SelectionDAGBuilder.cpp revision 102b00fc58d0800bf8821316133b3c908d51f904
1//===-- SelectionDAGBuilder.cpp - Selection-DAG building ------------------===// 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 implements routines for translating from LLVM IR into SelectionDAG IR. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "isel" 15#include "SDNodeDbgValue.h" 16#include "SelectionDAGBuilder.h" 17#include "FunctionLoweringInfo.h" 18#include "llvm/ADT/BitVector.h" 19#include "llvm/ADT/SmallSet.h" 20#include "llvm/Analysis/AliasAnalysis.h" 21#include "llvm/Analysis/ConstantFolding.h" 22#include "llvm/Constants.h" 23#include "llvm/CallingConv.h" 24#include "llvm/DerivedTypes.h" 25#include "llvm/Function.h" 26#include "llvm/GlobalVariable.h" 27#include "llvm/InlineAsm.h" 28#include "llvm/Instructions.h" 29#include "llvm/Intrinsics.h" 30#include "llvm/IntrinsicInst.h" 31#include "llvm/LLVMContext.h" 32#include "llvm/Module.h" 33#include "llvm/CodeGen/Analysis.h" 34#include "llvm/CodeGen/FastISel.h" 35#include "llvm/CodeGen/GCStrategy.h" 36#include "llvm/CodeGen/GCMetadata.h" 37#include "llvm/CodeGen/MachineFunction.h" 38#include "llvm/CodeGen/MachineFrameInfo.h" 39#include "llvm/CodeGen/MachineInstrBuilder.h" 40#include "llvm/CodeGen/MachineJumpTableInfo.h" 41#include "llvm/CodeGen/MachineModuleInfo.h" 42#include "llvm/CodeGen/MachineRegisterInfo.h" 43#include "llvm/CodeGen/PseudoSourceValue.h" 44#include "llvm/CodeGen/SelectionDAG.h" 45#include "llvm/Analysis/DebugInfo.h" 46#include "llvm/Target/TargetRegisterInfo.h" 47#include "llvm/Target/TargetData.h" 48#include "llvm/Target/TargetFrameInfo.h" 49#include "llvm/Target/TargetInstrInfo.h" 50#include "llvm/Target/TargetIntrinsicInfo.h" 51#include "llvm/Target/TargetLowering.h" 52#include "llvm/Target/TargetOptions.h" 53#include "llvm/Support/Compiler.h" 54#include "llvm/Support/CommandLine.h" 55#include "llvm/Support/Debug.h" 56#include "llvm/Support/ErrorHandling.h" 57#include "llvm/Support/MathExtras.h" 58#include "llvm/Support/raw_ostream.h" 59#include <algorithm> 60using namespace llvm; 61 62/// LimitFloatPrecision - Generate low-precision inline sequences for 63/// some float libcalls (6, 8 or 12 bits). 64static unsigned LimitFloatPrecision; 65 66static cl::opt<unsigned, true> 67LimitFPPrecision("limit-float-precision", 68 cl::desc("Generate low-precision inline sequences " 69 "for some float libcalls"), 70 cl::location(LimitFloatPrecision), 71 cl::init(0)); 72 73/// getCopyFromParts - Create a value that contains the specified legal parts 74/// combined into the value they represent. If the parts combine to a type 75/// larger then ValueVT then AssertOp can be used to specify whether the extra 76/// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT 77/// (ISD::AssertSext). 78static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, 79 const SDValue *Parts, 80 unsigned NumParts, EVT PartVT, EVT ValueVT, 81 ISD::NodeType AssertOp = ISD::DELETED_NODE) { 82 assert(NumParts > 0 && "No parts to assemble!"); 83 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 84 SDValue Val = Parts[0]; 85 86 if (NumParts > 1) { 87 // Assemble the value from multiple parts. 88 if (!ValueVT.isVector() && ValueVT.isInteger()) { 89 unsigned PartBits = PartVT.getSizeInBits(); 90 unsigned ValueBits = ValueVT.getSizeInBits(); 91 92 // Assemble the power of 2 part. 93 unsigned RoundParts = NumParts & (NumParts - 1) ? 94 1 << Log2_32(NumParts) : NumParts; 95 unsigned RoundBits = PartBits * RoundParts; 96 EVT RoundVT = RoundBits == ValueBits ? 97 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits); 98 SDValue Lo, Hi; 99 100 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2); 101 102 if (RoundParts > 2) { 103 Lo = getCopyFromParts(DAG, dl, Parts, RoundParts / 2, 104 PartVT, HalfVT); 105 Hi = getCopyFromParts(DAG, dl, Parts + RoundParts / 2, 106 RoundParts / 2, PartVT, HalfVT); 107 } else { 108 Lo = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[0]); 109 Hi = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[1]); 110 } 111 112 if (TLI.isBigEndian()) 113 std::swap(Lo, Hi); 114 115 Val = DAG.getNode(ISD::BUILD_PAIR, dl, RoundVT, Lo, Hi); 116 117 if (RoundParts < NumParts) { 118 // Assemble the trailing non-power-of-2 part. 119 unsigned OddParts = NumParts - RoundParts; 120 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits); 121 Hi = getCopyFromParts(DAG, dl, 122 Parts + RoundParts, OddParts, PartVT, OddVT); 123 124 // Combine the round and odd parts. 125 Lo = Val; 126 if (TLI.isBigEndian()) 127 std::swap(Lo, Hi); 128 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); 129 Hi = DAG.getNode(ISD::ANY_EXTEND, dl, TotalVT, Hi); 130 Hi = DAG.getNode(ISD::SHL, dl, TotalVT, Hi, 131 DAG.getConstant(Lo.getValueType().getSizeInBits(), 132 TLI.getPointerTy())); 133 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, TotalVT, Lo); 134 Val = DAG.getNode(ISD::OR, dl, TotalVT, Lo, Hi); 135 } 136 } else if (ValueVT.isVector()) { 137 // Handle a multi-element vector. 138 EVT IntermediateVT, RegisterVT; 139 unsigned NumIntermediates; 140 unsigned NumRegs = 141 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT, 142 NumIntermediates, RegisterVT); 143 assert(NumRegs == NumParts 144 && "Part count doesn't match vector breakdown!"); 145 NumParts = NumRegs; // Silence a compiler warning. 146 assert(RegisterVT == PartVT 147 && "Part type doesn't match vector breakdown!"); 148 assert(RegisterVT == Parts[0].getValueType() && 149 "Part type doesn't match part!"); 150 151 // Assemble the parts into intermediate operands. 152 SmallVector<SDValue, 8> Ops(NumIntermediates); 153 if (NumIntermediates == NumParts) { 154 // If the register was not expanded, truncate or copy the value, 155 // as appropriate. 156 for (unsigned i = 0; i != NumParts; ++i) 157 Ops[i] = getCopyFromParts(DAG, dl, &Parts[i], 1, 158 PartVT, IntermediateVT); 159 } else if (NumParts > 0) { 160 // If the intermediate type was expanded, build the intermediate 161 // operands from the parts. 162 assert(NumParts % NumIntermediates == 0 && 163 "Must expand into a divisible number of parts!"); 164 unsigned Factor = NumParts / NumIntermediates; 165 for (unsigned i = 0; i != NumIntermediates; ++i) 166 Ops[i] = getCopyFromParts(DAG, dl, &Parts[i * Factor], Factor, 167 PartVT, IntermediateVT); 168 } 169 170 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the 171 // intermediate operands. 172 Val = DAG.getNode(IntermediateVT.isVector() ? 173 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, dl, 174 ValueVT, &Ops[0], NumIntermediates); 175 } else if (PartVT.isFloatingPoint()) { 176 // FP split into multiple FP parts (for ppcf128) 177 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == EVT(MVT::f64) && 178 "Unexpected split"); 179 SDValue Lo, Hi; 180 Lo = DAG.getNode(ISD::BIT_CONVERT, dl, EVT(MVT::f64), Parts[0]); 181 Hi = DAG.getNode(ISD::BIT_CONVERT, dl, EVT(MVT::f64), Parts[1]); 182 if (TLI.isBigEndian()) 183 std::swap(Lo, Hi); 184 Val = DAG.getNode(ISD::BUILD_PAIR, dl, ValueVT, Lo, Hi); 185 } else { 186 // FP split into integer parts (soft fp) 187 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() && 188 !PartVT.isVector() && "Unexpected split"); 189 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits()); 190 Val = getCopyFromParts(DAG, dl, Parts, NumParts, PartVT, IntVT); 191 } 192 } 193 194 // There is now one part, held in Val. Correct it to match ValueVT. 195 PartVT = Val.getValueType(); 196 197 if (PartVT == ValueVT) 198 return Val; 199 200 if (PartVT.isVector()) { 201 assert(ValueVT.isVector() && "Unknown vector conversion!"); 202 return DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val); 203 } 204 205 if (ValueVT.isVector()) { 206 assert(ValueVT.getVectorElementType() == PartVT && 207 ValueVT.getVectorNumElements() == 1 && 208 "Only trivial scalar-to-vector conversions should get here!"); 209 return DAG.getNode(ISD::BUILD_VECTOR, dl, ValueVT, Val); 210 } 211 212 if (PartVT.isInteger() && 213 ValueVT.isInteger()) { 214 if (ValueVT.bitsLT(PartVT)) { 215 // For a truncate, see if we have any information to 216 // indicate whether the truncated bits will always be 217 // zero or sign-extension. 218 if (AssertOp != ISD::DELETED_NODE) 219 Val = DAG.getNode(AssertOp, dl, PartVT, Val, 220 DAG.getValueType(ValueVT)); 221 return DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val); 222 } else { 223 return DAG.getNode(ISD::ANY_EXTEND, dl, ValueVT, Val); 224 } 225 } 226 227 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { 228 if (ValueVT.bitsLT(Val.getValueType())) { 229 // FP_ROUND's are always exact here. 230 return DAG.getNode(ISD::FP_ROUND, dl, ValueVT, Val, 231 DAG.getIntPtrConstant(1)); 232 } 233 234 return DAG.getNode(ISD::FP_EXTEND, dl, ValueVT, Val); 235 } 236 237 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) 238 return DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val); 239 240 llvm_unreachable("Unknown mismatch!"); 241 return SDValue(); 242} 243 244/// getCopyToParts - Create a series of nodes that contain the specified value 245/// split into legal parts. If the parts contain more bits than Val, then, for 246/// integers, ExtendKind can be used to specify how to generate the extra bits. 247static void getCopyToParts(SelectionDAG &DAG, DebugLoc dl, 248 SDValue Val, SDValue *Parts, unsigned NumParts, 249 EVT PartVT, 250 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) { 251 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 252 EVT PtrVT = TLI.getPointerTy(); 253 EVT ValueVT = Val.getValueType(); 254 unsigned PartBits = PartVT.getSizeInBits(); 255 unsigned OrigNumParts = NumParts; 256 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!"); 257 258 if (!NumParts) 259 return; 260 261 if (!ValueVT.isVector()) { 262 if (PartVT == ValueVT) { 263 assert(NumParts == 1 && "No-op copy with multiple parts!"); 264 Parts[0] = Val; 265 return; 266 } 267 268 if (NumParts * PartBits > ValueVT.getSizeInBits()) { 269 // If the parts cover more bits than the value has, promote the value. 270 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { 271 assert(NumParts == 1 && "Do not know what to promote to!"); 272 Val = DAG.getNode(ISD::FP_EXTEND, dl, PartVT, Val); 273 } else if (PartVT.isInteger() && ValueVT.isInteger()) { 274 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); 275 Val = DAG.getNode(ExtendKind, dl, ValueVT, Val); 276 } else { 277 llvm_unreachable("Unknown mismatch!"); 278 } 279 } else if (PartBits == ValueVT.getSizeInBits()) { 280 // Different types of the same size. 281 assert(NumParts == 1 && PartVT != ValueVT); 282 Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val); 283 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) { 284 // If the parts cover less bits than value has, truncate the value. 285 if (PartVT.isInteger() && ValueVT.isInteger()) { 286 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); 287 Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val); 288 } else { 289 llvm_unreachable("Unknown mismatch!"); 290 } 291 } 292 293 // The value may have changed - recompute ValueVT. 294 ValueVT = Val.getValueType(); 295 assert(NumParts * PartBits == ValueVT.getSizeInBits() && 296 "Failed to tile the value with PartVT!"); 297 298 if (NumParts == 1) { 299 assert(PartVT == ValueVT && "Type conversion failed!"); 300 Parts[0] = Val; 301 return; 302 } 303 304 // Expand the value into multiple parts. 305 if (NumParts & (NumParts - 1)) { 306 // The number of parts is not a power of 2. Split off and copy the tail. 307 assert(PartVT.isInteger() && ValueVT.isInteger() && 308 "Do not know what to expand to!"); 309 unsigned RoundParts = 1 << Log2_32(NumParts); 310 unsigned RoundBits = RoundParts * PartBits; 311 unsigned OddParts = NumParts - RoundParts; 312 SDValue OddVal = DAG.getNode(ISD::SRL, dl, ValueVT, Val, 313 DAG.getConstant(RoundBits, 314 TLI.getPointerTy())); 315 getCopyToParts(DAG, dl, OddVal, Parts + RoundParts, 316 OddParts, PartVT); 317 318 if (TLI.isBigEndian()) 319 // The odd parts were reversed by getCopyToParts - unreverse them. 320 std::reverse(Parts + RoundParts, Parts + NumParts); 321 322 NumParts = RoundParts; 323 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); 324 Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val); 325 } 326 327 // The number of parts is a power of 2. Repeatedly bisect the value using 328 // EXTRACT_ELEMENT. 329 Parts[0] = DAG.getNode(ISD::BIT_CONVERT, dl, 330 EVT::getIntegerVT(*DAG.getContext(), 331 ValueVT.getSizeInBits()), 332 Val); 333 334 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) { 335 for (unsigned i = 0; i < NumParts; i += StepSize) { 336 unsigned ThisBits = StepSize * PartBits / 2; 337 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits); 338 SDValue &Part0 = Parts[i]; 339 SDValue &Part1 = Parts[i+StepSize/2]; 340 341 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, 342 ThisVT, Part0, 343 DAG.getConstant(1, PtrVT)); 344 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, 345 ThisVT, Part0, 346 DAG.getConstant(0, PtrVT)); 347 348 if (ThisBits == PartBits && ThisVT != PartVT) { 349 Part0 = DAG.getNode(ISD::BIT_CONVERT, dl, 350 PartVT, Part0); 351 Part1 = DAG.getNode(ISD::BIT_CONVERT, dl, 352 PartVT, Part1); 353 } 354 } 355 } 356 357 if (TLI.isBigEndian()) 358 std::reverse(Parts, Parts + OrigNumParts); 359 360 return; 361 } 362 363 // Vector ValueVT. 364 if (NumParts == 1) { 365 if (PartVT != ValueVT) { 366 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) { 367 Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val); 368 } else { 369 assert(ValueVT.getVectorElementType() == PartVT && 370 ValueVT.getVectorNumElements() == 1 && 371 "Only trivial vector-to-scalar conversions should get here!"); 372 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, 373 PartVT, Val, 374 DAG.getConstant(0, PtrVT)); 375 } 376 } 377 378 Parts[0] = Val; 379 return; 380 } 381 382 // Handle a multi-element vector. 383 EVT IntermediateVT, RegisterVT; 384 unsigned NumIntermediates; 385 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, 386 IntermediateVT, NumIntermediates, RegisterVT); 387 unsigned NumElements = ValueVT.getVectorNumElements(); 388 389 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!"); 390 NumParts = NumRegs; // Silence a compiler warning. 391 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!"); 392 393 // Split the vector into intermediate operands. 394 SmallVector<SDValue, 8> Ops(NumIntermediates); 395 for (unsigned i = 0; i != NumIntermediates; ++i) { 396 if (IntermediateVT.isVector()) 397 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, 398 IntermediateVT, Val, 399 DAG.getConstant(i * (NumElements / NumIntermediates), 400 PtrVT)); 401 else 402 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, 403 IntermediateVT, Val, 404 DAG.getConstant(i, PtrVT)); 405 } 406 407 // Split the intermediate operands into legal parts. 408 if (NumParts == NumIntermediates) { 409 // If the register was not expanded, promote or copy the value, 410 // as appropriate. 411 for (unsigned i = 0; i != NumParts; ++i) 412 getCopyToParts(DAG, dl, Ops[i], &Parts[i], 1, PartVT); 413 } else if (NumParts > 0) { 414 // If the intermediate type was expanded, split each the value into 415 // legal parts. 416 assert(NumParts % NumIntermediates == 0 && 417 "Must expand into a divisible number of parts!"); 418 unsigned Factor = NumParts / NumIntermediates; 419 for (unsigned i = 0; i != NumIntermediates; ++i) 420 getCopyToParts(DAG, dl, Ops[i], &Parts[i*Factor], Factor, PartVT); 421 } 422} 423 424namespace { 425 /// RegsForValue - This struct represents the registers (physical or virtual) 426 /// that a particular set of values is assigned, and the type information 427 /// about the value. The most common situation is to represent one value at a 428 /// time, but struct or array values are handled element-wise as multiple 429 /// values. The splitting of aggregates is performed recursively, so that we 430 /// never have aggregate-typed registers. The values at this point do not 431 /// necessarily have legal types, so each value may require one or more 432 /// registers of some legal type. 433 /// 434 struct RegsForValue { 435 /// ValueVTs - The value types of the values, which may not be legal, and 436 /// may need be promoted or synthesized from one or more registers. 437 /// 438 SmallVector<EVT, 4> ValueVTs; 439 440 /// RegVTs - The value types of the registers. This is the same size as 441 /// ValueVTs and it records, for each value, what the type of the assigned 442 /// register or registers are. (Individual values are never synthesized 443 /// from more than one type of register.) 444 /// 445 /// With virtual registers, the contents of RegVTs is redundant with TLI's 446 /// getRegisterType member function, however when with physical registers 447 /// it is necessary to have a separate record of the types. 448 /// 449 SmallVector<EVT, 4> RegVTs; 450 451 /// Regs - This list holds the registers assigned to the values. 452 /// Each legal or promoted value requires one register, and each 453 /// expanded value requires multiple registers. 454 /// 455 SmallVector<unsigned, 4> Regs; 456 457 RegsForValue() {} 458 459 RegsForValue(const SmallVector<unsigned, 4> ®s, 460 EVT regvt, EVT valuevt) 461 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {} 462 463 RegsForValue(const SmallVector<unsigned, 4> ®s, 464 const SmallVector<EVT, 4> ®vts, 465 const SmallVector<EVT, 4> &valuevts) 466 : ValueVTs(valuevts), RegVTs(regvts), Regs(regs) {} 467 468 RegsForValue(LLVMContext &Context, const TargetLowering &tli, 469 unsigned Reg, const Type *Ty) { 470 ComputeValueVTs(tli, Ty, ValueVTs); 471 472 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) { 473 EVT ValueVT = ValueVTs[Value]; 474 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT); 475 EVT RegisterVT = tli.getRegisterType(Context, ValueVT); 476 for (unsigned i = 0; i != NumRegs; ++i) 477 Regs.push_back(Reg + i); 478 RegVTs.push_back(RegisterVT); 479 Reg += NumRegs; 480 } 481 } 482 483 /// areValueTypesLegal - Return true if types of all the values are legal. 484 bool areValueTypesLegal(const TargetLowering &TLI) { 485 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) { 486 EVT RegisterVT = RegVTs[Value]; 487 if (!TLI.isTypeLegal(RegisterVT)) 488 return false; 489 } 490 return true; 491 } 492 493 /// append - Add the specified values to this one. 494 void append(const RegsForValue &RHS) { 495 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end()); 496 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end()); 497 Regs.append(RHS.Regs.begin(), RHS.Regs.end()); 498 } 499 500 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from 501 /// this value and returns the result as a ValueVTs value. This uses 502 /// Chain/Flag as the input and updates them for the output Chain/Flag. 503 /// If the Flag pointer is NULL, no flag is used. 504 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo, 505 DebugLoc dl, 506 SDValue &Chain, SDValue *Flag) const; 507 508 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the 509 /// specified value into the registers specified by this object. This uses 510 /// Chain/Flag as the input and updates them for the output Chain/Flag. 511 /// If the Flag pointer is NULL, no flag is used. 512 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl, 513 SDValue &Chain, SDValue *Flag) const; 514 515 /// AddInlineAsmOperands - Add this value to the specified inlineasm node 516 /// operand list. This adds the code marker, matching input operand index 517 /// (if applicable), and includes the number of values added into it. 518 void AddInlineAsmOperands(unsigned Kind, 519 bool HasMatching, unsigned MatchingIdx, 520 SelectionDAG &DAG, 521 std::vector<SDValue> &Ops) const; 522 }; 523} 524 525/// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from 526/// this value and returns the result as a ValueVT value. This uses 527/// Chain/Flag as the input and updates them for the output Chain/Flag. 528/// If the Flag pointer is NULL, no flag is used. 529SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG, 530 FunctionLoweringInfo &FuncInfo, 531 DebugLoc dl, 532 SDValue &Chain, SDValue *Flag) const { 533 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 534 535 // Assemble the legal parts into the final values. 536 SmallVector<SDValue, 4> Values(ValueVTs.size()); 537 SmallVector<SDValue, 8> Parts; 538 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) { 539 // Copy the legal parts from the registers. 540 EVT ValueVT = ValueVTs[Value]; 541 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT); 542 EVT RegisterVT = RegVTs[Value]; 543 544 Parts.resize(NumRegs); 545 for (unsigned i = 0; i != NumRegs; ++i) { 546 SDValue P; 547 if (Flag == 0) { 548 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT); 549 } else { 550 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag); 551 *Flag = P.getValue(2); 552 } 553 554 Chain = P.getValue(1); 555 556 // If the source register was virtual and if we know something about it, 557 // add an assert node. 558 if (TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) && 559 RegisterVT.isInteger() && !RegisterVT.isVector()) { 560 unsigned SlotNo = Regs[Part+i]-TargetRegisterInfo::FirstVirtualRegister; 561 if (FuncInfo.LiveOutRegInfo.size() > SlotNo) { 562 const FunctionLoweringInfo::LiveOutInfo &LOI = 563 FuncInfo.LiveOutRegInfo[SlotNo]; 564 565 unsigned RegSize = RegisterVT.getSizeInBits(); 566 unsigned NumSignBits = LOI.NumSignBits; 567 unsigned NumZeroBits = LOI.KnownZero.countLeadingOnes(); 568 569 // FIXME: We capture more information than the dag can represent. For 570 // now, just use the tightest assertzext/assertsext possible. 571 bool isSExt = true; 572 EVT FromVT(MVT::Other); 573 if (NumSignBits == RegSize) 574 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1 575 else if (NumZeroBits >= RegSize-1) 576 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1 577 else if (NumSignBits > RegSize-8) 578 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8 579 else if (NumZeroBits >= RegSize-8) 580 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8 581 else if (NumSignBits > RegSize-16) 582 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16 583 else if (NumZeroBits >= RegSize-16) 584 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16 585 else if (NumSignBits > RegSize-32) 586 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32 587 else if (NumZeroBits >= RegSize-32) 588 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32 589 590 if (FromVT != MVT::Other) 591 P = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl, 592 RegisterVT, P, DAG.getValueType(FromVT)); 593 } 594 } 595 596 Parts[i] = P; 597 } 598 599 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(), 600 NumRegs, RegisterVT, ValueVT); 601 Part += NumRegs; 602 Parts.clear(); 603 } 604 605 return DAG.getNode(ISD::MERGE_VALUES, dl, 606 DAG.getVTList(&ValueVTs[0], ValueVTs.size()), 607 &Values[0], ValueVTs.size()); 608} 609 610/// getCopyToRegs - Emit a series of CopyToReg nodes that copies the 611/// specified value into the registers specified by this object. This uses 612/// Chain/Flag as the input and updates them for the output Chain/Flag. 613/// If the Flag pointer is NULL, no flag is used. 614void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl, 615 SDValue &Chain, SDValue *Flag) const { 616 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 617 618 // Get the list of the values's legal parts. 619 unsigned NumRegs = Regs.size(); 620 SmallVector<SDValue, 8> Parts(NumRegs); 621 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) { 622 EVT ValueVT = ValueVTs[Value]; 623 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT); 624 EVT RegisterVT = RegVTs[Value]; 625 626 getCopyToParts(DAG, dl, 627 Val.getValue(Val.getResNo() + Value), 628 &Parts[Part], NumParts, RegisterVT); 629 Part += NumParts; 630 } 631 632 // Copy the parts into the registers. 633 SmallVector<SDValue, 8> Chains(NumRegs); 634 for (unsigned i = 0; i != NumRegs; ++i) { 635 SDValue Part; 636 if (Flag == 0) { 637 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]); 638 } else { 639 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag); 640 *Flag = Part.getValue(1); 641 } 642 643 Chains[i] = Part.getValue(0); 644 } 645 646 if (NumRegs == 1 || Flag) 647 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is 648 // flagged to it. That is the CopyToReg nodes and the user are considered 649 // a single scheduling unit. If we create a TokenFactor and return it as 650 // chain, then the TokenFactor is both a predecessor (operand) of the 651 // user as well as a successor (the TF operands are flagged to the user). 652 // c1, f1 = CopyToReg 653 // c2, f2 = CopyToReg 654 // c3 = TokenFactor c1, c2 655 // ... 656 // = op c3, ..., f2 657 Chain = Chains[NumRegs-1]; 658 else 659 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs); 660} 661 662/// AddInlineAsmOperands - Add this value to the specified inlineasm node 663/// operand list. This adds the code marker and includes the number of 664/// values added into it. 665void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching, 666 unsigned MatchingIdx, 667 SelectionDAG &DAG, 668 std::vector<SDValue> &Ops) const { 669 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 670 671 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size()); 672 if (HasMatching) 673 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx); 674 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32); 675 Ops.push_back(Res); 676 677 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) { 678 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]); 679 EVT RegisterVT = RegVTs[Value]; 680 for (unsigned i = 0; i != NumRegs; ++i) { 681 assert(Reg < Regs.size() && "Mismatch in # registers expected"); 682 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT)); 683 } 684 } 685} 686 687void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa) { 688 AA = &aa; 689 GFI = gfi; 690 TD = DAG.getTarget().getTargetData(); 691} 692 693/// clear - Clear out the current SelectionDAG and the associated 694/// state and prepare this SelectionDAGBuilder object to be used 695/// for a new block. This doesn't clear out information about 696/// additional blocks that are needed to complete switch lowering 697/// or PHI node updating; that information is cleared out as it is 698/// consumed. 699void SelectionDAGBuilder::clear() { 700 NodeMap.clear(); 701 UnusedArgNodeMap.clear(); 702 PendingLoads.clear(); 703 PendingExports.clear(); 704 CurDebugLoc = DebugLoc(); 705 HasTailCall = false; 706} 707 708/// getRoot - Return the current virtual root of the Selection DAG, 709/// flushing any PendingLoad items. This must be done before emitting 710/// a store or any other node that may need to be ordered after any 711/// prior load instructions. 712/// 713SDValue SelectionDAGBuilder::getRoot() { 714 if (PendingLoads.empty()) 715 return DAG.getRoot(); 716 717 if (PendingLoads.size() == 1) { 718 SDValue Root = PendingLoads[0]; 719 DAG.setRoot(Root); 720 PendingLoads.clear(); 721 return Root; 722 } 723 724 // Otherwise, we have to make a token factor node. 725 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other, 726 &PendingLoads[0], PendingLoads.size()); 727 PendingLoads.clear(); 728 DAG.setRoot(Root); 729 return Root; 730} 731 732/// getControlRoot - Similar to getRoot, but instead of flushing all the 733/// PendingLoad items, flush all the PendingExports items. It is necessary 734/// to do this before emitting a terminator instruction. 735/// 736SDValue SelectionDAGBuilder::getControlRoot() { 737 SDValue Root = DAG.getRoot(); 738 739 if (PendingExports.empty()) 740 return Root; 741 742 // Turn all of the CopyToReg chains into one factored node. 743 if (Root.getOpcode() != ISD::EntryToken) { 744 unsigned i = 0, e = PendingExports.size(); 745 for (; i != e; ++i) { 746 assert(PendingExports[i].getNode()->getNumOperands() > 1); 747 if (PendingExports[i].getNode()->getOperand(0) == Root) 748 break; // Don't add the root if we already indirectly depend on it. 749 } 750 751 if (i == e) 752 PendingExports.push_back(Root); 753 } 754 755 Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other, 756 &PendingExports[0], 757 PendingExports.size()); 758 PendingExports.clear(); 759 DAG.setRoot(Root); 760 return Root; 761} 762 763void SelectionDAGBuilder::AssignOrderingToNode(const SDNode *Node) { 764 if (DAG.GetOrdering(Node) != 0) return; // Already has ordering. 765 DAG.AssignOrdering(Node, SDNodeOrder); 766 767 for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I) 768 AssignOrderingToNode(Node->getOperand(I).getNode()); 769} 770 771void SelectionDAGBuilder::visit(const Instruction &I) { 772 // Set up outgoing PHI node register values before emitting the terminator. 773 if (isa<TerminatorInst>(&I)) 774 HandlePHINodesInSuccessorBlocks(I.getParent()); 775 776 CurDebugLoc = I.getDebugLoc(); 777 778 visit(I.getOpcode(), I); 779 780 if (!isa<TerminatorInst>(&I) && !HasTailCall) 781 CopyToExportRegsIfNeeded(&I); 782 783 CurDebugLoc = DebugLoc(); 784} 785 786void SelectionDAGBuilder::visitPHI(const PHINode &) { 787 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!"); 788} 789 790void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) { 791 // Note: this doesn't use InstVisitor, because it has to work with 792 // ConstantExpr's in addition to instructions. 793 switch (Opcode) { 794 default: llvm_unreachable("Unknown instruction type encountered!"); 795 // Build the switch statement using the Instruction.def file. 796#define HANDLE_INST(NUM, OPCODE, CLASS) \ 797 case Instruction::OPCODE: visit##OPCODE((CLASS&)I); break; 798#include "llvm/Instruction.def" 799 } 800 801 // Assign the ordering to the freshly created DAG nodes. 802 if (NodeMap.count(&I)) { 803 ++SDNodeOrder; 804 AssignOrderingToNode(getValue(&I).getNode()); 805 } 806} 807 808SDValue SelectionDAGBuilder::getValue(const Value *V) { 809 SDValue &N = NodeMap[V]; 810 if (N.getNode()) return N; 811 812 if (const Constant *C = dyn_cast<Constant>(V)) { 813 EVT VT = TLI.getValueType(V->getType(), true); 814 815 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C)) 816 return N = DAG.getConstant(*CI, VT); 817 818 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C)) 819 return N = DAG.getGlobalAddress(GV, VT); 820 821 if (isa<ConstantPointerNull>(C)) 822 return N = DAG.getConstant(0, TLI.getPointerTy()); 823 824 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C)) 825 return N = DAG.getConstantFP(*CFP, VT); 826 827 if (isa<UndefValue>(C) && !V->getType()->isAggregateType()) 828 return N = DAG.getUNDEF(VT); 829 830 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 831 visit(CE->getOpcode(), *CE); 832 SDValue N1 = NodeMap[V]; 833 assert(N1.getNode() && "visit didn't populate the NodeMap!"); 834 return N1; 835 } 836 837 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) { 838 SmallVector<SDValue, 4> Constants; 839 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end(); 840 OI != OE; ++OI) { 841 SDNode *Val = getValue(*OI).getNode(); 842 // If the operand is an empty aggregate, there are no values. 843 if (!Val) continue; 844 // Add each leaf value from the operand to the Constants list 845 // to form a flattened list of all the values. 846 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i) 847 Constants.push_back(SDValue(Val, i)); 848 } 849 850 return DAG.getMergeValues(&Constants[0], Constants.size(), 851 getCurDebugLoc()); 852 } 853 854 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) { 855 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) && 856 "Unknown struct or array constant!"); 857 858 SmallVector<EVT, 4> ValueVTs; 859 ComputeValueVTs(TLI, C->getType(), ValueVTs); 860 unsigned NumElts = ValueVTs.size(); 861 if (NumElts == 0) 862 return SDValue(); // empty struct 863 SmallVector<SDValue, 4> Constants(NumElts); 864 for (unsigned i = 0; i != NumElts; ++i) { 865 EVT EltVT = ValueVTs[i]; 866 if (isa<UndefValue>(C)) 867 Constants[i] = DAG.getUNDEF(EltVT); 868 else if (EltVT.isFloatingPoint()) 869 Constants[i] = DAG.getConstantFP(0, EltVT); 870 else 871 Constants[i] = DAG.getConstant(0, EltVT); 872 } 873 874 return DAG.getMergeValues(&Constants[0], NumElts, 875 getCurDebugLoc()); 876 } 877 878 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C)) 879 return DAG.getBlockAddress(BA, VT); 880 881 const VectorType *VecTy = cast<VectorType>(V->getType()); 882 unsigned NumElements = VecTy->getNumElements(); 883 884 // Now that we know the number and type of the elements, get that number of 885 // elements into the Ops array based on what kind of constant it is. 886 SmallVector<SDValue, 16> Ops; 887 if (const ConstantVector *CP = dyn_cast<ConstantVector>(C)) { 888 for (unsigned i = 0; i != NumElements; ++i) 889 Ops.push_back(getValue(CP->getOperand(i))); 890 } else { 891 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!"); 892 EVT EltVT = TLI.getValueType(VecTy->getElementType()); 893 894 SDValue Op; 895 if (EltVT.isFloatingPoint()) 896 Op = DAG.getConstantFP(0, EltVT); 897 else 898 Op = DAG.getConstant(0, EltVT); 899 Ops.assign(NumElements, Op); 900 } 901 902 // Create a BUILD_VECTOR node. 903 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(), 904 VT, &Ops[0], Ops.size()); 905 } 906 907 // If this is a static alloca, generate it as the frameindex instead of 908 // computation. 909 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) { 910 DenseMap<const AllocaInst*, int>::iterator SI = 911 FuncInfo.StaticAllocaMap.find(AI); 912 if (SI != FuncInfo.StaticAllocaMap.end()) 913 return DAG.getFrameIndex(SI->second, TLI.getPointerTy()); 914 } 915 916 unsigned InReg = FuncInfo.ValueMap[V]; 917 assert(InReg && "Value not in map!"); 918 919 RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType()); 920 SDValue Chain = DAG.getEntryNode(); 921 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL); 922} 923 924/// Get the EVTs and ArgFlags collections that represent the legalized return 925/// type of the given function. This does not require a DAG or a return value, 926/// and is suitable for use before any DAGs for the function are constructed. 927static void getReturnInfo(const Type* ReturnType, 928 Attributes attr, SmallVectorImpl<EVT> &OutVTs, 929 SmallVectorImpl<ISD::ArgFlagsTy> &OutFlags, 930 const TargetLowering &TLI, 931 SmallVectorImpl<uint64_t> *Offsets = 0) { 932 SmallVector<EVT, 4> ValueVTs; 933 ComputeValueVTs(TLI, ReturnType, ValueVTs); 934 unsigned NumValues = ValueVTs.size(); 935 if (NumValues == 0) return; 936 unsigned Offset = 0; 937 938 for (unsigned j = 0, f = NumValues; j != f; ++j) { 939 EVT VT = ValueVTs[j]; 940 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 941 942 if (attr & Attribute::SExt) 943 ExtendKind = ISD::SIGN_EXTEND; 944 else if (attr & Attribute::ZExt) 945 ExtendKind = ISD::ZERO_EXTEND; 946 947 // FIXME: C calling convention requires the return type to be promoted to 948 // at least 32-bit. But this is not necessary for non-C calling 949 // conventions. The frontend should mark functions whose return values 950 // require promoting with signext or zeroext attributes. 951 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 952 EVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 953 if (VT.bitsLT(MinVT)) 954 VT = MinVT; 955 } 956 957 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT); 958 EVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT); 959 unsigned PartSize = TLI.getTargetData()->getTypeAllocSize( 960 PartVT.getTypeForEVT(ReturnType->getContext())); 961 962 // 'inreg' on function refers to return value 963 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 964 if (attr & Attribute::InReg) 965 Flags.setInReg(); 966 967 // Propagate extension type if any 968 if (attr & Attribute::SExt) 969 Flags.setSExt(); 970 else if (attr & Attribute::ZExt) 971 Flags.setZExt(); 972 973 for (unsigned i = 0; i < NumParts; ++i) { 974 OutVTs.push_back(PartVT); 975 OutFlags.push_back(Flags); 976 if (Offsets) 977 { 978 Offsets->push_back(Offset); 979 Offset += PartSize; 980 } 981 } 982 } 983} 984 985void SelectionDAGBuilder::visitRet(const ReturnInst &I) { 986 SDValue Chain = getControlRoot(); 987 SmallVector<ISD::OutputArg, 8> Outs; 988 989 if (!FuncInfo.CanLowerReturn) { 990 unsigned DemoteReg = FuncInfo.DemoteRegister; 991 const Function *F = I.getParent()->getParent(); 992 993 // Emit a store of the return value through the virtual register. 994 // Leave Outs empty so that LowerReturn won't try to load return 995 // registers the usual way. 996 SmallVector<EVT, 1> PtrValueVTs; 997 ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()), 998 PtrValueVTs); 999 1000 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]); 1001 SDValue RetOp = getValue(I.getOperand(0)); 1002 1003 SmallVector<EVT, 4> ValueVTs; 1004 SmallVector<uint64_t, 4> Offsets; 1005 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets); 1006 unsigned NumValues = ValueVTs.size(); 1007 1008 SmallVector<SDValue, 4> Chains(NumValues); 1009 EVT PtrVT = PtrValueVTs[0]; 1010 for (unsigned i = 0; i != NumValues; ++i) { 1011 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, RetPtr, 1012 DAG.getConstant(Offsets[i], PtrVT)); 1013 Chains[i] = 1014 DAG.getStore(Chain, getCurDebugLoc(), 1015 SDValue(RetOp.getNode(), RetOp.getResNo() + i), 1016 Add, NULL, Offsets[i], false, false, 0); 1017 } 1018 1019 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), 1020 MVT::Other, &Chains[0], NumValues); 1021 } else if (I.getNumOperands() != 0) { 1022 SmallVector<EVT, 4> ValueVTs; 1023 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs); 1024 unsigned NumValues = ValueVTs.size(); 1025 if (NumValues) { 1026 SDValue RetOp = getValue(I.getOperand(0)); 1027 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1028 EVT VT = ValueVTs[j]; 1029 1030 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1031 1032 const Function *F = I.getParent()->getParent(); 1033 if (F->paramHasAttr(0, Attribute::SExt)) 1034 ExtendKind = ISD::SIGN_EXTEND; 1035 else if (F->paramHasAttr(0, Attribute::ZExt)) 1036 ExtendKind = ISD::ZERO_EXTEND; 1037 1038 // FIXME: C calling convention requires the return type to be promoted 1039 // to at least 32-bit. But this is not necessary for non-C calling 1040 // conventions. The frontend should mark functions whose return values 1041 // require promoting with signext or zeroext attributes. 1042 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1043 EVT MinVT = TLI.getRegisterType(*DAG.getContext(), MVT::i32); 1044 if (VT.bitsLT(MinVT)) 1045 VT = MinVT; 1046 } 1047 1048 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT); 1049 EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT); 1050 SmallVector<SDValue, 4> Parts(NumParts); 1051 getCopyToParts(DAG, getCurDebugLoc(), 1052 SDValue(RetOp.getNode(), RetOp.getResNo() + j), 1053 &Parts[0], NumParts, PartVT, ExtendKind); 1054 1055 // 'inreg' on function refers to return value 1056 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1057 if (F->paramHasAttr(0, Attribute::InReg)) 1058 Flags.setInReg(); 1059 1060 // Propagate extension type if any 1061 if (F->paramHasAttr(0, Attribute::SExt)) 1062 Flags.setSExt(); 1063 else if (F->paramHasAttr(0, Attribute::ZExt)) 1064 Flags.setZExt(); 1065 1066 for (unsigned i = 0; i < NumParts; ++i) 1067 Outs.push_back(ISD::OutputArg(Flags, Parts[i], /*isfixed=*/true)); 1068 } 1069 } 1070 } 1071 1072 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg(); 1073 CallingConv::ID CallConv = 1074 DAG.getMachineFunction().getFunction()->getCallingConv(); 1075 Chain = TLI.LowerReturn(Chain, CallConv, isVarArg, 1076 Outs, getCurDebugLoc(), DAG); 1077 1078 // Verify that the target's LowerReturn behaved as expected. 1079 assert(Chain.getNode() && Chain.getValueType() == MVT::Other && 1080 "LowerReturn didn't return a valid chain!"); 1081 1082 // Update the DAG with the new chain value resulting from return lowering. 1083 DAG.setRoot(Chain); 1084} 1085 1086/// CopyToExportRegsIfNeeded - If the given value has virtual registers 1087/// created for it, emit nodes to copy the value into the virtual 1088/// registers. 1089void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) { 1090 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V); 1091 if (VMI != FuncInfo.ValueMap.end()) { 1092 assert(!V->use_empty() && "Unused value assigned virtual registers!"); 1093 CopyValueToVirtualRegister(V, VMI->second); 1094 } 1095} 1096 1097/// ExportFromCurrentBlock - If this condition isn't known to be exported from 1098/// the current basic block, add it to ValueMap now so that we'll get a 1099/// CopyTo/FromReg. 1100void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) { 1101 // No need to export constants. 1102 if (!isa<Instruction>(V) && !isa<Argument>(V)) return; 1103 1104 // Already exported? 1105 if (FuncInfo.isExportedInst(V)) return; 1106 1107 unsigned Reg = FuncInfo.InitializeRegForValue(V); 1108 CopyValueToVirtualRegister(V, Reg); 1109} 1110 1111bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V, 1112 const BasicBlock *FromBB) { 1113 // The operands of the setcc have to be in this block. We don't know 1114 // how to export them from some other block. 1115 if (const Instruction *VI = dyn_cast<Instruction>(V)) { 1116 // Can export from current BB. 1117 if (VI->getParent() == FromBB) 1118 return true; 1119 1120 // Is already exported, noop. 1121 return FuncInfo.isExportedInst(V); 1122 } 1123 1124 // If this is an argument, we can export it if the BB is the entry block or 1125 // if it is already exported. 1126 if (isa<Argument>(V)) { 1127 if (FromBB == &FromBB->getParent()->getEntryBlock()) 1128 return true; 1129 1130 // Otherwise, can only export this if it is already exported. 1131 return FuncInfo.isExportedInst(V); 1132 } 1133 1134 // Otherwise, constants can always be exported. 1135 return true; 1136} 1137 1138static bool InBlock(const Value *V, const BasicBlock *BB) { 1139 if (const Instruction *I = dyn_cast<Instruction>(V)) 1140 return I->getParent() == BB; 1141 return true; 1142} 1143 1144/// EmitBranchForMergedCondition - Helper method for FindMergedConditions. 1145/// This function emits a branch and is used at the leaves of an OR or an 1146/// AND operator tree. 1147/// 1148void 1149SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond, 1150 MachineBasicBlock *TBB, 1151 MachineBasicBlock *FBB, 1152 MachineBasicBlock *CurBB, 1153 MachineBasicBlock *SwitchBB) { 1154 const BasicBlock *BB = CurBB->getBasicBlock(); 1155 1156 // If the leaf of the tree is a comparison, merge the condition into 1157 // the caseblock. 1158 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) { 1159 // The operands of the cmp have to be in this block. We don't know 1160 // how to export them from some other block. If this is the first block 1161 // of the sequence, no exporting is needed. 1162 if (CurBB == SwitchBB || 1163 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) && 1164 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) { 1165 ISD::CondCode Condition; 1166 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) { 1167 Condition = getICmpCondCode(IC->getPredicate()); 1168 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) { 1169 Condition = getFCmpCondCode(FC->getPredicate()); 1170 } else { 1171 Condition = ISD::SETEQ; // silence warning. 1172 llvm_unreachable("Unknown compare instruction"); 1173 } 1174 1175 CaseBlock CB(Condition, BOp->getOperand(0), 1176 BOp->getOperand(1), NULL, TBB, FBB, CurBB); 1177 SwitchCases.push_back(CB); 1178 return; 1179 } 1180 } 1181 1182 // Create a CaseBlock record representing this branch. 1183 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()), 1184 NULL, TBB, FBB, CurBB); 1185 SwitchCases.push_back(CB); 1186} 1187 1188/// FindMergedConditions - If Cond is an expression like 1189void SelectionDAGBuilder::FindMergedConditions(const Value *Cond, 1190 MachineBasicBlock *TBB, 1191 MachineBasicBlock *FBB, 1192 MachineBasicBlock *CurBB, 1193 MachineBasicBlock *SwitchBB, 1194 unsigned Opc) { 1195 // If this node is not part of the or/and tree, emit it as a branch. 1196 const Instruction *BOp = dyn_cast<Instruction>(Cond); 1197 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) || 1198 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() || 1199 BOp->getParent() != CurBB->getBasicBlock() || 1200 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) || 1201 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) { 1202 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB); 1203 return; 1204 } 1205 1206 // Create TmpBB after CurBB. 1207 MachineFunction::iterator BBI = CurBB; 1208 MachineFunction &MF = DAG.getMachineFunction(); 1209 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock()); 1210 CurBB->getParent()->insert(++BBI, TmpBB); 1211 1212 if (Opc == Instruction::Or) { 1213 // Codegen X | Y as: 1214 // jmp_if_X TBB 1215 // jmp TmpBB 1216 // TmpBB: 1217 // jmp_if_Y TBB 1218 // jmp FBB 1219 // 1220 1221 // Emit the LHS condition. 1222 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc); 1223 1224 // Emit the RHS condition into TmpBB. 1225 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc); 1226 } else { 1227 assert(Opc == Instruction::And && "Unknown merge op!"); 1228 // Codegen X & Y as: 1229 // jmp_if_X TmpBB 1230 // jmp FBB 1231 // TmpBB: 1232 // jmp_if_Y TBB 1233 // jmp FBB 1234 // 1235 // This requires creation of TmpBB after CurBB. 1236 1237 // Emit the LHS condition. 1238 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc); 1239 1240 // Emit the RHS condition into TmpBB. 1241 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc); 1242 } 1243} 1244 1245/// If the set of cases should be emitted as a series of branches, return true. 1246/// If we should emit this as a bunch of and/or'd together conditions, return 1247/// false. 1248bool 1249SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){ 1250 if (Cases.size() != 2) return true; 1251 1252 // If this is two comparisons of the same values or'd or and'd together, they 1253 // will get folded into a single comparison, so don't emit two blocks. 1254 if ((Cases[0].CmpLHS == Cases[1].CmpLHS && 1255 Cases[0].CmpRHS == Cases[1].CmpRHS) || 1256 (Cases[0].CmpRHS == Cases[1].CmpLHS && 1257 Cases[0].CmpLHS == Cases[1].CmpRHS)) { 1258 return false; 1259 } 1260 1261 // Handle: (X != null) | (Y != null) --> (X|Y) != 0 1262 // Handle: (X == null) & (Y == null) --> (X|Y) == 0 1263 if (Cases[0].CmpRHS == Cases[1].CmpRHS && 1264 Cases[0].CC == Cases[1].CC && 1265 isa<Constant>(Cases[0].CmpRHS) && 1266 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) { 1267 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB) 1268 return false; 1269 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB) 1270 return false; 1271 } 1272 1273 return true; 1274} 1275 1276void SelectionDAGBuilder::visitBr(const BranchInst &I) { 1277 MachineBasicBlock *BrMBB = FuncInfo.MBBMap[I.getParent()]; 1278 1279 // Update machine-CFG edges. 1280 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)]; 1281 1282 // Figure out which block is immediately after the current one. 1283 MachineBasicBlock *NextBlock = 0; 1284 MachineFunction::iterator BBI = BrMBB; 1285 if (++BBI != FuncInfo.MF->end()) 1286 NextBlock = BBI; 1287 1288 if (I.isUnconditional()) { 1289 // Update machine-CFG edges. 1290 BrMBB->addSuccessor(Succ0MBB); 1291 1292 // If this is not a fall-through branch, emit the branch. 1293 if (Succ0MBB != NextBlock) 1294 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(), 1295 MVT::Other, getControlRoot(), 1296 DAG.getBasicBlock(Succ0MBB))); 1297 1298 return; 1299 } 1300 1301 // If this condition is one of the special cases we handle, do special stuff 1302 // now. 1303 const Value *CondVal = I.getCondition(); 1304 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)]; 1305 1306 // If this is a series of conditions that are or'd or and'd together, emit 1307 // this as a sequence of branches instead of setcc's with and/or operations. 1308 // For example, instead of something like: 1309 // cmp A, B 1310 // C = seteq 1311 // cmp D, E 1312 // F = setle 1313 // or C, F 1314 // jnz foo 1315 // Emit: 1316 // cmp A, B 1317 // je foo 1318 // cmp D, E 1319 // jle foo 1320 // 1321 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) { 1322 if (BOp->hasOneUse() && 1323 (BOp->getOpcode() == Instruction::And || 1324 BOp->getOpcode() == Instruction::Or)) { 1325 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB, 1326 BOp->getOpcode()); 1327 // If the compares in later blocks need to use values not currently 1328 // exported from this block, export them now. This block should always 1329 // be the first entry. 1330 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!"); 1331 1332 // Allow some cases to be rejected. 1333 if (ShouldEmitAsBranches(SwitchCases)) { 1334 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) { 1335 ExportFromCurrentBlock(SwitchCases[i].CmpLHS); 1336 ExportFromCurrentBlock(SwitchCases[i].CmpRHS); 1337 } 1338 1339 // Emit the branch for this block. 1340 visitSwitchCase(SwitchCases[0], BrMBB); 1341 SwitchCases.erase(SwitchCases.begin()); 1342 return; 1343 } 1344 1345 // Okay, we decided not to do this, remove any inserted MBB's and clear 1346 // SwitchCases. 1347 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) 1348 FuncInfo.MF->erase(SwitchCases[i].ThisBB); 1349 1350 SwitchCases.clear(); 1351 } 1352 } 1353 1354 // Create a CaseBlock record representing this branch. 1355 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()), 1356 NULL, Succ0MBB, Succ1MBB, BrMBB); 1357 1358 // Use visitSwitchCase to actually insert the fast branch sequence for this 1359 // cond branch. 1360 visitSwitchCase(CB, BrMBB); 1361} 1362 1363/// visitSwitchCase - Emits the necessary code to represent a single node in 1364/// the binary search tree resulting from lowering a switch instruction. 1365void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB, 1366 MachineBasicBlock *SwitchBB) { 1367 SDValue Cond; 1368 SDValue CondLHS = getValue(CB.CmpLHS); 1369 DebugLoc dl = getCurDebugLoc(); 1370 1371 // Build the setcc now. 1372 if (CB.CmpMHS == NULL) { 1373 // Fold "(X == true)" to X and "(X == false)" to !X to 1374 // handle common cases produced by branch lowering. 1375 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) && 1376 CB.CC == ISD::SETEQ) 1377 Cond = CondLHS; 1378 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) && 1379 CB.CC == ISD::SETEQ) { 1380 SDValue True = DAG.getConstant(1, CondLHS.getValueType()); 1381 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True); 1382 } else 1383 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC); 1384 } else { 1385 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now"); 1386 1387 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue(); 1388 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue(); 1389 1390 SDValue CmpOp = getValue(CB.CmpMHS); 1391 EVT VT = CmpOp.getValueType(); 1392 1393 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) { 1394 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT), 1395 ISD::SETLE); 1396 } else { 1397 SDValue SUB = DAG.getNode(ISD::SUB, dl, 1398 VT, CmpOp, DAG.getConstant(Low, VT)); 1399 Cond = DAG.getSetCC(dl, MVT::i1, SUB, 1400 DAG.getConstant(High-Low, VT), ISD::SETULE); 1401 } 1402 } 1403 1404 // Update successor info 1405 SwitchBB->addSuccessor(CB.TrueBB); 1406 SwitchBB->addSuccessor(CB.FalseBB); 1407 1408 // Set NextBlock to be the MBB immediately after the current one, if any. 1409 // This is used to avoid emitting unnecessary branches to the next block. 1410 MachineBasicBlock *NextBlock = 0; 1411 MachineFunction::iterator BBI = SwitchBB; 1412 if (++BBI != FuncInfo.MF->end()) 1413 NextBlock = BBI; 1414 1415 // If the lhs block is the next block, invert the condition so that we can 1416 // fall through to the lhs instead of the rhs block. 1417 if (CB.TrueBB == NextBlock) { 1418 std::swap(CB.TrueBB, CB.FalseBB); 1419 SDValue True = DAG.getConstant(1, Cond.getValueType()); 1420 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True); 1421 } 1422 1423 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl, 1424 MVT::Other, getControlRoot(), Cond, 1425 DAG.getBasicBlock(CB.TrueBB)); 1426 1427 // Insert the false branch. 1428 if (CB.FalseBB != NextBlock) 1429 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond, 1430 DAG.getBasicBlock(CB.FalseBB)); 1431 1432 DAG.setRoot(BrCond); 1433} 1434 1435/// visitJumpTable - Emit JumpTable node in the current MBB 1436void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) { 1437 // Emit the code for the jump table 1438 assert(JT.Reg != -1U && "Should lower JT Header first!"); 1439 EVT PTy = TLI.getPointerTy(); 1440 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), 1441 JT.Reg, PTy); 1442 SDValue Table = DAG.getJumpTable(JT.JTI, PTy); 1443 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurDebugLoc(), 1444 MVT::Other, Index.getValue(1), 1445 Table, Index); 1446 DAG.setRoot(BrJumpTable); 1447} 1448 1449/// visitJumpTableHeader - This function emits necessary code to produce index 1450/// in the JumpTable from switch case. 1451void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT, 1452 JumpTableHeader &JTH, 1453 MachineBasicBlock *SwitchBB) { 1454 // Subtract the lowest switch case value from the value being switched on and 1455 // conditional branch to default mbb if the result is greater than the 1456 // difference between smallest and largest cases. 1457 SDValue SwitchOp = getValue(JTH.SValue); 1458 EVT VT = SwitchOp.getValueType(); 1459 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp, 1460 DAG.getConstant(JTH.First, VT)); 1461 1462 // The SDNode we just created, which holds the value being switched on minus 1463 // the smallest case value, needs to be copied to a virtual register so it 1464 // can be used as an index into the jump table in a subsequent basic block. 1465 // This value may be smaller or larger than the target's pointer type, and 1466 // therefore require extension or truncating. 1467 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), TLI.getPointerTy()); 1468 1469 unsigned JumpTableReg = FuncInfo.MakeReg(TLI.getPointerTy()); 1470 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(), 1471 JumpTableReg, SwitchOp); 1472 JT.Reg = JumpTableReg; 1473 1474 // Emit the range check for the jump table, and branch to the default block 1475 // for the switch statement if the value being switched on exceeds the largest 1476 // case in the switch. 1477 SDValue CMP = DAG.getSetCC(getCurDebugLoc(), 1478 TLI.getSetCCResultType(Sub.getValueType()), Sub, 1479 DAG.getConstant(JTH.Last-JTH.First,VT), 1480 ISD::SETUGT); 1481 1482 // Set NextBlock to be the MBB immediately after the current one, if any. 1483 // This is used to avoid emitting unnecessary branches to the next block. 1484 MachineBasicBlock *NextBlock = 0; 1485 MachineFunction::iterator BBI = SwitchBB; 1486 1487 if (++BBI != FuncInfo.MF->end()) 1488 NextBlock = BBI; 1489 1490 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(), 1491 MVT::Other, CopyTo, CMP, 1492 DAG.getBasicBlock(JT.Default)); 1493 1494 if (JT.MBB != NextBlock) 1495 BrCond = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond, 1496 DAG.getBasicBlock(JT.MBB)); 1497 1498 DAG.setRoot(BrCond); 1499} 1500 1501/// visitBitTestHeader - This function emits necessary code to produce value 1502/// suitable for "bit tests" 1503void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B, 1504 MachineBasicBlock *SwitchBB) { 1505 // Subtract the minimum value 1506 SDValue SwitchOp = getValue(B.SValue); 1507 EVT VT = SwitchOp.getValueType(); 1508 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp, 1509 DAG.getConstant(B.First, VT)); 1510 1511 // Check range 1512 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(), 1513 TLI.getSetCCResultType(Sub.getValueType()), 1514 Sub, DAG.getConstant(B.Range, VT), 1515 ISD::SETUGT); 1516 1517 SDValue ShiftOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), 1518 TLI.getPointerTy()); 1519 1520 B.Reg = FuncInfo.MakeReg(TLI.getPointerTy()); 1521 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(), 1522 B.Reg, ShiftOp); 1523 1524 // Set NextBlock to be the MBB immediately after the current one, if any. 1525 // This is used to avoid emitting unnecessary branches to the next block. 1526 MachineBasicBlock *NextBlock = 0; 1527 MachineFunction::iterator BBI = SwitchBB; 1528 if (++BBI != FuncInfo.MF->end()) 1529 NextBlock = BBI; 1530 1531 MachineBasicBlock* MBB = B.Cases[0].ThisBB; 1532 1533 SwitchBB->addSuccessor(B.Default); 1534 SwitchBB->addSuccessor(MBB); 1535 1536 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(), 1537 MVT::Other, CopyTo, RangeCmp, 1538 DAG.getBasicBlock(B.Default)); 1539 1540 if (MBB != NextBlock) 1541 BrRange = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo, 1542 DAG.getBasicBlock(MBB)); 1543 1544 DAG.setRoot(BrRange); 1545} 1546 1547/// visitBitTestCase - this function produces one "bit test" 1548void SelectionDAGBuilder::visitBitTestCase(MachineBasicBlock* NextMBB, 1549 unsigned Reg, 1550 BitTestCase &B, 1551 MachineBasicBlock *SwitchBB) { 1552 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), Reg, 1553 TLI.getPointerTy()); 1554 SDValue Cmp; 1555 if (CountPopulation_64(B.Mask) == 1) { 1556 // Testing for a single bit; just compare the shift count with what it 1557 // would need to be to shift a 1 bit in that position. 1558 Cmp = DAG.getSetCC(getCurDebugLoc(), 1559 TLI.getSetCCResultType(ShiftOp.getValueType()), 1560 ShiftOp, 1561 DAG.getConstant(CountTrailingZeros_64(B.Mask), 1562 TLI.getPointerTy()), 1563 ISD::SETEQ); 1564 } else { 1565 // Make desired shift 1566 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(), 1567 TLI.getPointerTy(), 1568 DAG.getConstant(1, TLI.getPointerTy()), 1569 ShiftOp); 1570 1571 // Emit bit tests and jumps 1572 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(), 1573 TLI.getPointerTy(), SwitchVal, 1574 DAG.getConstant(B.Mask, TLI.getPointerTy())); 1575 Cmp = DAG.getSetCC(getCurDebugLoc(), 1576 TLI.getSetCCResultType(AndOp.getValueType()), 1577 AndOp, DAG.getConstant(0, TLI.getPointerTy()), 1578 ISD::SETNE); 1579 } 1580 1581 SwitchBB->addSuccessor(B.TargetBB); 1582 SwitchBB->addSuccessor(NextMBB); 1583 1584 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(), 1585 MVT::Other, getControlRoot(), 1586 Cmp, DAG.getBasicBlock(B.TargetBB)); 1587 1588 // Set NextBlock to be the MBB immediately after the current one, if any. 1589 // This is used to avoid emitting unnecessary branches to the next block. 1590 MachineBasicBlock *NextBlock = 0; 1591 MachineFunction::iterator BBI = SwitchBB; 1592 if (++BBI != FuncInfo.MF->end()) 1593 NextBlock = BBI; 1594 1595 if (NextMBB != NextBlock) 1596 BrAnd = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd, 1597 DAG.getBasicBlock(NextMBB)); 1598 1599 DAG.setRoot(BrAnd); 1600} 1601 1602void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) { 1603 MachineBasicBlock *InvokeMBB = FuncInfo.MBBMap[I.getParent()]; 1604 1605 // Retrieve successors. 1606 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)]; 1607 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)]; 1608 1609 const Value *Callee(I.getCalledValue()); 1610 if (isa<InlineAsm>(Callee)) 1611 visitInlineAsm(&I); 1612 else 1613 LowerCallTo(&I, getValue(Callee), false, LandingPad); 1614 1615 // If the value of the invoke is used outside of its defining block, make it 1616 // available as a virtual register. 1617 CopyToExportRegsIfNeeded(&I); 1618 1619 // Update successor info 1620 InvokeMBB->addSuccessor(Return); 1621 InvokeMBB->addSuccessor(LandingPad); 1622 1623 // Drop into normal successor. 1624 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(), 1625 MVT::Other, getControlRoot(), 1626 DAG.getBasicBlock(Return))); 1627} 1628 1629void SelectionDAGBuilder::visitUnwind(const UnwindInst &I) { 1630} 1631 1632/// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for 1633/// small case ranges). 1634bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR, 1635 CaseRecVector& WorkList, 1636 const Value* SV, 1637 MachineBasicBlock *Default, 1638 MachineBasicBlock *SwitchBB) { 1639 Case& BackCase = *(CR.Range.second-1); 1640 1641 // Size is the number of Cases represented by this range. 1642 size_t Size = CR.Range.second - CR.Range.first; 1643 if (Size > 3) 1644 return false; 1645 1646 // Get the MachineFunction which holds the current MBB. This is used when 1647 // inserting any additional MBBs necessary to represent the switch. 1648 MachineFunction *CurMF = FuncInfo.MF; 1649 1650 // Figure out which block is immediately after the current one. 1651 MachineBasicBlock *NextBlock = 0; 1652 MachineFunction::iterator BBI = CR.CaseBB; 1653 1654 if (++BBI != FuncInfo.MF->end()) 1655 NextBlock = BBI; 1656 1657 // TODO: If any two of the cases has the same destination, and if one value 1658 // is the same as the other, but has one bit unset that the other has set, 1659 // use bit manipulation to do two compares at once. For example: 1660 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)" 1661 1662 // Rearrange the case blocks so that the last one falls through if possible. 1663 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) { 1664 // The last case block won't fall through into 'NextBlock' if we emit the 1665 // branches in this order. See if rearranging a case value would help. 1666 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) { 1667 if (I->BB == NextBlock) { 1668 std::swap(*I, BackCase); 1669 break; 1670 } 1671 } 1672 } 1673 1674 // Create a CaseBlock record representing a conditional branch to 1675 // the Case's target mbb if the value being switched on SV is equal 1676 // to C. 1677 MachineBasicBlock *CurBlock = CR.CaseBB; 1678 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) { 1679 MachineBasicBlock *FallThrough; 1680 if (I != E-1) { 1681 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock()); 1682 CurMF->insert(BBI, FallThrough); 1683 1684 // Put SV in a virtual register to make it available from the new blocks. 1685 ExportFromCurrentBlock(SV); 1686 } else { 1687 // If the last case doesn't match, go to the default block. 1688 FallThrough = Default; 1689 } 1690 1691 const Value *RHS, *LHS, *MHS; 1692 ISD::CondCode CC; 1693 if (I->High == I->Low) { 1694 // This is just small small case range :) containing exactly 1 case 1695 CC = ISD::SETEQ; 1696 LHS = SV; RHS = I->High; MHS = NULL; 1697 } else { 1698 CC = ISD::SETLE; 1699 LHS = I->Low; MHS = SV; RHS = I->High; 1700 } 1701 CaseBlock CB(CC, LHS, RHS, MHS, I->BB, FallThrough, CurBlock); 1702 1703 // If emitting the first comparison, just call visitSwitchCase to emit the 1704 // code into the current block. Otherwise, push the CaseBlock onto the 1705 // vector to be later processed by SDISel, and insert the node's MBB 1706 // before the next MBB. 1707 if (CurBlock == SwitchBB) 1708 visitSwitchCase(CB, SwitchBB); 1709 else 1710 SwitchCases.push_back(CB); 1711 1712 CurBlock = FallThrough; 1713 } 1714 1715 return true; 1716} 1717 1718static inline bool areJTsAllowed(const TargetLowering &TLI) { 1719 return !DisableJumpTables && 1720 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) || 1721 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other)); 1722} 1723 1724static APInt ComputeRange(const APInt &First, const APInt &Last) { 1725 APInt LastExt(Last), FirstExt(First); 1726 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1; 1727 LastExt.sext(BitWidth); FirstExt.sext(BitWidth); 1728 return (LastExt - FirstExt + 1ULL); 1729} 1730 1731/// handleJTSwitchCase - Emit jumptable for current switch case range 1732bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec& CR, 1733 CaseRecVector& WorkList, 1734 const Value* SV, 1735 MachineBasicBlock* Default, 1736 MachineBasicBlock *SwitchBB) { 1737 Case& FrontCase = *CR.Range.first; 1738 Case& BackCase = *(CR.Range.second-1); 1739 1740 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue(); 1741 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue(); 1742 1743 APInt TSize(First.getBitWidth(), 0); 1744 for (CaseItr I = CR.Range.first, E = CR.Range.second; 1745 I!=E; ++I) 1746 TSize += I->size(); 1747 1748 if (!areJTsAllowed(TLI) || TSize.ult(4)) 1749 return false; 1750 1751 APInt Range = ComputeRange(First, Last); 1752 double Density = TSize.roundToDouble() / Range.roundToDouble(); 1753 if (Density < 0.4) 1754 return false; 1755 1756 DEBUG(dbgs() << "Lowering jump table\n" 1757 << "First entry: " << First << ". Last entry: " << Last << '\n' 1758 << "Range: " << Range 1759 << "Size: " << TSize << ". Density: " << Density << "\n\n"); 1760 1761 // Get the MachineFunction which holds the current MBB. This is used when 1762 // inserting any additional MBBs necessary to represent the switch. 1763 MachineFunction *CurMF = FuncInfo.MF; 1764 1765 // Figure out which block is immediately after the current one. 1766 MachineFunction::iterator BBI = CR.CaseBB; 1767 ++BBI; 1768 1769 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); 1770 1771 // Create a new basic block to hold the code for loading the address 1772 // of the jump table, and jumping to it. Update successor information; 1773 // we will either branch to the default case for the switch, or the jump 1774 // table. 1775 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB); 1776 CurMF->insert(BBI, JumpTableBB); 1777 CR.CaseBB->addSuccessor(Default); 1778 CR.CaseBB->addSuccessor(JumpTableBB); 1779 1780 // Build a vector of destination BBs, corresponding to each target 1781 // of the jump table. If the value of the jump table slot corresponds to 1782 // a case statement, push the case's BB onto the vector, otherwise, push 1783 // the default BB. 1784 std::vector<MachineBasicBlock*> DestBBs; 1785 APInt TEI = First; 1786 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) { 1787 const APInt &Low = cast<ConstantInt>(I->Low)->getValue(); 1788 const APInt &High = cast<ConstantInt>(I->High)->getValue(); 1789 1790 if (Low.sle(TEI) && TEI.sle(High)) { 1791 DestBBs.push_back(I->BB); 1792 if (TEI==High) 1793 ++I; 1794 } else { 1795 DestBBs.push_back(Default); 1796 } 1797 } 1798 1799 // Update successor info. Add one edge to each unique successor. 1800 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs()); 1801 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(), 1802 E = DestBBs.end(); I != E; ++I) { 1803 if (!SuccsHandled[(*I)->getNumber()]) { 1804 SuccsHandled[(*I)->getNumber()] = true; 1805 JumpTableBB->addSuccessor(*I); 1806 } 1807 } 1808 1809 // Create a jump table index for this jump table. 1810 unsigned JTEncoding = TLI.getJumpTableEncoding(); 1811 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding) 1812 ->createJumpTableIndex(DestBBs); 1813 1814 // Set the jump table information so that we can codegen it as a second 1815 // MachineBasicBlock 1816 JumpTable JT(-1U, JTI, JumpTableBB, Default); 1817 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB)); 1818 if (CR.CaseBB == SwitchBB) 1819 visitJumpTableHeader(JT, JTH, SwitchBB); 1820 1821 JTCases.push_back(JumpTableBlock(JTH, JT)); 1822 1823 return true; 1824} 1825 1826/// handleBTSplitSwitchCase - emit comparison and split binary search tree into 1827/// 2 subtrees. 1828bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR, 1829 CaseRecVector& WorkList, 1830 const Value* SV, 1831 MachineBasicBlock *Default, 1832 MachineBasicBlock *SwitchBB) { 1833 // Get the MachineFunction which holds the current MBB. This is used when 1834 // inserting any additional MBBs necessary to represent the switch. 1835 MachineFunction *CurMF = FuncInfo.MF; 1836 1837 // Figure out which block is immediately after the current one. 1838 MachineFunction::iterator BBI = CR.CaseBB; 1839 ++BBI; 1840 1841 Case& FrontCase = *CR.Range.first; 1842 Case& BackCase = *(CR.Range.second-1); 1843 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); 1844 1845 // Size is the number of Cases represented by this range. 1846 unsigned Size = CR.Range.second - CR.Range.first; 1847 1848 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue(); 1849 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue(); 1850 double FMetric = 0; 1851 CaseItr Pivot = CR.Range.first + Size/2; 1852 1853 // Select optimal pivot, maximizing sum density of LHS and RHS. This will 1854 // (heuristically) allow us to emit JumpTable's later. 1855 APInt TSize(First.getBitWidth(), 0); 1856 for (CaseItr I = CR.Range.first, E = CR.Range.second; 1857 I!=E; ++I) 1858 TSize += I->size(); 1859 1860 APInt LSize = FrontCase.size(); 1861 APInt RSize = TSize-LSize; 1862 DEBUG(dbgs() << "Selecting best pivot: \n" 1863 << "First: " << First << ", Last: " << Last <<'\n' 1864 << "LSize: " << LSize << ", RSize: " << RSize << '\n'); 1865 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second; 1866 J!=E; ++I, ++J) { 1867 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue(); 1868 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue(); 1869 APInt Range = ComputeRange(LEnd, RBegin); 1870 assert((Range - 2ULL).isNonNegative() && 1871 "Invalid case distance"); 1872 double LDensity = (double)LSize.roundToDouble() / 1873 (LEnd - First + 1ULL).roundToDouble(); 1874 double RDensity = (double)RSize.roundToDouble() / 1875 (Last - RBegin + 1ULL).roundToDouble(); 1876 double Metric = Range.logBase2()*(LDensity+RDensity); 1877 // Should always split in some non-trivial place 1878 DEBUG(dbgs() <<"=>Step\n" 1879 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n' 1880 << "LDensity: " << LDensity 1881 << ", RDensity: " << RDensity << '\n' 1882 << "Metric: " << Metric << '\n'); 1883 if (FMetric < Metric) { 1884 Pivot = J; 1885 FMetric = Metric; 1886 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n'); 1887 } 1888 1889 LSize += J->size(); 1890 RSize -= J->size(); 1891 } 1892 if (areJTsAllowed(TLI)) { 1893 // If our case is dense we *really* should handle it earlier! 1894 assert((FMetric > 0) && "Should handle dense range earlier!"); 1895 } else { 1896 Pivot = CR.Range.first + Size/2; 1897 } 1898 1899 CaseRange LHSR(CR.Range.first, Pivot); 1900 CaseRange RHSR(Pivot, CR.Range.second); 1901 Constant *C = Pivot->Low; 1902 MachineBasicBlock *FalseBB = 0, *TrueBB = 0; 1903 1904 // We know that we branch to the LHS if the Value being switched on is 1905 // less than the Pivot value, C. We use this to optimize our binary 1906 // tree a bit, by recognizing that if SV is greater than or equal to the 1907 // LHS's Case Value, and that Case Value is exactly one less than the 1908 // Pivot's Value, then we can branch directly to the LHS's Target, 1909 // rather than creating a leaf node for it. 1910 if ((LHSR.second - LHSR.first) == 1 && 1911 LHSR.first->High == CR.GE && 1912 cast<ConstantInt>(C)->getValue() == 1913 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) { 1914 TrueBB = LHSR.first->BB; 1915 } else { 1916 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB); 1917 CurMF->insert(BBI, TrueBB); 1918 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR)); 1919 1920 // Put SV in a virtual register to make it available from the new blocks. 1921 ExportFromCurrentBlock(SV); 1922 } 1923 1924 // Similar to the optimization above, if the Value being switched on is 1925 // known to be less than the Constant CR.LT, and the current Case Value 1926 // is CR.LT - 1, then we can branch directly to the target block for 1927 // the current Case Value, rather than emitting a RHS leaf node for it. 1928 if ((RHSR.second - RHSR.first) == 1 && CR.LT && 1929 cast<ConstantInt>(RHSR.first->Low)->getValue() == 1930 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) { 1931 FalseBB = RHSR.first->BB; 1932 } else { 1933 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB); 1934 CurMF->insert(BBI, FalseBB); 1935 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR)); 1936 1937 // Put SV in a virtual register to make it available from the new blocks. 1938 ExportFromCurrentBlock(SV); 1939 } 1940 1941 // Create a CaseBlock record representing a conditional branch to 1942 // the LHS node if the value being switched on SV is less than C. 1943 // Otherwise, branch to LHS. 1944 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB); 1945 1946 if (CR.CaseBB == SwitchBB) 1947 visitSwitchCase(CB, SwitchBB); 1948 else 1949 SwitchCases.push_back(CB); 1950 1951 return true; 1952} 1953 1954/// handleBitTestsSwitchCase - if current case range has few destination and 1955/// range span less, than machine word bitwidth, encode case range into series 1956/// of masks and emit bit tests with these masks. 1957bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR, 1958 CaseRecVector& WorkList, 1959 const Value* SV, 1960 MachineBasicBlock* Default, 1961 MachineBasicBlock *SwitchBB){ 1962 EVT PTy = TLI.getPointerTy(); 1963 unsigned IntPtrBits = PTy.getSizeInBits(); 1964 1965 Case& FrontCase = *CR.Range.first; 1966 Case& BackCase = *(CR.Range.second-1); 1967 1968 // Get the MachineFunction which holds the current MBB. This is used when 1969 // inserting any additional MBBs necessary to represent the switch. 1970 MachineFunction *CurMF = FuncInfo.MF; 1971 1972 // If target does not have legal shift left, do not emit bit tests at all. 1973 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy())) 1974 return false; 1975 1976 size_t numCmps = 0; 1977 for (CaseItr I = CR.Range.first, E = CR.Range.second; 1978 I!=E; ++I) { 1979 // Single case counts one, case range - two. 1980 numCmps += (I->Low == I->High ? 1 : 2); 1981 } 1982 1983 // Count unique destinations 1984 SmallSet<MachineBasicBlock*, 4> Dests; 1985 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) { 1986 Dests.insert(I->BB); 1987 if (Dests.size() > 3) 1988 // Don't bother the code below, if there are too much unique destinations 1989 return false; 1990 } 1991 DEBUG(dbgs() << "Total number of unique destinations: " 1992 << Dests.size() << '\n' 1993 << "Total number of comparisons: " << numCmps << '\n'); 1994 1995 // Compute span of values. 1996 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue(); 1997 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue(); 1998 APInt cmpRange = maxValue - minValue; 1999 2000 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n' 2001 << "Low bound: " << minValue << '\n' 2002 << "High bound: " << maxValue << '\n'); 2003 2004 if (cmpRange.uge(IntPtrBits) || 2005 (!(Dests.size() == 1 && numCmps >= 3) && 2006 !(Dests.size() == 2 && numCmps >= 5) && 2007 !(Dests.size() >= 3 && numCmps >= 6))) 2008 return false; 2009 2010 DEBUG(dbgs() << "Emitting bit tests\n"); 2011 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth()); 2012 2013 // Optimize the case where all the case values fit in a 2014 // word without having to subtract minValue. In this case, 2015 // we can optimize away the subtraction. 2016 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) { 2017 cmpRange = maxValue; 2018 } else { 2019 lowBound = minValue; 2020 } 2021 2022 CaseBitsVector CasesBits; 2023 unsigned i, count = 0; 2024 2025 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) { 2026 MachineBasicBlock* Dest = I->BB; 2027 for (i = 0; i < count; ++i) 2028 if (Dest == CasesBits[i].BB) 2029 break; 2030 2031 if (i == count) { 2032 assert((count < 3) && "Too much destinations to test!"); 2033 CasesBits.push_back(CaseBits(0, Dest, 0)); 2034 count++; 2035 } 2036 2037 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue(); 2038 const APInt& highValue = cast<ConstantInt>(I->High)->getValue(); 2039 2040 uint64_t lo = (lowValue - lowBound).getZExtValue(); 2041 uint64_t hi = (highValue - lowBound).getZExtValue(); 2042 2043 for (uint64_t j = lo; j <= hi; j++) { 2044 CasesBits[i].Mask |= 1ULL << j; 2045 CasesBits[i].Bits++; 2046 } 2047 2048 } 2049 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp()); 2050 2051 BitTestInfo BTC; 2052 2053 // Figure out which block is immediately after the current one. 2054 MachineFunction::iterator BBI = CR.CaseBB; 2055 ++BBI; 2056 2057 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); 2058 2059 DEBUG(dbgs() << "Cases:\n"); 2060 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) { 2061 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask 2062 << ", Bits: " << CasesBits[i].Bits 2063 << ", BB: " << CasesBits[i].BB << '\n'); 2064 2065 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB); 2066 CurMF->insert(BBI, CaseBB); 2067 BTC.push_back(BitTestCase(CasesBits[i].Mask, 2068 CaseBB, 2069 CasesBits[i].BB)); 2070 2071 // Put SV in a virtual register to make it available from the new blocks. 2072 ExportFromCurrentBlock(SV); 2073 } 2074 2075 BitTestBlock BTB(lowBound, cmpRange, SV, 2076 -1U, (CR.CaseBB == SwitchBB), 2077 CR.CaseBB, Default, BTC); 2078 2079 if (CR.CaseBB == SwitchBB) 2080 visitBitTestHeader(BTB, SwitchBB); 2081 2082 BitTestCases.push_back(BTB); 2083 2084 return true; 2085} 2086 2087/// Clusterify - Transform simple list of Cases into list of CaseRange's 2088size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases, 2089 const SwitchInst& SI) { 2090 size_t numCmps = 0; 2091 2092 // Start with "simple" cases 2093 for (size_t i = 1; i < SI.getNumSuccessors(); ++i) { 2094 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)]; 2095 Cases.push_back(Case(SI.getSuccessorValue(i), 2096 SI.getSuccessorValue(i), 2097 SMBB)); 2098 } 2099 std::sort(Cases.begin(), Cases.end(), CaseCmp()); 2100 2101 // Merge case into clusters 2102 if (Cases.size() >= 2) 2103 // Must recompute end() each iteration because it may be 2104 // invalidated by erase if we hold on to it 2105 for (CaseItr I = Cases.begin(), J = ++(Cases.begin()); J != Cases.end(); ) { 2106 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue(); 2107 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue(); 2108 MachineBasicBlock* nextBB = J->BB; 2109 MachineBasicBlock* currentBB = I->BB; 2110 2111 // If the two neighboring cases go to the same destination, merge them 2112 // into a single case. 2113 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) { 2114 I->High = J->High; 2115 J = Cases.erase(J); 2116 } else { 2117 I = J++; 2118 } 2119 } 2120 2121 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) { 2122 if (I->Low != I->High) 2123 // A range counts double, since it requires two compares. 2124 ++numCmps; 2125 } 2126 2127 return numCmps; 2128} 2129 2130void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) { 2131 MachineBasicBlock *SwitchMBB = FuncInfo.MBBMap[SI.getParent()]; 2132 2133 // Figure out which block is immediately after the current one. 2134 MachineBasicBlock *NextBlock = 0; 2135 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()]; 2136 2137 // If there is only the default destination, branch to it if it is not the 2138 // next basic block. Otherwise, just fall through. 2139 if (SI.getNumOperands() == 2) { 2140 // Update machine-CFG edges. 2141 2142 // If this is not a fall-through branch, emit the branch. 2143 SwitchMBB->addSuccessor(Default); 2144 if (Default != NextBlock) 2145 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(), 2146 MVT::Other, getControlRoot(), 2147 DAG.getBasicBlock(Default))); 2148 2149 return; 2150 } 2151 2152 // If there are any non-default case statements, create a vector of Cases 2153 // representing each one, and sort the vector so that we can efficiently 2154 // create a binary search tree from them. 2155 CaseVector Cases; 2156 size_t numCmps = Clusterify(Cases, SI); 2157 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size() 2158 << ". Total compares: " << numCmps << '\n'); 2159 numCmps = 0; 2160 2161 // Get the Value to be switched on and default basic blocks, which will be 2162 // inserted into CaseBlock records, representing basic blocks in the binary 2163 // search tree. 2164 const Value *SV = SI.getOperand(0); 2165 2166 // Push the initial CaseRec onto the worklist 2167 CaseRecVector WorkList; 2168 WorkList.push_back(CaseRec(SwitchMBB,0,0, 2169 CaseRange(Cases.begin(),Cases.end()))); 2170 2171 while (!WorkList.empty()) { 2172 // Grab a record representing a case range to process off the worklist 2173 CaseRec CR = WorkList.back(); 2174 WorkList.pop_back(); 2175 2176 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB)) 2177 continue; 2178 2179 // If the range has few cases (two or less) emit a series of specific 2180 // tests. 2181 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB)) 2182 continue; 2183 2184 // If the switch has more than 5 blocks, and at least 40% dense, and the 2185 // target supports indirect branches, then emit a jump table rather than 2186 // lowering the switch to a binary tree of conditional branches. 2187 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB)) 2188 continue; 2189 2190 // Emit binary tree. We need to pick a pivot, and push left and right ranges 2191 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call. 2192 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB); 2193 } 2194} 2195 2196void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) { 2197 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBBMap[I.getParent()]; 2198 2199 // Update machine-CFG edges with unique successors. 2200 SmallVector<BasicBlock*, 32> succs; 2201 succs.reserve(I.getNumSuccessors()); 2202 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) 2203 succs.push_back(I.getSuccessor(i)); 2204 array_pod_sort(succs.begin(), succs.end()); 2205 succs.erase(std::unique(succs.begin(), succs.end()), succs.end()); 2206 for (unsigned i = 0, e = succs.size(); i != e; ++i) 2207 IndirectBrMBB->addSuccessor(FuncInfo.MBBMap[succs[i]]); 2208 2209 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurDebugLoc(), 2210 MVT::Other, getControlRoot(), 2211 getValue(I.getAddress()))); 2212} 2213 2214void SelectionDAGBuilder::visitFSub(const User &I) { 2215 // -0.0 - X --> fneg 2216 const Type *Ty = I.getType(); 2217 if (Ty->isVectorTy()) { 2218 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) { 2219 const VectorType *DestTy = cast<VectorType>(I.getType()); 2220 const Type *ElTy = DestTy->getElementType(); 2221 unsigned VL = DestTy->getNumElements(); 2222 std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy)); 2223 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size()); 2224 if (CV == CNZ) { 2225 SDValue Op2 = getValue(I.getOperand(1)); 2226 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(), 2227 Op2.getValueType(), Op2)); 2228 return; 2229 } 2230 } 2231 } 2232 2233 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0))) 2234 if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) { 2235 SDValue Op2 = getValue(I.getOperand(1)); 2236 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(), 2237 Op2.getValueType(), Op2)); 2238 return; 2239 } 2240 2241 visitBinary(I, ISD::FSUB); 2242} 2243 2244void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) { 2245 SDValue Op1 = getValue(I.getOperand(0)); 2246 SDValue Op2 = getValue(I.getOperand(1)); 2247 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(), 2248 Op1.getValueType(), Op1, Op2)); 2249} 2250 2251void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) { 2252 SDValue Op1 = getValue(I.getOperand(0)); 2253 SDValue Op2 = getValue(I.getOperand(1)); 2254 if (!I.getType()->isVectorTy() && 2255 Op2.getValueType() != TLI.getShiftAmountTy()) { 2256 // If the operand is smaller than the shift count type, promote it. 2257 EVT PTy = TLI.getPointerTy(); 2258 EVT STy = TLI.getShiftAmountTy(); 2259 if (STy.bitsGT(Op2.getValueType())) 2260 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(), 2261 TLI.getShiftAmountTy(), Op2); 2262 // If the operand is larger than the shift count type but the shift 2263 // count type has enough bits to represent any shift value, truncate 2264 // it now. This is a common case and it exposes the truncate to 2265 // optimization early. 2266 else if (STy.getSizeInBits() >= 2267 Log2_32_Ceil(Op2.getValueType().getSizeInBits())) 2268 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), 2269 TLI.getShiftAmountTy(), Op2); 2270 // Otherwise we'll need to temporarily settle for some other 2271 // convenient type; type legalization will make adjustments as 2272 // needed. 2273 else if (PTy.bitsLT(Op2.getValueType())) 2274 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), 2275 TLI.getPointerTy(), Op2); 2276 else if (PTy.bitsGT(Op2.getValueType())) 2277 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(), 2278 TLI.getPointerTy(), Op2); 2279 } 2280 2281 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(), 2282 Op1.getValueType(), Op1, Op2)); 2283} 2284 2285void SelectionDAGBuilder::visitICmp(const User &I) { 2286 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE; 2287 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I)) 2288 predicate = IC->getPredicate(); 2289 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I)) 2290 predicate = ICmpInst::Predicate(IC->getPredicate()); 2291 SDValue Op1 = getValue(I.getOperand(0)); 2292 SDValue Op2 = getValue(I.getOperand(1)); 2293 ISD::CondCode Opcode = getICmpCondCode(predicate); 2294 2295 EVT DestVT = TLI.getValueType(I.getType()); 2296 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode)); 2297} 2298 2299void SelectionDAGBuilder::visitFCmp(const User &I) { 2300 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE; 2301 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I)) 2302 predicate = FC->getPredicate(); 2303 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I)) 2304 predicate = FCmpInst::Predicate(FC->getPredicate()); 2305 SDValue Op1 = getValue(I.getOperand(0)); 2306 SDValue Op2 = getValue(I.getOperand(1)); 2307 ISD::CondCode Condition = getFCmpCondCode(predicate); 2308 EVT DestVT = TLI.getValueType(I.getType()); 2309 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition)); 2310} 2311 2312void SelectionDAGBuilder::visitSelect(const User &I) { 2313 SmallVector<EVT, 4> ValueVTs; 2314 ComputeValueVTs(TLI, I.getType(), ValueVTs); 2315 unsigned NumValues = ValueVTs.size(); 2316 if (NumValues == 0) return; 2317 2318 SmallVector<SDValue, 4> Values(NumValues); 2319 SDValue Cond = getValue(I.getOperand(0)); 2320 SDValue TrueVal = getValue(I.getOperand(1)); 2321 SDValue FalseVal = getValue(I.getOperand(2)); 2322 2323 for (unsigned i = 0; i != NumValues; ++i) 2324 Values[i] = DAG.getNode(ISD::SELECT, getCurDebugLoc(), 2325 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i), 2326 Cond, 2327 SDValue(TrueVal.getNode(), 2328 TrueVal.getResNo() + i), 2329 SDValue(FalseVal.getNode(), 2330 FalseVal.getResNo() + i)); 2331 2332 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(), 2333 DAG.getVTList(&ValueVTs[0], NumValues), 2334 &Values[0], NumValues)); 2335} 2336 2337void SelectionDAGBuilder::visitTrunc(const User &I) { 2338 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest). 2339 SDValue N = getValue(I.getOperand(0)); 2340 EVT DestVT = TLI.getValueType(I.getType()); 2341 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N)); 2342} 2343 2344void SelectionDAGBuilder::visitZExt(const User &I) { 2345 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest). 2346 // ZExt also can't be a cast to bool for same reason. So, nothing much to do 2347 SDValue N = getValue(I.getOperand(0)); 2348 EVT DestVT = TLI.getValueType(I.getType()); 2349 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N)); 2350} 2351 2352void SelectionDAGBuilder::visitSExt(const User &I) { 2353 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest). 2354 // SExt also can't be a cast to bool for same reason. So, nothing much to do 2355 SDValue N = getValue(I.getOperand(0)); 2356 EVT DestVT = TLI.getValueType(I.getType()); 2357 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N)); 2358} 2359 2360void SelectionDAGBuilder::visitFPTrunc(const User &I) { 2361 // FPTrunc is never a no-op cast, no need to check 2362 SDValue N = getValue(I.getOperand(0)); 2363 EVT DestVT = TLI.getValueType(I.getType()); 2364 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(), 2365 DestVT, N, DAG.getIntPtrConstant(0))); 2366} 2367 2368void SelectionDAGBuilder::visitFPExt(const User &I){ 2369 // FPTrunc is never a no-op cast, no need to check 2370 SDValue N = getValue(I.getOperand(0)); 2371 EVT DestVT = TLI.getValueType(I.getType()); 2372 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N)); 2373} 2374 2375void SelectionDAGBuilder::visitFPToUI(const User &I) { 2376 // FPToUI is never a no-op cast, no need to check 2377 SDValue N = getValue(I.getOperand(0)); 2378 EVT DestVT = TLI.getValueType(I.getType()); 2379 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N)); 2380} 2381 2382void SelectionDAGBuilder::visitFPToSI(const User &I) { 2383 // FPToSI is never a no-op cast, no need to check 2384 SDValue N = getValue(I.getOperand(0)); 2385 EVT DestVT = TLI.getValueType(I.getType()); 2386 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N)); 2387} 2388 2389void SelectionDAGBuilder::visitUIToFP(const User &I) { 2390 // UIToFP is never a no-op cast, no need to check 2391 SDValue N = getValue(I.getOperand(0)); 2392 EVT DestVT = TLI.getValueType(I.getType()); 2393 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N)); 2394} 2395 2396void SelectionDAGBuilder::visitSIToFP(const User &I){ 2397 // SIToFP is never a no-op cast, no need to check 2398 SDValue N = getValue(I.getOperand(0)); 2399 EVT DestVT = TLI.getValueType(I.getType()); 2400 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N)); 2401} 2402 2403void SelectionDAGBuilder::visitPtrToInt(const User &I) { 2404 // What to do depends on the size of the integer and the size of the pointer. 2405 // We can either truncate, zero extend, or no-op, accordingly. 2406 SDValue N = getValue(I.getOperand(0)); 2407 EVT SrcVT = N.getValueType(); 2408 EVT DestVT = TLI.getValueType(I.getType()); 2409 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT)); 2410} 2411 2412void SelectionDAGBuilder::visitIntToPtr(const User &I) { 2413 // What to do depends on the size of the integer and the size of the pointer. 2414 // We can either truncate, zero extend, or no-op, accordingly. 2415 SDValue N = getValue(I.getOperand(0)); 2416 EVT SrcVT = N.getValueType(); 2417 EVT DestVT = TLI.getValueType(I.getType()); 2418 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT)); 2419} 2420 2421void SelectionDAGBuilder::visitBitCast(const User &I) { 2422 SDValue N = getValue(I.getOperand(0)); 2423 EVT DestVT = TLI.getValueType(I.getType()); 2424 2425 // BitCast assures us that source and destination are the same size so this is 2426 // either a BIT_CONVERT or a no-op. 2427 if (DestVT != N.getValueType()) 2428 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), 2429 DestVT, N)); // convert types. 2430 else 2431 setValue(&I, N); // noop cast. 2432} 2433 2434void SelectionDAGBuilder::visitInsertElement(const User &I) { 2435 SDValue InVec = getValue(I.getOperand(0)); 2436 SDValue InVal = getValue(I.getOperand(1)); 2437 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), 2438 TLI.getPointerTy(), 2439 getValue(I.getOperand(2))); 2440 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(), 2441 TLI.getValueType(I.getType()), 2442 InVec, InVal, InIdx)); 2443} 2444 2445void SelectionDAGBuilder::visitExtractElement(const User &I) { 2446 SDValue InVec = getValue(I.getOperand(0)); 2447 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), 2448 TLI.getPointerTy(), 2449 getValue(I.getOperand(1))); 2450 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(), 2451 TLI.getValueType(I.getType()), InVec, InIdx)); 2452} 2453 2454// Utility for visitShuffleVector - Returns true if the mask is mask starting 2455// from SIndx and increasing to the element length (undefs are allowed). 2456static bool SequentialMask(SmallVectorImpl<int> &Mask, unsigned SIndx) { 2457 unsigned MaskNumElts = Mask.size(); 2458 for (unsigned i = 0; i != MaskNumElts; ++i) 2459 if ((Mask[i] >= 0) && (Mask[i] != (int)(i + SIndx))) 2460 return false; 2461 return true; 2462} 2463 2464void SelectionDAGBuilder::visitShuffleVector(const User &I) { 2465 SmallVector<int, 8> Mask; 2466 SDValue Src1 = getValue(I.getOperand(0)); 2467 SDValue Src2 = getValue(I.getOperand(1)); 2468 2469 // Convert the ConstantVector mask operand into an array of ints, with -1 2470 // representing undef values. 2471 SmallVector<Constant*, 8> MaskElts; 2472 cast<Constant>(I.getOperand(2))->getVectorElements(MaskElts); 2473 unsigned MaskNumElts = MaskElts.size(); 2474 for (unsigned i = 0; i != MaskNumElts; ++i) { 2475 if (isa<UndefValue>(MaskElts[i])) 2476 Mask.push_back(-1); 2477 else 2478 Mask.push_back(cast<ConstantInt>(MaskElts[i])->getSExtValue()); 2479 } 2480 2481 EVT VT = TLI.getValueType(I.getType()); 2482 EVT SrcVT = Src1.getValueType(); 2483 unsigned SrcNumElts = SrcVT.getVectorNumElements(); 2484 2485 if (SrcNumElts == MaskNumElts) { 2486 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2, 2487 &Mask[0])); 2488 return; 2489 } 2490 2491 // Normalize the shuffle vector since mask and vector length don't match. 2492 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) { 2493 // Mask is longer than the source vectors and is a multiple of the source 2494 // vectors. We can use concatenate vector to make the mask and vectors 2495 // lengths match. 2496 if (SrcNumElts*2 == MaskNumElts && SequentialMask(Mask, 0)) { 2497 // The shuffle is concatenating two vectors together. 2498 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(), 2499 VT, Src1, Src2)); 2500 return; 2501 } 2502 2503 // Pad both vectors with undefs to make them the same length as the mask. 2504 unsigned NumConcat = MaskNumElts / SrcNumElts; 2505 bool Src1U = Src1.getOpcode() == ISD::UNDEF; 2506 bool Src2U = Src2.getOpcode() == ISD::UNDEF; 2507 SDValue UndefVal = DAG.getUNDEF(SrcVT); 2508 2509 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal); 2510 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal); 2511 MOps1[0] = Src1; 2512 MOps2[0] = Src2; 2513 2514 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS, 2515 getCurDebugLoc(), VT, 2516 &MOps1[0], NumConcat); 2517 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS, 2518 getCurDebugLoc(), VT, 2519 &MOps2[0], NumConcat); 2520 2521 // Readjust mask for new input vector length. 2522 SmallVector<int, 8> MappedOps; 2523 for (unsigned i = 0; i != MaskNumElts; ++i) { 2524 int Idx = Mask[i]; 2525 if (Idx < (int)SrcNumElts) 2526 MappedOps.push_back(Idx); 2527 else 2528 MappedOps.push_back(Idx + MaskNumElts - SrcNumElts); 2529 } 2530 2531 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2, 2532 &MappedOps[0])); 2533 return; 2534 } 2535 2536 if (SrcNumElts > MaskNumElts) { 2537 // Analyze the access pattern of the vector to see if we can extract 2538 // two subvectors and do the shuffle. The analysis is done by calculating 2539 // the range of elements the mask access on both vectors. 2540 int MinRange[2] = { SrcNumElts+1, SrcNumElts+1}; 2541 int MaxRange[2] = {-1, -1}; 2542 2543 for (unsigned i = 0; i != MaskNumElts; ++i) { 2544 int Idx = Mask[i]; 2545 int Input = 0; 2546 if (Idx < 0) 2547 continue; 2548 2549 if (Idx >= (int)SrcNumElts) { 2550 Input = 1; 2551 Idx -= SrcNumElts; 2552 } 2553 if (Idx > MaxRange[Input]) 2554 MaxRange[Input] = Idx; 2555 if (Idx < MinRange[Input]) 2556 MinRange[Input] = Idx; 2557 } 2558 2559 // Check if the access is smaller than the vector size and can we find 2560 // a reasonable extract index. 2561 int RangeUse[2] = { 2, 2 }; // 0 = Unused, 1 = Extract, 2 = Can not 2562 // Extract. 2563 int StartIdx[2]; // StartIdx to extract from 2564 for (int Input=0; Input < 2; ++Input) { 2565 if (MinRange[Input] == (int)(SrcNumElts+1) && MaxRange[Input] == -1) { 2566 RangeUse[Input] = 0; // Unused 2567 StartIdx[Input] = 0; 2568 } else if (MaxRange[Input] - MinRange[Input] < (int)MaskNumElts) { 2569 // Fits within range but we should see if we can find a good 2570 // start index that is a multiple of the mask length. 2571 if (MaxRange[Input] < (int)MaskNumElts) { 2572 RangeUse[Input] = 1; // Extract from beginning of the vector 2573 StartIdx[Input] = 0; 2574 } else { 2575 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts; 2576 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts && 2577 StartIdx[Input] + MaskNumElts < SrcNumElts) 2578 RangeUse[Input] = 1; // Extract from a multiple of the mask length. 2579 } 2580 } 2581 } 2582 2583 if (RangeUse[0] == 0 && RangeUse[1] == 0) { 2584 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used. 2585 return; 2586 } 2587 else if (RangeUse[0] < 2 && RangeUse[1] < 2) { 2588 // Extract appropriate subvector and generate a vector shuffle 2589 for (int Input=0; Input < 2; ++Input) { 2590 SDValue &Src = Input == 0 ? Src1 : Src2; 2591 if (RangeUse[Input] == 0) 2592 Src = DAG.getUNDEF(VT); 2593 else 2594 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT, 2595 Src, DAG.getIntPtrConstant(StartIdx[Input])); 2596 } 2597 2598 // Calculate new mask. 2599 SmallVector<int, 8> MappedOps; 2600 for (unsigned i = 0; i != MaskNumElts; ++i) { 2601 int Idx = Mask[i]; 2602 if (Idx < 0) 2603 MappedOps.push_back(Idx); 2604 else if (Idx < (int)SrcNumElts) 2605 MappedOps.push_back(Idx - StartIdx[0]); 2606 else 2607 MappedOps.push_back(Idx - SrcNumElts - StartIdx[1] + MaskNumElts); 2608 } 2609 2610 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2, 2611 &MappedOps[0])); 2612 return; 2613 } 2614 } 2615 2616 // We can't use either concat vectors or extract subvectors so fall back to 2617 // replacing the shuffle with extract and build vector. 2618 // to insert and build vector. 2619 EVT EltVT = VT.getVectorElementType(); 2620 EVT PtrVT = TLI.getPointerTy(); 2621 SmallVector<SDValue,8> Ops; 2622 for (unsigned i = 0; i != MaskNumElts; ++i) { 2623 if (Mask[i] < 0) { 2624 Ops.push_back(DAG.getUNDEF(EltVT)); 2625 } else { 2626 int Idx = Mask[i]; 2627 SDValue Res; 2628 2629 if (Idx < (int)SrcNumElts) 2630 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(), 2631 EltVT, Src1, DAG.getConstant(Idx, PtrVT)); 2632 else 2633 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(), 2634 EltVT, Src2, 2635 DAG.getConstant(Idx - SrcNumElts, PtrVT)); 2636 2637 Ops.push_back(Res); 2638 } 2639 } 2640 2641 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(), 2642 VT, &Ops[0], Ops.size())); 2643} 2644 2645void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) { 2646 const Value *Op0 = I.getOperand(0); 2647 const Value *Op1 = I.getOperand(1); 2648 const Type *AggTy = I.getType(); 2649 const Type *ValTy = Op1->getType(); 2650 bool IntoUndef = isa<UndefValue>(Op0); 2651 bool FromUndef = isa<UndefValue>(Op1); 2652 2653 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy, 2654 I.idx_begin(), I.idx_end()); 2655 2656 SmallVector<EVT, 4> AggValueVTs; 2657 ComputeValueVTs(TLI, AggTy, AggValueVTs); 2658 SmallVector<EVT, 4> ValValueVTs; 2659 ComputeValueVTs(TLI, ValTy, ValValueVTs); 2660 2661 unsigned NumAggValues = AggValueVTs.size(); 2662 unsigned NumValValues = ValValueVTs.size(); 2663 SmallVector<SDValue, 4> Values(NumAggValues); 2664 2665 SDValue Agg = getValue(Op0); 2666 SDValue Val = getValue(Op1); 2667 unsigned i = 0; 2668 // Copy the beginning value(s) from the original aggregate. 2669 for (; i != LinearIndex; ++i) 2670 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) : 2671 SDValue(Agg.getNode(), Agg.getResNo() + i); 2672 // Copy values from the inserted value(s). 2673 for (; i != LinearIndex + NumValValues; ++i) 2674 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) : 2675 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex); 2676 // Copy remaining value(s) from the original aggregate. 2677 for (; i != NumAggValues; ++i) 2678 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) : 2679 SDValue(Agg.getNode(), Agg.getResNo() + i); 2680 2681 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(), 2682 DAG.getVTList(&AggValueVTs[0], NumAggValues), 2683 &Values[0], NumAggValues)); 2684} 2685 2686void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) { 2687 const Value *Op0 = I.getOperand(0); 2688 const Type *AggTy = Op0->getType(); 2689 const Type *ValTy = I.getType(); 2690 bool OutOfUndef = isa<UndefValue>(Op0); 2691 2692 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy, 2693 I.idx_begin(), I.idx_end()); 2694 2695 SmallVector<EVT, 4> ValValueVTs; 2696 ComputeValueVTs(TLI, ValTy, ValValueVTs); 2697 2698 unsigned NumValValues = ValValueVTs.size(); 2699 SmallVector<SDValue, 4> Values(NumValValues); 2700 2701 SDValue Agg = getValue(Op0); 2702 // Copy out the selected value(s). 2703 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i) 2704 Values[i - LinearIndex] = 2705 OutOfUndef ? 2706 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) : 2707 SDValue(Agg.getNode(), Agg.getResNo() + i); 2708 2709 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(), 2710 DAG.getVTList(&ValValueVTs[0], NumValValues), 2711 &Values[0], NumValValues)); 2712} 2713 2714void SelectionDAGBuilder::visitGetElementPtr(const User &I) { 2715 SDValue N = getValue(I.getOperand(0)); 2716 const Type *Ty = I.getOperand(0)->getType(); 2717 2718 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end(); 2719 OI != E; ++OI) { 2720 const Value *Idx = *OI; 2721 if (const StructType *StTy = dyn_cast<StructType>(Ty)) { 2722 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue(); 2723 if (Field) { 2724 // N = N + Offset 2725 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field); 2726 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N, 2727 DAG.getIntPtrConstant(Offset)); 2728 } 2729 2730 Ty = StTy->getElementType(Field); 2731 } else if (const UnionType *UnTy = dyn_cast<UnionType>(Ty)) { 2732 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue(); 2733 2734 // Offset canonically 0 for unions, but type changes 2735 Ty = UnTy->getElementType(Field); 2736 } else { 2737 Ty = cast<SequentialType>(Ty)->getElementType(); 2738 2739 // If this is a constant subscript, handle it quickly. 2740 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) { 2741 if (CI->isZero()) continue; 2742 uint64_t Offs = 2743 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue(); 2744 SDValue OffsVal; 2745 EVT PTy = TLI.getPointerTy(); 2746 unsigned PtrBits = PTy.getSizeInBits(); 2747 if (PtrBits < 64) 2748 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), 2749 TLI.getPointerTy(), 2750 DAG.getConstant(Offs, MVT::i64)); 2751 else 2752 OffsVal = DAG.getIntPtrConstant(Offs); 2753 2754 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N, 2755 OffsVal); 2756 continue; 2757 } 2758 2759 // N = N + Idx * ElementSize; 2760 APInt ElementSize = APInt(TLI.getPointerTy().getSizeInBits(), 2761 TD->getTypeAllocSize(Ty)); 2762 SDValue IdxN = getValue(Idx); 2763 2764 // If the index is smaller or larger than intptr_t, truncate or extend 2765 // it. 2766 IdxN = DAG.getSExtOrTrunc(IdxN, getCurDebugLoc(), N.getValueType()); 2767 2768 // If this is a multiply by a power of two, turn it into a shl 2769 // immediately. This is a very common case. 2770 if (ElementSize != 1) { 2771 if (ElementSize.isPowerOf2()) { 2772 unsigned Amt = ElementSize.logBase2(); 2773 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(), 2774 N.getValueType(), IdxN, 2775 DAG.getConstant(Amt, TLI.getPointerTy())); 2776 } else { 2777 SDValue Scale = DAG.getConstant(ElementSize, TLI.getPointerTy()); 2778 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(), 2779 N.getValueType(), IdxN, Scale); 2780 } 2781 } 2782 2783 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), 2784 N.getValueType(), N, IdxN); 2785 } 2786 } 2787 2788 setValue(&I, N); 2789} 2790 2791void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) { 2792 // If this is a fixed sized alloca in the entry block of the function, 2793 // allocate it statically on the stack. 2794 if (FuncInfo.StaticAllocaMap.count(&I)) 2795 return; // getValue will auto-populate this. 2796 2797 const Type *Ty = I.getAllocatedType(); 2798 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty); 2799 unsigned Align = 2800 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), 2801 I.getAlignment()); 2802 2803 SDValue AllocSize = getValue(I.getArraySize()); 2804 2805 EVT IntPtr = TLI.getPointerTy(); 2806 if (AllocSize.getValueType() != IntPtr) 2807 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurDebugLoc(), IntPtr); 2808 2809 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), IntPtr, 2810 AllocSize, 2811 DAG.getConstant(TySize, IntPtr)); 2812 2813 // Handle alignment. If the requested alignment is less than or equal to 2814 // the stack alignment, ignore it. If the size is greater than or equal to 2815 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node. 2816 unsigned StackAlign = TM.getFrameInfo()->getStackAlignment(); 2817 if (Align <= StackAlign) 2818 Align = 0; 2819 2820 // Round the size of the allocation up to the stack alignment size 2821 // by add SA-1 to the size. 2822 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(), 2823 AllocSize.getValueType(), AllocSize, 2824 DAG.getIntPtrConstant(StackAlign-1)); 2825 2826 // Mask out the low bits for alignment purposes. 2827 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(), 2828 AllocSize.getValueType(), AllocSize, 2829 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1))); 2830 2831 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) }; 2832 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other); 2833 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(), 2834 VTs, Ops, 3); 2835 setValue(&I, DSA); 2836 DAG.setRoot(DSA.getValue(1)); 2837 2838 // Inform the Frame Information that we have just allocated a variable-sized 2839 // object. 2840 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(); 2841} 2842 2843void SelectionDAGBuilder::visitLoad(const LoadInst &I) { 2844 const Value *SV = I.getOperand(0); 2845 SDValue Ptr = getValue(SV); 2846 2847 const Type *Ty = I.getType(); 2848 2849 bool isVolatile = I.isVolatile(); 2850 bool isNonTemporal = I.getMetadata("nontemporal") != 0; 2851 unsigned Alignment = I.getAlignment(); 2852 2853 SmallVector<EVT, 4> ValueVTs; 2854 SmallVector<uint64_t, 4> Offsets; 2855 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets); 2856 unsigned NumValues = ValueVTs.size(); 2857 if (NumValues == 0) 2858 return; 2859 2860 SDValue Root; 2861 bool ConstantMemory = false; 2862 if (I.isVolatile()) 2863 // Serialize volatile loads with other side effects. 2864 Root = getRoot(); 2865 else if (AA->pointsToConstantMemory(SV)) { 2866 // Do not serialize (non-volatile) loads of constant memory with anything. 2867 Root = DAG.getEntryNode(); 2868 ConstantMemory = true; 2869 } else { 2870 // Do not serialize non-volatile loads against each other. 2871 Root = DAG.getRoot(); 2872 } 2873 2874 SmallVector<SDValue, 4> Values(NumValues); 2875 SmallVector<SDValue, 4> Chains(NumValues); 2876 EVT PtrVT = Ptr.getValueType(); 2877 for (unsigned i = 0; i != NumValues; ++i) { 2878 SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(), 2879 PtrVT, Ptr, 2880 DAG.getConstant(Offsets[i], PtrVT)); 2881 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root, 2882 A, SV, Offsets[i], isVolatile, 2883 isNonTemporal, Alignment); 2884 2885 Values[i] = L; 2886 Chains[i] = L.getValue(1); 2887 } 2888 2889 if (!ConstantMemory) { 2890 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), 2891 MVT::Other, &Chains[0], NumValues); 2892 if (isVolatile) 2893 DAG.setRoot(Chain); 2894 else 2895 PendingLoads.push_back(Chain); 2896 } 2897 2898 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(), 2899 DAG.getVTList(&ValueVTs[0], NumValues), 2900 &Values[0], NumValues)); 2901} 2902 2903void SelectionDAGBuilder::visitStore(const StoreInst &I) { 2904 const Value *SrcV = I.getOperand(0); 2905 const Value *PtrV = I.getOperand(1); 2906 2907 SmallVector<EVT, 4> ValueVTs; 2908 SmallVector<uint64_t, 4> Offsets; 2909 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets); 2910 unsigned NumValues = ValueVTs.size(); 2911 if (NumValues == 0) 2912 return; 2913 2914 // Get the lowered operands. Note that we do this after 2915 // checking if NumResults is zero, because with zero results 2916 // the operands won't have values in the map. 2917 SDValue Src = getValue(SrcV); 2918 SDValue Ptr = getValue(PtrV); 2919 2920 SDValue Root = getRoot(); 2921 SmallVector<SDValue, 4> Chains(NumValues); 2922 EVT PtrVT = Ptr.getValueType(); 2923 bool isVolatile = I.isVolatile(); 2924 bool isNonTemporal = I.getMetadata("nontemporal") != 0; 2925 unsigned Alignment = I.getAlignment(); 2926 2927 for (unsigned i = 0; i != NumValues; ++i) { 2928 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr, 2929 DAG.getConstant(Offsets[i], PtrVT)); 2930 Chains[i] = DAG.getStore(Root, getCurDebugLoc(), 2931 SDValue(Src.getNode(), Src.getResNo() + i), 2932 Add, PtrV, Offsets[i], isVolatile, 2933 isNonTemporal, Alignment); 2934 } 2935 2936 DAG.setRoot(DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), 2937 MVT::Other, &Chains[0], NumValues)); 2938} 2939 2940/// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC 2941/// node. 2942void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I, 2943 unsigned Intrinsic) { 2944 bool HasChain = !I.doesNotAccessMemory(); 2945 bool OnlyLoad = HasChain && I.onlyReadsMemory(); 2946 2947 // Build the operand list. 2948 SmallVector<SDValue, 8> Ops; 2949 if (HasChain) { // If this intrinsic has side-effects, chainify it. 2950 if (OnlyLoad) { 2951 // We don't need to serialize loads against other loads. 2952 Ops.push_back(DAG.getRoot()); 2953 } else { 2954 Ops.push_back(getRoot()); 2955 } 2956 } 2957 2958 // Info is set by getTgtMemInstrinsic 2959 TargetLowering::IntrinsicInfo Info; 2960 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic); 2961 2962 // Add the intrinsic ID as an integer operand if it's not a target intrinsic. 2963 if (!IsTgtIntrinsic) 2964 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy())); 2965 2966 // Add all operands of the call to the operand list. 2967 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) { 2968 SDValue Op = getValue(I.getArgOperand(i)); 2969 assert(TLI.isTypeLegal(Op.getValueType()) && 2970 "Intrinsic uses a non-legal type?"); 2971 Ops.push_back(Op); 2972 } 2973 2974 SmallVector<EVT, 4> ValueVTs; 2975 ComputeValueVTs(TLI, I.getType(), ValueVTs); 2976#ifndef NDEBUG 2977 for (unsigned Val = 0, E = ValueVTs.size(); Val != E; ++Val) { 2978 assert(TLI.isTypeLegal(ValueVTs[Val]) && 2979 "Intrinsic uses a non-legal type?"); 2980 } 2981#endif // NDEBUG 2982 2983 if (HasChain) 2984 ValueVTs.push_back(MVT::Other); 2985 2986 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size()); 2987 2988 // Create the node. 2989 SDValue Result; 2990 if (IsTgtIntrinsic) { 2991 // This is target intrinsic that touches memory 2992 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(), 2993 VTs, &Ops[0], Ops.size(), 2994 Info.memVT, Info.ptrVal, Info.offset, 2995 Info.align, Info.vol, 2996 Info.readMem, Info.writeMem); 2997 } else if (!HasChain) { 2998 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(), 2999 VTs, &Ops[0], Ops.size()); 3000 } else if (!I.getType()->isVoidTy()) { 3001 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(), 3002 VTs, &Ops[0], Ops.size()); 3003 } else { 3004 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(), 3005 VTs, &Ops[0], Ops.size()); 3006 } 3007 3008 if (HasChain) { 3009 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1); 3010 if (OnlyLoad) 3011 PendingLoads.push_back(Chain); 3012 else 3013 DAG.setRoot(Chain); 3014 } 3015 3016 if (!I.getType()->isVoidTy()) { 3017 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) { 3018 EVT VT = TLI.getValueType(PTy); 3019 Result = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), VT, Result); 3020 } 3021 3022 setValue(&I, Result); 3023 } 3024} 3025 3026/// GetSignificand - Get the significand and build it into a floating-point 3027/// number with exponent of 1: 3028/// 3029/// Op = (Op & 0x007fffff) | 0x3f800000; 3030/// 3031/// where Op is the hexidecimal representation of floating point value. 3032static SDValue 3033GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) { 3034 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op, 3035 DAG.getConstant(0x007fffff, MVT::i32)); 3036 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1, 3037 DAG.getConstant(0x3f800000, MVT::i32)); 3038 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t2); 3039} 3040 3041/// GetExponent - Get the exponent: 3042/// 3043/// (float)(int)(((Op & 0x7f800000) >> 23) - 127); 3044/// 3045/// where Op is the hexidecimal representation of floating point value. 3046static SDValue 3047GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI, 3048 DebugLoc dl) { 3049 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op, 3050 DAG.getConstant(0x7f800000, MVT::i32)); 3051 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0, 3052 DAG.getConstant(23, TLI.getPointerTy())); 3053 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1, 3054 DAG.getConstant(127, MVT::i32)); 3055 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2); 3056} 3057 3058/// getF32Constant - Get 32-bit floating point constant. 3059static SDValue 3060getF32Constant(SelectionDAG &DAG, unsigned Flt) { 3061 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32); 3062} 3063 3064/// Inlined utility function to implement binary input atomic intrinsics for 3065/// visitIntrinsicCall: I is a call instruction 3066/// Op is the associated NodeType for I 3067const char * 3068SelectionDAGBuilder::implVisitBinaryAtomic(const CallInst& I, 3069 ISD::NodeType Op) { 3070 SDValue Root = getRoot(); 3071 SDValue L = 3072 DAG.getAtomic(Op, getCurDebugLoc(), 3073 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(), 3074 Root, 3075 getValue(I.getArgOperand(0)), 3076 getValue(I.getArgOperand(1)), 3077 I.getArgOperand(0)); 3078 setValue(&I, L); 3079 DAG.setRoot(L.getValue(1)); 3080 return 0; 3081} 3082 3083// implVisitAluOverflow - Lower arithmetic overflow instrinsics. 3084const char * 3085SelectionDAGBuilder::implVisitAluOverflow(const CallInst &I, ISD::NodeType Op) { 3086 SDValue Op1 = getValue(I.getArgOperand(0)); 3087 SDValue Op2 = getValue(I.getArgOperand(1)); 3088 3089 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1); 3090 setValue(&I, DAG.getNode(Op, getCurDebugLoc(), VTs, Op1, Op2)); 3091 return 0; 3092} 3093 3094/// visitExp - Lower an exp intrinsic. Handles the special sequences for 3095/// limited-precision mode. 3096void 3097SelectionDAGBuilder::visitExp(const CallInst &I) { 3098 SDValue result; 3099 DebugLoc dl = getCurDebugLoc(); 3100 3101 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 && 3102 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3103 SDValue Op = getValue(I.getArgOperand(0)); 3104 3105 // Put the exponent in the right bit position for later addition to the 3106 // final result: 3107 // 3108 // #define LOG2OFe 1.4426950f 3109 // IntegerPartOfX = ((int32_t)(X * LOG2OFe)); 3110 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op, 3111 getF32Constant(DAG, 0x3fb8aa3b)); 3112 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0); 3113 3114 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX; 3115 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); 3116 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1); 3117 3118 // IntegerPartOfX <<= 23; 3119 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, 3120 DAG.getConstant(23, TLI.getPointerTy())); 3121 3122 if (LimitFloatPrecision <= 6) { 3123 // For floating-point precision of 6: 3124 // 3125 // TwoToFractionalPartOfX = 3126 // 0.997535578f + 3127 // (0.735607626f + 0.252464424f * x) * x; 3128 // 3129 // error 0.0144103317, which is 6 bits 3130 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3131 getF32Constant(DAG, 0x3e814304)); 3132 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3133 getF32Constant(DAG, 0x3f3c50c8)); 3134 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3135 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3136 getF32Constant(DAG, 0x3f7f5e7e)); 3137 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t5); 3138 3139 // Add the exponent into the result in integer domain. 3140 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32, 3141 TwoToFracPartOfX, IntegerPartOfX); 3142 3143 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t6); 3144 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) { 3145 // For floating-point precision of 12: 3146 // 3147 // TwoToFractionalPartOfX = 3148 // 0.999892986f + 3149 // (0.696457318f + 3150 // (0.224338339f + 0.792043434e-1f * x) * x) * x; 3151 // 3152 // 0.000107046256 error, which is 13 to 14 bits 3153 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3154 getF32Constant(DAG, 0x3da235e3)); 3155 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3156 getF32Constant(DAG, 0x3e65b8f3)); 3157 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3158 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3159 getF32Constant(DAG, 0x3f324b07)); 3160 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3161 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3162 getF32Constant(DAG, 0x3f7ff8fd)); 3163 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t7); 3164 3165 // Add the exponent into the result in integer domain. 3166 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32, 3167 TwoToFracPartOfX, IntegerPartOfX); 3168 3169 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t8); 3170 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18 3171 // For floating-point precision of 18: 3172 // 3173 // TwoToFractionalPartOfX = 3174 // 0.999999982f + 3175 // (0.693148872f + 3176 // (0.240227044f + 3177 // (0.554906021e-1f + 3178 // (0.961591928e-2f + 3179 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; 3180 // 3181 // error 2.47208000*10^(-7), which is better than 18 bits 3182 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3183 getF32Constant(DAG, 0x3924b03e)); 3184 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3185 getF32Constant(DAG, 0x3ab24b87)); 3186 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3187 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3188 getF32Constant(DAG, 0x3c1d8c17)); 3189 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3190 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3191 getF32Constant(DAG, 0x3d634a1d)); 3192 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3193 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 3194 getF32Constant(DAG, 0x3e75fe14)); 3195 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 3196 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, 3197 getF32Constant(DAG, 0x3f317234)); 3198 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); 3199 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, 3200 getF32Constant(DAG, 0x3f800000)); 3201 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl, 3202 MVT::i32, t13); 3203 3204 // Add the exponent into the result in integer domain. 3205 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32, 3206 TwoToFracPartOfX, IntegerPartOfX); 3207 3208 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t14); 3209 } 3210 } else { 3211 // No special expansion. 3212 result = DAG.getNode(ISD::FEXP, dl, 3213 getValue(I.getArgOperand(0)).getValueType(), 3214 getValue(I.getArgOperand(0))); 3215 } 3216 3217 setValue(&I, result); 3218} 3219 3220/// visitLog - Lower a log intrinsic. Handles the special sequences for 3221/// limited-precision mode. 3222void 3223SelectionDAGBuilder::visitLog(const CallInst &I) { 3224 SDValue result; 3225 DebugLoc dl = getCurDebugLoc(); 3226 3227 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 && 3228 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3229 SDValue Op = getValue(I.getArgOperand(0)); 3230 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op); 3231 3232 // Scale the exponent by log(2) [0.69314718f]. 3233 SDValue Exp = GetExponent(DAG, Op1, TLI, dl); 3234 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp, 3235 getF32Constant(DAG, 0x3f317218)); 3236 3237 // Get the significand and build it into a floating-point number with 3238 // exponent of 1. 3239 SDValue X = GetSignificand(DAG, Op1, dl); 3240 3241 if (LimitFloatPrecision <= 6) { 3242 // For floating-point precision of 6: 3243 // 3244 // LogofMantissa = 3245 // -1.1609546f + 3246 // (1.4034025f - 0.23903021f * x) * x; 3247 // 3248 // error 0.0034276066, which is better than 8 bits 3249 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3250 getF32Constant(DAG, 0xbe74c456)); 3251 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3252 getF32Constant(DAG, 0x3fb3a2b1)); 3253 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3254 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3255 getF32Constant(DAG, 0x3f949a29)); 3256 3257 result = DAG.getNode(ISD::FADD, dl, 3258 MVT::f32, LogOfExponent, LogOfMantissa); 3259 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) { 3260 // For floating-point precision of 12: 3261 // 3262 // LogOfMantissa = 3263 // -1.7417939f + 3264 // (2.8212026f + 3265 // (-1.4699568f + 3266 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x; 3267 // 3268 // error 0.000061011436, which is 14 bits 3269 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3270 getF32Constant(DAG, 0xbd67b6d6)); 3271 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3272 getF32Constant(DAG, 0x3ee4f4b8)); 3273 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3274 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3275 getF32Constant(DAG, 0x3fbc278b)); 3276 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3277 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3278 getF32Constant(DAG, 0x40348e95)); 3279 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3280 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, 3281 getF32Constant(DAG, 0x3fdef31a)); 3282 3283 result = DAG.getNode(ISD::FADD, dl, 3284 MVT::f32, LogOfExponent, LogOfMantissa); 3285 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18 3286 // For floating-point precision of 18: 3287 // 3288 // LogOfMantissa = 3289 // -2.1072184f + 3290 // (4.2372794f + 3291 // (-3.7029485f + 3292 // (2.2781945f + 3293 // (-0.87823314f + 3294 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x; 3295 // 3296 // error 0.0000023660568, which is better than 18 bits 3297 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3298 getF32Constant(DAG, 0xbc91e5ac)); 3299 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3300 getF32Constant(DAG, 0x3e4350aa)); 3301 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3302 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3303 getF32Constant(DAG, 0x3f60d3e3)); 3304 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3305 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3306 getF32Constant(DAG, 0x4011cdf0)); 3307 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3308 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, 3309 getF32Constant(DAG, 0x406cfd1c)); 3310 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3311 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 3312 getF32Constant(DAG, 0x408797cb)); 3313 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 3314 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10, 3315 getF32Constant(DAG, 0x4006dcab)); 3316 3317 result = DAG.getNode(ISD::FADD, dl, 3318 MVT::f32, LogOfExponent, LogOfMantissa); 3319 } 3320 } else { 3321 // No special expansion. 3322 result = DAG.getNode(ISD::FLOG, dl, 3323 getValue(I.getArgOperand(0)).getValueType(), 3324 getValue(I.getArgOperand(0))); 3325 } 3326 3327 setValue(&I, result); 3328} 3329 3330/// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for 3331/// limited-precision mode. 3332void 3333SelectionDAGBuilder::visitLog2(const CallInst &I) { 3334 SDValue result; 3335 DebugLoc dl = getCurDebugLoc(); 3336 3337 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 && 3338 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3339 SDValue Op = getValue(I.getArgOperand(0)); 3340 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op); 3341 3342 // Get the exponent. 3343 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl); 3344 3345 // Get the significand and build it into a floating-point number with 3346 // exponent of 1. 3347 SDValue X = GetSignificand(DAG, Op1, dl); 3348 3349 // Different possible minimax approximations of significand in 3350 // floating-point for various degrees of accuracy over [1,2]. 3351 if (LimitFloatPrecision <= 6) { 3352 // For floating-point precision of 6: 3353 // 3354 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x; 3355 // 3356 // error 0.0049451742, which is more than 7 bits 3357 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3358 getF32Constant(DAG, 0xbeb08fe0)); 3359 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3360 getF32Constant(DAG, 0x40019463)); 3361 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3362 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3363 getF32Constant(DAG, 0x3fd6633d)); 3364 3365 result = DAG.getNode(ISD::FADD, dl, 3366 MVT::f32, LogOfExponent, Log2ofMantissa); 3367 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) { 3368 // For floating-point precision of 12: 3369 // 3370 // Log2ofMantissa = 3371 // -2.51285454f + 3372 // (4.07009056f + 3373 // (-2.12067489f + 3374 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x; 3375 // 3376 // error 0.0000876136000, which is better than 13 bits 3377 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3378 getF32Constant(DAG, 0xbda7262e)); 3379 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3380 getF32Constant(DAG, 0x3f25280b)); 3381 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3382 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3383 getF32Constant(DAG, 0x4007b923)); 3384 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3385 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3386 getF32Constant(DAG, 0x40823e2f)); 3387 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3388 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, 3389 getF32Constant(DAG, 0x4020d29c)); 3390 3391 result = DAG.getNode(ISD::FADD, dl, 3392 MVT::f32, LogOfExponent, Log2ofMantissa); 3393 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18 3394 // For floating-point precision of 18: 3395 // 3396 // Log2ofMantissa = 3397 // -3.0400495f + 3398 // (6.1129976f + 3399 // (-5.3420409f + 3400 // (3.2865683f + 3401 // (-1.2669343f + 3402 // (0.27515199f - 3403 // 0.25691327e-1f * x) * x) * x) * x) * x) * x; 3404 // 3405 // error 0.0000018516, which is better than 18 bits 3406 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3407 getF32Constant(DAG, 0xbcd2769e)); 3408 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3409 getF32Constant(DAG, 0x3e8ce0b9)); 3410 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3411 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3412 getF32Constant(DAG, 0x3fa22ae7)); 3413 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3414 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3415 getF32Constant(DAG, 0x40525723)); 3416 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3417 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, 3418 getF32Constant(DAG, 0x40aaf200)); 3419 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3420 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 3421 getF32Constant(DAG, 0x40c39dad)); 3422 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 3423 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10, 3424 getF32Constant(DAG, 0x4042902c)); 3425 3426 result = DAG.getNode(ISD::FADD, dl, 3427 MVT::f32, LogOfExponent, Log2ofMantissa); 3428 } 3429 } else { 3430 // No special expansion. 3431 result = DAG.getNode(ISD::FLOG2, dl, 3432 getValue(I.getArgOperand(0)).getValueType(), 3433 getValue(I.getArgOperand(0))); 3434 } 3435 3436 setValue(&I, result); 3437} 3438 3439/// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for 3440/// limited-precision mode. 3441void 3442SelectionDAGBuilder::visitLog10(const CallInst &I) { 3443 SDValue result; 3444 DebugLoc dl = getCurDebugLoc(); 3445 3446 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 && 3447 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3448 SDValue Op = getValue(I.getArgOperand(0)); 3449 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op); 3450 3451 // Scale the exponent by log10(2) [0.30102999f]. 3452 SDValue Exp = GetExponent(DAG, Op1, TLI, dl); 3453 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp, 3454 getF32Constant(DAG, 0x3e9a209a)); 3455 3456 // Get the significand and build it into a floating-point number with 3457 // exponent of 1. 3458 SDValue X = GetSignificand(DAG, Op1, dl); 3459 3460 if (LimitFloatPrecision <= 6) { 3461 // For floating-point precision of 6: 3462 // 3463 // Log10ofMantissa = 3464 // -0.50419619f + 3465 // (0.60948995f - 0.10380950f * x) * x; 3466 // 3467 // error 0.0014886165, which is 6 bits 3468 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3469 getF32Constant(DAG, 0xbdd49a13)); 3470 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3471 getF32Constant(DAG, 0x3f1c0789)); 3472 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3473 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3474 getF32Constant(DAG, 0x3f011300)); 3475 3476 result = DAG.getNode(ISD::FADD, dl, 3477 MVT::f32, LogOfExponent, Log10ofMantissa); 3478 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) { 3479 // For floating-point precision of 12: 3480 // 3481 // Log10ofMantissa = 3482 // -0.64831180f + 3483 // (0.91751397f + 3484 // (-0.31664806f + 0.47637168e-1f * x) * x) * x; 3485 // 3486 // error 0.00019228036, which is better than 12 bits 3487 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3488 getF32Constant(DAG, 0x3d431f31)); 3489 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, 3490 getF32Constant(DAG, 0x3ea21fb2)); 3491 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3492 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3493 getF32Constant(DAG, 0x3f6ae232)); 3494 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3495 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4, 3496 getF32Constant(DAG, 0x3f25f7c3)); 3497 3498 result = DAG.getNode(ISD::FADD, dl, 3499 MVT::f32, LogOfExponent, Log10ofMantissa); 3500 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18 3501 // For floating-point precision of 18: 3502 // 3503 // Log10ofMantissa = 3504 // -0.84299375f + 3505 // (1.5327582f + 3506 // (-1.0688956f + 3507 // (0.49102474f + 3508 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x; 3509 // 3510 // error 0.0000037995730, which is better than 18 bits 3511 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3512 getF32Constant(DAG, 0x3c5d51ce)); 3513 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, 3514 getF32Constant(DAG, 0x3e00685a)); 3515 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3516 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3517 getF32Constant(DAG, 0x3efb6798)); 3518 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3519 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4, 3520 getF32Constant(DAG, 0x3f88d192)); 3521 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3522 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3523 getF32Constant(DAG, 0x3fc4316c)); 3524 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3525 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8, 3526 getF32Constant(DAG, 0x3f57ce70)); 3527 3528 result = DAG.getNode(ISD::FADD, dl, 3529 MVT::f32, LogOfExponent, Log10ofMantissa); 3530 } 3531 } else { 3532 // No special expansion. 3533 result = DAG.getNode(ISD::FLOG10, dl, 3534 getValue(I.getArgOperand(0)).getValueType(), 3535 getValue(I.getArgOperand(0))); 3536 } 3537 3538 setValue(&I, result); 3539} 3540 3541/// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for 3542/// limited-precision mode. 3543void 3544SelectionDAGBuilder::visitExp2(const CallInst &I) { 3545 SDValue result; 3546 DebugLoc dl = getCurDebugLoc(); 3547 3548 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 && 3549 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3550 SDValue Op = getValue(I.getArgOperand(0)); 3551 3552 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op); 3553 3554 // FractionalPartOfX = x - (float)IntegerPartOfX; 3555 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); 3556 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1); 3557 3558 // IntegerPartOfX <<= 23; 3559 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, 3560 DAG.getConstant(23, TLI.getPointerTy())); 3561 3562 if (LimitFloatPrecision <= 6) { 3563 // For floating-point precision of 6: 3564 // 3565 // TwoToFractionalPartOfX = 3566 // 0.997535578f + 3567 // (0.735607626f + 0.252464424f * x) * x; 3568 // 3569 // error 0.0144103317, which is 6 bits 3570 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3571 getF32Constant(DAG, 0x3e814304)); 3572 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3573 getF32Constant(DAG, 0x3f3c50c8)); 3574 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3575 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3576 getF32Constant(DAG, 0x3f7f5e7e)); 3577 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5); 3578 SDValue TwoToFractionalPartOfX = 3579 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX); 3580 3581 result = DAG.getNode(ISD::BIT_CONVERT, dl, 3582 MVT::f32, TwoToFractionalPartOfX); 3583 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) { 3584 // For floating-point precision of 12: 3585 // 3586 // TwoToFractionalPartOfX = 3587 // 0.999892986f + 3588 // (0.696457318f + 3589 // (0.224338339f + 0.792043434e-1f * x) * x) * x; 3590 // 3591 // error 0.000107046256, which is 13 to 14 bits 3592 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3593 getF32Constant(DAG, 0x3da235e3)); 3594 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3595 getF32Constant(DAG, 0x3e65b8f3)); 3596 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3597 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3598 getF32Constant(DAG, 0x3f324b07)); 3599 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3600 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3601 getF32Constant(DAG, 0x3f7ff8fd)); 3602 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7); 3603 SDValue TwoToFractionalPartOfX = 3604 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX); 3605 3606 result = DAG.getNode(ISD::BIT_CONVERT, dl, 3607 MVT::f32, TwoToFractionalPartOfX); 3608 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18 3609 // For floating-point precision of 18: 3610 // 3611 // TwoToFractionalPartOfX = 3612 // 0.999999982f + 3613 // (0.693148872f + 3614 // (0.240227044f + 3615 // (0.554906021e-1f + 3616 // (0.961591928e-2f + 3617 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; 3618 // error 2.47208000*10^(-7), which is better than 18 bits 3619 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3620 getF32Constant(DAG, 0x3924b03e)); 3621 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3622 getF32Constant(DAG, 0x3ab24b87)); 3623 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3624 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3625 getF32Constant(DAG, 0x3c1d8c17)); 3626 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3627 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3628 getF32Constant(DAG, 0x3d634a1d)); 3629 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3630 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 3631 getF32Constant(DAG, 0x3e75fe14)); 3632 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 3633 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, 3634 getF32Constant(DAG, 0x3f317234)); 3635 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); 3636 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, 3637 getF32Constant(DAG, 0x3f800000)); 3638 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13); 3639 SDValue TwoToFractionalPartOfX = 3640 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX); 3641 3642 result = DAG.getNode(ISD::BIT_CONVERT, dl, 3643 MVT::f32, TwoToFractionalPartOfX); 3644 } 3645 } else { 3646 // No special expansion. 3647 result = DAG.getNode(ISD::FEXP2, dl, 3648 getValue(I.getArgOperand(0)).getValueType(), 3649 getValue(I.getArgOperand(0))); 3650 } 3651 3652 setValue(&I, result); 3653} 3654 3655/// visitPow - Lower a pow intrinsic. Handles the special sequences for 3656/// limited-precision mode with x == 10.0f. 3657void 3658SelectionDAGBuilder::visitPow(const CallInst &I) { 3659 SDValue result; 3660 const Value *Val = I.getArgOperand(0); 3661 DebugLoc dl = getCurDebugLoc(); 3662 bool IsExp10 = false; 3663 3664 if (getValue(Val).getValueType() == MVT::f32 && 3665 getValue(I.getArgOperand(1)).getValueType() == MVT::f32 && 3666 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3667 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) { 3668 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 3669 APFloat Ten(10.0f); 3670 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten); 3671 } 3672 } 3673 } 3674 3675 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3676 SDValue Op = getValue(I.getArgOperand(1)); 3677 3678 // Put the exponent in the right bit position for later addition to the 3679 // final result: 3680 // 3681 // #define LOG2OF10 3.3219281f 3682 // IntegerPartOfX = (int32_t)(x * LOG2OF10); 3683 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op, 3684 getF32Constant(DAG, 0x40549a78)); 3685 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0); 3686 3687 // FractionalPartOfX = x - (float)IntegerPartOfX; 3688 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); 3689 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1); 3690 3691 // IntegerPartOfX <<= 23; 3692 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, 3693 DAG.getConstant(23, TLI.getPointerTy())); 3694 3695 if (LimitFloatPrecision <= 6) { 3696 // For floating-point precision of 6: 3697 // 3698 // twoToFractionalPartOfX = 3699 // 0.997535578f + 3700 // (0.735607626f + 0.252464424f * x) * x; 3701 // 3702 // error 0.0144103317, which is 6 bits 3703 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3704 getF32Constant(DAG, 0x3e814304)); 3705 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3706 getF32Constant(DAG, 0x3f3c50c8)); 3707 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3708 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3709 getF32Constant(DAG, 0x3f7f5e7e)); 3710 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5); 3711 SDValue TwoToFractionalPartOfX = 3712 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX); 3713 3714 result = DAG.getNode(ISD::BIT_CONVERT, dl, 3715 MVT::f32, TwoToFractionalPartOfX); 3716 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) { 3717 // For floating-point precision of 12: 3718 // 3719 // TwoToFractionalPartOfX = 3720 // 0.999892986f + 3721 // (0.696457318f + 3722 // (0.224338339f + 0.792043434e-1f * x) * x) * x; 3723 // 3724 // error 0.000107046256, which is 13 to 14 bits 3725 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3726 getF32Constant(DAG, 0x3da235e3)); 3727 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3728 getF32Constant(DAG, 0x3e65b8f3)); 3729 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3730 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3731 getF32Constant(DAG, 0x3f324b07)); 3732 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3733 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3734 getF32Constant(DAG, 0x3f7ff8fd)); 3735 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7); 3736 SDValue TwoToFractionalPartOfX = 3737 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX); 3738 3739 result = DAG.getNode(ISD::BIT_CONVERT, dl, 3740 MVT::f32, TwoToFractionalPartOfX); 3741 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18 3742 // For floating-point precision of 18: 3743 // 3744 // TwoToFractionalPartOfX = 3745 // 0.999999982f + 3746 // (0.693148872f + 3747 // (0.240227044f + 3748 // (0.554906021e-1f + 3749 // (0.961591928e-2f + 3750 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; 3751 // error 2.47208000*10^(-7), which is better than 18 bits 3752 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3753 getF32Constant(DAG, 0x3924b03e)); 3754 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3755 getF32Constant(DAG, 0x3ab24b87)); 3756 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3757 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3758 getF32Constant(DAG, 0x3c1d8c17)); 3759 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3760 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3761 getF32Constant(DAG, 0x3d634a1d)); 3762 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3763 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 3764 getF32Constant(DAG, 0x3e75fe14)); 3765 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 3766 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, 3767 getF32Constant(DAG, 0x3f317234)); 3768 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); 3769 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, 3770 getF32Constant(DAG, 0x3f800000)); 3771 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13); 3772 SDValue TwoToFractionalPartOfX = 3773 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX); 3774 3775 result = DAG.getNode(ISD::BIT_CONVERT, dl, 3776 MVT::f32, TwoToFractionalPartOfX); 3777 } 3778 } else { 3779 // No special expansion. 3780 result = DAG.getNode(ISD::FPOW, dl, 3781 getValue(I.getArgOperand(0)).getValueType(), 3782 getValue(I.getArgOperand(0)), 3783 getValue(I.getArgOperand(1))); 3784 } 3785 3786 setValue(&I, result); 3787} 3788 3789 3790/// ExpandPowI - Expand a llvm.powi intrinsic. 3791static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS, 3792 SelectionDAG &DAG) { 3793 // If RHS is a constant, we can expand this out to a multiplication tree, 3794 // otherwise we end up lowering to a call to __powidf2 (for example). When 3795 // optimizing for size, we only want to do this if the expansion would produce 3796 // a small number of multiplies, otherwise we do the full expansion. 3797 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) { 3798 // Get the exponent as a positive value. 3799 unsigned Val = RHSC->getSExtValue(); 3800 if ((int)Val < 0) Val = -Val; 3801 3802 // powi(x, 0) -> 1.0 3803 if (Val == 0) 3804 return DAG.getConstantFP(1.0, LHS.getValueType()); 3805 3806 const Function *F = DAG.getMachineFunction().getFunction(); 3807 if (!F->hasFnAttr(Attribute::OptimizeForSize) || 3808 // If optimizing for size, don't insert too many multiplies. This 3809 // inserts up to 5 multiplies. 3810 CountPopulation_32(Val)+Log2_32(Val) < 7) { 3811 // We use the simple binary decomposition method to generate the multiply 3812 // sequence. There are more optimal ways to do this (for example, 3813 // powi(x,15) generates one more multiply than it should), but this has 3814 // the benefit of being both really simple and much better than a libcall. 3815 SDValue Res; // Logically starts equal to 1.0 3816 SDValue CurSquare = LHS; 3817 while (Val) { 3818 if (Val & 1) { 3819 if (Res.getNode()) 3820 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare); 3821 else 3822 Res = CurSquare; // 1.0*CurSquare. 3823 } 3824 3825 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(), 3826 CurSquare, CurSquare); 3827 Val >>= 1; 3828 } 3829 3830 // If the original was negative, invert the result, producing 1/(x*x*x). 3831 if (RHSC->getSExtValue() < 0) 3832 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(), 3833 DAG.getConstantFP(1.0, LHS.getValueType()), Res); 3834 return Res; 3835 } 3836 } 3837 3838 // Otherwise, expand to a libcall. 3839 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS); 3840} 3841 3842/// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function 3843/// argument, create the corresponding DBG_VALUE machine instruction for it now. 3844/// At the end of instruction selection, they will be inserted to the entry BB. 3845bool 3846SelectionDAGBuilder::EmitFuncArgumentDbgValue(const DbgValueInst &DI, 3847 const Value *V, MDNode *Variable, 3848 uint64_t Offset, 3849 const SDValue &N) { 3850 if (!isa<Argument>(V)) 3851 return false; 3852 3853 MachineFunction &MF = DAG.getMachineFunction(); 3854 // Ignore inlined function arguments here. 3855 DIVariable DV(Variable); 3856 if (DV.isInlinedFnArgument(MF.getFunction())) 3857 return false; 3858 3859 MachineBasicBlock *MBB = FuncInfo.MBBMap[DI.getParent()]; 3860 if (MBB != &MF.front()) 3861 return false; 3862 3863 unsigned Reg = 0; 3864 if (N.getOpcode() == ISD::CopyFromReg) { 3865 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg(); 3866 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) { 3867 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 3868 unsigned PR = RegInfo.getLiveInPhysReg(Reg); 3869 if (PR) 3870 Reg = PR; 3871 } 3872 } 3873 3874 if (!Reg) { 3875 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V); 3876 if (VMI == FuncInfo.ValueMap.end()) 3877 return false; 3878 Reg = VMI->second; 3879 } 3880 3881 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo(); 3882 MachineInstrBuilder MIB = BuildMI(MF, getCurDebugLoc(), 3883 TII->get(TargetOpcode::DBG_VALUE)) 3884 .addReg(Reg, RegState::Debug).addImm(Offset).addMetadata(Variable); 3885 FuncInfo.ArgDbgValues.push_back(&*MIB); 3886 return true; 3887} 3888 3889// VisualStudio defines setjmp as _setjmp 3890#if defined(_MSC_VER) && defined(setjmp) 3891#define setjmp_undefined_for_visual_studio 3892#undef setjmp 3893#endif 3894 3895/// visitIntrinsicCall - Lower the call to the specified intrinsic function. If 3896/// we want to emit this as a call to a named external function, return the name 3897/// otherwise lower it and return null. 3898const char * 3899SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) { 3900 DebugLoc dl = getCurDebugLoc(); 3901 SDValue Res; 3902 3903 switch (Intrinsic) { 3904 default: 3905 // By default, turn this into a target intrinsic node. 3906 visitTargetIntrinsic(I, Intrinsic); 3907 return 0; 3908 case Intrinsic::vastart: visitVAStart(I); return 0; 3909 case Intrinsic::vaend: visitVAEnd(I); return 0; 3910 case Intrinsic::vacopy: visitVACopy(I); return 0; 3911 case Intrinsic::returnaddress: 3912 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(), 3913 getValue(I.getArgOperand(0)))); 3914 return 0; 3915 case Intrinsic::frameaddress: 3916 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(), 3917 getValue(I.getArgOperand(0)))); 3918 return 0; 3919 case Intrinsic::setjmp: 3920 return "_setjmp"+!TLI.usesUnderscoreSetJmp(); 3921 case Intrinsic::longjmp: 3922 return "_longjmp"+!TLI.usesUnderscoreLongJmp(); 3923 case Intrinsic::memcpy: { 3924 // Assert for address < 256 since we support only user defined address 3925 // spaces. 3926 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() 3927 < 256 && 3928 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() 3929 < 256 && 3930 "Unknown address space"); 3931 SDValue Op1 = getValue(I.getArgOperand(0)); 3932 SDValue Op2 = getValue(I.getArgOperand(1)); 3933 SDValue Op3 = getValue(I.getArgOperand(2)); 3934 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); 3935 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); 3936 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, false, 3937 I.getArgOperand(0), 0, I.getArgOperand(1), 0)); 3938 return 0; 3939 } 3940 case Intrinsic::memset: { 3941 // Assert for address < 256 since we support only user defined address 3942 // spaces. 3943 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() 3944 < 256 && 3945 "Unknown address space"); 3946 SDValue Op1 = getValue(I.getArgOperand(0)); 3947 SDValue Op2 = getValue(I.getArgOperand(1)); 3948 SDValue Op3 = getValue(I.getArgOperand(2)); 3949 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); 3950 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); 3951 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align, isVol, 3952 I.getArgOperand(0), 0)); 3953 return 0; 3954 } 3955 case Intrinsic::memmove: { 3956 // Assert for address < 256 since we support only user defined address 3957 // spaces. 3958 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() 3959 < 256 && 3960 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() 3961 < 256 && 3962 "Unknown address space"); 3963 SDValue Op1 = getValue(I.getArgOperand(0)); 3964 SDValue Op2 = getValue(I.getArgOperand(1)); 3965 SDValue Op3 = getValue(I.getArgOperand(2)); 3966 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); 3967 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); 3968 3969 // If the source and destination are known to not be aliases, we can 3970 // lower memmove as memcpy. 3971 uint64_t Size = -1ULL; 3972 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3)) 3973 Size = C->getZExtValue(); 3974 if (AA->alias(I.getArgOperand(0), Size, I.getArgOperand(1), Size) == 3975 AliasAnalysis::NoAlias) { 3976 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, 3977 false, I.getArgOperand(0), 0, I.getArgOperand(1), 0)); 3978 return 0; 3979 } 3980 3981 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align, isVol, 3982 I.getArgOperand(0), 0, I.getArgOperand(1), 0)); 3983 return 0; 3984 } 3985 case Intrinsic::dbg_declare: { 3986 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I); 3987 if (!DIVariable(DI.getVariable()).Verify()) 3988 return 0; 3989 3990 MDNode *Variable = DI.getVariable(); 3991 // Parameters are handled specially. 3992 bool isParameter = 3993 DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable; 3994 const Value *Address = DI.getAddress(); 3995 if (!Address) 3996 return 0; 3997 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address)) 3998 Address = BCI->getOperand(0); 3999 const AllocaInst *AI = dyn_cast<AllocaInst>(Address); 4000 if (AI) { 4001 // Don't handle byval arguments or VLAs, for example. 4002 // Non-byval arguments are handled here (they refer to the stack temporary 4003 // alloca at this point). 4004 DenseMap<const AllocaInst*, int>::iterator SI = 4005 FuncInfo.StaticAllocaMap.find(AI); 4006 if (SI == FuncInfo.StaticAllocaMap.end()) 4007 return 0; // VLAs. 4008 int FI = SI->second; 4009 4010 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 4011 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo()) 4012 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc()); 4013 } 4014 4015 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder 4016 // but do not always have a corresponding SDNode built. The SDNodeOrder 4017 // absolute, but not relative, values are different depending on whether 4018 // debug info exists. 4019 ++SDNodeOrder; 4020 SDValue &N = NodeMap[Address]; 4021 SDDbgValue *SDV; 4022 if (N.getNode()) { 4023 if (isParameter && !AI) { 4024 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode()); 4025 if (FINode) 4026 // Byval parameter. We have a frame index at this point. 4027 SDV = DAG.getDbgValue(Variable, FINode->getIndex(), 4028 0, dl, SDNodeOrder); 4029 else 4030 // Can't do anything with other non-AI cases yet. This might be a 4031 // parameter of a callee function that got inlined, for example. 4032 return 0; 4033 } else if (AI) 4034 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(), 4035 0, dl, SDNodeOrder); 4036 else 4037 // Can't do anything with other non-AI cases yet. 4038 return 0; 4039 DAG.AddDbgValue(SDV, N.getNode(), isParameter); 4040 } else { 4041 // This isn't useful, but it shows what we're missing. 4042 SDV = DAG.getDbgValue(Variable, UndefValue::get(Address->getType()), 4043 0, dl, SDNodeOrder); 4044 DAG.AddDbgValue(SDV, 0, isParameter); 4045 } 4046 return 0; 4047 } 4048 case Intrinsic::dbg_value: { 4049 const DbgValueInst &DI = cast<DbgValueInst>(I); 4050 if (!DIVariable(DI.getVariable()).Verify()) 4051 return 0; 4052 4053 MDNode *Variable = DI.getVariable(); 4054 uint64_t Offset = DI.getOffset(); 4055 const Value *V = DI.getValue(); 4056 if (!V) 4057 return 0; 4058 4059 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder 4060 // but do not always have a corresponding SDNode built. The SDNodeOrder 4061 // absolute, but not relative, values are different depending on whether 4062 // debug info exists. 4063 ++SDNodeOrder; 4064 SDDbgValue *SDV; 4065 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) { 4066 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder); 4067 DAG.AddDbgValue(SDV, 0, false); 4068 } else { 4069 bool createUndef = false; 4070 // FIXME : Why not use getValue() directly ? 4071 SDValue N = NodeMap[V]; 4072 if (!N.getNode() && isa<Argument>(V)) 4073 // Check unused arguments map. 4074 N = UnusedArgNodeMap[V]; 4075 if (N.getNode()) { 4076 if (!EmitFuncArgumentDbgValue(DI, V, Variable, Offset, N)) { 4077 SDV = DAG.getDbgValue(Variable, N.getNode(), 4078 N.getResNo(), Offset, dl, SDNodeOrder); 4079 DAG.AddDbgValue(SDV, N.getNode(), false); 4080 } 4081 } else if (isa<PHINode>(V) && !V->use_empty()) { 4082 SDValue N = getValue(V); 4083 if (N.getNode()) { 4084 if (!EmitFuncArgumentDbgValue(DI, V, Variable, Offset, N)) { 4085 SDV = DAG.getDbgValue(Variable, N.getNode(), 4086 N.getResNo(), Offset, dl, SDNodeOrder); 4087 DAG.AddDbgValue(SDV, N.getNode(), false); 4088 } 4089 } else 4090 createUndef = true; 4091 } else 4092 createUndef = true; 4093 if (createUndef) { 4094 // We may expand this to cover more cases. One case where we have no 4095 // data available is an unreferenced parameter; we need this fallback. 4096 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()), 4097 Offset, dl, SDNodeOrder); 4098 DAG.AddDbgValue(SDV, 0, false); 4099 } 4100 } 4101 4102 // Build a debug info table entry. 4103 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V)) 4104 V = BCI->getOperand(0); 4105 const AllocaInst *AI = dyn_cast<AllocaInst>(V); 4106 // Don't handle byval struct arguments or VLAs, for example. 4107 if (!AI) 4108 return 0; 4109 DenseMap<const AllocaInst*, int>::iterator SI = 4110 FuncInfo.StaticAllocaMap.find(AI); 4111 if (SI == FuncInfo.StaticAllocaMap.end()) 4112 return 0; // VLAs. 4113 int FI = SI->second; 4114 4115 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 4116 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo()) 4117 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc()); 4118 return 0; 4119 } 4120 case Intrinsic::eh_exception: { 4121 // Insert the EXCEPTIONADDR instruction. 4122 assert(FuncInfo.MBBMap[I.getParent()]->isLandingPad() && 4123 "Call to eh.exception not in landing pad!"); 4124 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other); 4125 SDValue Ops[1]; 4126 Ops[0] = DAG.getRoot(); 4127 SDValue Op = DAG.getNode(ISD::EXCEPTIONADDR, dl, VTs, Ops, 1); 4128 setValue(&I, Op); 4129 DAG.setRoot(Op.getValue(1)); 4130 return 0; 4131 } 4132 4133 case Intrinsic::eh_selector: { 4134 MachineBasicBlock *CallMBB = FuncInfo.MBBMap[I.getParent()]; 4135 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 4136 if (CallMBB->isLandingPad()) 4137 AddCatchInfo(I, &MMI, CallMBB); 4138 else { 4139#ifndef NDEBUG 4140 FuncInfo.CatchInfoLost.insert(&I); 4141#endif 4142 // FIXME: Mark exception selector register as live in. Hack for PR1508. 4143 unsigned Reg = TLI.getExceptionSelectorRegister(); 4144 if (Reg) FuncInfo.MBBMap[I.getParent()]->addLiveIn(Reg); 4145 } 4146 4147 // Insert the EHSELECTION instruction. 4148 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other); 4149 SDValue Ops[2]; 4150 Ops[0] = getValue(I.getArgOperand(0)); 4151 Ops[1] = getRoot(); 4152 SDValue Op = DAG.getNode(ISD::EHSELECTION, dl, VTs, Ops, 2); 4153 DAG.setRoot(Op.getValue(1)); 4154 setValue(&I, DAG.getSExtOrTrunc(Op, dl, MVT::i32)); 4155 return 0; 4156 } 4157 4158 case Intrinsic::eh_typeid_for: { 4159 // Find the type id for the given typeinfo. 4160 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0)); 4161 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV); 4162 Res = DAG.getConstant(TypeID, MVT::i32); 4163 setValue(&I, Res); 4164 return 0; 4165 } 4166 4167 case Intrinsic::eh_return_i32: 4168 case Intrinsic::eh_return_i64: 4169 DAG.getMachineFunction().getMMI().setCallsEHReturn(true); 4170 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl, 4171 MVT::Other, 4172 getControlRoot(), 4173 getValue(I.getArgOperand(0)), 4174 getValue(I.getArgOperand(1)))); 4175 return 0; 4176 case Intrinsic::eh_unwind_init: 4177 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true); 4178 return 0; 4179 case Intrinsic::eh_dwarf_cfa: { 4180 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), dl, 4181 TLI.getPointerTy()); 4182 SDValue Offset = DAG.getNode(ISD::ADD, dl, 4183 TLI.getPointerTy(), 4184 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl, 4185 TLI.getPointerTy()), 4186 CfaArg); 4187 SDValue FA = DAG.getNode(ISD::FRAMEADDR, dl, 4188 TLI.getPointerTy(), 4189 DAG.getConstant(0, TLI.getPointerTy())); 4190 setValue(&I, DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), 4191 FA, Offset)); 4192 return 0; 4193 } 4194 case Intrinsic::eh_sjlj_callsite: { 4195 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 4196 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0)); 4197 assert(CI && "Non-constant call site value in eh.sjlj.callsite!"); 4198 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!"); 4199 4200 MMI.setCurrentCallSite(CI->getZExtValue()); 4201 return 0; 4202 } 4203 case Intrinsic::eh_sjlj_setjmp: { 4204 setValue(&I, DAG.getNode(ISD::EH_SJLJ_SETJMP, dl, MVT::i32, getRoot(), 4205 getValue(I.getArgOperand(0)))); 4206 return 0; 4207 } 4208 case Intrinsic::eh_sjlj_longjmp: { 4209 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, dl, MVT::Other, 4210 getRoot(), 4211 getValue(I.getArgOperand(0)))); 4212 return 0; 4213 } 4214 4215 case Intrinsic::convertff: 4216 case Intrinsic::convertfsi: 4217 case Intrinsic::convertfui: 4218 case Intrinsic::convertsif: 4219 case Intrinsic::convertuif: 4220 case Intrinsic::convertss: 4221 case Intrinsic::convertsu: 4222 case Intrinsic::convertus: 4223 case Intrinsic::convertuu: { 4224 ISD::CvtCode Code = ISD::CVT_INVALID; 4225 switch (Intrinsic) { 4226 case Intrinsic::convertff: Code = ISD::CVT_FF; break; 4227 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break; 4228 case Intrinsic::convertfui: Code = ISD::CVT_FU; break; 4229 case Intrinsic::convertsif: Code = ISD::CVT_SF; break; 4230 case Intrinsic::convertuif: Code = ISD::CVT_UF; break; 4231 case Intrinsic::convertss: Code = ISD::CVT_SS; break; 4232 case Intrinsic::convertsu: Code = ISD::CVT_SU; break; 4233 case Intrinsic::convertus: Code = ISD::CVT_US; break; 4234 case Intrinsic::convertuu: Code = ISD::CVT_UU; break; 4235 } 4236 EVT DestVT = TLI.getValueType(I.getType()); 4237 const Value *Op1 = I.getArgOperand(0); 4238 Res = DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1), 4239 DAG.getValueType(DestVT), 4240 DAG.getValueType(getValue(Op1).getValueType()), 4241 getValue(I.getArgOperand(1)), 4242 getValue(I.getArgOperand(2)), 4243 Code); 4244 setValue(&I, Res); 4245 return 0; 4246 } 4247 case Intrinsic::sqrt: 4248 setValue(&I, DAG.getNode(ISD::FSQRT, dl, 4249 getValue(I.getArgOperand(0)).getValueType(), 4250 getValue(I.getArgOperand(0)))); 4251 return 0; 4252 case Intrinsic::powi: 4253 setValue(&I, ExpandPowI(dl, getValue(I.getArgOperand(0)), 4254 getValue(I.getArgOperand(1)), DAG)); 4255 return 0; 4256 case Intrinsic::sin: 4257 setValue(&I, DAG.getNode(ISD::FSIN, dl, 4258 getValue(I.getArgOperand(0)).getValueType(), 4259 getValue(I.getArgOperand(0)))); 4260 return 0; 4261 case Intrinsic::cos: 4262 setValue(&I, DAG.getNode(ISD::FCOS, dl, 4263 getValue(I.getArgOperand(0)).getValueType(), 4264 getValue(I.getArgOperand(0)))); 4265 return 0; 4266 case Intrinsic::log: 4267 visitLog(I); 4268 return 0; 4269 case Intrinsic::log2: 4270 visitLog2(I); 4271 return 0; 4272 case Intrinsic::log10: 4273 visitLog10(I); 4274 return 0; 4275 case Intrinsic::exp: 4276 visitExp(I); 4277 return 0; 4278 case Intrinsic::exp2: 4279 visitExp2(I); 4280 return 0; 4281 case Intrinsic::pow: 4282 visitPow(I); 4283 return 0; 4284 case Intrinsic::convert_to_fp16: 4285 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, dl, 4286 MVT::i16, getValue(I.getArgOperand(0)))); 4287 return 0; 4288 case Intrinsic::convert_from_fp16: 4289 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, dl, 4290 MVT::f32, getValue(I.getArgOperand(0)))); 4291 return 0; 4292 case Intrinsic::pcmarker: { 4293 SDValue Tmp = getValue(I.getArgOperand(0)); 4294 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp)); 4295 return 0; 4296 } 4297 case Intrinsic::readcyclecounter: { 4298 SDValue Op = getRoot(); 4299 Res = DAG.getNode(ISD::READCYCLECOUNTER, dl, 4300 DAG.getVTList(MVT::i64, MVT::Other), 4301 &Op, 1); 4302 setValue(&I, Res); 4303 DAG.setRoot(Res.getValue(1)); 4304 return 0; 4305 } 4306 case Intrinsic::bswap: 4307 setValue(&I, DAG.getNode(ISD::BSWAP, dl, 4308 getValue(I.getArgOperand(0)).getValueType(), 4309 getValue(I.getArgOperand(0)))); 4310 return 0; 4311 case Intrinsic::cttz: { 4312 SDValue Arg = getValue(I.getArgOperand(0)); 4313 EVT Ty = Arg.getValueType(); 4314 setValue(&I, DAG.getNode(ISD::CTTZ, dl, Ty, Arg)); 4315 return 0; 4316 } 4317 case Intrinsic::ctlz: { 4318 SDValue Arg = getValue(I.getArgOperand(0)); 4319 EVT Ty = Arg.getValueType(); 4320 setValue(&I, DAG.getNode(ISD::CTLZ, dl, Ty, Arg)); 4321 return 0; 4322 } 4323 case Intrinsic::ctpop: { 4324 SDValue Arg = getValue(I.getArgOperand(0)); 4325 EVT Ty = Arg.getValueType(); 4326 setValue(&I, DAG.getNode(ISD::CTPOP, dl, Ty, Arg)); 4327 return 0; 4328 } 4329 case Intrinsic::stacksave: { 4330 SDValue Op = getRoot(); 4331 Res = DAG.getNode(ISD::STACKSAVE, dl, 4332 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1); 4333 setValue(&I, Res); 4334 DAG.setRoot(Res.getValue(1)); 4335 return 0; 4336 } 4337 case Intrinsic::stackrestore: { 4338 Res = getValue(I.getArgOperand(0)); 4339 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Res)); 4340 return 0; 4341 } 4342 case Intrinsic::stackprotector: { 4343 // Emit code into the DAG to store the stack guard onto the stack. 4344 MachineFunction &MF = DAG.getMachineFunction(); 4345 MachineFrameInfo *MFI = MF.getFrameInfo(); 4346 EVT PtrTy = TLI.getPointerTy(); 4347 4348 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value. 4349 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1)); 4350 4351 int FI = FuncInfo.StaticAllocaMap[Slot]; 4352 MFI->setStackProtectorIndex(FI); 4353 4354 SDValue FIN = DAG.getFrameIndex(FI, PtrTy); 4355 4356 // Store the stack protector onto the stack. 4357 Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN, 4358 PseudoSourceValue::getFixedStack(FI), 4359 0, true, false, 0); 4360 setValue(&I, Res); 4361 DAG.setRoot(Res); 4362 return 0; 4363 } 4364 case Intrinsic::objectsize: { 4365 // If we don't know by now, we're never going to know. 4366 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1)); 4367 4368 assert(CI && "Non-constant type in __builtin_object_size?"); 4369 4370 SDValue Arg = getValue(I.getCalledValue()); 4371 EVT Ty = Arg.getValueType(); 4372 4373 if (CI->isZero()) 4374 Res = DAG.getConstant(-1ULL, Ty); 4375 else 4376 Res = DAG.getConstant(0, Ty); 4377 4378 setValue(&I, Res); 4379 return 0; 4380 } 4381 case Intrinsic::var_annotation: 4382 // Discard annotate attributes 4383 return 0; 4384 4385 case Intrinsic::init_trampoline: { 4386 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts()); 4387 4388 SDValue Ops[6]; 4389 Ops[0] = getRoot(); 4390 Ops[1] = getValue(I.getArgOperand(0)); 4391 Ops[2] = getValue(I.getArgOperand(1)); 4392 Ops[3] = getValue(I.getArgOperand(2)); 4393 Ops[4] = DAG.getSrcValue(I.getArgOperand(0)); 4394 Ops[5] = DAG.getSrcValue(F); 4395 4396 Res = DAG.getNode(ISD::TRAMPOLINE, dl, 4397 DAG.getVTList(TLI.getPointerTy(), MVT::Other), 4398 Ops, 6); 4399 4400 setValue(&I, Res); 4401 DAG.setRoot(Res.getValue(1)); 4402 return 0; 4403 } 4404 case Intrinsic::gcroot: 4405 if (GFI) { 4406 const Value *Alloca = I.getArgOperand(0); 4407 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1)); 4408 4409 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode()); 4410 GFI->addStackRoot(FI->getIndex(), TypeMap); 4411 } 4412 return 0; 4413 case Intrinsic::gcread: 4414 case Intrinsic::gcwrite: 4415 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!"); 4416 return 0; 4417 case Intrinsic::flt_rounds: 4418 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32)); 4419 return 0; 4420 case Intrinsic::trap: 4421 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot())); 4422 return 0; 4423 case Intrinsic::uadd_with_overflow: 4424 return implVisitAluOverflow(I, ISD::UADDO); 4425 case Intrinsic::sadd_with_overflow: 4426 return implVisitAluOverflow(I, ISD::SADDO); 4427 case Intrinsic::usub_with_overflow: 4428 return implVisitAluOverflow(I, ISD::USUBO); 4429 case Intrinsic::ssub_with_overflow: 4430 return implVisitAluOverflow(I, ISD::SSUBO); 4431 case Intrinsic::umul_with_overflow: 4432 return implVisitAluOverflow(I, ISD::UMULO); 4433 case Intrinsic::smul_with_overflow: 4434 return implVisitAluOverflow(I, ISD::SMULO); 4435 4436 case Intrinsic::prefetch: { 4437 SDValue Ops[4]; 4438 Ops[0] = getRoot(); 4439 Ops[1] = getValue(I.getArgOperand(0)); 4440 Ops[2] = getValue(I.getArgOperand(1)); 4441 Ops[3] = getValue(I.getArgOperand(2)); 4442 DAG.setRoot(DAG.getNode(ISD::PREFETCH, dl, MVT::Other, &Ops[0], 4)); 4443 return 0; 4444 } 4445 4446 case Intrinsic::memory_barrier: { 4447 SDValue Ops[6]; 4448 Ops[0] = getRoot(); 4449 for (int x = 1; x < 6; ++x) 4450 Ops[x] = getValue(I.getArgOperand(x - 1)); 4451 4452 DAG.setRoot(DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, &Ops[0], 6)); 4453 return 0; 4454 } 4455 case Intrinsic::atomic_cmp_swap: { 4456 SDValue Root = getRoot(); 4457 SDValue L = 4458 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, getCurDebugLoc(), 4459 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(), 4460 Root, 4461 getValue(I.getArgOperand(0)), 4462 getValue(I.getArgOperand(1)), 4463 getValue(I.getArgOperand(2)), 4464 I.getArgOperand(0)); 4465 setValue(&I, L); 4466 DAG.setRoot(L.getValue(1)); 4467 return 0; 4468 } 4469 case Intrinsic::atomic_load_add: 4470 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD); 4471 case Intrinsic::atomic_load_sub: 4472 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB); 4473 case Intrinsic::atomic_load_or: 4474 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR); 4475 case Intrinsic::atomic_load_xor: 4476 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR); 4477 case Intrinsic::atomic_load_and: 4478 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND); 4479 case Intrinsic::atomic_load_nand: 4480 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND); 4481 case Intrinsic::atomic_load_max: 4482 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX); 4483 case Intrinsic::atomic_load_min: 4484 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN); 4485 case Intrinsic::atomic_load_umin: 4486 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN); 4487 case Intrinsic::atomic_load_umax: 4488 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX); 4489 case Intrinsic::atomic_swap: 4490 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP); 4491 4492 case Intrinsic::invariant_start: 4493 case Intrinsic::lifetime_start: 4494 // Discard region information. 4495 setValue(&I, DAG.getUNDEF(TLI.getPointerTy())); 4496 return 0; 4497 case Intrinsic::invariant_end: 4498 case Intrinsic::lifetime_end: 4499 // Discard region information. 4500 return 0; 4501 } 4502} 4503 4504void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee, 4505 bool isTailCall, 4506 MachineBasicBlock *LandingPad) { 4507 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType()); 4508 const FunctionType *FTy = cast<FunctionType>(PT->getElementType()); 4509 const Type *RetTy = FTy->getReturnType(); 4510 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 4511 MCSymbol *BeginLabel = 0; 4512 4513 TargetLowering::ArgListTy Args; 4514 TargetLowering::ArgListEntry Entry; 4515 Args.reserve(CS.arg_size()); 4516 4517 // Check whether the function can return without sret-demotion. 4518 SmallVector<EVT, 4> OutVTs; 4519 SmallVector<ISD::ArgFlagsTy, 4> OutsFlags; 4520 SmallVector<uint64_t, 4> Offsets; 4521 getReturnInfo(RetTy, CS.getAttributes().getRetAttributes(), 4522 OutVTs, OutsFlags, TLI, &Offsets); 4523 4524 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(), 4525 FTy->isVarArg(), OutVTs, OutsFlags, DAG); 4526 4527 SDValue DemoteStackSlot; 4528 4529 if (!CanLowerReturn) { 4530 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize( 4531 FTy->getReturnType()); 4532 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment( 4533 FTy->getReturnType()); 4534 MachineFunction &MF = DAG.getMachineFunction(); 4535 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false); 4536 const Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType()); 4537 4538 DemoteStackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy()); 4539 Entry.Node = DemoteStackSlot; 4540 Entry.Ty = StackSlotPtrType; 4541 Entry.isSExt = false; 4542 Entry.isZExt = false; 4543 Entry.isInReg = false; 4544 Entry.isSRet = true; 4545 Entry.isNest = false; 4546 Entry.isByVal = false; 4547 Entry.Alignment = Align; 4548 Args.push_back(Entry); 4549 RetTy = Type::getVoidTy(FTy->getContext()); 4550 } 4551 4552 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end(); 4553 i != e; ++i) { 4554 SDValue ArgNode = getValue(*i); 4555 Entry.Node = ArgNode; Entry.Ty = (*i)->getType(); 4556 4557 unsigned attrInd = i - CS.arg_begin() + 1; 4558 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt); 4559 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt); 4560 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg); 4561 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet); 4562 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest); 4563 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal); 4564 Entry.Alignment = CS.getParamAlignment(attrInd); 4565 Args.push_back(Entry); 4566 } 4567 4568 if (LandingPad) { 4569 // Insert a label before the invoke call to mark the try range. This can be 4570 // used to detect deletion of the invoke via the MachineModuleInfo. 4571 BeginLabel = MMI.getContext().CreateTempSymbol(); 4572 4573 // For SjLj, keep track of which landing pads go with which invokes 4574 // so as to maintain the ordering of pads in the LSDA. 4575 unsigned CallSiteIndex = MMI.getCurrentCallSite(); 4576 if (CallSiteIndex) { 4577 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex); 4578 // Now that the call site is handled, stop tracking it. 4579 MMI.setCurrentCallSite(0); 4580 } 4581 4582 // Both PendingLoads and PendingExports must be flushed here; 4583 // this call might not return. 4584 (void)getRoot(); 4585 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getControlRoot(), BeginLabel)); 4586 } 4587 4588 // Check if target-independent constraints permit a tail call here. 4589 // Target-dependent constraints are checked within TLI.LowerCallTo. 4590 if (isTailCall && 4591 !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI)) 4592 isTailCall = false; 4593 4594 std::pair<SDValue,SDValue> Result = 4595 TLI.LowerCallTo(getRoot(), RetTy, 4596 CS.paramHasAttr(0, Attribute::SExt), 4597 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(), 4598 CS.paramHasAttr(0, Attribute::InReg), FTy->getNumParams(), 4599 CS.getCallingConv(), 4600 isTailCall, 4601 !CS.getInstruction()->use_empty(), 4602 Callee, Args, DAG, getCurDebugLoc()); 4603 assert((isTailCall || Result.second.getNode()) && 4604 "Non-null chain expected with non-tail call!"); 4605 assert((Result.second.getNode() || !Result.first.getNode()) && 4606 "Null value expected with tail call!"); 4607 if (Result.first.getNode()) { 4608 setValue(CS.getInstruction(), Result.first); 4609 } else if (!CanLowerReturn && Result.second.getNode()) { 4610 // The instruction result is the result of loading from the 4611 // hidden sret parameter. 4612 SmallVector<EVT, 1> PVTs; 4613 const Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType()); 4614 4615 ComputeValueVTs(TLI, PtrRetTy, PVTs); 4616 assert(PVTs.size() == 1 && "Pointers should fit in one register"); 4617 EVT PtrVT = PVTs[0]; 4618 unsigned NumValues = OutVTs.size(); 4619 SmallVector<SDValue, 4> Values(NumValues); 4620 SmallVector<SDValue, 4> Chains(NumValues); 4621 4622 for (unsigned i = 0; i < NumValues; ++i) { 4623 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, 4624 DemoteStackSlot, 4625 DAG.getConstant(Offsets[i], PtrVT)); 4626 SDValue L = DAG.getLoad(OutVTs[i], getCurDebugLoc(), Result.second, 4627 Add, NULL, Offsets[i], false, false, 1); 4628 Values[i] = L; 4629 Chains[i] = L.getValue(1); 4630 } 4631 4632 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), 4633 MVT::Other, &Chains[0], NumValues); 4634 PendingLoads.push_back(Chain); 4635 4636 // Collect the legal value parts into potentially illegal values 4637 // that correspond to the original function's return values. 4638 SmallVector<EVT, 4> RetTys; 4639 RetTy = FTy->getReturnType(); 4640 ComputeValueVTs(TLI, RetTy, RetTys); 4641 ISD::NodeType AssertOp = ISD::DELETED_NODE; 4642 SmallVector<SDValue, 4> ReturnValues; 4643 unsigned CurReg = 0; 4644 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) { 4645 EVT VT = RetTys[I]; 4646 EVT RegisterVT = TLI.getRegisterType(RetTy->getContext(), VT); 4647 unsigned NumRegs = TLI.getNumRegisters(RetTy->getContext(), VT); 4648 4649 SDValue ReturnValue = 4650 getCopyFromParts(DAG, getCurDebugLoc(), &Values[CurReg], NumRegs, 4651 RegisterVT, VT, AssertOp); 4652 ReturnValues.push_back(ReturnValue); 4653 CurReg += NumRegs; 4654 } 4655 4656 setValue(CS.getInstruction(), 4657 DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(), 4658 DAG.getVTList(&RetTys[0], RetTys.size()), 4659 &ReturnValues[0], ReturnValues.size())); 4660 4661 } 4662 4663 // As a special case, a null chain means that a tail call has been emitted and 4664 // the DAG root is already updated. 4665 if (Result.second.getNode()) 4666 DAG.setRoot(Result.second); 4667 else 4668 HasTailCall = true; 4669 4670 if (LandingPad) { 4671 // Insert a label at the end of the invoke call to mark the try range. This 4672 // can be used to detect deletion of the invoke via the MachineModuleInfo. 4673 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol(); 4674 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getRoot(), EndLabel)); 4675 4676 // Inform MachineModuleInfo of range. 4677 MMI.addInvoke(LandingPad, BeginLabel, EndLabel); 4678 } 4679} 4680 4681/// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the 4682/// value is equal or not-equal to zero. 4683static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) { 4684 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); 4685 UI != E; ++UI) { 4686 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI)) 4687 if (IC->isEquality()) 4688 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1))) 4689 if (C->isNullValue()) 4690 continue; 4691 // Unknown instruction. 4692 return false; 4693 } 4694 return true; 4695} 4696 4697static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT, 4698 const Type *LoadTy, 4699 SelectionDAGBuilder &Builder) { 4700 4701 // Check to see if this load can be trivially constant folded, e.g. if the 4702 // input is from a string literal. 4703 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) { 4704 // Cast pointer to the type we really want to load. 4705 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput), 4706 PointerType::getUnqual(LoadTy)); 4707 4708 if (const Constant *LoadCst = 4709 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput), 4710 Builder.TD)) 4711 return Builder.getValue(LoadCst); 4712 } 4713 4714 // Otherwise, we have to emit the load. If the pointer is to unfoldable but 4715 // still constant memory, the input chain can be the entry node. 4716 SDValue Root; 4717 bool ConstantMemory = false; 4718 4719 // Do not serialize (non-volatile) loads of constant memory with anything. 4720 if (Builder.AA->pointsToConstantMemory(PtrVal)) { 4721 Root = Builder.DAG.getEntryNode(); 4722 ConstantMemory = true; 4723 } else { 4724 // Do not serialize non-volatile loads against each other. 4725 Root = Builder.DAG.getRoot(); 4726 } 4727 4728 SDValue Ptr = Builder.getValue(PtrVal); 4729 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root, 4730 Ptr, PtrVal /*SrcValue*/, 0/*SVOffset*/, 4731 false /*volatile*/, 4732 false /*nontemporal*/, 1 /* align=1 */); 4733 4734 if (!ConstantMemory) 4735 Builder.PendingLoads.push_back(LoadVal.getValue(1)); 4736 return LoadVal; 4737} 4738 4739 4740/// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form. 4741/// If so, return true and lower it, otherwise return false and it will be 4742/// lowered like a normal call. 4743bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) { 4744 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t) 4745 if (I.getNumOperands() != 4) 4746 return false; 4747 4748 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1); 4749 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() || 4750 !I.getArgOperand(2)->getType()->isIntegerTy() || 4751 !I.getType()->isIntegerTy()) 4752 return false; 4753 4754 const ConstantInt *Size = dyn_cast<ConstantInt>(I.getArgOperand(2)); 4755 4756 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0 4757 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0 4758 if (Size && IsOnlyUsedInZeroEqualityComparison(&I)) { 4759 bool ActuallyDoIt = true; 4760 MVT LoadVT; 4761 const Type *LoadTy; 4762 switch (Size->getZExtValue()) { 4763 default: 4764 LoadVT = MVT::Other; 4765 LoadTy = 0; 4766 ActuallyDoIt = false; 4767 break; 4768 case 2: 4769 LoadVT = MVT::i16; 4770 LoadTy = Type::getInt16Ty(Size->getContext()); 4771 break; 4772 case 4: 4773 LoadVT = MVT::i32; 4774 LoadTy = Type::getInt32Ty(Size->getContext()); 4775 break; 4776 case 8: 4777 LoadVT = MVT::i64; 4778 LoadTy = Type::getInt64Ty(Size->getContext()); 4779 break; 4780 /* 4781 case 16: 4782 LoadVT = MVT::v4i32; 4783 LoadTy = Type::getInt32Ty(Size->getContext()); 4784 LoadTy = VectorType::get(LoadTy, 4); 4785 break; 4786 */ 4787 } 4788 4789 // This turns into unaligned loads. We only do this if the target natively 4790 // supports the MVT we'll be loading or if it is small enough (<= 4) that 4791 // we'll only produce a small number of byte loads. 4792 4793 // Require that we can find a legal MVT, and only do this if the target 4794 // supports unaligned loads of that type. Expanding into byte loads would 4795 // bloat the code. 4796 if (ActuallyDoIt && Size->getZExtValue() > 4) { 4797 // TODO: Handle 5 byte compare as 4-byte + 1 byte. 4798 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads. 4799 if (!TLI.isTypeLegal(LoadVT) ||!TLI.allowsUnalignedMemoryAccesses(LoadVT)) 4800 ActuallyDoIt = false; 4801 } 4802 4803 if (ActuallyDoIt) { 4804 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this); 4805 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this); 4806 4807 SDValue Res = DAG.getSetCC(getCurDebugLoc(), MVT::i1, LHSVal, RHSVal, 4808 ISD::SETNE); 4809 EVT CallVT = TLI.getValueType(I.getType(), true); 4810 setValue(&I, DAG.getZExtOrTrunc(Res, getCurDebugLoc(), CallVT)); 4811 return true; 4812 } 4813 } 4814 4815 4816 return false; 4817} 4818 4819 4820void SelectionDAGBuilder::visitCall(const CallInst &I) { 4821 const char *RenameFn = 0; 4822 if (Function *F = I.getCalledFunction()) { 4823 if (F->isDeclaration()) { 4824 const TargetIntrinsicInfo *II = TM.getIntrinsicInfo(); 4825 if (II) { 4826 if (unsigned IID = II->getIntrinsicID(F)) { 4827 RenameFn = visitIntrinsicCall(I, IID); 4828 if (!RenameFn) 4829 return; 4830 } 4831 } 4832 if (unsigned IID = F->getIntrinsicID()) { 4833 RenameFn = visitIntrinsicCall(I, IID); 4834 if (!RenameFn) 4835 return; 4836 } 4837 } 4838 4839 // Check for well-known libc/libm calls. If the function is internal, it 4840 // can't be a library call. 4841 if (!F->hasLocalLinkage() && F->hasName()) { 4842 StringRef Name = F->getName(); 4843 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl") { 4844 if (I.getNumOperands() == 3 && // Basic sanity checks. 4845 I.getArgOperand(0)->getType()->isFloatingPointTy() && 4846 I.getType() == I.getArgOperand(0)->getType() && 4847 I.getType() == I.getArgOperand(1)->getType()) { 4848 SDValue LHS = getValue(I.getArgOperand(0)); 4849 SDValue RHS = getValue(I.getArgOperand(1)); 4850 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(), 4851 LHS.getValueType(), LHS, RHS)); 4852 return; 4853 } 4854 } else if (Name == "fabs" || Name == "fabsf" || Name == "fabsl") { 4855 if (I.getNumOperands() == 2 && // Basic sanity checks. 4856 I.getArgOperand(0)->getType()->isFloatingPointTy() && 4857 I.getType() == I.getArgOperand(0)->getType()) { 4858 SDValue Tmp = getValue(I.getArgOperand(0)); 4859 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(), 4860 Tmp.getValueType(), Tmp)); 4861 return; 4862 } 4863 } else if (Name == "sin" || Name == "sinf" || Name == "sinl") { 4864 if (I.getNumOperands() == 2 && // Basic sanity checks. 4865 I.getArgOperand(0)->getType()->isFloatingPointTy() && 4866 I.getType() == I.getArgOperand(0)->getType() && 4867 I.onlyReadsMemory()) { 4868 SDValue Tmp = getValue(I.getArgOperand(0)); 4869 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(), 4870 Tmp.getValueType(), Tmp)); 4871 return; 4872 } 4873 } else if (Name == "cos" || Name == "cosf" || Name == "cosl") { 4874 if (I.getNumOperands() == 2 && // Basic sanity checks. 4875 I.getArgOperand(0)->getType()->isFloatingPointTy() && 4876 I.getType() == I.getArgOperand(0)->getType() && 4877 I.onlyReadsMemory()) { 4878 SDValue Tmp = getValue(I.getArgOperand(0)); 4879 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(), 4880 Tmp.getValueType(), Tmp)); 4881 return; 4882 } 4883 } else if (Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") { 4884 if (I.getNumOperands() == 2 && // Basic sanity checks. 4885 I.getArgOperand(0)->getType()->isFloatingPointTy() && 4886 I.getType() == I.getArgOperand(0)->getType() && 4887 I.onlyReadsMemory()) { 4888 SDValue Tmp = getValue(I.getArgOperand(0)); 4889 setValue(&I, DAG.getNode(ISD::FSQRT, getCurDebugLoc(), 4890 Tmp.getValueType(), Tmp)); 4891 return; 4892 } 4893 } else if (Name == "memcmp") { 4894 if (visitMemCmpCall(I)) 4895 return; 4896 } 4897 } 4898 } else if (isa<InlineAsm>(I.getCalledValue())) { 4899 visitInlineAsm(&I); 4900 return; 4901 } 4902 4903 SDValue Callee; 4904 if (!RenameFn) 4905 Callee = getValue(I.getCalledValue()); 4906 else 4907 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy()); 4908 4909 // Check if we can potentially perform a tail call. More detailed checking is 4910 // be done within LowerCallTo, after more information about the call is known. 4911 LowerCallTo(&I, Callee, I.isTailCall()); 4912} 4913 4914namespace llvm { 4915 4916/// AsmOperandInfo - This contains information for each constraint that we are 4917/// lowering. 4918class LLVM_LIBRARY_VISIBILITY SDISelAsmOperandInfo : 4919 public TargetLowering::AsmOperandInfo { 4920public: 4921 /// CallOperand - If this is the result output operand or a clobber 4922 /// this is null, otherwise it is the incoming operand to the CallInst. 4923 /// This gets modified as the asm is processed. 4924 SDValue CallOperand; 4925 4926 /// AssignedRegs - If this is a register or register class operand, this 4927 /// contains the set of register corresponding to the operand. 4928 RegsForValue AssignedRegs; 4929 4930 explicit SDISelAsmOperandInfo(const InlineAsm::ConstraintInfo &info) 4931 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) { 4932 } 4933 4934 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers 4935 /// busy in OutputRegs/InputRegs. 4936 void MarkAllocatedRegs(bool isOutReg, bool isInReg, 4937 std::set<unsigned> &OutputRegs, 4938 std::set<unsigned> &InputRegs, 4939 const TargetRegisterInfo &TRI) const { 4940 if (isOutReg) { 4941 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i) 4942 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI); 4943 } 4944 if (isInReg) { 4945 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i) 4946 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI); 4947 } 4948 } 4949 4950 /// getCallOperandValEVT - Return the EVT of the Value* that this operand 4951 /// corresponds to. If there is no Value* for this operand, it returns 4952 /// MVT::Other. 4953 EVT getCallOperandValEVT(LLVMContext &Context, 4954 const TargetLowering &TLI, 4955 const TargetData *TD) const { 4956 if (CallOperandVal == 0) return MVT::Other; 4957 4958 if (isa<BasicBlock>(CallOperandVal)) 4959 return TLI.getPointerTy(); 4960 4961 const llvm::Type *OpTy = CallOperandVal->getType(); 4962 4963 // If this is an indirect operand, the operand is a pointer to the 4964 // accessed type. 4965 if (isIndirect) { 4966 const llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy); 4967 if (!PtrTy) 4968 report_fatal_error("Indirect operand for inline asm not a pointer!"); 4969 OpTy = PtrTy->getElementType(); 4970 } 4971 4972 // If OpTy is not a single value, it may be a struct/union that we 4973 // can tile with integers. 4974 if (!OpTy->isSingleValueType() && OpTy->isSized()) { 4975 unsigned BitSize = TD->getTypeSizeInBits(OpTy); 4976 switch (BitSize) { 4977 default: break; 4978 case 1: 4979 case 8: 4980 case 16: 4981 case 32: 4982 case 64: 4983 case 128: 4984 OpTy = IntegerType::get(Context, BitSize); 4985 break; 4986 } 4987 } 4988 4989 return TLI.getValueType(OpTy, true); 4990 } 4991 4992private: 4993 /// MarkRegAndAliases - Mark the specified register and all aliases in the 4994 /// specified set. 4995 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs, 4996 const TargetRegisterInfo &TRI) { 4997 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg"); 4998 Regs.insert(Reg); 4999 if (const unsigned *Aliases = TRI.getAliasSet(Reg)) 5000 for (; *Aliases; ++Aliases) 5001 Regs.insert(*Aliases); 5002 } 5003}; 5004 5005} // end llvm namespace. 5006 5007/// isAllocatableRegister - If the specified register is safe to allocate, 5008/// i.e. it isn't a stack pointer or some other special register, return the 5009/// register class for the register. Otherwise, return null. 5010static const TargetRegisterClass * 5011isAllocatableRegister(unsigned Reg, MachineFunction &MF, 5012 const TargetLowering &TLI, 5013 const TargetRegisterInfo *TRI) { 5014 EVT FoundVT = MVT::Other; 5015 const TargetRegisterClass *FoundRC = 0; 5016 for (TargetRegisterInfo::regclass_iterator RCI = TRI->regclass_begin(), 5017 E = TRI->regclass_end(); RCI != E; ++RCI) { 5018 EVT ThisVT = MVT::Other; 5019 5020 const TargetRegisterClass *RC = *RCI; 5021 // If none of the value types for this register class are valid, we 5022 // can't use it. For example, 64-bit reg classes on 32-bit targets. 5023 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end(); 5024 I != E; ++I) { 5025 if (TLI.isTypeLegal(*I)) { 5026 // If we have already found this register in a different register class, 5027 // choose the one with the largest VT specified. For example, on 5028 // PowerPC, we favor f64 register classes over f32. 5029 if (FoundVT == MVT::Other || FoundVT.bitsLT(*I)) { 5030 ThisVT = *I; 5031 break; 5032 } 5033 } 5034 } 5035 5036 if (ThisVT == MVT::Other) continue; 5037 5038 // NOTE: This isn't ideal. In particular, this might allocate the 5039 // frame pointer in functions that need it (due to them not being taken 5040 // out of allocation, because a variable sized allocation hasn't been seen 5041 // yet). This is a slight code pessimization, but should still work. 5042 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF), 5043 E = RC->allocation_order_end(MF); I != E; ++I) 5044 if (*I == Reg) { 5045 // We found a matching register class. Keep looking at others in case 5046 // we find one with larger registers that this physreg is also in. 5047 FoundRC = RC; 5048 FoundVT = ThisVT; 5049 break; 5050 } 5051 } 5052 return FoundRC; 5053} 5054 5055/// GetRegistersForValue - Assign registers (virtual or physical) for the 5056/// specified operand. We prefer to assign virtual registers, to allow the 5057/// register allocator to handle the assignment process. However, if the asm 5058/// uses features that we can't model on machineinstrs, we have SDISel do the 5059/// allocation. This produces generally horrible, but correct, code. 5060/// 5061/// OpInfo describes the operand. 5062/// Input and OutputRegs are the set of already allocated physical registers. 5063/// 5064void SelectionDAGBuilder:: 5065GetRegistersForValue(SDISelAsmOperandInfo &OpInfo, 5066 std::set<unsigned> &OutputRegs, 5067 std::set<unsigned> &InputRegs) { 5068 LLVMContext &Context = FuncInfo.Fn->getContext(); 5069 5070 // Compute whether this value requires an input register, an output register, 5071 // or both. 5072 bool isOutReg = false; 5073 bool isInReg = false; 5074 switch (OpInfo.Type) { 5075 case InlineAsm::isOutput: 5076 isOutReg = true; 5077 5078 // If there is an input constraint that matches this, we need to reserve 5079 // the input register so no other inputs allocate to it. 5080 isInReg = OpInfo.hasMatchingInput(); 5081 break; 5082 case InlineAsm::isInput: 5083 isInReg = true; 5084 isOutReg = false; 5085 break; 5086 case InlineAsm::isClobber: 5087 isOutReg = true; 5088 isInReg = true; 5089 break; 5090 } 5091 5092 5093 MachineFunction &MF = DAG.getMachineFunction(); 5094 SmallVector<unsigned, 4> Regs; 5095 5096 // If this is a constraint for a single physreg, or a constraint for a 5097 // register class, find it. 5098 std::pair<unsigned, const TargetRegisterClass*> PhysReg = 5099 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode, 5100 OpInfo.ConstraintVT); 5101 5102 unsigned NumRegs = 1; 5103 if (OpInfo.ConstraintVT != MVT::Other) { 5104 // If this is a FP input in an integer register (or visa versa) insert a bit 5105 // cast of the input value. More generally, handle any case where the input 5106 // value disagrees with the register class we plan to stick this in. 5107 if (OpInfo.Type == InlineAsm::isInput && 5108 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) { 5109 // Try to convert to the first EVT that the reg class contains. If the 5110 // types are identical size, use a bitcast to convert (e.g. two differing 5111 // vector types). 5112 EVT RegVT = *PhysReg.second->vt_begin(); 5113 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) { 5114 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), 5115 RegVT, OpInfo.CallOperand); 5116 OpInfo.ConstraintVT = RegVT; 5117 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) { 5118 // If the input is a FP value and we want it in FP registers, do a 5119 // bitcast to the corresponding integer type. This turns an f64 value 5120 // into i64, which can be passed with two i32 values on a 32-bit 5121 // machine. 5122 RegVT = EVT::getIntegerVT(Context, 5123 OpInfo.ConstraintVT.getSizeInBits()); 5124 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), 5125 RegVT, OpInfo.CallOperand); 5126 OpInfo.ConstraintVT = RegVT; 5127 } 5128 } 5129 5130 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT); 5131 } 5132 5133 EVT RegVT; 5134 EVT ValueVT = OpInfo.ConstraintVT; 5135 5136 // If this is a constraint for a specific physical register, like {r17}, 5137 // assign it now. 5138 if (unsigned AssignedReg = PhysReg.first) { 5139 const TargetRegisterClass *RC = PhysReg.second; 5140 if (OpInfo.ConstraintVT == MVT::Other) 5141 ValueVT = *RC->vt_begin(); 5142 5143 // Get the actual register value type. This is important, because the user 5144 // may have asked for (e.g.) the AX register in i32 type. We need to 5145 // remember that AX is actually i16 to get the right extension. 5146 RegVT = *RC->vt_begin(); 5147 5148 // This is a explicit reference to a physical register. 5149 Regs.push_back(AssignedReg); 5150 5151 // If this is an expanded reference, add the rest of the regs to Regs. 5152 if (NumRegs != 1) { 5153 TargetRegisterClass::iterator I = RC->begin(); 5154 for (; *I != AssignedReg; ++I) 5155 assert(I != RC->end() && "Didn't find reg!"); 5156 5157 // Already added the first reg. 5158 --NumRegs; ++I; 5159 for (; NumRegs; --NumRegs, ++I) { 5160 assert(I != RC->end() && "Ran out of registers to allocate!"); 5161 Regs.push_back(*I); 5162 } 5163 } 5164 5165 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT); 5166 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo(); 5167 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI); 5168 return; 5169 } 5170 5171 // Otherwise, if this was a reference to an LLVM register class, create vregs 5172 // for this reference. 5173 if (const TargetRegisterClass *RC = PhysReg.second) { 5174 RegVT = *RC->vt_begin(); 5175 if (OpInfo.ConstraintVT == MVT::Other) 5176 ValueVT = RegVT; 5177 5178 // Create the appropriate number of virtual registers. 5179 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 5180 for (; NumRegs; --NumRegs) 5181 Regs.push_back(RegInfo.createVirtualRegister(RC)); 5182 5183 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT); 5184 return; 5185 } 5186 5187 // This is a reference to a register class that doesn't directly correspond 5188 // to an LLVM register class. Allocate NumRegs consecutive, available, 5189 // registers from the class. 5190 std::vector<unsigned> RegClassRegs 5191 = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode, 5192 OpInfo.ConstraintVT); 5193 5194 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo(); 5195 unsigned NumAllocated = 0; 5196 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) { 5197 unsigned Reg = RegClassRegs[i]; 5198 // See if this register is available. 5199 if ((isOutReg && OutputRegs.count(Reg)) || // Already used. 5200 (isInReg && InputRegs.count(Reg))) { // Already used. 5201 // Make sure we find consecutive registers. 5202 NumAllocated = 0; 5203 continue; 5204 } 5205 5206 // Check to see if this register is allocatable (i.e. don't give out the 5207 // stack pointer). 5208 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, TRI); 5209 if (!RC) { // Couldn't allocate this register. 5210 // Reset NumAllocated to make sure we return consecutive registers. 5211 NumAllocated = 0; 5212 continue; 5213 } 5214 5215 // Okay, this register is good, we can use it. 5216 ++NumAllocated; 5217 5218 // If we allocated enough consecutive registers, succeed. 5219 if (NumAllocated == NumRegs) { 5220 unsigned RegStart = (i-NumAllocated)+1; 5221 unsigned RegEnd = i+1; 5222 // Mark all of the allocated registers used. 5223 for (unsigned i = RegStart; i != RegEnd; ++i) 5224 Regs.push_back(RegClassRegs[i]); 5225 5226 OpInfo.AssignedRegs = RegsForValue(Regs, *RC->vt_begin(), 5227 OpInfo.ConstraintVT); 5228 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI); 5229 return; 5230 } 5231 } 5232 5233 // Otherwise, we couldn't allocate enough registers for this. 5234} 5235 5236/// visitInlineAsm - Handle a call to an InlineAsm object. 5237/// 5238void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) { 5239 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 5240 5241 /// ConstraintOperands - Information about all of the constraints. 5242 std::vector<SDISelAsmOperandInfo> ConstraintOperands; 5243 5244 std::set<unsigned> OutputRegs, InputRegs; 5245 5246 // Do a prepass over the constraints, canonicalizing them, and building up the 5247 // ConstraintOperands list. 5248 std::vector<InlineAsm::ConstraintInfo> 5249 ConstraintInfos = IA->ParseConstraints(); 5250 5251 bool hasMemory = hasInlineAsmMemConstraint(ConstraintInfos, TLI); 5252 5253 SDValue Chain, Flag; 5254 5255 // We won't need to flush pending loads if this asm doesn't touch 5256 // memory and is nonvolatile. 5257 if (hasMemory || IA->hasSideEffects()) 5258 Chain = getRoot(); 5259 else 5260 Chain = DAG.getRoot(); 5261 5262 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 5263 unsigned ResNo = 0; // ResNo - The result number of the next output. 5264 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 5265 ConstraintOperands.push_back(SDISelAsmOperandInfo(ConstraintInfos[i])); 5266 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back(); 5267 5268 EVT OpVT = MVT::Other; 5269 5270 // Compute the value type for each operand. 5271 switch (OpInfo.Type) { 5272 case InlineAsm::isOutput: 5273 // Indirect outputs just consume an argument. 5274 if (OpInfo.isIndirect) { 5275 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 5276 break; 5277 } 5278 5279 // The return value of the call is this value. As such, there is no 5280 // corresponding argument. 5281 assert(!CS.getType()->isVoidTy() && 5282 "Bad inline asm!"); 5283 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) { 5284 OpVT = TLI.getValueType(STy->getElementType(ResNo)); 5285 } else { 5286 assert(ResNo == 0 && "Asm only has one result!"); 5287 OpVT = TLI.getValueType(CS.getType()); 5288 } 5289 ++ResNo; 5290 break; 5291 case InlineAsm::isInput: 5292 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 5293 break; 5294 case InlineAsm::isClobber: 5295 // Nothing to do. 5296 break; 5297 } 5298 5299 // If this is an input or an indirect output, process the call argument. 5300 // BasicBlocks are labels, currently appearing only in asm's. 5301 if (OpInfo.CallOperandVal) { 5302 // Strip bitcasts, if any. This mostly comes up for functions. 5303 OpInfo.CallOperandVal = OpInfo.CallOperandVal->stripPointerCasts(); 5304 5305 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) { 5306 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]); 5307 } else { 5308 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal); 5309 } 5310 5311 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, TD); 5312 } 5313 5314 OpInfo.ConstraintVT = OpVT; 5315 } 5316 5317 // Second pass over the constraints: compute which constraint option to use 5318 // and assign registers to constraints that want a specific physreg. 5319 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 5320 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; 5321 5322 // If this is an output operand with a matching input operand, look up the 5323 // matching input. If their types mismatch, e.g. one is an integer, the 5324 // other is floating point, or their sizes are different, flag it as an 5325 // error. 5326 if (OpInfo.hasMatchingInput()) { 5327 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; 5328 5329 if (OpInfo.ConstraintVT != Input.ConstraintVT) { 5330 if ((OpInfo.ConstraintVT.isInteger() != 5331 Input.ConstraintVT.isInteger()) || 5332 (OpInfo.ConstraintVT.getSizeInBits() != 5333 Input.ConstraintVT.getSizeInBits())) { 5334 report_fatal_error("Unsupported asm: input constraint" 5335 " with a matching output constraint of" 5336 " incompatible type!"); 5337 } 5338 Input.ConstraintVT = OpInfo.ConstraintVT; 5339 } 5340 } 5341 5342 // Compute the constraint code and ConstraintType to use. 5343 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, hasMemory, &DAG); 5344 5345 // If this is a memory input, and if the operand is not indirect, do what we 5346 // need to to provide an address for the memory input. 5347 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 5348 !OpInfo.isIndirect) { 5349 assert(OpInfo.Type == InlineAsm::isInput && 5350 "Can only indirectify direct input operands!"); 5351 5352 // Memory operands really want the address of the value. If we don't have 5353 // an indirect input, put it in the constpool if we can, otherwise spill 5354 // it to a stack slot. 5355 5356 // If the operand is a float, integer, or vector constant, spill to a 5357 // constant pool entry to get its address. 5358 const Value *OpVal = OpInfo.CallOperandVal; 5359 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) || 5360 isa<ConstantVector>(OpVal)) { 5361 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal), 5362 TLI.getPointerTy()); 5363 } else { 5364 // Otherwise, create a stack slot and emit a store to it before the 5365 // asm. 5366 const Type *Ty = OpVal->getType(); 5367 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty); 5368 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty); 5369 MachineFunction &MF = DAG.getMachineFunction(); 5370 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false); 5371 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy()); 5372 Chain = DAG.getStore(Chain, getCurDebugLoc(), 5373 OpInfo.CallOperand, StackSlot, NULL, 0, 5374 false, false, 0); 5375 OpInfo.CallOperand = StackSlot; 5376 } 5377 5378 // There is no longer a Value* corresponding to this operand. 5379 OpInfo.CallOperandVal = 0; 5380 5381 // It is now an indirect operand. 5382 OpInfo.isIndirect = true; 5383 } 5384 5385 // If this constraint is for a specific register, allocate it before 5386 // anything else. 5387 if (OpInfo.ConstraintType == TargetLowering::C_Register) 5388 GetRegistersForValue(OpInfo, OutputRegs, InputRegs); 5389 } 5390 5391 ConstraintInfos.clear(); 5392 5393 // Second pass - Loop over all of the operands, assigning virtual or physregs 5394 // to register class operands. 5395 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { 5396 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; 5397 5398 // C_Register operands have already been allocated, Other/Memory don't need 5399 // to be. 5400 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass) 5401 GetRegistersForValue(OpInfo, OutputRegs, InputRegs); 5402 } 5403 5404 // AsmNodeOperands - The operands for the ISD::INLINEASM node. 5405 std::vector<SDValue> AsmNodeOperands; 5406 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain 5407 AsmNodeOperands.push_back( 5408 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(), 5409 TLI.getPointerTy())); 5410 5411 // If we have a !srcloc metadata node associated with it, we want to attach 5412 // this to the ultimately generated inline asm machineinstr. To do this, we 5413 // pass in the third operand as this (potentially null) inline asm MDNode. 5414 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc"); 5415 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc)); 5416 5417 // Loop over all of the inputs, copying the operand values into the 5418 // appropriate registers and processing the output regs. 5419 RegsForValue RetValRegs; 5420 5421 // IndirectStoresToEmit - The set of stores to emit after the inline asm node. 5422 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit; 5423 5424 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { 5425 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; 5426 5427 switch (OpInfo.Type) { 5428 case InlineAsm::isOutput: { 5429 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass && 5430 OpInfo.ConstraintType != TargetLowering::C_Register) { 5431 // Memory output, or 'other' output (e.g. 'X' constraint). 5432 assert(OpInfo.isIndirect && "Memory output must be indirect operand"); 5433 5434 // Add information to the INLINEASM node to know about this output. 5435 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1); 5436 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags, 5437 TLI.getPointerTy())); 5438 AsmNodeOperands.push_back(OpInfo.CallOperand); 5439 break; 5440 } 5441 5442 // Otherwise, this is a register or register class output. 5443 5444 // Copy the output from the appropriate register. Find a register that 5445 // we can use. 5446 if (OpInfo.AssignedRegs.Regs.empty()) 5447 report_fatal_error("Couldn't allocate output reg for constraint '" + 5448 Twine(OpInfo.ConstraintCode) + "'!"); 5449 5450 // If this is an indirect operand, store through the pointer after the 5451 // asm. 5452 if (OpInfo.isIndirect) { 5453 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs, 5454 OpInfo.CallOperandVal)); 5455 } else { 5456 // This is the result value of the call. 5457 assert(!CS.getType()->isVoidTy() && "Bad inline asm!"); 5458 // Concatenate this output onto the outputs list. 5459 RetValRegs.append(OpInfo.AssignedRegs); 5460 } 5461 5462 // Add information to the INLINEASM node to know that this register is 5463 // set. 5464 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ? 5465 InlineAsm::Kind_RegDefEarlyClobber : 5466 InlineAsm::Kind_RegDef, 5467 false, 5468 0, 5469 DAG, 5470 AsmNodeOperands); 5471 break; 5472 } 5473 case InlineAsm::isInput: { 5474 SDValue InOperandVal = OpInfo.CallOperand; 5475 5476 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint? 5477 // If this is required to match an output register we have already set, 5478 // just use its register. 5479 unsigned OperandNo = OpInfo.getMatchedOperand(); 5480 5481 // Scan until we find the definition we already emitted of this operand. 5482 // When we find it, create a RegsForValue operand. 5483 unsigned CurOp = InlineAsm::Op_FirstOperand; 5484 for (; OperandNo; --OperandNo) { 5485 // Advance to the next operand. 5486 unsigned OpFlag = 5487 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue(); 5488 assert((InlineAsm::isRegDefKind(OpFlag) || 5489 InlineAsm::isRegDefEarlyClobberKind(OpFlag) || 5490 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?"); 5491 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1; 5492 } 5493 5494 unsigned OpFlag = 5495 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue(); 5496 if (InlineAsm::isRegDefKind(OpFlag) || 5497 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) { 5498 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs. 5499 if (OpInfo.isIndirect) { 5500 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c 5501 LLVMContext &Ctx = *DAG.getContext(); 5502 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:" 5503 " don't know how to handle tied " 5504 "indirect register inputs"); 5505 } 5506 5507 RegsForValue MatchedRegs; 5508 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType()); 5509 EVT RegVT = AsmNodeOperands[CurOp+1].getValueType(); 5510 MatchedRegs.RegVTs.push_back(RegVT); 5511 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo(); 5512 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag); 5513 i != e; ++i) 5514 MatchedRegs.Regs.push_back 5515 (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT))); 5516 5517 // Use the produced MatchedRegs object to 5518 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(), 5519 Chain, &Flag); 5520 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, 5521 true, OpInfo.getMatchedOperand(), 5522 DAG, AsmNodeOperands); 5523 break; 5524 } 5525 5526 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!"); 5527 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 && 5528 "Unexpected number of operands"); 5529 // Add information to the INLINEASM node to know about this input. 5530 // See InlineAsm.h isUseOperandTiedToDef. 5531 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag, 5532 OpInfo.getMatchedOperand()); 5533 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag, 5534 TLI.getPointerTy())); 5535 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]); 5536 break; 5537 } 5538 5539 if (OpInfo.ConstraintType == TargetLowering::C_Other) { 5540 assert(!OpInfo.isIndirect && 5541 "Don't know how to handle indirect other inputs yet!"); 5542 5543 std::vector<SDValue> Ops; 5544 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0], 5545 hasMemory, Ops, DAG); 5546 if (Ops.empty()) 5547 report_fatal_error("Invalid operand for inline asm constraint '" + 5548 Twine(OpInfo.ConstraintCode) + "'!"); 5549 5550 // Add information to the INLINEASM node to know about this input. 5551 unsigned ResOpType = 5552 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size()); 5553 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType, 5554 TLI.getPointerTy())); 5555 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end()); 5556 break; 5557 } 5558 5559 if (OpInfo.ConstraintType == TargetLowering::C_Memory) { 5560 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!"); 5561 assert(InOperandVal.getValueType() == TLI.getPointerTy() && 5562 "Memory operands expect pointer values"); 5563 5564 // Add information to the INLINEASM node to know about this input. 5565 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1); 5566 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType, 5567 TLI.getPointerTy())); 5568 AsmNodeOperands.push_back(InOperandVal); 5569 break; 5570 } 5571 5572 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass || 5573 OpInfo.ConstraintType == TargetLowering::C_Register) && 5574 "Unknown constraint type!"); 5575 assert(!OpInfo.isIndirect && 5576 "Don't know how to handle indirect register inputs yet!"); 5577 5578 // Copy the input into the appropriate registers. 5579 if (OpInfo.AssignedRegs.Regs.empty() || 5580 !OpInfo.AssignedRegs.areValueTypesLegal(TLI)) 5581 report_fatal_error("Couldn't allocate input reg for constraint '" + 5582 Twine(OpInfo.ConstraintCode) + "'!"); 5583 5584 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(), 5585 Chain, &Flag); 5586 5587 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0, 5588 DAG, AsmNodeOperands); 5589 break; 5590 } 5591 case InlineAsm::isClobber: { 5592 // Add the clobbered value to the operand list, so that the register 5593 // allocator is aware that the physreg got clobbered. 5594 if (!OpInfo.AssignedRegs.Regs.empty()) 5595 OpInfo.AssignedRegs.AddInlineAsmOperands( 5596 InlineAsm::Kind_RegDefEarlyClobber, 5597 false, 0, DAG, 5598 AsmNodeOperands); 5599 break; 5600 } 5601 } 5602 } 5603 5604 // Finish up input operands. Set the input chain and add the flag last. 5605 AsmNodeOperands[0] = Chain; 5606 if (Flag.getNode()) AsmNodeOperands.push_back(Flag); 5607 5608 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(), 5609 DAG.getVTList(MVT::Other, MVT::Flag), 5610 &AsmNodeOperands[0], AsmNodeOperands.size()); 5611 Flag = Chain.getValue(1); 5612 5613 // If this asm returns a register value, copy the result from that register 5614 // and set it as the value of the call. 5615 if (!RetValRegs.Regs.empty()) { 5616 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), 5617 Chain, &Flag); 5618 5619 // FIXME: Why don't we do this for inline asms with MRVs? 5620 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) { 5621 EVT ResultType = TLI.getValueType(CS.getType()); 5622 5623 // If any of the results of the inline asm is a vector, it may have the 5624 // wrong width/num elts. This can happen for register classes that can 5625 // contain multiple different value types. The preg or vreg allocated may 5626 // not have the same VT as was expected. Convert it to the right type 5627 // with bit_convert. 5628 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) { 5629 Val = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), 5630 ResultType, Val); 5631 5632 } else if (ResultType != Val.getValueType() && 5633 ResultType.isInteger() && Val.getValueType().isInteger()) { 5634 // If a result value was tied to an input value, the computed result may 5635 // have a wider width than the expected result. Extract the relevant 5636 // portion. 5637 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val); 5638 } 5639 5640 assert(ResultType == Val.getValueType() && "Asm result value mismatch!"); 5641 } 5642 5643 setValue(CS.getInstruction(), Val); 5644 // Don't need to use this as a chain in this case. 5645 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty()) 5646 return; 5647 } 5648 5649 std::vector<std::pair<SDValue, const Value *> > StoresToEmit; 5650 5651 // Process indirect outputs, first output all of the flagged copies out of 5652 // physregs. 5653 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) { 5654 RegsForValue &OutRegs = IndirectStoresToEmit[i].first; 5655 const Value *Ptr = IndirectStoresToEmit[i].second; 5656 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), 5657 Chain, &Flag); 5658 StoresToEmit.push_back(std::make_pair(OutVal, Ptr)); 5659 } 5660 5661 // Emit the non-flagged stores from the physregs. 5662 SmallVector<SDValue, 8> OutChains; 5663 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) { 5664 SDValue Val = DAG.getStore(Chain, getCurDebugLoc(), 5665 StoresToEmit[i].first, 5666 getValue(StoresToEmit[i].second), 5667 StoresToEmit[i].second, 0, 5668 false, false, 0); 5669 OutChains.push_back(Val); 5670 } 5671 5672 if (!OutChains.empty()) 5673 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other, 5674 &OutChains[0], OutChains.size()); 5675 5676 DAG.setRoot(Chain); 5677} 5678 5679void SelectionDAGBuilder::visitVAStart(const CallInst &I) { 5680 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(), 5681 MVT::Other, getRoot(), 5682 getValue(I.getArgOperand(0)), 5683 DAG.getSrcValue(I.getArgOperand(0)))); 5684} 5685 5686void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) { 5687 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(), 5688 getRoot(), getValue(I.getOperand(0)), 5689 DAG.getSrcValue(I.getOperand(0))); 5690 setValue(&I, V); 5691 DAG.setRoot(V.getValue(1)); 5692} 5693 5694void SelectionDAGBuilder::visitVAEnd(const CallInst &I) { 5695 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(), 5696 MVT::Other, getRoot(), 5697 getValue(I.getArgOperand(0)), 5698 DAG.getSrcValue(I.getArgOperand(0)))); 5699} 5700 5701void SelectionDAGBuilder::visitVACopy(const CallInst &I) { 5702 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(), 5703 MVT::Other, getRoot(), 5704 getValue(I.getArgOperand(0)), 5705 getValue(I.getArgOperand(1)), 5706 DAG.getSrcValue(I.getArgOperand(0)), 5707 DAG.getSrcValue(I.getArgOperand(1)))); 5708} 5709 5710/// TargetLowering::LowerCallTo - This is the default LowerCallTo 5711/// implementation, which just calls LowerCall. 5712/// FIXME: When all targets are 5713/// migrated to using LowerCall, this hook should be integrated into SDISel. 5714std::pair<SDValue, SDValue> 5715TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy, 5716 bool RetSExt, bool RetZExt, bool isVarArg, 5717 bool isInreg, unsigned NumFixedArgs, 5718 CallingConv::ID CallConv, bool isTailCall, 5719 bool isReturnValueUsed, 5720 SDValue Callee, 5721 ArgListTy &Args, SelectionDAG &DAG, 5722 DebugLoc dl) const { 5723 // Handle all of the outgoing arguments. 5724 SmallVector<ISD::OutputArg, 32> Outs; 5725 for (unsigned i = 0, e = Args.size(); i != e; ++i) { 5726 SmallVector<EVT, 4> ValueVTs; 5727 ComputeValueVTs(*this, Args[i].Ty, ValueVTs); 5728 for (unsigned Value = 0, NumValues = ValueVTs.size(); 5729 Value != NumValues; ++Value) { 5730 EVT VT = ValueVTs[Value]; 5731 const Type *ArgTy = VT.getTypeForEVT(RetTy->getContext()); 5732 SDValue Op = SDValue(Args[i].Node.getNode(), 5733 Args[i].Node.getResNo() + Value); 5734 ISD::ArgFlagsTy Flags; 5735 unsigned OriginalAlignment = 5736 getTargetData()->getABITypeAlignment(ArgTy); 5737 5738 if (Args[i].isZExt) 5739 Flags.setZExt(); 5740 if (Args[i].isSExt) 5741 Flags.setSExt(); 5742 if (Args[i].isInReg) 5743 Flags.setInReg(); 5744 if (Args[i].isSRet) 5745 Flags.setSRet(); 5746 if (Args[i].isByVal) { 5747 Flags.setByVal(); 5748 const PointerType *Ty = cast<PointerType>(Args[i].Ty); 5749 const Type *ElementTy = Ty->getElementType(); 5750 unsigned FrameAlign = getByValTypeAlignment(ElementTy); 5751 unsigned FrameSize = getTargetData()->getTypeAllocSize(ElementTy); 5752 // For ByVal, alignment should come from FE. BE will guess if this 5753 // info is not there but there are cases it cannot get right. 5754 if (Args[i].Alignment) 5755 FrameAlign = Args[i].Alignment; 5756 Flags.setByValAlign(FrameAlign); 5757 Flags.setByValSize(FrameSize); 5758 } 5759 if (Args[i].isNest) 5760 Flags.setNest(); 5761 Flags.setOrigAlign(OriginalAlignment); 5762 5763 EVT PartVT = getRegisterType(RetTy->getContext(), VT); 5764 unsigned NumParts = getNumRegisters(RetTy->getContext(), VT); 5765 SmallVector<SDValue, 4> Parts(NumParts); 5766 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 5767 5768 if (Args[i].isSExt) 5769 ExtendKind = ISD::SIGN_EXTEND; 5770 else if (Args[i].isZExt) 5771 ExtendKind = ISD::ZERO_EXTEND; 5772 5773 getCopyToParts(DAG, dl, Op, &Parts[0], NumParts, 5774 PartVT, ExtendKind); 5775 5776 for (unsigned j = 0; j != NumParts; ++j) { 5777 // if it isn't first piece, alignment must be 1 5778 ISD::OutputArg MyFlags(Flags, Parts[j], i < NumFixedArgs); 5779 if (NumParts > 1 && j == 0) 5780 MyFlags.Flags.setSplit(); 5781 else if (j != 0) 5782 MyFlags.Flags.setOrigAlign(1); 5783 5784 Outs.push_back(MyFlags); 5785 } 5786 } 5787 } 5788 5789 // Handle the incoming return values from the call. 5790 SmallVector<ISD::InputArg, 32> Ins; 5791 SmallVector<EVT, 4> RetTys; 5792 ComputeValueVTs(*this, RetTy, RetTys); 5793 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) { 5794 EVT VT = RetTys[I]; 5795 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT); 5796 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT); 5797 for (unsigned i = 0; i != NumRegs; ++i) { 5798 ISD::InputArg MyFlags; 5799 MyFlags.VT = RegisterVT; 5800 MyFlags.Used = isReturnValueUsed; 5801 if (RetSExt) 5802 MyFlags.Flags.setSExt(); 5803 if (RetZExt) 5804 MyFlags.Flags.setZExt(); 5805 if (isInreg) 5806 MyFlags.Flags.setInReg(); 5807 Ins.push_back(MyFlags); 5808 } 5809 } 5810 5811 SmallVector<SDValue, 4> InVals; 5812 Chain = LowerCall(Chain, Callee, CallConv, isVarArg, isTailCall, 5813 Outs, Ins, dl, DAG, InVals); 5814 5815 // Verify that the target's LowerCall behaved as expected. 5816 assert(Chain.getNode() && Chain.getValueType() == MVT::Other && 5817 "LowerCall didn't return a valid chain!"); 5818 assert((!isTailCall || InVals.empty()) && 5819 "LowerCall emitted a return value for a tail call!"); 5820 assert((isTailCall || InVals.size() == Ins.size()) && 5821 "LowerCall didn't emit the correct number of values!"); 5822 5823 // For a tail call, the return value is merely live-out and there aren't 5824 // any nodes in the DAG representing it. Return a special value to 5825 // indicate that a tail call has been emitted and no more Instructions 5826 // should be processed in the current block. 5827 if (isTailCall) { 5828 DAG.setRoot(Chain); 5829 return std::make_pair(SDValue(), SDValue()); 5830 } 5831 5832 DEBUG(for (unsigned i = 0, e = Ins.size(); i != e; ++i) { 5833 assert(InVals[i].getNode() && 5834 "LowerCall emitted a null value!"); 5835 assert(Ins[i].VT == InVals[i].getValueType() && 5836 "LowerCall emitted a value with the wrong type!"); 5837 }); 5838 5839 // Collect the legal value parts into potentially illegal values 5840 // that correspond to the original function's return values. 5841 ISD::NodeType AssertOp = ISD::DELETED_NODE; 5842 if (RetSExt) 5843 AssertOp = ISD::AssertSext; 5844 else if (RetZExt) 5845 AssertOp = ISD::AssertZext; 5846 SmallVector<SDValue, 4> ReturnValues; 5847 unsigned CurReg = 0; 5848 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) { 5849 EVT VT = RetTys[I]; 5850 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT); 5851 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT); 5852 5853 ReturnValues.push_back(getCopyFromParts(DAG, dl, &InVals[CurReg], 5854 NumRegs, RegisterVT, VT, 5855 AssertOp)); 5856 CurReg += NumRegs; 5857 } 5858 5859 // For a function returning void, there is no return value. We can't create 5860 // such a node, so we just return a null return value in that case. In 5861 // that case, nothing will actualy look at the value. 5862 if (ReturnValues.empty()) 5863 return std::make_pair(SDValue(), Chain); 5864 5865 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl, 5866 DAG.getVTList(&RetTys[0], RetTys.size()), 5867 &ReturnValues[0], ReturnValues.size()); 5868 return std::make_pair(Res, Chain); 5869} 5870 5871void TargetLowering::LowerOperationWrapper(SDNode *N, 5872 SmallVectorImpl<SDValue> &Results, 5873 SelectionDAG &DAG) const { 5874 SDValue Res = LowerOperation(SDValue(N, 0), DAG); 5875 if (Res.getNode()) 5876 Results.push_back(Res); 5877} 5878 5879SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { 5880 llvm_unreachable("LowerOperation not implemented for this target!"); 5881 return SDValue(); 5882} 5883 5884void 5885SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) { 5886 SDValue Op = getValue(V); 5887 assert((Op.getOpcode() != ISD::CopyFromReg || 5888 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) && 5889 "Copy from a reg to the same reg!"); 5890 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg"); 5891 5892 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType()); 5893 SDValue Chain = DAG.getEntryNode(); 5894 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0); 5895 PendingExports.push_back(Chain); 5896} 5897 5898#include "llvm/CodeGen/SelectionDAGISel.h" 5899 5900void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) { 5901 // If this is the entry block, emit arguments. 5902 const Function &F = *LLVMBB->getParent(); 5903 SelectionDAG &DAG = SDB->DAG; 5904 SDValue OldRoot = DAG.getRoot(); 5905 DebugLoc dl = SDB->getCurDebugLoc(); 5906 const TargetData *TD = TLI.getTargetData(); 5907 SmallVector<ISD::InputArg, 16> Ins; 5908 5909 // Check whether the function can return without sret-demotion. 5910 SmallVector<EVT, 4> OutVTs; 5911 SmallVector<ISD::ArgFlagsTy, 4> OutsFlags; 5912 getReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(), 5913 OutVTs, OutsFlags, TLI); 5914 5915 FuncInfo->CanLowerReturn = TLI.CanLowerReturn(F.getCallingConv(), 5916 F.isVarArg(), 5917 OutVTs, OutsFlags, DAG); 5918 if (!FuncInfo->CanLowerReturn) { 5919 // Put in an sret pointer parameter before all the other parameters. 5920 SmallVector<EVT, 1> ValueVTs; 5921 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs); 5922 5923 // NOTE: Assuming that a pointer will never break down to more than one VT 5924 // or one register. 5925 ISD::ArgFlagsTy Flags; 5926 Flags.setSRet(); 5927 EVT RegisterVT = TLI.getRegisterType(*DAG.getContext(), ValueVTs[0]); 5928 ISD::InputArg RetArg(Flags, RegisterVT, true); 5929 Ins.push_back(RetArg); 5930 } 5931 5932 // Set up the incoming argument description vector. 5933 unsigned Idx = 1; 5934 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); 5935 I != E; ++I, ++Idx) { 5936 SmallVector<EVT, 4> ValueVTs; 5937 ComputeValueVTs(TLI, I->getType(), ValueVTs); 5938 bool isArgValueUsed = !I->use_empty(); 5939 for (unsigned Value = 0, NumValues = ValueVTs.size(); 5940 Value != NumValues; ++Value) { 5941 EVT VT = ValueVTs[Value]; 5942 const Type *ArgTy = VT.getTypeForEVT(*DAG.getContext()); 5943 ISD::ArgFlagsTy Flags; 5944 unsigned OriginalAlignment = 5945 TD->getABITypeAlignment(ArgTy); 5946 5947 if (F.paramHasAttr(Idx, Attribute::ZExt)) 5948 Flags.setZExt(); 5949 if (F.paramHasAttr(Idx, Attribute::SExt)) 5950 Flags.setSExt(); 5951 if (F.paramHasAttr(Idx, Attribute::InReg)) 5952 Flags.setInReg(); 5953 if (F.paramHasAttr(Idx, Attribute::StructRet)) 5954 Flags.setSRet(); 5955 if (F.paramHasAttr(Idx, Attribute::ByVal)) { 5956 Flags.setByVal(); 5957 const PointerType *Ty = cast<PointerType>(I->getType()); 5958 const Type *ElementTy = Ty->getElementType(); 5959 unsigned FrameAlign = TLI.getByValTypeAlignment(ElementTy); 5960 unsigned FrameSize = TD->getTypeAllocSize(ElementTy); 5961 // For ByVal, alignment should be passed from FE. BE will guess if 5962 // this info is not there but there are cases it cannot get right. 5963 if (F.getParamAlignment(Idx)) 5964 FrameAlign = F.getParamAlignment(Idx); 5965 Flags.setByValAlign(FrameAlign); 5966 Flags.setByValSize(FrameSize); 5967 } 5968 if (F.paramHasAttr(Idx, Attribute::Nest)) 5969 Flags.setNest(); 5970 Flags.setOrigAlign(OriginalAlignment); 5971 5972 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), VT); 5973 unsigned NumRegs = TLI.getNumRegisters(*CurDAG->getContext(), VT); 5974 for (unsigned i = 0; i != NumRegs; ++i) { 5975 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed); 5976 if (NumRegs > 1 && i == 0) 5977 MyFlags.Flags.setSplit(); 5978 // if it isn't first piece, alignment must be 1 5979 else if (i > 0) 5980 MyFlags.Flags.setOrigAlign(1); 5981 Ins.push_back(MyFlags); 5982 } 5983 } 5984 } 5985 5986 // Call the target to set up the argument values. 5987 SmallVector<SDValue, 8> InVals; 5988 SDValue NewRoot = TLI.LowerFormalArguments(DAG.getRoot(), F.getCallingConv(), 5989 F.isVarArg(), Ins, 5990 dl, DAG, InVals); 5991 5992 // Verify that the target's LowerFormalArguments behaved as expected. 5993 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other && 5994 "LowerFormalArguments didn't return a valid chain!"); 5995 assert(InVals.size() == Ins.size() && 5996 "LowerFormalArguments didn't emit the correct number of values!"); 5997 DEBUG({ 5998 for (unsigned i = 0, e = Ins.size(); i != e; ++i) { 5999 assert(InVals[i].getNode() && 6000 "LowerFormalArguments emitted a null value!"); 6001 assert(Ins[i].VT == InVals[i].getValueType() && 6002 "LowerFormalArguments emitted a value with the wrong type!"); 6003 } 6004 }); 6005 6006 // Update the DAG with the new chain value resulting from argument lowering. 6007 DAG.setRoot(NewRoot); 6008 6009 // Set up the argument values. 6010 unsigned i = 0; 6011 Idx = 1; 6012 if (!FuncInfo->CanLowerReturn) { 6013 // Create a virtual register for the sret pointer, and put in a copy 6014 // from the sret argument into it. 6015 SmallVector<EVT, 1> ValueVTs; 6016 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs); 6017 EVT VT = ValueVTs[0]; 6018 EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT); 6019 ISD::NodeType AssertOp = ISD::DELETED_NODE; 6020 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1, 6021 RegVT, VT, AssertOp); 6022 6023 MachineFunction& MF = SDB->DAG.getMachineFunction(); 6024 MachineRegisterInfo& RegInfo = MF.getRegInfo(); 6025 unsigned SRetReg = RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)); 6026 FuncInfo->DemoteRegister = SRetReg; 6027 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurDebugLoc(), 6028 SRetReg, ArgValue); 6029 DAG.setRoot(NewRoot); 6030 6031 // i indexes lowered arguments. Bump it past the hidden sret argument. 6032 // Idx indexes LLVM arguments. Don't touch it. 6033 ++i; 6034 } 6035 6036 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; 6037 ++I, ++Idx) { 6038 SmallVector<SDValue, 4> ArgValues; 6039 SmallVector<EVT, 4> ValueVTs; 6040 ComputeValueVTs(TLI, I->getType(), ValueVTs); 6041 unsigned NumValues = ValueVTs.size(); 6042 6043 // If this argument is unused then remember its value. It is used to generate 6044 // debugging information. 6045 if (I->use_empty() && NumValues) 6046 SDB->setUnusedArgValue(I, InVals[i]); 6047 6048 for (unsigned Value = 0; Value != NumValues; ++Value) { 6049 EVT VT = ValueVTs[Value]; 6050 EVT PartVT = TLI.getRegisterType(*CurDAG->getContext(), VT); 6051 unsigned NumParts = TLI.getNumRegisters(*CurDAG->getContext(), VT); 6052 6053 if (!I->use_empty()) { 6054 ISD::NodeType AssertOp = ISD::DELETED_NODE; 6055 if (F.paramHasAttr(Idx, Attribute::SExt)) 6056 AssertOp = ISD::AssertSext; 6057 else if (F.paramHasAttr(Idx, Attribute::ZExt)) 6058 AssertOp = ISD::AssertZext; 6059 6060 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i], 6061 NumParts, PartVT, VT, 6062 AssertOp)); 6063 } 6064 6065 i += NumParts; 6066 } 6067 6068 if (!I->use_empty()) { 6069 SDValue Res; 6070 if (!ArgValues.empty()) 6071 Res = DAG.getMergeValues(&ArgValues[0], NumValues, 6072 SDB->getCurDebugLoc()); 6073 SDB->setValue(I, Res); 6074 6075 // If this argument is live outside of the entry block, insert a copy from 6076 // whereever we got it to the vreg that other BB's will reference it as. 6077 SDB->CopyToExportRegsIfNeeded(I); 6078 } 6079 } 6080 6081 assert(i == InVals.size() && "Argument register count mismatch!"); 6082 6083 // Finally, if the target has anything special to do, allow it to do so. 6084 // FIXME: this should insert code into the DAG! 6085 EmitFunctionEntryCode(); 6086} 6087 6088/// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to 6089/// ensure constants are generated when needed. Remember the virtual registers 6090/// that need to be added to the Machine PHI nodes as input. We cannot just 6091/// directly add them, because expansion might result in multiple MBB's for one 6092/// BB. As such, the start of the BB might correspond to a different MBB than 6093/// the end. 6094/// 6095void 6096SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) { 6097 const TerminatorInst *TI = LLVMBB->getTerminator(); 6098 6099 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled; 6100 6101 // Check successor nodes' PHI nodes that expect a constant to be available 6102 // from this block. 6103 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) { 6104 const BasicBlock *SuccBB = TI->getSuccessor(succ); 6105 if (!isa<PHINode>(SuccBB->begin())) continue; 6106 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB]; 6107 6108 // If this terminator has multiple identical successors (common for 6109 // switches), only handle each succ once. 6110 if (!SuccsHandled.insert(SuccMBB)) continue; 6111 6112 MachineBasicBlock::iterator MBBI = SuccMBB->begin(); 6113 6114 // At this point we know that there is a 1-1 correspondence between LLVM PHI 6115 // nodes and Machine PHI nodes, but the incoming operands have not been 6116 // emitted yet. 6117 for (BasicBlock::const_iterator I = SuccBB->begin(); 6118 const PHINode *PN = dyn_cast<PHINode>(I); ++I) { 6119 // Ignore dead phi's. 6120 if (PN->use_empty()) continue; 6121 6122 unsigned Reg; 6123 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB); 6124 6125 if (const Constant *C = dyn_cast<Constant>(PHIOp)) { 6126 unsigned &RegOut = ConstantsOut[C]; 6127 if (RegOut == 0) { 6128 RegOut = FuncInfo.CreateRegForValue(C); 6129 CopyValueToVirtualRegister(C, RegOut); 6130 } 6131 Reg = RegOut; 6132 } else { 6133 Reg = FuncInfo.ValueMap[PHIOp]; 6134 if (Reg == 0) { 6135 assert(isa<AllocaInst>(PHIOp) && 6136 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) && 6137 "Didn't codegen value into a register!??"); 6138 Reg = FuncInfo.CreateRegForValue(PHIOp); 6139 CopyValueToVirtualRegister(PHIOp, Reg); 6140 } 6141 } 6142 6143 // Remember that this register needs to added to the machine PHI node as 6144 // the input for this MBB. 6145 SmallVector<EVT, 4> ValueVTs; 6146 ComputeValueVTs(TLI, PN->getType(), ValueVTs); 6147 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) { 6148 EVT VT = ValueVTs[vti]; 6149 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT); 6150 for (unsigned i = 0, e = NumRegisters; i != e; ++i) 6151 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i)); 6152 Reg += NumRegisters; 6153 } 6154 } 6155 } 6156 ConstantsOut.clear(); 6157} 6158