1 //===-- LegalizeVectorOps.cpp - Implement SelectionDAG::LegalizeVectors ---===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the SelectionDAG::LegalizeVectors method. 11 // 12 // The vector legalizer looks for vector operations which might need to be 13 // scalarized and legalizes them. This is a separate step from Legalize because 14 // scalarizing can introduce illegal types. For example, suppose we have an 15 // ISD::SDIV of type v2i64 on x86-32. The type is legal (for example, addition 16 // on a v2i64 is legal), but ISD::SDIV isn't legal, so we have to unroll the 17 // operation, which introduces nodes with the illegal type i64 which must be 18 // expanded. Similarly, suppose we have an ISD::SRA of type v16i8 on PowerPC; 19 // the operation must be unrolled, which introduces nodes with the illegal 20 // type i8 which must be promoted. 21 // 22 // This does not legalize vector manipulations like ISD::BUILD_VECTOR, 23 // or operations that happen to take a vector which are custom-lowered; 24 // the legalization for such operations never produces nodes 25 // with illegal types, so it's okay to put off legalizing them until 26 // SelectionDAG::Legalize runs. 27 // 28 //===----------------------------------------------------------------------===// 29 30 #include "llvm/CodeGen/SelectionDAG.h" 31 #include "llvm/Target/TargetLowering.h" 32 using namespace llvm; 33 34 namespace { 35 class VectorLegalizer { 36 SelectionDAG& DAG; 37 const TargetLowering &TLI; 38 bool Changed; // Keep track of whether anything changed 39 40 /// LegalizedNodes - For nodes that are of legal width, and that have more 41 /// than one use, this map indicates what regularized operand to use. This 42 /// allows us to avoid legalizing the same thing more than once. 43 DenseMap<SDValue, SDValue> LegalizedNodes; 44 45 // Adds a node to the translation cache 46 void AddLegalizedOperand(SDValue From, SDValue To) { 47 LegalizedNodes.insert(std::make_pair(From, To)); 48 // If someone requests legalization of the new node, return itself. 49 if (From != To) 50 LegalizedNodes.insert(std::make_pair(To, To)); 51 } 52 53 // Legalizes the given node 54 SDValue LegalizeOp(SDValue Op); 55 // Assuming the node is legal, "legalize" the results 56 SDValue TranslateLegalizeResults(SDValue Op, SDValue Result); 57 // Implements unrolling a VSETCC. 58 SDValue UnrollVSETCC(SDValue Op); 59 // Implements expansion for FNEG; falls back to UnrollVectorOp if FSUB 60 // isn't legal. 61 // Implements expansion for UINT_TO_FLOAT; falls back to UnrollVectorOp if 62 // SINT_TO_FLOAT and SHR on vectors isn't legal. 63 SDValue ExpandUINT_TO_FLOAT(SDValue Op); 64 // Implement vselect in terms of XOR, AND, OR when blend is not supported 65 // by the target. 66 SDValue ExpandVSELECT(SDValue Op); 67 SDValue ExpandFNEG(SDValue Op); 68 // Implements vector promotion; this is essentially just bitcasting the 69 // operands to a different type and bitcasting the result back to the 70 // original type. 71 SDValue PromoteVectorOp(SDValue Op); 72 73 public: 74 bool Run(); 75 VectorLegalizer(SelectionDAG& dag) : 76 DAG(dag), TLI(dag.getTargetLoweringInfo()), Changed(false) {} 77 }; 78 79 bool VectorLegalizer::Run() { 80 // The legalize process is inherently a bottom-up recursive process (users 81 // legalize their uses before themselves). Given infinite stack space, we 82 // could just start legalizing on the root and traverse the whole graph. In 83 // practice however, this causes us to run out of stack space on large basic 84 // blocks. To avoid this problem, compute an ordering of the nodes where each 85 // node is only legalized after all of its operands are legalized. 86 DAG.AssignTopologicalOrder(); 87 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(), 88 E = prior(DAG.allnodes_end()); I != llvm::next(E); ++I) 89 LegalizeOp(SDValue(I, 0)); 90 91 // Finally, it's possible the root changed. Get the new root. 92 SDValue OldRoot = DAG.getRoot(); 93 assert(LegalizedNodes.count(OldRoot) && "Root didn't get legalized?"); 94 DAG.setRoot(LegalizedNodes[OldRoot]); 95 96 LegalizedNodes.clear(); 97 98 // Remove dead nodes now. 99 DAG.RemoveDeadNodes(); 100 101 return Changed; 102 } 103 104 SDValue VectorLegalizer::TranslateLegalizeResults(SDValue Op, SDValue Result) { 105 // Generic legalization: just pass the operand through. 106 for (unsigned i = 0, e = Op.getNode()->getNumValues(); i != e; ++i) 107 AddLegalizedOperand(Op.getValue(i), Result.getValue(i)); 108 return Result.getValue(Op.getResNo()); 109 } 110 111 SDValue VectorLegalizer::LegalizeOp(SDValue Op) { 112 // Note that LegalizeOp may be reentered even from single-use nodes, which 113 // means that we always must cache transformed nodes. 114 DenseMap<SDValue, SDValue>::iterator I = LegalizedNodes.find(Op); 115 if (I != LegalizedNodes.end()) return I->second; 116 117 SDNode* Node = Op.getNode(); 118 119 // Legalize the operands 120 SmallVector<SDValue, 8> Ops; 121 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) 122 Ops.push_back(LegalizeOp(Node->getOperand(i))); 123 124 SDValue Result = 125 SDValue(DAG.UpdateNodeOperands(Op.getNode(), Ops.data(), Ops.size()), 0); 126 127 bool HasVectorValue = false; 128 for (SDNode::value_iterator J = Node->value_begin(), E = Node->value_end(); 129 J != E; 130 ++J) 131 HasVectorValue |= J->isVector(); 132 if (!HasVectorValue) 133 return TranslateLegalizeResults(Op, Result); 134 135 EVT QueryType; 136 switch (Op.getOpcode()) { 137 default: 138 return TranslateLegalizeResults(Op, Result); 139 case ISD::ADD: 140 case ISD::SUB: 141 case ISD::MUL: 142 case ISD::SDIV: 143 case ISD::UDIV: 144 case ISD::SREM: 145 case ISD::UREM: 146 case ISD::FADD: 147 case ISD::FSUB: 148 case ISD::FMUL: 149 case ISD::FDIV: 150 case ISD::FREM: 151 case ISD::AND: 152 case ISD::OR: 153 case ISD::XOR: 154 case ISD::SHL: 155 case ISD::SRA: 156 case ISD::SRL: 157 case ISD::ROTL: 158 case ISD::ROTR: 159 case ISD::CTTZ: 160 case ISD::CTLZ: 161 case ISD::CTPOP: 162 case ISD::SELECT: 163 case ISD::VSELECT: 164 case ISD::SELECT_CC: 165 case ISD::SETCC: 166 case ISD::ZERO_EXTEND: 167 case ISD::ANY_EXTEND: 168 case ISD::TRUNCATE: 169 case ISD::SIGN_EXTEND: 170 case ISD::FP_TO_SINT: 171 case ISD::FP_TO_UINT: 172 case ISD::FNEG: 173 case ISD::FABS: 174 case ISD::FSQRT: 175 case ISD::FSIN: 176 case ISD::FCOS: 177 case ISD::FPOWI: 178 case ISD::FPOW: 179 case ISD::FLOG: 180 case ISD::FLOG2: 181 case ISD::FLOG10: 182 case ISD::FEXP: 183 case ISD::FEXP2: 184 case ISD::FCEIL: 185 case ISD::FTRUNC: 186 case ISD::FRINT: 187 case ISD::FNEARBYINT: 188 case ISD::FFLOOR: 189 case ISD::SIGN_EXTEND_INREG: 190 QueryType = Node->getValueType(0); 191 break; 192 case ISD::FP_ROUND_INREG: 193 QueryType = cast<VTSDNode>(Node->getOperand(1))->getVT(); 194 break; 195 case ISD::SINT_TO_FP: 196 case ISD::UINT_TO_FP: 197 QueryType = Node->getOperand(0).getValueType(); 198 break; 199 } 200 201 switch (TLI.getOperationAction(Node->getOpcode(), QueryType)) { 202 case TargetLowering::Promote: 203 // "Promote" the operation by bitcasting 204 Result = PromoteVectorOp(Op); 205 Changed = true; 206 break; 207 case TargetLowering::Legal: break; 208 case TargetLowering::Custom: { 209 SDValue Tmp1 = TLI.LowerOperation(Op, DAG); 210 if (Tmp1.getNode()) { 211 Result = Tmp1; 212 break; 213 } 214 // FALL THROUGH 215 } 216 case TargetLowering::Expand: 217 if (Node->getOpcode() == ISD::VSELECT) 218 Result = ExpandVSELECT(Op); 219 else if (Node->getOpcode() == ISD::UINT_TO_FP) 220 Result = ExpandUINT_TO_FLOAT(Op); 221 else if (Node->getOpcode() == ISD::FNEG) 222 Result = ExpandFNEG(Op); 223 else if (Node->getOpcode() == ISD::SETCC) 224 Result = UnrollVSETCC(Op); 225 else 226 Result = DAG.UnrollVectorOp(Op.getNode()); 227 break; 228 } 229 230 // Make sure that the generated code is itself legal. 231 if (Result != Op) { 232 Result = LegalizeOp(Result); 233 Changed = true; 234 } 235 236 // Note that LegalizeOp may be reentered even from single-use nodes, which 237 // means that we always must cache transformed nodes. 238 AddLegalizedOperand(Op, Result); 239 return Result; 240 } 241 242 SDValue VectorLegalizer::PromoteVectorOp(SDValue Op) { 243 // Vector "promotion" is basically just bitcasting and doing the operation 244 // in a different type. For example, x86 promotes ISD::AND on v2i32 to 245 // v1i64. 246 EVT VT = Op.getValueType(); 247 assert(Op.getNode()->getNumValues() == 1 && 248 "Can't promote a vector with multiple results!"); 249 EVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT); 250 DebugLoc dl = Op.getDebugLoc(); 251 SmallVector<SDValue, 4> Operands(Op.getNumOperands()); 252 253 for (unsigned j = 0; j != Op.getNumOperands(); ++j) { 254 if (Op.getOperand(j).getValueType().isVector()) 255 Operands[j] = DAG.getNode(ISD::BITCAST, dl, NVT, Op.getOperand(j)); 256 else 257 Operands[j] = Op.getOperand(j); 258 } 259 260 Op = DAG.getNode(Op.getOpcode(), dl, NVT, &Operands[0], Operands.size()); 261 262 return DAG.getNode(ISD::BITCAST, dl, VT, Op); 263 } 264 265 SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) { 266 // Implement VSELECT in terms of XOR, AND, OR 267 // on platforms which do not support blend natively. 268 EVT VT = Op.getOperand(0).getValueType(); 269 DebugLoc DL = Op.getDebugLoc(); 270 271 SDValue Mask = Op.getOperand(0); 272 SDValue Op1 = Op.getOperand(1); 273 SDValue Op2 = Op.getOperand(2); 274 275 // If we can't even use the basic vector operations of 276 // AND,OR,XOR, we will have to scalarize the op. 277 if (!TLI.isOperationLegalOrCustom(ISD::AND, VT) || 278 !TLI.isOperationLegalOrCustom(ISD::XOR, VT) || 279 !TLI.isOperationLegalOrCustom(ISD::OR, VT)) 280 return DAG.UnrollVectorOp(Op.getNode()); 281 282 assert(VT.getSizeInBits() == Op.getOperand(1).getValueType().getSizeInBits() 283 && "Invalid mask size"); 284 // Bitcast the operands to be the same type as the mask. 285 // This is needed when we select between FP types because 286 // the mask is a vector of integers. 287 Op1 = DAG.getNode(ISD::BITCAST, DL, VT, Op1); 288 Op2 = DAG.getNode(ISD::BITCAST, DL, VT, Op2); 289 290 SDValue AllOnes = DAG.getConstant( 291 APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()), VT); 292 SDValue NotMask = DAG.getNode(ISD::XOR, DL, VT, Mask, AllOnes); 293 294 Op1 = DAG.getNode(ISD::AND, DL, VT, Op1, Mask); 295 Op2 = DAG.getNode(ISD::AND, DL, VT, Op2, NotMask); 296 return DAG.getNode(ISD::OR, DL, VT, Op1, Op2); 297 } 298 299 SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) { 300 EVT VT = Op.getOperand(0).getValueType(); 301 DebugLoc DL = Op.getDebugLoc(); 302 303 // Make sure that the SINT_TO_FP and SRL instructions are available. 304 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, VT) || 305 !TLI.isOperationLegalOrCustom(ISD::SRL, VT)) 306 return DAG.UnrollVectorOp(Op.getNode()); 307 308 EVT SVT = VT.getScalarType(); 309 assert((SVT.getSizeInBits() == 64 || SVT.getSizeInBits() == 32) && 310 "Elements in vector-UINT_TO_FP must be 32 or 64 bits wide"); 311 312 unsigned BW = SVT.getSizeInBits(); 313 SDValue HalfWord = DAG.getConstant(BW/2, VT); 314 315 // Constants to clear the upper part of the word. 316 // Notice that we can also use SHL+SHR, but using a constant is slightly 317 // faster on x86. 318 uint64_t HWMask = (SVT.getSizeInBits()==64)?0x00000000FFFFFFFF:0x0000FFFF; 319 SDValue HalfWordMask = DAG.getConstant(HWMask, VT); 320 321 // Two to the power of half-word-size. 322 SDValue TWOHW = DAG.getConstantFP((1<<(BW/2)), Op.getValueType()); 323 324 // Clear upper part of LO, lower HI 325 SDValue HI = DAG.getNode(ISD::SRL, DL, VT, Op.getOperand(0), HalfWord); 326 SDValue LO = DAG.getNode(ISD::AND, DL, VT, Op.getOperand(0), HalfWordMask); 327 328 // Convert hi and lo to floats 329 // Convert the hi part back to the upper values 330 SDValue fHI = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), HI); 331 fHI = DAG.getNode(ISD::FMUL, DL, Op.getValueType(), fHI, TWOHW); 332 SDValue fLO = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), LO); 333 334 // Add the two halves 335 return DAG.getNode(ISD::FADD, DL, Op.getValueType(), fHI, fLO); 336 } 337 338 339 SDValue VectorLegalizer::ExpandFNEG(SDValue Op) { 340 if (TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) { 341 SDValue Zero = DAG.getConstantFP(-0.0, Op.getValueType()); 342 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(), 343 Zero, Op.getOperand(0)); 344 } 345 return DAG.UnrollVectorOp(Op.getNode()); 346 } 347 348 SDValue VectorLegalizer::UnrollVSETCC(SDValue Op) { 349 EVT VT = Op.getValueType(); 350 unsigned NumElems = VT.getVectorNumElements(); 351 EVT EltVT = VT.getVectorElementType(); 352 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1), CC = Op.getOperand(2); 353 EVT TmpEltVT = LHS.getValueType().getVectorElementType(); 354 DebugLoc dl = Op.getDebugLoc(); 355 SmallVector<SDValue, 8> Ops(NumElems); 356 for (unsigned i = 0; i < NumElems; ++i) { 357 SDValue LHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS, 358 DAG.getIntPtrConstant(i)); 359 SDValue RHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS, 360 DAG.getIntPtrConstant(i)); 361 Ops[i] = DAG.getNode(ISD::SETCC, dl, TLI.getSetCCResultType(TmpEltVT), 362 LHSElem, RHSElem, CC); 363 Ops[i] = DAG.getNode(ISD::SELECT, dl, EltVT, Ops[i], 364 DAG.getConstant(APInt::getAllOnesValue 365 (EltVT.getSizeInBits()), EltVT), 366 DAG.getConstant(0, EltVT)); 367 } 368 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], NumElems); 369 } 370 371 } 372 373 bool SelectionDAG::LegalizeVectors() { 374 return VectorLegalizer(*this).Run(); 375 } 376