SelectionDAGBuilder.cpp revision 098e40cd1cb2e502cd9672aecf75cedde5713c35
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 "llvm/ADT/BitVector.h" 18#include "llvm/ADT/SmallSet.h" 19#include "llvm/Analysis/AliasAnalysis.h" 20#include "llvm/Analysis/ConstantFolding.h" 21#include "llvm/Constants.h" 22#include "llvm/CallingConv.h" 23#include "llvm/DerivedTypes.h" 24#include "llvm/Function.h" 25#include "llvm/GlobalVariable.h" 26#include "llvm/InlineAsm.h" 27#include "llvm/Instructions.h" 28#include "llvm/Intrinsics.h" 29#include "llvm/IntrinsicInst.h" 30#include "llvm/LLVMContext.h" 31#include "llvm/Module.h" 32#include "llvm/CodeGen/Analysis.h" 33#include "llvm/CodeGen/FastISel.h" 34#include "llvm/CodeGen/FunctionLoweringInfo.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 73static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL, 74 const SDValue *Parts, unsigned NumParts, 75 EVT PartVT, EVT ValueVT); 76 77/// getCopyFromParts - Create a value that contains the specified legal parts 78/// combined into the value they represent. If the parts combine to a type 79/// larger then ValueVT then AssertOp can be used to specify whether the extra 80/// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT 81/// (ISD::AssertSext). 82static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc DL, 83 const SDValue *Parts, 84 unsigned NumParts, EVT PartVT, EVT ValueVT, 85 ISD::NodeType AssertOp = ISD::DELETED_NODE) { 86 if (ValueVT.isVector()) 87 return getCopyFromPartsVector(DAG, DL, Parts, NumParts, PartVT, ValueVT); 88 89 assert(NumParts > 0 && "No parts to assemble!"); 90 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 91 SDValue Val = Parts[0]; 92 93 if (NumParts > 1) { 94 // Assemble the value from multiple parts. 95 if (ValueVT.isInteger()) { 96 unsigned PartBits = PartVT.getSizeInBits(); 97 unsigned ValueBits = ValueVT.getSizeInBits(); 98 99 // Assemble the power of 2 part. 100 unsigned RoundParts = NumParts & (NumParts - 1) ? 101 1 << Log2_32(NumParts) : NumParts; 102 unsigned RoundBits = PartBits * RoundParts; 103 EVT RoundVT = RoundBits == ValueBits ? 104 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits); 105 SDValue Lo, Hi; 106 107 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2); 108 109 if (RoundParts > 2) { 110 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2, 111 PartVT, HalfVT); 112 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2, 113 RoundParts / 2, PartVT, HalfVT); 114 } else { 115 Lo = DAG.getNode(ISD::BIT_CONVERT, DL, HalfVT, Parts[0]); 116 Hi = DAG.getNode(ISD::BIT_CONVERT, DL, HalfVT, Parts[1]); 117 } 118 119 if (TLI.isBigEndian()) 120 std::swap(Lo, Hi); 121 122 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi); 123 124 if (RoundParts < NumParts) { 125 // Assemble the trailing non-power-of-2 part. 126 unsigned OddParts = NumParts - RoundParts; 127 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits); 128 Hi = getCopyFromParts(DAG, DL, 129 Parts + RoundParts, OddParts, PartVT, OddVT); 130 131 // Combine the round and odd parts. 132 Lo = Val; 133 if (TLI.isBigEndian()) 134 std::swap(Lo, Hi); 135 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); 136 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi); 137 Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi, 138 DAG.getConstant(Lo.getValueType().getSizeInBits(), 139 TLI.getPointerTy())); 140 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo); 141 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi); 142 } 143 } else if (PartVT.isFloatingPoint()) { 144 // FP split into multiple FP parts (for ppcf128) 145 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == EVT(MVT::f64) && 146 "Unexpected split"); 147 SDValue Lo, Hi; 148 Lo = DAG.getNode(ISD::BIT_CONVERT, DL, EVT(MVT::f64), Parts[0]); 149 Hi = DAG.getNode(ISD::BIT_CONVERT, DL, EVT(MVT::f64), Parts[1]); 150 if (TLI.isBigEndian()) 151 std::swap(Lo, Hi); 152 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi); 153 } else { 154 // FP split into integer parts (soft fp) 155 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() && 156 !PartVT.isVector() && "Unexpected split"); 157 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits()); 158 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT); 159 } 160 } 161 162 // There is now one part, held in Val. Correct it to match ValueVT. 163 PartVT = Val.getValueType(); 164 165 if (PartVT == ValueVT) 166 return Val; 167 168 if (PartVT.isInteger() && ValueVT.isInteger()) { 169 if (ValueVT.bitsLT(PartVT)) { 170 // For a truncate, see if we have any information to 171 // indicate whether the truncated bits will always be 172 // zero or sign-extension. 173 if (AssertOp != ISD::DELETED_NODE) 174 Val = DAG.getNode(AssertOp, DL, PartVT, Val, 175 DAG.getValueType(ValueVT)); 176 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); 177 } 178 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val); 179 } 180 181 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { 182 // FP_ROUND's are always exact here. 183 if (ValueVT.bitsLT(Val.getValueType())) 184 return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val, 185 DAG.getIntPtrConstant(1)); 186 187 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val); 188 } 189 190 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) 191 return DAG.getNode(ISD::BIT_CONVERT, DL, ValueVT, Val); 192 193 llvm_unreachable("Unknown mismatch!"); 194 return SDValue(); 195} 196 197/// getCopyFromParts - Create a value that contains the specified legal parts 198/// combined into the value they represent. If the parts combine to a type 199/// larger then ValueVT then AssertOp can be used to specify whether the extra 200/// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT 201/// (ISD::AssertSext). 202static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL, 203 const SDValue *Parts, unsigned NumParts, 204 EVT PartVT, EVT ValueVT) { 205 assert(ValueVT.isVector() && "Not a vector value"); 206 assert(NumParts > 0 && "No parts to assemble!"); 207 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 208 SDValue Val = Parts[0]; 209 210 // Handle a multi-element vector. 211 if (NumParts > 1) { 212 EVT IntermediateVT, RegisterVT; 213 unsigned NumIntermediates; 214 unsigned NumRegs = 215 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT, 216 NumIntermediates, RegisterVT); 217 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!"); 218 NumParts = NumRegs; // Silence a compiler warning. 219 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!"); 220 assert(RegisterVT == Parts[0].getValueType() && 221 "Part type doesn't match part!"); 222 223 // Assemble the parts into intermediate operands. 224 SmallVector<SDValue, 8> Ops(NumIntermediates); 225 if (NumIntermediates == NumParts) { 226 // If the register was not expanded, truncate or copy the value, 227 // as appropriate. 228 for (unsigned i = 0; i != NumParts; ++i) 229 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1, 230 PartVT, IntermediateVT); 231 } else if (NumParts > 0) { 232 // If the intermediate type was expanded, build the intermediate 233 // operands from the parts. 234 assert(NumParts % NumIntermediates == 0 && 235 "Must expand into a divisible number of parts!"); 236 unsigned Factor = NumParts / NumIntermediates; 237 for (unsigned i = 0; i != NumIntermediates; ++i) 238 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor, 239 PartVT, IntermediateVT); 240 } 241 242 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the 243 // intermediate operands. 244 Val = DAG.getNode(IntermediateVT.isVector() ? 245 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL, 246 ValueVT, &Ops[0], NumIntermediates); 247 } 248 249 // There is now one part, held in Val. Correct it to match ValueVT. 250 PartVT = Val.getValueType(); 251 252 if (PartVT == ValueVT) 253 return Val; 254 255 if (PartVT.isVector()) { 256 // If the element type of the source/dest vectors are the same, but the 257 // parts vector has more elements than the value vector, then we have a 258 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the 259 // elements we want. 260 if (PartVT.getVectorElementType() == ValueVT.getVectorElementType()) { 261 assert(PartVT.getVectorNumElements() > ValueVT.getVectorNumElements() && 262 "Cannot narrow, it would be a lossy transformation"); 263 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val, 264 DAG.getIntPtrConstant(0)); 265 } 266 267 // Vector/Vector bitcast. 268 return DAG.getNode(ISD::BIT_CONVERT, DL, ValueVT, Val); 269 } 270 271 assert(ValueVT.getVectorElementType() == PartVT && 272 ValueVT.getVectorNumElements() == 1 && 273 "Only trivial scalar-to-vector conversions should get here!"); 274 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val); 275} 276 277 278 279 280static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc dl, 281 SDValue Val, SDValue *Parts, unsigned NumParts, 282 EVT PartVT); 283 284/// getCopyToParts - Create a series of nodes that contain the specified value 285/// split into legal parts. If the parts contain more bits than Val, then, for 286/// integers, ExtendKind can be used to specify how to generate the extra bits. 287static void getCopyToParts(SelectionDAG &DAG, DebugLoc DL, 288 SDValue Val, SDValue *Parts, unsigned NumParts, 289 EVT PartVT, 290 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) { 291 EVT ValueVT = Val.getValueType(); 292 293 // Handle the vector case separately. 294 if (ValueVT.isVector()) 295 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT); 296 297 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 298 unsigned PartBits = PartVT.getSizeInBits(); 299 unsigned OrigNumParts = NumParts; 300 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!"); 301 302 if (NumParts == 0) 303 return; 304 305 assert(!ValueVT.isVector() && "Vector case handled elsewhere"); 306 if (PartVT == ValueVT) { 307 assert(NumParts == 1 && "No-op copy with multiple parts!"); 308 Parts[0] = Val; 309 return; 310 } 311 312 if (NumParts * PartBits > ValueVT.getSizeInBits()) { 313 // If the parts cover more bits than the value has, promote the value. 314 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { 315 assert(NumParts == 1 && "Do not know what to promote to!"); 316 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val); 317 } else { 318 assert(PartVT.isInteger() && ValueVT.isInteger() && 319 "Unknown mismatch!"); 320 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); 321 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val); 322 } 323 } else if (PartBits == ValueVT.getSizeInBits()) { 324 // Different types of the same size. 325 assert(NumParts == 1 && PartVT != ValueVT); 326 Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val); 327 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) { 328 // If the parts cover less bits than value has, truncate the value. 329 assert(PartVT.isInteger() && ValueVT.isInteger() && 330 "Unknown mismatch!"); 331 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); 332 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); 333 } 334 335 // The value may have changed - recompute ValueVT. 336 ValueVT = Val.getValueType(); 337 assert(NumParts * PartBits == ValueVT.getSizeInBits() && 338 "Failed to tile the value with PartVT!"); 339 340 if (NumParts == 1) { 341 assert(PartVT == ValueVT && "Type conversion failed!"); 342 Parts[0] = Val; 343 return; 344 } 345 346 // Expand the value into multiple parts. 347 if (NumParts & (NumParts - 1)) { 348 // The number of parts is not a power of 2. Split off and copy the tail. 349 assert(PartVT.isInteger() && ValueVT.isInteger() && 350 "Do not know what to expand to!"); 351 unsigned RoundParts = 1 << Log2_32(NumParts); 352 unsigned RoundBits = RoundParts * PartBits; 353 unsigned OddParts = NumParts - RoundParts; 354 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val, 355 DAG.getIntPtrConstant(RoundBits)); 356 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT); 357 358 if (TLI.isBigEndian()) 359 // The odd parts were reversed by getCopyToParts - unreverse them. 360 std::reverse(Parts + RoundParts, Parts + NumParts); 361 362 NumParts = RoundParts; 363 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); 364 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); 365 } 366 367 // The number of parts is a power of 2. Repeatedly bisect the value using 368 // EXTRACT_ELEMENT. 369 Parts[0] = DAG.getNode(ISD::BIT_CONVERT, DL, 370 EVT::getIntegerVT(*DAG.getContext(), 371 ValueVT.getSizeInBits()), 372 Val); 373 374 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) { 375 for (unsigned i = 0; i < NumParts; i += StepSize) { 376 unsigned ThisBits = StepSize * PartBits / 2; 377 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits); 378 SDValue &Part0 = Parts[i]; 379 SDValue &Part1 = Parts[i+StepSize/2]; 380 381 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, 382 ThisVT, Part0, DAG.getIntPtrConstant(1)); 383 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, 384 ThisVT, Part0, DAG.getIntPtrConstant(0)); 385 386 if (ThisBits == PartBits && ThisVT != PartVT) { 387 Part0 = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Part0); 388 Part1 = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Part1); 389 } 390 } 391 } 392 393 if (TLI.isBigEndian()) 394 std::reverse(Parts, Parts + OrigNumParts); 395} 396 397 398/// getCopyToPartsVector - Create a series of nodes that contain the specified 399/// value split into legal parts. 400static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc DL, 401 SDValue Val, SDValue *Parts, unsigned NumParts, 402 EVT PartVT) { 403 EVT ValueVT = Val.getValueType(); 404 assert(ValueVT.isVector() && "Not a vector"); 405 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 406 407 if (NumParts == 1) { 408 if (PartVT == ValueVT) { 409 // Nothing to do. 410 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) { 411 // Bitconvert vector->vector case. 412 Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val); 413 } else if (PartVT.isVector() && 414 PartVT.getVectorElementType() == ValueVT.getVectorElementType()&& 415 PartVT.getVectorNumElements() > ValueVT.getVectorNumElements()) { 416 EVT ElementVT = PartVT.getVectorElementType(); 417 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in 418 // undef elements. 419 SmallVector<SDValue, 16> Ops; 420 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i) 421 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, 422 ElementVT, Val, DAG.getIntPtrConstant(i))); 423 424 for (unsigned i = ValueVT.getVectorNumElements(), 425 e = PartVT.getVectorNumElements(); i != e; ++i) 426 Ops.push_back(DAG.getUNDEF(ElementVT)); 427 428 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size()); 429 430 // FIXME: Use CONCAT for 2x -> 4x. 431 432 //SDValue UndefElts = DAG.getUNDEF(VectorTy); 433 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts); 434 } else { 435 // Vector -> scalar conversion. 436 assert(ValueVT.getVectorElementType() == PartVT && 437 ValueVT.getVectorNumElements() == 1 && 438 "Only trivial vector-to-scalar conversions should get here!"); 439 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, 440 PartVT, Val, DAG.getIntPtrConstant(0)); 441 } 442 443 Parts[0] = Val; 444 return; 445 } 446 447 // Handle a multi-element vector. 448 EVT IntermediateVT, RegisterVT; 449 unsigned NumIntermediates; 450 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, 451 IntermediateVT, 452 NumIntermediates, RegisterVT); 453 unsigned NumElements = ValueVT.getVectorNumElements(); 454 455 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!"); 456 NumParts = NumRegs; // Silence a compiler warning. 457 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!"); 458 459 // Split the vector into intermediate operands. 460 SmallVector<SDValue, 8> Ops(NumIntermediates); 461 for (unsigned i = 0; i != NumIntermediates; ++i) { 462 if (IntermediateVT.isVector()) 463 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, 464 IntermediateVT, Val, 465 DAG.getIntPtrConstant(i * (NumElements / NumIntermediates))); 466 else 467 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, 468 IntermediateVT, Val, DAG.getIntPtrConstant(i)); 469 } 470 471 // Split the intermediate operands into legal parts. 472 if (NumParts == NumIntermediates) { 473 // If the register was not expanded, promote or copy the value, 474 // as appropriate. 475 for (unsigned i = 0; i != NumParts; ++i) 476 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT); 477 } else if (NumParts > 0) { 478 // If the intermediate type was expanded, split each the value into 479 // legal parts. 480 assert(NumParts % NumIntermediates == 0 && 481 "Must expand into a divisible number of parts!"); 482 unsigned Factor = NumParts / NumIntermediates; 483 for (unsigned i = 0; i != NumIntermediates; ++i) 484 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT); 485 } 486} 487 488 489 490 491namespace { 492 /// RegsForValue - This struct represents the registers (physical or virtual) 493 /// that a particular set of values is assigned, and the type information 494 /// about the value. The most common situation is to represent one value at a 495 /// time, but struct or array values are handled element-wise as multiple 496 /// values. The splitting of aggregates is performed recursively, so that we 497 /// never have aggregate-typed registers. The values at this point do not 498 /// necessarily have legal types, so each value may require one or more 499 /// registers of some legal type. 500 /// 501 struct RegsForValue { 502 /// ValueVTs - The value types of the values, which may not be legal, and 503 /// may need be promoted or synthesized from one or more registers. 504 /// 505 SmallVector<EVT, 4> ValueVTs; 506 507 /// RegVTs - The value types of the registers. This is the same size as 508 /// ValueVTs and it records, for each value, what the type of the assigned 509 /// register or registers are. (Individual values are never synthesized 510 /// from more than one type of register.) 511 /// 512 /// With virtual registers, the contents of RegVTs is redundant with TLI's 513 /// getRegisterType member function, however when with physical registers 514 /// it is necessary to have a separate record of the types. 515 /// 516 SmallVector<EVT, 4> RegVTs; 517 518 /// Regs - This list holds the registers assigned to the values. 519 /// Each legal or promoted value requires one register, and each 520 /// expanded value requires multiple registers. 521 /// 522 SmallVector<unsigned, 4> Regs; 523 524 RegsForValue() {} 525 526 RegsForValue(const SmallVector<unsigned, 4> ®s, 527 EVT regvt, EVT valuevt) 528 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {} 529 530 RegsForValue(LLVMContext &Context, const TargetLowering &tli, 531 unsigned Reg, const Type *Ty) { 532 ComputeValueVTs(tli, Ty, ValueVTs); 533 534 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) { 535 EVT ValueVT = ValueVTs[Value]; 536 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT); 537 EVT RegisterVT = tli.getRegisterType(Context, ValueVT); 538 for (unsigned i = 0; i != NumRegs; ++i) 539 Regs.push_back(Reg + i); 540 RegVTs.push_back(RegisterVT); 541 Reg += NumRegs; 542 } 543 } 544 545 /// areValueTypesLegal - Return true if types of all the values are legal. 546 bool areValueTypesLegal(const TargetLowering &TLI) { 547 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) { 548 EVT RegisterVT = RegVTs[Value]; 549 if (!TLI.isTypeLegal(RegisterVT)) 550 return false; 551 } 552 return true; 553 } 554 555 /// append - Add the specified values to this one. 556 void append(const RegsForValue &RHS) { 557 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end()); 558 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end()); 559 Regs.append(RHS.Regs.begin(), RHS.Regs.end()); 560 } 561 562 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from 563 /// this value and returns the result as a ValueVTs value. This uses 564 /// Chain/Flag as the input and updates them for the output Chain/Flag. 565 /// If the Flag pointer is NULL, no flag is used. 566 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo, 567 DebugLoc dl, 568 SDValue &Chain, SDValue *Flag) const; 569 570 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the 571 /// specified value into the registers specified by this object. This uses 572 /// Chain/Flag as the input and updates them for the output Chain/Flag. 573 /// If the Flag pointer is NULL, no flag is used. 574 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl, 575 SDValue &Chain, SDValue *Flag) const; 576 577 /// AddInlineAsmOperands - Add this value to the specified inlineasm node 578 /// operand list. This adds the code marker, matching input operand index 579 /// (if applicable), and includes the number of values added into it. 580 void AddInlineAsmOperands(unsigned Kind, 581 bool HasMatching, unsigned MatchingIdx, 582 SelectionDAG &DAG, 583 std::vector<SDValue> &Ops) const; 584 }; 585} 586 587/// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from 588/// this value and returns the result as a ValueVT value. This uses 589/// Chain/Flag as the input and updates them for the output Chain/Flag. 590/// If the Flag pointer is NULL, no flag is used. 591SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG, 592 FunctionLoweringInfo &FuncInfo, 593 DebugLoc dl, 594 SDValue &Chain, SDValue *Flag) const { 595 // A Value with type {} or [0 x %t] needs no registers. 596 if (ValueVTs.empty()) 597 return SDValue(); 598 599 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 600 601 // Assemble the legal parts into the final values. 602 SmallVector<SDValue, 4> Values(ValueVTs.size()); 603 SmallVector<SDValue, 8> Parts; 604 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) { 605 // Copy the legal parts from the registers. 606 EVT ValueVT = ValueVTs[Value]; 607 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT); 608 EVT RegisterVT = RegVTs[Value]; 609 610 Parts.resize(NumRegs); 611 for (unsigned i = 0; i != NumRegs; ++i) { 612 SDValue P; 613 if (Flag == 0) { 614 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT); 615 } else { 616 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag); 617 *Flag = P.getValue(2); 618 } 619 620 Chain = P.getValue(1); 621 622 // If the source register was virtual and if we know something about it, 623 // add an assert node. 624 if (TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) && 625 RegisterVT.isInteger() && !RegisterVT.isVector()) { 626 unsigned SlotNo = Regs[Part+i]-TargetRegisterInfo::FirstVirtualRegister; 627 if (FuncInfo.LiveOutRegInfo.size() > SlotNo) { 628 const FunctionLoweringInfo::LiveOutInfo &LOI = 629 FuncInfo.LiveOutRegInfo[SlotNo]; 630 631 unsigned RegSize = RegisterVT.getSizeInBits(); 632 unsigned NumSignBits = LOI.NumSignBits; 633 unsigned NumZeroBits = LOI.KnownZero.countLeadingOnes(); 634 635 // FIXME: We capture more information than the dag can represent. For 636 // now, just use the tightest assertzext/assertsext possible. 637 bool isSExt = true; 638 EVT FromVT(MVT::Other); 639 if (NumSignBits == RegSize) 640 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1 641 else if (NumZeroBits >= RegSize-1) 642 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1 643 else if (NumSignBits > RegSize-8) 644 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8 645 else if (NumZeroBits >= RegSize-8) 646 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8 647 else if (NumSignBits > RegSize-16) 648 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16 649 else if (NumZeroBits >= RegSize-16) 650 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16 651 else if (NumSignBits > RegSize-32) 652 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32 653 else if (NumZeroBits >= RegSize-32) 654 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32 655 656 if (FromVT != MVT::Other) 657 P = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl, 658 RegisterVT, P, DAG.getValueType(FromVT)); 659 } 660 } 661 662 Parts[i] = P; 663 } 664 665 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(), 666 NumRegs, RegisterVT, ValueVT); 667 Part += NumRegs; 668 Parts.clear(); 669 } 670 671 return DAG.getNode(ISD::MERGE_VALUES, dl, 672 DAG.getVTList(&ValueVTs[0], ValueVTs.size()), 673 &Values[0], ValueVTs.size()); 674} 675 676/// getCopyToRegs - Emit a series of CopyToReg nodes that copies the 677/// specified value into the registers specified by this object. This uses 678/// Chain/Flag as the input and updates them for the output Chain/Flag. 679/// If the Flag pointer is NULL, no flag is used. 680void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl, 681 SDValue &Chain, SDValue *Flag) const { 682 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 683 684 // Get the list of the values's legal parts. 685 unsigned NumRegs = Regs.size(); 686 SmallVector<SDValue, 8> Parts(NumRegs); 687 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) { 688 EVT ValueVT = ValueVTs[Value]; 689 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT); 690 EVT RegisterVT = RegVTs[Value]; 691 692 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value), 693 &Parts[Part], NumParts, RegisterVT); 694 Part += NumParts; 695 } 696 697 // Copy the parts into the registers. 698 SmallVector<SDValue, 8> Chains(NumRegs); 699 for (unsigned i = 0; i != NumRegs; ++i) { 700 SDValue Part; 701 if (Flag == 0) { 702 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]); 703 } else { 704 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag); 705 *Flag = Part.getValue(1); 706 } 707 708 Chains[i] = Part.getValue(0); 709 } 710 711 if (NumRegs == 1 || Flag) 712 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is 713 // flagged to it. That is the CopyToReg nodes and the user are considered 714 // a single scheduling unit. If we create a TokenFactor and return it as 715 // chain, then the TokenFactor is both a predecessor (operand) of the 716 // user as well as a successor (the TF operands are flagged to the user). 717 // c1, f1 = CopyToReg 718 // c2, f2 = CopyToReg 719 // c3 = TokenFactor c1, c2 720 // ... 721 // = op c3, ..., f2 722 Chain = Chains[NumRegs-1]; 723 else 724 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs); 725} 726 727/// AddInlineAsmOperands - Add this value to the specified inlineasm node 728/// operand list. This adds the code marker and includes the number of 729/// values added into it. 730void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching, 731 unsigned MatchingIdx, 732 SelectionDAG &DAG, 733 std::vector<SDValue> &Ops) const { 734 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 735 736 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size()); 737 if (HasMatching) 738 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx); 739 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32); 740 Ops.push_back(Res); 741 742 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) { 743 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]); 744 EVT RegisterVT = RegVTs[Value]; 745 for (unsigned i = 0; i != NumRegs; ++i) { 746 assert(Reg < Regs.size() && "Mismatch in # registers expected"); 747 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT)); 748 } 749 } 750} 751 752void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa) { 753 AA = &aa; 754 GFI = gfi; 755 TD = DAG.getTarget().getTargetData(); 756} 757 758/// clear - Clear out the current SelectionDAG and the associated 759/// state and prepare this SelectionDAGBuilder object to be used 760/// for a new block. This doesn't clear out information about 761/// additional blocks that are needed to complete switch lowering 762/// or PHI node updating; that information is cleared out as it is 763/// consumed. 764void SelectionDAGBuilder::clear() { 765 NodeMap.clear(); 766 UnusedArgNodeMap.clear(); 767 PendingLoads.clear(); 768 PendingExports.clear(); 769 DanglingDebugInfoMap.clear(); 770 CurDebugLoc = DebugLoc(); 771 HasTailCall = false; 772} 773 774/// getRoot - Return the current virtual root of the Selection DAG, 775/// flushing any PendingLoad items. This must be done before emitting 776/// a store or any other node that may need to be ordered after any 777/// prior load instructions. 778/// 779SDValue SelectionDAGBuilder::getRoot() { 780 if (PendingLoads.empty()) 781 return DAG.getRoot(); 782 783 if (PendingLoads.size() == 1) { 784 SDValue Root = PendingLoads[0]; 785 DAG.setRoot(Root); 786 PendingLoads.clear(); 787 return Root; 788 } 789 790 // Otherwise, we have to make a token factor node. 791 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other, 792 &PendingLoads[0], PendingLoads.size()); 793 PendingLoads.clear(); 794 DAG.setRoot(Root); 795 return Root; 796} 797 798/// getControlRoot - Similar to getRoot, but instead of flushing all the 799/// PendingLoad items, flush all the PendingExports items. It is necessary 800/// to do this before emitting a terminator instruction. 801/// 802SDValue SelectionDAGBuilder::getControlRoot() { 803 SDValue Root = DAG.getRoot(); 804 805 if (PendingExports.empty()) 806 return Root; 807 808 // Turn all of the CopyToReg chains into one factored node. 809 if (Root.getOpcode() != ISD::EntryToken) { 810 unsigned i = 0, e = PendingExports.size(); 811 for (; i != e; ++i) { 812 assert(PendingExports[i].getNode()->getNumOperands() > 1); 813 if (PendingExports[i].getNode()->getOperand(0) == Root) 814 break; // Don't add the root if we already indirectly depend on it. 815 } 816 817 if (i == e) 818 PendingExports.push_back(Root); 819 } 820 821 Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other, 822 &PendingExports[0], 823 PendingExports.size()); 824 PendingExports.clear(); 825 DAG.setRoot(Root); 826 return Root; 827} 828 829void SelectionDAGBuilder::AssignOrderingToNode(const SDNode *Node) { 830 if (DAG.GetOrdering(Node) != 0) return; // Already has ordering. 831 DAG.AssignOrdering(Node, SDNodeOrder); 832 833 for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I) 834 AssignOrderingToNode(Node->getOperand(I).getNode()); 835} 836 837void SelectionDAGBuilder::visit(const Instruction &I) { 838 // Set up outgoing PHI node register values before emitting the terminator. 839 if (isa<TerminatorInst>(&I)) 840 HandlePHINodesInSuccessorBlocks(I.getParent()); 841 842 CurDebugLoc = I.getDebugLoc(); 843 844 visit(I.getOpcode(), I); 845 846 if (!isa<TerminatorInst>(&I) && !HasTailCall) 847 CopyToExportRegsIfNeeded(&I); 848 849 CurDebugLoc = DebugLoc(); 850} 851 852void SelectionDAGBuilder::visitPHI(const PHINode &) { 853 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!"); 854} 855 856void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) { 857 // Note: this doesn't use InstVisitor, because it has to work with 858 // ConstantExpr's in addition to instructions. 859 switch (Opcode) { 860 default: llvm_unreachable("Unknown instruction type encountered!"); 861 // Build the switch statement using the Instruction.def file. 862#define HANDLE_INST(NUM, OPCODE, CLASS) \ 863 case Instruction::OPCODE: visit##OPCODE((CLASS&)I); break; 864#include "llvm/Instruction.def" 865 } 866 867 // Assign the ordering to the freshly created DAG nodes. 868 if (NodeMap.count(&I)) { 869 ++SDNodeOrder; 870 AssignOrderingToNode(getValue(&I).getNode()); 871 } 872} 873 874// resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V, 875// generate the debug data structures now that we've seen its definition. 876void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V, 877 SDValue Val) { 878 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V]; 879 MDNode *Variable = NULL; 880 uint64_t Offset = 0; 881 882 if (const DbgValueInst *DI = dyn_cast_or_null<DbgValueInst>(DDI.getDI())) { 883 Variable = DI->getVariable(); 884 Offset = DI->getOffset(); 885 } else if (const DbgDeclareInst *DI = 886 dyn_cast_or_null<DbgDeclareInst>(DDI.getDI())) 887 Variable = DI->getVariable(); 888 else { 889 assert (DDI.getDI() == NULL && "Invalid debug info intrinsic!"); 890 return; 891 } 892 893 if (Variable) { 894 DebugLoc dl = DDI.getdl(); 895 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder(); 896 SDDbgValue *SDV; 897 if (Val.getNode()) { 898 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) { 899 SDV = DAG.getDbgValue(Variable, Val.getNode(), 900 Val.getResNo(), Offset, dl, DbgSDNodeOrder); 901 DAG.AddDbgValue(SDV, Val.getNode(), false); 902 } 903 } else { 904 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()), 905 Offset, dl, SDNodeOrder); 906 DAG.AddDbgValue(SDV, 0, false); 907 } 908 DanglingDebugInfoMap[V] = DanglingDebugInfo(); 909 } 910} 911 912// getValue - Return an SDValue for the given Value. 913SDValue SelectionDAGBuilder::getValue(const Value *V) { 914 // If we already have an SDValue for this value, use it. It's important 915 // to do this first, so that we don't create a CopyFromReg if we already 916 // have a regular SDValue. 917 SDValue &N = NodeMap[V]; 918 if (N.getNode()) return N; 919 920 // If there's a virtual register allocated and initialized for this 921 // value, use it. 922 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V); 923 if (It != FuncInfo.ValueMap.end()) { 924 unsigned InReg = It->second; 925 RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType()); 926 SDValue Chain = DAG.getEntryNode(); 927 return N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain,NULL); 928 } 929 930 // Otherwise create a new SDValue and remember it. 931 SDValue Val = getValueImpl(V); 932 NodeMap[V] = Val; 933 resolveDanglingDebugInfo(V, Val); 934 return Val; 935} 936 937/// getNonRegisterValue - Return an SDValue for the given Value, but 938/// don't look in FuncInfo.ValueMap for a virtual register. 939SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) { 940 // If we already have an SDValue for this value, use it. 941 SDValue &N = NodeMap[V]; 942 if (N.getNode()) return N; 943 944 // Otherwise create a new SDValue and remember it. 945 SDValue Val = getValueImpl(V); 946 NodeMap[V] = Val; 947 resolveDanglingDebugInfo(V, Val); 948 return Val; 949} 950 951/// getValueImpl - Helper function for getValue and getNonRegisterValue. 952/// Create an SDValue for the given value. 953SDValue SelectionDAGBuilder::getValueImpl(const Value *V) { 954 if (const Constant *C = dyn_cast<Constant>(V)) { 955 EVT VT = TLI.getValueType(V->getType(), true); 956 957 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C)) 958 return DAG.getConstant(*CI, VT); 959 960 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C)) 961 return DAG.getGlobalAddress(GV, getCurDebugLoc(), VT); 962 963 if (isa<ConstantPointerNull>(C)) 964 return DAG.getConstant(0, TLI.getPointerTy()); 965 966 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C)) 967 return DAG.getConstantFP(*CFP, VT); 968 969 if (isa<UndefValue>(C) && !V->getType()->isAggregateType()) 970 return DAG.getUNDEF(VT); 971 972 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 973 visit(CE->getOpcode(), *CE); 974 SDValue N1 = NodeMap[V]; 975 assert(N1.getNode() && "visit didn't populate the NodeMap!"); 976 return N1; 977 } 978 979 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) { 980 SmallVector<SDValue, 4> Constants; 981 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end(); 982 OI != OE; ++OI) { 983 SDNode *Val = getValue(*OI).getNode(); 984 // If the operand is an empty aggregate, there are no values. 985 if (!Val) continue; 986 // Add each leaf value from the operand to the Constants list 987 // to form a flattened list of all the values. 988 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i) 989 Constants.push_back(SDValue(Val, i)); 990 } 991 992 return DAG.getMergeValues(&Constants[0], Constants.size(), 993 getCurDebugLoc()); 994 } 995 996 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) { 997 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) && 998 "Unknown struct or array constant!"); 999 1000 SmallVector<EVT, 4> ValueVTs; 1001 ComputeValueVTs(TLI, C->getType(), ValueVTs); 1002 unsigned NumElts = ValueVTs.size(); 1003 if (NumElts == 0) 1004 return SDValue(); // empty struct 1005 SmallVector<SDValue, 4> Constants(NumElts); 1006 for (unsigned i = 0; i != NumElts; ++i) { 1007 EVT EltVT = ValueVTs[i]; 1008 if (isa<UndefValue>(C)) 1009 Constants[i] = DAG.getUNDEF(EltVT); 1010 else if (EltVT.isFloatingPoint()) 1011 Constants[i] = DAG.getConstantFP(0, EltVT); 1012 else 1013 Constants[i] = DAG.getConstant(0, EltVT); 1014 } 1015 1016 return DAG.getMergeValues(&Constants[0], NumElts, 1017 getCurDebugLoc()); 1018 } 1019 1020 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C)) 1021 return DAG.getBlockAddress(BA, VT); 1022 1023 const VectorType *VecTy = cast<VectorType>(V->getType()); 1024 unsigned NumElements = VecTy->getNumElements(); 1025 1026 // Now that we know the number and type of the elements, get that number of 1027 // elements into the Ops array based on what kind of constant it is. 1028 SmallVector<SDValue, 16> Ops; 1029 if (const ConstantVector *CP = dyn_cast<ConstantVector>(C)) { 1030 for (unsigned i = 0; i != NumElements; ++i) 1031 Ops.push_back(getValue(CP->getOperand(i))); 1032 } else { 1033 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!"); 1034 EVT EltVT = TLI.getValueType(VecTy->getElementType()); 1035 1036 SDValue Op; 1037 if (EltVT.isFloatingPoint()) 1038 Op = DAG.getConstantFP(0, EltVT); 1039 else 1040 Op = DAG.getConstant(0, EltVT); 1041 Ops.assign(NumElements, Op); 1042 } 1043 1044 // Create a BUILD_VECTOR node. 1045 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(), 1046 VT, &Ops[0], Ops.size()); 1047 } 1048 1049 // If this is a static alloca, generate it as the frameindex instead of 1050 // computation. 1051 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) { 1052 DenseMap<const AllocaInst*, int>::iterator SI = 1053 FuncInfo.StaticAllocaMap.find(AI); 1054 if (SI != FuncInfo.StaticAllocaMap.end()) 1055 return DAG.getFrameIndex(SI->second, TLI.getPointerTy()); 1056 } 1057 1058 // If this is an instruction which fast-isel has deferred, select it now. 1059 if (const Instruction *Inst = dyn_cast<Instruction>(V)) { 1060 unsigned InReg = FuncInfo.InitializeRegForValue(Inst); 1061 RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType()); 1062 SDValue Chain = DAG.getEntryNode(); 1063 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL); 1064 } 1065 1066 llvm_unreachable("Can't get register for value!"); 1067 return SDValue(); 1068} 1069 1070void SelectionDAGBuilder::visitRet(const ReturnInst &I) { 1071 SDValue Chain = getControlRoot(); 1072 SmallVector<ISD::OutputArg, 8> Outs; 1073 SmallVector<SDValue, 8> OutVals; 1074 1075 if (!FuncInfo.CanLowerReturn) { 1076 unsigned DemoteReg = FuncInfo.DemoteRegister; 1077 const Function *F = I.getParent()->getParent(); 1078 1079 // Emit a store of the return value through the virtual register. 1080 // Leave Outs empty so that LowerReturn won't try to load return 1081 // registers the usual way. 1082 SmallVector<EVT, 1> PtrValueVTs; 1083 ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()), 1084 PtrValueVTs); 1085 1086 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]); 1087 SDValue RetOp = getValue(I.getOperand(0)); 1088 1089 SmallVector<EVT, 4> ValueVTs; 1090 SmallVector<uint64_t, 4> Offsets; 1091 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets); 1092 unsigned NumValues = ValueVTs.size(); 1093 1094 SmallVector<SDValue, 4> Chains(NumValues); 1095 for (unsigned i = 0; i != NumValues; ++i) { 1096 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), 1097 RetPtr.getValueType(), RetPtr, 1098 DAG.getIntPtrConstant(Offsets[i])); 1099 Chains[i] = 1100 DAG.getStore(Chain, getCurDebugLoc(), 1101 SDValue(RetOp.getNode(), RetOp.getResNo() + i), 1102 Add, NULL, Offsets[i], false, false, 0); 1103 } 1104 1105 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), 1106 MVT::Other, &Chains[0], NumValues); 1107 } else if (I.getNumOperands() != 0) { 1108 SmallVector<EVT, 4> ValueVTs; 1109 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs); 1110 unsigned NumValues = ValueVTs.size(); 1111 if (NumValues) { 1112 SDValue RetOp = getValue(I.getOperand(0)); 1113 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1114 EVT VT = ValueVTs[j]; 1115 1116 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1117 1118 const Function *F = I.getParent()->getParent(); 1119 if (F->paramHasAttr(0, Attribute::SExt)) 1120 ExtendKind = ISD::SIGN_EXTEND; 1121 else if (F->paramHasAttr(0, Attribute::ZExt)) 1122 ExtendKind = ISD::ZERO_EXTEND; 1123 1124 // FIXME: C calling convention requires the return type to be promoted 1125 // to at least 32-bit. But this is not necessary for non-C calling 1126 // conventions. The frontend should mark functions whose return values 1127 // require promoting with signext or zeroext attributes. 1128 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1129 EVT MinVT = TLI.getRegisterType(*DAG.getContext(), MVT::i32); 1130 if (VT.bitsLT(MinVT)) 1131 VT = MinVT; 1132 } 1133 1134 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT); 1135 EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT); 1136 SmallVector<SDValue, 4> Parts(NumParts); 1137 getCopyToParts(DAG, getCurDebugLoc(), 1138 SDValue(RetOp.getNode(), RetOp.getResNo() + j), 1139 &Parts[0], NumParts, PartVT, ExtendKind); 1140 1141 // 'inreg' on function refers to return value 1142 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1143 if (F->paramHasAttr(0, Attribute::InReg)) 1144 Flags.setInReg(); 1145 1146 // Propagate extension type if any 1147 if (F->paramHasAttr(0, Attribute::SExt)) 1148 Flags.setSExt(); 1149 else if (F->paramHasAttr(0, Attribute::ZExt)) 1150 Flags.setZExt(); 1151 1152 for (unsigned i = 0; i < NumParts; ++i) { 1153 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(), 1154 /*isfixed=*/true)); 1155 OutVals.push_back(Parts[i]); 1156 } 1157 } 1158 } 1159 } 1160 1161 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg(); 1162 CallingConv::ID CallConv = 1163 DAG.getMachineFunction().getFunction()->getCallingConv(); 1164 Chain = TLI.LowerReturn(Chain, CallConv, isVarArg, 1165 Outs, OutVals, getCurDebugLoc(), DAG); 1166 1167 // Verify that the target's LowerReturn behaved as expected. 1168 assert(Chain.getNode() && Chain.getValueType() == MVT::Other && 1169 "LowerReturn didn't return a valid chain!"); 1170 1171 // Update the DAG with the new chain value resulting from return lowering. 1172 DAG.setRoot(Chain); 1173} 1174 1175/// CopyToExportRegsIfNeeded - If the given value has virtual registers 1176/// created for it, emit nodes to copy the value into the virtual 1177/// registers. 1178void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) { 1179 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V); 1180 if (VMI != FuncInfo.ValueMap.end()) { 1181 assert(!V->use_empty() && "Unused value assigned virtual registers!"); 1182 CopyValueToVirtualRegister(V, VMI->second); 1183 } 1184} 1185 1186/// ExportFromCurrentBlock - If this condition isn't known to be exported from 1187/// the current basic block, add it to ValueMap now so that we'll get a 1188/// CopyTo/FromReg. 1189void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) { 1190 // No need to export constants. 1191 if (!isa<Instruction>(V) && !isa<Argument>(V)) return; 1192 1193 // Already exported? 1194 if (FuncInfo.isExportedInst(V)) return; 1195 1196 unsigned Reg = FuncInfo.InitializeRegForValue(V); 1197 CopyValueToVirtualRegister(V, Reg); 1198} 1199 1200bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V, 1201 const BasicBlock *FromBB) { 1202 // The operands of the setcc have to be in this block. We don't know 1203 // how to export them from some other block. 1204 if (const Instruction *VI = dyn_cast<Instruction>(V)) { 1205 // Can export from current BB. 1206 if (VI->getParent() == FromBB) 1207 return true; 1208 1209 // Is already exported, noop. 1210 return FuncInfo.isExportedInst(V); 1211 } 1212 1213 // If this is an argument, we can export it if the BB is the entry block or 1214 // if it is already exported. 1215 if (isa<Argument>(V)) { 1216 if (FromBB == &FromBB->getParent()->getEntryBlock()) 1217 return true; 1218 1219 // Otherwise, can only export this if it is already exported. 1220 return FuncInfo.isExportedInst(V); 1221 } 1222 1223 // Otherwise, constants can always be exported. 1224 return true; 1225} 1226 1227static bool InBlock(const Value *V, const BasicBlock *BB) { 1228 if (const Instruction *I = dyn_cast<Instruction>(V)) 1229 return I->getParent() == BB; 1230 return true; 1231} 1232 1233/// EmitBranchForMergedCondition - Helper method for FindMergedConditions. 1234/// This function emits a branch and is used at the leaves of an OR or an 1235/// AND operator tree. 1236/// 1237void 1238SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond, 1239 MachineBasicBlock *TBB, 1240 MachineBasicBlock *FBB, 1241 MachineBasicBlock *CurBB, 1242 MachineBasicBlock *SwitchBB) { 1243 const BasicBlock *BB = CurBB->getBasicBlock(); 1244 1245 // If the leaf of the tree is a comparison, merge the condition into 1246 // the caseblock. 1247 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) { 1248 // The operands of the cmp have to be in this block. We don't know 1249 // how to export them from some other block. If this is the first block 1250 // of the sequence, no exporting is needed. 1251 if (CurBB == SwitchBB || 1252 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) && 1253 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) { 1254 ISD::CondCode Condition; 1255 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) { 1256 Condition = getICmpCondCode(IC->getPredicate()); 1257 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) { 1258 Condition = getFCmpCondCode(FC->getPredicate()); 1259 } else { 1260 Condition = ISD::SETEQ; // silence warning. 1261 llvm_unreachable("Unknown compare instruction"); 1262 } 1263 1264 CaseBlock CB(Condition, BOp->getOperand(0), 1265 BOp->getOperand(1), NULL, TBB, FBB, CurBB); 1266 SwitchCases.push_back(CB); 1267 return; 1268 } 1269 } 1270 1271 // Create a CaseBlock record representing this branch. 1272 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()), 1273 NULL, TBB, FBB, CurBB); 1274 SwitchCases.push_back(CB); 1275} 1276 1277/// FindMergedConditions - If Cond is an expression like 1278void SelectionDAGBuilder::FindMergedConditions(const Value *Cond, 1279 MachineBasicBlock *TBB, 1280 MachineBasicBlock *FBB, 1281 MachineBasicBlock *CurBB, 1282 MachineBasicBlock *SwitchBB, 1283 unsigned Opc) { 1284 // If this node is not part of the or/and tree, emit it as a branch. 1285 const Instruction *BOp = dyn_cast<Instruction>(Cond); 1286 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) || 1287 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() || 1288 BOp->getParent() != CurBB->getBasicBlock() || 1289 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) || 1290 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) { 1291 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB); 1292 return; 1293 } 1294 1295 // Create TmpBB after CurBB. 1296 MachineFunction::iterator BBI = CurBB; 1297 MachineFunction &MF = DAG.getMachineFunction(); 1298 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock()); 1299 CurBB->getParent()->insert(++BBI, TmpBB); 1300 1301 if (Opc == Instruction::Or) { 1302 // Codegen X | Y as: 1303 // jmp_if_X TBB 1304 // jmp TmpBB 1305 // TmpBB: 1306 // jmp_if_Y TBB 1307 // jmp FBB 1308 // 1309 1310 // Emit the LHS condition. 1311 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc); 1312 1313 // Emit the RHS condition into TmpBB. 1314 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc); 1315 } else { 1316 assert(Opc == Instruction::And && "Unknown merge op!"); 1317 // Codegen X & Y as: 1318 // jmp_if_X TmpBB 1319 // jmp FBB 1320 // TmpBB: 1321 // jmp_if_Y TBB 1322 // jmp FBB 1323 // 1324 // This requires creation of TmpBB after CurBB. 1325 1326 // Emit the LHS condition. 1327 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc); 1328 1329 // Emit the RHS condition into TmpBB. 1330 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc); 1331 } 1332} 1333 1334/// If the set of cases should be emitted as a series of branches, return true. 1335/// If we should emit this as a bunch of and/or'd together conditions, return 1336/// false. 1337bool 1338SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){ 1339 if (Cases.size() != 2) return true; 1340 1341 // If this is two comparisons of the same values or'd or and'd together, they 1342 // will get folded into a single comparison, so don't emit two blocks. 1343 if ((Cases[0].CmpLHS == Cases[1].CmpLHS && 1344 Cases[0].CmpRHS == Cases[1].CmpRHS) || 1345 (Cases[0].CmpRHS == Cases[1].CmpLHS && 1346 Cases[0].CmpLHS == Cases[1].CmpRHS)) { 1347 return false; 1348 } 1349 1350 // Handle: (X != null) | (Y != null) --> (X|Y) != 0 1351 // Handle: (X == null) & (Y == null) --> (X|Y) == 0 1352 if (Cases[0].CmpRHS == Cases[1].CmpRHS && 1353 Cases[0].CC == Cases[1].CC && 1354 isa<Constant>(Cases[0].CmpRHS) && 1355 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) { 1356 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB) 1357 return false; 1358 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB) 1359 return false; 1360 } 1361 1362 return true; 1363} 1364 1365void SelectionDAGBuilder::visitBr(const BranchInst &I) { 1366 MachineBasicBlock *BrMBB = FuncInfo.MBB; 1367 1368 // Update machine-CFG edges. 1369 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)]; 1370 1371 // Figure out which block is immediately after the current one. 1372 MachineBasicBlock *NextBlock = 0; 1373 MachineFunction::iterator BBI = BrMBB; 1374 if (++BBI != FuncInfo.MF->end()) 1375 NextBlock = BBI; 1376 1377 if (I.isUnconditional()) { 1378 // Update machine-CFG edges. 1379 BrMBB->addSuccessor(Succ0MBB); 1380 1381 // If this is not a fall-through branch, emit the branch. 1382 if (Succ0MBB != NextBlock) 1383 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(), 1384 MVT::Other, getControlRoot(), 1385 DAG.getBasicBlock(Succ0MBB))); 1386 1387 return; 1388 } 1389 1390 // If this condition is one of the special cases we handle, do special stuff 1391 // now. 1392 const Value *CondVal = I.getCondition(); 1393 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)]; 1394 1395 // If this is a series of conditions that are or'd or and'd together, emit 1396 // this as a sequence of branches instead of setcc's with and/or operations. 1397 // For example, instead of something like: 1398 // cmp A, B 1399 // C = seteq 1400 // cmp D, E 1401 // F = setle 1402 // or C, F 1403 // jnz foo 1404 // Emit: 1405 // cmp A, B 1406 // je foo 1407 // cmp D, E 1408 // jle foo 1409 // 1410 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) { 1411 if (BOp->hasOneUse() && 1412 (BOp->getOpcode() == Instruction::And || 1413 BOp->getOpcode() == Instruction::Or)) { 1414 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB, 1415 BOp->getOpcode()); 1416 // If the compares in later blocks need to use values not currently 1417 // exported from this block, export them now. This block should always 1418 // be the first entry. 1419 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!"); 1420 1421 // Allow some cases to be rejected. 1422 if (ShouldEmitAsBranches(SwitchCases)) { 1423 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) { 1424 ExportFromCurrentBlock(SwitchCases[i].CmpLHS); 1425 ExportFromCurrentBlock(SwitchCases[i].CmpRHS); 1426 } 1427 1428 // Emit the branch for this block. 1429 visitSwitchCase(SwitchCases[0], BrMBB); 1430 SwitchCases.erase(SwitchCases.begin()); 1431 return; 1432 } 1433 1434 // Okay, we decided not to do this, remove any inserted MBB's and clear 1435 // SwitchCases. 1436 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) 1437 FuncInfo.MF->erase(SwitchCases[i].ThisBB); 1438 1439 SwitchCases.clear(); 1440 } 1441 } 1442 1443 // Create a CaseBlock record representing this branch. 1444 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()), 1445 NULL, Succ0MBB, Succ1MBB, BrMBB); 1446 1447 // Use visitSwitchCase to actually insert the fast branch sequence for this 1448 // cond branch. 1449 visitSwitchCase(CB, BrMBB); 1450} 1451 1452/// visitSwitchCase - Emits the necessary code to represent a single node in 1453/// the binary search tree resulting from lowering a switch instruction. 1454void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB, 1455 MachineBasicBlock *SwitchBB) { 1456 SDValue Cond; 1457 SDValue CondLHS = getValue(CB.CmpLHS); 1458 DebugLoc dl = getCurDebugLoc(); 1459 1460 // Build the setcc now. 1461 if (CB.CmpMHS == NULL) { 1462 // Fold "(X == true)" to X and "(X == false)" to !X to 1463 // handle common cases produced by branch lowering. 1464 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) && 1465 CB.CC == ISD::SETEQ) 1466 Cond = CondLHS; 1467 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) && 1468 CB.CC == ISD::SETEQ) { 1469 SDValue True = DAG.getConstant(1, CondLHS.getValueType()); 1470 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True); 1471 } else 1472 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC); 1473 } else { 1474 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now"); 1475 1476 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue(); 1477 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue(); 1478 1479 SDValue CmpOp = getValue(CB.CmpMHS); 1480 EVT VT = CmpOp.getValueType(); 1481 1482 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) { 1483 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT), 1484 ISD::SETLE); 1485 } else { 1486 SDValue SUB = DAG.getNode(ISD::SUB, dl, 1487 VT, CmpOp, DAG.getConstant(Low, VT)); 1488 Cond = DAG.getSetCC(dl, MVT::i1, SUB, 1489 DAG.getConstant(High-Low, VT), ISD::SETULE); 1490 } 1491 } 1492 1493 // Update successor info 1494 SwitchBB->addSuccessor(CB.TrueBB); 1495 SwitchBB->addSuccessor(CB.FalseBB); 1496 1497 // Set NextBlock to be the MBB immediately after the current one, if any. 1498 // This is used to avoid emitting unnecessary branches to the next block. 1499 MachineBasicBlock *NextBlock = 0; 1500 MachineFunction::iterator BBI = SwitchBB; 1501 if (++BBI != FuncInfo.MF->end()) 1502 NextBlock = BBI; 1503 1504 // If the lhs block is the next block, invert the condition so that we can 1505 // fall through to the lhs instead of the rhs block. 1506 if (CB.TrueBB == NextBlock) { 1507 std::swap(CB.TrueBB, CB.FalseBB); 1508 SDValue True = DAG.getConstant(1, Cond.getValueType()); 1509 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True); 1510 } 1511 1512 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl, 1513 MVT::Other, getControlRoot(), Cond, 1514 DAG.getBasicBlock(CB.TrueBB)); 1515 1516 // Insert the false branch. 1517 if (CB.FalseBB != NextBlock) 1518 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond, 1519 DAG.getBasicBlock(CB.FalseBB)); 1520 1521 DAG.setRoot(BrCond); 1522} 1523 1524/// visitJumpTable - Emit JumpTable node in the current MBB 1525void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) { 1526 // Emit the code for the jump table 1527 assert(JT.Reg != -1U && "Should lower JT Header first!"); 1528 EVT PTy = TLI.getPointerTy(); 1529 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), 1530 JT.Reg, PTy); 1531 SDValue Table = DAG.getJumpTable(JT.JTI, PTy); 1532 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurDebugLoc(), 1533 MVT::Other, Index.getValue(1), 1534 Table, Index); 1535 DAG.setRoot(BrJumpTable); 1536} 1537 1538/// visitJumpTableHeader - This function emits necessary code to produce index 1539/// in the JumpTable from switch case. 1540void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT, 1541 JumpTableHeader &JTH, 1542 MachineBasicBlock *SwitchBB) { 1543 // Subtract the lowest switch case value from the value being switched on and 1544 // conditional branch to default mbb if the result is greater than the 1545 // difference between smallest and largest cases. 1546 SDValue SwitchOp = getValue(JTH.SValue); 1547 EVT VT = SwitchOp.getValueType(); 1548 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp, 1549 DAG.getConstant(JTH.First, VT)); 1550 1551 // The SDNode we just created, which holds the value being switched on minus 1552 // the smallest case value, needs to be copied to a virtual register so it 1553 // can be used as an index into the jump table in a subsequent basic block. 1554 // This value may be smaller or larger than the target's pointer type, and 1555 // therefore require extension or truncating. 1556 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), TLI.getPointerTy()); 1557 1558 unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy()); 1559 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(), 1560 JumpTableReg, SwitchOp); 1561 JT.Reg = JumpTableReg; 1562 1563 // Emit the range check for the jump table, and branch to the default block 1564 // for the switch statement if the value being switched on exceeds the largest 1565 // case in the switch. 1566 SDValue CMP = DAG.getSetCC(getCurDebugLoc(), 1567 TLI.getSetCCResultType(Sub.getValueType()), Sub, 1568 DAG.getConstant(JTH.Last-JTH.First,VT), 1569 ISD::SETUGT); 1570 1571 // Set NextBlock to be the MBB immediately after the current one, if any. 1572 // This is used to avoid emitting unnecessary branches to the next block. 1573 MachineBasicBlock *NextBlock = 0; 1574 MachineFunction::iterator BBI = SwitchBB; 1575 1576 if (++BBI != FuncInfo.MF->end()) 1577 NextBlock = BBI; 1578 1579 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(), 1580 MVT::Other, CopyTo, CMP, 1581 DAG.getBasicBlock(JT.Default)); 1582 1583 if (JT.MBB != NextBlock) 1584 BrCond = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond, 1585 DAG.getBasicBlock(JT.MBB)); 1586 1587 DAG.setRoot(BrCond); 1588} 1589 1590/// visitBitTestHeader - This function emits necessary code to produce value 1591/// suitable for "bit tests" 1592void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B, 1593 MachineBasicBlock *SwitchBB) { 1594 // Subtract the minimum value 1595 SDValue SwitchOp = getValue(B.SValue); 1596 EVT VT = SwitchOp.getValueType(); 1597 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp, 1598 DAG.getConstant(B.First, VT)); 1599 1600 // Check range 1601 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(), 1602 TLI.getSetCCResultType(Sub.getValueType()), 1603 Sub, DAG.getConstant(B.Range, VT), 1604 ISD::SETUGT); 1605 1606 SDValue ShiftOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), 1607 TLI.getPointerTy()); 1608 1609 B.Reg = FuncInfo.CreateReg(TLI.getPointerTy()); 1610 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(), 1611 B.Reg, ShiftOp); 1612 1613 // Set NextBlock to be the MBB immediately after the current one, if any. 1614 // This is used to avoid emitting unnecessary branches to the next block. 1615 MachineBasicBlock *NextBlock = 0; 1616 MachineFunction::iterator BBI = SwitchBB; 1617 if (++BBI != FuncInfo.MF->end()) 1618 NextBlock = BBI; 1619 1620 MachineBasicBlock* MBB = B.Cases[0].ThisBB; 1621 1622 SwitchBB->addSuccessor(B.Default); 1623 SwitchBB->addSuccessor(MBB); 1624 1625 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(), 1626 MVT::Other, CopyTo, RangeCmp, 1627 DAG.getBasicBlock(B.Default)); 1628 1629 if (MBB != NextBlock) 1630 BrRange = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo, 1631 DAG.getBasicBlock(MBB)); 1632 1633 DAG.setRoot(BrRange); 1634} 1635 1636/// visitBitTestCase - this function produces one "bit test" 1637void SelectionDAGBuilder::visitBitTestCase(MachineBasicBlock* NextMBB, 1638 unsigned Reg, 1639 BitTestCase &B, 1640 MachineBasicBlock *SwitchBB) { 1641 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), Reg, 1642 TLI.getPointerTy()); 1643 SDValue Cmp; 1644 if (CountPopulation_64(B.Mask) == 1) { 1645 // Testing for a single bit; just compare the shift count with what it 1646 // would need to be to shift a 1 bit in that position. 1647 Cmp = DAG.getSetCC(getCurDebugLoc(), 1648 TLI.getSetCCResultType(ShiftOp.getValueType()), 1649 ShiftOp, 1650 DAG.getConstant(CountTrailingZeros_64(B.Mask), 1651 TLI.getPointerTy()), 1652 ISD::SETEQ); 1653 } else { 1654 // Make desired shift 1655 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(), 1656 TLI.getPointerTy(), 1657 DAG.getConstant(1, TLI.getPointerTy()), 1658 ShiftOp); 1659 1660 // Emit bit tests and jumps 1661 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(), 1662 TLI.getPointerTy(), SwitchVal, 1663 DAG.getConstant(B.Mask, TLI.getPointerTy())); 1664 Cmp = DAG.getSetCC(getCurDebugLoc(), 1665 TLI.getSetCCResultType(AndOp.getValueType()), 1666 AndOp, DAG.getConstant(0, TLI.getPointerTy()), 1667 ISD::SETNE); 1668 } 1669 1670 SwitchBB->addSuccessor(B.TargetBB); 1671 SwitchBB->addSuccessor(NextMBB); 1672 1673 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(), 1674 MVT::Other, getControlRoot(), 1675 Cmp, DAG.getBasicBlock(B.TargetBB)); 1676 1677 // Set NextBlock to be the MBB immediately after the current one, if any. 1678 // This is used to avoid emitting unnecessary branches to the next block. 1679 MachineBasicBlock *NextBlock = 0; 1680 MachineFunction::iterator BBI = SwitchBB; 1681 if (++BBI != FuncInfo.MF->end()) 1682 NextBlock = BBI; 1683 1684 if (NextMBB != NextBlock) 1685 BrAnd = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd, 1686 DAG.getBasicBlock(NextMBB)); 1687 1688 DAG.setRoot(BrAnd); 1689} 1690 1691void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) { 1692 MachineBasicBlock *InvokeMBB = FuncInfo.MBB; 1693 1694 // Retrieve successors. 1695 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)]; 1696 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)]; 1697 1698 const Value *Callee(I.getCalledValue()); 1699 if (isa<InlineAsm>(Callee)) 1700 visitInlineAsm(&I); 1701 else 1702 LowerCallTo(&I, getValue(Callee), false, LandingPad); 1703 1704 // If the value of the invoke is used outside of its defining block, make it 1705 // available as a virtual register. 1706 CopyToExportRegsIfNeeded(&I); 1707 1708 // Update successor info 1709 InvokeMBB->addSuccessor(Return); 1710 InvokeMBB->addSuccessor(LandingPad); 1711 1712 // Drop into normal successor. 1713 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(), 1714 MVT::Other, getControlRoot(), 1715 DAG.getBasicBlock(Return))); 1716} 1717 1718void SelectionDAGBuilder::visitUnwind(const UnwindInst &I) { 1719} 1720 1721/// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for 1722/// small case ranges). 1723bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR, 1724 CaseRecVector& WorkList, 1725 const Value* SV, 1726 MachineBasicBlock *Default, 1727 MachineBasicBlock *SwitchBB) { 1728 Case& BackCase = *(CR.Range.second-1); 1729 1730 // Size is the number of Cases represented by this range. 1731 size_t Size = CR.Range.second - CR.Range.first; 1732 if (Size > 3) 1733 return false; 1734 1735 // Get the MachineFunction which holds the current MBB. This is used when 1736 // inserting any additional MBBs necessary to represent the switch. 1737 MachineFunction *CurMF = FuncInfo.MF; 1738 1739 // Figure out which block is immediately after the current one. 1740 MachineBasicBlock *NextBlock = 0; 1741 MachineFunction::iterator BBI = CR.CaseBB; 1742 1743 if (++BBI != FuncInfo.MF->end()) 1744 NextBlock = BBI; 1745 1746 // TODO: If any two of the cases has the same destination, and if one value 1747 // is the same as the other, but has one bit unset that the other has set, 1748 // use bit manipulation to do two compares at once. For example: 1749 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)" 1750 1751 // Rearrange the case blocks so that the last one falls through if possible. 1752 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) { 1753 // The last case block won't fall through into 'NextBlock' if we emit the 1754 // branches in this order. See if rearranging a case value would help. 1755 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) { 1756 if (I->BB == NextBlock) { 1757 std::swap(*I, BackCase); 1758 break; 1759 } 1760 } 1761 } 1762 1763 // Create a CaseBlock record representing a conditional branch to 1764 // the Case's target mbb if the value being switched on SV is equal 1765 // to C. 1766 MachineBasicBlock *CurBlock = CR.CaseBB; 1767 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) { 1768 MachineBasicBlock *FallThrough; 1769 if (I != E-1) { 1770 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock()); 1771 CurMF->insert(BBI, FallThrough); 1772 1773 // Put SV in a virtual register to make it available from the new blocks. 1774 ExportFromCurrentBlock(SV); 1775 } else { 1776 // If the last case doesn't match, go to the default block. 1777 FallThrough = Default; 1778 } 1779 1780 const Value *RHS, *LHS, *MHS; 1781 ISD::CondCode CC; 1782 if (I->High == I->Low) { 1783 // This is just small small case range :) containing exactly 1 case 1784 CC = ISD::SETEQ; 1785 LHS = SV; RHS = I->High; MHS = NULL; 1786 } else { 1787 CC = ISD::SETLE; 1788 LHS = I->Low; MHS = SV; RHS = I->High; 1789 } 1790 CaseBlock CB(CC, LHS, RHS, MHS, I->BB, FallThrough, CurBlock); 1791 1792 // If emitting the first comparison, just call visitSwitchCase to emit the 1793 // code into the current block. Otherwise, push the CaseBlock onto the 1794 // vector to be later processed by SDISel, and insert the node's MBB 1795 // before the next MBB. 1796 if (CurBlock == SwitchBB) 1797 visitSwitchCase(CB, SwitchBB); 1798 else 1799 SwitchCases.push_back(CB); 1800 1801 CurBlock = FallThrough; 1802 } 1803 1804 return true; 1805} 1806 1807static inline bool areJTsAllowed(const TargetLowering &TLI) { 1808 return !DisableJumpTables && 1809 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) || 1810 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other)); 1811} 1812 1813static APInt ComputeRange(const APInt &First, const APInt &Last) { 1814 APInt LastExt(Last), FirstExt(First); 1815 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1; 1816 LastExt.sext(BitWidth); FirstExt.sext(BitWidth); 1817 return (LastExt - FirstExt + 1ULL); 1818} 1819 1820/// handleJTSwitchCase - Emit jumptable for current switch case range 1821bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec& CR, 1822 CaseRecVector& WorkList, 1823 const Value* SV, 1824 MachineBasicBlock* Default, 1825 MachineBasicBlock *SwitchBB) { 1826 Case& FrontCase = *CR.Range.first; 1827 Case& BackCase = *(CR.Range.second-1); 1828 1829 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue(); 1830 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue(); 1831 1832 APInt TSize(First.getBitWidth(), 0); 1833 for (CaseItr I = CR.Range.first, E = CR.Range.second; 1834 I!=E; ++I) 1835 TSize += I->size(); 1836 1837 if (!areJTsAllowed(TLI) || TSize.ult(4)) 1838 return false; 1839 1840 APInt Range = ComputeRange(First, Last); 1841 double Density = TSize.roundToDouble() / Range.roundToDouble(); 1842 if (Density < 0.4) 1843 return false; 1844 1845 DEBUG(dbgs() << "Lowering jump table\n" 1846 << "First entry: " << First << ". Last entry: " << Last << '\n' 1847 << "Range: " << Range 1848 << "Size: " << TSize << ". Density: " << Density << "\n\n"); 1849 1850 // Get the MachineFunction which holds the current MBB. This is used when 1851 // inserting any additional MBBs necessary to represent the switch. 1852 MachineFunction *CurMF = FuncInfo.MF; 1853 1854 // Figure out which block is immediately after the current one. 1855 MachineFunction::iterator BBI = CR.CaseBB; 1856 ++BBI; 1857 1858 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); 1859 1860 // Create a new basic block to hold the code for loading the address 1861 // of the jump table, and jumping to it. Update successor information; 1862 // we will either branch to the default case for the switch, or the jump 1863 // table. 1864 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB); 1865 CurMF->insert(BBI, JumpTableBB); 1866 CR.CaseBB->addSuccessor(Default); 1867 CR.CaseBB->addSuccessor(JumpTableBB); 1868 1869 // Build a vector of destination BBs, corresponding to each target 1870 // of the jump table. If the value of the jump table slot corresponds to 1871 // a case statement, push the case's BB onto the vector, otherwise, push 1872 // the default BB. 1873 std::vector<MachineBasicBlock*> DestBBs; 1874 APInt TEI = First; 1875 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) { 1876 const APInt &Low = cast<ConstantInt>(I->Low)->getValue(); 1877 const APInt &High = cast<ConstantInt>(I->High)->getValue(); 1878 1879 if (Low.sle(TEI) && TEI.sle(High)) { 1880 DestBBs.push_back(I->BB); 1881 if (TEI==High) 1882 ++I; 1883 } else { 1884 DestBBs.push_back(Default); 1885 } 1886 } 1887 1888 // Update successor info. Add one edge to each unique successor. 1889 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs()); 1890 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(), 1891 E = DestBBs.end(); I != E; ++I) { 1892 if (!SuccsHandled[(*I)->getNumber()]) { 1893 SuccsHandled[(*I)->getNumber()] = true; 1894 JumpTableBB->addSuccessor(*I); 1895 } 1896 } 1897 1898 // Create a jump table index for this jump table. 1899 unsigned JTEncoding = TLI.getJumpTableEncoding(); 1900 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding) 1901 ->createJumpTableIndex(DestBBs); 1902 1903 // Set the jump table information so that we can codegen it as a second 1904 // MachineBasicBlock 1905 JumpTable JT(-1U, JTI, JumpTableBB, Default); 1906 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB)); 1907 if (CR.CaseBB == SwitchBB) 1908 visitJumpTableHeader(JT, JTH, SwitchBB); 1909 1910 JTCases.push_back(JumpTableBlock(JTH, JT)); 1911 1912 return true; 1913} 1914 1915/// handleBTSplitSwitchCase - emit comparison and split binary search tree into 1916/// 2 subtrees. 1917bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR, 1918 CaseRecVector& WorkList, 1919 const Value* SV, 1920 MachineBasicBlock *Default, 1921 MachineBasicBlock *SwitchBB) { 1922 // Get the MachineFunction which holds the current MBB. This is used when 1923 // inserting any additional MBBs necessary to represent the switch. 1924 MachineFunction *CurMF = FuncInfo.MF; 1925 1926 // Figure out which block is immediately after the current one. 1927 MachineFunction::iterator BBI = CR.CaseBB; 1928 ++BBI; 1929 1930 Case& FrontCase = *CR.Range.first; 1931 Case& BackCase = *(CR.Range.second-1); 1932 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); 1933 1934 // Size is the number of Cases represented by this range. 1935 unsigned Size = CR.Range.second - CR.Range.first; 1936 1937 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue(); 1938 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue(); 1939 double FMetric = 0; 1940 CaseItr Pivot = CR.Range.first + Size/2; 1941 1942 // Select optimal pivot, maximizing sum density of LHS and RHS. This will 1943 // (heuristically) allow us to emit JumpTable's later. 1944 APInt TSize(First.getBitWidth(), 0); 1945 for (CaseItr I = CR.Range.first, E = CR.Range.second; 1946 I!=E; ++I) 1947 TSize += I->size(); 1948 1949 APInt LSize = FrontCase.size(); 1950 APInt RSize = TSize-LSize; 1951 DEBUG(dbgs() << "Selecting best pivot: \n" 1952 << "First: " << First << ", Last: " << Last <<'\n' 1953 << "LSize: " << LSize << ", RSize: " << RSize << '\n'); 1954 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second; 1955 J!=E; ++I, ++J) { 1956 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue(); 1957 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue(); 1958 APInt Range = ComputeRange(LEnd, RBegin); 1959 assert((Range - 2ULL).isNonNegative() && 1960 "Invalid case distance"); 1961 double LDensity = (double)LSize.roundToDouble() / 1962 (LEnd - First + 1ULL).roundToDouble(); 1963 double RDensity = (double)RSize.roundToDouble() / 1964 (Last - RBegin + 1ULL).roundToDouble(); 1965 double Metric = Range.logBase2()*(LDensity+RDensity); 1966 // Should always split in some non-trivial place 1967 DEBUG(dbgs() <<"=>Step\n" 1968 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n' 1969 << "LDensity: " << LDensity 1970 << ", RDensity: " << RDensity << '\n' 1971 << "Metric: " << Metric << '\n'); 1972 if (FMetric < Metric) { 1973 Pivot = J; 1974 FMetric = Metric; 1975 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n'); 1976 } 1977 1978 LSize += J->size(); 1979 RSize -= J->size(); 1980 } 1981 if (areJTsAllowed(TLI)) { 1982 // If our case is dense we *really* should handle it earlier! 1983 assert((FMetric > 0) && "Should handle dense range earlier!"); 1984 } else { 1985 Pivot = CR.Range.first + Size/2; 1986 } 1987 1988 CaseRange LHSR(CR.Range.first, Pivot); 1989 CaseRange RHSR(Pivot, CR.Range.second); 1990 Constant *C = Pivot->Low; 1991 MachineBasicBlock *FalseBB = 0, *TrueBB = 0; 1992 1993 // We know that we branch to the LHS if the Value being switched on is 1994 // less than the Pivot value, C. We use this to optimize our binary 1995 // tree a bit, by recognizing that if SV is greater than or equal to the 1996 // LHS's Case Value, and that Case Value is exactly one less than the 1997 // Pivot's Value, then we can branch directly to the LHS's Target, 1998 // rather than creating a leaf node for it. 1999 if ((LHSR.second - LHSR.first) == 1 && 2000 LHSR.first->High == CR.GE && 2001 cast<ConstantInt>(C)->getValue() == 2002 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) { 2003 TrueBB = LHSR.first->BB; 2004 } else { 2005 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB); 2006 CurMF->insert(BBI, TrueBB); 2007 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR)); 2008 2009 // Put SV in a virtual register to make it available from the new blocks. 2010 ExportFromCurrentBlock(SV); 2011 } 2012 2013 // Similar to the optimization above, if the Value being switched on is 2014 // known to be less than the Constant CR.LT, and the current Case Value 2015 // is CR.LT - 1, then we can branch directly to the target block for 2016 // the current Case Value, rather than emitting a RHS leaf node for it. 2017 if ((RHSR.second - RHSR.first) == 1 && CR.LT && 2018 cast<ConstantInt>(RHSR.first->Low)->getValue() == 2019 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) { 2020 FalseBB = RHSR.first->BB; 2021 } else { 2022 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB); 2023 CurMF->insert(BBI, FalseBB); 2024 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR)); 2025 2026 // Put SV in a virtual register to make it available from the new blocks. 2027 ExportFromCurrentBlock(SV); 2028 } 2029 2030 // Create a CaseBlock record representing a conditional branch to 2031 // the LHS node if the value being switched on SV is less than C. 2032 // Otherwise, branch to LHS. 2033 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB); 2034 2035 if (CR.CaseBB == SwitchBB) 2036 visitSwitchCase(CB, SwitchBB); 2037 else 2038 SwitchCases.push_back(CB); 2039 2040 return true; 2041} 2042 2043/// handleBitTestsSwitchCase - if current case range has few destination and 2044/// range span less, than machine word bitwidth, encode case range into series 2045/// of masks and emit bit tests with these masks. 2046bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR, 2047 CaseRecVector& WorkList, 2048 const Value* SV, 2049 MachineBasicBlock* Default, 2050 MachineBasicBlock *SwitchBB){ 2051 EVT PTy = TLI.getPointerTy(); 2052 unsigned IntPtrBits = PTy.getSizeInBits(); 2053 2054 Case& FrontCase = *CR.Range.first; 2055 Case& BackCase = *(CR.Range.second-1); 2056 2057 // Get the MachineFunction which holds the current MBB. This is used when 2058 // inserting any additional MBBs necessary to represent the switch. 2059 MachineFunction *CurMF = FuncInfo.MF; 2060 2061 // If target does not have legal shift left, do not emit bit tests at all. 2062 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy())) 2063 return false; 2064 2065 size_t numCmps = 0; 2066 for (CaseItr I = CR.Range.first, E = CR.Range.second; 2067 I!=E; ++I) { 2068 // Single case counts one, case range - two. 2069 numCmps += (I->Low == I->High ? 1 : 2); 2070 } 2071 2072 // Count unique destinations 2073 SmallSet<MachineBasicBlock*, 4> Dests; 2074 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) { 2075 Dests.insert(I->BB); 2076 if (Dests.size() > 3) 2077 // Don't bother the code below, if there are too much unique destinations 2078 return false; 2079 } 2080 DEBUG(dbgs() << "Total number of unique destinations: " 2081 << Dests.size() << '\n' 2082 << "Total number of comparisons: " << numCmps << '\n'); 2083 2084 // Compute span of values. 2085 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue(); 2086 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue(); 2087 APInt cmpRange = maxValue - minValue; 2088 2089 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n' 2090 << "Low bound: " << minValue << '\n' 2091 << "High bound: " << maxValue << '\n'); 2092 2093 if (cmpRange.uge(IntPtrBits) || 2094 (!(Dests.size() == 1 && numCmps >= 3) && 2095 !(Dests.size() == 2 && numCmps >= 5) && 2096 !(Dests.size() >= 3 && numCmps >= 6))) 2097 return false; 2098 2099 DEBUG(dbgs() << "Emitting bit tests\n"); 2100 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth()); 2101 2102 // Optimize the case where all the case values fit in a 2103 // word without having to subtract minValue. In this case, 2104 // we can optimize away the subtraction. 2105 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) { 2106 cmpRange = maxValue; 2107 } else { 2108 lowBound = minValue; 2109 } 2110 2111 CaseBitsVector CasesBits; 2112 unsigned i, count = 0; 2113 2114 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) { 2115 MachineBasicBlock* Dest = I->BB; 2116 for (i = 0; i < count; ++i) 2117 if (Dest == CasesBits[i].BB) 2118 break; 2119 2120 if (i == count) { 2121 assert((count < 3) && "Too much destinations to test!"); 2122 CasesBits.push_back(CaseBits(0, Dest, 0)); 2123 count++; 2124 } 2125 2126 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue(); 2127 const APInt& highValue = cast<ConstantInt>(I->High)->getValue(); 2128 2129 uint64_t lo = (lowValue - lowBound).getZExtValue(); 2130 uint64_t hi = (highValue - lowBound).getZExtValue(); 2131 2132 for (uint64_t j = lo; j <= hi; j++) { 2133 CasesBits[i].Mask |= 1ULL << j; 2134 CasesBits[i].Bits++; 2135 } 2136 2137 } 2138 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp()); 2139 2140 BitTestInfo BTC; 2141 2142 // Figure out which block is immediately after the current one. 2143 MachineFunction::iterator BBI = CR.CaseBB; 2144 ++BBI; 2145 2146 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); 2147 2148 DEBUG(dbgs() << "Cases:\n"); 2149 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) { 2150 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask 2151 << ", Bits: " << CasesBits[i].Bits 2152 << ", BB: " << CasesBits[i].BB << '\n'); 2153 2154 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB); 2155 CurMF->insert(BBI, CaseBB); 2156 BTC.push_back(BitTestCase(CasesBits[i].Mask, 2157 CaseBB, 2158 CasesBits[i].BB)); 2159 2160 // Put SV in a virtual register to make it available from the new blocks. 2161 ExportFromCurrentBlock(SV); 2162 } 2163 2164 BitTestBlock BTB(lowBound, cmpRange, SV, 2165 -1U, (CR.CaseBB == SwitchBB), 2166 CR.CaseBB, Default, BTC); 2167 2168 if (CR.CaseBB == SwitchBB) 2169 visitBitTestHeader(BTB, SwitchBB); 2170 2171 BitTestCases.push_back(BTB); 2172 2173 return true; 2174} 2175 2176/// Clusterify - Transform simple list of Cases into list of CaseRange's 2177size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases, 2178 const SwitchInst& SI) { 2179 size_t numCmps = 0; 2180 2181 // Start with "simple" cases 2182 for (size_t i = 1; i < SI.getNumSuccessors(); ++i) { 2183 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)]; 2184 Cases.push_back(Case(SI.getSuccessorValue(i), 2185 SI.getSuccessorValue(i), 2186 SMBB)); 2187 } 2188 std::sort(Cases.begin(), Cases.end(), CaseCmp()); 2189 2190 // Merge case into clusters 2191 if (Cases.size() >= 2) 2192 // Must recompute end() each iteration because it may be 2193 // invalidated by erase if we hold on to it 2194 for (CaseItr I = Cases.begin(), J = ++(Cases.begin()); J != Cases.end(); ) { 2195 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue(); 2196 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue(); 2197 MachineBasicBlock* nextBB = J->BB; 2198 MachineBasicBlock* currentBB = I->BB; 2199 2200 // If the two neighboring cases go to the same destination, merge them 2201 // into a single case. 2202 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) { 2203 I->High = J->High; 2204 J = Cases.erase(J); 2205 } else { 2206 I = J++; 2207 } 2208 } 2209 2210 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) { 2211 if (I->Low != I->High) 2212 // A range counts double, since it requires two compares. 2213 ++numCmps; 2214 } 2215 2216 return numCmps; 2217} 2218 2219void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) { 2220 MachineBasicBlock *SwitchMBB = FuncInfo.MBB; 2221 2222 // Figure out which block is immediately after the current one. 2223 MachineBasicBlock *NextBlock = 0; 2224 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()]; 2225 2226 // If there is only the default destination, branch to it if it is not the 2227 // next basic block. Otherwise, just fall through. 2228 if (SI.getNumOperands() == 2) { 2229 // Update machine-CFG edges. 2230 2231 // If this is not a fall-through branch, emit the branch. 2232 SwitchMBB->addSuccessor(Default); 2233 if (Default != NextBlock) 2234 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(), 2235 MVT::Other, getControlRoot(), 2236 DAG.getBasicBlock(Default))); 2237 2238 return; 2239 } 2240 2241 // If there are any non-default case statements, create a vector of Cases 2242 // representing each one, and sort the vector so that we can efficiently 2243 // create a binary search tree from them. 2244 CaseVector Cases; 2245 size_t numCmps = Clusterify(Cases, SI); 2246 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size() 2247 << ". Total compares: " << numCmps << '\n'); 2248 numCmps = 0; 2249 2250 // Get the Value to be switched on and default basic blocks, which will be 2251 // inserted into CaseBlock records, representing basic blocks in the binary 2252 // search tree. 2253 const Value *SV = SI.getOperand(0); 2254 2255 // Push the initial CaseRec onto the worklist 2256 CaseRecVector WorkList; 2257 WorkList.push_back(CaseRec(SwitchMBB,0,0, 2258 CaseRange(Cases.begin(),Cases.end()))); 2259 2260 while (!WorkList.empty()) { 2261 // Grab a record representing a case range to process off the worklist 2262 CaseRec CR = WorkList.back(); 2263 WorkList.pop_back(); 2264 2265 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB)) 2266 continue; 2267 2268 // If the range has few cases (two or less) emit a series of specific 2269 // tests. 2270 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB)) 2271 continue; 2272 2273 // If the switch has more than 5 blocks, and at least 40% dense, and the 2274 // target supports indirect branches, then emit a jump table rather than 2275 // lowering the switch to a binary tree of conditional branches. 2276 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB)) 2277 continue; 2278 2279 // Emit binary tree. We need to pick a pivot, and push left and right ranges 2280 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call. 2281 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB); 2282 } 2283} 2284 2285void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) { 2286 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB; 2287 2288 // Update machine-CFG edges with unique successors. 2289 SmallVector<BasicBlock*, 32> succs; 2290 succs.reserve(I.getNumSuccessors()); 2291 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) 2292 succs.push_back(I.getSuccessor(i)); 2293 array_pod_sort(succs.begin(), succs.end()); 2294 succs.erase(std::unique(succs.begin(), succs.end()), succs.end()); 2295 for (unsigned i = 0, e = succs.size(); i != e; ++i) 2296 IndirectBrMBB->addSuccessor(FuncInfo.MBBMap[succs[i]]); 2297 2298 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurDebugLoc(), 2299 MVT::Other, getControlRoot(), 2300 getValue(I.getAddress()))); 2301} 2302 2303void SelectionDAGBuilder::visitFSub(const User &I) { 2304 // -0.0 - X --> fneg 2305 const Type *Ty = I.getType(); 2306 if (Ty->isVectorTy()) { 2307 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) { 2308 const VectorType *DestTy = cast<VectorType>(I.getType()); 2309 const Type *ElTy = DestTy->getElementType(); 2310 unsigned VL = DestTy->getNumElements(); 2311 std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy)); 2312 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size()); 2313 if (CV == CNZ) { 2314 SDValue Op2 = getValue(I.getOperand(1)); 2315 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(), 2316 Op2.getValueType(), Op2)); 2317 return; 2318 } 2319 } 2320 } 2321 2322 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0))) 2323 if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) { 2324 SDValue Op2 = getValue(I.getOperand(1)); 2325 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(), 2326 Op2.getValueType(), Op2)); 2327 return; 2328 } 2329 2330 visitBinary(I, ISD::FSUB); 2331} 2332 2333void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) { 2334 SDValue Op1 = getValue(I.getOperand(0)); 2335 SDValue Op2 = getValue(I.getOperand(1)); 2336 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(), 2337 Op1.getValueType(), Op1, Op2)); 2338} 2339 2340void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) { 2341 SDValue Op1 = getValue(I.getOperand(0)); 2342 SDValue Op2 = getValue(I.getOperand(1)); 2343 if (!I.getType()->isVectorTy() && 2344 Op2.getValueType() != TLI.getShiftAmountTy()) { 2345 // If the operand is smaller than the shift count type, promote it. 2346 EVT PTy = TLI.getPointerTy(); 2347 EVT STy = TLI.getShiftAmountTy(); 2348 if (STy.bitsGT(Op2.getValueType())) 2349 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(), 2350 TLI.getShiftAmountTy(), Op2); 2351 // If the operand is larger than the shift count type but the shift 2352 // count type has enough bits to represent any shift value, truncate 2353 // it now. This is a common case and it exposes the truncate to 2354 // optimization early. 2355 else if (STy.getSizeInBits() >= 2356 Log2_32_Ceil(Op2.getValueType().getSizeInBits())) 2357 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), 2358 TLI.getShiftAmountTy(), Op2); 2359 // Otherwise we'll need to temporarily settle for some other 2360 // convenient type; type legalization will make adjustments as 2361 // needed. 2362 else if (PTy.bitsLT(Op2.getValueType())) 2363 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), 2364 TLI.getPointerTy(), Op2); 2365 else if (PTy.bitsGT(Op2.getValueType())) 2366 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(), 2367 TLI.getPointerTy(), Op2); 2368 } 2369 2370 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(), 2371 Op1.getValueType(), Op1, Op2)); 2372} 2373 2374void SelectionDAGBuilder::visitICmp(const User &I) { 2375 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE; 2376 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I)) 2377 predicate = IC->getPredicate(); 2378 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I)) 2379 predicate = ICmpInst::Predicate(IC->getPredicate()); 2380 SDValue Op1 = getValue(I.getOperand(0)); 2381 SDValue Op2 = getValue(I.getOperand(1)); 2382 ISD::CondCode Opcode = getICmpCondCode(predicate); 2383 2384 EVT DestVT = TLI.getValueType(I.getType()); 2385 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode)); 2386} 2387 2388void SelectionDAGBuilder::visitFCmp(const User &I) { 2389 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE; 2390 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I)) 2391 predicate = FC->getPredicate(); 2392 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I)) 2393 predicate = FCmpInst::Predicate(FC->getPredicate()); 2394 SDValue Op1 = getValue(I.getOperand(0)); 2395 SDValue Op2 = getValue(I.getOperand(1)); 2396 ISD::CondCode Condition = getFCmpCondCode(predicate); 2397 EVT DestVT = TLI.getValueType(I.getType()); 2398 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition)); 2399} 2400 2401void SelectionDAGBuilder::visitSelect(const User &I) { 2402 SmallVector<EVT, 4> ValueVTs; 2403 ComputeValueVTs(TLI, I.getType(), ValueVTs); 2404 unsigned NumValues = ValueVTs.size(); 2405 if (NumValues == 0) return; 2406 2407 SmallVector<SDValue, 4> Values(NumValues); 2408 SDValue Cond = getValue(I.getOperand(0)); 2409 SDValue TrueVal = getValue(I.getOperand(1)); 2410 SDValue FalseVal = getValue(I.getOperand(2)); 2411 2412 for (unsigned i = 0; i != NumValues; ++i) 2413 Values[i] = DAG.getNode(ISD::SELECT, getCurDebugLoc(), 2414 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i), 2415 Cond, 2416 SDValue(TrueVal.getNode(), 2417 TrueVal.getResNo() + i), 2418 SDValue(FalseVal.getNode(), 2419 FalseVal.getResNo() + i)); 2420 2421 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(), 2422 DAG.getVTList(&ValueVTs[0], NumValues), 2423 &Values[0], NumValues)); 2424} 2425 2426void SelectionDAGBuilder::visitTrunc(const User &I) { 2427 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest). 2428 SDValue N = getValue(I.getOperand(0)); 2429 EVT DestVT = TLI.getValueType(I.getType()); 2430 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N)); 2431} 2432 2433void SelectionDAGBuilder::visitZExt(const User &I) { 2434 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest). 2435 // ZExt also can't be a cast to bool for same reason. So, nothing much to do 2436 SDValue N = getValue(I.getOperand(0)); 2437 EVT DestVT = TLI.getValueType(I.getType()); 2438 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N)); 2439} 2440 2441void SelectionDAGBuilder::visitSExt(const User &I) { 2442 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest). 2443 // SExt also can't be a cast to bool for same reason. So, nothing much to do 2444 SDValue N = getValue(I.getOperand(0)); 2445 EVT DestVT = TLI.getValueType(I.getType()); 2446 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N)); 2447} 2448 2449void SelectionDAGBuilder::visitFPTrunc(const User &I) { 2450 // FPTrunc is never a no-op cast, no need to check 2451 SDValue N = getValue(I.getOperand(0)); 2452 EVT DestVT = TLI.getValueType(I.getType()); 2453 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(), 2454 DestVT, N, DAG.getIntPtrConstant(0))); 2455} 2456 2457void SelectionDAGBuilder::visitFPExt(const User &I){ 2458 // FPTrunc is never a no-op cast, no need to check 2459 SDValue N = getValue(I.getOperand(0)); 2460 EVT DestVT = TLI.getValueType(I.getType()); 2461 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N)); 2462} 2463 2464void SelectionDAGBuilder::visitFPToUI(const User &I) { 2465 // FPToUI is never a no-op cast, no need to check 2466 SDValue N = getValue(I.getOperand(0)); 2467 EVT DestVT = TLI.getValueType(I.getType()); 2468 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N)); 2469} 2470 2471void SelectionDAGBuilder::visitFPToSI(const User &I) { 2472 // FPToSI is never a no-op cast, no need to check 2473 SDValue N = getValue(I.getOperand(0)); 2474 EVT DestVT = TLI.getValueType(I.getType()); 2475 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N)); 2476} 2477 2478void SelectionDAGBuilder::visitUIToFP(const User &I) { 2479 // UIToFP is never a no-op cast, no need to check 2480 SDValue N = getValue(I.getOperand(0)); 2481 EVT DestVT = TLI.getValueType(I.getType()); 2482 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N)); 2483} 2484 2485void SelectionDAGBuilder::visitSIToFP(const User &I){ 2486 // SIToFP is never a no-op cast, no need to check 2487 SDValue N = getValue(I.getOperand(0)); 2488 EVT DestVT = TLI.getValueType(I.getType()); 2489 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N)); 2490} 2491 2492void SelectionDAGBuilder::visitPtrToInt(const User &I) { 2493 // What to do depends on the size of the integer and the size of the pointer. 2494 // We can either truncate, zero extend, or no-op, accordingly. 2495 SDValue N = getValue(I.getOperand(0)); 2496 EVT DestVT = TLI.getValueType(I.getType()); 2497 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT)); 2498} 2499 2500void SelectionDAGBuilder::visitIntToPtr(const User &I) { 2501 // What to do depends on the size of the integer and the size of the pointer. 2502 // We can either truncate, zero extend, or no-op, accordingly. 2503 SDValue N = getValue(I.getOperand(0)); 2504 EVT DestVT = TLI.getValueType(I.getType()); 2505 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT)); 2506} 2507 2508void SelectionDAGBuilder::visitBitCast(const User &I) { 2509 SDValue N = getValue(I.getOperand(0)); 2510 EVT DestVT = TLI.getValueType(I.getType()); 2511 2512 // BitCast assures us that source and destination are the same size so this is 2513 // either a BIT_CONVERT or a no-op. 2514 if (DestVT != N.getValueType()) 2515 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), 2516 DestVT, N)); // convert types. 2517 else 2518 setValue(&I, N); // noop cast. 2519} 2520 2521void SelectionDAGBuilder::visitInsertElement(const User &I) { 2522 SDValue InVec = getValue(I.getOperand(0)); 2523 SDValue InVal = getValue(I.getOperand(1)); 2524 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), 2525 TLI.getPointerTy(), 2526 getValue(I.getOperand(2))); 2527 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(), 2528 TLI.getValueType(I.getType()), 2529 InVec, InVal, InIdx)); 2530} 2531 2532void SelectionDAGBuilder::visitExtractElement(const User &I) { 2533 SDValue InVec = getValue(I.getOperand(0)); 2534 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), 2535 TLI.getPointerTy(), 2536 getValue(I.getOperand(1))); 2537 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(), 2538 TLI.getValueType(I.getType()), InVec, InIdx)); 2539} 2540 2541// Utility for visitShuffleVector - Returns true if the mask is mask starting 2542// from SIndx and increasing to the element length (undefs are allowed). 2543static bool SequentialMask(SmallVectorImpl<int> &Mask, unsigned SIndx) { 2544 unsigned MaskNumElts = Mask.size(); 2545 for (unsigned i = 0; i != MaskNumElts; ++i) 2546 if ((Mask[i] >= 0) && (Mask[i] != (int)(i + SIndx))) 2547 return false; 2548 return true; 2549} 2550 2551void SelectionDAGBuilder::visitShuffleVector(const User &I) { 2552 SmallVector<int, 8> Mask; 2553 SDValue Src1 = getValue(I.getOperand(0)); 2554 SDValue Src2 = getValue(I.getOperand(1)); 2555 2556 // Convert the ConstantVector mask operand into an array of ints, with -1 2557 // representing undef values. 2558 SmallVector<Constant*, 8> MaskElts; 2559 cast<Constant>(I.getOperand(2))->getVectorElements(MaskElts); 2560 unsigned MaskNumElts = MaskElts.size(); 2561 for (unsigned i = 0; i != MaskNumElts; ++i) { 2562 if (isa<UndefValue>(MaskElts[i])) 2563 Mask.push_back(-1); 2564 else 2565 Mask.push_back(cast<ConstantInt>(MaskElts[i])->getSExtValue()); 2566 } 2567 2568 EVT VT = TLI.getValueType(I.getType()); 2569 EVT SrcVT = Src1.getValueType(); 2570 unsigned SrcNumElts = SrcVT.getVectorNumElements(); 2571 2572 if (SrcNumElts == MaskNumElts) { 2573 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2, 2574 &Mask[0])); 2575 return; 2576 } 2577 2578 // Normalize the shuffle vector since mask and vector length don't match. 2579 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) { 2580 // Mask is longer than the source vectors and is a multiple of the source 2581 // vectors. We can use concatenate vector to make the mask and vectors 2582 // lengths match. 2583 if (SrcNumElts*2 == MaskNumElts && SequentialMask(Mask, 0)) { 2584 // The shuffle is concatenating two vectors together. 2585 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(), 2586 VT, Src1, Src2)); 2587 return; 2588 } 2589 2590 // Pad both vectors with undefs to make them the same length as the mask. 2591 unsigned NumConcat = MaskNumElts / SrcNumElts; 2592 bool Src1U = Src1.getOpcode() == ISD::UNDEF; 2593 bool Src2U = Src2.getOpcode() == ISD::UNDEF; 2594 SDValue UndefVal = DAG.getUNDEF(SrcVT); 2595 2596 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal); 2597 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal); 2598 MOps1[0] = Src1; 2599 MOps2[0] = Src2; 2600 2601 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS, 2602 getCurDebugLoc(), VT, 2603 &MOps1[0], NumConcat); 2604 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS, 2605 getCurDebugLoc(), VT, 2606 &MOps2[0], NumConcat); 2607 2608 // Readjust mask for new input vector length. 2609 SmallVector<int, 8> MappedOps; 2610 for (unsigned i = 0; i != MaskNumElts; ++i) { 2611 int Idx = Mask[i]; 2612 if (Idx < (int)SrcNumElts) 2613 MappedOps.push_back(Idx); 2614 else 2615 MappedOps.push_back(Idx + MaskNumElts - SrcNumElts); 2616 } 2617 2618 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2, 2619 &MappedOps[0])); 2620 return; 2621 } 2622 2623 if (SrcNumElts > MaskNumElts) { 2624 // Analyze the access pattern of the vector to see if we can extract 2625 // two subvectors and do the shuffle. The analysis is done by calculating 2626 // the range of elements the mask access on both vectors. 2627 int MinRange[2] = { SrcNumElts+1, SrcNumElts+1}; 2628 int MaxRange[2] = {-1, -1}; 2629 2630 for (unsigned i = 0; i != MaskNumElts; ++i) { 2631 int Idx = Mask[i]; 2632 int Input = 0; 2633 if (Idx < 0) 2634 continue; 2635 2636 if (Idx >= (int)SrcNumElts) { 2637 Input = 1; 2638 Idx -= SrcNumElts; 2639 } 2640 if (Idx > MaxRange[Input]) 2641 MaxRange[Input] = Idx; 2642 if (Idx < MinRange[Input]) 2643 MinRange[Input] = Idx; 2644 } 2645 2646 // Check if the access is smaller than the vector size and can we find 2647 // a reasonable extract index. 2648 int RangeUse[2] = { 2, 2 }; // 0 = Unused, 1 = Extract, 2 = Can not 2649 // Extract. 2650 int StartIdx[2]; // StartIdx to extract from 2651 for (int Input=0; Input < 2; ++Input) { 2652 if (MinRange[Input] == (int)(SrcNumElts+1) && MaxRange[Input] == -1) { 2653 RangeUse[Input] = 0; // Unused 2654 StartIdx[Input] = 0; 2655 } else if (MaxRange[Input] - MinRange[Input] < (int)MaskNumElts) { 2656 // Fits within range but we should see if we can find a good 2657 // start index that is a multiple of the mask length. 2658 if (MaxRange[Input] < (int)MaskNumElts) { 2659 RangeUse[Input] = 1; // Extract from beginning of the vector 2660 StartIdx[Input] = 0; 2661 } else { 2662 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts; 2663 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts && 2664 StartIdx[Input] + MaskNumElts < SrcNumElts) 2665 RangeUse[Input] = 1; // Extract from a multiple of the mask length. 2666 } 2667 } 2668 } 2669 2670 if (RangeUse[0] == 0 && RangeUse[1] == 0) { 2671 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used. 2672 return; 2673 } 2674 else if (RangeUse[0] < 2 && RangeUse[1] < 2) { 2675 // Extract appropriate subvector and generate a vector shuffle 2676 for (int Input=0; Input < 2; ++Input) { 2677 SDValue &Src = Input == 0 ? Src1 : Src2; 2678 if (RangeUse[Input] == 0) 2679 Src = DAG.getUNDEF(VT); 2680 else 2681 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT, 2682 Src, DAG.getIntPtrConstant(StartIdx[Input])); 2683 } 2684 2685 // Calculate new mask. 2686 SmallVector<int, 8> MappedOps; 2687 for (unsigned i = 0; i != MaskNumElts; ++i) { 2688 int Idx = Mask[i]; 2689 if (Idx < 0) 2690 MappedOps.push_back(Idx); 2691 else if (Idx < (int)SrcNumElts) 2692 MappedOps.push_back(Idx - StartIdx[0]); 2693 else 2694 MappedOps.push_back(Idx - SrcNumElts - StartIdx[1] + MaskNumElts); 2695 } 2696 2697 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2, 2698 &MappedOps[0])); 2699 return; 2700 } 2701 } 2702 2703 // We can't use either concat vectors or extract subvectors so fall back to 2704 // replacing the shuffle with extract and build vector. 2705 // to insert and build vector. 2706 EVT EltVT = VT.getVectorElementType(); 2707 EVT PtrVT = TLI.getPointerTy(); 2708 SmallVector<SDValue,8> Ops; 2709 for (unsigned i = 0; i != MaskNumElts; ++i) { 2710 if (Mask[i] < 0) { 2711 Ops.push_back(DAG.getUNDEF(EltVT)); 2712 } else { 2713 int Idx = Mask[i]; 2714 SDValue Res; 2715 2716 if (Idx < (int)SrcNumElts) 2717 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(), 2718 EltVT, Src1, DAG.getConstant(Idx, PtrVT)); 2719 else 2720 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(), 2721 EltVT, Src2, 2722 DAG.getConstant(Idx - SrcNumElts, PtrVT)); 2723 2724 Ops.push_back(Res); 2725 } 2726 } 2727 2728 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(), 2729 VT, &Ops[0], Ops.size())); 2730} 2731 2732void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) { 2733 const Value *Op0 = I.getOperand(0); 2734 const Value *Op1 = I.getOperand(1); 2735 const Type *AggTy = I.getType(); 2736 const Type *ValTy = Op1->getType(); 2737 bool IntoUndef = isa<UndefValue>(Op0); 2738 bool FromUndef = isa<UndefValue>(Op1); 2739 2740 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy, 2741 I.idx_begin(), I.idx_end()); 2742 2743 SmallVector<EVT, 4> AggValueVTs; 2744 ComputeValueVTs(TLI, AggTy, AggValueVTs); 2745 SmallVector<EVT, 4> ValValueVTs; 2746 ComputeValueVTs(TLI, ValTy, ValValueVTs); 2747 2748 unsigned NumAggValues = AggValueVTs.size(); 2749 unsigned NumValValues = ValValueVTs.size(); 2750 SmallVector<SDValue, 4> Values(NumAggValues); 2751 2752 SDValue Agg = getValue(Op0); 2753 SDValue Val = getValue(Op1); 2754 unsigned i = 0; 2755 // Copy the beginning value(s) from the original aggregate. 2756 for (; i != LinearIndex; ++i) 2757 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) : 2758 SDValue(Agg.getNode(), Agg.getResNo() + i); 2759 // Copy values from the inserted value(s). 2760 for (; i != LinearIndex + NumValValues; ++i) 2761 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) : 2762 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex); 2763 // Copy remaining value(s) from the original aggregate. 2764 for (; i != NumAggValues; ++i) 2765 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) : 2766 SDValue(Agg.getNode(), Agg.getResNo() + i); 2767 2768 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(), 2769 DAG.getVTList(&AggValueVTs[0], NumAggValues), 2770 &Values[0], NumAggValues)); 2771} 2772 2773void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) { 2774 const Value *Op0 = I.getOperand(0); 2775 const Type *AggTy = Op0->getType(); 2776 const Type *ValTy = I.getType(); 2777 bool OutOfUndef = isa<UndefValue>(Op0); 2778 2779 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy, 2780 I.idx_begin(), I.idx_end()); 2781 2782 SmallVector<EVT, 4> ValValueVTs; 2783 ComputeValueVTs(TLI, ValTy, ValValueVTs); 2784 2785 unsigned NumValValues = ValValueVTs.size(); 2786 SmallVector<SDValue, 4> Values(NumValValues); 2787 2788 SDValue Agg = getValue(Op0); 2789 // Copy out the selected value(s). 2790 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i) 2791 Values[i - LinearIndex] = 2792 OutOfUndef ? 2793 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) : 2794 SDValue(Agg.getNode(), Agg.getResNo() + i); 2795 2796 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(), 2797 DAG.getVTList(&ValValueVTs[0], NumValValues), 2798 &Values[0], NumValValues)); 2799} 2800 2801void SelectionDAGBuilder::visitGetElementPtr(const User &I) { 2802 SDValue N = getValue(I.getOperand(0)); 2803 const Type *Ty = I.getOperand(0)->getType(); 2804 2805 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end(); 2806 OI != E; ++OI) { 2807 const Value *Idx = *OI; 2808 if (const StructType *StTy = dyn_cast<StructType>(Ty)) { 2809 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue(); 2810 if (Field) { 2811 // N = N + Offset 2812 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field); 2813 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N, 2814 DAG.getIntPtrConstant(Offset)); 2815 } 2816 2817 Ty = StTy->getElementType(Field); 2818 } else if (const UnionType *UnTy = dyn_cast<UnionType>(Ty)) { 2819 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue(); 2820 2821 // Offset canonically 0 for unions, but type changes 2822 Ty = UnTy->getElementType(Field); 2823 } else { 2824 Ty = cast<SequentialType>(Ty)->getElementType(); 2825 2826 // If this is a constant subscript, handle it quickly. 2827 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) { 2828 if (CI->isZero()) continue; 2829 uint64_t Offs = 2830 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue(); 2831 SDValue OffsVal; 2832 EVT PTy = TLI.getPointerTy(); 2833 unsigned PtrBits = PTy.getSizeInBits(); 2834 if (PtrBits < 64) 2835 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), 2836 TLI.getPointerTy(), 2837 DAG.getConstant(Offs, MVT::i64)); 2838 else 2839 OffsVal = DAG.getIntPtrConstant(Offs); 2840 2841 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N, 2842 OffsVal); 2843 continue; 2844 } 2845 2846 // N = N + Idx * ElementSize; 2847 APInt ElementSize = APInt(TLI.getPointerTy().getSizeInBits(), 2848 TD->getTypeAllocSize(Ty)); 2849 SDValue IdxN = getValue(Idx); 2850 2851 // If the index is smaller or larger than intptr_t, truncate or extend 2852 // it. 2853 IdxN = DAG.getSExtOrTrunc(IdxN, getCurDebugLoc(), N.getValueType()); 2854 2855 // If this is a multiply by a power of two, turn it into a shl 2856 // immediately. This is a very common case. 2857 if (ElementSize != 1) { 2858 if (ElementSize.isPowerOf2()) { 2859 unsigned Amt = ElementSize.logBase2(); 2860 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(), 2861 N.getValueType(), IdxN, 2862 DAG.getConstant(Amt, TLI.getPointerTy())); 2863 } else { 2864 SDValue Scale = DAG.getConstant(ElementSize, TLI.getPointerTy()); 2865 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(), 2866 N.getValueType(), IdxN, Scale); 2867 } 2868 } 2869 2870 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), 2871 N.getValueType(), N, IdxN); 2872 } 2873 } 2874 2875 setValue(&I, N); 2876} 2877 2878void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) { 2879 // If this is a fixed sized alloca in the entry block of the function, 2880 // allocate it statically on the stack. 2881 if (FuncInfo.StaticAllocaMap.count(&I)) 2882 return; // getValue will auto-populate this. 2883 2884 const Type *Ty = I.getAllocatedType(); 2885 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty); 2886 unsigned Align = 2887 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), 2888 I.getAlignment()); 2889 2890 SDValue AllocSize = getValue(I.getArraySize()); 2891 2892 EVT IntPtr = TLI.getPointerTy(); 2893 if (AllocSize.getValueType() != IntPtr) 2894 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurDebugLoc(), IntPtr); 2895 2896 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), IntPtr, 2897 AllocSize, 2898 DAG.getConstant(TySize, IntPtr)); 2899 2900 // Handle alignment. If the requested alignment is less than or equal to 2901 // the stack alignment, ignore it. If the size is greater than or equal to 2902 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node. 2903 unsigned StackAlign = TM.getFrameInfo()->getStackAlignment(); 2904 if (Align <= StackAlign) 2905 Align = 0; 2906 2907 // Round the size of the allocation up to the stack alignment size 2908 // by add SA-1 to the size. 2909 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(), 2910 AllocSize.getValueType(), AllocSize, 2911 DAG.getIntPtrConstant(StackAlign-1)); 2912 2913 // Mask out the low bits for alignment purposes. 2914 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(), 2915 AllocSize.getValueType(), AllocSize, 2916 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1))); 2917 2918 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) }; 2919 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other); 2920 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(), 2921 VTs, Ops, 3); 2922 setValue(&I, DSA); 2923 DAG.setRoot(DSA.getValue(1)); 2924 2925 // Inform the Frame Information that we have just allocated a variable-sized 2926 // object. 2927 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1); 2928} 2929 2930void SelectionDAGBuilder::visitLoad(const LoadInst &I) { 2931 const Value *SV = I.getOperand(0); 2932 SDValue Ptr = getValue(SV); 2933 2934 const Type *Ty = I.getType(); 2935 2936 bool isVolatile = I.isVolatile(); 2937 bool isNonTemporal = I.getMetadata("nontemporal") != 0; 2938 unsigned Alignment = I.getAlignment(); 2939 2940 SmallVector<EVT, 4> ValueVTs; 2941 SmallVector<uint64_t, 4> Offsets; 2942 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets); 2943 unsigned NumValues = ValueVTs.size(); 2944 if (NumValues == 0) 2945 return; 2946 2947 SDValue Root; 2948 bool ConstantMemory = false; 2949 if (I.isVolatile()) 2950 // Serialize volatile loads with other side effects. 2951 Root = getRoot(); 2952 else if (AA->pointsToConstantMemory(SV)) { 2953 // Do not serialize (non-volatile) loads of constant memory with anything. 2954 Root = DAG.getEntryNode(); 2955 ConstantMemory = true; 2956 } else { 2957 // Do not serialize non-volatile loads against each other. 2958 Root = DAG.getRoot(); 2959 } 2960 2961 SmallVector<SDValue, 4> Values(NumValues); 2962 SmallVector<SDValue, 4> Chains(NumValues); 2963 EVT PtrVT = Ptr.getValueType(); 2964 for (unsigned i = 0; i != NumValues; ++i) { 2965 SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(), 2966 PtrVT, Ptr, 2967 DAG.getConstant(Offsets[i], PtrVT)); 2968 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root, 2969 A, SV, Offsets[i], isVolatile, 2970 isNonTemporal, Alignment); 2971 2972 Values[i] = L; 2973 Chains[i] = L.getValue(1); 2974 } 2975 2976 if (!ConstantMemory) { 2977 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), 2978 MVT::Other, &Chains[0], NumValues); 2979 if (isVolatile) 2980 DAG.setRoot(Chain); 2981 else 2982 PendingLoads.push_back(Chain); 2983 } 2984 2985 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(), 2986 DAG.getVTList(&ValueVTs[0], NumValues), 2987 &Values[0], NumValues)); 2988} 2989 2990void SelectionDAGBuilder::visitStore(const StoreInst &I) { 2991 const Value *SrcV = I.getOperand(0); 2992 const Value *PtrV = I.getOperand(1); 2993 2994 SmallVector<EVT, 4> ValueVTs; 2995 SmallVector<uint64_t, 4> Offsets; 2996 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets); 2997 unsigned NumValues = ValueVTs.size(); 2998 if (NumValues == 0) 2999 return; 3000 3001 // Get the lowered operands. Note that we do this after 3002 // checking if NumResults is zero, because with zero results 3003 // the operands won't have values in the map. 3004 SDValue Src = getValue(SrcV); 3005 SDValue Ptr = getValue(PtrV); 3006 3007 SDValue Root = getRoot(); 3008 SmallVector<SDValue, 4> Chains(NumValues); 3009 EVT PtrVT = Ptr.getValueType(); 3010 bool isVolatile = I.isVolatile(); 3011 bool isNonTemporal = I.getMetadata("nontemporal") != 0; 3012 unsigned Alignment = I.getAlignment(); 3013 3014 for (unsigned i = 0; i != NumValues; ++i) { 3015 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr, 3016 DAG.getConstant(Offsets[i], PtrVT)); 3017 Chains[i] = DAG.getStore(Root, getCurDebugLoc(), 3018 SDValue(Src.getNode(), Src.getResNo() + i), 3019 Add, PtrV, Offsets[i], isVolatile, 3020 isNonTemporal, Alignment); 3021 } 3022 3023 DAG.setRoot(DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), 3024 MVT::Other, &Chains[0], NumValues)); 3025} 3026 3027/// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC 3028/// node. 3029void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I, 3030 unsigned Intrinsic) { 3031 bool HasChain = !I.doesNotAccessMemory(); 3032 bool OnlyLoad = HasChain && I.onlyReadsMemory(); 3033 3034 // Build the operand list. 3035 SmallVector<SDValue, 8> Ops; 3036 if (HasChain) { // If this intrinsic has side-effects, chainify it. 3037 if (OnlyLoad) { 3038 // We don't need to serialize loads against other loads. 3039 Ops.push_back(DAG.getRoot()); 3040 } else { 3041 Ops.push_back(getRoot()); 3042 } 3043 } 3044 3045 // Info is set by getTgtMemInstrinsic 3046 TargetLowering::IntrinsicInfo Info; 3047 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic); 3048 3049 // Add the intrinsic ID as an integer operand if it's not a target intrinsic. 3050 if (!IsTgtIntrinsic) 3051 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy())); 3052 3053 // Add all operands of the call to the operand list. 3054 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) { 3055 SDValue Op = getValue(I.getArgOperand(i)); 3056 assert(TLI.isTypeLegal(Op.getValueType()) && 3057 "Intrinsic uses a non-legal type?"); 3058 Ops.push_back(Op); 3059 } 3060 3061 SmallVector<EVT, 4> ValueVTs; 3062 ComputeValueVTs(TLI, I.getType(), ValueVTs); 3063#ifndef NDEBUG 3064 for (unsigned Val = 0, E = ValueVTs.size(); Val != E; ++Val) { 3065 assert(TLI.isTypeLegal(ValueVTs[Val]) && 3066 "Intrinsic uses a non-legal type?"); 3067 } 3068#endif // NDEBUG 3069 3070 if (HasChain) 3071 ValueVTs.push_back(MVT::Other); 3072 3073 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size()); 3074 3075 // Create the node. 3076 SDValue Result; 3077 if (IsTgtIntrinsic) { 3078 // This is target intrinsic that touches memory 3079 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(), 3080 VTs, &Ops[0], Ops.size(), 3081 Info.memVT, Info.ptrVal, Info.offset, 3082 Info.align, Info.vol, 3083 Info.readMem, Info.writeMem); 3084 } else if (!HasChain) { 3085 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(), 3086 VTs, &Ops[0], Ops.size()); 3087 } else if (!I.getType()->isVoidTy()) { 3088 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(), 3089 VTs, &Ops[0], Ops.size()); 3090 } else { 3091 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(), 3092 VTs, &Ops[0], Ops.size()); 3093 } 3094 3095 if (HasChain) { 3096 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1); 3097 if (OnlyLoad) 3098 PendingLoads.push_back(Chain); 3099 else 3100 DAG.setRoot(Chain); 3101 } 3102 3103 if (!I.getType()->isVoidTy()) { 3104 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) { 3105 EVT VT = TLI.getValueType(PTy); 3106 Result = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), VT, Result); 3107 } 3108 3109 setValue(&I, Result); 3110 } 3111} 3112 3113/// GetSignificand - Get the significand and build it into a floating-point 3114/// number with exponent of 1: 3115/// 3116/// Op = (Op & 0x007fffff) | 0x3f800000; 3117/// 3118/// where Op is the hexidecimal representation of floating point value. 3119static SDValue 3120GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) { 3121 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op, 3122 DAG.getConstant(0x007fffff, MVT::i32)); 3123 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1, 3124 DAG.getConstant(0x3f800000, MVT::i32)); 3125 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t2); 3126} 3127 3128/// GetExponent - Get the exponent: 3129/// 3130/// (float)(int)(((Op & 0x7f800000) >> 23) - 127); 3131/// 3132/// where Op is the hexidecimal representation of floating point value. 3133static SDValue 3134GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI, 3135 DebugLoc dl) { 3136 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op, 3137 DAG.getConstant(0x7f800000, MVT::i32)); 3138 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0, 3139 DAG.getConstant(23, TLI.getPointerTy())); 3140 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1, 3141 DAG.getConstant(127, MVT::i32)); 3142 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2); 3143} 3144 3145/// getF32Constant - Get 32-bit floating point constant. 3146static SDValue 3147getF32Constant(SelectionDAG &DAG, unsigned Flt) { 3148 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32); 3149} 3150 3151/// Inlined utility function to implement binary input atomic intrinsics for 3152/// visitIntrinsicCall: I is a call instruction 3153/// Op is the associated NodeType for I 3154const char * 3155SelectionDAGBuilder::implVisitBinaryAtomic(const CallInst& I, 3156 ISD::NodeType Op) { 3157 SDValue Root = getRoot(); 3158 SDValue L = 3159 DAG.getAtomic(Op, getCurDebugLoc(), 3160 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(), 3161 Root, 3162 getValue(I.getArgOperand(0)), 3163 getValue(I.getArgOperand(1)), 3164 I.getArgOperand(0)); 3165 setValue(&I, L); 3166 DAG.setRoot(L.getValue(1)); 3167 return 0; 3168} 3169 3170// implVisitAluOverflow - Lower arithmetic overflow instrinsics. 3171const char * 3172SelectionDAGBuilder::implVisitAluOverflow(const CallInst &I, ISD::NodeType Op) { 3173 SDValue Op1 = getValue(I.getArgOperand(0)); 3174 SDValue Op2 = getValue(I.getArgOperand(1)); 3175 3176 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1); 3177 setValue(&I, DAG.getNode(Op, getCurDebugLoc(), VTs, Op1, Op2)); 3178 return 0; 3179} 3180 3181/// visitExp - Lower an exp intrinsic. Handles the special sequences for 3182/// limited-precision mode. 3183void 3184SelectionDAGBuilder::visitExp(const CallInst &I) { 3185 SDValue result; 3186 DebugLoc dl = getCurDebugLoc(); 3187 3188 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 && 3189 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3190 SDValue Op = getValue(I.getArgOperand(0)); 3191 3192 // Put the exponent in the right bit position for later addition to the 3193 // final result: 3194 // 3195 // #define LOG2OFe 1.4426950f 3196 // IntegerPartOfX = ((int32_t)(X * LOG2OFe)); 3197 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op, 3198 getF32Constant(DAG, 0x3fb8aa3b)); 3199 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0); 3200 3201 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX; 3202 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); 3203 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1); 3204 3205 // IntegerPartOfX <<= 23; 3206 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, 3207 DAG.getConstant(23, TLI.getPointerTy())); 3208 3209 if (LimitFloatPrecision <= 6) { 3210 // For floating-point precision of 6: 3211 // 3212 // TwoToFractionalPartOfX = 3213 // 0.997535578f + 3214 // (0.735607626f + 0.252464424f * x) * x; 3215 // 3216 // error 0.0144103317, which is 6 bits 3217 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3218 getF32Constant(DAG, 0x3e814304)); 3219 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3220 getF32Constant(DAG, 0x3f3c50c8)); 3221 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3222 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3223 getF32Constant(DAG, 0x3f7f5e7e)); 3224 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t5); 3225 3226 // Add the exponent into the result in integer domain. 3227 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32, 3228 TwoToFracPartOfX, IntegerPartOfX); 3229 3230 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t6); 3231 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) { 3232 // For floating-point precision of 12: 3233 // 3234 // TwoToFractionalPartOfX = 3235 // 0.999892986f + 3236 // (0.696457318f + 3237 // (0.224338339f + 0.792043434e-1f * x) * x) * x; 3238 // 3239 // 0.000107046256 error, which is 13 to 14 bits 3240 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3241 getF32Constant(DAG, 0x3da235e3)); 3242 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3243 getF32Constant(DAG, 0x3e65b8f3)); 3244 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3245 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3246 getF32Constant(DAG, 0x3f324b07)); 3247 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3248 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3249 getF32Constant(DAG, 0x3f7ff8fd)); 3250 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t7); 3251 3252 // Add the exponent into the result in integer domain. 3253 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32, 3254 TwoToFracPartOfX, IntegerPartOfX); 3255 3256 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t8); 3257 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18 3258 // For floating-point precision of 18: 3259 // 3260 // TwoToFractionalPartOfX = 3261 // 0.999999982f + 3262 // (0.693148872f + 3263 // (0.240227044f + 3264 // (0.554906021e-1f + 3265 // (0.961591928e-2f + 3266 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; 3267 // 3268 // error 2.47208000*10^(-7), which is better than 18 bits 3269 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3270 getF32Constant(DAG, 0x3924b03e)); 3271 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3272 getF32Constant(DAG, 0x3ab24b87)); 3273 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3274 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3275 getF32Constant(DAG, 0x3c1d8c17)); 3276 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3277 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3278 getF32Constant(DAG, 0x3d634a1d)); 3279 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3280 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 3281 getF32Constant(DAG, 0x3e75fe14)); 3282 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 3283 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, 3284 getF32Constant(DAG, 0x3f317234)); 3285 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); 3286 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, 3287 getF32Constant(DAG, 0x3f800000)); 3288 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl, 3289 MVT::i32, t13); 3290 3291 // Add the exponent into the result in integer domain. 3292 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32, 3293 TwoToFracPartOfX, IntegerPartOfX); 3294 3295 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t14); 3296 } 3297 } else { 3298 // No special expansion. 3299 result = DAG.getNode(ISD::FEXP, dl, 3300 getValue(I.getArgOperand(0)).getValueType(), 3301 getValue(I.getArgOperand(0))); 3302 } 3303 3304 setValue(&I, result); 3305} 3306 3307/// visitLog - Lower a log intrinsic. Handles the special sequences for 3308/// limited-precision mode. 3309void 3310SelectionDAGBuilder::visitLog(const CallInst &I) { 3311 SDValue result; 3312 DebugLoc dl = getCurDebugLoc(); 3313 3314 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 && 3315 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3316 SDValue Op = getValue(I.getArgOperand(0)); 3317 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op); 3318 3319 // Scale the exponent by log(2) [0.69314718f]. 3320 SDValue Exp = GetExponent(DAG, Op1, TLI, dl); 3321 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp, 3322 getF32Constant(DAG, 0x3f317218)); 3323 3324 // Get the significand and build it into a floating-point number with 3325 // exponent of 1. 3326 SDValue X = GetSignificand(DAG, Op1, dl); 3327 3328 if (LimitFloatPrecision <= 6) { 3329 // For floating-point precision of 6: 3330 // 3331 // LogofMantissa = 3332 // -1.1609546f + 3333 // (1.4034025f - 0.23903021f * x) * x; 3334 // 3335 // error 0.0034276066, which is better than 8 bits 3336 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3337 getF32Constant(DAG, 0xbe74c456)); 3338 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3339 getF32Constant(DAG, 0x3fb3a2b1)); 3340 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3341 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3342 getF32Constant(DAG, 0x3f949a29)); 3343 3344 result = DAG.getNode(ISD::FADD, dl, 3345 MVT::f32, LogOfExponent, LogOfMantissa); 3346 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) { 3347 // For floating-point precision of 12: 3348 // 3349 // LogOfMantissa = 3350 // -1.7417939f + 3351 // (2.8212026f + 3352 // (-1.4699568f + 3353 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x; 3354 // 3355 // error 0.000061011436, which is 14 bits 3356 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3357 getF32Constant(DAG, 0xbd67b6d6)); 3358 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3359 getF32Constant(DAG, 0x3ee4f4b8)); 3360 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3361 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3362 getF32Constant(DAG, 0x3fbc278b)); 3363 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3364 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3365 getF32Constant(DAG, 0x40348e95)); 3366 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3367 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, 3368 getF32Constant(DAG, 0x3fdef31a)); 3369 3370 result = DAG.getNode(ISD::FADD, dl, 3371 MVT::f32, LogOfExponent, LogOfMantissa); 3372 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18 3373 // For floating-point precision of 18: 3374 // 3375 // LogOfMantissa = 3376 // -2.1072184f + 3377 // (4.2372794f + 3378 // (-3.7029485f + 3379 // (2.2781945f + 3380 // (-0.87823314f + 3381 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x; 3382 // 3383 // error 0.0000023660568, which is better than 18 bits 3384 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3385 getF32Constant(DAG, 0xbc91e5ac)); 3386 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3387 getF32Constant(DAG, 0x3e4350aa)); 3388 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3389 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3390 getF32Constant(DAG, 0x3f60d3e3)); 3391 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3392 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3393 getF32Constant(DAG, 0x4011cdf0)); 3394 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3395 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, 3396 getF32Constant(DAG, 0x406cfd1c)); 3397 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3398 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 3399 getF32Constant(DAG, 0x408797cb)); 3400 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 3401 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10, 3402 getF32Constant(DAG, 0x4006dcab)); 3403 3404 result = DAG.getNode(ISD::FADD, dl, 3405 MVT::f32, LogOfExponent, LogOfMantissa); 3406 } 3407 } else { 3408 // No special expansion. 3409 result = DAG.getNode(ISD::FLOG, dl, 3410 getValue(I.getArgOperand(0)).getValueType(), 3411 getValue(I.getArgOperand(0))); 3412 } 3413 3414 setValue(&I, result); 3415} 3416 3417/// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for 3418/// limited-precision mode. 3419void 3420SelectionDAGBuilder::visitLog2(const CallInst &I) { 3421 SDValue result; 3422 DebugLoc dl = getCurDebugLoc(); 3423 3424 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 && 3425 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3426 SDValue Op = getValue(I.getArgOperand(0)); 3427 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op); 3428 3429 // Get the exponent. 3430 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl); 3431 3432 // Get the significand and build it into a floating-point number with 3433 // exponent of 1. 3434 SDValue X = GetSignificand(DAG, Op1, dl); 3435 3436 // Different possible minimax approximations of significand in 3437 // floating-point for various degrees of accuracy over [1,2]. 3438 if (LimitFloatPrecision <= 6) { 3439 // For floating-point precision of 6: 3440 // 3441 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x; 3442 // 3443 // error 0.0049451742, which is more than 7 bits 3444 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3445 getF32Constant(DAG, 0xbeb08fe0)); 3446 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3447 getF32Constant(DAG, 0x40019463)); 3448 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3449 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3450 getF32Constant(DAG, 0x3fd6633d)); 3451 3452 result = DAG.getNode(ISD::FADD, dl, 3453 MVT::f32, LogOfExponent, Log2ofMantissa); 3454 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) { 3455 // For floating-point precision of 12: 3456 // 3457 // Log2ofMantissa = 3458 // -2.51285454f + 3459 // (4.07009056f + 3460 // (-2.12067489f + 3461 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x; 3462 // 3463 // error 0.0000876136000, which is better than 13 bits 3464 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3465 getF32Constant(DAG, 0xbda7262e)); 3466 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3467 getF32Constant(DAG, 0x3f25280b)); 3468 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3469 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3470 getF32Constant(DAG, 0x4007b923)); 3471 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3472 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3473 getF32Constant(DAG, 0x40823e2f)); 3474 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3475 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, 3476 getF32Constant(DAG, 0x4020d29c)); 3477 3478 result = DAG.getNode(ISD::FADD, dl, 3479 MVT::f32, LogOfExponent, Log2ofMantissa); 3480 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18 3481 // For floating-point precision of 18: 3482 // 3483 // Log2ofMantissa = 3484 // -3.0400495f + 3485 // (6.1129976f + 3486 // (-5.3420409f + 3487 // (3.2865683f + 3488 // (-1.2669343f + 3489 // (0.27515199f - 3490 // 0.25691327e-1f * x) * x) * x) * x) * x) * x; 3491 // 3492 // error 0.0000018516, which is better than 18 bits 3493 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3494 getF32Constant(DAG, 0xbcd2769e)); 3495 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3496 getF32Constant(DAG, 0x3e8ce0b9)); 3497 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3498 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3499 getF32Constant(DAG, 0x3fa22ae7)); 3500 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3501 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3502 getF32Constant(DAG, 0x40525723)); 3503 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3504 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, 3505 getF32Constant(DAG, 0x40aaf200)); 3506 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3507 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 3508 getF32Constant(DAG, 0x40c39dad)); 3509 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 3510 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10, 3511 getF32Constant(DAG, 0x4042902c)); 3512 3513 result = DAG.getNode(ISD::FADD, dl, 3514 MVT::f32, LogOfExponent, Log2ofMantissa); 3515 } 3516 } else { 3517 // No special expansion. 3518 result = DAG.getNode(ISD::FLOG2, dl, 3519 getValue(I.getArgOperand(0)).getValueType(), 3520 getValue(I.getArgOperand(0))); 3521 } 3522 3523 setValue(&I, result); 3524} 3525 3526/// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for 3527/// limited-precision mode. 3528void 3529SelectionDAGBuilder::visitLog10(const CallInst &I) { 3530 SDValue result; 3531 DebugLoc dl = getCurDebugLoc(); 3532 3533 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 && 3534 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3535 SDValue Op = getValue(I.getArgOperand(0)); 3536 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op); 3537 3538 // Scale the exponent by log10(2) [0.30102999f]. 3539 SDValue Exp = GetExponent(DAG, Op1, TLI, dl); 3540 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp, 3541 getF32Constant(DAG, 0x3e9a209a)); 3542 3543 // Get the significand and build it into a floating-point number with 3544 // exponent of 1. 3545 SDValue X = GetSignificand(DAG, Op1, dl); 3546 3547 if (LimitFloatPrecision <= 6) { 3548 // For floating-point precision of 6: 3549 // 3550 // Log10ofMantissa = 3551 // -0.50419619f + 3552 // (0.60948995f - 0.10380950f * x) * x; 3553 // 3554 // error 0.0014886165, which is 6 bits 3555 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3556 getF32Constant(DAG, 0xbdd49a13)); 3557 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3558 getF32Constant(DAG, 0x3f1c0789)); 3559 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3560 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3561 getF32Constant(DAG, 0x3f011300)); 3562 3563 result = DAG.getNode(ISD::FADD, dl, 3564 MVT::f32, LogOfExponent, Log10ofMantissa); 3565 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) { 3566 // For floating-point precision of 12: 3567 // 3568 // Log10ofMantissa = 3569 // -0.64831180f + 3570 // (0.91751397f + 3571 // (-0.31664806f + 0.47637168e-1f * x) * x) * x; 3572 // 3573 // error 0.00019228036, which is better than 12 bits 3574 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3575 getF32Constant(DAG, 0x3d431f31)); 3576 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, 3577 getF32Constant(DAG, 0x3ea21fb2)); 3578 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3579 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3580 getF32Constant(DAG, 0x3f6ae232)); 3581 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3582 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4, 3583 getF32Constant(DAG, 0x3f25f7c3)); 3584 3585 result = DAG.getNode(ISD::FADD, dl, 3586 MVT::f32, LogOfExponent, Log10ofMantissa); 3587 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18 3588 // For floating-point precision of 18: 3589 // 3590 // Log10ofMantissa = 3591 // -0.84299375f + 3592 // (1.5327582f + 3593 // (-1.0688956f + 3594 // (0.49102474f + 3595 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x; 3596 // 3597 // error 0.0000037995730, which is better than 18 bits 3598 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3599 getF32Constant(DAG, 0x3c5d51ce)); 3600 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, 3601 getF32Constant(DAG, 0x3e00685a)); 3602 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3603 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3604 getF32Constant(DAG, 0x3efb6798)); 3605 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3606 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4, 3607 getF32Constant(DAG, 0x3f88d192)); 3608 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3609 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3610 getF32Constant(DAG, 0x3fc4316c)); 3611 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3612 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8, 3613 getF32Constant(DAG, 0x3f57ce70)); 3614 3615 result = DAG.getNode(ISD::FADD, dl, 3616 MVT::f32, LogOfExponent, Log10ofMantissa); 3617 } 3618 } else { 3619 // No special expansion. 3620 result = DAG.getNode(ISD::FLOG10, dl, 3621 getValue(I.getArgOperand(0)).getValueType(), 3622 getValue(I.getArgOperand(0))); 3623 } 3624 3625 setValue(&I, result); 3626} 3627 3628/// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for 3629/// limited-precision mode. 3630void 3631SelectionDAGBuilder::visitExp2(const CallInst &I) { 3632 SDValue result; 3633 DebugLoc dl = getCurDebugLoc(); 3634 3635 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 && 3636 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3637 SDValue Op = getValue(I.getArgOperand(0)); 3638 3639 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op); 3640 3641 // FractionalPartOfX = x - (float)IntegerPartOfX; 3642 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); 3643 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1); 3644 3645 // IntegerPartOfX <<= 23; 3646 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, 3647 DAG.getConstant(23, TLI.getPointerTy())); 3648 3649 if (LimitFloatPrecision <= 6) { 3650 // For floating-point precision of 6: 3651 // 3652 // TwoToFractionalPartOfX = 3653 // 0.997535578f + 3654 // (0.735607626f + 0.252464424f * x) * x; 3655 // 3656 // error 0.0144103317, which is 6 bits 3657 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3658 getF32Constant(DAG, 0x3e814304)); 3659 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3660 getF32Constant(DAG, 0x3f3c50c8)); 3661 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3662 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3663 getF32Constant(DAG, 0x3f7f5e7e)); 3664 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5); 3665 SDValue TwoToFractionalPartOfX = 3666 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX); 3667 3668 result = DAG.getNode(ISD::BIT_CONVERT, dl, 3669 MVT::f32, TwoToFractionalPartOfX); 3670 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) { 3671 // For floating-point precision of 12: 3672 // 3673 // TwoToFractionalPartOfX = 3674 // 0.999892986f + 3675 // (0.696457318f + 3676 // (0.224338339f + 0.792043434e-1f * x) * x) * x; 3677 // 3678 // error 0.000107046256, which is 13 to 14 bits 3679 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3680 getF32Constant(DAG, 0x3da235e3)); 3681 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3682 getF32Constant(DAG, 0x3e65b8f3)); 3683 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3684 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3685 getF32Constant(DAG, 0x3f324b07)); 3686 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3687 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3688 getF32Constant(DAG, 0x3f7ff8fd)); 3689 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7); 3690 SDValue TwoToFractionalPartOfX = 3691 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX); 3692 3693 result = DAG.getNode(ISD::BIT_CONVERT, dl, 3694 MVT::f32, TwoToFractionalPartOfX); 3695 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18 3696 // For floating-point precision of 18: 3697 // 3698 // TwoToFractionalPartOfX = 3699 // 0.999999982f + 3700 // (0.693148872f + 3701 // (0.240227044f + 3702 // (0.554906021e-1f + 3703 // (0.961591928e-2f + 3704 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; 3705 // error 2.47208000*10^(-7), which is better than 18 bits 3706 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3707 getF32Constant(DAG, 0x3924b03e)); 3708 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3709 getF32Constant(DAG, 0x3ab24b87)); 3710 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3711 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3712 getF32Constant(DAG, 0x3c1d8c17)); 3713 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3714 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3715 getF32Constant(DAG, 0x3d634a1d)); 3716 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3717 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 3718 getF32Constant(DAG, 0x3e75fe14)); 3719 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 3720 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, 3721 getF32Constant(DAG, 0x3f317234)); 3722 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); 3723 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, 3724 getF32Constant(DAG, 0x3f800000)); 3725 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13); 3726 SDValue TwoToFractionalPartOfX = 3727 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX); 3728 3729 result = DAG.getNode(ISD::BIT_CONVERT, dl, 3730 MVT::f32, TwoToFractionalPartOfX); 3731 } 3732 } else { 3733 // No special expansion. 3734 result = DAG.getNode(ISD::FEXP2, dl, 3735 getValue(I.getArgOperand(0)).getValueType(), 3736 getValue(I.getArgOperand(0))); 3737 } 3738 3739 setValue(&I, result); 3740} 3741 3742/// visitPow - Lower a pow intrinsic. Handles the special sequences for 3743/// limited-precision mode with x == 10.0f. 3744void 3745SelectionDAGBuilder::visitPow(const CallInst &I) { 3746 SDValue result; 3747 const Value *Val = I.getArgOperand(0); 3748 DebugLoc dl = getCurDebugLoc(); 3749 bool IsExp10 = false; 3750 3751 if (getValue(Val).getValueType() == MVT::f32 && 3752 getValue(I.getArgOperand(1)).getValueType() == MVT::f32 && 3753 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3754 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) { 3755 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 3756 APFloat Ten(10.0f); 3757 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten); 3758 } 3759 } 3760 } 3761 3762 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3763 SDValue Op = getValue(I.getArgOperand(1)); 3764 3765 // Put the exponent in the right bit position for later addition to the 3766 // final result: 3767 // 3768 // #define LOG2OF10 3.3219281f 3769 // IntegerPartOfX = (int32_t)(x * LOG2OF10); 3770 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op, 3771 getF32Constant(DAG, 0x40549a78)); 3772 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0); 3773 3774 // FractionalPartOfX = x - (float)IntegerPartOfX; 3775 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); 3776 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1); 3777 3778 // IntegerPartOfX <<= 23; 3779 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, 3780 DAG.getConstant(23, TLI.getPointerTy())); 3781 3782 if (LimitFloatPrecision <= 6) { 3783 // For floating-point precision of 6: 3784 // 3785 // twoToFractionalPartOfX = 3786 // 0.997535578f + 3787 // (0.735607626f + 0.252464424f * x) * x; 3788 // 3789 // error 0.0144103317, which is 6 bits 3790 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3791 getF32Constant(DAG, 0x3e814304)); 3792 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3793 getF32Constant(DAG, 0x3f3c50c8)); 3794 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3795 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3796 getF32Constant(DAG, 0x3f7f5e7e)); 3797 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5); 3798 SDValue TwoToFractionalPartOfX = 3799 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX); 3800 3801 result = DAG.getNode(ISD::BIT_CONVERT, dl, 3802 MVT::f32, TwoToFractionalPartOfX); 3803 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) { 3804 // For floating-point precision of 12: 3805 // 3806 // TwoToFractionalPartOfX = 3807 // 0.999892986f + 3808 // (0.696457318f + 3809 // (0.224338339f + 0.792043434e-1f * x) * x) * x; 3810 // 3811 // error 0.000107046256, which is 13 to 14 bits 3812 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3813 getF32Constant(DAG, 0x3da235e3)); 3814 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3815 getF32Constant(DAG, 0x3e65b8f3)); 3816 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3817 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3818 getF32Constant(DAG, 0x3f324b07)); 3819 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3820 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3821 getF32Constant(DAG, 0x3f7ff8fd)); 3822 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7); 3823 SDValue TwoToFractionalPartOfX = 3824 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX); 3825 3826 result = DAG.getNode(ISD::BIT_CONVERT, dl, 3827 MVT::f32, TwoToFractionalPartOfX); 3828 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18 3829 // For floating-point precision of 18: 3830 // 3831 // TwoToFractionalPartOfX = 3832 // 0.999999982f + 3833 // (0.693148872f + 3834 // (0.240227044f + 3835 // (0.554906021e-1f + 3836 // (0.961591928e-2f + 3837 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; 3838 // error 2.47208000*10^(-7), which is better than 18 bits 3839 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3840 getF32Constant(DAG, 0x3924b03e)); 3841 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3842 getF32Constant(DAG, 0x3ab24b87)); 3843 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3844 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3845 getF32Constant(DAG, 0x3c1d8c17)); 3846 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3847 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3848 getF32Constant(DAG, 0x3d634a1d)); 3849 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3850 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 3851 getF32Constant(DAG, 0x3e75fe14)); 3852 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 3853 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, 3854 getF32Constant(DAG, 0x3f317234)); 3855 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); 3856 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, 3857 getF32Constant(DAG, 0x3f800000)); 3858 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13); 3859 SDValue TwoToFractionalPartOfX = 3860 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX); 3861 3862 result = DAG.getNode(ISD::BIT_CONVERT, dl, 3863 MVT::f32, TwoToFractionalPartOfX); 3864 } 3865 } else { 3866 // No special expansion. 3867 result = DAG.getNode(ISD::FPOW, dl, 3868 getValue(I.getArgOperand(0)).getValueType(), 3869 getValue(I.getArgOperand(0)), 3870 getValue(I.getArgOperand(1))); 3871 } 3872 3873 setValue(&I, result); 3874} 3875 3876 3877/// ExpandPowI - Expand a llvm.powi intrinsic. 3878static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS, 3879 SelectionDAG &DAG) { 3880 // If RHS is a constant, we can expand this out to a multiplication tree, 3881 // otherwise we end up lowering to a call to __powidf2 (for example). When 3882 // optimizing for size, we only want to do this if the expansion would produce 3883 // a small number of multiplies, otherwise we do the full expansion. 3884 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) { 3885 // Get the exponent as a positive value. 3886 unsigned Val = RHSC->getSExtValue(); 3887 if ((int)Val < 0) Val = -Val; 3888 3889 // powi(x, 0) -> 1.0 3890 if (Val == 0) 3891 return DAG.getConstantFP(1.0, LHS.getValueType()); 3892 3893 const Function *F = DAG.getMachineFunction().getFunction(); 3894 if (!F->hasFnAttr(Attribute::OptimizeForSize) || 3895 // If optimizing for size, don't insert too many multiplies. This 3896 // inserts up to 5 multiplies. 3897 CountPopulation_32(Val)+Log2_32(Val) < 7) { 3898 // We use the simple binary decomposition method to generate the multiply 3899 // sequence. There are more optimal ways to do this (for example, 3900 // powi(x,15) generates one more multiply than it should), but this has 3901 // the benefit of being both really simple and much better than a libcall. 3902 SDValue Res; // Logically starts equal to 1.0 3903 SDValue CurSquare = LHS; 3904 while (Val) { 3905 if (Val & 1) { 3906 if (Res.getNode()) 3907 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare); 3908 else 3909 Res = CurSquare; // 1.0*CurSquare. 3910 } 3911 3912 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(), 3913 CurSquare, CurSquare); 3914 Val >>= 1; 3915 } 3916 3917 // If the original was negative, invert the result, producing 1/(x*x*x). 3918 if (RHSC->getSExtValue() < 0) 3919 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(), 3920 DAG.getConstantFP(1.0, LHS.getValueType()), Res); 3921 return Res; 3922 } 3923 } 3924 3925 // Otherwise, expand to a libcall. 3926 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS); 3927} 3928 3929/// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function 3930/// argument, create the corresponding DBG_VALUE machine instruction for it now. 3931/// At the end of instruction selection, they will be inserted to the entry BB. 3932bool 3933SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable, 3934 uint64_t Offset, 3935 const SDValue &N) { 3936 if (!isa<Argument>(V)) 3937 return false; 3938 3939 MachineFunction &MF = DAG.getMachineFunction(); 3940 // Ignore inlined function arguments here. 3941 DIVariable DV(Variable); 3942 if (DV.isInlinedFnArgument(MF.getFunction())) 3943 return false; 3944 3945 MachineBasicBlock *MBB = FuncInfo.MBB; 3946 if (MBB != &MF.front()) 3947 return false; 3948 3949 unsigned Reg = 0; 3950 if (N.getOpcode() == ISD::CopyFromReg) { 3951 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg(); 3952 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) { 3953 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 3954 unsigned PR = RegInfo.getLiveInPhysReg(Reg); 3955 if (PR) 3956 Reg = PR; 3957 } 3958 } 3959 3960 if (!Reg) { 3961 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V); 3962 if (VMI == FuncInfo.ValueMap.end()) 3963 return false; 3964 Reg = VMI->second; 3965 } 3966 3967 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo(); 3968 MachineInstrBuilder MIB = BuildMI(MF, getCurDebugLoc(), 3969 TII->get(TargetOpcode::DBG_VALUE)) 3970 .addReg(Reg, RegState::Debug).addImm(Offset).addMetadata(Variable); 3971 FuncInfo.ArgDbgValues.push_back(&*MIB); 3972 return true; 3973} 3974 3975// VisualStudio defines setjmp as _setjmp 3976#if defined(_MSC_VER) && defined(setjmp) 3977#define setjmp_undefined_for_visual_studio 3978#undef setjmp 3979#endif 3980 3981/// visitIntrinsicCall - Lower the call to the specified intrinsic function. If 3982/// we want to emit this as a call to a named external function, return the name 3983/// otherwise lower it and return null. 3984const char * 3985SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) { 3986 DebugLoc dl = getCurDebugLoc(); 3987 SDValue Res; 3988 3989 switch (Intrinsic) { 3990 default: 3991 // By default, turn this into a target intrinsic node. 3992 visitTargetIntrinsic(I, Intrinsic); 3993 return 0; 3994 case Intrinsic::vastart: visitVAStart(I); return 0; 3995 case Intrinsic::vaend: visitVAEnd(I); return 0; 3996 case Intrinsic::vacopy: visitVACopy(I); return 0; 3997 case Intrinsic::returnaddress: 3998 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(), 3999 getValue(I.getArgOperand(0)))); 4000 return 0; 4001 case Intrinsic::frameaddress: 4002 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(), 4003 getValue(I.getArgOperand(0)))); 4004 return 0; 4005 case Intrinsic::setjmp: 4006 return "_setjmp"+!TLI.usesUnderscoreSetJmp(); 4007 case Intrinsic::longjmp: 4008 return "_longjmp"+!TLI.usesUnderscoreLongJmp(); 4009 case Intrinsic::memcpy: { 4010 // Assert for address < 256 since we support only user defined address 4011 // spaces. 4012 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() 4013 < 256 && 4014 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() 4015 < 256 && 4016 "Unknown address space"); 4017 SDValue Op1 = getValue(I.getArgOperand(0)); 4018 SDValue Op2 = getValue(I.getArgOperand(1)); 4019 SDValue Op3 = getValue(I.getArgOperand(2)); 4020 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); 4021 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); 4022 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, false, 4023 I.getArgOperand(0), 0, I.getArgOperand(1), 0)); 4024 return 0; 4025 } 4026 case Intrinsic::memset: { 4027 // Assert for address < 256 since we support only user defined address 4028 // spaces. 4029 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() 4030 < 256 && 4031 "Unknown address space"); 4032 SDValue Op1 = getValue(I.getArgOperand(0)); 4033 SDValue Op2 = getValue(I.getArgOperand(1)); 4034 SDValue Op3 = getValue(I.getArgOperand(2)); 4035 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); 4036 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); 4037 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align, isVol, 4038 I.getArgOperand(0), 0)); 4039 return 0; 4040 } 4041 case Intrinsic::memmove: { 4042 // Assert for address < 256 since we support only user defined address 4043 // spaces. 4044 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() 4045 < 256 && 4046 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() 4047 < 256 && 4048 "Unknown address space"); 4049 SDValue Op1 = getValue(I.getArgOperand(0)); 4050 SDValue Op2 = getValue(I.getArgOperand(1)); 4051 SDValue Op3 = getValue(I.getArgOperand(2)); 4052 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); 4053 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); 4054 4055 // If the source and destination are known to not be aliases, we can 4056 // lower memmove as memcpy. 4057 uint64_t Size = -1ULL; 4058 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3)) 4059 Size = C->getZExtValue(); 4060 if (AA->alias(I.getArgOperand(0), Size, I.getArgOperand(1), Size) == 4061 AliasAnalysis::NoAlias) { 4062 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, 4063 false, I.getArgOperand(0), 0, 4064 I.getArgOperand(1), 0)); 4065 return 0; 4066 } 4067 4068 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align, isVol, 4069 I.getArgOperand(0), 0, I.getArgOperand(1), 0)); 4070 return 0; 4071 } 4072 case Intrinsic::dbg_declare: { 4073 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I); 4074 if (!DIVariable(DI.getVariable()).Verify()) 4075 return 0; 4076 4077 MDNode *Variable = DI.getVariable(); 4078 // Parameters are handled specially. 4079 bool isParameter = 4080 DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable; 4081 const Value *Address = DI.getAddress(); 4082 if (!Address) 4083 return 0; 4084 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address)) 4085 Address = BCI->getOperand(0); 4086 const AllocaInst *AI = dyn_cast<AllocaInst>(Address); 4087 4088 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder 4089 // but do not always have a corresponding SDNode built. The SDNodeOrder 4090 // absolute, but not relative, values are different depending on whether 4091 // debug info exists. 4092 ++SDNodeOrder; 4093 SDValue &N = NodeMap[Address]; 4094 SDDbgValue *SDV; 4095 if (N.getNode()) { 4096 if (isParameter && !AI) { 4097 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode()); 4098 if (FINode) 4099 // Byval parameter. We have a frame index at this point. 4100 SDV = DAG.getDbgValue(Variable, FINode->getIndex(), 4101 0, dl, SDNodeOrder); 4102 else 4103 // Can't do anything with other non-AI cases yet. This might be a 4104 // parameter of a callee function that got inlined, for example. 4105 return 0; 4106 } else if (AI) 4107 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(), 4108 0, dl, SDNodeOrder); 4109 else 4110 // Can't do anything with other non-AI cases yet. 4111 return 0; 4112 DAG.AddDbgValue(SDV, N.getNode(), isParameter); 4113 } else { 4114 // This isn't useful, but it shows what we're missing. 4115 SDV = DAG.getDbgValue(Variable, UndefValue::get(Address->getType()), 4116 0, dl, SDNodeOrder); 4117 DAG.AddDbgValue(SDV, 0, isParameter); 4118 } 4119 return 0; 4120 } 4121 case Intrinsic::dbg_value: { 4122 const DbgValueInst &DI = cast<DbgValueInst>(I); 4123 if (!DIVariable(DI.getVariable()).Verify()) 4124 return 0; 4125 4126 MDNode *Variable = DI.getVariable(); 4127 uint64_t Offset = DI.getOffset(); 4128 const Value *V = DI.getValue(); 4129 if (!V) 4130 return 0; 4131 4132 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder 4133 // but do not always have a corresponding SDNode built. The SDNodeOrder 4134 // absolute, but not relative, values are different depending on whether 4135 // debug info exists. 4136 ++SDNodeOrder; 4137 SDDbgValue *SDV; 4138 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) { 4139 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder); 4140 DAG.AddDbgValue(SDV, 0, false); 4141 } else { 4142 bool createUndef = false; 4143 // Do not use getValue() in here; we don't want to generate code at 4144 // this point if it hasn't been done yet. 4145 SDValue N = NodeMap[V]; 4146 if (!N.getNode() && isa<Argument>(V)) 4147 // Check unused arguments map. 4148 N = UnusedArgNodeMap[V]; 4149 if (N.getNode()) { 4150 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) { 4151 SDV = DAG.getDbgValue(Variable, N.getNode(), 4152 N.getResNo(), Offset, dl, SDNodeOrder); 4153 DAG.AddDbgValue(SDV, N.getNode(), false); 4154 } 4155 } else if (isa<PHINode>(V) && !V->use_empty() ) { 4156 // Do not call getValue(V) yet, as we don't want to generate code. 4157 // Remember it for later. 4158 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder); 4159 DanglingDebugInfoMap[V] = DDI; 4160 } else 4161 createUndef = true; 4162 if (createUndef) { 4163 // We may expand this to cover more cases. One case where we have no 4164 // data available is an unreferenced parameter; we need this fallback. 4165 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()), 4166 Offset, dl, SDNodeOrder); 4167 DAG.AddDbgValue(SDV, 0, false); 4168 } 4169 } 4170 4171 // Build a debug info table entry. 4172 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V)) 4173 V = BCI->getOperand(0); 4174 const AllocaInst *AI = dyn_cast<AllocaInst>(V); 4175 // Don't handle byval struct arguments or VLAs, for example. 4176 if (!AI) 4177 return 0; 4178 DenseMap<const AllocaInst*, int>::iterator SI = 4179 FuncInfo.StaticAllocaMap.find(AI); 4180 if (SI == FuncInfo.StaticAllocaMap.end()) 4181 return 0; // VLAs. 4182 int FI = SI->second; 4183 4184 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 4185 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo()) 4186 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc()); 4187 return 0; 4188 } 4189 case Intrinsic::eh_exception: { 4190 // Insert the EXCEPTIONADDR instruction. 4191 assert(FuncInfo.MBB->isLandingPad() && 4192 "Call to eh.exception not in landing pad!"); 4193 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other); 4194 SDValue Ops[1]; 4195 Ops[0] = DAG.getRoot(); 4196 SDValue Op = DAG.getNode(ISD::EXCEPTIONADDR, dl, VTs, Ops, 1); 4197 setValue(&I, Op); 4198 DAG.setRoot(Op.getValue(1)); 4199 return 0; 4200 } 4201 4202 case Intrinsic::eh_selector: { 4203 MachineBasicBlock *CallMBB = FuncInfo.MBB; 4204 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 4205 if (CallMBB->isLandingPad()) 4206 AddCatchInfo(I, &MMI, CallMBB); 4207 else { 4208#ifndef NDEBUG 4209 FuncInfo.CatchInfoLost.insert(&I); 4210#endif 4211 // FIXME: Mark exception selector register as live in. Hack for PR1508. 4212 unsigned Reg = TLI.getExceptionSelectorRegister(); 4213 if (Reg) FuncInfo.MBB->addLiveIn(Reg); 4214 } 4215 4216 // Insert the EHSELECTION instruction. 4217 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other); 4218 SDValue Ops[2]; 4219 Ops[0] = getValue(I.getArgOperand(0)); 4220 Ops[1] = getRoot(); 4221 SDValue Op = DAG.getNode(ISD::EHSELECTION, dl, VTs, Ops, 2); 4222 DAG.setRoot(Op.getValue(1)); 4223 setValue(&I, DAG.getSExtOrTrunc(Op, dl, MVT::i32)); 4224 return 0; 4225 } 4226 4227 case Intrinsic::eh_typeid_for: { 4228 // Find the type id for the given typeinfo. 4229 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0)); 4230 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV); 4231 Res = DAG.getConstant(TypeID, MVT::i32); 4232 setValue(&I, Res); 4233 return 0; 4234 } 4235 4236 case Intrinsic::eh_return_i32: 4237 case Intrinsic::eh_return_i64: 4238 DAG.getMachineFunction().getMMI().setCallsEHReturn(true); 4239 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl, 4240 MVT::Other, 4241 getControlRoot(), 4242 getValue(I.getArgOperand(0)), 4243 getValue(I.getArgOperand(1)))); 4244 return 0; 4245 case Intrinsic::eh_unwind_init: 4246 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true); 4247 return 0; 4248 case Intrinsic::eh_dwarf_cfa: { 4249 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), dl, 4250 TLI.getPointerTy()); 4251 SDValue Offset = DAG.getNode(ISD::ADD, dl, 4252 TLI.getPointerTy(), 4253 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl, 4254 TLI.getPointerTy()), 4255 CfaArg); 4256 SDValue FA = DAG.getNode(ISD::FRAMEADDR, dl, 4257 TLI.getPointerTy(), 4258 DAG.getConstant(0, TLI.getPointerTy())); 4259 setValue(&I, DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), 4260 FA, Offset)); 4261 return 0; 4262 } 4263 case Intrinsic::eh_sjlj_callsite: { 4264 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 4265 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0)); 4266 assert(CI && "Non-constant call site value in eh.sjlj.callsite!"); 4267 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!"); 4268 4269 MMI.setCurrentCallSite(CI->getZExtValue()); 4270 return 0; 4271 } 4272 case Intrinsic::eh_sjlj_setjmp: { 4273 setValue(&I, DAG.getNode(ISD::EH_SJLJ_SETJMP, dl, MVT::i32, getRoot(), 4274 getValue(I.getArgOperand(0)))); 4275 return 0; 4276 } 4277 case Intrinsic::eh_sjlj_longjmp: { 4278 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, dl, MVT::Other, 4279 getRoot(), 4280 getValue(I.getArgOperand(0)))); 4281 return 0; 4282 } 4283 4284 case Intrinsic::convertff: 4285 case Intrinsic::convertfsi: 4286 case Intrinsic::convertfui: 4287 case Intrinsic::convertsif: 4288 case Intrinsic::convertuif: 4289 case Intrinsic::convertss: 4290 case Intrinsic::convertsu: 4291 case Intrinsic::convertus: 4292 case Intrinsic::convertuu: { 4293 ISD::CvtCode Code = ISD::CVT_INVALID; 4294 switch (Intrinsic) { 4295 case Intrinsic::convertff: Code = ISD::CVT_FF; break; 4296 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break; 4297 case Intrinsic::convertfui: Code = ISD::CVT_FU; break; 4298 case Intrinsic::convertsif: Code = ISD::CVT_SF; break; 4299 case Intrinsic::convertuif: Code = ISD::CVT_UF; break; 4300 case Intrinsic::convertss: Code = ISD::CVT_SS; break; 4301 case Intrinsic::convertsu: Code = ISD::CVT_SU; break; 4302 case Intrinsic::convertus: Code = ISD::CVT_US; break; 4303 case Intrinsic::convertuu: Code = ISD::CVT_UU; break; 4304 } 4305 EVT DestVT = TLI.getValueType(I.getType()); 4306 const Value *Op1 = I.getArgOperand(0); 4307 Res = DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1), 4308 DAG.getValueType(DestVT), 4309 DAG.getValueType(getValue(Op1).getValueType()), 4310 getValue(I.getArgOperand(1)), 4311 getValue(I.getArgOperand(2)), 4312 Code); 4313 setValue(&I, Res); 4314 return 0; 4315 } 4316 case Intrinsic::sqrt: 4317 setValue(&I, DAG.getNode(ISD::FSQRT, dl, 4318 getValue(I.getArgOperand(0)).getValueType(), 4319 getValue(I.getArgOperand(0)))); 4320 return 0; 4321 case Intrinsic::powi: 4322 setValue(&I, ExpandPowI(dl, getValue(I.getArgOperand(0)), 4323 getValue(I.getArgOperand(1)), DAG)); 4324 return 0; 4325 case Intrinsic::sin: 4326 setValue(&I, DAG.getNode(ISD::FSIN, dl, 4327 getValue(I.getArgOperand(0)).getValueType(), 4328 getValue(I.getArgOperand(0)))); 4329 return 0; 4330 case Intrinsic::cos: 4331 setValue(&I, DAG.getNode(ISD::FCOS, dl, 4332 getValue(I.getArgOperand(0)).getValueType(), 4333 getValue(I.getArgOperand(0)))); 4334 return 0; 4335 case Intrinsic::log: 4336 visitLog(I); 4337 return 0; 4338 case Intrinsic::log2: 4339 visitLog2(I); 4340 return 0; 4341 case Intrinsic::log10: 4342 visitLog10(I); 4343 return 0; 4344 case Intrinsic::exp: 4345 visitExp(I); 4346 return 0; 4347 case Intrinsic::exp2: 4348 visitExp2(I); 4349 return 0; 4350 case Intrinsic::pow: 4351 visitPow(I); 4352 return 0; 4353 case Intrinsic::convert_to_fp16: 4354 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, dl, 4355 MVT::i16, getValue(I.getArgOperand(0)))); 4356 return 0; 4357 case Intrinsic::convert_from_fp16: 4358 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, dl, 4359 MVT::f32, getValue(I.getArgOperand(0)))); 4360 return 0; 4361 case Intrinsic::pcmarker: { 4362 SDValue Tmp = getValue(I.getArgOperand(0)); 4363 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp)); 4364 return 0; 4365 } 4366 case Intrinsic::readcyclecounter: { 4367 SDValue Op = getRoot(); 4368 Res = DAG.getNode(ISD::READCYCLECOUNTER, dl, 4369 DAG.getVTList(MVT::i64, MVT::Other), 4370 &Op, 1); 4371 setValue(&I, Res); 4372 DAG.setRoot(Res.getValue(1)); 4373 return 0; 4374 } 4375 case Intrinsic::bswap: 4376 setValue(&I, DAG.getNode(ISD::BSWAP, dl, 4377 getValue(I.getArgOperand(0)).getValueType(), 4378 getValue(I.getArgOperand(0)))); 4379 return 0; 4380 case Intrinsic::cttz: { 4381 SDValue Arg = getValue(I.getArgOperand(0)); 4382 EVT Ty = Arg.getValueType(); 4383 setValue(&I, DAG.getNode(ISD::CTTZ, dl, Ty, Arg)); 4384 return 0; 4385 } 4386 case Intrinsic::ctlz: { 4387 SDValue Arg = getValue(I.getArgOperand(0)); 4388 EVT Ty = Arg.getValueType(); 4389 setValue(&I, DAG.getNode(ISD::CTLZ, dl, Ty, Arg)); 4390 return 0; 4391 } 4392 case Intrinsic::ctpop: { 4393 SDValue Arg = getValue(I.getArgOperand(0)); 4394 EVT Ty = Arg.getValueType(); 4395 setValue(&I, DAG.getNode(ISD::CTPOP, dl, Ty, Arg)); 4396 return 0; 4397 } 4398 case Intrinsic::stacksave: { 4399 SDValue Op = getRoot(); 4400 Res = DAG.getNode(ISD::STACKSAVE, dl, 4401 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1); 4402 setValue(&I, Res); 4403 DAG.setRoot(Res.getValue(1)); 4404 return 0; 4405 } 4406 case Intrinsic::stackrestore: { 4407 Res = getValue(I.getArgOperand(0)); 4408 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Res)); 4409 return 0; 4410 } 4411 case Intrinsic::stackprotector: { 4412 // Emit code into the DAG to store the stack guard onto the stack. 4413 MachineFunction &MF = DAG.getMachineFunction(); 4414 MachineFrameInfo *MFI = MF.getFrameInfo(); 4415 EVT PtrTy = TLI.getPointerTy(); 4416 4417 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value. 4418 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1)); 4419 4420 int FI = FuncInfo.StaticAllocaMap[Slot]; 4421 MFI->setStackProtectorIndex(FI); 4422 4423 SDValue FIN = DAG.getFrameIndex(FI, PtrTy); 4424 4425 // Store the stack protector onto the stack. 4426 Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN, 4427 PseudoSourceValue::getFixedStack(FI), 4428 0, true, false, 0); 4429 setValue(&I, Res); 4430 DAG.setRoot(Res); 4431 return 0; 4432 } 4433 case Intrinsic::objectsize: { 4434 // If we don't know by now, we're never going to know. 4435 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1)); 4436 4437 assert(CI && "Non-constant type in __builtin_object_size?"); 4438 4439 SDValue Arg = getValue(I.getCalledValue()); 4440 EVT Ty = Arg.getValueType(); 4441 4442 if (CI->isZero()) 4443 Res = DAG.getConstant(-1ULL, Ty); 4444 else 4445 Res = DAG.getConstant(0, Ty); 4446 4447 setValue(&I, Res); 4448 return 0; 4449 } 4450 case Intrinsic::var_annotation: 4451 // Discard annotate attributes 4452 return 0; 4453 4454 case Intrinsic::init_trampoline: { 4455 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts()); 4456 4457 SDValue Ops[6]; 4458 Ops[0] = getRoot(); 4459 Ops[1] = getValue(I.getArgOperand(0)); 4460 Ops[2] = getValue(I.getArgOperand(1)); 4461 Ops[3] = getValue(I.getArgOperand(2)); 4462 Ops[4] = DAG.getSrcValue(I.getArgOperand(0)); 4463 Ops[5] = DAG.getSrcValue(F); 4464 4465 Res = DAG.getNode(ISD::TRAMPOLINE, dl, 4466 DAG.getVTList(TLI.getPointerTy(), MVT::Other), 4467 Ops, 6); 4468 4469 setValue(&I, Res); 4470 DAG.setRoot(Res.getValue(1)); 4471 return 0; 4472 } 4473 case Intrinsic::gcroot: 4474 if (GFI) { 4475 const Value *Alloca = I.getArgOperand(0); 4476 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1)); 4477 4478 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode()); 4479 GFI->addStackRoot(FI->getIndex(), TypeMap); 4480 } 4481 return 0; 4482 case Intrinsic::gcread: 4483 case Intrinsic::gcwrite: 4484 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!"); 4485 return 0; 4486 case Intrinsic::flt_rounds: 4487 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32)); 4488 return 0; 4489 case Intrinsic::trap: 4490 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot())); 4491 return 0; 4492 case Intrinsic::uadd_with_overflow: 4493 return implVisitAluOverflow(I, ISD::UADDO); 4494 case Intrinsic::sadd_with_overflow: 4495 return implVisitAluOverflow(I, ISD::SADDO); 4496 case Intrinsic::usub_with_overflow: 4497 return implVisitAluOverflow(I, ISD::USUBO); 4498 case Intrinsic::ssub_with_overflow: 4499 return implVisitAluOverflow(I, ISD::SSUBO); 4500 case Intrinsic::umul_with_overflow: 4501 return implVisitAluOverflow(I, ISD::UMULO); 4502 case Intrinsic::smul_with_overflow: 4503 return implVisitAluOverflow(I, ISD::SMULO); 4504 4505 case Intrinsic::prefetch: { 4506 SDValue Ops[4]; 4507 Ops[0] = getRoot(); 4508 Ops[1] = getValue(I.getArgOperand(0)); 4509 Ops[2] = getValue(I.getArgOperand(1)); 4510 Ops[3] = getValue(I.getArgOperand(2)); 4511 DAG.setRoot(DAG.getNode(ISD::PREFETCH, dl, MVT::Other, &Ops[0], 4)); 4512 return 0; 4513 } 4514 4515 case Intrinsic::memory_barrier: { 4516 SDValue Ops[6]; 4517 Ops[0] = getRoot(); 4518 for (int x = 1; x < 6; ++x) 4519 Ops[x] = getValue(I.getArgOperand(x - 1)); 4520 4521 DAG.setRoot(DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, &Ops[0], 6)); 4522 return 0; 4523 } 4524 case Intrinsic::atomic_cmp_swap: { 4525 SDValue Root = getRoot(); 4526 SDValue L = 4527 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, getCurDebugLoc(), 4528 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(), 4529 Root, 4530 getValue(I.getArgOperand(0)), 4531 getValue(I.getArgOperand(1)), 4532 getValue(I.getArgOperand(2)), 4533 I.getArgOperand(0)); 4534 setValue(&I, L); 4535 DAG.setRoot(L.getValue(1)); 4536 return 0; 4537 } 4538 case Intrinsic::atomic_load_add: 4539 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD); 4540 case Intrinsic::atomic_load_sub: 4541 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB); 4542 case Intrinsic::atomic_load_or: 4543 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR); 4544 case Intrinsic::atomic_load_xor: 4545 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR); 4546 case Intrinsic::atomic_load_and: 4547 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND); 4548 case Intrinsic::atomic_load_nand: 4549 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND); 4550 case Intrinsic::atomic_load_max: 4551 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX); 4552 case Intrinsic::atomic_load_min: 4553 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN); 4554 case Intrinsic::atomic_load_umin: 4555 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN); 4556 case Intrinsic::atomic_load_umax: 4557 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX); 4558 case Intrinsic::atomic_swap: 4559 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP); 4560 4561 case Intrinsic::invariant_start: 4562 case Intrinsic::lifetime_start: 4563 // Discard region information. 4564 setValue(&I, DAG.getUNDEF(TLI.getPointerTy())); 4565 return 0; 4566 case Intrinsic::invariant_end: 4567 case Intrinsic::lifetime_end: 4568 // Discard region information. 4569 return 0; 4570 } 4571} 4572 4573void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee, 4574 bool isTailCall, 4575 MachineBasicBlock *LandingPad) { 4576 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType()); 4577 const FunctionType *FTy = cast<FunctionType>(PT->getElementType()); 4578 const Type *RetTy = FTy->getReturnType(); 4579 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 4580 MCSymbol *BeginLabel = 0; 4581 4582 TargetLowering::ArgListTy Args; 4583 TargetLowering::ArgListEntry Entry; 4584 Args.reserve(CS.arg_size()); 4585 4586 // Check whether the function can return without sret-demotion. 4587 SmallVector<ISD::OutputArg, 4> Outs; 4588 SmallVector<uint64_t, 4> Offsets; 4589 GetReturnInfo(RetTy, CS.getAttributes().getRetAttributes(), 4590 Outs, TLI, &Offsets); 4591 4592 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(), 4593 FTy->isVarArg(), Outs, FTy->getContext()); 4594 4595 SDValue DemoteStackSlot; 4596 4597 if (!CanLowerReturn) { 4598 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize( 4599 FTy->getReturnType()); 4600 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment( 4601 FTy->getReturnType()); 4602 MachineFunction &MF = DAG.getMachineFunction(); 4603 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false); 4604 const Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType()); 4605 4606 DemoteStackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy()); 4607 Entry.Node = DemoteStackSlot; 4608 Entry.Ty = StackSlotPtrType; 4609 Entry.isSExt = false; 4610 Entry.isZExt = false; 4611 Entry.isInReg = false; 4612 Entry.isSRet = true; 4613 Entry.isNest = false; 4614 Entry.isByVal = false; 4615 Entry.Alignment = Align; 4616 Args.push_back(Entry); 4617 RetTy = Type::getVoidTy(FTy->getContext()); 4618 } 4619 4620 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end(); 4621 i != e; ++i) { 4622 SDValue ArgNode = getValue(*i); 4623 Entry.Node = ArgNode; Entry.Ty = (*i)->getType(); 4624 4625 unsigned attrInd = i - CS.arg_begin() + 1; 4626 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt); 4627 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt); 4628 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg); 4629 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet); 4630 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest); 4631 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal); 4632 Entry.Alignment = CS.getParamAlignment(attrInd); 4633 Args.push_back(Entry); 4634 } 4635 4636 if (LandingPad) { 4637 // Insert a label before the invoke call to mark the try range. This can be 4638 // used to detect deletion of the invoke via the MachineModuleInfo. 4639 BeginLabel = MMI.getContext().CreateTempSymbol(); 4640 4641 // For SjLj, keep track of which landing pads go with which invokes 4642 // so as to maintain the ordering of pads in the LSDA. 4643 unsigned CallSiteIndex = MMI.getCurrentCallSite(); 4644 if (CallSiteIndex) { 4645 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex); 4646 // Now that the call site is handled, stop tracking it. 4647 MMI.setCurrentCallSite(0); 4648 } 4649 4650 // Both PendingLoads and PendingExports must be flushed here; 4651 // this call might not return. 4652 (void)getRoot(); 4653 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getControlRoot(), BeginLabel)); 4654 } 4655 4656 // Check if target-independent constraints permit a tail call here. 4657 // Target-dependent constraints are checked within TLI.LowerCallTo. 4658 if (isTailCall && 4659 !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI)) 4660 isTailCall = false; 4661 4662 std::pair<SDValue,SDValue> Result = 4663 TLI.LowerCallTo(getRoot(), RetTy, 4664 CS.paramHasAttr(0, Attribute::SExt), 4665 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(), 4666 CS.paramHasAttr(0, Attribute::InReg), FTy->getNumParams(), 4667 CS.getCallingConv(), 4668 isTailCall, 4669 !CS.getInstruction()->use_empty(), 4670 Callee, Args, DAG, getCurDebugLoc()); 4671 assert((isTailCall || Result.second.getNode()) && 4672 "Non-null chain expected with non-tail call!"); 4673 assert((Result.second.getNode() || !Result.first.getNode()) && 4674 "Null value expected with tail call!"); 4675 if (Result.first.getNode()) { 4676 setValue(CS.getInstruction(), Result.first); 4677 } else if (!CanLowerReturn && Result.second.getNode()) { 4678 // The instruction result is the result of loading from the 4679 // hidden sret parameter. 4680 SmallVector<EVT, 1> PVTs; 4681 const Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType()); 4682 4683 ComputeValueVTs(TLI, PtrRetTy, PVTs); 4684 assert(PVTs.size() == 1 && "Pointers should fit in one register"); 4685 EVT PtrVT = PVTs[0]; 4686 unsigned NumValues = Outs.size(); 4687 SmallVector<SDValue, 4> Values(NumValues); 4688 SmallVector<SDValue, 4> Chains(NumValues); 4689 4690 for (unsigned i = 0; i < NumValues; ++i) { 4691 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, 4692 DemoteStackSlot, 4693 DAG.getConstant(Offsets[i], PtrVT)); 4694 SDValue L = DAG.getLoad(Outs[i].VT, getCurDebugLoc(), Result.second, 4695 Add, NULL, Offsets[i], false, false, 1); 4696 Values[i] = L; 4697 Chains[i] = L.getValue(1); 4698 } 4699 4700 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), 4701 MVT::Other, &Chains[0], NumValues); 4702 PendingLoads.push_back(Chain); 4703 4704 // Collect the legal value parts into potentially illegal values 4705 // that correspond to the original function's return values. 4706 SmallVector<EVT, 4> RetTys; 4707 RetTy = FTy->getReturnType(); 4708 ComputeValueVTs(TLI, RetTy, RetTys); 4709 ISD::NodeType AssertOp = ISD::DELETED_NODE; 4710 SmallVector<SDValue, 4> ReturnValues; 4711 unsigned CurReg = 0; 4712 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) { 4713 EVT VT = RetTys[I]; 4714 EVT RegisterVT = TLI.getRegisterType(RetTy->getContext(), VT); 4715 unsigned NumRegs = TLI.getNumRegisters(RetTy->getContext(), VT); 4716 4717 SDValue ReturnValue = 4718 getCopyFromParts(DAG, getCurDebugLoc(), &Values[CurReg], NumRegs, 4719 RegisterVT, VT, AssertOp); 4720 ReturnValues.push_back(ReturnValue); 4721 CurReg += NumRegs; 4722 } 4723 4724 setValue(CS.getInstruction(), 4725 DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(), 4726 DAG.getVTList(&RetTys[0], RetTys.size()), 4727 &ReturnValues[0], ReturnValues.size())); 4728 4729 } 4730 4731 // As a special case, a null chain means that a tail call has been emitted and 4732 // the DAG root is already updated. 4733 if (Result.second.getNode()) 4734 DAG.setRoot(Result.second); 4735 else 4736 HasTailCall = true; 4737 4738 if (LandingPad) { 4739 // Insert a label at the end of the invoke call to mark the try range. This 4740 // can be used to detect deletion of the invoke via the MachineModuleInfo. 4741 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol(); 4742 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getRoot(), EndLabel)); 4743 4744 // Inform MachineModuleInfo of range. 4745 MMI.addInvoke(LandingPad, BeginLabel, EndLabel); 4746 } 4747} 4748 4749/// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the 4750/// value is equal or not-equal to zero. 4751static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) { 4752 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); 4753 UI != E; ++UI) { 4754 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI)) 4755 if (IC->isEquality()) 4756 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1))) 4757 if (C->isNullValue()) 4758 continue; 4759 // Unknown instruction. 4760 return false; 4761 } 4762 return true; 4763} 4764 4765static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT, 4766 const Type *LoadTy, 4767 SelectionDAGBuilder &Builder) { 4768 4769 // Check to see if this load can be trivially constant folded, e.g. if the 4770 // input is from a string literal. 4771 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) { 4772 // Cast pointer to the type we really want to load. 4773 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput), 4774 PointerType::getUnqual(LoadTy)); 4775 4776 if (const Constant *LoadCst = 4777 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput), 4778 Builder.TD)) 4779 return Builder.getValue(LoadCst); 4780 } 4781 4782 // Otherwise, we have to emit the load. If the pointer is to unfoldable but 4783 // still constant memory, the input chain can be the entry node. 4784 SDValue Root; 4785 bool ConstantMemory = false; 4786 4787 // Do not serialize (non-volatile) loads of constant memory with anything. 4788 if (Builder.AA->pointsToConstantMemory(PtrVal)) { 4789 Root = Builder.DAG.getEntryNode(); 4790 ConstantMemory = true; 4791 } else { 4792 // Do not serialize non-volatile loads against each other. 4793 Root = Builder.DAG.getRoot(); 4794 } 4795 4796 SDValue Ptr = Builder.getValue(PtrVal); 4797 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root, 4798 Ptr, PtrVal /*SrcValue*/, 0/*SVOffset*/, 4799 false /*volatile*/, 4800 false /*nontemporal*/, 1 /* align=1 */); 4801 4802 if (!ConstantMemory) 4803 Builder.PendingLoads.push_back(LoadVal.getValue(1)); 4804 return LoadVal; 4805} 4806 4807 4808/// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form. 4809/// If so, return true and lower it, otherwise return false and it will be 4810/// lowered like a normal call. 4811bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) { 4812 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t) 4813 if (I.getNumArgOperands() != 3) 4814 return false; 4815 4816 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1); 4817 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() || 4818 !I.getArgOperand(2)->getType()->isIntegerTy() || 4819 !I.getType()->isIntegerTy()) 4820 return false; 4821 4822 const ConstantInt *Size = dyn_cast<ConstantInt>(I.getArgOperand(2)); 4823 4824 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0 4825 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0 4826 if (Size && IsOnlyUsedInZeroEqualityComparison(&I)) { 4827 bool ActuallyDoIt = true; 4828 MVT LoadVT; 4829 const Type *LoadTy; 4830 switch (Size->getZExtValue()) { 4831 default: 4832 LoadVT = MVT::Other; 4833 LoadTy = 0; 4834 ActuallyDoIt = false; 4835 break; 4836 case 2: 4837 LoadVT = MVT::i16; 4838 LoadTy = Type::getInt16Ty(Size->getContext()); 4839 break; 4840 case 4: 4841 LoadVT = MVT::i32; 4842 LoadTy = Type::getInt32Ty(Size->getContext()); 4843 break; 4844 case 8: 4845 LoadVT = MVT::i64; 4846 LoadTy = Type::getInt64Ty(Size->getContext()); 4847 break; 4848 /* 4849 case 16: 4850 LoadVT = MVT::v4i32; 4851 LoadTy = Type::getInt32Ty(Size->getContext()); 4852 LoadTy = VectorType::get(LoadTy, 4); 4853 break; 4854 */ 4855 } 4856 4857 // This turns into unaligned loads. We only do this if the target natively 4858 // supports the MVT we'll be loading or if it is small enough (<= 4) that 4859 // we'll only produce a small number of byte loads. 4860 4861 // Require that we can find a legal MVT, and only do this if the target 4862 // supports unaligned loads of that type. Expanding into byte loads would 4863 // bloat the code. 4864 if (ActuallyDoIt && Size->getZExtValue() > 4) { 4865 // TODO: Handle 5 byte compare as 4-byte + 1 byte. 4866 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads. 4867 if (!TLI.isTypeLegal(LoadVT) ||!TLI.allowsUnalignedMemoryAccesses(LoadVT)) 4868 ActuallyDoIt = false; 4869 } 4870 4871 if (ActuallyDoIt) { 4872 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this); 4873 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this); 4874 4875 SDValue Res = DAG.getSetCC(getCurDebugLoc(), MVT::i1, LHSVal, RHSVal, 4876 ISD::SETNE); 4877 EVT CallVT = TLI.getValueType(I.getType(), true); 4878 setValue(&I, DAG.getZExtOrTrunc(Res, getCurDebugLoc(), CallVT)); 4879 return true; 4880 } 4881 } 4882 4883 4884 return false; 4885} 4886 4887 4888void SelectionDAGBuilder::visitCall(const CallInst &I) { 4889 // Handle inline assembly differently. 4890 if (isa<InlineAsm>(I.getCalledValue())) { 4891 visitInlineAsm(&I); 4892 return; 4893 } 4894 4895 const char *RenameFn = 0; 4896 if (Function *F = I.getCalledFunction()) { 4897 if (F->isDeclaration()) { 4898 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) { 4899 if (unsigned IID = II->getIntrinsicID(F)) { 4900 RenameFn = visitIntrinsicCall(I, IID); 4901 if (!RenameFn) 4902 return; 4903 } 4904 } 4905 if (unsigned IID = F->getIntrinsicID()) { 4906 RenameFn = visitIntrinsicCall(I, IID); 4907 if (!RenameFn) 4908 return; 4909 } 4910 } 4911 4912 // Check for well-known libc/libm calls. If the function is internal, it 4913 // can't be a library call. 4914 if (!F->hasLocalLinkage() && F->hasName()) { 4915 StringRef Name = F->getName(); 4916 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl") { 4917 if (I.getNumArgOperands() == 2 && // Basic sanity checks. 4918 I.getArgOperand(0)->getType()->isFloatingPointTy() && 4919 I.getType() == I.getArgOperand(0)->getType() && 4920 I.getType() == I.getArgOperand(1)->getType()) { 4921 SDValue LHS = getValue(I.getArgOperand(0)); 4922 SDValue RHS = getValue(I.getArgOperand(1)); 4923 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(), 4924 LHS.getValueType(), LHS, RHS)); 4925 return; 4926 } 4927 } else if (Name == "fabs" || Name == "fabsf" || Name == "fabsl") { 4928 if (I.getNumArgOperands() == 1 && // Basic sanity checks. 4929 I.getArgOperand(0)->getType()->isFloatingPointTy() && 4930 I.getType() == I.getArgOperand(0)->getType()) { 4931 SDValue Tmp = getValue(I.getArgOperand(0)); 4932 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(), 4933 Tmp.getValueType(), Tmp)); 4934 return; 4935 } 4936 } else if (Name == "sin" || Name == "sinf" || Name == "sinl") { 4937 if (I.getNumArgOperands() == 1 && // Basic sanity checks. 4938 I.getArgOperand(0)->getType()->isFloatingPointTy() && 4939 I.getType() == I.getArgOperand(0)->getType() && 4940 I.onlyReadsMemory()) { 4941 SDValue Tmp = getValue(I.getArgOperand(0)); 4942 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(), 4943 Tmp.getValueType(), Tmp)); 4944 return; 4945 } 4946 } else if (Name == "cos" || Name == "cosf" || Name == "cosl") { 4947 if (I.getNumArgOperands() == 1 && // Basic sanity checks. 4948 I.getArgOperand(0)->getType()->isFloatingPointTy() && 4949 I.getType() == I.getArgOperand(0)->getType() && 4950 I.onlyReadsMemory()) { 4951 SDValue Tmp = getValue(I.getArgOperand(0)); 4952 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(), 4953 Tmp.getValueType(), Tmp)); 4954 return; 4955 } 4956 } else if (Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") { 4957 if (I.getNumArgOperands() == 1 && // Basic sanity checks. 4958 I.getArgOperand(0)->getType()->isFloatingPointTy() && 4959 I.getType() == I.getArgOperand(0)->getType() && 4960 I.onlyReadsMemory()) { 4961 SDValue Tmp = getValue(I.getArgOperand(0)); 4962 setValue(&I, DAG.getNode(ISD::FSQRT, getCurDebugLoc(), 4963 Tmp.getValueType(), Tmp)); 4964 return; 4965 } 4966 } else if (Name == "memcmp") { 4967 if (visitMemCmpCall(I)) 4968 return; 4969 } 4970 } 4971 } 4972 4973 SDValue Callee; 4974 if (!RenameFn) 4975 Callee = getValue(I.getCalledValue()); 4976 else 4977 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy()); 4978 4979 // Check if we can potentially perform a tail call. More detailed checking is 4980 // be done within LowerCallTo, after more information about the call is known. 4981 LowerCallTo(&I, Callee, I.isTailCall()); 4982} 4983 4984namespace llvm { 4985 4986/// AsmOperandInfo - This contains information for each constraint that we are 4987/// lowering. 4988class LLVM_LIBRARY_VISIBILITY SDISelAsmOperandInfo : 4989 public TargetLowering::AsmOperandInfo { 4990public: 4991 /// CallOperand - If this is the result output operand or a clobber 4992 /// this is null, otherwise it is the incoming operand to the CallInst. 4993 /// This gets modified as the asm is processed. 4994 SDValue CallOperand; 4995 4996 /// AssignedRegs - If this is a register or register class operand, this 4997 /// contains the set of register corresponding to the operand. 4998 RegsForValue AssignedRegs; 4999 5000 explicit SDISelAsmOperandInfo(const InlineAsm::ConstraintInfo &info) 5001 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) { 5002 } 5003 5004 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers 5005 /// busy in OutputRegs/InputRegs. 5006 void MarkAllocatedRegs(bool isOutReg, bool isInReg, 5007 std::set<unsigned> &OutputRegs, 5008 std::set<unsigned> &InputRegs, 5009 const TargetRegisterInfo &TRI) const { 5010 if (isOutReg) { 5011 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i) 5012 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI); 5013 } 5014 if (isInReg) { 5015 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i) 5016 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI); 5017 } 5018 } 5019 5020 /// getCallOperandValEVT - Return the EVT of the Value* that this operand 5021 /// corresponds to. If there is no Value* for this operand, it returns 5022 /// MVT::Other. 5023 EVT getCallOperandValEVT(LLVMContext &Context, 5024 const TargetLowering &TLI, 5025 const TargetData *TD) const { 5026 if (CallOperandVal == 0) return MVT::Other; 5027 5028 if (isa<BasicBlock>(CallOperandVal)) 5029 return TLI.getPointerTy(); 5030 5031 const llvm::Type *OpTy = CallOperandVal->getType(); 5032 5033 // If this is an indirect operand, the operand is a pointer to the 5034 // accessed type. 5035 if (isIndirect) { 5036 const llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy); 5037 if (!PtrTy) 5038 report_fatal_error("Indirect operand for inline asm not a pointer!"); 5039 OpTy = PtrTy->getElementType(); 5040 } 5041 5042 // If OpTy is not a single value, it may be a struct/union that we 5043 // can tile with integers. 5044 if (!OpTy->isSingleValueType() && OpTy->isSized()) { 5045 unsigned BitSize = TD->getTypeSizeInBits(OpTy); 5046 switch (BitSize) { 5047 default: break; 5048 case 1: 5049 case 8: 5050 case 16: 5051 case 32: 5052 case 64: 5053 case 128: 5054 OpTy = IntegerType::get(Context, BitSize); 5055 break; 5056 } 5057 } 5058 5059 return TLI.getValueType(OpTy, true); 5060 } 5061 5062private: 5063 /// MarkRegAndAliases - Mark the specified register and all aliases in the 5064 /// specified set. 5065 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs, 5066 const TargetRegisterInfo &TRI) { 5067 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg"); 5068 Regs.insert(Reg); 5069 if (const unsigned *Aliases = TRI.getAliasSet(Reg)) 5070 for (; *Aliases; ++Aliases) 5071 Regs.insert(*Aliases); 5072 } 5073}; 5074 5075} // end llvm namespace. 5076 5077/// isAllocatableRegister - If the specified register is safe to allocate, 5078/// i.e. it isn't a stack pointer or some other special register, return the 5079/// register class for the register. Otherwise, return null. 5080static const TargetRegisterClass * 5081isAllocatableRegister(unsigned Reg, MachineFunction &MF, 5082 const TargetLowering &TLI, 5083 const TargetRegisterInfo *TRI) { 5084 EVT FoundVT = MVT::Other; 5085 const TargetRegisterClass *FoundRC = 0; 5086 for (TargetRegisterInfo::regclass_iterator RCI = TRI->regclass_begin(), 5087 E = TRI->regclass_end(); RCI != E; ++RCI) { 5088 EVT ThisVT = MVT::Other; 5089 5090 const TargetRegisterClass *RC = *RCI; 5091 // If none of the value types for this register class are valid, we 5092 // can't use it. For example, 64-bit reg classes on 32-bit targets. 5093 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end(); 5094 I != E; ++I) { 5095 if (TLI.isTypeLegal(*I)) { 5096 // If we have already found this register in a different register class, 5097 // choose the one with the largest VT specified. For example, on 5098 // PowerPC, we favor f64 register classes over f32. 5099 if (FoundVT == MVT::Other || FoundVT.bitsLT(*I)) { 5100 ThisVT = *I; 5101 break; 5102 } 5103 } 5104 } 5105 5106 if (ThisVT == MVT::Other) continue; 5107 5108 // NOTE: This isn't ideal. In particular, this might allocate the 5109 // frame pointer in functions that need it (due to them not being taken 5110 // out of allocation, because a variable sized allocation hasn't been seen 5111 // yet). This is a slight code pessimization, but should still work. 5112 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF), 5113 E = RC->allocation_order_end(MF); I != E; ++I) 5114 if (*I == Reg) { 5115 // We found a matching register class. Keep looking at others in case 5116 // we find one with larger registers that this physreg is also in. 5117 FoundRC = RC; 5118 FoundVT = ThisVT; 5119 break; 5120 } 5121 } 5122 return FoundRC; 5123} 5124 5125/// GetRegistersForValue - Assign registers (virtual or physical) for the 5126/// specified operand. We prefer to assign virtual registers, to allow the 5127/// register allocator to handle the assignment process. However, if the asm 5128/// uses features that we can't model on machineinstrs, we have SDISel do the 5129/// allocation. This produces generally horrible, but correct, code. 5130/// 5131/// OpInfo describes the operand. 5132/// Input and OutputRegs are the set of already allocated physical registers. 5133/// 5134void SelectionDAGBuilder:: 5135GetRegistersForValue(SDISelAsmOperandInfo &OpInfo, 5136 std::set<unsigned> &OutputRegs, 5137 std::set<unsigned> &InputRegs) { 5138 LLVMContext &Context = FuncInfo.Fn->getContext(); 5139 5140 // Compute whether this value requires an input register, an output register, 5141 // or both. 5142 bool isOutReg = false; 5143 bool isInReg = false; 5144 switch (OpInfo.Type) { 5145 case InlineAsm::isOutput: 5146 isOutReg = true; 5147 5148 // If there is an input constraint that matches this, we need to reserve 5149 // the input register so no other inputs allocate to it. 5150 isInReg = OpInfo.hasMatchingInput(); 5151 break; 5152 case InlineAsm::isInput: 5153 isInReg = true; 5154 isOutReg = false; 5155 break; 5156 case InlineAsm::isClobber: 5157 isOutReg = true; 5158 isInReg = true; 5159 break; 5160 } 5161 5162 5163 MachineFunction &MF = DAG.getMachineFunction(); 5164 SmallVector<unsigned, 4> Regs; 5165 5166 // If this is a constraint for a single physreg, or a constraint for a 5167 // register class, find it. 5168 std::pair<unsigned, const TargetRegisterClass*> PhysReg = 5169 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode, 5170 OpInfo.ConstraintVT); 5171 5172 unsigned NumRegs = 1; 5173 if (OpInfo.ConstraintVT != MVT::Other) { 5174 // If this is a FP input in an integer register (or visa versa) insert a bit 5175 // cast of the input value. More generally, handle any case where the input 5176 // value disagrees with the register class we plan to stick this in. 5177 if (OpInfo.Type == InlineAsm::isInput && 5178 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) { 5179 // Try to convert to the first EVT that the reg class contains. If the 5180 // types are identical size, use a bitcast to convert (e.g. two differing 5181 // vector types). 5182 EVT RegVT = *PhysReg.second->vt_begin(); 5183 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) { 5184 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), 5185 RegVT, OpInfo.CallOperand); 5186 OpInfo.ConstraintVT = RegVT; 5187 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) { 5188 // If the input is a FP value and we want it in FP registers, do a 5189 // bitcast to the corresponding integer type. This turns an f64 value 5190 // into i64, which can be passed with two i32 values on a 32-bit 5191 // machine. 5192 RegVT = EVT::getIntegerVT(Context, 5193 OpInfo.ConstraintVT.getSizeInBits()); 5194 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), 5195 RegVT, OpInfo.CallOperand); 5196 OpInfo.ConstraintVT = RegVT; 5197 } 5198 } 5199 5200 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT); 5201 } 5202 5203 EVT RegVT; 5204 EVT ValueVT = OpInfo.ConstraintVT; 5205 5206 // If this is a constraint for a specific physical register, like {r17}, 5207 // assign it now. 5208 if (unsigned AssignedReg = PhysReg.first) { 5209 const TargetRegisterClass *RC = PhysReg.second; 5210 if (OpInfo.ConstraintVT == MVT::Other) 5211 ValueVT = *RC->vt_begin(); 5212 5213 // Get the actual register value type. This is important, because the user 5214 // may have asked for (e.g.) the AX register in i32 type. We need to 5215 // remember that AX is actually i16 to get the right extension. 5216 RegVT = *RC->vt_begin(); 5217 5218 // This is a explicit reference to a physical register. 5219 Regs.push_back(AssignedReg); 5220 5221 // If this is an expanded reference, add the rest of the regs to Regs. 5222 if (NumRegs != 1) { 5223 TargetRegisterClass::iterator I = RC->begin(); 5224 for (; *I != AssignedReg; ++I) 5225 assert(I != RC->end() && "Didn't find reg!"); 5226 5227 // Already added the first reg. 5228 --NumRegs; ++I; 5229 for (; NumRegs; --NumRegs, ++I) { 5230 assert(I != RC->end() && "Ran out of registers to allocate!"); 5231 Regs.push_back(*I); 5232 } 5233 } 5234 5235 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT); 5236 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo(); 5237 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI); 5238 return; 5239 } 5240 5241 // Otherwise, if this was a reference to an LLVM register class, create vregs 5242 // for this reference. 5243 if (const TargetRegisterClass *RC = PhysReg.second) { 5244 RegVT = *RC->vt_begin(); 5245 if (OpInfo.ConstraintVT == MVT::Other) 5246 ValueVT = RegVT; 5247 5248 // Create the appropriate number of virtual registers. 5249 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 5250 for (; NumRegs; --NumRegs) 5251 Regs.push_back(RegInfo.createVirtualRegister(RC)); 5252 5253 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT); 5254 return; 5255 } 5256 5257 // This is a reference to a register class that doesn't directly correspond 5258 // to an LLVM register class. Allocate NumRegs consecutive, available, 5259 // registers from the class. 5260 std::vector<unsigned> RegClassRegs 5261 = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode, 5262 OpInfo.ConstraintVT); 5263 5264 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo(); 5265 unsigned NumAllocated = 0; 5266 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) { 5267 unsigned Reg = RegClassRegs[i]; 5268 // See if this register is available. 5269 if ((isOutReg && OutputRegs.count(Reg)) || // Already used. 5270 (isInReg && InputRegs.count(Reg))) { // Already used. 5271 // Make sure we find consecutive registers. 5272 NumAllocated = 0; 5273 continue; 5274 } 5275 5276 // Check to see if this register is allocatable (i.e. don't give out the 5277 // stack pointer). 5278 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, TRI); 5279 if (!RC) { // Couldn't allocate this register. 5280 // Reset NumAllocated to make sure we return consecutive registers. 5281 NumAllocated = 0; 5282 continue; 5283 } 5284 5285 // Okay, this register is good, we can use it. 5286 ++NumAllocated; 5287 5288 // If we allocated enough consecutive registers, succeed. 5289 if (NumAllocated == NumRegs) { 5290 unsigned RegStart = (i-NumAllocated)+1; 5291 unsigned RegEnd = i+1; 5292 // Mark all of the allocated registers used. 5293 for (unsigned i = RegStart; i != RegEnd; ++i) 5294 Regs.push_back(RegClassRegs[i]); 5295 5296 OpInfo.AssignedRegs = RegsForValue(Regs, *RC->vt_begin(), 5297 OpInfo.ConstraintVT); 5298 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI); 5299 return; 5300 } 5301 } 5302 5303 // Otherwise, we couldn't allocate enough registers for this. 5304} 5305 5306/// visitInlineAsm - Handle a call to an InlineAsm object. 5307/// 5308void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) { 5309 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 5310 5311 /// ConstraintOperands - Information about all of the constraints. 5312 std::vector<SDISelAsmOperandInfo> ConstraintOperands; 5313 5314 std::set<unsigned> OutputRegs, InputRegs; 5315 5316 // Do a prepass over the constraints, canonicalizing them, and building up the 5317 // ConstraintOperands list. 5318 std::vector<InlineAsm::ConstraintInfo> 5319 ConstraintInfos = IA->ParseConstraints(); 5320 5321 bool hasMemory = hasInlineAsmMemConstraint(ConstraintInfos, TLI); 5322 5323 SDValue Chain, Flag; 5324 5325 // We won't need to flush pending loads if this asm doesn't touch 5326 // memory and is nonvolatile. 5327 if (hasMemory || IA->hasSideEffects()) 5328 Chain = getRoot(); 5329 else 5330 Chain = DAG.getRoot(); 5331 5332 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 5333 unsigned ResNo = 0; // ResNo - The result number of the next output. 5334 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 5335 ConstraintOperands.push_back(SDISelAsmOperandInfo(ConstraintInfos[i])); 5336 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back(); 5337 5338 EVT OpVT = MVT::Other; 5339 5340 // Compute the value type for each operand. 5341 switch (OpInfo.Type) { 5342 case InlineAsm::isOutput: 5343 // Indirect outputs just consume an argument. 5344 if (OpInfo.isIndirect) { 5345 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 5346 break; 5347 } 5348 5349 // The return value of the call is this value. As such, there is no 5350 // corresponding argument. 5351 assert(!CS.getType()->isVoidTy() && 5352 "Bad inline asm!"); 5353 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) { 5354 OpVT = TLI.getValueType(STy->getElementType(ResNo)); 5355 } else { 5356 assert(ResNo == 0 && "Asm only has one result!"); 5357 OpVT = TLI.getValueType(CS.getType()); 5358 } 5359 ++ResNo; 5360 break; 5361 case InlineAsm::isInput: 5362 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 5363 break; 5364 case InlineAsm::isClobber: 5365 // Nothing to do. 5366 break; 5367 } 5368 5369 // If this is an input or an indirect output, process the call argument. 5370 // BasicBlocks are labels, currently appearing only in asm's. 5371 if (OpInfo.CallOperandVal) { 5372 // Strip bitcasts, if any. This mostly comes up for functions. 5373 OpInfo.CallOperandVal = OpInfo.CallOperandVal->stripPointerCasts(); 5374 5375 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) { 5376 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]); 5377 } else { 5378 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal); 5379 } 5380 5381 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, TD); 5382 } 5383 5384 OpInfo.ConstraintVT = OpVT; 5385 } 5386 5387 // Second pass over the constraints: compute which constraint option to use 5388 // and assign registers to constraints that want a specific physreg. 5389 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 5390 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; 5391 5392 // If this is an output operand with a matching input operand, look up the 5393 // matching input. If their types mismatch, e.g. one is an integer, the 5394 // other is floating point, or their sizes are different, flag it as an 5395 // error. 5396 if (OpInfo.hasMatchingInput()) { 5397 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; 5398 5399 if (OpInfo.ConstraintVT != Input.ConstraintVT) { 5400 if ((OpInfo.ConstraintVT.isInteger() != 5401 Input.ConstraintVT.isInteger()) || 5402 (OpInfo.ConstraintVT.getSizeInBits() != 5403 Input.ConstraintVT.getSizeInBits())) { 5404 report_fatal_error("Unsupported asm: input constraint" 5405 " with a matching output constraint of" 5406 " incompatible type!"); 5407 } 5408 Input.ConstraintVT = OpInfo.ConstraintVT; 5409 } 5410 } 5411 5412 // Compute the constraint code and ConstraintType to use. 5413 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG); 5414 5415 // If this is a memory input, and if the operand is not indirect, do what we 5416 // need to to provide an address for the memory input. 5417 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 5418 !OpInfo.isIndirect) { 5419 assert(OpInfo.Type == InlineAsm::isInput && 5420 "Can only indirectify direct input operands!"); 5421 5422 // Memory operands really want the address of the value. If we don't have 5423 // an indirect input, put it in the constpool if we can, otherwise spill 5424 // it to a stack slot. 5425 5426 // If the operand is a float, integer, or vector constant, spill to a 5427 // constant pool entry to get its address. 5428 const Value *OpVal = OpInfo.CallOperandVal; 5429 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) || 5430 isa<ConstantVector>(OpVal)) { 5431 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal), 5432 TLI.getPointerTy()); 5433 } else { 5434 // Otherwise, create a stack slot and emit a store to it before the 5435 // asm. 5436 const Type *Ty = OpVal->getType(); 5437 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty); 5438 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty); 5439 MachineFunction &MF = DAG.getMachineFunction(); 5440 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false); 5441 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy()); 5442 Chain = DAG.getStore(Chain, getCurDebugLoc(), 5443 OpInfo.CallOperand, StackSlot, NULL, 0, 5444 false, false, 0); 5445 OpInfo.CallOperand = StackSlot; 5446 } 5447 5448 // There is no longer a Value* corresponding to this operand. 5449 OpInfo.CallOperandVal = 0; 5450 5451 // It is now an indirect operand. 5452 OpInfo.isIndirect = true; 5453 } 5454 5455 // If this constraint is for a specific register, allocate it before 5456 // anything else. 5457 if (OpInfo.ConstraintType == TargetLowering::C_Register) 5458 GetRegistersForValue(OpInfo, OutputRegs, InputRegs); 5459 } 5460 5461 ConstraintInfos.clear(); 5462 5463 // Second pass - Loop over all of the operands, assigning virtual or physregs 5464 // to register class operands. 5465 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { 5466 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; 5467 5468 // C_Register operands have already been allocated, Other/Memory don't need 5469 // to be. 5470 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass) 5471 GetRegistersForValue(OpInfo, OutputRegs, InputRegs); 5472 } 5473 5474 // AsmNodeOperands - The operands for the ISD::INLINEASM node. 5475 std::vector<SDValue> AsmNodeOperands; 5476 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain 5477 AsmNodeOperands.push_back( 5478 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(), 5479 TLI.getPointerTy())); 5480 5481 // If we have a !srcloc metadata node associated with it, we want to attach 5482 // this to the ultimately generated inline asm machineinstr. To do this, we 5483 // pass in the third operand as this (potentially null) inline asm MDNode. 5484 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc"); 5485 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc)); 5486 5487 // Remember the AlignStack bit as operand 3. 5488 AsmNodeOperands.push_back(DAG.getTargetConstant(IA->isAlignStack() ? 1 : 0, 5489 MVT::i1)); 5490 5491 // Loop over all of the inputs, copying the operand values into the 5492 // appropriate registers and processing the output regs. 5493 RegsForValue RetValRegs; 5494 5495 // IndirectStoresToEmit - The set of stores to emit after the inline asm node. 5496 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit; 5497 5498 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { 5499 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; 5500 5501 switch (OpInfo.Type) { 5502 case InlineAsm::isOutput: { 5503 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass && 5504 OpInfo.ConstraintType != TargetLowering::C_Register) { 5505 // Memory output, or 'other' output (e.g. 'X' constraint). 5506 assert(OpInfo.isIndirect && "Memory output must be indirect operand"); 5507 5508 // Add information to the INLINEASM node to know about this output. 5509 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1); 5510 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags, 5511 TLI.getPointerTy())); 5512 AsmNodeOperands.push_back(OpInfo.CallOperand); 5513 break; 5514 } 5515 5516 // Otherwise, this is a register or register class output. 5517 5518 // Copy the output from the appropriate register. Find a register that 5519 // we can use. 5520 if (OpInfo.AssignedRegs.Regs.empty()) 5521 report_fatal_error("Couldn't allocate output reg for constraint '" + 5522 Twine(OpInfo.ConstraintCode) + "'!"); 5523 5524 // If this is an indirect operand, store through the pointer after the 5525 // asm. 5526 if (OpInfo.isIndirect) { 5527 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs, 5528 OpInfo.CallOperandVal)); 5529 } else { 5530 // This is the result value of the call. 5531 assert(!CS.getType()->isVoidTy() && "Bad inline asm!"); 5532 // Concatenate this output onto the outputs list. 5533 RetValRegs.append(OpInfo.AssignedRegs); 5534 } 5535 5536 // Add information to the INLINEASM node to know that this register is 5537 // set. 5538 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ? 5539 InlineAsm::Kind_RegDefEarlyClobber : 5540 InlineAsm::Kind_RegDef, 5541 false, 5542 0, 5543 DAG, 5544 AsmNodeOperands); 5545 break; 5546 } 5547 case InlineAsm::isInput: { 5548 SDValue InOperandVal = OpInfo.CallOperand; 5549 5550 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint? 5551 // If this is required to match an output register we have already set, 5552 // just use its register. 5553 unsigned OperandNo = OpInfo.getMatchedOperand(); 5554 5555 // Scan until we find the definition we already emitted of this operand. 5556 // When we find it, create a RegsForValue operand. 5557 unsigned CurOp = InlineAsm::Op_FirstOperand; 5558 for (; OperandNo; --OperandNo) { 5559 // Advance to the next operand. 5560 unsigned OpFlag = 5561 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue(); 5562 assert((InlineAsm::isRegDefKind(OpFlag) || 5563 InlineAsm::isRegDefEarlyClobberKind(OpFlag) || 5564 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?"); 5565 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1; 5566 } 5567 5568 unsigned OpFlag = 5569 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue(); 5570 if (InlineAsm::isRegDefKind(OpFlag) || 5571 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) { 5572 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs. 5573 if (OpInfo.isIndirect) { 5574 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c 5575 LLVMContext &Ctx = *DAG.getContext(); 5576 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:" 5577 " don't know how to handle tied " 5578 "indirect register inputs"); 5579 } 5580 5581 RegsForValue MatchedRegs; 5582 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType()); 5583 EVT RegVT = AsmNodeOperands[CurOp+1].getValueType(); 5584 MatchedRegs.RegVTs.push_back(RegVT); 5585 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo(); 5586 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag); 5587 i != e; ++i) 5588 MatchedRegs.Regs.push_back 5589 (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT))); 5590 5591 // Use the produced MatchedRegs object to 5592 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(), 5593 Chain, &Flag); 5594 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, 5595 true, OpInfo.getMatchedOperand(), 5596 DAG, AsmNodeOperands); 5597 break; 5598 } 5599 5600 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!"); 5601 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 && 5602 "Unexpected number of operands"); 5603 // Add information to the INLINEASM node to know about this input. 5604 // See InlineAsm.h isUseOperandTiedToDef. 5605 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag, 5606 OpInfo.getMatchedOperand()); 5607 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag, 5608 TLI.getPointerTy())); 5609 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]); 5610 break; 5611 } 5612 5613 // Treat indirect 'X' constraint as memory. 5614 if (OpInfo.ConstraintType == TargetLowering::C_Other && 5615 OpInfo.isIndirect) 5616 OpInfo.ConstraintType = TargetLowering::C_Memory; 5617 5618 if (OpInfo.ConstraintType == TargetLowering::C_Other) { 5619 std::vector<SDValue> Ops; 5620 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0], 5621 Ops, DAG); 5622 if (Ops.empty()) 5623 report_fatal_error("Invalid operand for inline asm constraint '" + 5624 Twine(OpInfo.ConstraintCode) + "'!"); 5625 5626 // Add information to the INLINEASM node to know about this input. 5627 unsigned ResOpType = 5628 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size()); 5629 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType, 5630 TLI.getPointerTy())); 5631 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end()); 5632 break; 5633 } 5634 5635 if (OpInfo.ConstraintType == TargetLowering::C_Memory) { 5636 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!"); 5637 assert(InOperandVal.getValueType() == TLI.getPointerTy() && 5638 "Memory operands expect pointer values"); 5639 5640 // Add information to the INLINEASM node to know about this input. 5641 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1); 5642 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType, 5643 TLI.getPointerTy())); 5644 AsmNodeOperands.push_back(InOperandVal); 5645 break; 5646 } 5647 5648 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass || 5649 OpInfo.ConstraintType == TargetLowering::C_Register) && 5650 "Unknown constraint type!"); 5651 assert(!OpInfo.isIndirect && 5652 "Don't know how to handle indirect register inputs yet!"); 5653 5654 // Copy the input into the appropriate registers. 5655 if (OpInfo.AssignedRegs.Regs.empty() || 5656 !OpInfo.AssignedRegs.areValueTypesLegal(TLI)) 5657 report_fatal_error("Couldn't allocate input reg for constraint '" + 5658 Twine(OpInfo.ConstraintCode) + "'!"); 5659 5660 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(), 5661 Chain, &Flag); 5662 5663 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0, 5664 DAG, AsmNodeOperands); 5665 break; 5666 } 5667 case InlineAsm::isClobber: { 5668 // Add the clobbered value to the operand list, so that the register 5669 // allocator is aware that the physreg got clobbered. 5670 if (!OpInfo.AssignedRegs.Regs.empty()) 5671 OpInfo.AssignedRegs.AddInlineAsmOperands( 5672 InlineAsm::Kind_RegDefEarlyClobber, 5673 false, 0, DAG, 5674 AsmNodeOperands); 5675 break; 5676 } 5677 } 5678 } 5679 5680 // Finish up input operands. Set the input chain and add the flag last. 5681 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain; 5682 if (Flag.getNode()) AsmNodeOperands.push_back(Flag); 5683 5684 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(), 5685 DAG.getVTList(MVT::Other, MVT::Flag), 5686 &AsmNodeOperands[0], AsmNodeOperands.size()); 5687 Flag = Chain.getValue(1); 5688 5689 // If this asm returns a register value, copy the result from that register 5690 // and set it as the value of the call. 5691 if (!RetValRegs.Regs.empty()) { 5692 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), 5693 Chain, &Flag); 5694 5695 // FIXME: Why don't we do this for inline asms with MRVs? 5696 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) { 5697 EVT ResultType = TLI.getValueType(CS.getType()); 5698 5699 // If any of the results of the inline asm is a vector, it may have the 5700 // wrong width/num elts. This can happen for register classes that can 5701 // contain multiple different value types. The preg or vreg allocated may 5702 // not have the same VT as was expected. Convert it to the right type 5703 // with bit_convert. 5704 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) { 5705 Val = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), 5706 ResultType, Val); 5707 5708 } else if (ResultType != Val.getValueType() && 5709 ResultType.isInteger() && Val.getValueType().isInteger()) { 5710 // If a result value was tied to an input value, the computed result may 5711 // have a wider width than the expected result. Extract the relevant 5712 // portion. 5713 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val); 5714 } 5715 5716 assert(ResultType == Val.getValueType() && "Asm result value mismatch!"); 5717 } 5718 5719 setValue(CS.getInstruction(), Val); 5720 // Don't need to use this as a chain in this case. 5721 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty()) 5722 return; 5723 } 5724 5725 std::vector<std::pair<SDValue, const Value *> > StoresToEmit; 5726 5727 // Process indirect outputs, first output all of the flagged copies out of 5728 // physregs. 5729 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) { 5730 RegsForValue &OutRegs = IndirectStoresToEmit[i].first; 5731 const Value *Ptr = IndirectStoresToEmit[i].second; 5732 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), 5733 Chain, &Flag); 5734 StoresToEmit.push_back(std::make_pair(OutVal, Ptr)); 5735 } 5736 5737 // Emit the non-flagged stores from the physregs. 5738 SmallVector<SDValue, 8> OutChains; 5739 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) { 5740 SDValue Val = DAG.getStore(Chain, getCurDebugLoc(), 5741 StoresToEmit[i].first, 5742 getValue(StoresToEmit[i].second), 5743 StoresToEmit[i].second, 0, 5744 false, false, 0); 5745 OutChains.push_back(Val); 5746 } 5747 5748 if (!OutChains.empty()) 5749 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other, 5750 &OutChains[0], OutChains.size()); 5751 5752 DAG.setRoot(Chain); 5753} 5754 5755void SelectionDAGBuilder::visitVAStart(const CallInst &I) { 5756 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(), 5757 MVT::Other, getRoot(), 5758 getValue(I.getArgOperand(0)), 5759 DAG.getSrcValue(I.getArgOperand(0)))); 5760} 5761 5762void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) { 5763 const TargetData &TD = *TLI.getTargetData(); 5764 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(), 5765 getRoot(), getValue(I.getOperand(0)), 5766 DAG.getSrcValue(I.getOperand(0)), 5767 TD.getABITypeAlignment(I.getType())); 5768 setValue(&I, V); 5769 DAG.setRoot(V.getValue(1)); 5770} 5771 5772void SelectionDAGBuilder::visitVAEnd(const CallInst &I) { 5773 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(), 5774 MVT::Other, getRoot(), 5775 getValue(I.getArgOperand(0)), 5776 DAG.getSrcValue(I.getArgOperand(0)))); 5777} 5778 5779void SelectionDAGBuilder::visitVACopy(const CallInst &I) { 5780 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(), 5781 MVT::Other, getRoot(), 5782 getValue(I.getArgOperand(0)), 5783 getValue(I.getArgOperand(1)), 5784 DAG.getSrcValue(I.getArgOperand(0)), 5785 DAG.getSrcValue(I.getArgOperand(1)))); 5786} 5787 5788/// TargetLowering::LowerCallTo - This is the default LowerCallTo 5789/// implementation, which just calls LowerCall. 5790/// FIXME: When all targets are 5791/// migrated to using LowerCall, this hook should be integrated into SDISel. 5792std::pair<SDValue, SDValue> 5793TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy, 5794 bool RetSExt, bool RetZExt, bool isVarArg, 5795 bool isInreg, unsigned NumFixedArgs, 5796 CallingConv::ID CallConv, bool isTailCall, 5797 bool isReturnValueUsed, 5798 SDValue Callee, 5799 ArgListTy &Args, SelectionDAG &DAG, 5800 DebugLoc dl) const { 5801 // Handle all of the outgoing arguments. 5802 SmallVector<ISD::OutputArg, 32> Outs; 5803 SmallVector<SDValue, 32> OutVals; 5804 for (unsigned i = 0, e = Args.size(); i != e; ++i) { 5805 SmallVector<EVT, 4> ValueVTs; 5806 ComputeValueVTs(*this, Args[i].Ty, ValueVTs); 5807 for (unsigned Value = 0, NumValues = ValueVTs.size(); 5808 Value != NumValues; ++Value) { 5809 EVT VT = ValueVTs[Value]; 5810 const Type *ArgTy = VT.getTypeForEVT(RetTy->getContext()); 5811 SDValue Op = SDValue(Args[i].Node.getNode(), 5812 Args[i].Node.getResNo() + Value); 5813 ISD::ArgFlagsTy Flags; 5814 unsigned OriginalAlignment = 5815 getTargetData()->getABITypeAlignment(ArgTy); 5816 5817 if (Args[i].isZExt) 5818 Flags.setZExt(); 5819 if (Args[i].isSExt) 5820 Flags.setSExt(); 5821 if (Args[i].isInReg) 5822 Flags.setInReg(); 5823 if (Args[i].isSRet) 5824 Flags.setSRet(); 5825 if (Args[i].isByVal) { 5826 Flags.setByVal(); 5827 const PointerType *Ty = cast<PointerType>(Args[i].Ty); 5828 const Type *ElementTy = Ty->getElementType(); 5829 unsigned FrameAlign = getByValTypeAlignment(ElementTy); 5830 unsigned FrameSize = getTargetData()->getTypeAllocSize(ElementTy); 5831 // For ByVal, alignment should come from FE. BE will guess if this 5832 // info is not there but there are cases it cannot get right. 5833 if (Args[i].Alignment) 5834 FrameAlign = Args[i].Alignment; 5835 Flags.setByValAlign(FrameAlign); 5836 Flags.setByValSize(FrameSize); 5837 } 5838 if (Args[i].isNest) 5839 Flags.setNest(); 5840 Flags.setOrigAlign(OriginalAlignment); 5841 5842 EVT PartVT = getRegisterType(RetTy->getContext(), VT); 5843 unsigned NumParts = getNumRegisters(RetTy->getContext(), VT); 5844 SmallVector<SDValue, 4> Parts(NumParts); 5845 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 5846 5847 if (Args[i].isSExt) 5848 ExtendKind = ISD::SIGN_EXTEND; 5849 else if (Args[i].isZExt) 5850 ExtendKind = ISD::ZERO_EXTEND; 5851 5852 getCopyToParts(DAG, dl, Op, &Parts[0], NumParts, 5853 PartVT, ExtendKind); 5854 5855 for (unsigned j = 0; j != NumParts; ++j) { 5856 // if it isn't first piece, alignment must be 1 5857 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(), 5858 i < NumFixedArgs); 5859 if (NumParts > 1 && j == 0) 5860 MyFlags.Flags.setSplit(); 5861 else if (j != 0) 5862 MyFlags.Flags.setOrigAlign(1); 5863 5864 Outs.push_back(MyFlags); 5865 OutVals.push_back(Parts[j]); 5866 } 5867 } 5868 } 5869 5870 // Handle the incoming return values from the call. 5871 SmallVector<ISD::InputArg, 32> Ins; 5872 SmallVector<EVT, 4> RetTys; 5873 ComputeValueVTs(*this, RetTy, RetTys); 5874 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) { 5875 EVT VT = RetTys[I]; 5876 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT); 5877 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT); 5878 for (unsigned i = 0; i != NumRegs; ++i) { 5879 ISD::InputArg MyFlags; 5880 MyFlags.VT = RegisterVT; 5881 MyFlags.Used = isReturnValueUsed; 5882 if (RetSExt) 5883 MyFlags.Flags.setSExt(); 5884 if (RetZExt) 5885 MyFlags.Flags.setZExt(); 5886 if (isInreg) 5887 MyFlags.Flags.setInReg(); 5888 Ins.push_back(MyFlags); 5889 } 5890 } 5891 5892 SmallVector<SDValue, 4> InVals; 5893 Chain = LowerCall(Chain, Callee, CallConv, isVarArg, isTailCall, 5894 Outs, OutVals, Ins, dl, DAG, InVals); 5895 5896 // Verify that the target's LowerCall behaved as expected. 5897 assert(Chain.getNode() && Chain.getValueType() == MVT::Other && 5898 "LowerCall didn't return a valid chain!"); 5899 assert((!isTailCall || InVals.empty()) && 5900 "LowerCall emitted a return value for a tail call!"); 5901 assert((isTailCall || InVals.size() == Ins.size()) && 5902 "LowerCall didn't emit the correct number of values!"); 5903 5904 // For a tail call, the return value is merely live-out and there aren't 5905 // any nodes in the DAG representing it. Return a special value to 5906 // indicate that a tail call has been emitted and no more Instructions 5907 // should be processed in the current block. 5908 if (isTailCall) { 5909 DAG.setRoot(Chain); 5910 return std::make_pair(SDValue(), SDValue()); 5911 } 5912 5913 DEBUG(for (unsigned i = 0, e = Ins.size(); i != e; ++i) { 5914 assert(InVals[i].getNode() && 5915 "LowerCall emitted a null value!"); 5916 assert(Ins[i].VT == InVals[i].getValueType() && 5917 "LowerCall emitted a value with the wrong type!"); 5918 }); 5919 5920 // Collect the legal value parts into potentially illegal values 5921 // that correspond to the original function's return values. 5922 ISD::NodeType AssertOp = ISD::DELETED_NODE; 5923 if (RetSExt) 5924 AssertOp = ISD::AssertSext; 5925 else if (RetZExt) 5926 AssertOp = ISD::AssertZext; 5927 SmallVector<SDValue, 4> ReturnValues; 5928 unsigned CurReg = 0; 5929 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) { 5930 EVT VT = RetTys[I]; 5931 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT); 5932 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT); 5933 5934 ReturnValues.push_back(getCopyFromParts(DAG, dl, &InVals[CurReg], 5935 NumRegs, RegisterVT, VT, 5936 AssertOp)); 5937 CurReg += NumRegs; 5938 } 5939 5940 // For a function returning void, there is no return value. We can't create 5941 // such a node, so we just return a null return value in that case. In 5942 // that case, nothing will actualy look at the value. 5943 if (ReturnValues.empty()) 5944 return std::make_pair(SDValue(), Chain); 5945 5946 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl, 5947 DAG.getVTList(&RetTys[0], RetTys.size()), 5948 &ReturnValues[0], ReturnValues.size()); 5949 return std::make_pair(Res, Chain); 5950} 5951 5952void TargetLowering::LowerOperationWrapper(SDNode *N, 5953 SmallVectorImpl<SDValue> &Results, 5954 SelectionDAG &DAG) const { 5955 SDValue Res = LowerOperation(SDValue(N, 0), DAG); 5956 if (Res.getNode()) 5957 Results.push_back(Res); 5958} 5959 5960SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { 5961 llvm_unreachable("LowerOperation not implemented for this target!"); 5962 return SDValue(); 5963} 5964 5965void 5966SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) { 5967 SDValue Op = getNonRegisterValue(V); 5968 assert((Op.getOpcode() != ISD::CopyFromReg || 5969 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) && 5970 "Copy from a reg to the same reg!"); 5971 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg"); 5972 5973 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType()); 5974 SDValue Chain = DAG.getEntryNode(); 5975 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0); 5976 PendingExports.push_back(Chain); 5977} 5978 5979#include "llvm/CodeGen/SelectionDAGISel.h" 5980 5981void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) { 5982 // If this is the entry block, emit arguments. 5983 const Function &F = *LLVMBB->getParent(); 5984 SelectionDAG &DAG = SDB->DAG; 5985 DebugLoc dl = SDB->getCurDebugLoc(); 5986 const TargetData *TD = TLI.getTargetData(); 5987 SmallVector<ISD::InputArg, 16> Ins; 5988 5989 // Check whether the function can return without sret-demotion. 5990 SmallVector<ISD::OutputArg, 4> Outs; 5991 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(), 5992 Outs, TLI); 5993 5994 if (!FuncInfo->CanLowerReturn) { 5995 // Put in an sret pointer parameter before all the other parameters. 5996 SmallVector<EVT, 1> ValueVTs; 5997 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs); 5998 5999 // NOTE: Assuming that a pointer will never break down to more than one VT 6000 // or one register. 6001 ISD::ArgFlagsTy Flags; 6002 Flags.setSRet(); 6003 EVT RegisterVT = TLI.getRegisterType(*DAG.getContext(), ValueVTs[0]); 6004 ISD::InputArg RetArg(Flags, RegisterVT, true); 6005 Ins.push_back(RetArg); 6006 } 6007 6008 // Set up the incoming argument description vector. 6009 unsigned Idx = 1; 6010 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); 6011 I != E; ++I, ++Idx) { 6012 SmallVector<EVT, 4> ValueVTs; 6013 ComputeValueVTs(TLI, I->getType(), ValueVTs); 6014 bool isArgValueUsed = !I->use_empty(); 6015 for (unsigned Value = 0, NumValues = ValueVTs.size(); 6016 Value != NumValues; ++Value) { 6017 EVT VT = ValueVTs[Value]; 6018 const Type *ArgTy = VT.getTypeForEVT(*DAG.getContext()); 6019 ISD::ArgFlagsTy Flags; 6020 unsigned OriginalAlignment = 6021 TD->getABITypeAlignment(ArgTy); 6022 6023 if (F.paramHasAttr(Idx, Attribute::ZExt)) 6024 Flags.setZExt(); 6025 if (F.paramHasAttr(Idx, Attribute::SExt)) 6026 Flags.setSExt(); 6027 if (F.paramHasAttr(Idx, Attribute::InReg)) 6028 Flags.setInReg(); 6029 if (F.paramHasAttr(Idx, Attribute::StructRet)) 6030 Flags.setSRet(); 6031 if (F.paramHasAttr(Idx, Attribute::ByVal)) { 6032 Flags.setByVal(); 6033 const PointerType *Ty = cast<PointerType>(I->getType()); 6034 const Type *ElementTy = Ty->getElementType(); 6035 unsigned FrameAlign = TLI.getByValTypeAlignment(ElementTy); 6036 unsigned FrameSize = TD->getTypeAllocSize(ElementTy); 6037 // For ByVal, alignment should be passed from FE. BE will guess if 6038 // this info is not there but there are cases it cannot get right. 6039 if (F.getParamAlignment(Idx)) 6040 FrameAlign = F.getParamAlignment(Idx); 6041 Flags.setByValAlign(FrameAlign); 6042 Flags.setByValSize(FrameSize); 6043 } 6044 if (F.paramHasAttr(Idx, Attribute::Nest)) 6045 Flags.setNest(); 6046 Flags.setOrigAlign(OriginalAlignment); 6047 6048 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), VT); 6049 unsigned NumRegs = TLI.getNumRegisters(*CurDAG->getContext(), VT); 6050 for (unsigned i = 0; i != NumRegs; ++i) { 6051 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed); 6052 if (NumRegs > 1 && i == 0) 6053 MyFlags.Flags.setSplit(); 6054 // if it isn't first piece, alignment must be 1 6055 else if (i > 0) 6056 MyFlags.Flags.setOrigAlign(1); 6057 Ins.push_back(MyFlags); 6058 } 6059 } 6060 } 6061 6062 // Call the target to set up the argument values. 6063 SmallVector<SDValue, 8> InVals; 6064 SDValue NewRoot = TLI.LowerFormalArguments(DAG.getRoot(), F.getCallingConv(), 6065 F.isVarArg(), Ins, 6066 dl, DAG, InVals); 6067 6068 // Verify that the target's LowerFormalArguments behaved as expected. 6069 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other && 6070 "LowerFormalArguments didn't return a valid chain!"); 6071 assert(InVals.size() == Ins.size() && 6072 "LowerFormalArguments didn't emit the correct number of values!"); 6073 DEBUG({ 6074 for (unsigned i = 0, e = Ins.size(); i != e; ++i) { 6075 assert(InVals[i].getNode() && 6076 "LowerFormalArguments emitted a null value!"); 6077 assert(Ins[i].VT == InVals[i].getValueType() && 6078 "LowerFormalArguments emitted a value with the wrong type!"); 6079 } 6080 }); 6081 6082 // Update the DAG with the new chain value resulting from argument lowering. 6083 DAG.setRoot(NewRoot); 6084 6085 // Set up the argument values. 6086 unsigned i = 0; 6087 Idx = 1; 6088 if (!FuncInfo->CanLowerReturn) { 6089 // Create a virtual register for the sret pointer, and put in a copy 6090 // from the sret argument into it. 6091 SmallVector<EVT, 1> ValueVTs; 6092 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs); 6093 EVT VT = ValueVTs[0]; 6094 EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT); 6095 ISD::NodeType AssertOp = ISD::DELETED_NODE; 6096 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1, 6097 RegVT, VT, AssertOp); 6098 6099 MachineFunction& MF = SDB->DAG.getMachineFunction(); 6100 MachineRegisterInfo& RegInfo = MF.getRegInfo(); 6101 unsigned SRetReg = RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)); 6102 FuncInfo->DemoteRegister = SRetReg; 6103 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurDebugLoc(), 6104 SRetReg, ArgValue); 6105 DAG.setRoot(NewRoot); 6106 6107 // i indexes lowered arguments. Bump it past the hidden sret argument. 6108 // Idx indexes LLVM arguments. Don't touch it. 6109 ++i; 6110 } 6111 6112 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; 6113 ++I, ++Idx) { 6114 SmallVector<SDValue, 4> ArgValues; 6115 SmallVector<EVT, 4> ValueVTs; 6116 ComputeValueVTs(TLI, I->getType(), ValueVTs); 6117 unsigned NumValues = ValueVTs.size(); 6118 6119 // If this argument is unused then remember its value. It is used to generate 6120 // debugging information. 6121 if (I->use_empty() && NumValues) 6122 SDB->setUnusedArgValue(I, InVals[i]); 6123 6124 for (unsigned Value = 0; Value != NumValues; ++Value) { 6125 EVT VT = ValueVTs[Value]; 6126 EVT PartVT = TLI.getRegisterType(*CurDAG->getContext(), VT); 6127 unsigned NumParts = TLI.getNumRegisters(*CurDAG->getContext(), VT); 6128 6129 if (!I->use_empty()) { 6130 ISD::NodeType AssertOp = ISD::DELETED_NODE; 6131 if (F.paramHasAttr(Idx, Attribute::SExt)) 6132 AssertOp = ISD::AssertSext; 6133 else if (F.paramHasAttr(Idx, Attribute::ZExt)) 6134 AssertOp = ISD::AssertZext; 6135 6136 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i], 6137 NumParts, PartVT, VT, 6138 AssertOp)); 6139 } 6140 6141 i += NumParts; 6142 } 6143 6144 if (!I->use_empty()) { 6145 SDValue Res; 6146 if (!ArgValues.empty()) 6147 Res = DAG.getMergeValues(&ArgValues[0], NumValues, 6148 SDB->getCurDebugLoc()); 6149 SDB->setValue(I, Res); 6150 6151 // If this argument is live outside of the entry block, insert a copy from 6152 // whereever we got it to the vreg that other BB's will reference it as. 6153 SDB->CopyToExportRegsIfNeeded(I); 6154 } 6155 } 6156 6157 assert(i == InVals.size() && "Argument register count mismatch!"); 6158 6159 // Finally, if the target has anything special to do, allow it to do so. 6160 // FIXME: this should insert code into the DAG! 6161 EmitFunctionEntryCode(); 6162} 6163 6164/// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to 6165/// ensure constants are generated when needed. Remember the virtual registers 6166/// that need to be added to the Machine PHI nodes as input. We cannot just 6167/// directly add them, because expansion might result in multiple MBB's for one 6168/// BB. As such, the start of the BB might correspond to a different MBB than 6169/// the end. 6170/// 6171void 6172SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) { 6173 const TerminatorInst *TI = LLVMBB->getTerminator(); 6174 6175 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled; 6176 6177 // Check successor nodes' PHI nodes that expect a constant to be available 6178 // from this block. 6179 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) { 6180 const BasicBlock *SuccBB = TI->getSuccessor(succ); 6181 if (!isa<PHINode>(SuccBB->begin())) continue; 6182 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB]; 6183 6184 // If this terminator has multiple identical successors (common for 6185 // switches), only handle each succ once. 6186 if (!SuccsHandled.insert(SuccMBB)) continue; 6187 6188 MachineBasicBlock::iterator MBBI = SuccMBB->begin(); 6189 6190 // At this point we know that there is a 1-1 correspondence between LLVM PHI 6191 // nodes and Machine PHI nodes, but the incoming operands have not been 6192 // emitted yet. 6193 for (BasicBlock::const_iterator I = SuccBB->begin(); 6194 const PHINode *PN = dyn_cast<PHINode>(I); ++I) { 6195 // Ignore dead phi's. 6196 if (PN->use_empty()) continue; 6197 6198 unsigned Reg; 6199 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB); 6200 6201 if (const Constant *C = dyn_cast<Constant>(PHIOp)) { 6202 unsigned &RegOut = ConstantsOut[C]; 6203 if (RegOut == 0) { 6204 RegOut = FuncInfo.CreateRegs(C->getType()); 6205 CopyValueToVirtualRegister(C, RegOut); 6206 } 6207 Reg = RegOut; 6208 } else { 6209 DenseMap<const Value *, unsigned>::iterator I = 6210 FuncInfo.ValueMap.find(PHIOp); 6211 if (I != FuncInfo.ValueMap.end()) 6212 Reg = I->second; 6213 else { 6214 assert(isa<AllocaInst>(PHIOp) && 6215 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) && 6216 "Didn't codegen value into a register!??"); 6217 Reg = FuncInfo.CreateRegs(PHIOp->getType()); 6218 CopyValueToVirtualRegister(PHIOp, Reg); 6219 } 6220 } 6221 6222 // Remember that this register needs to added to the machine PHI node as 6223 // the input for this MBB. 6224 SmallVector<EVT, 4> ValueVTs; 6225 ComputeValueVTs(TLI, PN->getType(), ValueVTs); 6226 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) { 6227 EVT VT = ValueVTs[vti]; 6228 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT); 6229 for (unsigned i = 0, e = NumRegisters; i != e; ++i) 6230 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i)); 6231 Reg += NumRegisters; 6232 } 6233 } 6234 } 6235 ConstantsOut.clear(); 6236} 6237