADD: |
| 7706 | case ISD::VECREDUCE_SEQ_FMUL: |
| 7707 | Res = WidenVecOp_VECREDUCE_SEQ(N); |
| 7708 | break; |
| 7709 | case ISD::VP_REDUCE_FADD: |
| 7710 | case ISD::VP_REDUCE_SEQ_FADD: |
| 7711 | case ISD::VP_REDUCE_FMUL: |
| 7712 | case ISD::VP_REDUCE_SEQ_FMUL: |
| 7713 | case ISD::VP_REDUCE_ADD: |
| 7714 | case ISD::VP_REDUCE_MUL: |
| 7715 | case ISD::VP_REDUCE_AND: |
| 7716 | case ISD::VP_REDUCE_OR: |
| 7717 | case ISD::VP_REDUCE_XOR: |
| 7718 | case ISD::VP_REDUCE_SMAX: |
| 7719 | case ISD::VP_REDUCE_SMIN: |
| 7720 | case ISD::VP_REDUCE_UMAX: |
| 7721 | case ISD::VP_REDUCE_UMIN: |
| 7722 | case ISD::VP_REDUCE_FMAX: |
| 7723 | case ISD::VP_REDUCE_FMIN: |
| 7724 | case ISD::VP_REDUCE_FMAXIMUM: |
| 7725 | case ISD::VP_REDUCE_FMINIMUM: |
| 7726 | Res = WidenVecOp_VP_REDUCE(N); |
| 7727 | break; |
| 7728 | case ISD::CTTZ_ELTS: |
| 7729 | case ISD::CTTZ_ELTS_ZERO_POISON: |
| 7730 | Res = WidenVecOp_CttzElements(N); |
| 7731 | break; |
| 7732 | case ISD::VP_CTTZ_ELTS: |
| 7733 | case ISD::VP_CTTZ_ELTS_ZERO_POISON: |
| 7734 | Res = WidenVecOp_VP_CttzElements(N); |
| 7735 | break; |
| 7736 | case ISD::VECTOR_FIND_LAST_ACTIVE: |
| 7737 | Res = WidenVecOp_VECTOR_FIND_LAST_ACTIVE(N); |
| 7738 | break; |
| 7739 | } |
| 7740 | |
| 7741 | // If Res is null, the sub-method took care of registering the result. |
| 7742 | if (!Res.getNode()) return false; |
| 7743 | |
| 7744 | // If the result is N, the sub-method updated N in place. Tell the legalizer |
| 7745 | // core about this. |
| 7746 | if (Res.getNode() == N) |
| 7747 | return true; |
| 7748 | |
| 7749 | |
| 7750 | if (N->isStrictFPOpcode()) |
| 7751 | assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 2 && |
| 7752 | "Invalid operand expansion" ); |
| 7753 | else |
| 7754 | assert(Res.getValueType() == N->getValueType(0) && N->getNumValues() == 1 && |
| 7755 | "Invalid operand expansion" ); |
| 7756 | |
| 7757 | ReplaceValueWith(From: SDValue(N, 0), To: Res); |
| 7758 | return false; |
| 7759 | } |
| 7760 | |
| 7761 | SDValue DAGTypeLegalizer::WidenVecOp_EXTEND(SDNode *N) { |
| 7762 | SDLoc DL(N); |
| 7763 | EVT VT = N->getValueType(ResNo: 0); |
| 7764 | |
| 7765 | SDValue InOp = N->getOperand(Num: 0); |
| 7766 | assert(getTypeAction(InOp.getValueType()) == |
| 7767 | TargetLowering::TypeWidenVector && |
| 7768 | "Unexpected type action" ); |
| 7769 | InOp = GetWidenedVector(Op: InOp); |
| 7770 | assert(VT.getVectorNumElements() < |
| 7771 | InOp.getValueType().getVectorNumElements() && |
| 7772 | "Input wasn't widened!" ); |
| 7773 | |
| 7774 | // We may need to further widen the operand until it has the same total |
| 7775 | // vector size as the result. |
| 7776 | EVT InVT = InOp.getValueType(); |
| 7777 | if (InVT.getSizeInBits() != VT.getSizeInBits()) { |
| 7778 | EVT InEltVT = InVT.getVectorElementType(); |
| 7779 | for (EVT FixedVT : MVT::vector_valuetypes()) { |
| 7780 | EVT FixedEltVT = FixedVT.getVectorElementType(); |
| 7781 | if (TLI.isTypeLegal(VT: FixedVT) && |
| 7782 | FixedVT.getSizeInBits() == VT.getSizeInBits() && |
| 7783 | FixedEltVT == InEltVT) { |
| 7784 | assert(FixedVT.getVectorNumElements() >= VT.getVectorNumElements() && |
| 7785 | "Not enough elements in the fixed type for the operand!" ); |
| 7786 | assert(FixedVT.getVectorNumElements() != InVT.getVectorNumElements() && |
| 7787 | "We can't have the same type as we started with!" ); |
| 7788 | if (FixedVT.getVectorNumElements() > InVT.getVectorNumElements()) |
| 7789 | InOp = DAG.getInsertSubvector(DL, Vec: DAG.getPOISON(VT: FixedVT), SubVec: InOp, Idx: 0); |
| 7790 | else |
| 7791 | InOp = DAG.getExtractSubvector(DL, VT: FixedVT, Vec: InOp, Idx: 0); |
| 7792 | break; |
| 7793 | } |
| 7794 | } |
| 7795 | InVT = InOp.getValueType(); |
| 7796 | if (InVT.getSizeInBits() != VT.getSizeInBits()) |
| 7797 | // We couldn't find a legal vector type that was a widening of the input |
| 7798 | // and could be extended in-register to the result type, so we have to |
| 7799 | // scalarize. |
| 7800 | return WidenVecOp_Convert(N); |
| 7801 | } |
| 7802 | |
| 7803 | // Use special DAG nodes to represent the operation of extending the |
| 7804 | // low lanes. |
| 7805 | switch (N->getOpcode()) { |
| 7806 | default: |
| 7807 | llvm_unreachable("Extend legalization on extend operation!" ); |
| 7808 | case ISD::ANY_EXTEND: |
| 7809 | return DAG.getNode(Opcode: ISD::ANY_EXTEND_VECTOR_INREG, DL, VT, Operand: InOp); |
| 7810 | case ISD::SIGN_EXTEND: |
| 7811 | return DAG.getNode(Opcode: ISD::SIGN_EXTEND_VECTOR_INREG, DL, VT, Operand: InOp); |
| 7812 | case ISD::ZERO_EXTEND: |
| 7813 | return DAG.getNode(Opcode: ISD::ZERO_EXTEND_VECTOR_INREG, DL, VT, Operand: InOp); |
| 7814 | } |
| 7815 | } |
| 7816 | |
| 7817 | SDValue DAGTypeLegalizer::WidenVecOp_CMP(SDNode *N) { |
| 7818 | SDLoc dl(N); |
| 7819 | |
| 7820 | EVT OpVT = N->getOperand(Num: 0).getValueType(); |
| 7821 | EVT ResVT = N->getValueType(ResNo: 0); |
| 7822 | SDValue LHS = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 7823 | SDValue RHS = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 7824 | |
| 7825 | // 1. EXTRACT_SUBVECTOR |
| 7826 | // 2. SIGN_EXTEND/ZERO_EXTEND |
| 7827 | // 3. CMP |
| 7828 | LHS = DAG.getExtractSubvector(DL: dl, VT: OpVT, Vec: LHS, Idx: 0); |
| 7829 | RHS = DAG.getExtractSubvector(DL: dl, VT: OpVT, Vec: RHS, Idx: 0); |
| 7830 | |
| 7831 | // At this point the result type is guaranteed to be valid, so we can use it |
| 7832 | // as the operand type by extending it appropriately |
| 7833 | ISD::NodeType ExtendOpcode = |
| 7834 | N->getOpcode() == ISD::SCMP ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; |
| 7835 | LHS = DAG.getNode(Opcode: ExtendOpcode, DL: dl, VT: ResVT, Operand: LHS); |
| 7836 | RHS = DAG.getNode(Opcode: ExtendOpcode, DL: dl, VT: ResVT, Operand: RHS); |
| 7837 | |
| 7838 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: ResVT, N1: LHS, N2: RHS); |
| 7839 | } |
| 7840 | |
| 7841 | SDValue DAGTypeLegalizer::WidenVecOp_UnrollVectorOp(SDNode *N) { |
| 7842 | // The result (and first input) is legal, but the second input is illegal. |
| 7843 | // We can't do much to fix that, so just unroll and let the extracts off of |
| 7844 | // the second input be widened as needed later. |
| 7845 | return DAG.UnrollVectorOp(N); |
| 7846 | } |
| 7847 | |
| 7848 | SDValue DAGTypeLegalizer::WidenVecOp_IS_FPCLASS(SDNode *N) { |
| 7849 | SDLoc DL(N); |
| 7850 | EVT ResultVT = N->getValueType(ResNo: 0); |
| 7851 | SDValue Test = N->getOperand(Num: 1); |
| 7852 | SDValue WideArg = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 7853 | |
| 7854 | // Process this node similarly to SETCC. |
| 7855 | EVT WideResultVT = getSetCCResultType(VT: WideArg.getValueType()); |
| 7856 | if (ResultVT.getScalarType() == MVT::i1) |
| 7857 | WideResultVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: MVT::i1, |
| 7858 | NumElements: WideResultVT.getVectorNumElements()); |
| 7859 | |
| 7860 | SDValue WideNode = DAG.getNode(Opcode: ISD::IS_FPCLASS, DL, VT: WideResultVT, |
| 7861 | Ops: {WideArg, Test}, Flags: N->getFlags()); |
| 7862 | |
| 7863 | // Extract the needed results from the result vector. |
| 7864 | EVT ResVT = |
| 7865 | EVT::getVectorVT(Context&: *DAG.getContext(), VT: WideResultVT.getVectorElementType(), |
| 7866 | NumElements: ResultVT.getVectorNumElements()); |
| 7867 | SDValue CC = DAG.getExtractSubvector(DL, VT: ResVT, Vec: WideNode, Idx: 0); |
| 7868 | |
| 7869 | EVT OpVT = N->getOperand(Num: 0).getValueType(); |
| 7870 | ISD::NodeType ExtendCode = |
| 7871 | TargetLowering::getExtendForContent(Content: TLI.getBooleanContents(Type: OpVT)); |
| 7872 | return DAG.getNode(Opcode: ExtendCode, DL, VT: ResultVT, Operand: CC); |
| 7873 | } |
| 7874 | |
| 7875 | SDValue DAGTypeLegalizer::WidenVecOp_Convert(SDNode *N) { |
| 7876 | // Since the result is legal and the input is illegal. |
| 7877 | EVT VT = N->getValueType(ResNo: 0); |
| 7878 | EVT EltVT = VT.getVectorElementType(); |
| 7879 | SDLoc dl(N); |
| 7880 | SDValue InOp = N->getOperand(Num: N->isStrictFPOpcode() ? 1 : 0); |
| 7881 | assert(getTypeAction(InOp.getValueType()) == |
| 7882 | TargetLowering::TypeWidenVector && |
| 7883 | "Unexpected type action" ); |
| 7884 | InOp = GetWidenedVector(Op: InOp); |
| 7885 | EVT InVT = InOp.getValueType(); |
| 7886 | unsigned Opcode = N->getOpcode(); |
| 7887 | |
| 7888 | // Helper to build a convert node with all scalar trailing operands. |
| 7889 | auto MakeConvertNode = [&](EVT VT, SDValue Op) -> SDValue { |
| 7890 | if (Opcode == ISD::CONVERT_TO_ARBITRARY_FP) |
| 7891 | return DAG.getNode(Opcode, DL: dl, VT, N1: Op, N2: N->getOperand(Num: 1), N3: N->getOperand(Num: 2), |
| 7892 | N4: N->getOperand(Num: 3)); |
| 7893 | if (Opcode == ISD::FP_ROUND || Opcode == ISD::CONVERT_FROM_ARBITRARY_FP) |
| 7894 | return DAG.getNode(Opcode, DL: dl, VT, N1: Op, N2: N->getOperand(Num: 1)); |
| 7895 | return DAG.getNode(Opcode, DL: dl, VT, Operand: Op); |
| 7896 | }; |
| 7897 | |
| 7898 | // See if a widened result type would be legal, if so widen the node. |
| 7899 | // FIXME: This isn't safe for StrictFP. Other optimization here is needed. |
| 7900 | EVT WideVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: EltVT, |
| 7901 | EC: InVT.getVectorElementCount()); |
| 7902 | if (TLI.isTypeLegal(VT: WideVT) && !N->isStrictFPOpcode()) { |
| 7903 | SDValue Res; |
| 7904 | if (N->isStrictFPOpcode()) { |
| 7905 | if (Opcode == ISD::STRICT_FP_ROUND) |
| 7906 | Res = DAG.getNode(Opcode, DL: dl, ResultTys: { WideVT, MVT::Other }, |
| 7907 | Ops: { N->getOperand(Num: 0), InOp, N->getOperand(Num: 2) }); |
| 7908 | else |
| 7909 | Res = DAG.getNode(Opcode, DL: dl, ResultTys: { WideVT, MVT::Other }, |
| 7910 | Ops: { N->getOperand(Num: 0), InOp }); |
| 7911 | // Legalize the chain result - switch anything that used the old chain to |
| 7912 | // use the new one. |
| 7913 | ReplaceValueWith(From: SDValue(N, 1), To: Res.getValue(R: 1)); |
| 7914 | } else { |
| 7915 | Res = MakeConvertNode(WideVT, InOp); |
| 7916 | } |
| 7917 | return DAG.getExtractSubvector(DL: dl, VT, Vec: Res, Idx: 0); |
| 7918 | } |
| 7919 | |
| 7920 | EVT InEltVT = InVT.getVectorElementType(); |
| 7921 | |
| 7922 | // Unroll the convert into some scalar code and create a nasty build vector. |
| 7923 | unsigned NumElts = VT.getVectorNumElements(); |
| 7924 | SmallVector<SDValue, 16> Ops(NumElts); |
| 7925 | if (N->isStrictFPOpcode()) { |
| 7926 | SmallVector<SDValue, 4> NewOps(N->ops()); |
| 7927 | SmallVector<SDValue, 32> OpChains; |
| 7928 | for (unsigned i=0; i < NumElts; ++i) { |
| 7929 | NewOps[1] = DAG.getExtractVectorElt(DL: dl, VT: InEltVT, Vec: InOp, Idx: i); |
| 7930 | Ops[i] = DAG.getNode(Opcode, DL: dl, ResultTys: { EltVT, MVT::Other }, Ops: NewOps); |
| 7931 | OpChains.push_back(Elt: Ops[i].getValue(R: 1)); |
| 7932 | } |
| 7933 | SDValue NewChain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: OpChains); |
| 7934 | ReplaceValueWith(From: SDValue(N, 1), To: NewChain); |
| 7935 | } else { |
| 7936 | for (unsigned i = 0; i < NumElts; ++i) { |
| 7937 | SDValue Elt = DAG.getExtractVectorElt(DL: dl, VT: InEltVT, Vec: InOp, Idx: i); |
| 7938 | Ops[i] = MakeConvertNode(EltVT, Elt); |
| 7939 | } |
| 7940 | } |
| 7941 | |
| 7942 | return DAG.getBuildVector(VT, DL: dl, Ops); |
| 7943 | } |
| 7944 | |
| 7945 | SDValue DAGTypeLegalizer::WidenVecOp_FP_TO_XINT_SAT(SDNode *N) { |
| 7946 | EVT DstVT = N->getValueType(ResNo: 0); |
| 7947 | SDValue Src = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 7948 | EVT SrcVT = Src.getValueType(); |
| 7949 | ElementCount WideNumElts = SrcVT.getVectorElementCount(); |
| 7950 | SDLoc dl(N); |
| 7951 | |
| 7952 | // See if a widened result type would be legal, if so widen the node. |
| 7953 | EVT WideDstVT = EVT::getVectorVT(Context&: *DAG.getContext(), |
| 7954 | VT: DstVT.getVectorElementType(), EC: WideNumElts); |
| 7955 | if (TLI.isTypeLegal(VT: WideDstVT)) { |
| 7956 | SDValue Res = |
| 7957 | DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: WideDstVT, N1: Src, N2: N->getOperand(Num: 1)); |
| 7958 | return DAG.getNode( |
| 7959 | Opcode: ISD::EXTRACT_SUBVECTOR, DL: dl, VT: DstVT, N1: Res, |
| 7960 | N2: DAG.getConstant(Val: 0, DL: dl, VT: TLI.getVectorIdxTy(DL: DAG.getDataLayout()))); |
| 7961 | } |
| 7962 | |
| 7963 | // Give up and unroll. |
| 7964 | return DAG.UnrollVectorOp(N); |
| 7965 | } |
| 7966 | |
| 7967 | SDValue DAGTypeLegalizer::WidenVecOp_BITCAST(SDNode *N) { |
| 7968 | EVT VT = N->getValueType(ResNo: 0); |
| 7969 | SDValue InOp = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 7970 | EVT InWidenVT = InOp.getValueType(); |
| 7971 | SDLoc dl(N); |
| 7972 | |
| 7973 | // Check if we can convert between two legal vector types and extract. |
| 7974 | TypeSize InWidenSize = InWidenVT.getSizeInBits(); |
| 7975 | TypeSize Size = VT.getSizeInBits(); |
| 7976 | // x86mmx is not an acceptable vector element type, so don't try. |
| 7977 | if (!VT.isVector() && VT != MVT::x86mmx && |
| 7978 | InWidenSize.hasKnownScalarFactor(RHS: Size)) { |
| 7979 | unsigned NewNumElts = InWidenSize.getKnownScalarFactor(RHS: Size); |
| 7980 | EVT NewVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT, NumElements: NewNumElts); |
| 7981 | if (TLI.isTypeLegal(VT: NewVT)) { |
| 7982 | SDValue BitOp = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: NewVT, Operand: InOp); |
| 7983 | return DAG.getExtractVectorElt(DL: dl, VT, Vec: BitOp, Idx: 0); |
| 7984 | } |
| 7985 | } |
| 7986 | |
| 7987 | // Handle a case like bitcast v12i8 -> v3i32. Normally that would get widened |
| 7988 | // to v16i8 -> v4i32, but for a target where v3i32 is legal but v12i8 is not, |
| 7989 | // we end up here. Handling the case here with EXTRACT_SUBVECTOR avoids |
| 7990 | // having to copy via memory. |
| 7991 | if (VT.isVector()) { |
| 7992 | EVT EltVT = VT.getVectorElementType(); |
| 7993 | unsigned EltSize = EltVT.getFixedSizeInBits(); |
| 7994 | if (InWidenSize.isKnownMultipleOf(RHS: EltSize)) { |
| 7995 | ElementCount NewNumElts = |
| 7996 | (InWidenVT.getVectorElementCount() * InWidenVT.getScalarSizeInBits()) |
| 7997 | .divideCoefficientBy(RHS: EltSize); |
| 7998 | EVT NewVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: EltVT, EC: NewNumElts); |
| 7999 | if (TLI.isTypeLegal(VT: NewVT)) { |
| 8000 | SDValue BitOp = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: NewVT, Operand: InOp); |
| 8001 | return DAG.getExtractSubvector(DL: dl, VT, Vec: BitOp, Idx: 0); |
| 8002 | } |
| 8003 | } |
| 8004 | } |
| 8005 | |
| 8006 | return CreateStackStoreLoad(Op: InOp, DestVT: VT); |
| 8007 | } |
| 8008 | |
| 8009 | // Vectors with sizes that are not powers of 2 need to be widened to the |
| 8010 | // next largest power of 2. For example, we may get a vector of 3 32-bit |
| 8011 | // integers or of 6 16-bit integers, both of which have to be widened to a |
| 8012 | // 128-bit vector. |
| 8013 | SDValue DAGTypeLegalizer::WidenVecOp_FAKE_USE(SDNode *N) { |
| 8014 | SDValue WidenedOp = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 8015 | return DAG.getNode(Opcode: ISD::FAKE_USE, DL: SDLoc(), VT: MVT::Other, N1: N->getOperand(Num: 0), |
| 8016 | N2: WidenedOp); |
| 8017 | } |
| 8018 | |
| 8019 | SDValue DAGTypeLegalizer::WidenVecOp_CONCAT_VECTORS(SDNode *N) { |
| 8020 | EVT VT = N->getValueType(ResNo: 0); |
| 8021 | EVT EltVT = VT.getVectorElementType(); |
| 8022 | EVT InVT = N->getOperand(Num: 0).getValueType(); |
| 8023 | SDLoc dl(N); |
| 8024 | |
| 8025 | // If the widen width for this operand is the same as the width of the concat |
| 8026 | // and all but the first operand is undef, just use the widened operand. |
| 8027 | unsigned NumOperands = N->getNumOperands(); |
| 8028 | if (VT == TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: InVT)) { |
| 8029 | unsigned i; |
| 8030 | for (i = 1; i < NumOperands; ++i) |
| 8031 | if (!N->getOperand(Num: i).isUndef()) |
| 8032 | break; |
| 8033 | |
| 8034 | if (i == NumOperands) |
| 8035 | return GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 8036 | } |
| 8037 | |
| 8038 | // Otherwise, fall back to a nasty build vector. |
| 8039 | unsigned NumElts = VT.getVectorNumElements(); |
| 8040 | SmallVector<SDValue, 16> Ops(NumElts); |
| 8041 | |
| 8042 | unsigned NumInElts = InVT.getVectorNumElements(); |
| 8043 | |
| 8044 | unsigned Idx = 0; |
| 8045 | for (unsigned i=0; i < NumOperands; ++i) { |
| 8046 | SDValue InOp = N->getOperand(Num: i); |
| 8047 | assert(getTypeAction(InOp.getValueType()) == |
| 8048 | TargetLowering::TypeWidenVector && |
| 8049 | "Unexpected type action" ); |
| 8050 | InOp = GetWidenedVector(Op: InOp); |
| 8051 | for (unsigned j = 0; j < NumInElts; ++j) |
| 8052 | Ops[Idx++] = DAG.getExtractVectorElt(DL: dl, VT: EltVT, Vec: InOp, Idx: j); |
| 8053 | } |
| 8054 | return DAG.getBuildVector(VT, DL: dl, Ops); |
| 8055 | } |
| 8056 | |
| 8057 | SDValue DAGTypeLegalizer::WidenVecOp_INSERT_SUBVECTOR(SDNode *N) { |
| 8058 | EVT VT = N->getValueType(ResNo: 0); |
| 8059 | SDValue SubVec = N->getOperand(Num: 1); |
| 8060 | SDValue InVec = N->getOperand(Num: 0); |
| 8061 | |
| 8062 | EVT OrigVT = SubVec.getValueType(); |
| 8063 | SubVec = GetWidenedVector(Op: SubVec); |
| 8064 | EVT SubVT = SubVec.getValueType(); |
| 8065 | |
| 8066 | // Whether or not all the elements of the widened SubVec will be inserted into |
| 8067 | // valid indices of VT. |
| 8068 | bool IndicesValid = false; |
| 8069 | // If we statically know that VT can fit SubVT, the indices are valid. |
| 8070 | if (VT.knownBitsGE(VT: SubVT)) |
| 8071 | IndicesValid = true; |
| 8072 | else if (VT.isScalableVector() && SubVT.isFixedLengthVector()) { |
| 8073 | // Otherwise, if we're inserting a fixed vector into a scalable vector and |
| 8074 | // we know the minimum vscale we can work out if it's valid ourselves. |
| 8075 | Attribute Attr = DAG.getMachineFunction().getFunction().getFnAttribute( |
| 8076 | Kind: Attribute::VScaleRange); |
| 8077 | if (Attr.isValid()) { |
| 8078 | unsigned VScaleMin = Attr.getVScaleRangeMin(); |
| 8079 | if (VT.getSizeInBits().getKnownMinValue() * VScaleMin >= |
| 8080 | SubVT.getFixedSizeInBits()) |
| 8081 | IndicesValid = true; |
| 8082 | } |
| 8083 | } |
| 8084 | |
| 8085 | if (!IndicesValid) |
| 8086 | report_fatal_error( |
| 8087 | reason: "Don't know how to widen the operands for INSERT_SUBVECTOR" ); |
| 8088 | |
| 8089 | SDLoc DL(N); |
| 8090 | |
| 8091 | // We need to make sure that the indices are still valid, otherwise we might |
| 8092 | // widen what was previously well-defined to something undefined. |
| 8093 | if (InVec.isUndef() && N->getConstantOperandVal(Num: 2) == 0) |
| 8094 | return DAG.getNode(Opcode: ISD::INSERT_SUBVECTOR, DL, VT, N1: InVec, N2: SubVec, |
| 8095 | N3: N->getOperand(Num: 2)); |
| 8096 | |
| 8097 | if (OrigVT.isScalableVector()) { |
| 8098 | // When the widened types match, overwriting the start of a vector is |
| 8099 | // effectively a merge operation that can be implement as a vselect. |
| 8100 | if (SubVT == VT && N->getConstantOperandVal(Num: 2) == 0) { |
| 8101 | SDValue Mask = |
| 8102 | DAG.getMaskFromElementCount(DL, VT, Len: OrigVT.getVectorElementCount()); |
| 8103 | return DAG.getNode(Opcode: ISD::VSELECT, DL, VT, N1: Mask, N2: SubVec, N3: InVec); |
| 8104 | } |
| 8105 | |
| 8106 | // Fallback to inserting through memory. |
| 8107 | Align Alignment = DAG.getReducedAlign(VT, /*UseABI=*/false); |
| 8108 | SDValue StackPtr = DAG.CreateStackTemporary(Bytes: VT.getStoreSize(), Alignment); |
| 8109 | MachineFunction &MF = DAG.getMachineFunction(); |
| 8110 | int FrameIndex = cast<FrameIndexSDNode>(Val: StackPtr.getNode())->getIndex(); |
| 8111 | auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FI: FrameIndex); |
| 8112 | |
| 8113 | MachineMemOperand *StoreMMO = MF.getMachineMemOperand( |
| 8114 | PtrInfo, F: MachineMemOperand::MOStore, |
| 8115 | Size: LocationSize::beforeOrAfterPointer(), BaseAlignment: Alignment); |
| 8116 | MachineMemOperand *LoadMMO = MF.getMachineMemOperand( |
| 8117 | PtrInfo, F: MachineMemOperand::MOLoad, |
| 8118 | Size: LocationSize::beforeOrAfterPointer(), BaseAlignment: Alignment); |
| 8119 | |
| 8120 | // Write out the vector being inserting into. |
| 8121 | SDValue Ch = |
| 8122 | DAG.getStore(Chain: DAG.getEntryNode(), dl: DL, Val: InVec, Ptr: StackPtr, MMO: StoreMMO); |
| 8123 | |
| 8124 | // Build a mask to match the length of the sub-vector. |
| 8125 | SDValue Mask = |
| 8126 | DAG.getMaskFromElementCount(DL, VT: SubVT, Len: OrigVT.getVectorElementCount()); |
| 8127 | |
| 8128 | // Overwrite the sub-vector at the required offset. |
| 8129 | SDValue SubVecPtr = |
| 8130 | TLI.getVectorSubVecPointer(DAG, VecPtr: StackPtr, VecVT: VT, SubVecVT: OrigVT, Index: N->getOperand(Num: 2)); |
| 8131 | Ch = DAG.getMaskedStore(Chain: Ch, dl: DL, Val: SubVec, Base: SubVecPtr, |
| 8132 | Offset: DAG.getPOISON(VT: SubVecPtr.getValueType()), Mask, MemVT: VT, |
| 8133 | MMO: StoreMMO, AM: ISD::UNINDEXED, IsTruncating: ISD::NON_EXTLOAD); |
| 8134 | |
| 8135 | // Read back the result. |
| 8136 | return DAG.getLoad(VT, dl: DL, Chain: Ch, Ptr: StackPtr, MMO: LoadMMO); |
| 8137 | } |
| 8138 | |
| 8139 | // If the operands can't be widened legally, just replace the INSERT_SUBVECTOR |
| 8140 | // with a series of INSERT_VECTOR_ELT |
| 8141 | unsigned Idx = N->getConstantOperandVal(Num: 2); |
| 8142 | |
| 8143 | SDValue InsertElt = InVec; |
| 8144 | for (unsigned I = 0, E = OrigVT.getVectorNumElements(); I != E; ++I) { |
| 8145 | SDValue = |
| 8146 | DAG.getExtractVectorElt(DL, VT: VT.getVectorElementType(), Vec: SubVec, Idx: I); |
| 8147 | InsertElt = DAG.getInsertVectorElt(DL, Vec: InsertElt, Elt: ExtractElt, Idx: I + Idx); |
| 8148 | } |
| 8149 | |
| 8150 | return InsertElt; |
| 8151 | } |
| 8152 | |
| 8153 | SDValue DAGTypeLegalizer::(SDNode *N) { |
| 8154 | SDValue InOp = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 8155 | return DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL: SDLoc(N), |
| 8156 | VT: N->getValueType(ResNo: 0), N1: InOp, N2: N->getOperand(Num: 1)); |
| 8157 | } |
| 8158 | |
| 8159 | SDValue DAGTypeLegalizer::(SDNode *N) { |
| 8160 | SDValue InOp = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 8161 | return DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: SDLoc(N), |
| 8162 | VT: N->getValueType(ResNo: 0), N1: InOp, N2: N->getOperand(Num: 1)); |
| 8163 | } |
| 8164 | |
| 8165 | SDValue DAGTypeLegalizer::WidenVecOp_EXTEND_VECTOR_INREG(SDNode *N) { |
| 8166 | SDLoc DL(N); |
| 8167 | EVT ResVT = N->getValueType(ResNo: 0); |
| 8168 | |
| 8169 | // Widen the input as requested by the legalizer. |
| 8170 | SDValue WideInOp = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 8171 | EVT WideInVT = WideInOp.getValueType(); |
| 8172 | |
| 8173 | // Simple case: if widened input is still smaller than or equal to result, |
| 8174 | // just use it directly. |
| 8175 | if (WideInVT.getSizeInBits() <= ResVT.getSizeInBits()) |
| 8176 | return DAG.getNode(Opcode: N->getOpcode(), DL, VT: ResVT, Operand: WideInOp); |
| 8177 | |
| 8178 | // EXTEND_VECTOR_INREG requires input bits <= result bits. |
| 8179 | // If widening makes the input larger than the original result, widen the |
| 8180 | // result to match, then extract back down. |
| 8181 | EVT ResEltVT = ResVT.getVectorElementType(); |
| 8182 | unsigned EltBits = ResEltVT.getSizeInBits(); |
| 8183 | assert((WideInVT.getSizeInBits() % EltBits) == 0 && |
| 8184 | "Widened input size must be a multiple of result element size" ); |
| 8185 | |
| 8186 | unsigned WideNumElts = WideInVT.getSizeInBits() / EltBits; |
| 8187 | EVT WideResVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: ResEltVT, NumElements: WideNumElts); |
| 8188 | |
| 8189 | SDValue WideRes = DAG.getNode(Opcode: N->getOpcode(), DL, VT: WideResVT, Operand: WideInOp); |
| 8190 | return DAG.getExtractSubvector(DL, VT: ResVT, Vec: WideRes, Idx: 0); |
| 8191 | } |
| 8192 | |
| 8193 | SDValue DAGTypeLegalizer::WidenVecOp_STORE(SDNode *N) { |
| 8194 | // We have to widen the value, but we want only to store the original |
| 8195 | // vector type. |
| 8196 | StoreSDNode *ST = cast<StoreSDNode>(Val: N); |
| 8197 | |
| 8198 | if (!ST->getMemoryVT().getScalarType().isByteSized()) |
| 8199 | return TLI.scalarizeVectorStore(ST, DAG); |
| 8200 | |
| 8201 | if (ST->isTruncatingStore()) |
| 8202 | return TLI.scalarizeVectorStore(ST, DAG); |
| 8203 | |
| 8204 | // Generate a vector-predicated store if it is custom/legal on the target. |
| 8205 | // To avoid possible recursion, only do this if the widened mask type is |
| 8206 | // legal. |
| 8207 | // FIXME: Not all targets may support EVL in VP_STORE. These will have been |
| 8208 | // removed from the IR by the ExpandVectorPredication pass but we're |
| 8209 | // reintroducing them here. |
| 8210 | SDValue StVal = ST->getValue(); |
| 8211 | EVT StVT = StVal.getValueType(); |
| 8212 | EVT WideVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: StVT); |
| 8213 | EVT WideMaskVT = getSetCCResultType(VT: WideVT); |
| 8214 | |
| 8215 | if (TLI.isOperationLegalOrCustom(Op: ISD::VP_STORE, VT: WideVT) && |
| 8216 | TLI.isTypeLegal(VT: WideMaskVT)) { |
| 8217 | // Widen the value. |
| 8218 | SDLoc DL(N); |
| 8219 | StVal = GetWidenedVector(Op: StVal); |
| 8220 | SDValue Mask = DAG.getAllOnesConstant(DL, VT: WideMaskVT); |
| 8221 | SDValue EVL = DAG.getElementCount(DL, VT: TLI.getVPExplicitVectorLengthTy(), |
| 8222 | EC: StVT.getVectorElementCount()); |
| 8223 | return DAG.getStoreVP(Chain: ST->getChain(), dl: DL, Val: StVal, Ptr: ST->getBasePtr(), |
| 8224 | Offset: ST->getOffset(), Mask, EVL, MemVT: StVT, MMO: ST->getMemOperand(), |
| 8225 | AM: ST->getAddressingMode()); |
| 8226 | } |
| 8227 | |
| 8228 | SmallVector<SDValue, 16> StChain; |
| 8229 | if (GenWidenVectorStores(StChain, ST)) { |
| 8230 | if (StChain.size() == 1) |
| 8231 | return StChain[0]; |
| 8232 | |
| 8233 | return DAG.getNode(Opcode: ISD::TokenFactor, DL: SDLoc(ST), VT: MVT::Other, Ops: StChain); |
| 8234 | } |
| 8235 | |
| 8236 | if (StVT.isVector()) { |
| 8237 | // If all else fails replace the store with a wide masked store. |
| 8238 | SDLoc DL(N); |
| 8239 | SDValue WideStVal = GetWidenedVector(Op: StVal); |
| 8240 | SDValue Mask = |
| 8241 | DAG.getMaskFromElementCount(DL, VT: WideVT, Len: StVT.getVectorElementCount()); |
| 8242 | |
| 8243 | return DAG.getMaskedStore(Chain: ST->getChain(), dl: DL, Val: WideStVal, Base: ST->getBasePtr(), |
| 8244 | Offset: ST->getOffset(), Mask, MemVT: ST->getMemoryVT(), |
| 8245 | MMO: ST->getMemOperand(), AM: ST->getAddressingMode(), |
| 8246 | IsTruncating: ST->isTruncatingStore()); |
| 8247 | } |
| 8248 | |
| 8249 | report_fatal_error(reason: "Unable to widen vector store" ); |
| 8250 | } |
| 8251 | |
| 8252 | SDValue DAGTypeLegalizer::WidenVecOp_ATOMIC_STORE(AtomicSDNode *ST) { |
| 8253 | EVT StVT = ST->getMemoryVT(); |
| 8254 | SDLoc dl(ST); |
| 8255 | |
| 8256 | SDValue StVal = GetWidenedVector(Op: ST->getVal()); |
| 8257 | EVT WidenVT = StVal.getValueType(); |
| 8258 | |
| 8259 | TypeSize StWidth = StVT.getSizeInBits(); |
| 8260 | TypeSize WidenWidth = WidenVT.getSizeInBits(); |
| 8261 | TypeSize WidthDiff = WidenWidth - StWidth; |
| 8262 | |
| 8263 | // Find the vector type that can store the original memory width in one |
| 8264 | // atomic operation. Pass StAlign=0 (like atomic loads); a real align would |
| 8265 | // let findMemType widen the access past the value (e.g. <2 x i8> at align 4 |
| 8266 | // implies a 4-byte movl, writing undef bytes past its object). |
| 8267 | std::optional<EVT> FirstVT = |
| 8268 | findMemType(DAG, TLI, Width: StWidth.getKnownMinValue(), WidenVT, /*StAlign=*/Align: 0, |
| 8269 | WidenEx: WidthDiff.getKnownMinValue()); |
| 8270 | if (!FirstVT) |
| 8271 | return SDValue(); |
| 8272 | |
| 8273 | TypeSize FirstVTWidth = FirstVT->getSizeInBits(); |
| 8274 | |
| 8275 | SDValue StOp = |
| 8276 | coerceStoredValue(StVal, FirstVT: *FirstVT, WidenVT, FirstVTWidth, dl, DAG); |
| 8277 | |
| 8278 | return DAG.getAtomic(Opcode: ISD::ATOMIC_STORE, dl, MemVT: *FirstVT, Chain: ST->getChain(), Ptr: StOp, |
| 8279 | Val: ST->getBasePtr(), MMO: ST->getMemOperand()); |
| 8280 | } |
| 8281 | |
| 8282 | SDValue DAGTypeLegalizer::WidenVecOp_VP_STORE(SDNode *N, unsigned OpNo) { |
| 8283 | assert((OpNo == 1 || OpNo == 3) && |
| 8284 | "Can widen only data or mask operand of vp_store" ); |
| 8285 | VPStoreSDNode *ST = cast<VPStoreSDNode>(Val: N); |
| 8286 | SDValue Mask = ST->getMask(); |
| 8287 | SDValue StVal = ST->getValue(); |
| 8288 | SDLoc dl(N); |
| 8289 | |
| 8290 | if (OpNo == 1) { |
| 8291 | // Widen the value. |
| 8292 | StVal = GetWidenedVector(Op: StVal); |
| 8293 | |
| 8294 | // We only handle the case where the mask needs widening to an |
| 8295 | // identically-sized type as the vector inputs. |
| 8296 | assert(getTypeAction(Mask.getValueType()) == |
| 8297 | TargetLowering::TypeWidenVector && |
| 8298 | "Unable to widen VP store" ); |
| 8299 | Mask = GetWidenedVector(Op: Mask); |
| 8300 | } else { |
| 8301 | Mask = GetWidenedVector(Op: Mask); |
| 8302 | |
| 8303 | // We only handle the case where the stored value needs widening to an |
| 8304 | // identically-sized type as the mask. |
| 8305 | assert(getTypeAction(StVal.getValueType()) == |
| 8306 | TargetLowering::TypeWidenVector && |
| 8307 | "Unable to widen VP store" ); |
| 8308 | StVal = GetWidenedVector(Op: StVal); |
| 8309 | } |
| 8310 | |
| 8311 | assert(Mask.getValueType().getVectorElementCount() == |
| 8312 | StVal.getValueType().getVectorElementCount() && |
| 8313 | "Mask and data vectors should have the same number of elements" ); |
| 8314 | return DAG.getStoreVP(Chain: ST->getChain(), dl, Val: StVal, Ptr: ST->getBasePtr(), |
| 8315 | Offset: ST->getOffset(), Mask, EVL: ST->getVectorLength(), |
| 8316 | MemVT: ST->getMemoryVT(), MMO: ST->getMemOperand(), |
| 8317 | AM: ST->getAddressingMode(), IsTruncating: ST->isTruncatingStore(), |
| 8318 | IsCompressing: ST->isCompressingStore()); |
| 8319 | } |
| 8320 | |
| 8321 | SDValue DAGTypeLegalizer::WidenVecOp_VP_STRIDED_STORE(SDNode *N, |
| 8322 | unsigned OpNo) { |
| 8323 | assert((OpNo == 1 || OpNo == 4) && |
| 8324 | "Can widen only data or mask operand of vp_strided_store" ); |
| 8325 | VPStridedStoreSDNode *SST = cast<VPStridedStoreSDNode>(Val: N); |
| 8326 | SDValue Mask = SST->getMask(); |
| 8327 | SDValue StVal = SST->getValue(); |
| 8328 | SDLoc DL(N); |
| 8329 | |
| 8330 | if (OpNo == 1) |
| 8331 | assert(getTypeAction(Mask.getValueType()) == |
| 8332 | TargetLowering::TypeWidenVector && |
| 8333 | "Unable to widen VP strided store" ); |
| 8334 | else |
| 8335 | assert(getTypeAction(StVal.getValueType()) == |
| 8336 | TargetLowering::TypeWidenVector && |
| 8337 | "Unable to widen VP strided store" ); |
| 8338 | |
| 8339 | StVal = GetWidenedVector(Op: StVal); |
| 8340 | Mask = GetWidenedVector(Op: Mask); |
| 8341 | |
| 8342 | assert(StVal.getValueType().getVectorElementCount() == |
| 8343 | Mask.getValueType().getVectorElementCount() && |
| 8344 | "Data and mask vectors should have the same number of elements" ); |
| 8345 | |
| 8346 | return DAG.getStridedStoreVP( |
| 8347 | Chain: SST->getChain(), DL, Val: StVal, Ptr: SST->getBasePtr(), Offset: SST->getOffset(), |
| 8348 | Stride: SST->getStride(), Mask, EVL: SST->getVectorLength(), MemVT: SST->getMemoryVT(), |
| 8349 | MMO: SST->getMemOperand(), AM: SST->getAddressingMode(), IsTruncating: SST->isTruncatingStore(), |
| 8350 | IsCompressing: SST->isCompressingStore()); |
| 8351 | } |
| 8352 | |
| 8353 | SDValue DAGTypeLegalizer::WidenVecOp_MSTORE(SDNode *N, unsigned OpNo) { |
| 8354 | assert((OpNo == 1 || OpNo == 4) && |
| 8355 | "Can widen only data or mask operand of mstore" ); |
| 8356 | MaskedStoreSDNode *MST = cast<MaskedStoreSDNode>(Val: N); |
| 8357 | SDValue Mask = MST->getMask(); |
| 8358 | EVT MaskVT = Mask.getValueType(); |
| 8359 | SDValue StVal = MST->getValue(); |
| 8360 | EVT VT = StVal.getValueType(); |
| 8361 | SDLoc dl(N); |
| 8362 | |
| 8363 | EVT WideVT, WideMaskVT; |
| 8364 | if (OpNo == 1) { |
| 8365 | // Widen the value. |
| 8366 | StVal = GetWidenedVector(Op: StVal); |
| 8367 | |
| 8368 | WideVT = StVal.getValueType(); |
| 8369 | WideMaskVT = |
| 8370 | EVT::getVectorVT(Context&: *DAG.getContext(), VT: MaskVT.getVectorElementType(), |
| 8371 | EC: WideVT.getVectorElementCount()); |
| 8372 | } else { |
| 8373 | WideMaskVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: MaskVT); |
| 8374 | |
| 8375 | EVT ValueVT = StVal.getValueType(); |
| 8376 | WideVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: ValueVT.getVectorElementType(), |
| 8377 | EC: WideMaskVT.getVectorElementCount()); |
| 8378 | } |
| 8379 | |
| 8380 | if (TLI.isOperationLegalOrCustom(Op: ISD::VP_STORE, VT: WideVT) && |
| 8381 | TLI.isTypeLegal(VT: WideMaskVT) && !MST->isCompressingStore()) { |
| 8382 | Mask = DAG.getInsertSubvector(DL: dl, Vec: DAG.getPOISON(VT: WideMaskVT), SubVec: Mask, Idx: 0); |
| 8383 | SDValue EVL = DAG.getElementCount(DL: dl, VT: TLI.getVPExplicitVectorLengthTy(), |
| 8384 | EC: VT.getVectorElementCount()); |
| 8385 | return DAG.getStoreVP(Chain: MST->getChain(), dl, Val: StVal, Ptr: MST->getBasePtr(), |
| 8386 | Offset: MST->getOffset(), Mask, EVL, MemVT: MST->getMemoryVT(), |
| 8387 | MMO: MST->getMemOperand(), AM: MST->getAddressingMode()); |
| 8388 | } |
| 8389 | |
| 8390 | if (OpNo == 1) { |
| 8391 | // The mask should be widened as well. |
| 8392 | Mask = ModifyToType(InOp: Mask, NVT: WideMaskVT, FillWithZeroes: true); |
| 8393 | } else { |
| 8394 | // Widen the mask. |
| 8395 | Mask = ModifyToType(InOp: Mask, NVT: WideMaskVT, FillWithZeroes: true); |
| 8396 | |
| 8397 | StVal = ModifyToType(InOp: StVal, NVT: WideVT); |
| 8398 | } |
| 8399 | |
| 8400 | assert(Mask.getValueType().getVectorElementCount() == |
| 8401 | StVal.getValueType().getVectorElementCount() && |
| 8402 | "Mask and data vectors should have the same number of elements" ); |
| 8403 | return DAG.getMaskedStore(Chain: MST->getChain(), dl, Val: StVal, Base: MST->getBasePtr(), |
| 8404 | Offset: MST->getOffset(), Mask, MemVT: MST->getMemoryVT(), |
| 8405 | MMO: MST->getMemOperand(), AM: MST->getAddressingMode(), |
| 8406 | IsTruncating: false, IsCompressing: MST->isCompressingStore()); |
| 8407 | } |
| 8408 | |
| 8409 | SDValue DAGTypeLegalizer::WidenVecOp_MGATHER(SDNode *N, unsigned OpNo) { |
| 8410 | assert(OpNo == 4 && "Can widen only the index of mgather" ); |
| 8411 | auto *MG = cast<MaskedGatherSDNode>(Val: N); |
| 8412 | SDValue DataOp = MG->getPassThru(); |
| 8413 | SDValue Mask = MG->getMask(); |
| 8414 | SDValue Scale = MG->getScale(); |
| 8415 | |
| 8416 | // Just widen the index. It's allowed to have extra elements. |
| 8417 | SDValue Index = GetWidenedVector(Op: MG->getIndex()); |
| 8418 | |
| 8419 | SDLoc dl(N); |
| 8420 | SDValue Ops[] = {MG->getChain(), DataOp, Mask, MG->getBasePtr(), Index, |
| 8421 | Scale}; |
| 8422 | SDValue Res = DAG.getMaskedGather(VTs: MG->getVTList(), MemVT: MG->getMemoryVT(), dl, Ops, |
| 8423 | MMO: MG->getMemOperand(), IndexType: MG->getIndexType(), |
| 8424 | ExtTy: MG->getExtensionType()); |
| 8425 | ReplaceValueWith(From: SDValue(N, 1), To: Res.getValue(R: 1)); |
| 8426 | ReplaceValueWith(From: SDValue(N, 0), To: Res.getValue(R: 0)); |
| 8427 | return SDValue(); |
| 8428 | } |
| 8429 | |
| 8430 | SDValue DAGTypeLegalizer::WidenVecOp_MSCATTER(SDNode *N, unsigned OpNo) { |
| 8431 | MaskedScatterSDNode *MSC = cast<MaskedScatterSDNode>(Val: N); |
| 8432 | SDValue DataOp = MSC->getValue(); |
| 8433 | SDValue Mask = MSC->getMask(); |
| 8434 | SDValue Index = MSC->getIndex(); |
| 8435 | SDValue Scale = MSC->getScale(); |
| 8436 | EVT WideMemVT = MSC->getMemoryVT(); |
| 8437 | |
| 8438 | if (OpNo == 1) { |
| 8439 | DataOp = GetWidenedVector(Op: DataOp); |
| 8440 | ElementCount WideEC = DataOp.getValueType().getVectorElementCount(); |
| 8441 | |
| 8442 | // Widen index. |
| 8443 | EVT IndexVT = Index.getValueType(); |
| 8444 | EVT WideIndexVT = EVT::getVectorVT(Context&: *DAG.getContext(), |
| 8445 | VT: IndexVT.getVectorElementType(), EC: WideEC); |
| 8446 | Index = ModifyToType(InOp: Index, NVT: WideIndexVT); |
| 8447 | |
| 8448 | // The mask should be widened as well. |
| 8449 | EVT MaskVT = Mask.getValueType(); |
| 8450 | EVT WideMaskVT = EVT::getVectorVT(Context&: *DAG.getContext(), |
| 8451 | VT: MaskVT.getVectorElementType(), EC: WideEC); |
| 8452 | Mask = ModifyToType(InOp: Mask, NVT: WideMaskVT, FillWithZeroes: true); |
| 8453 | |
| 8454 | // Widen the MemoryType |
| 8455 | WideMemVT = EVT::getVectorVT(Context&: *DAG.getContext(), |
| 8456 | VT: MSC->getMemoryVT().getScalarType(), EC: WideEC); |
| 8457 | } else if (OpNo == 4) { |
| 8458 | // Just widen the index. It's allowed to have extra elements. |
| 8459 | Index = GetWidenedVector(Op: Index); |
| 8460 | } else |
| 8461 | llvm_unreachable("Can't widen this operand of mscatter" ); |
| 8462 | |
| 8463 | SDValue Ops[] = {MSC->getChain(), DataOp, Mask, MSC->getBasePtr(), Index, |
| 8464 | Scale}; |
| 8465 | return DAG.getMaskedScatter(VTs: DAG.getVTList(VT: MVT::Other), MemVT: WideMemVT, dl: SDLoc(N), |
| 8466 | Ops, MMO: MSC->getMemOperand(), IndexType: MSC->getIndexType(), |
| 8467 | IsTruncating: MSC->isTruncatingStore()); |
| 8468 | } |
| 8469 | |
| 8470 | SDValue DAGTypeLegalizer::WidenVecOp_VP_SCATTER(SDNode *N, unsigned OpNo) { |
| 8471 | VPScatterSDNode *VPSC = cast<VPScatterSDNode>(Val: N); |
| 8472 | SDValue DataOp = VPSC->getValue(); |
| 8473 | SDValue Mask = VPSC->getMask(); |
| 8474 | SDValue Index = VPSC->getIndex(); |
| 8475 | SDValue Scale = VPSC->getScale(); |
| 8476 | EVT WideMemVT = VPSC->getMemoryVT(); |
| 8477 | |
| 8478 | if (OpNo == 1) { |
| 8479 | DataOp = GetWidenedVector(Op: DataOp); |
| 8480 | Index = GetWidenedVector(Op: Index); |
| 8481 | const auto WideEC = DataOp.getValueType().getVectorElementCount(); |
| 8482 | Mask = GetWidenedMask(Mask, EC: WideEC); |
| 8483 | WideMemVT = EVT::getVectorVT(Context&: *DAG.getContext(), |
| 8484 | VT: VPSC->getMemoryVT().getScalarType(), EC: WideEC); |
| 8485 | } else if (OpNo == 3) { |
| 8486 | // Just widen the index. It's allowed to have extra elements. |
| 8487 | Index = GetWidenedVector(Op: Index); |
| 8488 | } else |
| 8489 | llvm_unreachable("Can't widen this operand of VP_SCATTER" ); |
| 8490 | |
| 8491 | SDValue Ops[] = { |
| 8492 | VPSC->getChain(), DataOp, VPSC->getBasePtr(), Index, Scale, Mask, |
| 8493 | VPSC->getVectorLength()}; |
| 8494 | return DAG.getScatterVP(VTs: DAG.getVTList(VT: MVT::Other), VT: WideMemVT, dl: SDLoc(N), Ops, |
| 8495 | MMO: VPSC->getMemOperand(), IndexType: VPSC->getIndexType()); |
| 8496 | } |
| 8497 | |
| 8498 | SDValue DAGTypeLegalizer::WidenVecOp_SETCC(SDNode *N) { |
| 8499 | SDValue InOp0 = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 8500 | SDValue InOp1 = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 8501 | SDLoc dl(N); |
| 8502 | EVT VT = N->getValueType(ResNo: 0); |
| 8503 | |
| 8504 | // WARNING: In this code we widen the compare instruction with garbage. |
| 8505 | // This garbage may contain denormal floats which may be slow. Is this a real |
| 8506 | // concern ? Should we zero the unused lanes if this is a float compare ? |
| 8507 | |
| 8508 | // Get a new SETCC node to compare the newly widened operands. |
| 8509 | // Only some of the compared elements are legal. |
| 8510 | EVT SVT = getSetCCResultType(VT: InOp0.getValueType()); |
| 8511 | // The result type is legal, if its vXi1, keep vXi1 for the new SETCC. |
| 8512 | if (VT.getScalarType() == MVT::i1) |
| 8513 | SVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: MVT::i1, |
| 8514 | EC: SVT.getVectorElementCount()); |
| 8515 | |
| 8516 | SDValue WideSETCC = DAG.getNode(Opcode: ISD::SETCC, DL: SDLoc(N), |
| 8517 | VT: SVT, N1: InOp0, N2: InOp1, N3: N->getOperand(Num: 2)); |
| 8518 | |
| 8519 | // Extract the needed results from the result vector. |
| 8520 | EVT ResVT = EVT::getVectorVT(Context&: *DAG.getContext(), |
| 8521 | VT: SVT.getVectorElementType(), |
| 8522 | EC: VT.getVectorElementCount()); |
| 8523 | SDValue CC = DAG.getExtractSubvector(DL: dl, VT: ResVT, Vec: WideSETCC, Idx: 0); |
| 8524 | |
| 8525 | EVT OpVT = N->getOperand(Num: 0).getValueType(); |
| 8526 | ISD::NodeType ExtendCode = |
| 8527 | TargetLowering::getExtendForContent(Content: TLI.getBooleanContents(Type: OpVT)); |
| 8528 | return DAG.getNode(Opcode: ExtendCode, DL: dl, VT, Operand: CC); |
| 8529 | } |
| 8530 | |
| 8531 | SDValue DAGTypeLegalizer::WidenVecOp_STRICT_FSETCC(SDNode *N) { |
| 8532 | SDValue Chain = N->getOperand(Num: 0); |
| 8533 | SDValue LHS = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 8534 | SDValue RHS = GetWidenedVector(Op: N->getOperand(Num: 2)); |
| 8535 | SDValue CC = N->getOperand(Num: 3); |
| 8536 | SDLoc dl(N); |
| 8537 | |
| 8538 | EVT VT = N->getValueType(ResNo: 0); |
| 8539 | EVT EltVT = VT.getVectorElementType(); |
| 8540 | EVT TmpEltVT = LHS.getValueType().getVectorElementType(); |
| 8541 | unsigned NumElts = VT.getVectorNumElements(); |
| 8542 | |
| 8543 | // Unroll into a build vector. |
| 8544 | SmallVector<SDValue, 8> Scalars(NumElts); |
| 8545 | SmallVector<SDValue, 8> Chains(NumElts); |
| 8546 | |
| 8547 | for (unsigned i = 0; i != NumElts; ++i) { |
| 8548 | SDValue LHSElem = DAG.getExtractVectorElt(DL: dl, VT: TmpEltVT, Vec: LHS, Idx: i); |
| 8549 | SDValue RHSElem = DAG.getExtractVectorElt(DL: dl, VT: TmpEltVT, Vec: RHS, Idx: i); |
| 8550 | |
| 8551 | Scalars[i] = DAG.getNode(Opcode: N->getOpcode(), DL: dl, ResultTys: {MVT::i1, MVT::Other}, |
| 8552 | Ops: {Chain, LHSElem, RHSElem, CC}); |
| 8553 | Chains[i] = Scalars[i].getValue(R: 1); |
| 8554 | Scalars[i] = DAG.getSelect(DL: dl, VT: EltVT, Cond: Scalars[i], |
| 8555 | LHS: DAG.getBoolConstant(V: true, DL: dl, VT: EltVT, OpVT: VT), |
| 8556 | RHS: DAG.getBoolConstant(V: false, DL: dl, VT: EltVT, OpVT: VT)); |
| 8557 | } |
| 8558 | |
| 8559 | SDValue NewChain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: Chains); |
| 8560 | ReplaceValueWith(From: SDValue(N, 1), To: NewChain); |
| 8561 | |
| 8562 | return DAG.getBuildVector(VT, DL: dl, Ops: Scalars); |
| 8563 | } |
| 8564 | |
| 8565 | static unsigned getExtendForIntVecReduction(unsigned Opc) { |
| 8566 | switch (Opc) { |
| 8567 | default: |
| 8568 | llvm_unreachable("Expected integer vector reduction" ); |
| 8569 | case ISD::VECREDUCE_ADD: |
| 8570 | case ISD::VECREDUCE_MUL: |
| 8571 | case ISD::VECREDUCE_AND: |
| 8572 | case ISD::VECREDUCE_OR: |
| 8573 | case ISD::VECREDUCE_XOR: |
| 8574 | return ISD::ANY_EXTEND; |
| 8575 | case ISD::VECREDUCE_SMAX: |
| 8576 | case ISD::VECREDUCE_SMIN: |
| 8577 | return ISD::SIGN_EXTEND; |
| 8578 | case ISD::VECREDUCE_UMAX: |
| 8579 | case ISD::VECREDUCE_UMIN: |
| 8580 | return ISD::ZERO_EXTEND; |
| 8581 | } |
| 8582 | } |
| 8583 | |
| 8584 | SDValue DAGTypeLegalizer::WidenVecOp_VECREDUCE(SDNode *N) { |
| 8585 | SDLoc dl(N); |
| 8586 | SDValue Op = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 8587 | EVT VT = N->getValueType(ResNo: 0); |
| 8588 | EVT OrigVT = N->getOperand(Num: 0).getValueType(); |
| 8589 | EVT WideVT = Op.getValueType(); |
| 8590 | EVT ElemVT = OrigVT.getVectorElementType(); |
| 8591 | SDNodeFlags Flags = N->getFlags(); |
| 8592 | |
| 8593 | unsigned Opc = N->getOpcode(); |
| 8594 | unsigned BaseOpc = ISD::getVecReduceBaseOpcode(VecReduceOpcode: Opc); |
| 8595 | SDValue NeutralElem = DAG.getIdentityElement(Opcode: BaseOpc, DL: dl, VT: ElemVT, Flags); |
| 8596 | assert(NeutralElem && "Neutral element must exist" ); |
| 8597 | |
| 8598 | // Pad the vector with the neutral element. |
| 8599 | unsigned OrigElts = OrigVT.getVectorMinNumElements(); |
| 8600 | unsigned WideElts = WideVT.getVectorMinNumElements(); |
| 8601 | |
| 8602 | // Generate a vp.reduce_op if it is custom/legal for the target. This avoids |
| 8603 | // needing to pad the source vector, because the inactive lanes can simply be |
| 8604 | // disabled and not contribute to the result. |
| 8605 | if (auto VPOpcode = ISD::getVPForBaseOpcode(Opcode: Opc); |
| 8606 | VPOpcode && TLI.isOperationLegalOrCustom(Op: *VPOpcode, VT: WideVT)) { |
| 8607 | SDValue Start = NeutralElem; |
| 8608 | if (VT.isInteger()) |
| 8609 | Start = DAG.getNode(Opcode: getExtendForIntVecReduction(Opc), DL: dl, VT, Operand: Start); |
| 8610 | assert(Start.getValueType() == VT); |
| 8611 | EVT WideMaskVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: MVT::i1, |
| 8612 | EC: WideVT.getVectorElementCount()); |
| 8613 | SDValue Mask = DAG.getAllOnesConstant(DL: dl, VT: WideMaskVT); |
| 8614 | SDValue EVL = DAG.getElementCount(DL: dl, VT: TLI.getVPExplicitVectorLengthTy(), |
| 8615 | EC: OrigVT.getVectorElementCount()); |
| 8616 | return DAG.getNode(Opcode: *VPOpcode, DL: dl, VT, Ops: {Start, Op, Mask, EVL}, Flags); |
| 8617 | } |
| 8618 | |
| 8619 | if (WideVT.isScalableVector()) { |
| 8620 | unsigned GCD = std::gcd(m: OrigElts, n: WideElts); |
| 8621 | EVT SplatVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: ElemVT, |
| 8622 | EC: ElementCount::getScalable(MinVal: GCD)); |
| 8623 | SDValue SplatNeutral = DAG.getSplatVector(VT: SplatVT, DL: dl, Op: NeutralElem); |
| 8624 | for (unsigned Idx = OrigElts; Idx < WideElts; Idx = Idx + GCD) |
| 8625 | Op = DAG.getInsertSubvector(DL: dl, Vec: Op, SubVec: SplatNeutral, Idx); |
| 8626 | return DAG.getNode(Opcode: Opc, DL: dl, VT, Operand: Op, Flags); |
| 8627 | } |
| 8628 | |
| 8629 | for (unsigned Idx = OrigElts; Idx < WideElts; Idx++) |
| 8630 | Op = DAG.getInsertVectorElt(DL: dl, Vec: Op, Elt: NeutralElem, Idx); |
| 8631 | |
| 8632 | return DAG.getNode(Opcode: Opc, DL: dl, VT, Operand: Op, Flags); |
| 8633 | } |
| 8634 | |
| 8635 | SDValue DAGTypeLegalizer::WidenVecOp_VECREDUCE_SEQ(SDNode *N) { |
| 8636 | SDLoc dl(N); |
| 8637 | SDValue AccOp = N->getOperand(Num: 0); |
| 8638 | SDValue VecOp = N->getOperand(Num: 1); |
| 8639 | SDValue Op = GetWidenedVector(Op: VecOp); |
| 8640 | |
| 8641 | EVT VT = N->getValueType(ResNo: 0); |
| 8642 | EVT OrigVT = VecOp.getValueType(); |
| 8643 | EVT WideVT = Op.getValueType(); |
| 8644 | EVT ElemVT = OrigVT.getVectorElementType(); |
| 8645 | SDNodeFlags Flags = N->getFlags(); |
| 8646 | |
| 8647 | unsigned Opc = N->getOpcode(); |
| 8648 | unsigned BaseOpc = ISD::getVecReduceBaseOpcode(VecReduceOpcode: Opc); |
| 8649 | SDValue NeutralElem = DAG.getIdentityElement(Opcode: BaseOpc, DL: dl, VT: ElemVT, Flags); |
| 8650 | |
| 8651 | // Pad the vector with the neutral element. |
| 8652 | unsigned OrigElts = OrigVT.getVectorMinNumElements(); |
| 8653 | unsigned WideElts = WideVT.getVectorMinNumElements(); |
| 8654 | |
| 8655 | // Generate a vp.reduce_op if it is custom/legal for the target. This avoids |
| 8656 | // needing to pad the source vector, because the inactive lanes can simply be |
| 8657 | // disabled and not contribute to the result. |
| 8658 | if (auto VPOpcode = ISD::getVPForBaseOpcode(Opcode: Opc); |
| 8659 | VPOpcode && TLI.isOperationLegalOrCustom(Op: *VPOpcode, VT: WideVT)) { |
| 8660 | EVT WideMaskVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: MVT::i1, |
| 8661 | EC: WideVT.getVectorElementCount()); |
| 8662 | SDValue Mask = DAG.getAllOnesConstant(DL: dl, VT: WideMaskVT); |
| 8663 | SDValue EVL = DAG.getElementCount(DL: dl, VT: TLI.getVPExplicitVectorLengthTy(), |
| 8664 | EC: OrigVT.getVectorElementCount()); |
| 8665 | return DAG.getNode(Opcode: *VPOpcode, DL: dl, VT, Ops: {AccOp, Op, Mask, EVL}, Flags); |
| 8666 | } |
| 8667 | |
| 8668 | if (WideVT.isScalableVector()) { |
| 8669 | unsigned GCD = std::gcd(m: OrigElts, n: WideElts); |
| 8670 | EVT SplatVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: ElemVT, |
| 8671 | EC: ElementCount::getScalable(MinVal: GCD)); |
| 8672 | SDValue SplatNeutral = DAG.getSplatVector(VT: SplatVT, DL: dl, Op: NeutralElem); |
| 8673 | for (unsigned Idx = OrigElts; Idx < WideElts; Idx = Idx + GCD) |
| 8674 | Op = DAG.getInsertSubvector(DL: dl, Vec: Op, SubVec: SplatNeutral, Idx); |
| 8675 | return DAG.getNode(Opcode: Opc, DL: dl, VT, N1: AccOp, N2: Op, Flags); |
| 8676 | } |
| 8677 | |
| 8678 | for (unsigned Idx = OrigElts; Idx < WideElts; Idx++) |
| 8679 | Op = DAG.getInsertVectorElt(DL: dl, Vec: Op, Elt: NeutralElem, Idx); |
| 8680 | |
| 8681 | return DAG.getNode(Opcode: Opc, DL: dl, VT, N1: AccOp, N2: Op, Flags); |
| 8682 | } |
| 8683 | |
| 8684 | SDValue DAGTypeLegalizer::WidenVecOp_VP_REDUCE(SDNode *N) { |
| 8685 | assert(N->isVPOpcode() && "Expected VP opcode" ); |
| 8686 | |
| 8687 | SDLoc dl(N); |
| 8688 | SDValue Op = GetWidenedVector(Op: N->getOperand(Num: 1)); |
| 8689 | SDValue Mask = GetWidenedMask(Mask: N->getOperand(Num: 2), |
| 8690 | EC: Op.getValueType().getVectorElementCount()); |
| 8691 | |
| 8692 | return DAG.getNode(Opcode: N->getOpcode(), DL: dl, VT: N->getValueType(ResNo: 0), |
| 8693 | Ops: {N->getOperand(Num: 0), Op, Mask, N->getOperand(Num: 3)}, |
| 8694 | Flags: N->getFlags()); |
| 8695 | } |
| 8696 | |
| 8697 | SDValue DAGTypeLegalizer::WidenVecOp_VSELECT(SDNode *N) { |
| 8698 | // This only gets called in the case that the left and right inputs and |
| 8699 | // result are of a legal odd vector type, and the condition is illegal i1 of |
| 8700 | // the same odd width that needs widening. |
| 8701 | EVT VT = N->getValueType(ResNo: 0); |
| 8702 | assert(VT.isVector() && !VT.isPow2VectorType() && isTypeLegal(VT)); |
| 8703 | |
| 8704 | SDValue Cond = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 8705 | SDValue LeftIn = DAG.WidenVector(N: N->getOperand(Num: 1), DL: SDLoc(N)); |
| 8706 | SDValue RightIn = DAG.WidenVector(N: N->getOperand(Num: 2), DL: SDLoc(N)); |
| 8707 | SDLoc DL(N); |
| 8708 | |
| 8709 | SDValue Select = DAG.getNode(Opcode: N->getOpcode(), DL, VT: LeftIn.getValueType(), N1: Cond, |
| 8710 | N2: LeftIn, N3: RightIn); |
| 8711 | return DAG.getExtractSubvector(DL, VT, Vec: Select, Idx: 0); |
| 8712 | } |
| 8713 | |
| 8714 | SDValue DAGTypeLegalizer::WidenVecOp_CttzElements(SDNode *N) { |
| 8715 | SDLoc DL(N); |
| 8716 | SDValue Source = N->getOperand(Num: 0); |
| 8717 | EVT WideVT = |
| 8718 | TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: Source.getValueType()); |
| 8719 | |
| 8720 | SDValue WideSource; |
| 8721 | if (N->getOpcode() == ISD::CTTZ_ELTS_ZERO_POISON) { |
| 8722 | WideSource = GetWidenedVector(Op: Source); |
| 8723 | } else { |
| 8724 | // Pad the widened portion with all-ones so the extra lanes appear as |
| 8725 | // active (non-zero) elements and do not contribute trailing zeros. |
| 8726 | SDValue AllOnes = DAG.getAllOnesConstant(DL, VT: WideVT); |
| 8727 | WideSource = DAG.getInsertSubvector(DL, Vec: AllOnes, SubVec: Source, Idx: 0); |
| 8728 | } |
| 8729 | |
| 8730 | return DAG.getNode(Opcode: N->getOpcode(), DL, VT: N->getValueType(ResNo: 0), Operand: WideSource, |
| 8731 | Flags: N->getFlags()); |
| 8732 | } |
| 8733 | |
| 8734 | SDValue DAGTypeLegalizer::WidenVecOp_VP_CttzElements(SDNode *N) { |
| 8735 | SDLoc DL(N); |
| 8736 | SDValue Source = GetWidenedVector(Op: N->getOperand(Num: 0)); |
| 8737 | EVT SrcVT = Source.getValueType(); |
| 8738 | SDValue Mask = |
| 8739 | GetWidenedMask(Mask: N->getOperand(Num: 1), EC: SrcVT.getVectorElementCount()); |
| 8740 | |
| 8741 | return DAG.getNode(Opcode: N->getOpcode(), DL, VT: N->getValueType(ResNo: 0), |
| 8742 | Ops: {Source, Mask, N->getOperand(Num: 2)}, Flags: N->getFlags()); |
| 8743 | } |
| 8744 | |
| 8745 | SDValue DAGTypeLegalizer::WidenVecOp_VECTOR_FIND_LAST_ACTIVE(SDNode *N) { |
| 8746 | SDLoc DL(N); |
| 8747 | SDValue Mask = N->getOperand(Num: 0); |
| 8748 | EVT OrigMaskVT = Mask.getValueType(); |
| 8749 | SDValue WideMask = GetWidenedVector(Op: Mask); |
| 8750 | EVT WideMaskVT = WideMask.getValueType(); |
| 8751 | |
| 8752 | // Pad the mask with zeros to ensure inactive lanes don't affect the result. |
| 8753 | unsigned OrigElts = OrigMaskVT.getVectorNumElements(); |
| 8754 | unsigned WideElts = WideMaskVT.getVectorNumElements(); |
| 8755 | if (OrigElts != WideElts) { |
| 8756 | SDValue ZeroMask = DAG.getConstant(Val: 0, DL, VT: WideMaskVT); |
| 8757 | WideMask = DAG.getNode(Opcode: ISD::INSERT_SUBVECTOR, DL, VT: WideMaskVT, N1: ZeroMask, |
| 8758 | N2: Mask, N3: DAG.getVectorIdxConstant(Val: 0, DL)); |
| 8759 | } |
| 8760 | |
| 8761 | return DAG.getNode(Opcode: ISD::VECTOR_FIND_LAST_ACTIVE, DL, VT: N->getValueType(ResNo: 0), |
| 8762 | Operand: WideMask); |
| 8763 | } |
| 8764 | |
| 8765 | //===----------------------------------------------------------------------===// |
| 8766 | // Vector Widening Utilities |
| 8767 | //===----------------------------------------------------------------------===// |
| 8768 | |
| 8769 | // Utility function to find the type to chop up a widen vector for load/store |
| 8770 | // TLI: Target lowering used to determine legal types. |
| 8771 | // Width: Width left need to load/store. |
| 8772 | // WidenVT: The widen vector type to load to/store from |
| 8773 | // Align: If 0, don't allow use of a wider type |
| 8774 | // WidenEx: If Align is not 0, the amount additional we can load/store from. |
| 8775 | |
| 8776 | static std::optional<EVT> findMemType(SelectionDAG &DAG, |
| 8777 | const TargetLowering &TLI, unsigned Width, |
| 8778 | EVT WidenVT, unsigned Align = 0, |
| 8779 | unsigned WidenEx = 0) { |
| 8780 | EVT WidenEltVT = WidenVT.getVectorElementType(); |
| 8781 | const bool Scalable = WidenVT.isScalableVector(); |
| 8782 | unsigned WidenWidth = WidenVT.getSizeInBits().getKnownMinValue(); |
| 8783 | unsigned WidenEltWidth = WidenEltVT.getSizeInBits(); |
| 8784 | unsigned AlignInBits = Align*8; |
| 8785 | |
| 8786 | EVT RetVT = WidenEltVT; |
| 8787 | // Don't bother looking for an integer type if the vector is scalable, skip |
| 8788 | // to vector types. |
| 8789 | if (!Scalable) { |
| 8790 | // If we have one element to load/store, return it. |
| 8791 | if (Width == WidenEltWidth) |
| 8792 | return RetVT; |
| 8793 | |
| 8794 | // See if there is larger legal integer than the element type to load/store. |
| 8795 | for (EVT MemVT : reverse(C: MVT::integer_valuetypes())) { |
| 8796 | unsigned MemVTWidth = MemVT.getSizeInBits(); |
| 8797 | if (MemVT.getSizeInBits() <= WidenEltWidth) |
| 8798 | break; |
| 8799 | auto Action = TLI.getTypeAction(Context&: *DAG.getContext(), VT: MemVT); |
| 8800 | if ((Action == TargetLowering::TypeLegal || |
| 8801 | Action == TargetLowering::TypePromoteInteger) && |
| 8802 | (WidenWidth % MemVTWidth) == 0 && |
| 8803 | isPowerOf2_32(Value: WidenWidth / MemVTWidth) && |
| 8804 | (MemVTWidth <= Width || |
| 8805 | (Align!=0 && MemVTWidth<=AlignInBits && MemVTWidth<=Width+WidenEx))) { |
| 8806 | if (MemVTWidth == WidenWidth) |
| 8807 | return MemVT; |
| 8808 | RetVT = MemVT; |
| 8809 | break; |
| 8810 | } |
| 8811 | } |
| 8812 | } |
| 8813 | |
| 8814 | // See if there is a larger vector type to load/store that has the same vector |
| 8815 | // element type and is evenly divisible with the WidenVT. |
| 8816 | for (EVT MemVT : reverse(C: MVT::vector_valuetypes())) { |
| 8817 | // Skip vector MVTs which don't match the scalable property of WidenVT. |
| 8818 | if (Scalable != MemVT.isScalableVector()) |
| 8819 | continue; |
| 8820 | unsigned MemVTWidth = MemVT.getSizeInBits().getKnownMinValue(); |
| 8821 | auto Action = TLI.getTypeAction(Context&: *DAG.getContext(), VT: MemVT); |
| 8822 | if ((Action == TargetLowering::TypeLegal || |
| 8823 | Action == TargetLowering::TypePromoteInteger) && |
| 8824 | WidenEltVT == MemVT.getVectorElementType() && |
| 8825 | (WidenWidth % MemVTWidth) == 0 && |
| 8826 | isPowerOf2_32(Value: WidenWidth / MemVTWidth) && |
| 8827 | (MemVTWidth <= Width || |
| 8828 | (Align!=0 && MemVTWidth<=AlignInBits && MemVTWidth<=Width+WidenEx))) { |
| 8829 | if (RetVT.getFixedSizeInBits() < MemVTWidth || MemVT == WidenVT) |
| 8830 | return MemVT; |
| 8831 | } |
| 8832 | } |
| 8833 | |
| 8834 | // Using element-wise loads and stores for widening operations is not |
| 8835 | // supported for scalable vectors |
| 8836 | if (Scalable) |
| 8837 | return std::nullopt; |
| 8838 | |
| 8839 | return RetVT; |
| 8840 | } |
| 8841 | |
| 8842 | // Builds a vector type from scalar loads |
| 8843 | // VecTy: Resulting Vector type |
| 8844 | // LDOps: Load operators to build a vector type |
| 8845 | // [Start,End) the list of loads to use. |
| 8846 | static SDValue BuildVectorFromScalar(SelectionDAG& DAG, EVT VecTy, |
| 8847 | SmallVectorImpl<SDValue> &LdOps, |
| 8848 | unsigned Start, unsigned End) { |
| 8849 | SDLoc dl(LdOps[Start]); |
| 8850 | EVT LdTy = LdOps[Start].getValueType(); |
| 8851 | unsigned Width = VecTy.getSizeInBits(); |
| 8852 | unsigned NumElts = Width / LdTy.getSizeInBits(); |
| 8853 | EVT NewVecVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: LdTy, NumElements: NumElts); |
| 8854 | |
| 8855 | unsigned Idx = 1; |
| 8856 | SDValue VecOp = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: dl, VT: NewVecVT,Operand: LdOps[Start]); |
| 8857 | |
| 8858 | for (unsigned i = Start + 1; i != End; ++i) { |
| 8859 | EVT NewLdTy = LdOps[i].getValueType(); |
| 8860 | if (NewLdTy != LdTy) { |
| 8861 | NumElts = Width / NewLdTy.getSizeInBits(); |
| 8862 | NewVecVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: NewLdTy, NumElements: NumElts); |
| 8863 | VecOp = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: NewVecVT, Operand: VecOp); |
| 8864 | // Readjust position and vector position based on new load type. |
| 8865 | Idx = Idx * LdTy.getSizeInBits() / NewLdTy.getSizeInBits(); |
| 8866 | LdTy = NewLdTy; |
| 8867 | } |
| 8868 | VecOp = DAG.getInsertVectorElt(DL: dl, Vec: VecOp, Elt: LdOps[i], Idx: Idx++); |
| 8869 | } |
| 8870 | return DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: VecTy, Operand: VecOp); |
| 8871 | } |
| 8872 | |
| 8873 | SDValue DAGTypeLegalizer::GenWidenVectorLoads(SmallVectorImpl<SDValue> &LdChain, |
| 8874 | LoadSDNode *LD) { |
| 8875 | // The strategy assumes that we can efficiently load power-of-two widths. |
| 8876 | // The routine chops the vector into the largest vector loads with the same |
| 8877 | // element type or scalar loads and then recombines it to the widen vector |
| 8878 | // type. |
| 8879 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(),VT: LD->getValueType(ResNo: 0)); |
| 8880 | EVT LdVT = LD->getMemoryVT(); |
| 8881 | SDLoc dl(LD); |
| 8882 | assert(LdVT.isVector() && WidenVT.isVector()); |
| 8883 | assert(LdVT.isScalableVector() == WidenVT.isScalableVector()); |
| 8884 | assert(LdVT.getVectorElementType() == WidenVT.getVectorElementType()); |
| 8885 | |
| 8886 | // Load information |
| 8887 | SDValue Chain = LD->getChain(); |
| 8888 | SDValue BasePtr = LD->getBasePtr(); |
| 8889 | MachineMemOperand::Flags MMOFlags = LD->getMemOperand()->getFlags(); |
| 8890 | AAMDNodes AAInfo = LD->getAAInfo(); |
| 8891 | |
| 8892 | TypeSize LdWidth = LdVT.getSizeInBits(); |
| 8893 | TypeSize WidenWidth = WidenVT.getSizeInBits(); |
| 8894 | TypeSize WidthDiff = WidenWidth - LdWidth; |
| 8895 | // Allow wider loads if they are sufficiently aligned to avoid memory faults |
| 8896 | // and if the original load is simple. |
| 8897 | unsigned LdAlign = |
| 8898 | (!LD->isSimple() || LdVT.isScalableVector()) ? 0 : LD->getAlign().value(); |
| 8899 | |
| 8900 | // Find the vector type that can load from. |
| 8901 | std::optional<EVT> FirstVT = |
| 8902 | findMemType(DAG, TLI, Width: LdWidth.getKnownMinValue(), WidenVT, Align: LdAlign, |
| 8903 | WidenEx: WidthDiff.getKnownMinValue()); |
| 8904 | |
| 8905 | if (!FirstVT) |
| 8906 | return SDValue(); |
| 8907 | |
| 8908 | SmallVector<EVT, 8> MemVTs; |
| 8909 | TypeSize FirstVTWidth = FirstVT->getSizeInBits(); |
| 8910 | |
| 8911 | // Unless we're able to load in one instruction we must work out how to load |
| 8912 | // the remainder. |
| 8913 | if (!TypeSize::isKnownLE(LHS: LdWidth, RHS: FirstVTWidth)) { |
| 8914 | std::optional<EVT> NewVT = FirstVT; |
| 8915 | TypeSize RemainingWidth = LdWidth; |
| 8916 | TypeSize NewVTWidth = FirstVTWidth; |
| 8917 | do { |
| 8918 | RemainingWidth -= NewVTWidth; |
| 8919 | if (TypeSize::isKnownLT(LHS: RemainingWidth, RHS: NewVTWidth)) { |
| 8920 | // The current type we are using is too large. Find a better size. |
| 8921 | NewVT = findMemType(DAG, TLI, Width: RemainingWidth.getKnownMinValue(), |
| 8922 | WidenVT, Align: LdAlign, WidenEx: WidthDiff.getKnownMinValue()); |
| 8923 | if (!NewVT) |
| 8924 | return SDValue(); |
| 8925 | NewVTWidth = NewVT->getSizeInBits(); |
| 8926 | } |
| 8927 | MemVTs.push_back(Elt: *NewVT); |
| 8928 | } while (TypeSize::isKnownGT(LHS: RemainingWidth, RHS: NewVTWidth)); |
| 8929 | } |
| 8930 | |
| 8931 | SDValue LdOp = DAG.getLoad(VT: *FirstVT, dl, Chain, Ptr: BasePtr, PtrInfo: LD->getPointerInfo(), |
| 8932 | Alignment: LD->getBaseAlign(), MMOFlags, AAInfo); |
| 8933 | LdChain.push_back(Elt: LdOp.getValue(R: 1)); |
| 8934 | |
| 8935 | // Check if we can load the element with one instruction. |
| 8936 | if (MemVTs.empty()) |
| 8937 | return coerceLoadedValue(LdOp, FirstVT: *FirstVT, WidenVT, LdWidth, FirstVTWidth, dl, |
| 8938 | DAG); |
| 8939 | |
| 8940 | // Load vector by using multiple loads from largest vector to scalar. |
| 8941 | SmallVector<SDValue, 16> LdOps; |
| 8942 | LdOps.push_back(Elt: LdOp); |
| 8943 | |
| 8944 | uint64_t ScaledOffset = 0; |
| 8945 | MachinePointerInfo MPI = LD->getPointerInfo(); |
| 8946 | |
| 8947 | // First incremement past the first load. |
| 8948 | IncrementPointer(N: cast<LoadSDNode>(Val&: LdOp), MemVT: *FirstVT, MPI, Ptr&: BasePtr, |
| 8949 | ScaledOffset: &ScaledOffset); |
| 8950 | |
| 8951 | for (EVT MemVT : MemVTs) { |
| 8952 | Align NewAlign = ScaledOffset == 0 |
| 8953 | ? LD->getBaseAlign() |
| 8954 | : commonAlignment(A: LD->getAlign(), Offset: ScaledOffset); |
| 8955 | SDValue L = |
| 8956 | DAG.getLoad(VT: MemVT, dl, Chain, Ptr: BasePtr, PtrInfo: MPI, Alignment: NewAlign, MMOFlags, AAInfo); |
| 8957 | |
| 8958 | LdOps.push_back(Elt: L); |
| 8959 | LdChain.push_back(Elt: L.getValue(R: 1)); |
| 8960 | IncrementPointer(N: cast<LoadSDNode>(Val&: L), MemVT, MPI, Ptr&: BasePtr, ScaledOffset: &ScaledOffset); |
| 8961 | } |
| 8962 | |
| 8963 | // Build the vector from the load operations. |
| 8964 | unsigned End = LdOps.size(); |
| 8965 | if (!LdOps[0].getValueType().isVector()) |
| 8966 | // All the loads are scalar loads. |
| 8967 | return BuildVectorFromScalar(DAG, VecTy: WidenVT, LdOps, Start: 0, End); |
| 8968 | |
| 8969 | // If the load contains vectors, build the vector using concat vector. |
| 8970 | // All of the vectors used to load are power-of-2, and the scalar loads can be |
| 8971 | // combined to make a power-of-2 vector. |
| 8972 | SmallVector<SDValue, 16> ConcatOps(End); |
| 8973 | int i = End - 1; |
| 8974 | int Idx = End; |
| 8975 | EVT LdTy = LdOps[i].getValueType(); |
| 8976 | // First, combine the scalar loads to a vector. |
| 8977 | if (!LdTy.isVector()) { |
| 8978 | for (--i; i >= 0; --i) { |
| 8979 | LdTy = LdOps[i].getValueType(); |
| 8980 | if (LdTy.isVector()) |
| 8981 | break; |
| 8982 | } |
| 8983 | ConcatOps[--Idx] = BuildVectorFromScalar(DAG, VecTy: LdTy, LdOps, Start: i + 1, End); |
| 8984 | } |
| 8985 | |
| 8986 | ConcatOps[--Idx] = LdOps[i]; |
| 8987 | for (--i; i >= 0; --i) { |
| 8988 | EVT NewLdTy = LdOps[i].getValueType(); |
| 8989 | if (NewLdTy != LdTy) { |
| 8990 | // Create a larger vector. |
| 8991 | TypeSize LdTySize = LdTy.getSizeInBits(); |
| 8992 | TypeSize NewLdTySize = NewLdTy.getSizeInBits(); |
| 8993 | assert(NewLdTySize.isScalable() == LdTySize.isScalable() && |
| 8994 | NewLdTySize.isKnownMultipleOf(LdTySize.getKnownMinValue())); |
| 8995 | unsigned NumOps = |
| 8996 | NewLdTySize.getKnownMinValue() / LdTySize.getKnownMinValue(); |
| 8997 | SmallVector<SDValue, 16> WidenOps(NumOps); |
| 8998 | unsigned j = 0; |
| 8999 | for (; j != End-Idx; ++j) |
| 9000 | WidenOps[j] = ConcatOps[Idx+j]; |
| 9001 | for (; j != NumOps; ++j) |
| 9002 | WidenOps[j] = DAG.getPOISON(VT: LdTy); |
| 9003 | |
| 9004 | ConcatOps[End-1] = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: NewLdTy, |
| 9005 | Ops: WidenOps); |
| 9006 | Idx = End - 1; |
| 9007 | LdTy = NewLdTy; |
| 9008 | } |
| 9009 | ConcatOps[--Idx] = LdOps[i]; |
| 9010 | } |
| 9011 | |
| 9012 | if (WidenWidth == LdTy.getSizeInBits() * (End - Idx)) |
| 9013 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: WidenVT, |
| 9014 | Ops: ArrayRef(&ConcatOps[Idx], End - Idx)); |
| 9015 | |
| 9016 | // We need to fill the rest with undefs to build the vector. |
| 9017 | unsigned NumOps = |
| 9018 | WidenWidth.getKnownMinValue() / LdTy.getSizeInBits().getKnownMinValue(); |
| 9019 | SmallVector<SDValue, 16> WidenOps(NumOps); |
| 9020 | SDValue UndefVal = DAG.getPOISON(VT: LdTy); |
| 9021 | { |
| 9022 | unsigned i = 0; |
| 9023 | for (; i != End-Idx; ++i) |
| 9024 | WidenOps[i] = ConcatOps[Idx+i]; |
| 9025 | for (; i != NumOps; ++i) |
| 9026 | WidenOps[i] = UndefVal; |
| 9027 | } |
| 9028 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: WidenVT, Ops: WidenOps); |
| 9029 | } |
| 9030 | |
| 9031 | SDValue |
| 9032 | DAGTypeLegalizer::GenWidenVectorExtLoads(SmallVectorImpl<SDValue> &LdChain, |
| 9033 | LoadSDNode *LD, |
| 9034 | ISD::LoadExtType ExtType) { |
| 9035 | // For extension loads, it may not be more efficient to chop up the vector |
| 9036 | // and then extend it. Instead, we unroll the load and build a new vector. |
| 9037 | EVT WidenVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(),VT: LD->getValueType(ResNo: 0)); |
| 9038 | EVT LdVT = LD->getMemoryVT(); |
| 9039 | SDLoc dl(LD); |
| 9040 | assert(LdVT.isVector() && WidenVT.isVector()); |
| 9041 | assert(LdVT.isScalableVector() == WidenVT.isScalableVector()); |
| 9042 | |
| 9043 | // Load information |
| 9044 | SDValue Chain = LD->getChain(); |
| 9045 | SDValue BasePtr = LD->getBasePtr(); |
| 9046 | MachineMemOperand::Flags MMOFlags = LD->getMemOperand()->getFlags(); |
| 9047 | AAMDNodes AAInfo = LD->getAAInfo(); |
| 9048 | |
| 9049 | if (LdVT.isScalableVector()) |
| 9050 | return SDValue(); |
| 9051 | |
| 9052 | EVT EltVT = WidenVT.getVectorElementType(); |
| 9053 | EVT LdEltVT = LdVT.getVectorElementType(); |
| 9054 | unsigned NumElts = LdVT.getVectorNumElements(); |
| 9055 | |
| 9056 | // Load each element and widen. |
| 9057 | unsigned WidenNumElts = WidenVT.getVectorNumElements(); |
| 9058 | SmallVector<SDValue, 16> Ops(WidenNumElts); |
| 9059 | unsigned Increment = LdEltVT.getSizeInBits() / 8; |
| 9060 | Ops[0] = |
| 9061 | DAG.getExtLoad(ExtType, dl, VT: EltVT, Chain, Ptr: BasePtr, PtrInfo: LD->getPointerInfo(), |
| 9062 | MemVT: LdEltVT, Alignment: LD->getBaseAlign(), MMOFlags, AAInfo); |
| 9063 | LdChain.push_back(Elt: Ops[0].getValue(R: 1)); |
| 9064 | unsigned i = 0, Offset = Increment; |
| 9065 | for (i=1; i < NumElts; ++i, Offset += Increment) { |
| 9066 | SDValue NewBasePtr = |
| 9067 | DAG.getObjectPtrOffset(SL: dl, Ptr: BasePtr, Offset: TypeSize::getFixed(ExactSize: Offset)); |
| 9068 | Ops[i] = DAG.getExtLoad(ExtType, dl, VT: EltVT, Chain, Ptr: NewBasePtr, |
| 9069 | PtrInfo: LD->getPointerInfo().getWithOffset(O: Offset), MemVT: LdEltVT, |
| 9070 | Alignment: LD->getBaseAlign(), MMOFlags, AAInfo); |
| 9071 | LdChain.push_back(Elt: Ops[i].getValue(R: 1)); |
| 9072 | } |
| 9073 | |
| 9074 | // Fill the rest with undefs. |
| 9075 | SDValue UndefVal = DAG.getPOISON(VT: EltVT); |
| 9076 | for (; i != WidenNumElts; ++i) |
| 9077 | Ops[i] = UndefVal; |
| 9078 | |
| 9079 | return DAG.getBuildVector(VT: WidenVT, DL: dl, Ops); |
| 9080 | } |
| 9081 | |
| 9082 | bool DAGTypeLegalizer::GenWidenVectorStores(SmallVectorImpl<SDValue> &StChain, |
| 9083 | StoreSDNode *ST) { |
| 9084 | // The strategy assumes that we can efficiently store power-of-two widths. |
| 9085 | // The routine chops the vector into the largest vector stores with the same |
| 9086 | // element type or scalar stores. |
| 9087 | SDValue Chain = ST->getChain(); |
| 9088 | SDValue BasePtr = ST->getBasePtr(); |
| 9089 | MachineMemOperand::Flags MMOFlags = ST->getMemOperand()->getFlags(); |
| 9090 | AAMDNodes AAInfo = ST->getAAInfo(); |
| 9091 | SDValue ValOp = GetWidenedVector(Op: ST->getValue()); |
| 9092 | SDLoc dl(ST); |
| 9093 | |
| 9094 | EVT StVT = ST->getMemoryVT(); |
| 9095 | TypeSize StWidth = StVT.getSizeInBits(); |
| 9096 | EVT ValVT = ValOp.getValueType(); |
| 9097 | TypeSize ValWidth = ValVT.getSizeInBits(); |
| 9098 | EVT ValEltVT = ValVT.getVectorElementType(); |
| 9099 | unsigned ValEltWidth = ValEltVT.getFixedSizeInBits(); |
| 9100 | assert(StVT.getVectorElementType() == ValEltVT); |
| 9101 | assert(StVT.isScalableVector() == ValVT.isScalableVector() && |
| 9102 | "Mismatch between store and value types" ); |
| 9103 | |
| 9104 | int Idx = 0; // current index to store |
| 9105 | |
| 9106 | MachinePointerInfo MPI = ST->getPointerInfo(); |
| 9107 | uint64_t ScaledOffset = 0; |
| 9108 | |
| 9109 | // A breakdown of how to widen this vector store. Each element of the vector |
| 9110 | // is a memory VT combined with the number of times it is to be stored to, |
| 9111 | // e,g., v5i32 -> {{v2i32,2},{i32,1}} |
| 9112 | SmallVector<std::pair<EVT, unsigned>, 4> MemVTs; |
| 9113 | |
| 9114 | while (StWidth.isNonZero()) { |
| 9115 | // Find the largest vector type we can store with. |
| 9116 | std::optional<EVT> NewVT = |
| 9117 | findMemType(DAG, TLI, Width: StWidth.getKnownMinValue(), WidenVT: ValVT); |
| 9118 | if (!NewVT) |
| 9119 | return false; |
| 9120 | MemVTs.push_back(Elt: {*NewVT, 0}); |
| 9121 | TypeSize NewVTWidth = NewVT->getSizeInBits(); |
| 9122 | |
| 9123 | do { |
| 9124 | StWidth -= NewVTWidth; |
| 9125 | MemVTs.back().second++; |
| 9126 | } while (StWidth.isNonZero() && TypeSize::isKnownGE(LHS: StWidth, RHS: NewVTWidth)); |
| 9127 | } |
| 9128 | |
| 9129 | for (const auto &Pair : MemVTs) { |
| 9130 | EVT NewVT = Pair.first; |
| 9131 | unsigned Count = Pair.second; |
| 9132 | TypeSize NewVTWidth = NewVT.getSizeInBits(); |
| 9133 | |
| 9134 | if (NewVT.isVector()) { |
| 9135 | unsigned NumVTElts = NewVT.getVectorMinNumElements(); |
| 9136 | do { |
| 9137 | Align NewAlign = ScaledOffset == 0 |
| 9138 | ? ST->getBaseAlign() |
| 9139 | : commonAlignment(A: ST->getAlign(), Offset: ScaledOffset); |
| 9140 | SDValue EOp = DAG.getExtractSubvector(DL: dl, VT: NewVT, Vec: ValOp, Idx); |
| 9141 | SDValue PartStore = DAG.getStore(Chain, dl, Val: EOp, Ptr: BasePtr, PtrInfo: MPI, Alignment: NewAlign, |
| 9142 | MMOFlags, AAInfo); |
| 9143 | StChain.push_back(Elt: PartStore); |
| 9144 | |
| 9145 | Idx += NumVTElts; |
| 9146 | IncrementPointer(N: cast<StoreSDNode>(Val&: PartStore), MemVT: NewVT, MPI, Ptr&: BasePtr, |
| 9147 | ScaledOffset: &ScaledOffset); |
| 9148 | } while (--Count); |
| 9149 | } else { |
| 9150 | // Cast the vector to the scalar type we can store. |
| 9151 | unsigned NumElts = ValWidth.getFixedValue() / NewVTWidth.getFixedValue(); |
| 9152 | EVT NewVecVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: NewVT, NumElements: NumElts); |
| 9153 | SDValue VecOp = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: NewVecVT, Operand: ValOp); |
| 9154 | // Readjust index position based on new vector type. |
| 9155 | Idx = Idx * ValEltWidth / NewVTWidth.getFixedValue(); |
| 9156 | do { |
| 9157 | SDValue EOp = DAG.getExtractVectorElt(DL: dl, VT: NewVT, Vec: VecOp, Idx: Idx++); |
| 9158 | SDValue PartStore = DAG.getStore(Chain, dl, Val: EOp, Ptr: BasePtr, PtrInfo: MPI, |
| 9159 | Alignment: ST->getBaseAlign(), MMOFlags, AAInfo); |
| 9160 | StChain.push_back(Elt: PartStore); |
| 9161 | |
| 9162 | IncrementPointer(N: cast<StoreSDNode>(Val&: PartStore), MemVT: NewVT, MPI, Ptr&: BasePtr); |
| 9163 | } while (--Count); |
| 9164 | // Restore index back to be relative to the original widen element type. |
| 9165 | Idx = Idx * NewVTWidth.getFixedValue() / ValEltWidth; |
| 9166 | } |
| 9167 | } |
| 9168 | |
| 9169 | return true; |
| 9170 | } |
| 9171 | |
| 9172 | /// Modifies a vector input (widen or narrows) to a vector of NVT. The |
| 9173 | /// input vector must have the same element type as NVT. |
| 9174 | /// FillWithZeroes specifies that the vector should be widened with zeroes. |
| 9175 | SDValue DAGTypeLegalizer::ModifyToType(SDValue InOp, EVT NVT, |
| 9176 | bool FillWithZeroes) { |
| 9177 | // Note that InOp might have been widened so it might already have |
| 9178 | // the right width or it might need be narrowed. |
| 9179 | EVT InVT = InOp.getValueType(); |
| 9180 | assert(InVT.getVectorElementType() == NVT.getVectorElementType() && |
| 9181 | "input and widen element type must match" ); |
| 9182 | assert(InVT.isScalableVector() == NVT.isScalableVector() && |
| 9183 | "cannot modify scalable vectors in this way" ); |
| 9184 | SDLoc dl(InOp); |
| 9185 | |
| 9186 | // Check if InOp already has the right width. |
| 9187 | if (InVT == NVT) |
| 9188 | return InOp; |
| 9189 | |
| 9190 | ElementCount InEC = InVT.getVectorElementCount(); |
| 9191 | ElementCount WidenEC = NVT.getVectorElementCount(); |
| 9192 | if (WidenEC.hasKnownScalarFactor(RHS: InEC)) { |
| 9193 | unsigned NumConcat = WidenEC.getKnownScalarFactor(RHS: InEC); |
| 9194 | SmallVector<SDValue, 16> Ops(NumConcat); |
| 9195 | SDValue FillVal = |
| 9196 | FillWithZeroes ? DAG.getConstant(Val: 0, DL: dl, VT: InVT) : DAG.getPOISON(VT: InVT); |
| 9197 | Ops[0] = InOp; |
| 9198 | for (unsigned i = 1; i != NumConcat; ++i) |
| 9199 | Ops[i] = FillVal; |
| 9200 | |
| 9201 | return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: dl, VT: NVT, Ops); |
| 9202 | } |
| 9203 | |
| 9204 | if (InEC.hasKnownScalarFactor(RHS: WidenEC)) |
| 9205 | return DAG.getExtractSubvector(DL: dl, VT: NVT, Vec: InOp, Idx: 0); |
| 9206 | |
| 9207 | assert(!InVT.isScalableVector() && !NVT.isScalableVector() && |
| 9208 | "Scalable vectors should have been handled already." ); |
| 9209 | |
| 9210 | unsigned InNumElts = InEC.getFixedValue(); |
| 9211 | unsigned WidenNumElts = WidenEC.getFixedValue(); |
| 9212 | |
| 9213 | // Fall back to extract and build (+ mask, if padding with zeros). |
| 9214 | SmallVector<SDValue, 16> Ops(WidenNumElts); |
| 9215 | EVT EltVT = NVT.getVectorElementType(); |
| 9216 | unsigned MinNumElts = std::min(a: WidenNumElts, b: InNumElts); |
| 9217 | unsigned Idx; |
| 9218 | for (Idx = 0; Idx < MinNumElts; ++Idx) |
| 9219 | Ops[Idx] = DAG.getExtractVectorElt(DL: dl, VT: EltVT, Vec: InOp, Idx); |
| 9220 | |
| 9221 | SDValue UndefVal = DAG.getPOISON(VT: EltVT); |
| 9222 | for (; Idx < WidenNumElts; ++Idx) |
| 9223 | Ops[Idx] = UndefVal; |
| 9224 | |
| 9225 | SDValue Widened = DAG.getBuildVector(VT: NVT, DL: dl, Ops); |
| 9226 | if (!FillWithZeroes) |
| 9227 | return Widened; |
| 9228 | |
| 9229 | assert(NVT.isInteger() && |
| 9230 | "We expect to never want to FillWithZeroes for non-integral types." ); |
| 9231 | |
| 9232 | SmallVector<SDValue, 16> MaskOps; |
| 9233 | MaskOps.append(NumInputs: MinNumElts, Elt: DAG.getAllOnesConstant(DL: dl, VT: EltVT)); |
| 9234 | MaskOps.append(NumInputs: WidenNumElts - MinNumElts, Elt: DAG.getConstant(Val: 0, DL: dl, VT: EltVT)); |
| 9235 | |
| 9236 | return DAG.getNode(Opcode: ISD::AND, DL: dl, VT: NVT, N1: Widened, |
| 9237 | N2: DAG.getBuildVector(VT: NVT, DL: dl, Ops: MaskOps)); |
| 9238 | } |
| 9239 | |