SelectionDAGBuilder.cpp revision 72cf01cc7c8f668a17e7fdfe6aaed50e164cac1b
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 "SelectionDAGBuilder.h" 16#include "SDNodeDbgValue.h" 17#include "llvm/ADT/BitVector.h" 18#include "llvm/ADT/Optional.h" 19#include "llvm/ADT/SmallSet.h" 20#include "llvm/Analysis/AliasAnalysis.h" 21#include "llvm/Analysis/BranchProbabilityInfo.h" 22#include "llvm/Analysis/ConstantFolding.h" 23#include "llvm/Analysis/ValueTracking.h" 24#include "llvm/CodeGen/Analysis.h" 25#include "llvm/CodeGen/FastISel.h" 26#include "llvm/CodeGen/FunctionLoweringInfo.h" 27#include "llvm/CodeGen/GCMetadata.h" 28#include "llvm/CodeGen/GCStrategy.h" 29#include "llvm/CodeGen/MachineFrameInfo.h" 30#include "llvm/CodeGen/MachineFunction.h" 31#include "llvm/CodeGen/MachineInstrBuilder.h" 32#include "llvm/CodeGen/MachineJumpTableInfo.h" 33#include "llvm/CodeGen/MachineModuleInfo.h" 34#include "llvm/CodeGen/MachineRegisterInfo.h" 35#include "llvm/CodeGen/SelectionDAG.h" 36#include "llvm/CodeGen/StackMaps.h" 37#include "llvm/DebugInfo.h" 38#include "llvm/IR/CallingConv.h" 39#include "llvm/IR/Constants.h" 40#include "llvm/IR/DataLayout.h" 41#include "llvm/IR/DerivedTypes.h" 42#include "llvm/IR/Function.h" 43#include "llvm/IR/GlobalVariable.h" 44#include "llvm/IR/InlineAsm.h" 45#include "llvm/IR/Instructions.h" 46#include "llvm/IR/IntrinsicInst.h" 47#include "llvm/IR/Intrinsics.h" 48#include "llvm/IR/LLVMContext.h" 49#include "llvm/IR/Module.h" 50#include "llvm/Support/CommandLine.h" 51#include "llvm/Support/Debug.h" 52#include "llvm/Support/ErrorHandling.h" 53#include "llvm/Support/MathExtras.h" 54#include "llvm/Support/raw_ostream.h" 55#include "llvm/Target/TargetFrameLowering.h" 56#include "llvm/Target/TargetInstrInfo.h" 57#include "llvm/Target/TargetIntrinsicInfo.h" 58#include "llvm/Target/TargetLibraryInfo.h" 59#include "llvm/Target/TargetLowering.h" 60#include "llvm/Target/TargetOptions.h" 61#include "llvm/Target/TargetSelectionDAGInfo.h" 62#include <algorithm> 63using namespace llvm; 64 65/// LimitFloatPrecision - Generate low-precision inline sequences for 66/// some float libcalls (6, 8 or 12 bits). 67static unsigned LimitFloatPrecision; 68 69static cl::opt<unsigned, true> 70LimitFPPrecision("limit-float-precision", 71 cl::desc("Generate low-precision inline sequences " 72 "for some float libcalls"), 73 cl::location(LimitFloatPrecision), 74 cl::init(0)); 75 76// Limit the width of DAG chains. This is important in general to prevent 77// prevent DAG-based analysis from blowing up. For example, alias analysis and 78// load clustering may not complete in reasonable time. It is difficult to 79// recognize and avoid this situation within each individual analysis, and 80// future analyses are likely to have the same behavior. Limiting DAG width is 81// the safe approach, and will be especially important with global DAGs. 82// 83// MaxParallelChains default is arbitrarily high to avoid affecting 84// optimization, but could be lowered to improve compile time. Any ld-ld-st-st 85// sequence over this should have been converted to llvm.memcpy by the 86// frontend. It easy to induce this behavior with .ll code such as: 87// %buffer = alloca [4096 x i8] 88// %data = load [4096 x i8]* %argPtr 89// store [4096 x i8] %data, [4096 x i8]* %buffer 90static const unsigned MaxParallelChains = 64; 91 92static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL, 93 const SDValue *Parts, unsigned NumParts, 94 MVT PartVT, EVT ValueVT, const Value *V); 95 96/// getCopyFromParts - Create a value that contains the specified legal parts 97/// combined into the value they represent. If the parts combine to a type 98/// larger then ValueVT then AssertOp can be used to specify whether the extra 99/// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT 100/// (ISD::AssertSext). 101static SDValue getCopyFromParts(SelectionDAG &DAG, SDLoc DL, 102 const SDValue *Parts, 103 unsigned NumParts, MVT PartVT, EVT ValueVT, 104 const Value *V, 105 ISD::NodeType AssertOp = ISD::DELETED_NODE) { 106 if (ValueVT.isVector()) 107 return getCopyFromPartsVector(DAG, DL, Parts, NumParts, 108 PartVT, ValueVT, V); 109 110 assert(NumParts > 0 && "No parts to assemble!"); 111 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 112 SDValue Val = Parts[0]; 113 114 if (NumParts > 1) { 115 // Assemble the value from multiple parts. 116 if (ValueVT.isInteger()) { 117 unsigned PartBits = PartVT.getSizeInBits(); 118 unsigned ValueBits = ValueVT.getSizeInBits(); 119 120 // Assemble the power of 2 part. 121 unsigned RoundParts = NumParts & (NumParts - 1) ? 122 1 << Log2_32(NumParts) : NumParts; 123 unsigned RoundBits = PartBits * RoundParts; 124 EVT RoundVT = RoundBits == ValueBits ? 125 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits); 126 SDValue Lo, Hi; 127 128 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2); 129 130 if (RoundParts > 2) { 131 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2, 132 PartVT, HalfVT, V); 133 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2, 134 RoundParts / 2, PartVT, HalfVT, V); 135 } else { 136 Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]); 137 Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]); 138 } 139 140 if (TLI.isBigEndian()) 141 std::swap(Lo, Hi); 142 143 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi); 144 145 if (RoundParts < NumParts) { 146 // Assemble the trailing non-power-of-2 part. 147 unsigned OddParts = NumParts - RoundParts; 148 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits); 149 Hi = getCopyFromParts(DAG, DL, 150 Parts + RoundParts, OddParts, PartVT, OddVT, V); 151 152 // Combine the round and odd parts. 153 Lo = Val; 154 if (TLI.isBigEndian()) 155 std::swap(Lo, Hi); 156 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); 157 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi); 158 Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi, 159 DAG.getConstant(Lo.getValueType().getSizeInBits(), 160 TLI.getPointerTy())); 161 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo); 162 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi); 163 } 164 } else if (PartVT.isFloatingPoint()) { 165 // FP split into multiple FP parts (for ppcf128) 166 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 && 167 "Unexpected split"); 168 SDValue Lo, Hi; 169 Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]); 170 Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]); 171 if (TLI.isBigEndian()) 172 std::swap(Lo, Hi); 173 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi); 174 } else { 175 // FP split into integer parts (soft fp) 176 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() && 177 !PartVT.isVector() && "Unexpected split"); 178 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits()); 179 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT, V); 180 } 181 } 182 183 // There is now one part, held in Val. Correct it to match ValueVT. 184 EVT PartEVT = Val.getValueType(); 185 186 if (PartEVT == ValueVT) 187 return Val; 188 189 if (PartEVT.isInteger() && ValueVT.isInteger()) { 190 if (ValueVT.bitsLT(PartEVT)) { 191 // For a truncate, see if we have any information to 192 // indicate whether the truncated bits will always be 193 // zero or sign-extension. 194 if (AssertOp != ISD::DELETED_NODE) 195 Val = DAG.getNode(AssertOp, DL, PartEVT, Val, 196 DAG.getValueType(ValueVT)); 197 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); 198 } 199 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val); 200 } 201 202 if (PartEVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { 203 // FP_ROUND's are always exact here. 204 if (ValueVT.bitsLT(Val.getValueType())) 205 return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val, 206 DAG.getTargetConstant(1, TLI.getPointerTy())); 207 208 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val); 209 } 210 211 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits()) 212 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val); 213 214 llvm_unreachable("Unknown mismatch!"); 215} 216 217/// getCopyFromPartsVector - Create a value that contains the specified legal 218/// parts combined into the value they represent. If the parts combine to a 219/// type larger then ValueVT then AssertOp can be used to specify whether the 220/// extra bits are known to be zero (ISD::AssertZext) or sign extended from 221/// ValueVT (ISD::AssertSext). 222static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL, 223 const SDValue *Parts, unsigned NumParts, 224 MVT PartVT, EVT ValueVT, const Value *V) { 225 assert(ValueVT.isVector() && "Not a vector value"); 226 assert(NumParts > 0 && "No parts to assemble!"); 227 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 228 SDValue Val = Parts[0]; 229 230 // Handle a multi-element vector. 231 if (NumParts > 1) { 232 EVT IntermediateVT; 233 MVT RegisterVT; 234 unsigned NumIntermediates; 235 unsigned NumRegs = 236 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT, 237 NumIntermediates, RegisterVT); 238 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!"); 239 NumParts = NumRegs; // Silence a compiler warning. 240 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!"); 241 assert(RegisterVT == Parts[0].getSimpleValueType() && 242 "Part type doesn't match part!"); 243 244 // Assemble the parts into intermediate operands. 245 SmallVector<SDValue, 8> Ops(NumIntermediates); 246 if (NumIntermediates == NumParts) { 247 // If the register was not expanded, truncate or copy the value, 248 // as appropriate. 249 for (unsigned i = 0; i != NumParts; ++i) 250 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1, 251 PartVT, IntermediateVT, V); 252 } else if (NumParts > 0) { 253 // If the intermediate type was expanded, build the intermediate 254 // operands from the parts. 255 assert(NumParts % NumIntermediates == 0 && 256 "Must expand into a divisible number of parts!"); 257 unsigned Factor = NumParts / NumIntermediates; 258 for (unsigned i = 0; i != NumIntermediates; ++i) 259 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor, 260 PartVT, IntermediateVT, V); 261 } 262 263 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the 264 // intermediate operands. 265 Val = DAG.getNode(IntermediateVT.isVector() ? 266 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL, 267 ValueVT, &Ops[0], NumIntermediates); 268 } 269 270 // There is now one part, held in Val. Correct it to match ValueVT. 271 EVT PartEVT = Val.getValueType(); 272 273 if (PartEVT == ValueVT) 274 return Val; 275 276 if (PartEVT.isVector()) { 277 // If the element type of the source/dest vectors are the same, but the 278 // parts vector has more elements than the value vector, then we have a 279 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the 280 // elements we want. 281 if (PartEVT.getVectorElementType() == ValueVT.getVectorElementType()) { 282 assert(PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() && 283 "Cannot narrow, it would be a lossy transformation"); 284 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val, 285 DAG.getConstant(0, TLI.getVectorIdxTy())); 286 } 287 288 // Vector/Vector bitcast. 289 if (ValueVT.getSizeInBits() == PartEVT.getSizeInBits()) 290 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val); 291 292 assert(PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() && 293 "Cannot handle this kind of promotion"); 294 // Promoted vector extract 295 bool Smaller = ValueVT.bitsLE(PartEVT); 296 return DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND), 297 DL, ValueVT, Val); 298 299 } 300 301 // Trivial bitcast if the types are the same size and the destination 302 // vector type is legal. 303 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits() && 304 TLI.isTypeLegal(ValueVT)) 305 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val); 306 307 // Handle cases such as i8 -> <1 x i1> 308 if (ValueVT.getVectorNumElements() != 1) { 309 LLVMContext &Ctx = *DAG.getContext(); 310 Twine ErrMsg("non-trivial scalar-to-vector conversion"); 311 if (const Instruction *I = dyn_cast_or_null<Instruction>(V)) { 312 if (const CallInst *CI = dyn_cast<CallInst>(I)) 313 if (isa<InlineAsm>(CI->getCalledValue())) 314 ErrMsg = ErrMsg + ", possible invalid constraint for vector type"; 315 Ctx.emitError(I, ErrMsg); 316 } else { 317 Ctx.emitError(ErrMsg); 318 } 319 return DAG.getUNDEF(ValueVT); 320 } 321 322 if (ValueVT.getVectorNumElements() == 1 && 323 ValueVT.getVectorElementType() != PartEVT) { 324 bool Smaller = ValueVT.bitsLE(PartEVT); 325 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND), 326 DL, ValueVT.getScalarType(), Val); 327 } 328 329 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val); 330} 331 332static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc dl, 333 SDValue Val, SDValue *Parts, unsigned NumParts, 334 MVT PartVT, const Value *V); 335 336/// getCopyToParts - Create a series of nodes that contain the specified value 337/// split into legal parts. If the parts contain more bits than Val, then, for 338/// integers, ExtendKind can be used to specify how to generate the extra bits. 339static void getCopyToParts(SelectionDAG &DAG, SDLoc DL, 340 SDValue Val, SDValue *Parts, unsigned NumParts, 341 MVT PartVT, const Value *V, 342 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) { 343 EVT ValueVT = Val.getValueType(); 344 345 // Handle the vector case separately. 346 if (ValueVT.isVector()) 347 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT, V); 348 349 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 350 unsigned PartBits = PartVT.getSizeInBits(); 351 unsigned OrigNumParts = NumParts; 352 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!"); 353 354 if (NumParts == 0) 355 return; 356 357 assert(!ValueVT.isVector() && "Vector case handled elsewhere"); 358 EVT PartEVT = PartVT; 359 if (PartEVT == ValueVT) { 360 assert(NumParts == 1 && "No-op copy with multiple parts!"); 361 Parts[0] = Val; 362 return; 363 } 364 365 if (NumParts * PartBits > ValueVT.getSizeInBits()) { 366 // If the parts cover more bits than the value has, promote the value. 367 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) { 368 assert(NumParts == 1 && "Do not know what to promote to!"); 369 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val); 370 } else { 371 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) && 372 ValueVT.isInteger() && 373 "Unknown mismatch!"); 374 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); 375 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val); 376 if (PartVT == MVT::x86mmx) 377 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val); 378 } 379 } else if (PartBits == ValueVT.getSizeInBits()) { 380 // Different types of the same size. 381 assert(NumParts == 1 && PartEVT != ValueVT); 382 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val); 383 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) { 384 // If the parts cover less bits than value has, truncate the value. 385 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) && 386 ValueVT.isInteger() && 387 "Unknown mismatch!"); 388 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); 389 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); 390 if (PartVT == MVT::x86mmx) 391 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val); 392 } 393 394 // The value may have changed - recompute ValueVT. 395 ValueVT = Val.getValueType(); 396 assert(NumParts * PartBits == ValueVT.getSizeInBits() && 397 "Failed to tile the value with PartVT!"); 398 399 if (NumParts == 1) { 400 if (PartEVT != ValueVT) { 401 LLVMContext &Ctx = *DAG.getContext(); 402 Twine ErrMsg("scalar-to-vector conversion failed"); 403 if (const Instruction *I = dyn_cast_or_null<Instruction>(V)) { 404 if (const CallInst *CI = dyn_cast<CallInst>(I)) 405 if (isa<InlineAsm>(CI->getCalledValue())) 406 ErrMsg = ErrMsg + ", possible invalid constraint for vector type"; 407 Ctx.emitError(I, ErrMsg); 408 } else { 409 Ctx.emitError(ErrMsg); 410 } 411 } 412 413 Parts[0] = Val; 414 return; 415 } 416 417 // Expand the value into multiple parts. 418 if (NumParts & (NumParts - 1)) { 419 // The number of parts is not a power of 2. Split off and copy the tail. 420 assert(PartVT.isInteger() && ValueVT.isInteger() && 421 "Do not know what to expand to!"); 422 unsigned RoundParts = 1 << Log2_32(NumParts); 423 unsigned RoundBits = RoundParts * PartBits; 424 unsigned OddParts = NumParts - RoundParts; 425 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val, 426 DAG.getIntPtrConstant(RoundBits)); 427 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT, V); 428 429 if (TLI.isBigEndian()) 430 // The odd parts were reversed by getCopyToParts - unreverse them. 431 std::reverse(Parts + RoundParts, Parts + NumParts); 432 433 NumParts = RoundParts; 434 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits); 435 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val); 436 } 437 438 // The number of parts is a power of 2. Repeatedly bisect the value using 439 // EXTRACT_ELEMENT. 440 Parts[0] = DAG.getNode(ISD::BITCAST, DL, 441 EVT::getIntegerVT(*DAG.getContext(), 442 ValueVT.getSizeInBits()), 443 Val); 444 445 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) { 446 for (unsigned i = 0; i < NumParts; i += StepSize) { 447 unsigned ThisBits = StepSize * PartBits / 2; 448 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits); 449 SDValue &Part0 = Parts[i]; 450 SDValue &Part1 = Parts[i+StepSize/2]; 451 452 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, 453 ThisVT, Part0, DAG.getIntPtrConstant(1)); 454 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, 455 ThisVT, Part0, DAG.getIntPtrConstant(0)); 456 457 if (ThisBits == PartBits && ThisVT != PartVT) { 458 Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0); 459 Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1); 460 } 461 } 462 } 463 464 if (TLI.isBigEndian()) 465 std::reverse(Parts, Parts + OrigNumParts); 466} 467 468 469/// getCopyToPartsVector - Create a series of nodes that contain the specified 470/// value split into legal parts. 471static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc DL, 472 SDValue Val, SDValue *Parts, unsigned NumParts, 473 MVT PartVT, const Value *V) { 474 EVT ValueVT = Val.getValueType(); 475 assert(ValueVT.isVector() && "Not a vector"); 476 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 477 478 if (NumParts == 1) { 479 EVT PartEVT = PartVT; 480 if (PartEVT == ValueVT) { 481 // Nothing to do. 482 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) { 483 // Bitconvert vector->vector case. 484 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val); 485 } else if (PartVT.isVector() && 486 PartEVT.getVectorElementType() == ValueVT.getVectorElementType() && 487 PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements()) { 488 EVT ElementVT = PartVT.getVectorElementType(); 489 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in 490 // undef elements. 491 SmallVector<SDValue, 16> Ops; 492 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i) 493 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, 494 ElementVT, Val, DAG.getConstant(i, 495 TLI.getVectorIdxTy()))); 496 497 for (unsigned i = ValueVT.getVectorNumElements(), 498 e = PartVT.getVectorNumElements(); i != e; ++i) 499 Ops.push_back(DAG.getUNDEF(ElementVT)); 500 501 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size()); 502 503 // FIXME: Use CONCAT for 2x -> 4x. 504 505 //SDValue UndefElts = DAG.getUNDEF(VectorTy); 506 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts); 507 } else if (PartVT.isVector() && 508 PartEVT.getVectorElementType().bitsGE( 509 ValueVT.getVectorElementType()) && 510 PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements()) { 511 512 // Promoted vector extract 513 bool Smaller = PartEVT.bitsLE(ValueVT); 514 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND), 515 DL, PartVT, Val); 516 } else{ 517 // Vector -> scalar conversion. 518 assert(ValueVT.getVectorNumElements() == 1 && 519 "Only trivial vector-to-scalar conversions should get here!"); 520 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, 521 PartVT, Val, DAG.getConstant(0, TLI.getVectorIdxTy())); 522 523 bool Smaller = ValueVT.bitsLE(PartVT); 524 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND), 525 DL, PartVT, Val); 526 } 527 528 Parts[0] = Val; 529 return; 530 } 531 532 // Handle a multi-element vector. 533 EVT IntermediateVT; 534 MVT RegisterVT; 535 unsigned NumIntermediates; 536 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, 537 IntermediateVT, 538 NumIntermediates, RegisterVT); 539 unsigned NumElements = ValueVT.getVectorNumElements(); 540 541 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!"); 542 NumParts = NumRegs; // Silence a compiler warning. 543 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!"); 544 545 // Split the vector into intermediate operands. 546 SmallVector<SDValue, 8> Ops(NumIntermediates); 547 for (unsigned i = 0; i != NumIntermediates; ++i) { 548 if (IntermediateVT.isVector()) 549 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, 550 IntermediateVT, Val, 551 DAG.getConstant(i * (NumElements / NumIntermediates), 552 TLI.getVectorIdxTy())); 553 else 554 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, 555 IntermediateVT, Val, 556 DAG.getConstant(i, TLI.getVectorIdxTy())); 557 } 558 559 // Split the intermediate operands into legal parts. 560 if (NumParts == NumIntermediates) { 561 // If the register was not expanded, promote or copy the value, 562 // as appropriate. 563 for (unsigned i = 0; i != NumParts; ++i) 564 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT, V); 565 } else if (NumParts > 0) { 566 // If the intermediate type was expanded, split each the value into 567 // legal parts. 568 assert(NumParts % NumIntermediates == 0 && 569 "Must expand into a divisible number of parts!"); 570 unsigned Factor = NumParts / NumIntermediates; 571 for (unsigned i = 0; i != NumIntermediates; ++i) 572 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT, V); 573 } 574} 575 576namespace { 577 /// RegsForValue - This struct represents the registers (physical or virtual) 578 /// that a particular set of values is assigned, and the type information 579 /// about the value. The most common situation is to represent one value at a 580 /// time, but struct or array values are handled element-wise as multiple 581 /// values. The splitting of aggregates is performed recursively, so that we 582 /// never have aggregate-typed registers. The values at this point do not 583 /// necessarily have legal types, so each value may require one or more 584 /// registers of some legal type. 585 /// 586 struct RegsForValue { 587 /// ValueVTs - The value types of the values, which may not be legal, and 588 /// may need be promoted or synthesized from one or more registers. 589 /// 590 SmallVector<EVT, 4> ValueVTs; 591 592 /// RegVTs - The value types of the registers. This is the same size as 593 /// ValueVTs and it records, for each value, what the type of the assigned 594 /// register or registers are. (Individual values are never synthesized 595 /// from more than one type of register.) 596 /// 597 /// With virtual registers, the contents of RegVTs is redundant with TLI's 598 /// getRegisterType member function, however when with physical registers 599 /// it is necessary to have a separate record of the types. 600 /// 601 SmallVector<MVT, 4> RegVTs; 602 603 /// Regs - This list holds the registers assigned to the values. 604 /// Each legal or promoted value requires one register, and each 605 /// expanded value requires multiple registers. 606 /// 607 SmallVector<unsigned, 4> Regs; 608 609 RegsForValue() {} 610 611 RegsForValue(const SmallVector<unsigned, 4> ®s, 612 MVT regvt, EVT valuevt) 613 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {} 614 615 RegsForValue(LLVMContext &Context, const TargetLowering &tli, 616 unsigned Reg, Type *Ty) { 617 ComputeValueVTs(tli, Ty, ValueVTs); 618 619 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) { 620 EVT ValueVT = ValueVTs[Value]; 621 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT); 622 MVT RegisterVT = tli.getRegisterType(Context, ValueVT); 623 for (unsigned i = 0; i != NumRegs; ++i) 624 Regs.push_back(Reg + i); 625 RegVTs.push_back(RegisterVT); 626 Reg += NumRegs; 627 } 628 } 629 630 /// areValueTypesLegal - Return true if types of all the values are legal. 631 bool areValueTypesLegal(const TargetLowering &TLI) { 632 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) { 633 MVT RegisterVT = RegVTs[Value]; 634 if (!TLI.isTypeLegal(RegisterVT)) 635 return false; 636 } 637 return true; 638 } 639 640 /// append - Add the specified values to this one. 641 void append(const RegsForValue &RHS) { 642 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end()); 643 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end()); 644 Regs.append(RHS.Regs.begin(), RHS.Regs.end()); 645 } 646 647 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from 648 /// this value and returns the result as a ValueVTs value. This uses 649 /// Chain/Flag as the input and updates them for the output Chain/Flag. 650 /// If the Flag pointer is NULL, no flag is used. 651 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo, 652 SDLoc dl, 653 SDValue &Chain, SDValue *Flag, 654 const Value *V = 0) const; 655 656 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the 657 /// specified value into the registers specified by this object. This uses 658 /// Chain/Flag as the input and updates them for the output Chain/Flag. 659 /// If the Flag pointer is NULL, no flag is used. 660 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl, 661 SDValue &Chain, SDValue *Flag, const Value *V) const; 662 663 /// AddInlineAsmOperands - Add this value to the specified inlineasm node 664 /// operand list. This adds the code marker, matching input operand index 665 /// (if applicable), and includes the number of values added into it. 666 void AddInlineAsmOperands(unsigned Kind, 667 bool HasMatching, unsigned MatchingIdx, 668 SelectionDAG &DAG, 669 std::vector<SDValue> &Ops) const; 670 }; 671} 672 673/// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from 674/// this value and returns the result as a ValueVT value. This uses 675/// Chain/Flag as the input and updates them for the output Chain/Flag. 676/// If the Flag pointer is NULL, no flag is used. 677SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG, 678 FunctionLoweringInfo &FuncInfo, 679 SDLoc dl, 680 SDValue &Chain, SDValue *Flag, 681 const Value *V) const { 682 // A Value with type {} or [0 x %t] needs no registers. 683 if (ValueVTs.empty()) 684 return SDValue(); 685 686 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 687 688 // Assemble the legal parts into the final values. 689 SmallVector<SDValue, 4> Values(ValueVTs.size()); 690 SmallVector<SDValue, 8> Parts; 691 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) { 692 // Copy the legal parts from the registers. 693 EVT ValueVT = ValueVTs[Value]; 694 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT); 695 MVT RegisterVT = RegVTs[Value]; 696 697 Parts.resize(NumRegs); 698 for (unsigned i = 0; i != NumRegs; ++i) { 699 SDValue P; 700 if (Flag == 0) { 701 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT); 702 } else { 703 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag); 704 *Flag = P.getValue(2); 705 } 706 707 Chain = P.getValue(1); 708 Parts[i] = P; 709 710 // If the source register was virtual and if we know something about it, 711 // add an assert node. 712 if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) || 713 !RegisterVT.isInteger() || RegisterVT.isVector()) 714 continue; 715 716 const FunctionLoweringInfo::LiveOutInfo *LOI = 717 FuncInfo.GetLiveOutRegInfo(Regs[Part+i]); 718 if (!LOI) 719 continue; 720 721 unsigned RegSize = RegisterVT.getSizeInBits(); 722 unsigned NumSignBits = LOI->NumSignBits; 723 unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes(); 724 725 if (NumZeroBits == RegSize) { 726 // The current value is a zero. 727 // Explicitly express that as it would be easier for 728 // optimizations to kick in. 729 Parts[i] = DAG.getConstant(0, RegisterVT); 730 continue; 731 } 732 733 // FIXME: We capture more information than the dag can represent. For 734 // now, just use the tightest assertzext/assertsext possible. 735 bool isSExt = true; 736 EVT FromVT(MVT::Other); 737 if (NumSignBits == RegSize) 738 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1 739 else if (NumZeroBits >= RegSize-1) 740 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1 741 else if (NumSignBits > RegSize-8) 742 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8 743 else if (NumZeroBits >= RegSize-8) 744 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8 745 else if (NumSignBits > RegSize-16) 746 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16 747 else if (NumZeroBits >= RegSize-16) 748 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16 749 else if (NumSignBits > RegSize-32) 750 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32 751 else if (NumZeroBits >= RegSize-32) 752 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32 753 else 754 continue; 755 756 // Add an assertion node. 757 assert(FromVT != MVT::Other); 758 Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl, 759 RegisterVT, P, DAG.getValueType(FromVT)); 760 } 761 762 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(), 763 NumRegs, RegisterVT, ValueVT, V); 764 Part += NumRegs; 765 Parts.clear(); 766 } 767 768 return DAG.getNode(ISD::MERGE_VALUES, dl, 769 DAG.getVTList(&ValueVTs[0], ValueVTs.size()), 770 &Values[0], ValueVTs.size()); 771} 772 773/// getCopyToRegs - Emit a series of CopyToReg nodes that copies the 774/// specified value into the registers specified by this object. This uses 775/// Chain/Flag as the input and updates them for the output Chain/Flag. 776/// If the Flag pointer is NULL, no flag is used. 777void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl, 778 SDValue &Chain, SDValue *Flag, 779 const Value *V) const { 780 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 781 782 // Get the list of the values's legal parts. 783 unsigned NumRegs = Regs.size(); 784 SmallVector<SDValue, 8> Parts(NumRegs); 785 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) { 786 EVT ValueVT = ValueVTs[Value]; 787 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT); 788 MVT RegisterVT = RegVTs[Value]; 789 ISD::NodeType ExtendKind = 790 TLI.isZExtFree(Val, RegisterVT)? ISD::ZERO_EXTEND: ISD::ANY_EXTEND; 791 792 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value), 793 &Parts[Part], NumParts, RegisterVT, V, ExtendKind); 794 Part += NumParts; 795 } 796 797 // Copy the parts into the registers. 798 SmallVector<SDValue, 8> Chains(NumRegs); 799 for (unsigned i = 0; i != NumRegs; ++i) { 800 SDValue Part; 801 if (Flag == 0) { 802 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]); 803 } else { 804 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag); 805 *Flag = Part.getValue(1); 806 } 807 808 Chains[i] = Part.getValue(0); 809 } 810 811 if (NumRegs == 1 || Flag) 812 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is 813 // flagged to it. That is the CopyToReg nodes and the user are considered 814 // a single scheduling unit. If we create a TokenFactor and return it as 815 // chain, then the TokenFactor is both a predecessor (operand) of the 816 // user as well as a successor (the TF operands are flagged to the user). 817 // c1, f1 = CopyToReg 818 // c2, f2 = CopyToReg 819 // c3 = TokenFactor c1, c2 820 // ... 821 // = op c3, ..., f2 822 Chain = Chains[NumRegs-1]; 823 else 824 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs); 825} 826 827/// AddInlineAsmOperands - Add this value to the specified inlineasm node 828/// operand list. This adds the code marker and includes the number of 829/// values added into it. 830void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching, 831 unsigned MatchingIdx, 832 SelectionDAG &DAG, 833 std::vector<SDValue> &Ops) const { 834 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 835 836 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size()); 837 if (HasMatching) 838 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx); 839 else if (!Regs.empty() && 840 TargetRegisterInfo::isVirtualRegister(Regs.front())) { 841 // Put the register class of the virtual registers in the flag word. That 842 // way, later passes can recompute register class constraints for inline 843 // assembly as well as normal instructions. 844 // Don't do this for tied operands that can use the regclass information 845 // from the def. 846 const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo(); 847 const TargetRegisterClass *RC = MRI.getRegClass(Regs.front()); 848 Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID()); 849 } 850 851 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32); 852 Ops.push_back(Res); 853 854 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) { 855 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]); 856 MVT RegisterVT = RegVTs[Value]; 857 for (unsigned i = 0; i != NumRegs; ++i) { 858 assert(Reg < Regs.size() && "Mismatch in # registers expected"); 859 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT)); 860 } 861 } 862} 863 864void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa, 865 const TargetLibraryInfo *li) { 866 AA = &aa; 867 GFI = gfi; 868 LibInfo = li; 869 TD = DAG.getTarget().getDataLayout(); 870 Context = DAG.getContext(); 871 LPadToCallSiteMap.clear(); 872} 873 874/// clear - Clear out the current SelectionDAG and the associated 875/// state and prepare this SelectionDAGBuilder object to be used 876/// for a new block. This doesn't clear out information about 877/// additional blocks that are needed to complete switch lowering 878/// or PHI node updating; that information is cleared out as it is 879/// consumed. 880void SelectionDAGBuilder::clear() { 881 NodeMap.clear(); 882 UnusedArgNodeMap.clear(); 883 PendingLoads.clear(); 884 PendingExports.clear(); 885 CurInst = NULL; 886 HasTailCall = false; 887} 888 889/// clearDanglingDebugInfo - Clear the dangling debug information 890/// map. This function is separated from the clear so that debug 891/// information that is dangling in a basic block can be properly 892/// resolved in a different basic block. This allows the 893/// SelectionDAG to resolve dangling debug information attached 894/// to PHI nodes. 895void SelectionDAGBuilder::clearDanglingDebugInfo() { 896 DanglingDebugInfoMap.clear(); 897} 898 899/// getRoot - Return the current virtual root of the Selection DAG, 900/// flushing any PendingLoad items. This must be done before emitting 901/// a store or any other node that may need to be ordered after any 902/// prior load instructions. 903/// 904SDValue SelectionDAGBuilder::getRoot() { 905 if (PendingLoads.empty()) 906 return DAG.getRoot(); 907 908 if (PendingLoads.size() == 1) { 909 SDValue Root = PendingLoads[0]; 910 DAG.setRoot(Root); 911 PendingLoads.clear(); 912 return Root; 913 } 914 915 // Otherwise, we have to make a token factor node. 916 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, 917 &PendingLoads[0], PendingLoads.size()); 918 PendingLoads.clear(); 919 DAG.setRoot(Root); 920 return Root; 921} 922 923/// getControlRoot - Similar to getRoot, but instead of flushing all the 924/// PendingLoad items, flush all the PendingExports items. It is necessary 925/// to do this before emitting a terminator instruction. 926/// 927SDValue SelectionDAGBuilder::getControlRoot() { 928 SDValue Root = DAG.getRoot(); 929 930 if (PendingExports.empty()) 931 return Root; 932 933 // Turn all of the CopyToReg chains into one factored node. 934 if (Root.getOpcode() != ISD::EntryToken) { 935 unsigned i = 0, e = PendingExports.size(); 936 for (; i != e; ++i) { 937 assert(PendingExports[i].getNode()->getNumOperands() > 1); 938 if (PendingExports[i].getNode()->getOperand(0) == Root) 939 break; // Don't add the root if we already indirectly depend on it. 940 } 941 942 if (i == e) 943 PendingExports.push_back(Root); 944 } 945 946 Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, 947 &PendingExports[0], 948 PendingExports.size()); 949 PendingExports.clear(); 950 DAG.setRoot(Root); 951 return Root; 952} 953 954void SelectionDAGBuilder::visit(const Instruction &I) { 955 // Set up outgoing PHI node register values before emitting the terminator. 956 if (isa<TerminatorInst>(&I)) 957 HandlePHINodesInSuccessorBlocks(I.getParent()); 958 959 ++SDNodeOrder; 960 961 CurInst = &I; 962 963 visit(I.getOpcode(), I); 964 965 if (!isa<TerminatorInst>(&I) && !HasTailCall) 966 CopyToExportRegsIfNeeded(&I); 967 968 CurInst = NULL; 969} 970 971void SelectionDAGBuilder::visitPHI(const PHINode &) { 972 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!"); 973} 974 975void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) { 976 // Note: this doesn't use InstVisitor, because it has to work with 977 // ConstantExpr's in addition to instructions. 978 switch (Opcode) { 979 default: llvm_unreachable("Unknown instruction type encountered!"); 980 // Build the switch statement using the Instruction.def file. 981#define HANDLE_INST(NUM, OPCODE, CLASS) \ 982 case Instruction::OPCODE: visit##OPCODE((const CLASS&)I); break; 983#include "llvm/IR/Instruction.def" 984 } 985} 986 987// resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V, 988// generate the debug data structures now that we've seen its definition. 989void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V, 990 SDValue Val) { 991 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V]; 992 if (DDI.getDI()) { 993 const DbgValueInst *DI = DDI.getDI(); 994 DebugLoc dl = DDI.getdl(); 995 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder(); 996 MDNode *Variable = DI->getVariable(); 997 uint64_t Offset = DI->getOffset(); 998 SDDbgValue *SDV; 999 if (Val.getNode()) { 1000 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) { 1001 SDV = DAG.getDbgValue(Variable, Val.getNode(), 1002 Val.getResNo(), Offset, dl, DbgSDNodeOrder); 1003 DAG.AddDbgValue(SDV, Val.getNode(), false); 1004 } 1005 } else 1006 DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n"); 1007 DanglingDebugInfoMap[V] = DanglingDebugInfo(); 1008 } 1009} 1010 1011/// getValue - Return an SDValue for the given Value. 1012SDValue SelectionDAGBuilder::getValue(const Value *V) { 1013 // If we already have an SDValue for this value, use it. It's important 1014 // to do this first, so that we don't create a CopyFromReg if we already 1015 // have a regular SDValue. 1016 SDValue &N = NodeMap[V]; 1017 if (N.getNode()) return N; 1018 1019 // If there's a virtual register allocated and initialized for this 1020 // value, use it. 1021 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V); 1022 if (It != FuncInfo.ValueMap.end()) { 1023 unsigned InReg = It->second; 1024 RegsForValue RFV(*DAG.getContext(), *TM.getTargetLowering(), 1025 InReg, V->getType()); 1026 SDValue Chain = DAG.getEntryNode(); 1027 N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, NULL, V); 1028 resolveDanglingDebugInfo(V, N); 1029 return N; 1030 } 1031 1032 // Otherwise create a new SDValue and remember it. 1033 SDValue Val = getValueImpl(V); 1034 NodeMap[V] = Val; 1035 resolveDanglingDebugInfo(V, Val); 1036 return Val; 1037} 1038 1039/// getNonRegisterValue - Return an SDValue for the given Value, but 1040/// don't look in FuncInfo.ValueMap for a virtual register. 1041SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) { 1042 // If we already have an SDValue for this value, use it. 1043 SDValue &N = NodeMap[V]; 1044 if (N.getNode()) return N; 1045 1046 // Otherwise create a new SDValue and remember it. 1047 SDValue Val = getValueImpl(V); 1048 NodeMap[V] = Val; 1049 resolveDanglingDebugInfo(V, Val); 1050 return Val; 1051} 1052 1053/// getValueImpl - Helper function for getValue and getNonRegisterValue. 1054/// Create an SDValue for the given value. 1055SDValue SelectionDAGBuilder::getValueImpl(const Value *V) { 1056 const TargetLowering *TLI = TM.getTargetLowering(); 1057 1058 if (const Constant *C = dyn_cast<Constant>(V)) { 1059 EVT VT = TLI->getValueType(V->getType(), true); 1060 1061 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C)) 1062 return DAG.getConstant(*CI, VT); 1063 1064 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C)) 1065 return DAG.getGlobalAddress(GV, getCurSDLoc(), VT); 1066 1067 if (isa<ConstantPointerNull>(C)) 1068 return DAG.getConstant(0, TLI->getPointerTy()); 1069 1070 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C)) 1071 return DAG.getConstantFP(*CFP, VT); 1072 1073 if (isa<UndefValue>(C) && !V->getType()->isAggregateType()) 1074 return DAG.getUNDEF(VT); 1075 1076 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 1077 visit(CE->getOpcode(), *CE); 1078 SDValue N1 = NodeMap[V]; 1079 assert(N1.getNode() && "visit didn't populate the NodeMap!"); 1080 return N1; 1081 } 1082 1083 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) { 1084 SmallVector<SDValue, 4> Constants; 1085 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end(); 1086 OI != OE; ++OI) { 1087 SDNode *Val = getValue(*OI).getNode(); 1088 // If the operand is an empty aggregate, there are no values. 1089 if (!Val) continue; 1090 // Add each leaf value from the operand to the Constants list 1091 // to form a flattened list of all the values. 1092 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i) 1093 Constants.push_back(SDValue(Val, i)); 1094 } 1095 1096 return DAG.getMergeValues(&Constants[0], Constants.size(), 1097 getCurSDLoc()); 1098 } 1099 1100 if (const ConstantDataSequential *CDS = 1101 dyn_cast<ConstantDataSequential>(C)) { 1102 SmallVector<SDValue, 4> Ops; 1103 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) { 1104 SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode(); 1105 // Add each leaf value from the operand to the Constants list 1106 // to form a flattened list of all the values. 1107 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i) 1108 Ops.push_back(SDValue(Val, i)); 1109 } 1110 1111 if (isa<ArrayType>(CDS->getType())) 1112 return DAG.getMergeValues(&Ops[0], Ops.size(), getCurSDLoc()); 1113 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), 1114 VT, &Ops[0], Ops.size()); 1115 } 1116 1117 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) { 1118 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) && 1119 "Unknown struct or array constant!"); 1120 1121 SmallVector<EVT, 4> ValueVTs; 1122 ComputeValueVTs(*TLI, C->getType(), ValueVTs); 1123 unsigned NumElts = ValueVTs.size(); 1124 if (NumElts == 0) 1125 return SDValue(); // empty struct 1126 SmallVector<SDValue, 4> Constants(NumElts); 1127 for (unsigned i = 0; i != NumElts; ++i) { 1128 EVT EltVT = ValueVTs[i]; 1129 if (isa<UndefValue>(C)) 1130 Constants[i] = DAG.getUNDEF(EltVT); 1131 else if (EltVT.isFloatingPoint()) 1132 Constants[i] = DAG.getConstantFP(0, EltVT); 1133 else 1134 Constants[i] = DAG.getConstant(0, EltVT); 1135 } 1136 1137 return DAG.getMergeValues(&Constants[0], NumElts, 1138 getCurSDLoc()); 1139 } 1140 1141 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C)) 1142 return DAG.getBlockAddress(BA, VT); 1143 1144 VectorType *VecTy = cast<VectorType>(V->getType()); 1145 unsigned NumElements = VecTy->getNumElements(); 1146 1147 // Now that we know the number and type of the elements, get that number of 1148 // elements into the Ops array based on what kind of constant it is. 1149 SmallVector<SDValue, 16> Ops; 1150 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) { 1151 for (unsigned i = 0; i != NumElements; ++i) 1152 Ops.push_back(getValue(CV->getOperand(i))); 1153 } else { 1154 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!"); 1155 EVT EltVT = TLI->getValueType(VecTy->getElementType()); 1156 1157 SDValue Op; 1158 if (EltVT.isFloatingPoint()) 1159 Op = DAG.getConstantFP(0, EltVT); 1160 else 1161 Op = DAG.getConstant(0, EltVT); 1162 Ops.assign(NumElements, Op); 1163 } 1164 1165 // Create a BUILD_VECTOR node. 1166 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), 1167 VT, &Ops[0], Ops.size()); 1168 } 1169 1170 // If this is a static alloca, generate it as the frameindex instead of 1171 // computation. 1172 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) { 1173 DenseMap<const AllocaInst*, int>::iterator SI = 1174 FuncInfo.StaticAllocaMap.find(AI); 1175 if (SI != FuncInfo.StaticAllocaMap.end()) 1176 return DAG.getFrameIndex(SI->second, TLI->getPointerTy()); 1177 } 1178 1179 // If this is an instruction which fast-isel has deferred, select it now. 1180 if (const Instruction *Inst = dyn_cast<Instruction>(V)) { 1181 unsigned InReg = FuncInfo.InitializeRegForValue(Inst); 1182 RegsForValue RFV(*DAG.getContext(), *TLI, InReg, Inst->getType()); 1183 SDValue Chain = DAG.getEntryNode(); 1184 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, NULL, V); 1185 } 1186 1187 llvm_unreachable("Can't get register for value!"); 1188} 1189 1190void SelectionDAGBuilder::visitRet(const ReturnInst &I) { 1191 const TargetLowering *TLI = TM.getTargetLowering(); 1192 SDValue Chain = getControlRoot(); 1193 SmallVector<ISD::OutputArg, 8> Outs; 1194 SmallVector<SDValue, 8> OutVals; 1195 1196 if (!FuncInfo.CanLowerReturn) { 1197 unsigned DemoteReg = FuncInfo.DemoteRegister; 1198 const Function *F = I.getParent()->getParent(); 1199 1200 // Emit a store of the return value through the virtual register. 1201 // Leave Outs empty so that LowerReturn won't try to load return 1202 // registers the usual way. 1203 SmallVector<EVT, 1> PtrValueVTs; 1204 ComputeValueVTs(*TLI, PointerType::getUnqual(F->getReturnType()), 1205 PtrValueVTs); 1206 1207 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]); 1208 SDValue RetOp = getValue(I.getOperand(0)); 1209 1210 SmallVector<EVT, 4> ValueVTs; 1211 SmallVector<uint64_t, 4> Offsets; 1212 ComputeValueVTs(*TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets); 1213 unsigned NumValues = ValueVTs.size(); 1214 1215 SmallVector<SDValue, 4> Chains(NumValues); 1216 for (unsigned i = 0; i != NumValues; ++i) { 1217 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(), 1218 RetPtr.getValueType(), RetPtr, 1219 DAG.getIntPtrConstant(Offsets[i])); 1220 Chains[i] = 1221 DAG.getStore(Chain, getCurSDLoc(), 1222 SDValue(RetOp.getNode(), RetOp.getResNo() + i), 1223 // FIXME: better loc info would be nice. 1224 Add, MachinePointerInfo(), false, false, 0); 1225 } 1226 1227 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), 1228 MVT::Other, &Chains[0], NumValues); 1229 } else if (I.getNumOperands() != 0) { 1230 SmallVector<EVT, 4> ValueVTs; 1231 ComputeValueVTs(*TLI, I.getOperand(0)->getType(), ValueVTs); 1232 unsigned NumValues = ValueVTs.size(); 1233 if (NumValues) { 1234 SDValue RetOp = getValue(I.getOperand(0)); 1235 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1236 EVT VT = ValueVTs[j]; 1237 1238 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1239 1240 const Function *F = I.getParent()->getParent(); 1241 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex, 1242 Attribute::SExt)) 1243 ExtendKind = ISD::SIGN_EXTEND; 1244 else if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex, 1245 Attribute::ZExt)) 1246 ExtendKind = ISD::ZERO_EXTEND; 1247 1248 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) 1249 VT = TLI->getTypeForExtArgOrReturn(VT.getSimpleVT(), ExtendKind); 1250 1251 unsigned NumParts = TLI->getNumRegisters(*DAG.getContext(), VT); 1252 MVT PartVT = TLI->getRegisterType(*DAG.getContext(), VT); 1253 SmallVector<SDValue, 4> Parts(NumParts); 1254 getCopyToParts(DAG, getCurSDLoc(), 1255 SDValue(RetOp.getNode(), RetOp.getResNo() + j), 1256 &Parts[0], NumParts, PartVT, &I, ExtendKind); 1257 1258 // 'inreg' on function refers to return value 1259 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1260 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex, 1261 Attribute::InReg)) 1262 Flags.setInReg(); 1263 1264 // Propagate extension type if any 1265 if (ExtendKind == ISD::SIGN_EXTEND) 1266 Flags.setSExt(); 1267 else if (ExtendKind == ISD::ZERO_EXTEND) 1268 Flags.setZExt(); 1269 1270 for (unsigned i = 0; i < NumParts; ++i) { 1271 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(), 1272 VT, /*isfixed=*/true, 0, 0)); 1273 OutVals.push_back(Parts[i]); 1274 } 1275 } 1276 } 1277 } 1278 1279 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg(); 1280 CallingConv::ID CallConv = 1281 DAG.getMachineFunction().getFunction()->getCallingConv(); 1282 Chain = TM.getTargetLowering()->LowerReturn(Chain, CallConv, isVarArg, 1283 Outs, OutVals, getCurSDLoc(), 1284 DAG); 1285 1286 // Verify that the target's LowerReturn behaved as expected. 1287 assert(Chain.getNode() && Chain.getValueType() == MVT::Other && 1288 "LowerReturn didn't return a valid chain!"); 1289 1290 // Update the DAG with the new chain value resulting from return lowering. 1291 DAG.setRoot(Chain); 1292} 1293 1294/// CopyToExportRegsIfNeeded - If the given value has virtual registers 1295/// created for it, emit nodes to copy the value into the virtual 1296/// registers. 1297void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) { 1298 // Skip empty types 1299 if (V->getType()->isEmptyTy()) 1300 return; 1301 1302 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V); 1303 if (VMI != FuncInfo.ValueMap.end()) { 1304 assert(!V->use_empty() && "Unused value assigned virtual registers!"); 1305 CopyValueToVirtualRegister(V, VMI->second); 1306 } 1307} 1308 1309/// ExportFromCurrentBlock - If this condition isn't known to be exported from 1310/// the current basic block, add it to ValueMap now so that we'll get a 1311/// CopyTo/FromReg. 1312void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) { 1313 // No need to export constants. 1314 if (!isa<Instruction>(V) && !isa<Argument>(V)) return; 1315 1316 // Already exported? 1317 if (FuncInfo.isExportedInst(V)) return; 1318 1319 unsigned Reg = FuncInfo.InitializeRegForValue(V); 1320 CopyValueToVirtualRegister(V, Reg); 1321} 1322 1323bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V, 1324 const BasicBlock *FromBB) { 1325 // The operands of the setcc have to be in this block. We don't know 1326 // how to export them from some other block. 1327 if (const Instruction *VI = dyn_cast<Instruction>(V)) { 1328 // Can export from current BB. 1329 if (VI->getParent() == FromBB) 1330 return true; 1331 1332 // Is already exported, noop. 1333 return FuncInfo.isExportedInst(V); 1334 } 1335 1336 // If this is an argument, we can export it if the BB is the entry block or 1337 // if it is already exported. 1338 if (isa<Argument>(V)) { 1339 if (FromBB == &FromBB->getParent()->getEntryBlock()) 1340 return true; 1341 1342 // Otherwise, can only export this if it is already exported. 1343 return FuncInfo.isExportedInst(V); 1344 } 1345 1346 // Otherwise, constants can always be exported. 1347 return true; 1348} 1349 1350/// Return branch probability calculated by BranchProbabilityInfo for IR blocks. 1351uint32_t SelectionDAGBuilder::getEdgeWeight(const MachineBasicBlock *Src, 1352 const MachineBasicBlock *Dst) const { 1353 BranchProbabilityInfo *BPI = FuncInfo.BPI; 1354 if (!BPI) 1355 return 0; 1356 const BasicBlock *SrcBB = Src->getBasicBlock(); 1357 const BasicBlock *DstBB = Dst->getBasicBlock(); 1358 return BPI->getEdgeWeight(SrcBB, DstBB); 1359} 1360 1361void SelectionDAGBuilder:: 1362addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst, 1363 uint32_t Weight /* = 0 */) { 1364 if (!Weight) 1365 Weight = getEdgeWeight(Src, Dst); 1366 Src->addSuccessor(Dst, Weight); 1367} 1368 1369 1370static bool InBlock(const Value *V, const BasicBlock *BB) { 1371 if (const Instruction *I = dyn_cast<Instruction>(V)) 1372 return I->getParent() == BB; 1373 return true; 1374} 1375 1376/// EmitBranchForMergedCondition - Helper method for FindMergedConditions. 1377/// This function emits a branch and is used at the leaves of an OR or an 1378/// AND operator tree. 1379/// 1380void 1381SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond, 1382 MachineBasicBlock *TBB, 1383 MachineBasicBlock *FBB, 1384 MachineBasicBlock *CurBB, 1385 MachineBasicBlock *SwitchBB) { 1386 const BasicBlock *BB = CurBB->getBasicBlock(); 1387 1388 // If the leaf of the tree is a comparison, merge the condition into 1389 // the caseblock. 1390 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) { 1391 // The operands of the cmp have to be in this block. We don't know 1392 // how to export them from some other block. If this is the first block 1393 // of the sequence, no exporting is needed. 1394 if (CurBB == SwitchBB || 1395 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) && 1396 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) { 1397 ISD::CondCode Condition; 1398 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) { 1399 Condition = getICmpCondCode(IC->getPredicate()); 1400 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) { 1401 Condition = getFCmpCondCode(FC->getPredicate()); 1402 if (TM.Options.NoNaNsFPMath) 1403 Condition = getFCmpCodeWithoutNaN(Condition); 1404 } else { 1405 Condition = ISD::SETEQ; // silence warning. 1406 llvm_unreachable("Unknown compare instruction"); 1407 } 1408 1409 CaseBlock CB(Condition, BOp->getOperand(0), 1410 BOp->getOperand(1), NULL, TBB, FBB, CurBB); 1411 SwitchCases.push_back(CB); 1412 return; 1413 } 1414 } 1415 1416 // Create a CaseBlock record representing this branch. 1417 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()), 1418 NULL, TBB, FBB, CurBB); 1419 SwitchCases.push_back(CB); 1420} 1421 1422/// FindMergedConditions - If Cond is an expression like 1423void SelectionDAGBuilder::FindMergedConditions(const Value *Cond, 1424 MachineBasicBlock *TBB, 1425 MachineBasicBlock *FBB, 1426 MachineBasicBlock *CurBB, 1427 MachineBasicBlock *SwitchBB, 1428 unsigned Opc) { 1429 // If this node is not part of the or/and tree, emit it as a branch. 1430 const Instruction *BOp = dyn_cast<Instruction>(Cond); 1431 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) || 1432 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() || 1433 BOp->getParent() != CurBB->getBasicBlock() || 1434 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) || 1435 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) { 1436 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB); 1437 return; 1438 } 1439 1440 // Create TmpBB after CurBB. 1441 MachineFunction::iterator BBI = CurBB; 1442 MachineFunction &MF = DAG.getMachineFunction(); 1443 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock()); 1444 CurBB->getParent()->insert(++BBI, TmpBB); 1445 1446 if (Opc == Instruction::Or) { 1447 // Codegen X | Y as: 1448 // jmp_if_X TBB 1449 // jmp TmpBB 1450 // TmpBB: 1451 // jmp_if_Y TBB 1452 // jmp FBB 1453 // 1454 1455 // Emit the LHS condition. 1456 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc); 1457 1458 // Emit the RHS condition into TmpBB. 1459 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc); 1460 } else { 1461 assert(Opc == Instruction::And && "Unknown merge op!"); 1462 // Codegen X & Y as: 1463 // jmp_if_X TmpBB 1464 // jmp FBB 1465 // TmpBB: 1466 // jmp_if_Y TBB 1467 // jmp FBB 1468 // 1469 // This requires creation of TmpBB after CurBB. 1470 1471 // Emit the LHS condition. 1472 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc); 1473 1474 // Emit the RHS condition into TmpBB. 1475 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc); 1476 } 1477} 1478 1479/// If the set of cases should be emitted as a series of branches, return true. 1480/// If we should emit this as a bunch of and/or'd together conditions, return 1481/// false. 1482bool 1483SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases) { 1484 if (Cases.size() != 2) return true; 1485 1486 // If this is two comparisons of the same values or'd or and'd together, they 1487 // will get folded into a single comparison, so don't emit two blocks. 1488 if ((Cases[0].CmpLHS == Cases[1].CmpLHS && 1489 Cases[0].CmpRHS == Cases[1].CmpRHS) || 1490 (Cases[0].CmpRHS == Cases[1].CmpLHS && 1491 Cases[0].CmpLHS == Cases[1].CmpRHS)) { 1492 return false; 1493 } 1494 1495 // Handle: (X != null) | (Y != null) --> (X|Y) != 0 1496 // Handle: (X == null) & (Y == null) --> (X|Y) == 0 1497 if (Cases[0].CmpRHS == Cases[1].CmpRHS && 1498 Cases[0].CC == Cases[1].CC && 1499 isa<Constant>(Cases[0].CmpRHS) && 1500 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) { 1501 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB) 1502 return false; 1503 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB) 1504 return false; 1505 } 1506 1507 return true; 1508} 1509 1510void SelectionDAGBuilder::visitBr(const BranchInst &I) { 1511 MachineBasicBlock *BrMBB = FuncInfo.MBB; 1512 1513 // Update machine-CFG edges. 1514 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)]; 1515 1516 // Figure out which block is immediately after the current one. 1517 MachineBasicBlock *NextBlock = 0; 1518 MachineFunction::iterator BBI = BrMBB; 1519 if (++BBI != FuncInfo.MF->end()) 1520 NextBlock = BBI; 1521 1522 if (I.isUnconditional()) { 1523 // Update machine-CFG edges. 1524 BrMBB->addSuccessor(Succ0MBB); 1525 1526 // If this is not a fall-through branch, emit the branch. 1527 if (Succ0MBB != NextBlock) 1528 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), 1529 MVT::Other, getControlRoot(), 1530 DAG.getBasicBlock(Succ0MBB))); 1531 1532 return; 1533 } 1534 1535 // If this condition is one of the special cases we handle, do special stuff 1536 // now. 1537 const Value *CondVal = I.getCondition(); 1538 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)]; 1539 1540 // If this is a series of conditions that are or'd or and'd together, emit 1541 // this as a sequence of branches instead of setcc's with and/or operations. 1542 // As long as jumps are not expensive, this should improve performance. 1543 // For example, instead of something like: 1544 // cmp A, B 1545 // C = seteq 1546 // cmp D, E 1547 // F = setle 1548 // or C, F 1549 // jnz foo 1550 // Emit: 1551 // cmp A, B 1552 // je foo 1553 // cmp D, E 1554 // jle foo 1555 // 1556 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) { 1557 if (!TM.getTargetLowering()->isJumpExpensive() && 1558 BOp->hasOneUse() && 1559 (BOp->getOpcode() == Instruction::And || 1560 BOp->getOpcode() == Instruction::Or)) { 1561 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB, 1562 BOp->getOpcode()); 1563 // If the compares in later blocks need to use values not currently 1564 // exported from this block, export them now. This block should always 1565 // be the first entry. 1566 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!"); 1567 1568 // Allow some cases to be rejected. 1569 if (ShouldEmitAsBranches(SwitchCases)) { 1570 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) { 1571 ExportFromCurrentBlock(SwitchCases[i].CmpLHS); 1572 ExportFromCurrentBlock(SwitchCases[i].CmpRHS); 1573 } 1574 1575 // Emit the branch for this block. 1576 visitSwitchCase(SwitchCases[0], BrMBB); 1577 SwitchCases.erase(SwitchCases.begin()); 1578 return; 1579 } 1580 1581 // Okay, we decided not to do this, remove any inserted MBB's and clear 1582 // SwitchCases. 1583 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) 1584 FuncInfo.MF->erase(SwitchCases[i].ThisBB); 1585 1586 SwitchCases.clear(); 1587 } 1588 } 1589 1590 // Create a CaseBlock record representing this branch. 1591 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()), 1592 NULL, Succ0MBB, Succ1MBB, BrMBB); 1593 1594 // Use visitSwitchCase to actually insert the fast branch sequence for this 1595 // cond branch. 1596 visitSwitchCase(CB, BrMBB); 1597} 1598 1599/// visitSwitchCase - Emits the necessary code to represent a single node in 1600/// the binary search tree resulting from lowering a switch instruction. 1601void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB, 1602 MachineBasicBlock *SwitchBB) { 1603 SDValue Cond; 1604 SDValue CondLHS = getValue(CB.CmpLHS); 1605 SDLoc dl = getCurSDLoc(); 1606 1607 // Build the setcc now. 1608 if (CB.CmpMHS == NULL) { 1609 // Fold "(X == true)" to X and "(X == false)" to !X to 1610 // handle common cases produced by branch lowering. 1611 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) && 1612 CB.CC == ISD::SETEQ) 1613 Cond = CondLHS; 1614 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) && 1615 CB.CC == ISD::SETEQ) { 1616 SDValue True = DAG.getConstant(1, CondLHS.getValueType()); 1617 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True); 1618 } else 1619 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC); 1620 } else { 1621 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now"); 1622 1623 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue(); 1624 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue(); 1625 1626 SDValue CmpOp = getValue(CB.CmpMHS); 1627 EVT VT = CmpOp.getValueType(); 1628 1629 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) { 1630 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT), 1631 ISD::SETLE); 1632 } else { 1633 SDValue SUB = DAG.getNode(ISD::SUB, dl, 1634 VT, CmpOp, DAG.getConstant(Low, VT)); 1635 Cond = DAG.getSetCC(dl, MVT::i1, SUB, 1636 DAG.getConstant(High-Low, VT), ISD::SETULE); 1637 } 1638 } 1639 1640 // Update successor info 1641 addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight); 1642 // TrueBB and FalseBB are always different unless the incoming IR is 1643 // degenerate. This only happens when running llc on weird IR. 1644 if (CB.TrueBB != CB.FalseBB) 1645 addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight); 1646 1647 // Set NextBlock to be the MBB immediately after the current one, if any. 1648 // This is used to avoid emitting unnecessary branches to the next block. 1649 MachineBasicBlock *NextBlock = 0; 1650 MachineFunction::iterator BBI = SwitchBB; 1651 if (++BBI != FuncInfo.MF->end()) 1652 NextBlock = BBI; 1653 1654 // If the lhs block is the next block, invert the condition so that we can 1655 // fall through to the lhs instead of the rhs block. 1656 if (CB.TrueBB == NextBlock) { 1657 std::swap(CB.TrueBB, CB.FalseBB); 1658 SDValue True = DAG.getConstant(1, Cond.getValueType()); 1659 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True); 1660 } 1661 1662 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl, 1663 MVT::Other, getControlRoot(), Cond, 1664 DAG.getBasicBlock(CB.TrueBB)); 1665 1666 // Insert the false branch. Do this even if it's a fall through branch, 1667 // this makes it easier to do DAG optimizations which require inverting 1668 // the branch condition. 1669 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond, 1670 DAG.getBasicBlock(CB.FalseBB)); 1671 1672 DAG.setRoot(BrCond); 1673} 1674 1675/// visitJumpTable - Emit JumpTable node in the current MBB 1676void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) { 1677 // Emit the code for the jump table 1678 assert(JT.Reg != -1U && "Should lower JT Header first!"); 1679 EVT PTy = TM.getTargetLowering()->getPointerTy(); 1680 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(), 1681 JT.Reg, PTy); 1682 SDValue Table = DAG.getJumpTable(JT.JTI, PTy); 1683 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurSDLoc(), 1684 MVT::Other, Index.getValue(1), 1685 Table, Index); 1686 DAG.setRoot(BrJumpTable); 1687} 1688 1689/// visitJumpTableHeader - This function emits necessary code to produce index 1690/// in the JumpTable from switch case. 1691void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT, 1692 JumpTableHeader &JTH, 1693 MachineBasicBlock *SwitchBB) { 1694 // Subtract the lowest switch case value from the value being switched on and 1695 // conditional branch to default mbb if the result is greater than the 1696 // difference between smallest and largest cases. 1697 SDValue SwitchOp = getValue(JTH.SValue); 1698 EVT VT = SwitchOp.getValueType(); 1699 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp, 1700 DAG.getConstant(JTH.First, VT)); 1701 1702 // The SDNode we just created, which holds the value being switched on minus 1703 // the smallest case value, needs to be copied to a virtual register so it 1704 // can be used as an index into the jump table in a subsequent basic block. 1705 // This value may be smaller or larger than the target's pointer type, and 1706 // therefore require extension or truncating. 1707 const TargetLowering *TLI = TM.getTargetLowering(); 1708 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), TLI->getPointerTy()); 1709 1710 unsigned JumpTableReg = FuncInfo.CreateReg(TLI->getPointerTy()); 1711 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(), 1712 JumpTableReg, SwitchOp); 1713 JT.Reg = JumpTableReg; 1714 1715 // Emit the range check for the jump table, and branch to the default block 1716 // for the switch statement if the value being switched on exceeds the largest 1717 // case in the switch. 1718 SDValue CMP = DAG.getSetCC(getCurSDLoc(), 1719 TLI->getSetCCResultType(*DAG.getContext(), 1720 Sub.getValueType()), 1721 Sub, 1722 DAG.getConstant(JTH.Last - JTH.First,VT), 1723 ISD::SETUGT); 1724 1725 // Set NextBlock to be the MBB immediately after the current one, if any. 1726 // This is used to avoid emitting unnecessary branches to the next block. 1727 MachineBasicBlock *NextBlock = 0; 1728 MachineFunction::iterator BBI = SwitchBB; 1729 1730 if (++BBI != FuncInfo.MF->end()) 1731 NextBlock = BBI; 1732 1733 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(), 1734 MVT::Other, CopyTo, CMP, 1735 DAG.getBasicBlock(JT.Default)); 1736 1737 if (JT.MBB != NextBlock) 1738 BrCond = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrCond, 1739 DAG.getBasicBlock(JT.MBB)); 1740 1741 DAG.setRoot(BrCond); 1742} 1743 1744/// Codegen a new tail for a stack protector check ParentMBB which has had its 1745/// tail spliced into a stack protector check success bb. 1746/// 1747/// For a high level explanation of how this fits into the stack protector 1748/// generation see the comment on the declaration of class 1749/// StackProtectorDescriptor. 1750void SelectionDAGBuilder::visitSPDescriptorParent(StackProtectorDescriptor &SPD, 1751 MachineBasicBlock *ParentBB) { 1752 1753 // First create the loads to the guard/stack slot for the comparison. 1754 const TargetLowering *TLI = TM.getTargetLowering(); 1755 EVT PtrTy = TLI->getPointerTy(); 1756 1757 MachineFrameInfo *MFI = ParentBB->getParent()->getFrameInfo(); 1758 int FI = MFI->getStackProtectorIndex(); 1759 1760 const Value *IRGuard = SPD.getGuard(); 1761 SDValue GuardPtr = getValue(IRGuard); 1762 SDValue StackSlotPtr = DAG.getFrameIndex(FI, PtrTy); 1763 1764 unsigned Align = 1765 TLI->getDataLayout()->getPrefTypeAlignment(IRGuard->getType()); 1766 SDValue Guard = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(), 1767 GuardPtr, MachinePointerInfo(IRGuard, 0), 1768 true, false, false, Align); 1769 1770 SDValue StackSlot = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(), 1771 StackSlotPtr, 1772 MachinePointerInfo::getFixedStack(FI), 1773 true, false, false, Align); 1774 1775 // Perform the comparison via a subtract/getsetcc. 1776 EVT VT = Guard.getValueType(); 1777 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, Guard, StackSlot); 1778 1779 SDValue Cmp = DAG.getSetCC(getCurSDLoc(), 1780 TLI->getSetCCResultType(*DAG.getContext(), 1781 Sub.getValueType()), 1782 Sub, DAG.getConstant(0, VT), 1783 ISD::SETNE); 1784 1785 // If the sub is not 0, then we know the guard/stackslot do not equal, so 1786 // branch to failure MBB. 1787 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(), 1788 MVT::Other, StackSlot.getOperand(0), 1789 Cmp, DAG.getBasicBlock(SPD.getFailureMBB())); 1790 // Otherwise branch to success MBB. 1791 SDValue Br = DAG.getNode(ISD::BR, getCurSDLoc(), 1792 MVT::Other, BrCond, 1793 DAG.getBasicBlock(SPD.getSuccessMBB())); 1794 1795 DAG.setRoot(Br); 1796} 1797 1798/// Codegen the failure basic block for a stack protector check. 1799/// 1800/// A failure stack protector machine basic block consists simply of a call to 1801/// __stack_chk_fail(). 1802/// 1803/// For a high level explanation of how this fits into the stack protector 1804/// generation see the comment on the declaration of class 1805/// StackProtectorDescriptor. 1806void 1807SelectionDAGBuilder::visitSPDescriptorFailure(StackProtectorDescriptor &SPD) { 1808 const TargetLowering *TLI = TM.getTargetLowering(); 1809 SDValue Chain = TLI->makeLibCall(DAG, RTLIB::STACKPROTECTOR_CHECK_FAIL, 1810 MVT::isVoid, 0, 0, false, getCurSDLoc(), 1811 false, false).second; 1812 DAG.setRoot(Chain); 1813} 1814 1815/// visitBitTestHeader - This function emits necessary code to produce value 1816/// suitable for "bit tests" 1817void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B, 1818 MachineBasicBlock *SwitchBB) { 1819 // Subtract the minimum value 1820 SDValue SwitchOp = getValue(B.SValue); 1821 EVT VT = SwitchOp.getValueType(); 1822 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp, 1823 DAG.getConstant(B.First, VT)); 1824 1825 // Check range 1826 const TargetLowering *TLI = TM.getTargetLowering(); 1827 SDValue RangeCmp = DAG.getSetCC(getCurSDLoc(), 1828 TLI->getSetCCResultType(*DAG.getContext(), 1829 Sub.getValueType()), 1830 Sub, DAG.getConstant(B.Range, VT), 1831 ISD::SETUGT); 1832 1833 // Determine the type of the test operands. 1834 bool UsePtrType = false; 1835 if (!TLI->isTypeLegal(VT)) 1836 UsePtrType = true; 1837 else { 1838 for (unsigned i = 0, e = B.Cases.size(); i != e; ++i) 1839 if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) { 1840 // Switch table case range are encoded into series of masks. 1841 // Just use pointer type, it's guaranteed to fit. 1842 UsePtrType = true; 1843 break; 1844 } 1845 } 1846 if (UsePtrType) { 1847 VT = TLI->getPointerTy(); 1848 Sub = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), VT); 1849 } 1850 1851 B.RegVT = VT.getSimpleVT(); 1852 B.Reg = FuncInfo.CreateReg(B.RegVT); 1853 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(), 1854 B.Reg, Sub); 1855 1856 // Set NextBlock to be the MBB immediately after the current one, if any. 1857 // This is used to avoid emitting unnecessary branches to the next block. 1858 MachineBasicBlock *NextBlock = 0; 1859 MachineFunction::iterator BBI = SwitchBB; 1860 if (++BBI != FuncInfo.MF->end()) 1861 NextBlock = BBI; 1862 1863 MachineBasicBlock* MBB = B.Cases[0].ThisBB; 1864 1865 addSuccessorWithWeight(SwitchBB, B.Default); 1866 addSuccessorWithWeight(SwitchBB, MBB); 1867 1868 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurSDLoc(), 1869 MVT::Other, CopyTo, RangeCmp, 1870 DAG.getBasicBlock(B.Default)); 1871 1872 if (MBB != NextBlock) 1873 BrRange = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, CopyTo, 1874 DAG.getBasicBlock(MBB)); 1875 1876 DAG.setRoot(BrRange); 1877} 1878 1879/// visitBitTestCase - this function produces one "bit test" 1880void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB, 1881 MachineBasicBlock* NextMBB, 1882 uint32_t BranchWeightToNext, 1883 unsigned Reg, 1884 BitTestCase &B, 1885 MachineBasicBlock *SwitchBB) { 1886 MVT VT = BB.RegVT; 1887 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(), 1888 Reg, VT); 1889 SDValue Cmp; 1890 unsigned PopCount = CountPopulation_64(B.Mask); 1891 const TargetLowering *TLI = TM.getTargetLowering(); 1892 if (PopCount == 1) { 1893 // Testing for a single bit; just compare the shift count with what it 1894 // would need to be to shift a 1 bit in that position. 1895 Cmp = DAG.getSetCC(getCurSDLoc(), 1896 TLI->getSetCCResultType(*DAG.getContext(), VT), 1897 ShiftOp, 1898 DAG.getConstant(countTrailingZeros(B.Mask), VT), 1899 ISD::SETEQ); 1900 } else if (PopCount == BB.Range) { 1901 // There is only one zero bit in the range, test for it directly. 1902 Cmp = DAG.getSetCC(getCurSDLoc(), 1903 TLI->getSetCCResultType(*DAG.getContext(), VT), 1904 ShiftOp, 1905 DAG.getConstant(CountTrailingOnes_64(B.Mask), VT), 1906 ISD::SETNE); 1907 } else { 1908 // Make desired shift 1909 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurSDLoc(), VT, 1910 DAG.getConstant(1, VT), ShiftOp); 1911 1912 // Emit bit tests and jumps 1913 SDValue AndOp = DAG.getNode(ISD::AND, getCurSDLoc(), 1914 VT, SwitchVal, DAG.getConstant(B.Mask, VT)); 1915 Cmp = DAG.getSetCC(getCurSDLoc(), 1916 TLI->getSetCCResultType(*DAG.getContext(), VT), 1917 AndOp, DAG.getConstant(0, VT), 1918 ISD::SETNE); 1919 } 1920 1921 // The branch weight from SwitchBB to B.TargetBB is B.ExtraWeight. 1922 addSuccessorWithWeight(SwitchBB, B.TargetBB, B.ExtraWeight); 1923 // The branch weight from SwitchBB to NextMBB is BranchWeightToNext. 1924 addSuccessorWithWeight(SwitchBB, NextMBB, BranchWeightToNext); 1925 1926 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurSDLoc(), 1927 MVT::Other, getControlRoot(), 1928 Cmp, DAG.getBasicBlock(B.TargetBB)); 1929 1930 // Set NextBlock to be the MBB immediately after the current one, if any. 1931 // This is used to avoid emitting unnecessary branches to the next block. 1932 MachineBasicBlock *NextBlock = 0; 1933 MachineFunction::iterator BBI = SwitchBB; 1934 if (++BBI != FuncInfo.MF->end()) 1935 NextBlock = BBI; 1936 1937 if (NextMBB != NextBlock) 1938 BrAnd = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrAnd, 1939 DAG.getBasicBlock(NextMBB)); 1940 1941 DAG.setRoot(BrAnd); 1942} 1943 1944void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) { 1945 MachineBasicBlock *InvokeMBB = FuncInfo.MBB; 1946 1947 // Retrieve successors. 1948 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)]; 1949 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)]; 1950 1951 const Value *Callee(I.getCalledValue()); 1952 const Function *Fn = dyn_cast<Function>(Callee); 1953 if (isa<InlineAsm>(Callee)) 1954 visitInlineAsm(&I); 1955 else if (Fn && Fn->isIntrinsic()) { 1956 assert(Fn->getIntrinsicID() == Intrinsic::donothing); 1957 // Ignore invokes to @llvm.donothing: jump directly to the next BB. 1958 } else 1959 LowerCallTo(&I, getValue(Callee), false, LandingPad); 1960 1961 // If the value of the invoke is used outside of its defining block, make it 1962 // available as a virtual register. 1963 CopyToExportRegsIfNeeded(&I); 1964 1965 // Update successor info 1966 addSuccessorWithWeight(InvokeMBB, Return); 1967 addSuccessorWithWeight(InvokeMBB, LandingPad); 1968 1969 // Drop into normal successor. 1970 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), 1971 MVT::Other, getControlRoot(), 1972 DAG.getBasicBlock(Return))); 1973} 1974 1975void SelectionDAGBuilder::visitResume(const ResumeInst &RI) { 1976 llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!"); 1977} 1978 1979void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) { 1980 assert(FuncInfo.MBB->isLandingPad() && 1981 "Call to landingpad not in landing pad!"); 1982 1983 MachineBasicBlock *MBB = FuncInfo.MBB; 1984 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 1985 AddLandingPadInfo(LP, MMI, MBB); 1986 1987 // If there aren't registers to copy the values into (e.g., during SjLj 1988 // exceptions), then don't bother to create these DAG nodes. 1989 const TargetLowering *TLI = TM.getTargetLowering(); 1990 if (TLI->getExceptionPointerRegister() == 0 && 1991 TLI->getExceptionSelectorRegister() == 0) 1992 return; 1993 1994 SmallVector<EVT, 2> ValueVTs; 1995 ComputeValueVTs(*TLI, LP.getType(), ValueVTs); 1996 assert(ValueVTs.size() == 2 && "Only two-valued landingpads are supported"); 1997 1998 // Get the two live-in registers as SDValues. The physregs have already been 1999 // copied into virtual registers. 2000 SDValue Ops[2]; 2001 Ops[0] = DAG.getZExtOrTrunc( 2002 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(), 2003 FuncInfo.ExceptionPointerVirtReg, TLI->getPointerTy()), 2004 getCurSDLoc(), ValueVTs[0]); 2005 Ops[1] = DAG.getZExtOrTrunc( 2006 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(), 2007 FuncInfo.ExceptionSelectorVirtReg, TLI->getPointerTy()), 2008 getCurSDLoc(), ValueVTs[1]); 2009 2010 // Merge into one. 2011 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), 2012 DAG.getVTList(&ValueVTs[0], ValueVTs.size()), 2013 &Ops[0], 2); 2014 setValue(&LP, Res); 2015} 2016 2017/// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for 2018/// small case ranges). 2019bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR, 2020 CaseRecVector& WorkList, 2021 const Value* SV, 2022 MachineBasicBlock *Default, 2023 MachineBasicBlock *SwitchBB) { 2024 // Size is the number of Cases represented by this range. 2025 size_t Size = CR.Range.second - CR.Range.first; 2026 if (Size > 3) 2027 return false; 2028 2029 // Get the MachineFunction which holds the current MBB. This is used when 2030 // inserting any additional MBBs necessary to represent the switch. 2031 MachineFunction *CurMF = FuncInfo.MF; 2032 2033 // Figure out which block is immediately after the current one. 2034 MachineBasicBlock *NextBlock = 0; 2035 MachineFunction::iterator BBI = CR.CaseBB; 2036 2037 if (++BBI != FuncInfo.MF->end()) 2038 NextBlock = BBI; 2039 2040 BranchProbabilityInfo *BPI = FuncInfo.BPI; 2041 // If any two of the cases has the same destination, and if one value 2042 // is the same as the other, but has one bit unset that the other has set, 2043 // use bit manipulation to do two compares at once. For example: 2044 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)" 2045 // TODO: This could be extended to merge any 2 cases in switches with 3 cases. 2046 // TODO: Handle cases where CR.CaseBB != SwitchBB. 2047 if (Size == 2 && CR.CaseBB == SwitchBB) { 2048 Case &Small = *CR.Range.first; 2049 Case &Big = *(CR.Range.second-1); 2050 2051 if (Small.Low == Small.High && Big.Low == Big.High && Small.BB == Big.BB) { 2052 const APInt& SmallValue = cast<ConstantInt>(Small.Low)->getValue(); 2053 const APInt& BigValue = cast<ConstantInt>(Big.Low)->getValue(); 2054 2055 // Check that there is only one bit different. 2056 if (BigValue.countPopulation() == SmallValue.countPopulation() + 1 && 2057 (SmallValue | BigValue) == BigValue) { 2058 // Isolate the common bit. 2059 APInt CommonBit = BigValue & ~SmallValue; 2060 assert((SmallValue | CommonBit) == BigValue && 2061 CommonBit.countPopulation() == 1 && "Not a common bit?"); 2062 2063 SDValue CondLHS = getValue(SV); 2064 EVT VT = CondLHS.getValueType(); 2065 SDLoc DL = getCurSDLoc(); 2066 2067 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS, 2068 DAG.getConstant(CommonBit, VT)); 2069 SDValue Cond = DAG.getSetCC(DL, MVT::i1, 2070 Or, DAG.getConstant(BigValue, VT), 2071 ISD::SETEQ); 2072 2073 // Update successor info. 2074 // Both Small and Big will jump to Small.BB, so we sum up the weights. 2075 addSuccessorWithWeight(SwitchBB, Small.BB, 2076 Small.ExtraWeight + Big.ExtraWeight); 2077 addSuccessorWithWeight(SwitchBB, Default, 2078 // The default destination is the first successor in IR. 2079 BPI ? BPI->getEdgeWeight(SwitchBB->getBasicBlock(), (unsigned)0) : 0); 2080 2081 // Insert the true branch. 2082 SDValue BrCond = DAG.getNode(ISD::BRCOND, DL, MVT::Other, 2083 getControlRoot(), Cond, 2084 DAG.getBasicBlock(Small.BB)); 2085 2086 // Insert the false branch. 2087 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond, 2088 DAG.getBasicBlock(Default)); 2089 2090 DAG.setRoot(BrCond); 2091 return true; 2092 } 2093 } 2094 } 2095 2096 // Order cases by weight so the most likely case will be checked first. 2097 uint32_t UnhandledWeights = 0; 2098 if (BPI) { 2099 for (CaseItr I = CR.Range.first, IE = CR.Range.second; I != IE; ++I) { 2100 uint32_t IWeight = I->ExtraWeight; 2101 UnhandledWeights += IWeight; 2102 for (CaseItr J = CR.Range.first; J < I; ++J) { 2103 uint32_t JWeight = J->ExtraWeight; 2104 if (IWeight > JWeight) 2105 std::swap(*I, *J); 2106 } 2107 } 2108 } 2109 // Rearrange the case blocks so that the last one falls through if possible. 2110 Case &BackCase = *(CR.Range.second-1); 2111 if (Size > 1 && 2112 NextBlock && Default != NextBlock && BackCase.BB != NextBlock) { 2113 // The last case block won't fall through into 'NextBlock' if we emit the 2114 // branches in this order. See if rearranging a case value would help. 2115 // We start at the bottom as it's the case with the least weight. 2116 for (Case *I = &*(CR.Range.second-2), *E = &*CR.Range.first-1; I != E; --I) 2117 if (I->BB == NextBlock) { 2118 std::swap(*I, BackCase); 2119 break; 2120 } 2121 } 2122 2123 // Create a CaseBlock record representing a conditional branch to 2124 // the Case's target mbb if the value being switched on SV is equal 2125 // to C. 2126 MachineBasicBlock *CurBlock = CR.CaseBB; 2127 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) { 2128 MachineBasicBlock *FallThrough; 2129 if (I != E-1) { 2130 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock()); 2131 CurMF->insert(BBI, FallThrough); 2132 2133 // Put SV in a virtual register to make it available from the new blocks. 2134 ExportFromCurrentBlock(SV); 2135 } else { 2136 // If the last case doesn't match, go to the default block. 2137 FallThrough = Default; 2138 } 2139 2140 const Value *RHS, *LHS, *MHS; 2141 ISD::CondCode CC; 2142 if (I->High == I->Low) { 2143 // This is just small small case range :) containing exactly 1 case 2144 CC = ISD::SETEQ; 2145 LHS = SV; RHS = I->High; MHS = NULL; 2146 } else { 2147 CC = ISD::SETLE; 2148 LHS = I->Low; MHS = SV; RHS = I->High; 2149 } 2150 2151 // The false weight should be sum of all un-handled cases. 2152 UnhandledWeights -= I->ExtraWeight; 2153 CaseBlock CB(CC, LHS, RHS, MHS, /* truebb */ I->BB, /* falsebb */ FallThrough, 2154 /* me */ CurBlock, 2155 /* trueweight */ I->ExtraWeight, 2156 /* falseweight */ UnhandledWeights); 2157 2158 // If emitting the first comparison, just call visitSwitchCase to emit the 2159 // code into the current block. Otherwise, push the CaseBlock onto the 2160 // vector to be later processed by SDISel, and insert the node's MBB 2161 // before the next MBB. 2162 if (CurBlock == SwitchBB) 2163 visitSwitchCase(CB, SwitchBB); 2164 else 2165 SwitchCases.push_back(CB); 2166 2167 CurBlock = FallThrough; 2168 } 2169 2170 return true; 2171} 2172 2173static inline bool areJTsAllowed(const TargetLowering &TLI) { 2174 return TLI.supportJumpTables() && 2175 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) || 2176 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other)); 2177} 2178 2179static APInt ComputeRange(const APInt &First, const APInt &Last) { 2180 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1; 2181 APInt LastExt = Last.sext(BitWidth), FirstExt = First.sext(BitWidth); 2182 return (LastExt - FirstExt + 1ULL); 2183} 2184 2185/// handleJTSwitchCase - Emit jumptable for current switch case range 2186bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec &CR, 2187 CaseRecVector &WorkList, 2188 const Value *SV, 2189 MachineBasicBlock *Default, 2190 MachineBasicBlock *SwitchBB) { 2191 Case& FrontCase = *CR.Range.first; 2192 Case& BackCase = *(CR.Range.second-1); 2193 2194 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue(); 2195 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue(); 2196 2197 APInt TSize(First.getBitWidth(), 0); 2198 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) 2199 TSize += I->size(); 2200 2201 const TargetLowering *TLI = TM.getTargetLowering(); 2202 if (!areJTsAllowed(*TLI) || TSize.ult(TLI->getMinimumJumpTableEntries())) 2203 return false; 2204 2205 APInt Range = ComputeRange(First, Last); 2206 // The density is TSize / Range. Require at least 40%. 2207 // It should not be possible for IntTSize to saturate for sane code, but make 2208 // sure we handle Range saturation correctly. 2209 uint64_t IntRange = Range.getLimitedValue(UINT64_MAX/10); 2210 uint64_t IntTSize = TSize.getLimitedValue(UINT64_MAX/10); 2211 if (IntTSize * 10 < IntRange * 4) 2212 return false; 2213 2214 DEBUG(dbgs() << "Lowering jump table\n" 2215 << "First entry: " << First << ". Last entry: " << Last << '\n' 2216 << "Range: " << Range << ". Size: " << TSize << ".\n\n"); 2217 2218 // Get the MachineFunction which holds the current MBB. This is used when 2219 // inserting any additional MBBs necessary to represent the switch. 2220 MachineFunction *CurMF = FuncInfo.MF; 2221 2222 // Figure out which block is immediately after the current one. 2223 MachineFunction::iterator BBI = CR.CaseBB; 2224 ++BBI; 2225 2226 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); 2227 2228 // Create a new basic block to hold the code for loading the address 2229 // of the jump table, and jumping to it. Update successor information; 2230 // we will either branch to the default case for the switch, or the jump 2231 // table. 2232 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB); 2233 CurMF->insert(BBI, JumpTableBB); 2234 2235 addSuccessorWithWeight(CR.CaseBB, Default); 2236 addSuccessorWithWeight(CR.CaseBB, JumpTableBB); 2237 2238 // Build a vector of destination BBs, corresponding to each target 2239 // of the jump table. If the value of the jump table slot corresponds to 2240 // a case statement, push the case's BB onto the vector, otherwise, push 2241 // the default BB. 2242 std::vector<MachineBasicBlock*> DestBBs; 2243 APInt TEI = First; 2244 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) { 2245 const APInt &Low = cast<ConstantInt>(I->Low)->getValue(); 2246 const APInt &High = cast<ConstantInt>(I->High)->getValue(); 2247 2248 if (Low.sle(TEI) && TEI.sle(High)) { 2249 DestBBs.push_back(I->BB); 2250 if (TEI==High) 2251 ++I; 2252 } else { 2253 DestBBs.push_back(Default); 2254 } 2255 } 2256 2257 // Calculate weight for each unique destination in CR. 2258 DenseMap<MachineBasicBlock*, uint32_t> DestWeights; 2259 if (FuncInfo.BPI) 2260 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) { 2261 DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr = 2262 DestWeights.find(I->BB); 2263 if (Itr != DestWeights.end()) 2264 Itr->second += I->ExtraWeight; 2265 else 2266 DestWeights[I->BB] = I->ExtraWeight; 2267 } 2268 2269 // Update successor info. Add one edge to each unique successor. 2270 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs()); 2271 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(), 2272 E = DestBBs.end(); I != E; ++I) { 2273 if (!SuccsHandled[(*I)->getNumber()]) { 2274 SuccsHandled[(*I)->getNumber()] = true; 2275 DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr = 2276 DestWeights.find(*I); 2277 addSuccessorWithWeight(JumpTableBB, *I, 2278 Itr != DestWeights.end() ? Itr->second : 0); 2279 } 2280 } 2281 2282 // Create a jump table index for this jump table. 2283 unsigned JTEncoding = TLI->getJumpTableEncoding(); 2284 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding) 2285 ->createJumpTableIndex(DestBBs); 2286 2287 // Set the jump table information so that we can codegen it as a second 2288 // MachineBasicBlock 2289 JumpTable JT(-1U, JTI, JumpTableBB, Default); 2290 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB)); 2291 if (CR.CaseBB == SwitchBB) 2292 visitJumpTableHeader(JT, JTH, SwitchBB); 2293 2294 JTCases.push_back(JumpTableBlock(JTH, JT)); 2295 return true; 2296} 2297 2298/// handleBTSplitSwitchCase - emit comparison and split binary search tree into 2299/// 2 subtrees. 2300bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR, 2301 CaseRecVector& WorkList, 2302 const Value* SV, 2303 MachineBasicBlock* Default, 2304 MachineBasicBlock* SwitchBB) { 2305 // Get the MachineFunction which holds the current MBB. This is used when 2306 // inserting any additional MBBs necessary to represent the switch. 2307 MachineFunction *CurMF = FuncInfo.MF; 2308 2309 // Figure out which block is immediately after the current one. 2310 MachineFunction::iterator BBI = CR.CaseBB; 2311 ++BBI; 2312 2313 Case& FrontCase = *CR.Range.first; 2314 Case& BackCase = *(CR.Range.second-1); 2315 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); 2316 2317 // Size is the number of Cases represented by this range. 2318 unsigned Size = CR.Range.second - CR.Range.first; 2319 2320 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue(); 2321 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue(); 2322 double FMetric = 0; 2323 CaseItr Pivot = CR.Range.first + Size/2; 2324 2325 // Select optimal pivot, maximizing sum density of LHS and RHS. This will 2326 // (heuristically) allow us to emit JumpTable's later. 2327 APInt TSize(First.getBitWidth(), 0); 2328 for (CaseItr I = CR.Range.first, E = CR.Range.second; 2329 I!=E; ++I) 2330 TSize += I->size(); 2331 2332 APInt LSize = FrontCase.size(); 2333 APInt RSize = TSize-LSize; 2334 DEBUG(dbgs() << "Selecting best pivot: \n" 2335 << "First: " << First << ", Last: " << Last <<'\n' 2336 << "LSize: " << LSize << ", RSize: " << RSize << '\n'); 2337 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second; 2338 J!=E; ++I, ++J) { 2339 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue(); 2340 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue(); 2341 APInt Range = ComputeRange(LEnd, RBegin); 2342 assert((Range - 2ULL).isNonNegative() && 2343 "Invalid case distance"); 2344 // Use volatile double here to avoid excess precision issues on some hosts, 2345 // e.g. that use 80-bit X87 registers. 2346 volatile double LDensity = 2347 (double)LSize.roundToDouble() / 2348 (LEnd - First + 1ULL).roundToDouble(); 2349 volatile double RDensity = 2350 (double)RSize.roundToDouble() / 2351 (Last - RBegin + 1ULL).roundToDouble(); 2352 double Metric = Range.logBase2()*(LDensity+RDensity); 2353 // Should always split in some non-trivial place 2354 DEBUG(dbgs() <<"=>Step\n" 2355 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n' 2356 << "LDensity: " << LDensity 2357 << ", RDensity: " << RDensity << '\n' 2358 << "Metric: " << Metric << '\n'); 2359 if (FMetric < Metric) { 2360 Pivot = J; 2361 FMetric = Metric; 2362 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n'); 2363 } 2364 2365 LSize += J->size(); 2366 RSize -= J->size(); 2367 } 2368 2369 const TargetLowering *TLI = TM.getTargetLowering(); 2370 if (areJTsAllowed(*TLI)) { 2371 // If our case is dense we *really* should handle it earlier! 2372 assert((FMetric > 0) && "Should handle dense range earlier!"); 2373 } else { 2374 Pivot = CR.Range.first + Size/2; 2375 } 2376 2377 CaseRange LHSR(CR.Range.first, Pivot); 2378 CaseRange RHSR(Pivot, CR.Range.second); 2379 const Constant *C = Pivot->Low; 2380 MachineBasicBlock *FalseBB = 0, *TrueBB = 0; 2381 2382 // We know that we branch to the LHS if the Value being switched on is 2383 // less than the Pivot value, C. We use this to optimize our binary 2384 // tree a bit, by recognizing that if SV is greater than or equal to the 2385 // LHS's Case Value, and that Case Value is exactly one less than the 2386 // Pivot's Value, then we can branch directly to the LHS's Target, 2387 // rather than creating a leaf node for it. 2388 if ((LHSR.second - LHSR.first) == 1 && 2389 LHSR.first->High == CR.GE && 2390 cast<ConstantInt>(C)->getValue() == 2391 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) { 2392 TrueBB = LHSR.first->BB; 2393 } else { 2394 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB); 2395 CurMF->insert(BBI, TrueBB); 2396 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR)); 2397 2398 // Put SV in a virtual register to make it available from the new blocks. 2399 ExportFromCurrentBlock(SV); 2400 } 2401 2402 // Similar to the optimization above, if the Value being switched on is 2403 // known to be less than the Constant CR.LT, and the current Case Value 2404 // is CR.LT - 1, then we can branch directly to the target block for 2405 // the current Case Value, rather than emitting a RHS leaf node for it. 2406 if ((RHSR.second - RHSR.first) == 1 && CR.LT && 2407 cast<ConstantInt>(RHSR.first->Low)->getValue() == 2408 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) { 2409 FalseBB = RHSR.first->BB; 2410 } else { 2411 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB); 2412 CurMF->insert(BBI, FalseBB); 2413 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR)); 2414 2415 // Put SV in a virtual register to make it available from the new blocks. 2416 ExportFromCurrentBlock(SV); 2417 } 2418 2419 // Create a CaseBlock record representing a conditional branch to 2420 // the LHS node if the value being switched on SV is less than C. 2421 // Otherwise, branch to LHS. 2422 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB); 2423 2424 if (CR.CaseBB == SwitchBB) 2425 visitSwitchCase(CB, SwitchBB); 2426 else 2427 SwitchCases.push_back(CB); 2428 2429 return true; 2430} 2431 2432/// handleBitTestsSwitchCase - if current case range has few destination and 2433/// range span less, than machine word bitwidth, encode case range into series 2434/// of masks and emit bit tests with these masks. 2435bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR, 2436 CaseRecVector& WorkList, 2437 const Value* SV, 2438 MachineBasicBlock* Default, 2439 MachineBasicBlock* SwitchBB) { 2440 const TargetLowering *TLI = TM.getTargetLowering(); 2441 EVT PTy = TLI->getPointerTy(); 2442 unsigned IntPtrBits = PTy.getSizeInBits(); 2443 2444 Case& FrontCase = *CR.Range.first; 2445 Case& BackCase = *(CR.Range.second-1); 2446 2447 // Get the MachineFunction which holds the current MBB. This is used when 2448 // inserting any additional MBBs necessary to represent the switch. 2449 MachineFunction *CurMF = FuncInfo.MF; 2450 2451 // If target does not have legal shift left, do not emit bit tests at all. 2452 if (!TLI->isOperationLegal(ISD::SHL, PTy)) 2453 return false; 2454 2455 size_t numCmps = 0; 2456 for (CaseItr I = CR.Range.first, E = CR.Range.second; 2457 I!=E; ++I) { 2458 // Single case counts one, case range - two. 2459 numCmps += (I->Low == I->High ? 1 : 2); 2460 } 2461 2462 // Count unique destinations 2463 SmallSet<MachineBasicBlock*, 4> Dests; 2464 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) { 2465 Dests.insert(I->BB); 2466 if (Dests.size() > 3) 2467 // Don't bother the code below, if there are too much unique destinations 2468 return false; 2469 } 2470 DEBUG(dbgs() << "Total number of unique destinations: " 2471 << Dests.size() << '\n' 2472 << "Total number of comparisons: " << numCmps << '\n'); 2473 2474 // Compute span of values. 2475 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue(); 2476 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue(); 2477 APInt cmpRange = maxValue - minValue; 2478 2479 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n' 2480 << "Low bound: " << minValue << '\n' 2481 << "High bound: " << maxValue << '\n'); 2482 2483 if (cmpRange.uge(IntPtrBits) || 2484 (!(Dests.size() == 1 && numCmps >= 3) && 2485 !(Dests.size() == 2 && numCmps >= 5) && 2486 !(Dests.size() >= 3 && numCmps >= 6))) 2487 return false; 2488 2489 DEBUG(dbgs() << "Emitting bit tests\n"); 2490 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth()); 2491 2492 // Optimize the case where all the case values fit in a 2493 // word without having to subtract minValue. In this case, 2494 // we can optimize away the subtraction. 2495 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) { 2496 cmpRange = maxValue; 2497 } else { 2498 lowBound = minValue; 2499 } 2500 2501 CaseBitsVector CasesBits; 2502 unsigned i, count = 0; 2503 2504 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) { 2505 MachineBasicBlock* Dest = I->BB; 2506 for (i = 0; i < count; ++i) 2507 if (Dest == CasesBits[i].BB) 2508 break; 2509 2510 if (i == count) { 2511 assert((count < 3) && "Too much destinations to test!"); 2512 CasesBits.push_back(CaseBits(0, Dest, 0, 0/*Weight*/)); 2513 count++; 2514 } 2515 2516 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue(); 2517 const APInt& highValue = cast<ConstantInt>(I->High)->getValue(); 2518 2519 uint64_t lo = (lowValue - lowBound).getZExtValue(); 2520 uint64_t hi = (highValue - lowBound).getZExtValue(); 2521 CasesBits[i].ExtraWeight += I->ExtraWeight; 2522 2523 for (uint64_t j = lo; j <= hi; j++) { 2524 CasesBits[i].Mask |= 1ULL << j; 2525 CasesBits[i].Bits++; 2526 } 2527 2528 } 2529 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp()); 2530 2531 BitTestInfo BTC; 2532 2533 // Figure out which block is immediately after the current one. 2534 MachineFunction::iterator BBI = CR.CaseBB; 2535 ++BBI; 2536 2537 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); 2538 2539 DEBUG(dbgs() << "Cases:\n"); 2540 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) { 2541 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask 2542 << ", Bits: " << CasesBits[i].Bits 2543 << ", BB: " << CasesBits[i].BB << '\n'); 2544 2545 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB); 2546 CurMF->insert(BBI, CaseBB); 2547 BTC.push_back(BitTestCase(CasesBits[i].Mask, 2548 CaseBB, 2549 CasesBits[i].BB, CasesBits[i].ExtraWeight)); 2550 2551 // Put SV in a virtual register to make it available from the new blocks. 2552 ExportFromCurrentBlock(SV); 2553 } 2554 2555 BitTestBlock BTB(lowBound, cmpRange, SV, 2556 -1U, MVT::Other, (CR.CaseBB == SwitchBB), 2557 CR.CaseBB, Default, BTC); 2558 2559 if (CR.CaseBB == SwitchBB) 2560 visitBitTestHeader(BTB, SwitchBB); 2561 2562 BitTestCases.push_back(BTB); 2563 2564 return true; 2565} 2566 2567/// Clusterify - Transform simple list of Cases into list of CaseRange's 2568size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases, 2569 const SwitchInst& SI) { 2570 size_t numCmps = 0; 2571 2572 BranchProbabilityInfo *BPI = FuncInfo.BPI; 2573 // Start with "simple" cases 2574 for (SwitchInst::ConstCaseIt i = SI.case_begin(), e = SI.case_end(); 2575 i != e; ++i) { 2576 const BasicBlock *SuccBB = i.getCaseSuccessor(); 2577 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SuccBB]; 2578 2579 uint32_t ExtraWeight = 2580 BPI ? BPI->getEdgeWeight(SI.getParent(), i.getSuccessorIndex()) : 0; 2581 2582 Cases.push_back(Case(i.getCaseValue(), i.getCaseValue(), 2583 SMBB, ExtraWeight)); 2584 } 2585 std::sort(Cases.begin(), Cases.end(), CaseCmp()); 2586 2587 // Merge case into clusters 2588 if (Cases.size() >= 2) 2589 // Must recompute end() each iteration because it may be 2590 // invalidated by erase if we hold on to it 2591 for (CaseItr I = Cases.begin(), J = llvm::next(Cases.begin()); 2592 J != Cases.end(); ) { 2593 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue(); 2594 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue(); 2595 MachineBasicBlock* nextBB = J->BB; 2596 MachineBasicBlock* currentBB = I->BB; 2597 2598 // If the two neighboring cases go to the same destination, merge them 2599 // into a single case. 2600 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) { 2601 I->High = J->High; 2602 I->ExtraWeight += J->ExtraWeight; 2603 J = Cases.erase(J); 2604 } else { 2605 I = J++; 2606 } 2607 } 2608 2609 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) { 2610 if (I->Low != I->High) 2611 // A range counts double, since it requires two compares. 2612 ++numCmps; 2613 } 2614 2615 return numCmps; 2616} 2617 2618void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First, 2619 MachineBasicBlock *Last) { 2620 // Update JTCases. 2621 for (unsigned i = 0, e = JTCases.size(); i != e; ++i) 2622 if (JTCases[i].first.HeaderBB == First) 2623 JTCases[i].first.HeaderBB = Last; 2624 2625 // Update BitTestCases. 2626 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i) 2627 if (BitTestCases[i].Parent == First) 2628 BitTestCases[i].Parent = Last; 2629} 2630 2631void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) { 2632 MachineBasicBlock *SwitchMBB = FuncInfo.MBB; 2633 2634 // Figure out which block is immediately after the current one. 2635 MachineBasicBlock *NextBlock = 0; 2636 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()]; 2637 2638 // If there is only the default destination, branch to it if it is not the 2639 // next basic block. Otherwise, just fall through. 2640 if (!SI.getNumCases()) { 2641 // Update machine-CFG edges. 2642 2643 // If this is not a fall-through branch, emit the branch. 2644 SwitchMBB->addSuccessor(Default); 2645 if (Default != NextBlock) 2646 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), 2647 MVT::Other, getControlRoot(), 2648 DAG.getBasicBlock(Default))); 2649 2650 return; 2651 } 2652 2653 // If there are any non-default case statements, create a vector of Cases 2654 // representing each one, and sort the vector so that we can efficiently 2655 // create a binary search tree from them. 2656 CaseVector Cases; 2657 size_t numCmps = Clusterify(Cases, SI); 2658 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size() 2659 << ". Total compares: " << numCmps << '\n'); 2660 (void)numCmps; 2661 2662 // Get the Value to be switched on and default basic blocks, which will be 2663 // inserted into CaseBlock records, representing basic blocks in the binary 2664 // search tree. 2665 const Value *SV = SI.getCondition(); 2666 2667 // Push the initial CaseRec onto the worklist 2668 CaseRecVector WorkList; 2669 WorkList.push_back(CaseRec(SwitchMBB,0,0, 2670 CaseRange(Cases.begin(),Cases.end()))); 2671 2672 while (!WorkList.empty()) { 2673 // Grab a record representing a case range to process off the worklist 2674 CaseRec CR = WorkList.back(); 2675 WorkList.pop_back(); 2676 2677 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB)) 2678 continue; 2679 2680 // If the range has few cases (two or less) emit a series of specific 2681 // tests. 2682 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB)) 2683 continue; 2684 2685 // If the switch has more than N blocks, and is at least 40% dense, and the 2686 // target supports indirect branches, then emit a jump table rather than 2687 // lowering the switch to a binary tree of conditional branches. 2688 // N defaults to 4 and is controlled via TLS.getMinimumJumpTableEntries(). 2689 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB)) 2690 continue; 2691 2692 // Emit binary tree. We need to pick a pivot, and push left and right ranges 2693 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call. 2694 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB); 2695 } 2696} 2697 2698void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) { 2699 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB; 2700 2701 // Update machine-CFG edges with unique successors. 2702 SmallSet<BasicBlock*, 32> Done; 2703 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) { 2704 BasicBlock *BB = I.getSuccessor(i); 2705 bool Inserted = Done.insert(BB); 2706 if (!Inserted) 2707 continue; 2708 2709 MachineBasicBlock *Succ = FuncInfo.MBBMap[BB]; 2710 addSuccessorWithWeight(IndirectBrMBB, Succ); 2711 } 2712 2713 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurSDLoc(), 2714 MVT::Other, getControlRoot(), 2715 getValue(I.getAddress()))); 2716} 2717 2718void SelectionDAGBuilder::visitFSub(const User &I) { 2719 // -0.0 - X --> fneg 2720 Type *Ty = I.getType(); 2721 if (isa<Constant>(I.getOperand(0)) && 2722 I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) { 2723 SDValue Op2 = getValue(I.getOperand(1)); 2724 setValue(&I, DAG.getNode(ISD::FNEG, getCurSDLoc(), 2725 Op2.getValueType(), Op2)); 2726 return; 2727 } 2728 2729 visitBinary(I, ISD::FSUB); 2730} 2731 2732void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) { 2733 SDValue Op1 = getValue(I.getOperand(0)); 2734 SDValue Op2 = getValue(I.getOperand(1)); 2735 setValue(&I, DAG.getNode(OpCode, getCurSDLoc(), 2736 Op1.getValueType(), Op1, Op2)); 2737} 2738 2739void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) { 2740 SDValue Op1 = getValue(I.getOperand(0)); 2741 SDValue Op2 = getValue(I.getOperand(1)); 2742 2743 EVT ShiftTy = TM.getTargetLowering()->getShiftAmountTy(Op2.getValueType()); 2744 2745 // Coerce the shift amount to the right type if we can. 2746 if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) { 2747 unsigned ShiftSize = ShiftTy.getSizeInBits(); 2748 unsigned Op2Size = Op2.getValueType().getSizeInBits(); 2749 SDLoc DL = getCurSDLoc(); 2750 2751 // If the operand is smaller than the shift count type, promote it. 2752 if (ShiftSize > Op2Size) 2753 Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2); 2754 2755 // If the operand is larger than the shift count type but the shift 2756 // count type has enough bits to represent any shift value, truncate 2757 // it now. This is a common case and it exposes the truncate to 2758 // optimization early. 2759 else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits())) 2760 Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2); 2761 // Otherwise we'll need to temporarily settle for some other convenient 2762 // type. Type legalization will make adjustments once the shiftee is split. 2763 else 2764 Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32); 2765 } 2766 2767 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), 2768 Op1.getValueType(), Op1, Op2)); 2769} 2770 2771void SelectionDAGBuilder::visitSDiv(const User &I) { 2772 SDValue Op1 = getValue(I.getOperand(0)); 2773 SDValue Op2 = getValue(I.getOperand(1)); 2774 2775 // Turn exact SDivs into multiplications. 2776 // FIXME: This should be in DAGCombiner, but it doesn't have access to the 2777 // exact bit. 2778 if (isa<BinaryOperator>(&I) && cast<BinaryOperator>(&I)->isExact() && 2779 !isa<ConstantSDNode>(Op1) && 2780 isa<ConstantSDNode>(Op2) && !cast<ConstantSDNode>(Op2)->isNullValue()) 2781 setValue(&I, TM.getTargetLowering()->BuildExactSDIV(Op1, Op2, 2782 getCurSDLoc(), DAG)); 2783 else 2784 setValue(&I, DAG.getNode(ISD::SDIV, getCurSDLoc(), Op1.getValueType(), 2785 Op1, Op2)); 2786} 2787 2788void SelectionDAGBuilder::visitICmp(const User &I) { 2789 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE; 2790 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I)) 2791 predicate = IC->getPredicate(); 2792 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I)) 2793 predicate = ICmpInst::Predicate(IC->getPredicate()); 2794 SDValue Op1 = getValue(I.getOperand(0)); 2795 SDValue Op2 = getValue(I.getOperand(1)); 2796 ISD::CondCode Opcode = getICmpCondCode(predicate); 2797 2798 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType()); 2799 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Opcode)); 2800} 2801 2802void SelectionDAGBuilder::visitFCmp(const User &I) { 2803 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE; 2804 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I)) 2805 predicate = FC->getPredicate(); 2806 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I)) 2807 predicate = FCmpInst::Predicate(FC->getPredicate()); 2808 SDValue Op1 = getValue(I.getOperand(0)); 2809 SDValue Op2 = getValue(I.getOperand(1)); 2810 ISD::CondCode Condition = getFCmpCondCode(predicate); 2811 if (TM.Options.NoNaNsFPMath) 2812 Condition = getFCmpCodeWithoutNaN(Condition); 2813 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType()); 2814 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Condition)); 2815} 2816 2817void SelectionDAGBuilder::visitSelect(const User &I) { 2818 SmallVector<EVT, 4> ValueVTs; 2819 ComputeValueVTs(*TM.getTargetLowering(), I.getType(), ValueVTs); 2820 unsigned NumValues = ValueVTs.size(); 2821 if (NumValues == 0) return; 2822 2823 SmallVector<SDValue, 4> Values(NumValues); 2824 SDValue Cond = getValue(I.getOperand(0)); 2825 SDValue TrueVal = getValue(I.getOperand(1)); 2826 SDValue FalseVal = getValue(I.getOperand(2)); 2827 ISD::NodeType OpCode = Cond.getValueType().isVector() ? 2828 ISD::VSELECT : ISD::SELECT; 2829 2830 for (unsigned i = 0; i != NumValues; ++i) 2831 Values[i] = DAG.getNode(OpCode, getCurSDLoc(), 2832 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i), 2833 Cond, 2834 SDValue(TrueVal.getNode(), 2835 TrueVal.getResNo() + i), 2836 SDValue(FalseVal.getNode(), 2837 FalseVal.getResNo() + i)); 2838 2839 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), 2840 DAG.getVTList(&ValueVTs[0], NumValues), 2841 &Values[0], NumValues)); 2842} 2843 2844void SelectionDAGBuilder::visitTrunc(const User &I) { 2845 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest). 2846 SDValue N = getValue(I.getOperand(0)); 2847 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType()); 2848 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), DestVT, N)); 2849} 2850 2851void SelectionDAGBuilder::visitZExt(const User &I) { 2852 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest). 2853 // ZExt also can't be a cast to bool for same reason. So, nothing much to do 2854 SDValue N = getValue(I.getOperand(0)); 2855 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType()); 2856 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurSDLoc(), DestVT, N)); 2857} 2858 2859void SelectionDAGBuilder::visitSExt(const User &I) { 2860 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest). 2861 // SExt also can't be a cast to bool for same reason. So, nothing much to do 2862 SDValue N = getValue(I.getOperand(0)); 2863 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType()); 2864 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurSDLoc(), DestVT, N)); 2865} 2866 2867void SelectionDAGBuilder::visitFPTrunc(const User &I) { 2868 // FPTrunc is never a no-op cast, no need to check 2869 SDValue N = getValue(I.getOperand(0)); 2870 const TargetLowering *TLI = TM.getTargetLowering(); 2871 EVT DestVT = TLI->getValueType(I.getType()); 2872 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurSDLoc(), 2873 DestVT, N, 2874 DAG.getTargetConstant(0, TLI->getPointerTy()))); 2875} 2876 2877void SelectionDAGBuilder::visitFPExt(const User &I) { 2878 // FPExt is never a no-op cast, no need to check 2879 SDValue N = getValue(I.getOperand(0)); 2880 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType()); 2881 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurSDLoc(), DestVT, N)); 2882} 2883 2884void SelectionDAGBuilder::visitFPToUI(const User &I) { 2885 // FPToUI is never a no-op cast, no need to check 2886 SDValue N = getValue(I.getOperand(0)); 2887 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType()); 2888 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurSDLoc(), DestVT, N)); 2889} 2890 2891void SelectionDAGBuilder::visitFPToSI(const User &I) { 2892 // FPToSI is never a no-op cast, no need to check 2893 SDValue N = getValue(I.getOperand(0)); 2894 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType()); 2895 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurSDLoc(), DestVT, N)); 2896} 2897 2898void SelectionDAGBuilder::visitUIToFP(const User &I) { 2899 // UIToFP is never a no-op cast, no need to check 2900 SDValue N = getValue(I.getOperand(0)); 2901 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType()); 2902 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurSDLoc(), DestVT, N)); 2903} 2904 2905void SelectionDAGBuilder::visitSIToFP(const User &I) { 2906 // SIToFP is never a no-op cast, no need to check 2907 SDValue N = getValue(I.getOperand(0)); 2908 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType()); 2909 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurSDLoc(), DestVT, N)); 2910} 2911 2912void SelectionDAGBuilder::visitPtrToInt(const User &I) { 2913 // What to do depends on the size of the integer and the size of the pointer. 2914 // We can either truncate, zero extend, or no-op, accordingly. 2915 SDValue N = getValue(I.getOperand(0)); 2916 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType()); 2917 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT)); 2918} 2919 2920void SelectionDAGBuilder::visitIntToPtr(const User &I) { 2921 // What to do depends on the size of the integer and the size of the pointer. 2922 // We can either truncate, zero extend, or no-op, accordingly. 2923 SDValue N = getValue(I.getOperand(0)); 2924 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType()); 2925 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT)); 2926} 2927 2928void SelectionDAGBuilder::visitBitCast(const User &I) { 2929 SDValue N = getValue(I.getOperand(0)); 2930 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType()); 2931 2932 // BitCast assures us that source and destination are the same size so this is 2933 // either a BITCAST or a no-op. 2934 if (DestVT != N.getValueType()) 2935 setValue(&I, DAG.getNode(ISD::BITCAST, getCurSDLoc(), 2936 DestVT, N)); // convert types. 2937 else 2938 setValue(&I, N); // noop cast. 2939} 2940 2941void SelectionDAGBuilder::visitInsertElement(const User &I) { 2942 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 2943 SDValue InVec = getValue(I.getOperand(0)); 2944 SDValue InVal = getValue(I.getOperand(1)); 2945 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(2)), 2946 getCurSDLoc(), TLI.getVectorIdxTy()); 2947 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurSDLoc(), 2948 TM.getTargetLowering()->getValueType(I.getType()), 2949 InVec, InVal, InIdx)); 2950} 2951 2952void SelectionDAGBuilder::visitExtractElement(const User &I) { 2953 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 2954 SDValue InVec = getValue(I.getOperand(0)); 2955 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(1)), 2956 getCurSDLoc(), TLI.getVectorIdxTy()); 2957 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(), 2958 TM.getTargetLowering()->getValueType(I.getType()), 2959 InVec, InIdx)); 2960} 2961 2962// Utility for visitShuffleVector - Return true if every element in Mask, 2963// beginning from position Pos and ending in Pos+Size, falls within the 2964// specified sequential range [L, L+Pos). or is undef. 2965static bool isSequentialInRange(const SmallVectorImpl<int> &Mask, 2966 unsigned Pos, unsigned Size, int Low) { 2967 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low) 2968 if (Mask[i] >= 0 && Mask[i] != Low) 2969 return false; 2970 return true; 2971} 2972 2973void SelectionDAGBuilder::visitShuffleVector(const User &I) { 2974 SDValue Src1 = getValue(I.getOperand(0)); 2975 SDValue Src2 = getValue(I.getOperand(1)); 2976 2977 SmallVector<int, 8> Mask; 2978 ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask); 2979 unsigned MaskNumElts = Mask.size(); 2980 2981 const TargetLowering *TLI = TM.getTargetLowering(); 2982 EVT VT = TLI->getValueType(I.getType()); 2983 EVT SrcVT = Src1.getValueType(); 2984 unsigned SrcNumElts = SrcVT.getVectorNumElements(); 2985 2986 if (SrcNumElts == MaskNumElts) { 2987 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2, 2988 &Mask[0])); 2989 return; 2990 } 2991 2992 // Normalize the shuffle vector since mask and vector length don't match. 2993 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) { 2994 // Mask is longer than the source vectors and is a multiple of the source 2995 // vectors. We can use concatenate vector to make the mask and vectors 2996 // lengths match. 2997 if (SrcNumElts*2 == MaskNumElts) { 2998 // First check for Src1 in low and Src2 in high 2999 if (isSequentialInRange(Mask, 0, SrcNumElts, 0) && 3000 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) { 3001 // The shuffle is concatenating two vectors together. 3002 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(), 3003 VT, Src1, Src2)); 3004 return; 3005 } 3006 // Then check for Src2 in low and Src1 in high 3007 if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) && 3008 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) { 3009 // The shuffle is concatenating two vectors together. 3010 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(), 3011 VT, Src2, Src1)); 3012 return; 3013 } 3014 } 3015 3016 // Pad both vectors with undefs to make them the same length as the mask. 3017 unsigned NumConcat = MaskNumElts / SrcNumElts; 3018 bool Src1U = Src1.getOpcode() == ISD::UNDEF; 3019 bool Src2U = Src2.getOpcode() == ISD::UNDEF; 3020 SDValue UndefVal = DAG.getUNDEF(SrcVT); 3021 3022 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal); 3023 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal); 3024 MOps1[0] = Src1; 3025 MOps2[0] = Src2; 3026 3027 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS, 3028 getCurSDLoc(), VT, 3029 &MOps1[0], NumConcat); 3030 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS, 3031 getCurSDLoc(), VT, 3032 &MOps2[0], NumConcat); 3033 3034 // Readjust mask for new input vector length. 3035 SmallVector<int, 8> MappedOps; 3036 for (unsigned i = 0; i != MaskNumElts; ++i) { 3037 int Idx = Mask[i]; 3038 if (Idx >= (int)SrcNumElts) 3039 Idx -= SrcNumElts - MaskNumElts; 3040 MappedOps.push_back(Idx); 3041 } 3042 3043 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2, 3044 &MappedOps[0])); 3045 return; 3046 } 3047 3048 if (SrcNumElts > MaskNumElts) { 3049 // Analyze the access pattern of the vector to see if we can extract 3050 // two subvectors and do the shuffle. The analysis is done by calculating 3051 // the range of elements the mask access on both vectors. 3052 int MinRange[2] = { static_cast<int>(SrcNumElts), 3053 static_cast<int>(SrcNumElts)}; 3054 int MaxRange[2] = {-1, -1}; 3055 3056 for (unsigned i = 0; i != MaskNumElts; ++i) { 3057 int Idx = Mask[i]; 3058 unsigned Input = 0; 3059 if (Idx < 0) 3060 continue; 3061 3062 if (Idx >= (int)SrcNumElts) { 3063 Input = 1; 3064 Idx -= SrcNumElts; 3065 } 3066 if (Idx > MaxRange[Input]) 3067 MaxRange[Input] = Idx; 3068 if (Idx < MinRange[Input]) 3069 MinRange[Input] = Idx; 3070 } 3071 3072 // Check if the access is smaller than the vector size and can we find 3073 // a reasonable extract index. 3074 int RangeUse[2] = { -1, -1 }; // 0 = Unused, 1 = Extract, -1 = Can not 3075 // Extract. 3076 int StartIdx[2]; // StartIdx to extract from 3077 for (unsigned Input = 0; Input < 2; ++Input) { 3078 if (MinRange[Input] >= (int)SrcNumElts && MaxRange[Input] < 0) { 3079 RangeUse[Input] = 0; // Unused 3080 StartIdx[Input] = 0; 3081 continue; 3082 } 3083 3084 // Find a good start index that is a multiple of the mask length. Then 3085 // see if the rest of the elements are in range. 3086 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts; 3087 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts && 3088 StartIdx[Input] + MaskNumElts <= SrcNumElts) 3089 RangeUse[Input] = 1; // Extract from a multiple of the mask length. 3090 } 3091 3092 if (RangeUse[0] == 0 && RangeUse[1] == 0) { 3093 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used. 3094 return; 3095 } 3096 if (RangeUse[0] >= 0 && RangeUse[1] >= 0) { 3097 // Extract appropriate subvector and generate a vector shuffle 3098 for (unsigned Input = 0; Input < 2; ++Input) { 3099 SDValue &Src = Input == 0 ? Src1 : Src2; 3100 if (RangeUse[Input] == 0) 3101 Src = DAG.getUNDEF(VT); 3102 else 3103 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurSDLoc(), VT, 3104 Src, DAG.getConstant(StartIdx[Input], 3105 TLI->getVectorIdxTy())); 3106 } 3107 3108 // Calculate new mask. 3109 SmallVector<int, 8> MappedOps; 3110 for (unsigned i = 0; i != MaskNumElts; ++i) { 3111 int Idx = Mask[i]; 3112 if (Idx >= 0) { 3113 if (Idx < (int)SrcNumElts) 3114 Idx -= StartIdx[0]; 3115 else 3116 Idx -= SrcNumElts + StartIdx[1] - MaskNumElts; 3117 } 3118 MappedOps.push_back(Idx); 3119 } 3120 3121 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2, 3122 &MappedOps[0])); 3123 return; 3124 } 3125 } 3126 3127 // We can't use either concat vectors or extract subvectors so fall back to 3128 // replacing the shuffle with extract and build vector. 3129 // to insert and build vector. 3130 EVT EltVT = VT.getVectorElementType(); 3131 EVT IdxVT = TLI->getVectorIdxTy(); 3132 SmallVector<SDValue,8> Ops; 3133 for (unsigned i = 0; i != MaskNumElts; ++i) { 3134 int Idx = Mask[i]; 3135 SDValue Res; 3136 3137 if (Idx < 0) { 3138 Res = DAG.getUNDEF(EltVT); 3139 } else { 3140 SDValue &Src = Idx < (int)SrcNumElts ? Src1 : Src2; 3141 if (Idx >= (int)SrcNumElts) Idx -= SrcNumElts; 3142 3143 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(), 3144 EltVT, Src, DAG.getConstant(Idx, IdxVT)); 3145 } 3146 3147 Ops.push_back(Res); 3148 } 3149 3150 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), 3151 VT, &Ops[0], Ops.size())); 3152} 3153 3154void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) { 3155 const Value *Op0 = I.getOperand(0); 3156 const Value *Op1 = I.getOperand(1); 3157 Type *AggTy = I.getType(); 3158 Type *ValTy = Op1->getType(); 3159 bool IntoUndef = isa<UndefValue>(Op0); 3160 bool FromUndef = isa<UndefValue>(Op1); 3161 3162 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices()); 3163 3164 const TargetLowering *TLI = TM.getTargetLowering(); 3165 SmallVector<EVT, 4> AggValueVTs; 3166 ComputeValueVTs(*TLI, AggTy, AggValueVTs); 3167 SmallVector<EVT, 4> ValValueVTs; 3168 ComputeValueVTs(*TLI, ValTy, ValValueVTs); 3169 3170 unsigned NumAggValues = AggValueVTs.size(); 3171 unsigned NumValValues = ValValueVTs.size(); 3172 SmallVector<SDValue, 4> Values(NumAggValues); 3173 3174 SDValue Agg = getValue(Op0); 3175 unsigned i = 0; 3176 // Copy the beginning value(s) from the original aggregate. 3177 for (; i != LinearIndex; ++i) 3178 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) : 3179 SDValue(Agg.getNode(), Agg.getResNo() + i); 3180 // Copy values from the inserted value(s). 3181 if (NumValValues) { 3182 SDValue Val = getValue(Op1); 3183 for (; i != LinearIndex + NumValValues; ++i) 3184 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) : 3185 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex); 3186 } 3187 // Copy remaining value(s) from the original aggregate. 3188 for (; i != NumAggValues; ++i) 3189 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) : 3190 SDValue(Agg.getNode(), Agg.getResNo() + i); 3191 3192 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), 3193 DAG.getVTList(&AggValueVTs[0], NumAggValues), 3194 &Values[0], NumAggValues)); 3195} 3196 3197void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) { 3198 const Value *Op0 = I.getOperand(0); 3199 Type *AggTy = Op0->getType(); 3200 Type *ValTy = I.getType(); 3201 bool OutOfUndef = isa<UndefValue>(Op0); 3202 3203 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices()); 3204 3205 const TargetLowering *TLI = TM.getTargetLowering(); 3206 SmallVector<EVT, 4> ValValueVTs; 3207 ComputeValueVTs(*TLI, ValTy, ValValueVTs); 3208 3209 unsigned NumValValues = ValValueVTs.size(); 3210 3211 // Ignore a extractvalue that produces an empty object 3212 if (!NumValValues) { 3213 setValue(&I, DAG.getUNDEF(MVT(MVT::Other))); 3214 return; 3215 } 3216 3217 SmallVector<SDValue, 4> Values(NumValValues); 3218 3219 SDValue Agg = getValue(Op0); 3220 // Copy out the selected value(s). 3221 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i) 3222 Values[i - LinearIndex] = 3223 OutOfUndef ? 3224 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) : 3225 SDValue(Agg.getNode(), Agg.getResNo() + i); 3226 3227 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), 3228 DAG.getVTList(&ValValueVTs[0], NumValValues), 3229 &Values[0], NumValValues)); 3230} 3231 3232void SelectionDAGBuilder::visitGetElementPtr(const User &I) { 3233 Value *Op0 = I.getOperand(0); 3234 // Note that the pointer operand may be a vector of pointers. Take the scalar 3235 // element which holds a pointer. 3236 Type *Ty = Op0->getType()->getScalarType(); 3237 unsigned AS = Ty->getPointerAddressSpace(); 3238 SDValue N = getValue(Op0); 3239 3240 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end(); 3241 OI != E; ++OI) { 3242 const Value *Idx = *OI; 3243 if (StructType *StTy = dyn_cast<StructType>(Ty)) { 3244 unsigned Field = cast<Constant>(Idx)->getUniqueInteger().getZExtValue(); 3245 if (Field) { 3246 // N = N + Offset 3247 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field); 3248 N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N, 3249 DAG.getConstant(Offset, N.getValueType())); 3250 } 3251 3252 Ty = StTy->getElementType(Field); 3253 } else { 3254 Ty = cast<SequentialType>(Ty)->getElementType(); 3255 3256 // If this is a constant subscript, handle it quickly. 3257 const TargetLowering *TLI = TM.getTargetLowering(); 3258 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) { 3259 if (CI->isZero()) continue; 3260 uint64_t Offs = 3261 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue(); 3262 SDValue OffsVal; 3263 EVT PTy = TLI->getPointerTy(AS); 3264 unsigned PtrBits = PTy.getSizeInBits(); 3265 if (PtrBits < 64) 3266 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), PTy, 3267 DAG.getConstant(Offs, MVT::i64)); 3268 else 3269 OffsVal = DAG.getConstant(Offs, PTy); 3270 3271 N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N, 3272 OffsVal); 3273 continue; 3274 } 3275 3276 // N = N + Idx * ElementSize; 3277 APInt ElementSize = APInt(TLI->getPointerSizeInBits(AS), 3278 TD->getTypeAllocSize(Ty)); 3279 SDValue IdxN = getValue(Idx); 3280 3281 // If the index is smaller or larger than intptr_t, truncate or extend 3282 // it. 3283 IdxN = DAG.getSExtOrTrunc(IdxN, getCurSDLoc(), N.getValueType()); 3284 3285 // If this is a multiply by a power of two, turn it into a shl 3286 // immediately. This is a very common case. 3287 if (ElementSize != 1) { 3288 if (ElementSize.isPowerOf2()) { 3289 unsigned Amt = ElementSize.logBase2(); 3290 IdxN = DAG.getNode(ISD::SHL, getCurSDLoc(), 3291 N.getValueType(), IdxN, 3292 DAG.getConstant(Amt, IdxN.getValueType())); 3293 } else { 3294 SDValue Scale = DAG.getConstant(ElementSize, IdxN.getValueType()); 3295 IdxN = DAG.getNode(ISD::MUL, getCurSDLoc(), 3296 N.getValueType(), IdxN, Scale); 3297 } 3298 } 3299 3300 N = DAG.getNode(ISD::ADD, getCurSDLoc(), 3301 N.getValueType(), N, IdxN); 3302 } 3303 } 3304 3305 setValue(&I, N); 3306} 3307 3308void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) { 3309 // If this is a fixed sized alloca in the entry block of the function, 3310 // allocate it statically on the stack. 3311 if (FuncInfo.StaticAllocaMap.count(&I)) 3312 return; // getValue will auto-populate this. 3313 3314 Type *Ty = I.getAllocatedType(); 3315 const TargetLowering *TLI = TM.getTargetLowering(); 3316 uint64_t TySize = TLI->getDataLayout()->getTypeAllocSize(Ty); 3317 unsigned Align = 3318 std::max((unsigned)TLI->getDataLayout()->getPrefTypeAlignment(Ty), 3319 I.getAlignment()); 3320 3321 SDValue AllocSize = getValue(I.getArraySize()); 3322 3323 EVT IntPtr = TLI->getPointerTy(); 3324 if (AllocSize.getValueType() != IntPtr) 3325 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurSDLoc(), IntPtr); 3326 3327 AllocSize = DAG.getNode(ISD::MUL, getCurSDLoc(), IntPtr, 3328 AllocSize, 3329 DAG.getConstant(TySize, IntPtr)); 3330 3331 // Handle alignment. If the requested alignment is less than or equal to 3332 // the stack alignment, ignore it. If the size is greater than or equal to 3333 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node. 3334 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment(); 3335 if (Align <= StackAlign) 3336 Align = 0; 3337 3338 // Round the size of the allocation up to the stack alignment size 3339 // by add SA-1 to the size. 3340 AllocSize = DAG.getNode(ISD::ADD, getCurSDLoc(), 3341 AllocSize.getValueType(), AllocSize, 3342 DAG.getIntPtrConstant(StackAlign-1)); 3343 3344 // Mask out the low bits for alignment purposes. 3345 AllocSize = DAG.getNode(ISD::AND, getCurSDLoc(), 3346 AllocSize.getValueType(), AllocSize, 3347 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1))); 3348 3349 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) }; 3350 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other); 3351 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurSDLoc(), 3352 VTs, Ops, 3); 3353 setValue(&I, DSA); 3354 DAG.setRoot(DSA.getValue(1)); 3355 3356 // Inform the Frame Information that we have just allocated a variable-sized 3357 // object. 3358 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1); 3359} 3360 3361void SelectionDAGBuilder::visitLoad(const LoadInst &I) { 3362 if (I.isAtomic()) 3363 return visitAtomicLoad(I); 3364 3365 const Value *SV = I.getOperand(0); 3366 SDValue Ptr = getValue(SV); 3367 3368 Type *Ty = I.getType(); 3369 3370 bool isVolatile = I.isVolatile(); 3371 bool isNonTemporal = I.getMetadata("nontemporal") != 0; 3372 bool isInvariant = I.getMetadata("invariant.load") != 0; 3373 unsigned Alignment = I.getAlignment(); 3374 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa); 3375 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range); 3376 3377 SmallVector<EVT, 4> ValueVTs; 3378 SmallVector<uint64_t, 4> Offsets; 3379 ComputeValueVTs(*TM.getTargetLowering(), Ty, ValueVTs, &Offsets); 3380 unsigned NumValues = ValueVTs.size(); 3381 if (NumValues == 0) 3382 return; 3383 3384 SDValue Root; 3385 bool ConstantMemory = false; 3386 if (I.isVolatile() || NumValues > MaxParallelChains) 3387 // Serialize volatile loads with other side effects. 3388 Root = getRoot(); 3389 else if (AA->pointsToConstantMemory( 3390 AliasAnalysis::Location(SV, AA->getTypeStoreSize(Ty), TBAAInfo))) { 3391 // Do not serialize (non-volatile) loads of constant memory with anything. 3392 Root = DAG.getEntryNode(); 3393 ConstantMemory = true; 3394 } else { 3395 // Do not serialize non-volatile loads against each other. 3396 Root = DAG.getRoot(); 3397 } 3398 3399 SmallVector<SDValue, 4> Values(NumValues); 3400 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains), 3401 NumValues)); 3402 EVT PtrVT = Ptr.getValueType(); 3403 unsigned ChainI = 0; 3404 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) { 3405 // Serializing loads here may result in excessive register pressure, and 3406 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling 3407 // could recover a bit by hoisting nodes upward in the chain by recognizing 3408 // they are side-effect free or do not alias. The optimizer should really 3409 // avoid this case by converting large object/array copies to llvm.memcpy 3410 // (MaxParallelChains should always remain as failsafe). 3411 if (ChainI == MaxParallelChains) { 3412 assert(PendingLoads.empty() && "PendingLoads must be serialized first"); 3413 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), 3414 MVT::Other, &Chains[0], ChainI); 3415 Root = Chain; 3416 ChainI = 0; 3417 } 3418 SDValue A = DAG.getNode(ISD::ADD, getCurSDLoc(), 3419 PtrVT, Ptr, 3420 DAG.getConstant(Offsets[i], PtrVT)); 3421 SDValue L = DAG.getLoad(ValueVTs[i], getCurSDLoc(), Root, 3422 A, MachinePointerInfo(SV, Offsets[i]), isVolatile, 3423 isNonTemporal, isInvariant, Alignment, TBAAInfo, 3424 Ranges); 3425 3426 Values[i] = L; 3427 Chains[ChainI] = L.getValue(1); 3428 } 3429 3430 if (!ConstantMemory) { 3431 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), 3432 MVT::Other, &Chains[0], ChainI); 3433 if (isVolatile) 3434 DAG.setRoot(Chain); 3435 else 3436 PendingLoads.push_back(Chain); 3437 } 3438 3439 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), 3440 DAG.getVTList(&ValueVTs[0], NumValues), 3441 &Values[0], NumValues)); 3442} 3443 3444void SelectionDAGBuilder::visitStore(const StoreInst &I) { 3445 if (I.isAtomic()) 3446 return visitAtomicStore(I); 3447 3448 const Value *SrcV = I.getOperand(0); 3449 const Value *PtrV = I.getOperand(1); 3450 3451 SmallVector<EVT, 4> ValueVTs; 3452 SmallVector<uint64_t, 4> Offsets; 3453 ComputeValueVTs(*TM.getTargetLowering(), SrcV->getType(), ValueVTs, &Offsets); 3454 unsigned NumValues = ValueVTs.size(); 3455 if (NumValues == 0) 3456 return; 3457 3458 // Get the lowered operands. Note that we do this after 3459 // checking if NumResults is zero, because with zero results 3460 // the operands won't have values in the map. 3461 SDValue Src = getValue(SrcV); 3462 SDValue Ptr = getValue(PtrV); 3463 3464 SDValue Root = getRoot(); 3465 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains), 3466 NumValues)); 3467 EVT PtrVT = Ptr.getValueType(); 3468 bool isVolatile = I.isVolatile(); 3469 bool isNonTemporal = I.getMetadata("nontemporal") != 0; 3470 unsigned Alignment = I.getAlignment(); 3471 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa); 3472 3473 unsigned ChainI = 0; 3474 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) { 3475 // See visitLoad comments. 3476 if (ChainI == MaxParallelChains) { 3477 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), 3478 MVT::Other, &Chains[0], ChainI); 3479 Root = Chain; 3480 ChainI = 0; 3481 } 3482 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(), PtrVT, Ptr, 3483 DAG.getConstant(Offsets[i], PtrVT)); 3484 SDValue St = DAG.getStore(Root, getCurSDLoc(), 3485 SDValue(Src.getNode(), Src.getResNo() + i), 3486 Add, MachinePointerInfo(PtrV, Offsets[i]), 3487 isVolatile, isNonTemporal, Alignment, TBAAInfo); 3488 Chains[ChainI] = St; 3489 } 3490 3491 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), 3492 MVT::Other, &Chains[0], ChainI); 3493 DAG.setRoot(StoreNode); 3494} 3495 3496static SDValue InsertFenceForAtomic(SDValue Chain, AtomicOrdering Order, 3497 SynchronizationScope Scope, 3498 bool Before, SDLoc dl, 3499 SelectionDAG &DAG, 3500 const TargetLowering &TLI) { 3501 // Fence, if necessary 3502 if (Before) { 3503 if (Order == AcquireRelease || Order == SequentiallyConsistent) 3504 Order = Release; 3505 else if (Order == Acquire || Order == Monotonic) 3506 return Chain; 3507 } else { 3508 if (Order == AcquireRelease) 3509 Order = Acquire; 3510 else if (Order == Release || Order == Monotonic) 3511 return Chain; 3512 } 3513 SDValue Ops[3]; 3514 Ops[0] = Chain; 3515 Ops[1] = DAG.getConstant(Order, TLI.getPointerTy()); 3516 Ops[2] = DAG.getConstant(Scope, TLI.getPointerTy()); 3517 return DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3); 3518} 3519 3520void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) { 3521 SDLoc dl = getCurSDLoc(); 3522 AtomicOrdering Order = I.getOrdering(); 3523 SynchronizationScope Scope = I.getSynchScope(); 3524 3525 SDValue InChain = getRoot(); 3526 3527 const TargetLowering *TLI = TM.getTargetLowering(); 3528 if (TLI->getInsertFencesForAtomic()) 3529 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl, 3530 DAG, *TLI); 3531 3532 SDValue L = 3533 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl, 3534 getValue(I.getCompareOperand()).getSimpleValueType(), 3535 InChain, 3536 getValue(I.getPointerOperand()), 3537 getValue(I.getCompareOperand()), 3538 getValue(I.getNewValOperand()), 3539 MachinePointerInfo(I.getPointerOperand()), 0 /* Alignment */, 3540 TLI->getInsertFencesForAtomic() ? Monotonic : Order, 3541 Scope); 3542 3543 SDValue OutChain = L.getValue(1); 3544 3545 if (TLI->getInsertFencesForAtomic()) 3546 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl, 3547 DAG, *TLI); 3548 3549 setValue(&I, L); 3550 DAG.setRoot(OutChain); 3551} 3552 3553void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) { 3554 SDLoc dl = getCurSDLoc(); 3555 ISD::NodeType NT; 3556 switch (I.getOperation()) { 3557 default: llvm_unreachable("Unknown atomicrmw operation"); 3558 case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break; 3559 case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break; 3560 case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break; 3561 case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break; 3562 case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break; 3563 case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break; 3564 case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break; 3565 case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break; 3566 case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break; 3567 case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break; 3568 case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break; 3569 } 3570 AtomicOrdering Order = I.getOrdering(); 3571 SynchronizationScope Scope = I.getSynchScope(); 3572 3573 SDValue InChain = getRoot(); 3574 3575 const TargetLowering *TLI = TM.getTargetLowering(); 3576 if (TLI->getInsertFencesForAtomic()) 3577 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl, 3578 DAG, *TLI); 3579 3580 SDValue L = 3581 DAG.getAtomic(NT, dl, 3582 getValue(I.getValOperand()).getSimpleValueType(), 3583 InChain, 3584 getValue(I.getPointerOperand()), 3585 getValue(I.getValOperand()), 3586 I.getPointerOperand(), 0 /* Alignment */, 3587 TLI->getInsertFencesForAtomic() ? Monotonic : Order, 3588 Scope); 3589 3590 SDValue OutChain = L.getValue(1); 3591 3592 if (TLI->getInsertFencesForAtomic()) 3593 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl, 3594 DAG, *TLI); 3595 3596 setValue(&I, L); 3597 DAG.setRoot(OutChain); 3598} 3599 3600void SelectionDAGBuilder::visitFence(const FenceInst &I) { 3601 SDLoc dl = getCurSDLoc(); 3602 const TargetLowering *TLI = TM.getTargetLowering(); 3603 SDValue Ops[3]; 3604 Ops[0] = getRoot(); 3605 Ops[1] = DAG.getConstant(I.getOrdering(), TLI->getPointerTy()); 3606 Ops[2] = DAG.getConstant(I.getSynchScope(), TLI->getPointerTy()); 3607 DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3)); 3608} 3609 3610void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) { 3611 SDLoc dl = getCurSDLoc(); 3612 AtomicOrdering Order = I.getOrdering(); 3613 SynchronizationScope Scope = I.getSynchScope(); 3614 3615 SDValue InChain = getRoot(); 3616 3617 const TargetLowering *TLI = TM.getTargetLowering(); 3618 EVT VT = TLI->getValueType(I.getType()); 3619 3620 if (I.getAlignment() < VT.getSizeInBits() / 8) 3621 report_fatal_error("Cannot generate unaligned atomic load"); 3622 3623 SDValue L = 3624 DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain, 3625 getValue(I.getPointerOperand()), 3626 I.getPointerOperand(), I.getAlignment(), 3627 TLI->getInsertFencesForAtomic() ? Monotonic : Order, 3628 Scope); 3629 3630 SDValue OutChain = L.getValue(1); 3631 3632 if (TLI->getInsertFencesForAtomic()) 3633 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl, 3634 DAG, *TLI); 3635 3636 setValue(&I, L); 3637 DAG.setRoot(OutChain); 3638} 3639 3640void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) { 3641 SDLoc dl = getCurSDLoc(); 3642 3643 AtomicOrdering Order = I.getOrdering(); 3644 SynchronizationScope Scope = I.getSynchScope(); 3645 3646 SDValue InChain = getRoot(); 3647 3648 const TargetLowering *TLI = TM.getTargetLowering(); 3649 EVT VT = TLI->getValueType(I.getValueOperand()->getType()); 3650 3651 if (I.getAlignment() < VT.getSizeInBits() / 8) 3652 report_fatal_error("Cannot generate unaligned atomic store"); 3653 3654 if (TLI->getInsertFencesForAtomic()) 3655 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl, 3656 DAG, *TLI); 3657 3658 SDValue OutChain = 3659 DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT, 3660 InChain, 3661 getValue(I.getPointerOperand()), 3662 getValue(I.getValueOperand()), 3663 I.getPointerOperand(), I.getAlignment(), 3664 TLI->getInsertFencesForAtomic() ? Monotonic : Order, 3665 Scope); 3666 3667 if (TLI->getInsertFencesForAtomic()) 3668 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl, 3669 DAG, *TLI); 3670 3671 DAG.setRoot(OutChain); 3672} 3673 3674/// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC 3675/// node. 3676void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I, 3677 unsigned Intrinsic) { 3678 bool HasChain = !I.doesNotAccessMemory(); 3679 bool OnlyLoad = HasChain && I.onlyReadsMemory(); 3680 3681 // Build the operand list. 3682 SmallVector<SDValue, 8> Ops; 3683 if (HasChain) { // If this intrinsic has side-effects, chainify it. 3684 if (OnlyLoad) { 3685 // We don't need to serialize loads against other loads. 3686 Ops.push_back(DAG.getRoot()); 3687 } else { 3688 Ops.push_back(getRoot()); 3689 } 3690 } 3691 3692 // Info is set by getTgtMemInstrinsic 3693 TargetLowering::IntrinsicInfo Info; 3694 const TargetLowering *TLI = TM.getTargetLowering(); 3695 bool IsTgtIntrinsic = TLI->getTgtMemIntrinsic(Info, I, Intrinsic); 3696 3697 // Add the intrinsic ID as an integer operand if it's not a target intrinsic. 3698 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID || 3699 Info.opc == ISD::INTRINSIC_W_CHAIN) 3700 Ops.push_back(DAG.getTargetConstant(Intrinsic, TLI->getPointerTy())); 3701 3702 // Add all operands of the call to the operand list. 3703 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) { 3704 SDValue Op = getValue(I.getArgOperand(i)); 3705 Ops.push_back(Op); 3706 } 3707 3708 SmallVector<EVT, 4> ValueVTs; 3709 ComputeValueVTs(*TLI, I.getType(), ValueVTs); 3710 3711 if (HasChain) 3712 ValueVTs.push_back(MVT::Other); 3713 3714 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size()); 3715 3716 // Create the node. 3717 SDValue Result; 3718 if (IsTgtIntrinsic) { 3719 // This is target intrinsic that touches memory 3720 Result = DAG.getMemIntrinsicNode(Info.opc, getCurSDLoc(), 3721 VTs, &Ops[0], Ops.size(), 3722 Info.memVT, 3723 MachinePointerInfo(Info.ptrVal, Info.offset), 3724 Info.align, Info.vol, 3725 Info.readMem, Info.writeMem); 3726 } else if (!HasChain) { 3727 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurSDLoc(), 3728 VTs, &Ops[0], Ops.size()); 3729 } else if (!I.getType()->isVoidTy()) { 3730 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurSDLoc(), 3731 VTs, &Ops[0], Ops.size()); 3732 } else { 3733 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurSDLoc(), 3734 VTs, &Ops[0], Ops.size()); 3735 } 3736 3737 if (HasChain) { 3738 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1); 3739 if (OnlyLoad) 3740 PendingLoads.push_back(Chain); 3741 else 3742 DAG.setRoot(Chain); 3743 } 3744 3745 if (!I.getType()->isVoidTy()) { 3746 if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) { 3747 EVT VT = TLI->getValueType(PTy); 3748 Result = DAG.getNode(ISD::BITCAST, getCurSDLoc(), VT, Result); 3749 } 3750 3751 setValue(&I, Result); 3752 } 3753} 3754 3755/// GetSignificand - Get the significand and build it into a floating-point 3756/// number with exponent of 1: 3757/// 3758/// Op = (Op & 0x007fffff) | 0x3f800000; 3759/// 3760/// where Op is the hexadecimal representation of floating point value. 3761static SDValue 3762GetSignificand(SelectionDAG &DAG, SDValue Op, SDLoc dl) { 3763 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op, 3764 DAG.getConstant(0x007fffff, MVT::i32)); 3765 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1, 3766 DAG.getConstant(0x3f800000, MVT::i32)); 3767 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2); 3768} 3769 3770/// GetExponent - Get the exponent: 3771/// 3772/// (float)(int)(((Op & 0x7f800000) >> 23) - 127); 3773/// 3774/// where Op is the hexadecimal representation of floating point value. 3775static SDValue 3776GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI, 3777 SDLoc dl) { 3778 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op, 3779 DAG.getConstant(0x7f800000, MVT::i32)); 3780 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0, 3781 DAG.getConstant(23, TLI.getPointerTy())); 3782 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1, 3783 DAG.getConstant(127, MVT::i32)); 3784 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2); 3785} 3786 3787/// getF32Constant - Get 32-bit floating point constant. 3788static SDValue 3789getF32Constant(SelectionDAG &DAG, unsigned Flt) { 3790 return DAG.getConstantFP(APFloat(APFloat::IEEEsingle, APInt(32, Flt)), 3791 MVT::f32); 3792} 3793 3794/// expandExp - Lower an exp intrinsic. Handles the special sequences for 3795/// limited-precision mode. 3796static SDValue expandExp(SDLoc dl, SDValue Op, SelectionDAG &DAG, 3797 const TargetLowering &TLI) { 3798 if (Op.getValueType() == MVT::f32 && 3799 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3800 3801 // Put the exponent in the right bit position for later addition to the 3802 // final result: 3803 // 3804 // #define LOG2OFe 1.4426950f 3805 // IntegerPartOfX = ((int32_t)(X * LOG2OFe)); 3806 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op, 3807 getF32Constant(DAG, 0x3fb8aa3b)); 3808 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0); 3809 3810 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX; 3811 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); 3812 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1); 3813 3814 // IntegerPartOfX <<= 23; 3815 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, 3816 DAG.getConstant(23, TLI.getPointerTy())); 3817 3818 SDValue TwoToFracPartOfX; 3819 if (LimitFloatPrecision <= 6) { 3820 // For floating-point precision of 6: 3821 // 3822 // TwoToFractionalPartOfX = 3823 // 0.997535578f + 3824 // (0.735607626f + 0.252464424f * x) * x; 3825 // 3826 // error 0.0144103317, which is 6 bits 3827 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3828 getF32Constant(DAG, 0x3e814304)); 3829 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3830 getF32Constant(DAG, 0x3f3c50c8)); 3831 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3832 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3833 getF32Constant(DAG, 0x3f7f5e7e)); 3834 } else if (LimitFloatPrecision <= 12) { 3835 // For floating-point precision of 12: 3836 // 3837 // TwoToFractionalPartOfX = 3838 // 0.999892986f + 3839 // (0.696457318f + 3840 // (0.224338339f + 0.792043434e-1f * x) * x) * x; 3841 // 3842 // 0.000107046256 error, which is 13 to 14 bits 3843 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3844 getF32Constant(DAG, 0x3da235e3)); 3845 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3846 getF32Constant(DAG, 0x3e65b8f3)); 3847 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3848 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3849 getF32Constant(DAG, 0x3f324b07)); 3850 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3851 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3852 getF32Constant(DAG, 0x3f7ff8fd)); 3853 } else { // LimitFloatPrecision <= 18 3854 // For floating-point precision of 18: 3855 // 3856 // TwoToFractionalPartOfX = 3857 // 0.999999982f + 3858 // (0.693148872f + 3859 // (0.240227044f + 3860 // (0.554906021e-1f + 3861 // (0.961591928e-2f + 3862 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; 3863 // 3864 // error 2.47208000*10^(-7), which is better than 18 bits 3865 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3866 getF32Constant(DAG, 0x3924b03e)); 3867 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 3868 getF32Constant(DAG, 0x3ab24b87)); 3869 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3870 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3871 getF32Constant(DAG, 0x3c1d8c17)); 3872 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3873 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 3874 getF32Constant(DAG, 0x3d634a1d)); 3875 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3876 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 3877 getF32Constant(DAG, 0x3e75fe14)); 3878 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 3879 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, 3880 getF32Constant(DAG, 0x3f317234)); 3881 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); 3882 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, 3883 getF32Constant(DAG, 0x3f800000)); 3884 } 3885 3886 // Add the exponent into the result in integer domain. 3887 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, TwoToFracPartOfX); 3888 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, 3889 DAG.getNode(ISD::ADD, dl, MVT::i32, 3890 t13, IntegerPartOfX)); 3891 } 3892 3893 // No special expansion. 3894 return DAG.getNode(ISD::FEXP, dl, Op.getValueType(), Op); 3895} 3896 3897/// expandLog - Lower a log intrinsic. Handles the special sequences for 3898/// limited-precision mode. 3899static SDValue expandLog(SDLoc dl, SDValue Op, SelectionDAG &DAG, 3900 const TargetLowering &TLI) { 3901 if (Op.getValueType() == MVT::f32 && 3902 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3903 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); 3904 3905 // Scale the exponent by log(2) [0.69314718f]. 3906 SDValue Exp = GetExponent(DAG, Op1, TLI, dl); 3907 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp, 3908 getF32Constant(DAG, 0x3f317218)); 3909 3910 // Get the significand and build it into a floating-point number with 3911 // exponent of 1. 3912 SDValue X = GetSignificand(DAG, Op1, dl); 3913 3914 SDValue LogOfMantissa; 3915 if (LimitFloatPrecision <= 6) { 3916 // For floating-point precision of 6: 3917 // 3918 // LogofMantissa = 3919 // -1.1609546f + 3920 // (1.4034025f - 0.23903021f * x) * x; 3921 // 3922 // error 0.0034276066, which is better than 8 bits 3923 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3924 getF32Constant(DAG, 0xbe74c456)); 3925 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3926 getF32Constant(DAG, 0x3fb3a2b1)); 3927 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3928 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3929 getF32Constant(DAG, 0x3f949a29)); 3930 } else if (LimitFloatPrecision <= 12) { 3931 // For floating-point precision of 12: 3932 // 3933 // LogOfMantissa = 3934 // -1.7417939f + 3935 // (2.8212026f + 3936 // (-1.4699568f + 3937 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x; 3938 // 3939 // error 0.000061011436, which is 14 bits 3940 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3941 getF32Constant(DAG, 0xbd67b6d6)); 3942 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3943 getF32Constant(DAG, 0x3ee4f4b8)); 3944 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3945 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3946 getF32Constant(DAG, 0x3fbc278b)); 3947 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3948 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3949 getF32Constant(DAG, 0x40348e95)); 3950 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3951 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, 3952 getF32Constant(DAG, 0x3fdef31a)); 3953 } else { // LimitFloatPrecision <= 18 3954 // For floating-point precision of 18: 3955 // 3956 // LogOfMantissa = 3957 // -2.1072184f + 3958 // (4.2372794f + 3959 // (-3.7029485f + 3960 // (2.2781945f + 3961 // (-0.87823314f + 3962 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x; 3963 // 3964 // error 0.0000023660568, which is better than 18 bits 3965 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 3966 getF32Constant(DAG, 0xbc91e5ac)); 3967 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 3968 getF32Constant(DAG, 0x3e4350aa)); 3969 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 3970 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 3971 getF32Constant(DAG, 0x3f60d3e3)); 3972 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 3973 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 3974 getF32Constant(DAG, 0x4011cdf0)); 3975 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 3976 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, 3977 getF32Constant(DAG, 0x406cfd1c)); 3978 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 3979 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 3980 getF32Constant(DAG, 0x408797cb)); 3981 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 3982 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10, 3983 getF32Constant(DAG, 0x4006dcab)); 3984 } 3985 3986 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, LogOfMantissa); 3987 } 3988 3989 // No special expansion. 3990 return DAG.getNode(ISD::FLOG, dl, Op.getValueType(), Op); 3991} 3992 3993/// expandLog2 - Lower a log2 intrinsic. Handles the special sequences for 3994/// limited-precision mode. 3995static SDValue expandLog2(SDLoc dl, SDValue Op, SelectionDAG &DAG, 3996 const TargetLowering &TLI) { 3997 if (Op.getValueType() == MVT::f32 && 3998 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 3999 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); 4000 4001 // Get the exponent. 4002 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl); 4003 4004 // Get the significand and build it into a floating-point number with 4005 // exponent of 1. 4006 SDValue X = GetSignificand(DAG, Op1, dl); 4007 4008 // Different possible minimax approximations of significand in 4009 // floating-point for various degrees of accuracy over [1,2]. 4010 SDValue Log2ofMantissa; 4011 if (LimitFloatPrecision <= 6) { 4012 // For floating-point precision of 6: 4013 // 4014 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x; 4015 // 4016 // error 0.0049451742, which is more than 7 bits 4017 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 4018 getF32Constant(DAG, 0xbeb08fe0)); 4019 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 4020 getF32Constant(DAG, 0x40019463)); 4021 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 4022 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 4023 getF32Constant(DAG, 0x3fd6633d)); 4024 } else if (LimitFloatPrecision <= 12) { 4025 // For floating-point precision of 12: 4026 // 4027 // Log2ofMantissa = 4028 // -2.51285454f + 4029 // (4.07009056f + 4030 // (-2.12067489f + 4031 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x; 4032 // 4033 // error 0.0000876136000, which is better than 13 bits 4034 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 4035 getF32Constant(DAG, 0xbda7262e)); 4036 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 4037 getF32Constant(DAG, 0x3f25280b)); 4038 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 4039 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 4040 getF32Constant(DAG, 0x4007b923)); 4041 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 4042 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 4043 getF32Constant(DAG, 0x40823e2f)); 4044 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 4045 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, 4046 getF32Constant(DAG, 0x4020d29c)); 4047 } else { // LimitFloatPrecision <= 18 4048 // For floating-point precision of 18: 4049 // 4050 // Log2ofMantissa = 4051 // -3.0400495f + 4052 // (6.1129976f + 4053 // (-5.3420409f + 4054 // (3.2865683f + 4055 // (-1.2669343f + 4056 // (0.27515199f - 4057 // 0.25691327e-1f * x) * x) * x) * x) * x) * x; 4058 // 4059 // error 0.0000018516, which is better than 18 bits 4060 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 4061 getF32Constant(DAG, 0xbcd2769e)); 4062 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 4063 getF32Constant(DAG, 0x3e8ce0b9)); 4064 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 4065 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 4066 getF32Constant(DAG, 0x3fa22ae7)); 4067 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 4068 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 4069 getF32Constant(DAG, 0x40525723)); 4070 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 4071 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6, 4072 getF32Constant(DAG, 0x40aaf200)); 4073 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 4074 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 4075 getF32Constant(DAG, 0x40c39dad)); 4076 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 4077 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10, 4078 getF32Constant(DAG, 0x4042902c)); 4079 } 4080 4081 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log2ofMantissa); 4082 } 4083 4084 // No special expansion. 4085 return DAG.getNode(ISD::FLOG2, dl, Op.getValueType(), Op); 4086} 4087 4088/// expandLog10 - Lower a log10 intrinsic. Handles the special sequences for 4089/// limited-precision mode. 4090static SDValue expandLog10(SDLoc dl, SDValue Op, SelectionDAG &DAG, 4091 const TargetLowering &TLI) { 4092 if (Op.getValueType() == MVT::f32 && 4093 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 4094 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); 4095 4096 // Scale the exponent by log10(2) [0.30102999f]. 4097 SDValue Exp = GetExponent(DAG, Op1, TLI, dl); 4098 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp, 4099 getF32Constant(DAG, 0x3e9a209a)); 4100 4101 // Get the significand and build it into a floating-point number with 4102 // exponent of 1. 4103 SDValue X = GetSignificand(DAG, Op1, dl); 4104 4105 SDValue Log10ofMantissa; 4106 if (LimitFloatPrecision <= 6) { 4107 // For floating-point precision of 6: 4108 // 4109 // Log10ofMantissa = 4110 // -0.50419619f + 4111 // (0.60948995f - 0.10380950f * x) * x; 4112 // 4113 // error 0.0014886165, which is 6 bits 4114 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 4115 getF32Constant(DAG, 0xbdd49a13)); 4116 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0, 4117 getF32Constant(DAG, 0x3f1c0789)); 4118 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 4119 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2, 4120 getF32Constant(DAG, 0x3f011300)); 4121 } else if (LimitFloatPrecision <= 12) { 4122 // For floating-point precision of 12: 4123 // 4124 // Log10ofMantissa = 4125 // -0.64831180f + 4126 // (0.91751397f + 4127 // (-0.31664806f + 0.47637168e-1f * x) * x) * x; 4128 // 4129 // error 0.00019228036, which is better than 12 bits 4130 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 4131 getF32Constant(DAG, 0x3d431f31)); 4132 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, 4133 getF32Constant(DAG, 0x3ea21fb2)); 4134 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 4135 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 4136 getF32Constant(DAG, 0x3f6ae232)); 4137 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 4138 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4, 4139 getF32Constant(DAG, 0x3f25f7c3)); 4140 } else { // LimitFloatPrecision <= 18 4141 // For floating-point precision of 18: 4142 // 4143 // Log10ofMantissa = 4144 // -0.84299375f + 4145 // (1.5327582f + 4146 // (-1.0688956f + 4147 // (0.49102474f + 4148 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x; 4149 // 4150 // error 0.0000037995730, which is better than 18 bits 4151 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 4152 getF32Constant(DAG, 0x3c5d51ce)); 4153 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, 4154 getF32Constant(DAG, 0x3e00685a)); 4155 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X); 4156 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 4157 getF32Constant(DAG, 0x3efb6798)); 4158 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 4159 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4, 4160 getF32Constant(DAG, 0x3f88d192)); 4161 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 4162 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 4163 getF32Constant(DAG, 0x3fc4316c)); 4164 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 4165 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8, 4166 getF32Constant(DAG, 0x3f57ce70)); 4167 } 4168 4169 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log10ofMantissa); 4170 } 4171 4172 // No special expansion. 4173 return DAG.getNode(ISD::FLOG10, dl, Op.getValueType(), Op); 4174} 4175 4176/// expandExp2 - Lower an exp2 intrinsic. Handles the special sequences for 4177/// limited-precision mode. 4178static SDValue expandExp2(SDLoc dl, SDValue Op, SelectionDAG &DAG, 4179 const TargetLowering &TLI) { 4180 if (Op.getValueType() == MVT::f32 && 4181 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 4182 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op); 4183 4184 // FractionalPartOfX = x - (float)IntegerPartOfX; 4185 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); 4186 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1); 4187 4188 // IntegerPartOfX <<= 23; 4189 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, 4190 DAG.getConstant(23, TLI.getPointerTy())); 4191 4192 SDValue TwoToFractionalPartOfX; 4193 if (LimitFloatPrecision <= 6) { 4194 // For floating-point precision of 6: 4195 // 4196 // TwoToFractionalPartOfX = 4197 // 0.997535578f + 4198 // (0.735607626f + 0.252464424f * x) * x; 4199 // 4200 // error 0.0144103317, which is 6 bits 4201 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 4202 getF32Constant(DAG, 0x3e814304)); 4203 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 4204 getF32Constant(DAG, 0x3f3c50c8)); 4205 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 4206 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 4207 getF32Constant(DAG, 0x3f7f5e7e)); 4208 } else if (LimitFloatPrecision <= 12) { 4209 // For floating-point precision of 12: 4210 // 4211 // TwoToFractionalPartOfX = 4212 // 0.999892986f + 4213 // (0.696457318f + 4214 // (0.224338339f + 0.792043434e-1f * x) * x) * x; 4215 // 4216 // error 0.000107046256, which is 13 to 14 bits 4217 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 4218 getF32Constant(DAG, 0x3da235e3)); 4219 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 4220 getF32Constant(DAG, 0x3e65b8f3)); 4221 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 4222 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 4223 getF32Constant(DAG, 0x3f324b07)); 4224 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 4225 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 4226 getF32Constant(DAG, 0x3f7ff8fd)); 4227 } else { // LimitFloatPrecision <= 18 4228 // For floating-point precision of 18: 4229 // 4230 // TwoToFractionalPartOfX = 4231 // 0.999999982f + 4232 // (0.693148872f + 4233 // (0.240227044f + 4234 // (0.554906021e-1f + 4235 // (0.961591928e-2f + 4236 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; 4237 // error 2.47208000*10^(-7), which is better than 18 bits 4238 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 4239 getF32Constant(DAG, 0x3924b03e)); 4240 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 4241 getF32Constant(DAG, 0x3ab24b87)); 4242 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 4243 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 4244 getF32Constant(DAG, 0x3c1d8c17)); 4245 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 4246 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 4247 getF32Constant(DAG, 0x3d634a1d)); 4248 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 4249 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 4250 getF32Constant(DAG, 0x3e75fe14)); 4251 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 4252 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, 4253 getF32Constant(DAG, 0x3f317234)); 4254 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); 4255 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, 4256 getF32Constant(DAG, 0x3f800000)); 4257 } 4258 4259 // Add the exponent into the result in integer domain. 4260 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, 4261 TwoToFractionalPartOfX); 4262 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, 4263 DAG.getNode(ISD::ADD, dl, MVT::i32, 4264 t13, IntegerPartOfX)); 4265 } 4266 4267 // No special expansion. 4268 return DAG.getNode(ISD::FEXP2, dl, Op.getValueType(), Op); 4269} 4270 4271/// visitPow - Lower a pow intrinsic. Handles the special sequences for 4272/// limited-precision mode with x == 10.0f. 4273static SDValue expandPow(SDLoc dl, SDValue LHS, SDValue RHS, 4274 SelectionDAG &DAG, const TargetLowering &TLI) { 4275 bool IsExp10 = false; 4276 if (LHS.getValueType() == MVT::f32 && LHS.getValueType() == MVT::f32 && 4277 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { 4278 if (ConstantFPSDNode *LHSC = dyn_cast<ConstantFPSDNode>(LHS)) { 4279 APFloat Ten(10.0f); 4280 IsExp10 = LHSC->isExactlyValue(Ten); 4281 } 4282 } 4283 4284 if (IsExp10) { 4285 // Put the exponent in the right bit position for later addition to the 4286 // final result: 4287 // 4288 // #define LOG2OF10 3.3219281f 4289 // IntegerPartOfX = (int32_t)(x * LOG2OF10); 4290 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, RHS, 4291 getF32Constant(DAG, 0x40549a78)); 4292 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0); 4293 4294 // FractionalPartOfX = x - (float)IntegerPartOfX; 4295 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX); 4296 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1); 4297 4298 // IntegerPartOfX <<= 23; 4299 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX, 4300 DAG.getConstant(23, TLI.getPointerTy())); 4301 4302 SDValue TwoToFractionalPartOfX; 4303 if (LimitFloatPrecision <= 6) { 4304 // For floating-point precision of 6: 4305 // 4306 // twoToFractionalPartOfX = 4307 // 0.997535578f + 4308 // (0.735607626f + 0.252464424f * x) * x; 4309 // 4310 // error 0.0144103317, which is 6 bits 4311 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 4312 getF32Constant(DAG, 0x3e814304)); 4313 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 4314 getF32Constant(DAG, 0x3f3c50c8)); 4315 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 4316 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 4317 getF32Constant(DAG, 0x3f7f5e7e)); 4318 } else if (LimitFloatPrecision <= 12) { 4319 // For floating-point precision of 12: 4320 // 4321 // TwoToFractionalPartOfX = 4322 // 0.999892986f + 4323 // (0.696457318f + 4324 // (0.224338339f + 0.792043434e-1f * x) * x) * x; 4325 // 4326 // error 0.000107046256, which is 13 to 14 bits 4327 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 4328 getF32Constant(DAG, 0x3da235e3)); 4329 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 4330 getF32Constant(DAG, 0x3e65b8f3)); 4331 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 4332 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 4333 getF32Constant(DAG, 0x3f324b07)); 4334 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 4335 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 4336 getF32Constant(DAG, 0x3f7ff8fd)); 4337 } else { // LimitFloatPrecision <= 18 4338 // For floating-point precision of 18: 4339 // 4340 // TwoToFractionalPartOfX = 4341 // 0.999999982f + 4342 // (0.693148872f + 4343 // (0.240227044f + 4344 // (0.554906021e-1f + 4345 // (0.961591928e-2f + 4346 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; 4347 // error 2.47208000*10^(-7), which is better than 18 bits 4348 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X, 4349 getF32Constant(DAG, 0x3924b03e)); 4350 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2, 4351 getF32Constant(DAG, 0x3ab24b87)); 4352 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X); 4353 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4, 4354 getF32Constant(DAG, 0x3c1d8c17)); 4355 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X); 4356 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6, 4357 getF32Constant(DAG, 0x3d634a1d)); 4358 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X); 4359 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8, 4360 getF32Constant(DAG, 0x3e75fe14)); 4361 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X); 4362 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10, 4363 getF32Constant(DAG, 0x3f317234)); 4364 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X); 4365 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12, 4366 getF32Constant(DAG, 0x3f800000)); 4367 } 4368 4369 SDValue t13 = DAG.getNode(ISD::BITCAST, dl,MVT::i32,TwoToFractionalPartOfX); 4370 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, 4371 DAG.getNode(ISD::ADD, dl, MVT::i32, 4372 t13, IntegerPartOfX)); 4373 } 4374 4375 // No special expansion. 4376 return DAG.getNode(ISD::FPOW, dl, LHS.getValueType(), LHS, RHS); 4377} 4378 4379 4380/// ExpandPowI - Expand a llvm.powi intrinsic. 4381static SDValue ExpandPowI(SDLoc DL, SDValue LHS, SDValue RHS, 4382 SelectionDAG &DAG) { 4383 // If RHS is a constant, we can expand this out to a multiplication tree, 4384 // otherwise we end up lowering to a call to __powidf2 (for example). When 4385 // optimizing for size, we only want to do this if the expansion would produce 4386 // a small number of multiplies, otherwise we do the full expansion. 4387 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) { 4388 // Get the exponent as a positive value. 4389 unsigned Val = RHSC->getSExtValue(); 4390 if ((int)Val < 0) Val = -Val; 4391 4392 // powi(x, 0) -> 1.0 4393 if (Val == 0) 4394 return DAG.getConstantFP(1.0, LHS.getValueType()); 4395 4396 const Function *F = DAG.getMachineFunction().getFunction(); 4397 if (!F->getAttributes().hasAttribute(AttributeSet::FunctionIndex, 4398 Attribute::OptimizeForSize) || 4399 // If optimizing for size, don't insert too many multiplies. This 4400 // inserts up to 5 multiplies. 4401 CountPopulation_32(Val)+Log2_32(Val) < 7) { 4402 // We use the simple binary decomposition method to generate the multiply 4403 // sequence. There are more optimal ways to do this (for example, 4404 // powi(x,15) generates one more multiply than it should), but this has 4405 // the benefit of being both really simple and much better than a libcall. 4406 SDValue Res; // Logically starts equal to 1.0 4407 SDValue CurSquare = LHS; 4408 while (Val) { 4409 if (Val & 1) { 4410 if (Res.getNode()) 4411 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare); 4412 else 4413 Res = CurSquare; // 1.0*CurSquare. 4414 } 4415 4416 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(), 4417 CurSquare, CurSquare); 4418 Val >>= 1; 4419 } 4420 4421 // If the original was negative, invert the result, producing 1/(x*x*x). 4422 if (RHSC->getSExtValue() < 0) 4423 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(), 4424 DAG.getConstantFP(1.0, LHS.getValueType()), Res); 4425 return Res; 4426 } 4427 } 4428 4429 // Otherwise, expand to a libcall. 4430 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS); 4431} 4432 4433// getTruncatedArgReg - Find underlying register used for an truncated 4434// argument. 4435static unsigned getTruncatedArgReg(const SDValue &N) { 4436 if (N.getOpcode() != ISD::TRUNCATE) 4437 return 0; 4438 4439 const SDValue &Ext = N.getOperand(0); 4440 if (Ext.getOpcode() == ISD::AssertZext || 4441 Ext.getOpcode() == ISD::AssertSext) { 4442 const SDValue &CFR = Ext.getOperand(0); 4443 if (CFR.getOpcode() == ISD::CopyFromReg) 4444 return cast<RegisterSDNode>(CFR.getOperand(1))->getReg(); 4445 if (CFR.getOpcode() == ISD::TRUNCATE) 4446 return getTruncatedArgReg(CFR); 4447 } 4448 return 0; 4449} 4450 4451/// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function 4452/// argument, create the corresponding DBG_VALUE machine instruction for it now. 4453/// At the end of instruction selection, they will be inserted to the entry BB. 4454bool 4455SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable, 4456 int64_t Offset, 4457 const SDValue &N) { 4458 const Argument *Arg = dyn_cast<Argument>(V); 4459 if (!Arg) 4460 return false; 4461 4462 MachineFunction &MF = DAG.getMachineFunction(); 4463 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo(); 4464 4465 // Ignore inlined function arguments here. 4466 DIVariable DV(Variable); 4467 if (DV.isInlinedFnArgument(MF.getFunction())) 4468 return false; 4469 4470 Optional<MachineOperand> Op; 4471 // Some arguments' frame index is recorded during argument lowering. 4472 if (int FI = FuncInfo.getArgumentFrameIndex(Arg)) 4473 Op = MachineOperand::CreateFI(FI); 4474 4475 if (!Op && N.getNode()) { 4476 unsigned Reg; 4477 if (N.getOpcode() == ISD::CopyFromReg) 4478 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg(); 4479 else 4480 Reg = getTruncatedArgReg(N); 4481 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) { 4482 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 4483 unsigned PR = RegInfo.getLiveInPhysReg(Reg); 4484 if (PR) 4485 Reg = PR; 4486 } 4487 if (Reg) 4488 Op = MachineOperand::CreateReg(Reg, false); 4489 } 4490 4491 if (!Op) { 4492 // Check if ValueMap has reg number. 4493 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V); 4494 if (VMI != FuncInfo.ValueMap.end()) 4495 Op = MachineOperand::CreateReg(VMI->second, false); 4496 } 4497 4498 if (!Op && N.getNode()) 4499 // Check if frame index is available. 4500 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode())) 4501 if (FrameIndexSDNode *FINode = 4502 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode())) 4503 Op = MachineOperand::CreateFI(FINode->getIndex()); 4504 4505 if (!Op) 4506 return false; 4507 4508 // FIXME: This does not handle register-indirect values at offset 0. 4509 bool IsIndirect = Offset != 0; 4510 if (Op->isReg()) 4511 FuncInfo.ArgDbgValues.push_back(BuildMI(MF, getCurDebugLoc(), 4512 TII->get(TargetOpcode::DBG_VALUE), 4513 IsIndirect, 4514 Op->getReg(), Offset, Variable)); 4515 else 4516 FuncInfo.ArgDbgValues.push_back( 4517 BuildMI(MF, getCurDebugLoc(), TII->get(TargetOpcode::DBG_VALUE)) 4518 .addOperand(*Op).addImm(Offset).addMetadata(Variable)); 4519 4520 return true; 4521} 4522 4523// VisualStudio defines setjmp as _setjmp 4524#if defined(_MSC_VER) && defined(setjmp) && \ 4525 !defined(setjmp_undefined_for_msvc) 4526# pragma push_macro("setjmp") 4527# undef setjmp 4528# define setjmp_undefined_for_msvc 4529#endif 4530 4531/// visitIntrinsicCall - Lower the call to the specified intrinsic function. If 4532/// we want to emit this as a call to a named external function, return the name 4533/// otherwise lower it and return null. 4534const char * 4535SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) { 4536 const TargetLowering *TLI = TM.getTargetLowering(); 4537 SDLoc sdl = getCurSDLoc(); 4538 DebugLoc dl = getCurDebugLoc(); 4539 SDValue Res; 4540 4541 switch (Intrinsic) { 4542 default: 4543 // By default, turn this into a target intrinsic node. 4544 visitTargetIntrinsic(I, Intrinsic); 4545 return 0; 4546 case Intrinsic::vastart: visitVAStart(I); return 0; 4547 case Intrinsic::vaend: visitVAEnd(I); return 0; 4548 case Intrinsic::vacopy: visitVACopy(I); return 0; 4549 case Intrinsic::returnaddress: 4550 setValue(&I, DAG.getNode(ISD::RETURNADDR, sdl, TLI->getPointerTy(), 4551 getValue(I.getArgOperand(0)))); 4552 return 0; 4553 case Intrinsic::frameaddress: 4554 setValue(&I, DAG.getNode(ISD::FRAMEADDR, sdl, TLI->getPointerTy(), 4555 getValue(I.getArgOperand(0)))); 4556 return 0; 4557 case Intrinsic::setjmp: 4558 return &"_setjmp"[!TLI->usesUnderscoreSetJmp()]; 4559 case Intrinsic::longjmp: 4560 return &"_longjmp"[!TLI->usesUnderscoreLongJmp()]; 4561 case Intrinsic::memcpy: { 4562 // Assert for address < 256 since we support only user defined address 4563 // spaces. 4564 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() 4565 < 256 && 4566 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() 4567 < 256 && 4568 "Unknown address space"); 4569 SDValue Op1 = getValue(I.getArgOperand(0)); 4570 SDValue Op2 = getValue(I.getArgOperand(1)); 4571 SDValue Op3 = getValue(I.getArgOperand(2)); 4572 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); 4573 if (!Align) 4574 Align = 1; // @llvm.memcpy defines 0 and 1 to both mean no alignment. 4575 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); 4576 DAG.setRoot(DAG.getMemcpy(getRoot(), sdl, Op1, Op2, Op3, Align, isVol, false, 4577 MachinePointerInfo(I.getArgOperand(0)), 4578 MachinePointerInfo(I.getArgOperand(1)))); 4579 return 0; 4580 } 4581 case Intrinsic::memset: { 4582 // Assert for address < 256 since we support only user defined address 4583 // spaces. 4584 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() 4585 < 256 && 4586 "Unknown address space"); 4587 SDValue Op1 = getValue(I.getArgOperand(0)); 4588 SDValue Op2 = getValue(I.getArgOperand(1)); 4589 SDValue Op3 = getValue(I.getArgOperand(2)); 4590 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); 4591 if (!Align) 4592 Align = 1; // @llvm.memset defines 0 and 1 to both mean no alignment. 4593 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); 4594 DAG.setRoot(DAG.getMemset(getRoot(), sdl, Op1, Op2, Op3, Align, isVol, 4595 MachinePointerInfo(I.getArgOperand(0)))); 4596 return 0; 4597 } 4598 case Intrinsic::memmove: { 4599 // Assert for address < 256 since we support only user defined address 4600 // spaces. 4601 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace() 4602 < 256 && 4603 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace() 4604 < 256 && 4605 "Unknown address space"); 4606 SDValue Op1 = getValue(I.getArgOperand(0)); 4607 SDValue Op2 = getValue(I.getArgOperand(1)); 4608 SDValue Op3 = getValue(I.getArgOperand(2)); 4609 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue(); 4610 if (!Align) 4611 Align = 1; // @llvm.memmove defines 0 and 1 to both mean no alignment. 4612 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue(); 4613 DAG.setRoot(DAG.getMemmove(getRoot(), sdl, Op1, Op2, Op3, Align, isVol, 4614 MachinePointerInfo(I.getArgOperand(0)), 4615 MachinePointerInfo(I.getArgOperand(1)))); 4616 return 0; 4617 } 4618 case Intrinsic::dbg_declare: { 4619 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I); 4620 MDNode *Variable = DI.getVariable(); 4621 const Value *Address = DI.getAddress(); 4622 DIVariable DIVar(Variable); 4623 assert((!DIVar || DIVar.isVariable()) && 4624 "Variable in DbgDeclareInst should be either null or a DIVariable."); 4625 if (!Address || !DIVar) { 4626 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n"); 4627 return 0; 4628 } 4629 4630 // Check if address has undef value. 4631 if (isa<UndefValue>(Address) || 4632 (Address->use_empty() && !isa<Argument>(Address))) { 4633 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n"); 4634 return 0; 4635 } 4636 4637 SDValue &N = NodeMap[Address]; 4638 if (!N.getNode() && isa<Argument>(Address)) 4639 // Check unused arguments map. 4640 N = UnusedArgNodeMap[Address]; 4641 SDDbgValue *SDV; 4642 if (N.getNode()) { 4643 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address)) 4644 Address = BCI->getOperand(0); 4645 // Parameters are handled specially. 4646 bool isParameter = 4647 (DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable || 4648 isa<Argument>(Address)); 4649 4650 const AllocaInst *AI = dyn_cast<AllocaInst>(Address); 4651 4652 if (isParameter && !AI) { 4653 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode()); 4654 if (FINode) 4655 // Byval parameter. We have a frame index at this point. 4656 SDV = DAG.getDbgValue(Variable, FINode->getIndex(), 4657 0, dl, SDNodeOrder); 4658 else { 4659 // Address is an argument, so try to emit its dbg value using 4660 // virtual register info from the FuncInfo.ValueMap. 4661 EmitFuncArgumentDbgValue(Address, Variable, 0, N); 4662 return 0; 4663 } 4664 } else if (AI) 4665 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(), 4666 0, dl, SDNodeOrder); 4667 else { 4668 // Can't do anything with other non-AI cases yet. 4669 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n"); 4670 DEBUG(dbgs() << "non-AllocaInst issue for Address: \n\t"); 4671 DEBUG(Address->dump()); 4672 return 0; 4673 } 4674 DAG.AddDbgValue(SDV, N.getNode(), isParameter); 4675 } else { 4676 // If Address is an argument then try to emit its dbg value using 4677 // virtual register info from the FuncInfo.ValueMap. 4678 if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) { 4679 // If variable is pinned by a alloca in dominating bb then 4680 // use StaticAllocaMap. 4681 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) { 4682 if (AI->getParent() != DI.getParent()) { 4683 DenseMap<const AllocaInst*, int>::iterator SI = 4684 FuncInfo.StaticAllocaMap.find(AI); 4685 if (SI != FuncInfo.StaticAllocaMap.end()) { 4686 SDV = DAG.getDbgValue(Variable, SI->second, 4687 0, dl, SDNodeOrder); 4688 DAG.AddDbgValue(SDV, 0, false); 4689 return 0; 4690 } 4691 } 4692 } 4693 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n"); 4694 } 4695 } 4696 return 0; 4697 } 4698 case Intrinsic::dbg_value: { 4699 const DbgValueInst &DI = cast<DbgValueInst>(I); 4700 DIVariable DIVar(DI.getVariable()); 4701 assert((!DIVar || DIVar.isVariable()) && 4702 "Variable in DbgValueInst should be either null or a DIVariable."); 4703 if (!DIVar) 4704 return 0; 4705 4706 MDNode *Variable = DI.getVariable(); 4707 uint64_t Offset = DI.getOffset(); 4708 const Value *V = DI.getValue(); 4709 if (!V) 4710 return 0; 4711 4712 SDDbgValue *SDV; 4713 if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) { 4714 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder); 4715 DAG.AddDbgValue(SDV, 0, false); 4716 } else { 4717 // Do not use getValue() in here; we don't want to generate code at 4718 // this point if it hasn't been done yet. 4719 SDValue N = NodeMap[V]; 4720 if (!N.getNode() && isa<Argument>(V)) 4721 // Check unused arguments map. 4722 N = UnusedArgNodeMap[V]; 4723 if (N.getNode()) { 4724 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) { 4725 SDV = DAG.getDbgValue(Variable, N.getNode(), 4726 N.getResNo(), Offset, dl, SDNodeOrder); 4727 DAG.AddDbgValue(SDV, N.getNode(), false); 4728 } 4729 } else if (!V->use_empty() ) { 4730 // Do not call getValue(V) yet, as we don't want to generate code. 4731 // Remember it for later. 4732 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder); 4733 DanglingDebugInfoMap[V] = DDI; 4734 } else { 4735 // We may expand this to cover more cases. One case where we have no 4736 // data available is an unreferenced parameter. 4737 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n"); 4738 } 4739 } 4740 4741 // Build a debug info table entry. 4742 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V)) 4743 V = BCI->getOperand(0); 4744 const AllocaInst *AI = dyn_cast<AllocaInst>(V); 4745 // Don't handle byval struct arguments or VLAs, for example. 4746 if (!AI) { 4747 DEBUG(dbgs() << "Dropping debug location info for:\n " << DI << "\n"); 4748 DEBUG(dbgs() << " Last seen at:\n " << *V << "\n"); 4749 return 0; 4750 } 4751 DenseMap<const AllocaInst*, int>::iterator SI = 4752 FuncInfo.StaticAllocaMap.find(AI); 4753 if (SI == FuncInfo.StaticAllocaMap.end()) 4754 return 0; // VLAs. 4755 int FI = SI->second; 4756 4757 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 4758 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo()) 4759 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc()); 4760 return 0; 4761 } 4762 4763 case Intrinsic::eh_typeid_for: { 4764 // Find the type id for the given typeinfo. 4765 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0)); 4766 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV); 4767 Res = DAG.getConstant(TypeID, MVT::i32); 4768 setValue(&I, Res); 4769 return 0; 4770 } 4771 4772 case Intrinsic::eh_return_i32: 4773 case Intrinsic::eh_return_i64: 4774 DAG.getMachineFunction().getMMI().setCallsEHReturn(true); 4775 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, sdl, 4776 MVT::Other, 4777 getControlRoot(), 4778 getValue(I.getArgOperand(0)), 4779 getValue(I.getArgOperand(1)))); 4780 return 0; 4781 case Intrinsic::eh_unwind_init: 4782 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true); 4783 return 0; 4784 case Intrinsic::eh_dwarf_cfa: { 4785 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), sdl, 4786 TLI->getPointerTy()); 4787 SDValue Offset = DAG.getNode(ISD::ADD, sdl, 4788 CfaArg.getValueType(), 4789 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, sdl, 4790 CfaArg.getValueType()), 4791 CfaArg); 4792 SDValue FA = DAG.getNode(ISD::FRAMEADDR, sdl, 4793 TLI->getPointerTy(), 4794 DAG.getConstant(0, TLI->getPointerTy())); 4795 setValue(&I, DAG.getNode(ISD::ADD, sdl, FA.getValueType(), 4796 FA, Offset)); 4797 return 0; 4798 } 4799 case Intrinsic::eh_sjlj_callsite: { 4800 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 4801 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0)); 4802 assert(CI && "Non-constant call site value in eh.sjlj.callsite!"); 4803 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!"); 4804 4805 MMI.setCurrentCallSite(CI->getZExtValue()); 4806 return 0; 4807 } 4808 case Intrinsic::eh_sjlj_functioncontext: { 4809 // Get and store the index of the function context. 4810 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 4811 AllocaInst *FnCtx = 4812 cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts()); 4813 int FI = FuncInfo.StaticAllocaMap[FnCtx]; 4814 MFI->setFunctionContextIndex(FI); 4815 return 0; 4816 } 4817 case Intrinsic::eh_sjlj_setjmp: { 4818 SDValue Ops[2]; 4819 Ops[0] = getRoot(); 4820 Ops[1] = getValue(I.getArgOperand(0)); 4821 SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, sdl, 4822 DAG.getVTList(MVT::i32, MVT::Other), 4823 Ops, 2); 4824 setValue(&I, Op.getValue(0)); 4825 DAG.setRoot(Op.getValue(1)); 4826 return 0; 4827 } 4828 case Intrinsic::eh_sjlj_longjmp: { 4829 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, sdl, MVT::Other, 4830 getRoot(), getValue(I.getArgOperand(0)))); 4831 return 0; 4832 } 4833 4834 case Intrinsic::x86_mmx_pslli_w: 4835 case Intrinsic::x86_mmx_pslli_d: 4836 case Intrinsic::x86_mmx_pslli_q: 4837 case Intrinsic::x86_mmx_psrli_w: 4838 case Intrinsic::x86_mmx_psrli_d: 4839 case Intrinsic::x86_mmx_psrli_q: 4840 case Intrinsic::x86_mmx_psrai_w: 4841 case Intrinsic::x86_mmx_psrai_d: { 4842 SDValue ShAmt = getValue(I.getArgOperand(1)); 4843 if (isa<ConstantSDNode>(ShAmt)) { 4844 visitTargetIntrinsic(I, Intrinsic); 4845 return 0; 4846 } 4847 unsigned NewIntrinsic = 0; 4848 EVT ShAmtVT = MVT::v2i32; 4849 switch (Intrinsic) { 4850 case Intrinsic::x86_mmx_pslli_w: 4851 NewIntrinsic = Intrinsic::x86_mmx_psll_w; 4852 break; 4853 case Intrinsic::x86_mmx_pslli_d: 4854 NewIntrinsic = Intrinsic::x86_mmx_psll_d; 4855 break; 4856 case Intrinsic::x86_mmx_pslli_q: 4857 NewIntrinsic = Intrinsic::x86_mmx_psll_q; 4858 break; 4859 case Intrinsic::x86_mmx_psrli_w: 4860 NewIntrinsic = Intrinsic::x86_mmx_psrl_w; 4861 break; 4862 case Intrinsic::x86_mmx_psrli_d: 4863 NewIntrinsic = Intrinsic::x86_mmx_psrl_d; 4864 break; 4865 case Intrinsic::x86_mmx_psrli_q: 4866 NewIntrinsic = Intrinsic::x86_mmx_psrl_q; 4867 break; 4868 case Intrinsic::x86_mmx_psrai_w: 4869 NewIntrinsic = Intrinsic::x86_mmx_psra_w; 4870 break; 4871 case Intrinsic::x86_mmx_psrai_d: 4872 NewIntrinsic = Intrinsic::x86_mmx_psra_d; 4873 break; 4874 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here. 4875 } 4876 4877 // The vector shift intrinsics with scalars uses 32b shift amounts but 4878 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits 4879 // to be zero. 4880 // We must do this early because v2i32 is not a legal type. 4881 SDValue ShOps[2]; 4882 ShOps[0] = ShAmt; 4883 ShOps[1] = DAG.getConstant(0, MVT::i32); 4884 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, sdl, ShAmtVT, &ShOps[0], 2); 4885 EVT DestVT = TLI->getValueType(I.getType()); 4886 ShAmt = DAG.getNode(ISD::BITCAST, sdl, DestVT, ShAmt); 4887 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, sdl, DestVT, 4888 DAG.getConstant(NewIntrinsic, MVT::i32), 4889 getValue(I.getArgOperand(0)), ShAmt); 4890 setValue(&I, Res); 4891 return 0; 4892 } 4893 case Intrinsic::x86_avx_vinsertf128_pd_256: 4894 case Intrinsic::x86_avx_vinsertf128_ps_256: 4895 case Intrinsic::x86_avx_vinsertf128_si_256: 4896 case Intrinsic::x86_avx2_vinserti128: { 4897 EVT DestVT = TLI->getValueType(I.getType()); 4898 EVT ElVT = TLI->getValueType(I.getArgOperand(1)->getType()); 4899 uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(2))->getZExtValue() & 1) * 4900 ElVT.getVectorNumElements(); 4901 Res = DAG.getNode(ISD::INSERT_SUBVECTOR, sdl, DestVT, 4902 getValue(I.getArgOperand(0)), 4903 getValue(I.getArgOperand(1)), 4904 DAG.getConstant(Idx, TLI->getVectorIdxTy())); 4905 setValue(&I, Res); 4906 return 0; 4907 } 4908 case Intrinsic::x86_avx_vextractf128_pd_256: 4909 case Intrinsic::x86_avx_vextractf128_ps_256: 4910 case Intrinsic::x86_avx_vextractf128_si_256: 4911 case Intrinsic::x86_avx2_vextracti128: { 4912 EVT DestVT = TLI->getValueType(I.getType()); 4913 uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(1))->getZExtValue() & 1) * 4914 DestVT.getVectorNumElements(); 4915 Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, sdl, DestVT, 4916 getValue(I.getArgOperand(0)), 4917 DAG.getConstant(Idx, TLI->getVectorIdxTy())); 4918 setValue(&I, Res); 4919 return 0; 4920 } 4921 case Intrinsic::convertff: 4922 case Intrinsic::convertfsi: 4923 case Intrinsic::convertfui: 4924 case Intrinsic::convertsif: 4925 case Intrinsic::convertuif: 4926 case Intrinsic::convertss: 4927 case Intrinsic::convertsu: 4928 case Intrinsic::convertus: 4929 case Intrinsic::convertuu: { 4930 ISD::CvtCode Code = ISD::CVT_INVALID; 4931 switch (Intrinsic) { 4932 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here. 4933 case Intrinsic::convertff: Code = ISD::CVT_FF; break; 4934 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break; 4935 case Intrinsic::convertfui: Code = ISD::CVT_FU; break; 4936 case Intrinsic::convertsif: Code = ISD::CVT_SF; break; 4937 case Intrinsic::convertuif: Code = ISD::CVT_UF; break; 4938 case Intrinsic::convertss: Code = ISD::CVT_SS; break; 4939 case Intrinsic::convertsu: Code = ISD::CVT_SU; break; 4940 case Intrinsic::convertus: Code = ISD::CVT_US; break; 4941 case Intrinsic::convertuu: Code = ISD::CVT_UU; break; 4942 } 4943 EVT DestVT = TLI->getValueType(I.getType()); 4944 const Value *Op1 = I.getArgOperand(0); 4945 Res = DAG.getConvertRndSat(DestVT, sdl, getValue(Op1), 4946 DAG.getValueType(DestVT), 4947 DAG.getValueType(getValue(Op1).getValueType()), 4948 getValue(I.getArgOperand(1)), 4949 getValue(I.getArgOperand(2)), 4950 Code); 4951 setValue(&I, Res); 4952 return 0; 4953 } 4954 case Intrinsic::powi: 4955 setValue(&I, ExpandPowI(sdl, getValue(I.getArgOperand(0)), 4956 getValue(I.getArgOperand(1)), DAG)); 4957 return 0; 4958 case Intrinsic::log: 4959 setValue(&I, expandLog(sdl, getValue(I.getArgOperand(0)), DAG, *TLI)); 4960 return 0; 4961 case Intrinsic::log2: 4962 setValue(&I, expandLog2(sdl, getValue(I.getArgOperand(0)), DAG, *TLI)); 4963 return 0; 4964 case Intrinsic::log10: 4965 setValue(&I, expandLog10(sdl, getValue(I.getArgOperand(0)), DAG, *TLI)); 4966 return 0; 4967 case Intrinsic::exp: 4968 setValue(&I, expandExp(sdl, getValue(I.getArgOperand(0)), DAG, *TLI)); 4969 return 0; 4970 case Intrinsic::exp2: 4971 setValue(&I, expandExp2(sdl, getValue(I.getArgOperand(0)), DAG, *TLI)); 4972 return 0; 4973 case Intrinsic::pow: 4974 setValue(&I, expandPow(sdl, getValue(I.getArgOperand(0)), 4975 getValue(I.getArgOperand(1)), DAG, *TLI)); 4976 return 0; 4977 case Intrinsic::sqrt: 4978 case Intrinsic::fabs: 4979 case Intrinsic::sin: 4980 case Intrinsic::cos: 4981 case Intrinsic::floor: 4982 case Intrinsic::ceil: 4983 case Intrinsic::trunc: 4984 case Intrinsic::rint: 4985 case Intrinsic::nearbyint: 4986 case Intrinsic::round: { 4987 unsigned Opcode; 4988 switch (Intrinsic) { 4989 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here. 4990 case Intrinsic::sqrt: Opcode = ISD::FSQRT; break; 4991 case Intrinsic::fabs: Opcode = ISD::FABS; break; 4992 case Intrinsic::sin: Opcode = ISD::FSIN; break; 4993 case Intrinsic::cos: Opcode = ISD::FCOS; break; 4994 case Intrinsic::floor: Opcode = ISD::FFLOOR; break; 4995 case Intrinsic::ceil: Opcode = ISD::FCEIL; break; 4996 case Intrinsic::trunc: Opcode = ISD::FTRUNC; break; 4997 case Intrinsic::rint: Opcode = ISD::FRINT; break; 4998 case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break; 4999 case Intrinsic::round: Opcode = ISD::FROUND; break; 5000 } 5001 5002 setValue(&I, DAG.getNode(Opcode, sdl, 5003 getValue(I.getArgOperand(0)).getValueType(), 5004 getValue(I.getArgOperand(0)))); 5005 return 0; 5006 } 5007 case Intrinsic::copysign: 5008 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, sdl, 5009 getValue(I.getArgOperand(0)).getValueType(), 5010 getValue(I.getArgOperand(0)), 5011 getValue(I.getArgOperand(1)))); 5012 return 0; 5013 case Intrinsic::fma: 5014 setValue(&I, DAG.getNode(ISD::FMA, sdl, 5015 getValue(I.getArgOperand(0)).getValueType(), 5016 getValue(I.getArgOperand(0)), 5017 getValue(I.getArgOperand(1)), 5018 getValue(I.getArgOperand(2)))); 5019 return 0; 5020 case Intrinsic::fmuladd: { 5021 EVT VT = TLI->getValueType(I.getType()); 5022 if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict && 5023 TLI->isFMAFasterThanFMulAndFAdd(VT)) { 5024 setValue(&I, DAG.getNode(ISD::FMA, sdl, 5025 getValue(I.getArgOperand(0)).getValueType(), 5026 getValue(I.getArgOperand(0)), 5027 getValue(I.getArgOperand(1)), 5028 getValue(I.getArgOperand(2)))); 5029 } else { 5030 SDValue Mul = DAG.getNode(ISD::FMUL, sdl, 5031 getValue(I.getArgOperand(0)).getValueType(), 5032 getValue(I.getArgOperand(0)), 5033 getValue(I.getArgOperand(1))); 5034 SDValue Add = DAG.getNode(ISD::FADD, sdl, 5035 getValue(I.getArgOperand(0)).getValueType(), 5036 Mul, 5037 getValue(I.getArgOperand(2))); 5038 setValue(&I, Add); 5039 } 5040 return 0; 5041 } 5042 case Intrinsic::convert_to_fp16: 5043 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, sdl, 5044 MVT::i16, getValue(I.getArgOperand(0)))); 5045 return 0; 5046 case Intrinsic::convert_from_fp16: 5047 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, sdl, 5048 MVT::f32, getValue(I.getArgOperand(0)))); 5049 return 0; 5050 case Intrinsic::pcmarker: { 5051 SDValue Tmp = getValue(I.getArgOperand(0)); 5052 DAG.setRoot(DAG.getNode(ISD::PCMARKER, sdl, MVT::Other, getRoot(), Tmp)); 5053 return 0; 5054 } 5055 case Intrinsic::readcyclecounter: { 5056 SDValue Op = getRoot(); 5057 Res = DAG.getNode(ISD::READCYCLECOUNTER, sdl, 5058 DAG.getVTList(MVT::i64, MVT::Other), 5059 &Op, 1); 5060 setValue(&I, Res); 5061 DAG.setRoot(Res.getValue(1)); 5062 return 0; 5063 } 5064 case Intrinsic::bswap: 5065 setValue(&I, DAG.getNode(ISD::BSWAP, sdl, 5066 getValue(I.getArgOperand(0)).getValueType(), 5067 getValue(I.getArgOperand(0)))); 5068 return 0; 5069 case Intrinsic::cttz: { 5070 SDValue Arg = getValue(I.getArgOperand(0)); 5071 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1)); 5072 EVT Ty = Arg.getValueType(); 5073 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF, 5074 sdl, Ty, Arg)); 5075 return 0; 5076 } 5077 case Intrinsic::ctlz: { 5078 SDValue Arg = getValue(I.getArgOperand(0)); 5079 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1)); 5080 EVT Ty = Arg.getValueType(); 5081 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF, 5082 sdl, Ty, Arg)); 5083 return 0; 5084 } 5085 case Intrinsic::ctpop: { 5086 SDValue Arg = getValue(I.getArgOperand(0)); 5087 EVT Ty = Arg.getValueType(); 5088 setValue(&I, DAG.getNode(ISD::CTPOP, sdl, Ty, Arg)); 5089 return 0; 5090 } 5091 case Intrinsic::stacksave: { 5092 SDValue Op = getRoot(); 5093 Res = DAG.getNode(ISD::STACKSAVE, sdl, 5094 DAG.getVTList(TLI->getPointerTy(), MVT::Other), &Op, 1); 5095 setValue(&I, Res); 5096 DAG.setRoot(Res.getValue(1)); 5097 return 0; 5098 } 5099 case Intrinsic::stackrestore: { 5100 Res = getValue(I.getArgOperand(0)); 5101 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, sdl, MVT::Other, getRoot(), Res)); 5102 return 0; 5103 } 5104 case Intrinsic::stackprotector: { 5105 // Emit code into the DAG to store the stack guard onto the stack. 5106 MachineFunction &MF = DAG.getMachineFunction(); 5107 MachineFrameInfo *MFI = MF.getFrameInfo(); 5108 EVT PtrTy = TLI->getPointerTy(); 5109 5110 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value. 5111 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1)); 5112 5113 int FI = FuncInfo.StaticAllocaMap[Slot]; 5114 MFI->setStackProtectorIndex(FI); 5115 5116 SDValue FIN = DAG.getFrameIndex(FI, PtrTy); 5117 5118 // Store the stack protector onto the stack. 5119 Res = DAG.getStore(getRoot(), sdl, Src, FIN, 5120 MachinePointerInfo::getFixedStack(FI), 5121 true, false, 0); 5122 setValue(&I, Res); 5123 DAG.setRoot(Res); 5124 return 0; 5125 } 5126 case Intrinsic::objectsize: { 5127 // If we don't know by now, we're never going to know. 5128 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1)); 5129 5130 assert(CI && "Non-constant type in __builtin_object_size?"); 5131 5132 SDValue Arg = getValue(I.getCalledValue()); 5133 EVT Ty = Arg.getValueType(); 5134 5135 if (CI->isZero()) 5136 Res = DAG.getConstant(-1ULL, Ty); 5137 else 5138 Res = DAG.getConstant(0, Ty); 5139 5140 setValue(&I, Res); 5141 return 0; 5142 } 5143 case Intrinsic::annotation: 5144 case Intrinsic::ptr_annotation: 5145 // Drop the intrinsic, but forward the value 5146 setValue(&I, getValue(I.getOperand(0))); 5147 return 0; 5148 case Intrinsic::var_annotation: 5149 // Discard annotate attributes 5150 return 0; 5151 5152 case Intrinsic::init_trampoline: { 5153 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts()); 5154 5155 SDValue Ops[6]; 5156 Ops[0] = getRoot(); 5157 Ops[1] = getValue(I.getArgOperand(0)); 5158 Ops[2] = getValue(I.getArgOperand(1)); 5159 Ops[3] = getValue(I.getArgOperand(2)); 5160 Ops[4] = DAG.getSrcValue(I.getArgOperand(0)); 5161 Ops[5] = DAG.getSrcValue(F); 5162 5163 Res = DAG.getNode(ISD::INIT_TRAMPOLINE, sdl, MVT::Other, Ops, 6); 5164 5165 DAG.setRoot(Res); 5166 return 0; 5167 } 5168 case Intrinsic::adjust_trampoline: { 5169 setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, sdl, 5170 TLI->getPointerTy(), 5171 getValue(I.getArgOperand(0)))); 5172 return 0; 5173 } 5174 case Intrinsic::gcroot: 5175 if (GFI) { 5176 const Value *Alloca = I.getArgOperand(0)->stripPointerCasts(); 5177 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1)); 5178 5179 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode()); 5180 GFI->addStackRoot(FI->getIndex(), TypeMap); 5181 } 5182 return 0; 5183 case Intrinsic::gcread: 5184 case Intrinsic::gcwrite: 5185 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!"); 5186 case Intrinsic::flt_rounds: 5187 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, sdl, MVT::i32)); 5188 return 0; 5189 5190 case Intrinsic::expect: { 5191 // Just replace __builtin_expect(exp, c) with EXP. 5192 setValue(&I, getValue(I.getArgOperand(0))); 5193 return 0; 5194 } 5195 5196 case Intrinsic::debugtrap: 5197 case Intrinsic::trap: { 5198 StringRef TrapFuncName = TM.Options.getTrapFunctionName(); 5199 if (TrapFuncName.empty()) { 5200 ISD::NodeType Op = (Intrinsic == Intrinsic::trap) ? 5201 ISD::TRAP : ISD::DEBUGTRAP; 5202 DAG.setRoot(DAG.getNode(Op, sdl,MVT::Other, getRoot())); 5203 return 0; 5204 } 5205 TargetLowering::ArgListTy Args; 5206 TargetLowering:: 5207 CallLoweringInfo CLI(getRoot(), I.getType(), 5208 false, false, false, false, 0, CallingConv::C, 5209 /*isTailCall=*/false, 5210 /*doesNotRet=*/false, /*isReturnValueUsed=*/true, 5211 DAG.getExternalSymbol(TrapFuncName.data(), 5212 TLI->getPointerTy()), 5213 Args, DAG, sdl); 5214 std::pair<SDValue, SDValue> Result = TLI->LowerCallTo(CLI); 5215 DAG.setRoot(Result.second); 5216 return 0; 5217 } 5218 5219 case Intrinsic::uadd_with_overflow: 5220 case Intrinsic::sadd_with_overflow: 5221 case Intrinsic::usub_with_overflow: 5222 case Intrinsic::ssub_with_overflow: 5223 case Intrinsic::umul_with_overflow: 5224 case Intrinsic::smul_with_overflow: { 5225 ISD::NodeType Op; 5226 switch (Intrinsic) { 5227 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here. 5228 case Intrinsic::uadd_with_overflow: Op = ISD::UADDO; break; 5229 case Intrinsic::sadd_with_overflow: Op = ISD::SADDO; break; 5230 case Intrinsic::usub_with_overflow: Op = ISD::USUBO; break; 5231 case Intrinsic::ssub_with_overflow: Op = ISD::SSUBO; break; 5232 case Intrinsic::umul_with_overflow: Op = ISD::UMULO; break; 5233 case Intrinsic::smul_with_overflow: Op = ISD::SMULO; break; 5234 } 5235 SDValue Op1 = getValue(I.getArgOperand(0)); 5236 SDValue Op2 = getValue(I.getArgOperand(1)); 5237 5238 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1); 5239 setValue(&I, DAG.getNode(Op, sdl, VTs, Op1, Op2)); 5240 return 0; 5241 } 5242 case Intrinsic::prefetch: { 5243 SDValue Ops[5]; 5244 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue(); 5245 Ops[0] = getRoot(); 5246 Ops[1] = getValue(I.getArgOperand(0)); 5247 Ops[2] = getValue(I.getArgOperand(1)); 5248 Ops[3] = getValue(I.getArgOperand(2)); 5249 Ops[4] = getValue(I.getArgOperand(3)); 5250 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, sdl, 5251 DAG.getVTList(MVT::Other), 5252 &Ops[0], 5, 5253 EVT::getIntegerVT(*Context, 8), 5254 MachinePointerInfo(I.getArgOperand(0)), 5255 0, /* align */ 5256 false, /* volatile */ 5257 rw==0, /* read */ 5258 rw==1)); /* write */ 5259 return 0; 5260 } 5261 case Intrinsic::lifetime_start: 5262 case Intrinsic::lifetime_end: { 5263 bool IsStart = (Intrinsic == Intrinsic::lifetime_start); 5264 // Stack coloring is not enabled in O0, discard region information. 5265 if (TM.getOptLevel() == CodeGenOpt::None) 5266 return 0; 5267 5268 SmallVector<Value *, 4> Allocas; 5269 GetUnderlyingObjects(I.getArgOperand(1), Allocas, TD); 5270 5271 for (SmallVectorImpl<Value*>::iterator Object = Allocas.begin(), 5272 E = Allocas.end(); Object != E; ++Object) { 5273 AllocaInst *LifetimeObject = dyn_cast_or_null<AllocaInst>(*Object); 5274 5275 // Could not find an Alloca. 5276 if (!LifetimeObject) 5277 continue; 5278 5279 int FI = FuncInfo.StaticAllocaMap[LifetimeObject]; 5280 5281 SDValue Ops[2]; 5282 Ops[0] = getRoot(); 5283 Ops[1] = DAG.getFrameIndex(FI, TLI->getPointerTy(), true); 5284 unsigned Opcode = (IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END); 5285 5286 Res = DAG.getNode(Opcode, sdl, MVT::Other, Ops, 2); 5287 DAG.setRoot(Res); 5288 } 5289 return 0; 5290 } 5291 case Intrinsic::invariant_start: 5292 // Discard region information. 5293 setValue(&I, DAG.getUNDEF(TLI->getPointerTy())); 5294 return 0; 5295 case Intrinsic::invariant_end: 5296 // Discard region information. 5297 return 0; 5298 case Intrinsic::stackprotectorcheck: { 5299 // Do not actually emit anything for this basic block. Instead we initialize 5300 // the stack protector descriptor and export the guard variable so we can 5301 // access it in FinishBasicBlock. 5302 const BasicBlock *BB = I.getParent(); 5303 SPDescriptor.initialize(BB, FuncInfo.MBBMap[BB], I); 5304 ExportFromCurrentBlock(SPDescriptor.getGuard()); 5305 5306 // Flush our exports since we are going to process a terminator. 5307 (void)getControlRoot(); 5308 return 0; 5309 } 5310 case Intrinsic::donothing: 5311 // ignore 5312 return 0; 5313 case Intrinsic::experimental_stackmap: { 5314 visitStackmap(I); 5315 return 0; 5316 } 5317 case Intrinsic::experimental_patchpoint_void: 5318 case Intrinsic::experimental_patchpoint_i64: { 5319 visitPatchpoint(I); 5320 return 0; 5321 } 5322 } 5323} 5324 5325void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee, 5326 bool isTailCall, 5327 MachineBasicBlock *LandingPad) { 5328 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType()); 5329 FunctionType *FTy = cast<FunctionType>(PT->getElementType()); 5330 Type *RetTy = FTy->getReturnType(); 5331 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 5332 MCSymbol *BeginLabel = 0; 5333 5334 TargetLowering::ArgListTy Args; 5335 TargetLowering::ArgListEntry Entry; 5336 Args.reserve(CS.arg_size()); 5337 5338 // Check whether the function can return without sret-demotion. 5339 SmallVector<ISD::OutputArg, 4> Outs; 5340 const TargetLowering *TLI = TM.getTargetLowering(); 5341 GetReturnInfo(RetTy, CS.getAttributes(), Outs, *TLI); 5342 5343 bool CanLowerReturn = TLI->CanLowerReturn(CS.getCallingConv(), 5344 DAG.getMachineFunction(), 5345 FTy->isVarArg(), Outs, 5346 FTy->getContext()); 5347 5348 SDValue DemoteStackSlot; 5349 int DemoteStackIdx = -100; 5350 5351 if (!CanLowerReturn) { 5352 uint64_t TySize = TLI->getDataLayout()->getTypeAllocSize( 5353 FTy->getReturnType()); 5354 unsigned Align = TLI->getDataLayout()->getPrefTypeAlignment( 5355 FTy->getReturnType()); 5356 MachineFunction &MF = DAG.getMachineFunction(); 5357 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false); 5358 Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType()); 5359 5360 DemoteStackSlot = DAG.getFrameIndex(DemoteStackIdx, TLI->getPointerTy()); 5361 Entry.Node = DemoteStackSlot; 5362 Entry.Ty = StackSlotPtrType; 5363 Entry.isSExt = false; 5364 Entry.isZExt = false; 5365 Entry.isInReg = false; 5366 Entry.isSRet = true; 5367 Entry.isNest = false; 5368 Entry.isByVal = false; 5369 Entry.isReturned = false; 5370 Entry.Alignment = Align; 5371 Args.push_back(Entry); 5372 RetTy = Type::getVoidTy(FTy->getContext()); 5373 } 5374 5375 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end(); 5376 i != e; ++i) { 5377 const Value *V = *i; 5378 5379 // Skip empty types 5380 if (V->getType()->isEmptyTy()) 5381 continue; 5382 5383 SDValue ArgNode = getValue(V); 5384 Entry.Node = ArgNode; Entry.Ty = V->getType(); 5385 5386 // Skip the first return-type Attribute to get to params. 5387 Entry.setAttributes(&CS, i - CS.arg_begin() + 1); 5388 Args.push_back(Entry); 5389 } 5390 5391 if (LandingPad) { 5392 // Insert a label before the invoke call to mark the try range. This can be 5393 // used to detect deletion of the invoke via the MachineModuleInfo. 5394 BeginLabel = MMI.getContext().CreateTempSymbol(); 5395 5396 // For SjLj, keep track of which landing pads go with which invokes 5397 // so as to maintain the ordering of pads in the LSDA. 5398 unsigned CallSiteIndex = MMI.getCurrentCallSite(); 5399 if (CallSiteIndex) { 5400 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex); 5401 LPadToCallSiteMap[LandingPad].push_back(CallSiteIndex); 5402 5403 // Now that the call site is handled, stop tracking it. 5404 MMI.setCurrentCallSite(0); 5405 } 5406 5407 // Both PendingLoads and PendingExports must be flushed here; 5408 // this call might not return. 5409 (void)getRoot(); 5410 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getControlRoot(), BeginLabel)); 5411 } 5412 5413 // Check if target-independent constraints permit a tail call here. 5414 // Target-dependent constraints are checked within TLI->LowerCallTo. 5415 if (isTailCall && !isInTailCallPosition(CS, *TLI)) 5416 isTailCall = false; 5417 5418 TargetLowering:: 5419 CallLoweringInfo CLI(getRoot(), RetTy, FTy, isTailCall, Callee, Args, DAG, 5420 getCurSDLoc(), CS); 5421 std::pair<SDValue,SDValue> Result = TLI->LowerCallTo(CLI); 5422 assert((isTailCall || Result.second.getNode()) && 5423 "Non-null chain expected with non-tail call!"); 5424 assert((Result.second.getNode() || !Result.first.getNode()) && 5425 "Null value expected with tail call!"); 5426 if (Result.first.getNode()) { 5427 setValue(CS.getInstruction(), Result.first); 5428 } else if (!CanLowerReturn && Result.second.getNode()) { 5429 // The instruction result is the result of loading from the 5430 // hidden sret parameter. 5431 SmallVector<EVT, 1> PVTs; 5432 Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType()); 5433 5434 ComputeValueVTs(*TLI, PtrRetTy, PVTs); 5435 assert(PVTs.size() == 1 && "Pointers should fit in one register"); 5436 EVT PtrVT = PVTs[0]; 5437 5438 SmallVector<EVT, 4> RetTys; 5439 SmallVector<uint64_t, 4> Offsets; 5440 RetTy = FTy->getReturnType(); 5441 ComputeValueVTs(*TLI, RetTy, RetTys, &Offsets); 5442 5443 unsigned NumValues = RetTys.size(); 5444 SmallVector<SDValue, 4> Values(NumValues); 5445 SmallVector<SDValue, 4> Chains(NumValues); 5446 5447 for (unsigned i = 0; i < NumValues; ++i) { 5448 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(), PtrVT, 5449 DemoteStackSlot, 5450 DAG.getConstant(Offsets[i], PtrVT)); 5451 SDValue L = DAG.getLoad(RetTys[i], getCurSDLoc(), Result.second, Add, 5452 MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]), 5453 false, false, false, 1); 5454 Values[i] = L; 5455 Chains[i] = L.getValue(1); 5456 } 5457 5458 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), 5459 MVT::Other, &Chains[0], NumValues); 5460 PendingLoads.push_back(Chain); 5461 5462 setValue(CS.getInstruction(), 5463 DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(), 5464 DAG.getVTList(&RetTys[0], RetTys.size()), 5465 &Values[0], Values.size())); 5466 } 5467 5468 if (!Result.second.getNode()) { 5469 // As a special case, a null chain means that a tail call has been emitted 5470 // and the DAG root is already updated. 5471 HasTailCall = true; 5472 5473 // Since there's no actual continuation from this block, nothing can be 5474 // relying on us setting vregs for them. 5475 PendingExports.clear(); 5476 } else { 5477 DAG.setRoot(Result.second); 5478 } 5479 5480 if (LandingPad) { 5481 // Insert a label at the end of the invoke call to mark the try range. This 5482 // can be used to detect deletion of the invoke via the MachineModuleInfo. 5483 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol(); 5484 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getRoot(), EndLabel)); 5485 5486 // Inform MachineModuleInfo of range. 5487 MMI.addInvoke(LandingPad, BeginLabel, EndLabel); 5488 } 5489} 5490 5491/// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the 5492/// value is equal or not-equal to zero. 5493static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) { 5494 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); 5495 UI != E; ++UI) { 5496 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI)) 5497 if (IC->isEquality()) 5498 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1))) 5499 if (C->isNullValue()) 5500 continue; 5501 // Unknown instruction. 5502 return false; 5503 } 5504 return true; 5505} 5506 5507static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT, 5508 Type *LoadTy, 5509 SelectionDAGBuilder &Builder) { 5510 5511 // Check to see if this load can be trivially constant folded, e.g. if the 5512 // input is from a string literal. 5513 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) { 5514 // Cast pointer to the type we really want to load. 5515 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput), 5516 PointerType::getUnqual(LoadTy)); 5517 5518 if (const Constant *LoadCst = 5519 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput), 5520 Builder.TD)) 5521 return Builder.getValue(LoadCst); 5522 } 5523 5524 // Otherwise, we have to emit the load. If the pointer is to unfoldable but 5525 // still constant memory, the input chain can be the entry node. 5526 SDValue Root; 5527 bool ConstantMemory = false; 5528 5529 // Do not serialize (non-volatile) loads of constant memory with anything. 5530 if (Builder.AA->pointsToConstantMemory(PtrVal)) { 5531 Root = Builder.DAG.getEntryNode(); 5532 ConstantMemory = true; 5533 } else { 5534 // Do not serialize non-volatile loads against each other. 5535 Root = Builder.DAG.getRoot(); 5536 } 5537 5538 SDValue Ptr = Builder.getValue(PtrVal); 5539 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurSDLoc(), Root, 5540 Ptr, MachinePointerInfo(PtrVal), 5541 false /*volatile*/, 5542 false /*nontemporal*/, 5543 false /*isinvariant*/, 1 /* align=1 */); 5544 5545 if (!ConstantMemory) 5546 Builder.PendingLoads.push_back(LoadVal.getValue(1)); 5547 return LoadVal; 5548} 5549 5550/// processIntegerCallValue - Record the value for an instruction that 5551/// produces an integer result, converting the type where necessary. 5552void SelectionDAGBuilder::processIntegerCallValue(const Instruction &I, 5553 SDValue Value, 5554 bool IsSigned) { 5555 EVT VT = TM.getTargetLowering()->getValueType(I.getType(), true); 5556 if (IsSigned) 5557 Value = DAG.getSExtOrTrunc(Value, getCurSDLoc(), VT); 5558 else 5559 Value = DAG.getZExtOrTrunc(Value, getCurSDLoc(), VT); 5560 setValue(&I, Value); 5561} 5562 5563/// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form. 5564/// If so, return true and lower it, otherwise return false and it will be 5565/// lowered like a normal call. 5566bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) { 5567 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t) 5568 if (I.getNumArgOperands() != 3) 5569 return false; 5570 5571 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1); 5572 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() || 5573 !I.getArgOperand(2)->getType()->isIntegerTy() || 5574 !I.getType()->isIntegerTy()) 5575 return false; 5576 5577 const Value *Size = I.getArgOperand(2); 5578 const ConstantInt *CSize = dyn_cast<ConstantInt>(Size); 5579 if (CSize && CSize->getZExtValue() == 0) { 5580 EVT CallVT = TM.getTargetLowering()->getValueType(I.getType(), true); 5581 setValue(&I, DAG.getConstant(0, CallVT)); 5582 return true; 5583 } 5584 5585 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); 5586 std::pair<SDValue, SDValue> Res = 5587 TSI.EmitTargetCodeForMemcmp(DAG, getCurSDLoc(), DAG.getRoot(), 5588 getValue(LHS), getValue(RHS), getValue(Size), 5589 MachinePointerInfo(LHS), 5590 MachinePointerInfo(RHS)); 5591 if (Res.first.getNode()) { 5592 processIntegerCallValue(I, Res.first, true); 5593 PendingLoads.push_back(Res.second); 5594 return true; 5595 } 5596 5597 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0 5598 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0 5599 if (CSize && IsOnlyUsedInZeroEqualityComparison(&I)) { 5600 bool ActuallyDoIt = true; 5601 MVT LoadVT; 5602 Type *LoadTy; 5603 switch (CSize->getZExtValue()) { 5604 default: 5605 LoadVT = MVT::Other; 5606 LoadTy = 0; 5607 ActuallyDoIt = false; 5608 break; 5609 case 2: 5610 LoadVT = MVT::i16; 5611 LoadTy = Type::getInt16Ty(CSize->getContext()); 5612 break; 5613 case 4: 5614 LoadVT = MVT::i32; 5615 LoadTy = Type::getInt32Ty(CSize->getContext()); 5616 break; 5617 case 8: 5618 LoadVT = MVT::i64; 5619 LoadTy = Type::getInt64Ty(CSize->getContext()); 5620 break; 5621 /* 5622 case 16: 5623 LoadVT = MVT::v4i32; 5624 LoadTy = Type::getInt32Ty(CSize->getContext()); 5625 LoadTy = VectorType::get(LoadTy, 4); 5626 break; 5627 */ 5628 } 5629 5630 // This turns into unaligned loads. We only do this if the target natively 5631 // supports the MVT we'll be loading or if it is small enough (<= 4) that 5632 // we'll only produce a small number of byte loads. 5633 5634 // Require that we can find a legal MVT, and only do this if the target 5635 // supports unaligned loads of that type. Expanding into byte loads would 5636 // bloat the code. 5637 const TargetLowering *TLI = TM.getTargetLowering(); 5638 if (ActuallyDoIt && CSize->getZExtValue() > 4) { 5639 // TODO: Handle 5 byte compare as 4-byte + 1 byte. 5640 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads. 5641 if (!TLI->isTypeLegal(LoadVT) ||!TLI->allowsUnalignedMemoryAccesses(LoadVT)) 5642 ActuallyDoIt = false; 5643 } 5644 5645 if (ActuallyDoIt) { 5646 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this); 5647 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this); 5648 5649 SDValue Res = DAG.getSetCC(getCurSDLoc(), MVT::i1, LHSVal, RHSVal, 5650 ISD::SETNE); 5651 processIntegerCallValue(I, Res, false); 5652 return true; 5653 } 5654 } 5655 5656 5657 return false; 5658} 5659 5660/// visitMemChrCall -- See if we can lower a memchr call into an optimized 5661/// form. If so, return true and lower it, otherwise return false and it 5662/// will be lowered like a normal call. 5663bool SelectionDAGBuilder::visitMemChrCall(const CallInst &I) { 5664 // Verify that the prototype makes sense. void *memchr(void *, int, size_t) 5665 if (I.getNumArgOperands() != 3) 5666 return false; 5667 5668 const Value *Src = I.getArgOperand(0); 5669 const Value *Char = I.getArgOperand(1); 5670 const Value *Length = I.getArgOperand(2); 5671 if (!Src->getType()->isPointerTy() || 5672 !Char->getType()->isIntegerTy() || 5673 !Length->getType()->isIntegerTy() || 5674 !I.getType()->isPointerTy()) 5675 return false; 5676 5677 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); 5678 std::pair<SDValue, SDValue> Res = 5679 TSI.EmitTargetCodeForMemchr(DAG, getCurSDLoc(), DAG.getRoot(), 5680 getValue(Src), getValue(Char), getValue(Length), 5681 MachinePointerInfo(Src)); 5682 if (Res.first.getNode()) { 5683 setValue(&I, Res.first); 5684 PendingLoads.push_back(Res.second); 5685 return true; 5686 } 5687 5688 return false; 5689} 5690 5691/// visitStrCpyCall -- See if we can lower a strcpy or stpcpy call into an 5692/// optimized form. If so, return true and lower it, otherwise return false 5693/// and it will be lowered like a normal call. 5694bool SelectionDAGBuilder::visitStrCpyCall(const CallInst &I, bool isStpcpy) { 5695 // Verify that the prototype makes sense. char *strcpy(char *, char *) 5696 if (I.getNumArgOperands() != 2) 5697 return false; 5698 5699 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1); 5700 if (!Arg0->getType()->isPointerTy() || 5701 !Arg1->getType()->isPointerTy() || 5702 !I.getType()->isPointerTy()) 5703 return false; 5704 5705 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); 5706 std::pair<SDValue, SDValue> Res = 5707 TSI.EmitTargetCodeForStrcpy(DAG, getCurSDLoc(), getRoot(), 5708 getValue(Arg0), getValue(Arg1), 5709 MachinePointerInfo(Arg0), 5710 MachinePointerInfo(Arg1), isStpcpy); 5711 if (Res.first.getNode()) { 5712 setValue(&I, Res.first); 5713 DAG.setRoot(Res.second); 5714 return true; 5715 } 5716 5717 return false; 5718} 5719 5720/// visitStrCmpCall - See if we can lower a call to strcmp in an optimized form. 5721/// If so, return true and lower it, otherwise return false and it will be 5722/// lowered like a normal call. 5723bool SelectionDAGBuilder::visitStrCmpCall(const CallInst &I) { 5724 // Verify that the prototype makes sense. int strcmp(void*,void*) 5725 if (I.getNumArgOperands() != 2) 5726 return false; 5727 5728 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1); 5729 if (!Arg0->getType()->isPointerTy() || 5730 !Arg1->getType()->isPointerTy() || 5731 !I.getType()->isIntegerTy()) 5732 return false; 5733 5734 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); 5735 std::pair<SDValue, SDValue> Res = 5736 TSI.EmitTargetCodeForStrcmp(DAG, getCurSDLoc(), DAG.getRoot(), 5737 getValue(Arg0), getValue(Arg1), 5738 MachinePointerInfo(Arg0), 5739 MachinePointerInfo(Arg1)); 5740 if (Res.first.getNode()) { 5741 processIntegerCallValue(I, Res.first, true); 5742 PendingLoads.push_back(Res.second); 5743 return true; 5744 } 5745 5746 return false; 5747} 5748 5749/// visitStrLenCall -- See if we can lower a strlen call into an optimized 5750/// form. If so, return true and lower it, otherwise return false and it 5751/// will be lowered like a normal call. 5752bool SelectionDAGBuilder::visitStrLenCall(const CallInst &I) { 5753 // Verify that the prototype makes sense. size_t strlen(char *) 5754 if (I.getNumArgOperands() != 1) 5755 return false; 5756 5757 const Value *Arg0 = I.getArgOperand(0); 5758 if (!Arg0->getType()->isPointerTy() || !I.getType()->isIntegerTy()) 5759 return false; 5760 5761 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); 5762 std::pair<SDValue, SDValue> Res = 5763 TSI.EmitTargetCodeForStrlen(DAG, getCurSDLoc(), DAG.getRoot(), 5764 getValue(Arg0), MachinePointerInfo(Arg0)); 5765 if (Res.first.getNode()) { 5766 processIntegerCallValue(I, Res.first, false); 5767 PendingLoads.push_back(Res.second); 5768 return true; 5769 } 5770 5771 return false; 5772} 5773 5774/// visitStrNLenCall -- See if we can lower a strnlen call into an optimized 5775/// form. If so, return true and lower it, otherwise return false and it 5776/// will be lowered like a normal call. 5777bool SelectionDAGBuilder::visitStrNLenCall(const CallInst &I) { 5778 // Verify that the prototype makes sense. size_t strnlen(char *, size_t) 5779 if (I.getNumArgOperands() != 2) 5780 return false; 5781 5782 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1); 5783 if (!Arg0->getType()->isPointerTy() || 5784 !Arg1->getType()->isIntegerTy() || 5785 !I.getType()->isIntegerTy()) 5786 return false; 5787 5788 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo(); 5789 std::pair<SDValue, SDValue> Res = 5790 TSI.EmitTargetCodeForStrnlen(DAG, getCurSDLoc(), DAG.getRoot(), 5791 getValue(Arg0), getValue(Arg1), 5792 MachinePointerInfo(Arg0)); 5793 if (Res.first.getNode()) { 5794 processIntegerCallValue(I, Res.first, false); 5795 PendingLoads.push_back(Res.second); 5796 return true; 5797 } 5798 5799 return false; 5800} 5801 5802/// visitUnaryFloatCall - If a call instruction is a unary floating-point 5803/// operation (as expected), translate it to an SDNode with the specified opcode 5804/// and return true. 5805bool SelectionDAGBuilder::visitUnaryFloatCall(const CallInst &I, 5806 unsigned Opcode) { 5807 // Sanity check that it really is a unary floating-point call. 5808 if (I.getNumArgOperands() != 1 || 5809 !I.getArgOperand(0)->getType()->isFloatingPointTy() || 5810 I.getType() != I.getArgOperand(0)->getType() || 5811 !I.onlyReadsMemory()) 5812 return false; 5813 5814 SDValue Tmp = getValue(I.getArgOperand(0)); 5815 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), Tmp.getValueType(), Tmp)); 5816 return true; 5817} 5818 5819void SelectionDAGBuilder::visitCall(const CallInst &I) { 5820 // Handle inline assembly differently. 5821 if (isa<InlineAsm>(I.getCalledValue())) { 5822 visitInlineAsm(&I); 5823 return; 5824 } 5825 5826 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI(); 5827 ComputeUsesVAFloatArgument(I, &MMI); 5828 5829 const char *RenameFn = 0; 5830 if (Function *F = I.getCalledFunction()) { 5831 if (F->isDeclaration()) { 5832 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) { 5833 if (unsigned IID = II->getIntrinsicID(F)) { 5834 RenameFn = visitIntrinsicCall(I, IID); 5835 if (!RenameFn) 5836 return; 5837 } 5838 } 5839 if (unsigned IID = F->getIntrinsicID()) { 5840 RenameFn = visitIntrinsicCall(I, IID); 5841 if (!RenameFn) 5842 return; 5843 } 5844 } 5845 5846 // Check for well-known libc/libm calls. If the function is internal, it 5847 // can't be a library call. 5848 LibFunc::Func Func; 5849 if (!F->hasLocalLinkage() && F->hasName() && 5850 LibInfo->getLibFunc(F->getName(), Func) && 5851 LibInfo->hasOptimizedCodeGen(Func)) { 5852 switch (Func) { 5853 default: break; 5854 case LibFunc::copysign: 5855 case LibFunc::copysignf: 5856 case LibFunc::copysignl: 5857 if (I.getNumArgOperands() == 2 && // Basic sanity checks. 5858 I.getArgOperand(0)->getType()->isFloatingPointTy() && 5859 I.getType() == I.getArgOperand(0)->getType() && 5860 I.getType() == I.getArgOperand(1)->getType() && 5861 I.onlyReadsMemory()) { 5862 SDValue LHS = getValue(I.getArgOperand(0)); 5863 SDValue RHS = getValue(I.getArgOperand(1)); 5864 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurSDLoc(), 5865 LHS.getValueType(), LHS, RHS)); 5866 return; 5867 } 5868 break; 5869 case LibFunc::fabs: 5870 case LibFunc::fabsf: 5871 case LibFunc::fabsl: 5872 if (visitUnaryFloatCall(I, ISD::FABS)) 5873 return; 5874 break; 5875 case LibFunc::sin: 5876 case LibFunc::sinf: 5877 case LibFunc::sinl: 5878 if (visitUnaryFloatCall(I, ISD::FSIN)) 5879 return; 5880 break; 5881 case LibFunc::cos: 5882 case LibFunc::cosf: 5883 case LibFunc::cosl: 5884 if (visitUnaryFloatCall(I, ISD::FCOS)) 5885 return; 5886 break; 5887 case LibFunc::sqrt: 5888 case LibFunc::sqrtf: 5889 case LibFunc::sqrtl: 5890 case LibFunc::sqrt_finite: 5891 case LibFunc::sqrtf_finite: 5892 case LibFunc::sqrtl_finite: 5893 if (visitUnaryFloatCall(I, ISD::FSQRT)) 5894 return; 5895 break; 5896 case LibFunc::floor: 5897 case LibFunc::floorf: 5898 case LibFunc::floorl: 5899 if (visitUnaryFloatCall(I, ISD::FFLOOR)) 5900 return; 5901 break; 5902 case LibFunc::nearbyint: 5903 case LibFunc::nearbyintf: 5904 case LibFunc::nearbyintl: 5905 if (visitUnaryFloatCall(I, ISD::FNEARBYINT)) 5906 return; 5907 break; 5908 case LibFunc::ceil: 5909 case LibFunc::ceilf: 5910 case LibFunc::ceill: 5911 if (visitUnaryFloatCall(I, ISD::FCEIL)) 5912 return; 5913 break; 5914 case LibFunc::rint: 5915 case LibFunc::rintf: 5916 case LibFunc::rintl: 5917 if (visitUnaryFloatCall(I, ISD::FRINT)) 5918 return; 5919 break; 5920 case LibFunc::round: 5921 case LibFunc::roundf: 5922 case LibFunc::roundl: 5923 if (visitUnaryFloatCall(I, ISD::FROUND)) 5924 return; 5925 break; 5926 case LibFunc::trunc: 5927 case LibFunc::truncf: 5928 case LibFunc::truncl: 5929 if (visitUnaryFloatCall(I, ISD::FTRUNC)) 5930 return; 5931 break; 5932 case LibFunc::log2: 5933 case LibFunc::log2f: 5934 case LibFunc::log2l: 5935 if (visitUnaryFloatCall(I, ISD::FLOG2)) 5936 return; 5937 break; 5938 case LibFunc::exp2: 5939 case LibFunc::exp2f: 5940 case LibFunc::exp2l: 5941 if (visitUnaryFloatCall(I, ISD::FEXP2)) 5942 return; 5943 break; 5944 case LibFunc::memcmp: 5945 if (visitMemCmpCall(I)) 5946 return; 5947 break; 5948 case LibFunc::memchr: 5949 if (visitMemChrCall(I)) 5950 return; 5951 break; 5952 case LibFunc::strcpy: 5953 if (visitStrCpyCall(I, false)) 5954 return; 5955 break; 5956 case LibFunc::stpcpy: 5957 if (visitStrCpyCall(I, true)) 5958 return; 5959 break; 5960 case LibFunc::strcmp: 5961 if (visitStrCmpCall(I)) 5962 return; 5963 break; 5964 case LibFunc::strlen: 5965 if (visitStrLenCall(I)) 5966 return; 5967 break; 5968 case LibFunc::strnlen: 5969 if (visitStrNLenCall(I)) 5970 return; 5971 break; 5972 } 5973 } 5974 } 5975 5976 SDValue Callee; 5977 if (!RenameFn) 5978 Callee = getValue(I.getCalledValue()); 5979 else 5980 Callee = DAG.getExternalSymbol(RenameFn, 5981 TM.getTargetLowering()->getPointerTy()); 5982 5983 // Check if we can potentially perform a tail call. More detailed checking is 5984 // be done within LowerCallTo, after more information about the call is known. 5985 LowerCallTo(&I, Callee, I.isTailCall()); 5986} 5987 5988namespace { 5989 5990/// AsmOperandInfo - This contains information for each constraint that we are 5991/// lowering. 5992class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo { 5993public: 5994 /// CallOperand - If this is the result output operand or a clobber 5995 /// this is null, otherwise it is the incoming operand to the CallInst. 5996 /// This gets modified as the asm is processed. 5997 SDValue CallOperand; 5998 5999 /// AssignedRegs - If this is a register or register class operand, this 6000 /// contains the set of register corresponding to the operand. 6001 RegsForValue AssignedRegs; 6002 6003 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info) 6004 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) { 6005 } 6006 6007 /// getCallOperandValEVT - Return the EVT of the Value* that this operand 6008 /// corresponds to. If there is no Value* for this operand, it returns 6009 /// MVT::Other. 6010 EVT getCallOperandValEVT(LLVMContext &Context, 6011 const TargetLowering &TLI, 6012 const DataLayout *TD) const { 6013 if (CallOperandVal == 0) return MVT::Other; 6014 6015 if (isa<BasicBlock>(CallOperandVal)) 6016 return TLI.getPointerTy(); 6017 6018 llvm::Type *OpTy = CallOperandVal->getType(); 6019 6020 // FIXME: code duplicated from TargetLowering::ParseConstraints(). 6021 // If this is an indirect operand, the operand is a pointer to the 6022 // accessed type. 6023 if (isIndirect) { 6024 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy); 6025 if (!PtrTy) 6026 report_fatal_error("Indirect operand for inline asm not a pointer!"); 6027 OpTy = PtrTy->getElementType(); 6028 } 6029 6030 // Look for vector wrapped in a struct. e.g. { <16 x i8> }. 6031 if (StructType *STy = dyn_cast<StructType>(OpTy)) 6032 if (STy->getNumElements() == 1) 6033 OpTy = STy->getElementType(0); 6034 6035 // If OpTy is not a single value, it may be a struct/union that we 6036 // can tile with integers. 6037 if (!OpTy->isSingleValueType() && OpTy->isSized()) { 6038 unsigned BitSize = TD->getTypeSizeInBits(OpTy); 6039 switch (BitSize) { 6040 default: break; 6041 case 1: 6042 case 8: 6043 case 16: 6044 case 32: 6045 case 64: 6046 case 128: 6047 OpTy = IntegerType::get(Context, BitSize); 6048 break; 6049 } 6050 } 6051 6052 return TLI.getValueType(OpTy, true); 6053 } 6054}; 6055 6056typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector; 6057 6058} // end anonymous namespace 6059 6060/// GetRegistersForValue - Assign registers (virtual or physical) for the 6061/// specified operand. We prefer to assign virtual registers, to allow the 6062/// register allocator to handle the assignment process. However, if the asm 6063/// uses features that we can't model on machineinstrs, we have SDISel do the 6064/// allocation. This produces generally horrible, but correct, code. 6065/// 6066/// OpInfo describes the operand. 6067/// 6068static void GetRegistersForValue(SelectionDAG &DAG, 6069 const TargetLowering &TLI, 6070 SDLoc DL, 6071 SDISelAsmOperandInfo &OpInfo) { 6072 LLVMContext &Context = *DAG.getContext(); 6073 6074 MachineFunction &MF = DAG.getMachineFunction(); 6075 SmallVector<unsigned, 4> Regs; 6076 6077 // If this is a constraint for a single physreg, or a constraint for a 6078 // register class, find it. 6079 std::pair<unsigned, const TargetRegisterClass*> PhysReg = 6080 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode, 6081 OpInfo.ConstraintVT); 6082 6083 unsigned NumRegs = 1; 6084 if (OpInfo.ConstraintVT != MVT::Other) { 6085 // If this is a FP input in an integer register (or visa versa) insert a bit 6086 // cast of the input value. More generally, handle any case where the input 6087 // value disagrees with the register class we plan to stick this in. 6088 if (OpInfo.Type == InlineAsm::isInput && 6089 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) { 6090 // Try to convert to the first EVT that the reg class contains. If the 6091 // types are identical size, use a bitcast to convert (e.g. two differing 6092 // vector types). 6093 MVT RegVT = *PhysReg.second->vt_begin(); 6094 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) { 6095 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL, 6096 RegVT, OpInfo.CallOperand); 6097 OpInfo.ConstraintVT = RegVT; 6098 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) { 6099 // If the input is a FP value and we want it in FP registers, do a 6100 // bitcast to the corresponding integer type. This turns an f64 value 6101 // into i64, which can be passed with two i32 values on a 32-bit 6102 // machine. 6103 RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits()); 6104 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL, 6105 RegVT, OpInfo.CallOperand); 6106 OpInfo.ConstraintVT = RegVT; 6107 } 6108 } 6109 6110 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT); 6111 } 6112 6113 MVT RegVT; 6114 EVT ValueVT = OpInfo.ConstraintVT; 6115 6116 // If this is a constraint for a specific physical register, like {r17}, 6117 // assign it now. 6118 if (unsigned AssignedReg = PhysReg.first) { 6119 const TargetRegisterClass *RC = PhysReg.second; 6120 if (OpInfo.ConstraintVT == MVT::Other) 6121 ValueVT = *RC->vt_begin(); 6122 6123 // Get the actual register value type. This is important, because the user 6124 // may have asked for (e.g.) the AX register in i32 type. We need to 6125 // remember that AX is actually i16 to get the right extension. 6126 RegVT = *RC->vt_begin(); 6127 6128 // This is a explicit reference to a physical register. 6129 Regs.push_back(AssignedReg); 6130 6131 // If this is an expanded reference, add the rest of the regs to Regs. 6132 if (NumRegs != 1) { 6133 TargetRegisterClass::iterator I = RC->begin(); 6134 for (; *I != AssignedReg; ++I) 6135 assert(I != RC->end() && "Didn't find reg!"); 6136 6137 // Already added the first reg. 6138 --NumRegs; ++I; 6139 for (; NumRegs; --NumRegs, ++I) { 6140 assert(I != RC->end() && "Ran out of registers to allocate!"); 6141 Regs.push_back(*I); 6142 } 6143 } 6144 6145 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT); 6146 return; 6147 } 6148 6149 // Otherwise, if this was a reference to an LLVM register class, create vregs 6150 // for this reference. 6151 if (const TargetRegisterClass *RC = PhysReg.second) { 6152 RegVT = *RC->vt_begin(); 6153 if (OpInfo.ConstraintVT == MVT::Other) 6154 ValueVT = RegVT; 6155 6156 // Create the appropriate number of virtual registers. 6157 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 6158 for (; NumRegs; --NumRegs) 6159 Regs.push_back(RegInfo.createVirtualRegister(RC)); 6160 6161 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT); 6162 return; 6163 } 6164 6165 // Otherwise, we couldn't allocate enough registers for this. 6166} 6167 6168/// visitInlineAsm - Handle a call to an InlineAsm object. 6169/// 6170void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) { 6171 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 6172 6173 /// ConstraintOperands - Information about all of the constraints. 6174 SDISelAsmOperandInfoVector ConstraintOperands; 6175 6176 const TargetLowering *TLI = TM.getTargetLowering(); 6177 TargetLowering::AsmOperandInfoVector 6178 TargetConstraints = TLI->ParseConstraints(CS); 6179 6180 bool hasMemory = false; 6181 6182 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 6183 unsigned ResNo = 0; // ResNo - The result number of the next output. 6184 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { 6185 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i])); 6186 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back(); 6187 6188 MVT OpVT = MVT::Other; 6189 6190 // Compute the value type for each operand. 6191 switch (OpInfo.Type) { 6192 case InlineAsm::isOutput: 6193 // Indirect outputs just consume an argument. 6194 if (OpInfo.isIndirect) { 6195 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 6196 break; 6197 } 6198 6199 // The return value of the call is this value. As such, there is no 6200 // corresponding argument. 6201 assert(!CS.getType()->isVoidTy() && "Bad inline asm!"); 6202 if (StructType *STy = dyn_cast<StructType>(CS.getType())) { 6203 OpVT = TLI->getSimpleValueType(STy->getElementType(ResNo)); 6204 } else { 6205 assert(ResNo == 0 && "Asm only has one result!"); 6206 OpVT = TLI->getSimpleValueType(CS.getType()); 6207 } 6208 ++ResNo; 6209 break; 6210 case InlineAsm::isInput: 6211 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 6212 break; 6213 case InlineAsm::isClobber: 6214 // Nothing to do. 6215 break; 6216 } 6217 6218 // If this is an input or an indirect output, process the call argument. 6219 // BasicBlocks are labels, currently appearing only in asm's. 6220 if (OpInfo.CallOperandVal) { 6221 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) { 6222 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]); 6223 } else { 6224 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal); 6225 } 6226 6227 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), *TLI, TD). 6228 getSimpleVT(); 6229 } 6230 6231 OpInfo.ConstraintVT = OpVT; 6232 6233 // Indirect operand accesses access memory. 6234 if (OpInfo.isIndirect) 6235 hasMemory = true; 6236 else { 6237 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) { 6238 TargetLowering::ConstraintType 6239 CType = TLI->getConstraintType(OpInfo.Codes[j]); 6240 if (CType == TargetLowering::C_Memory) { 6241 hasMemory = true; 6242 break; 6243 } 6244 } 6245 } 6246 } 6247 6248 SDValue Chain, Flag; 6249 6250 // We won't need to flush pending loads if this asm doesn't touch 6251 // memory and is nonvolatile. 6252 if (hasMemory || IA->hasSideEffects()) 6253 Chain = getRoot(); 6254 else 6255 Chain = DAG.getRoot(); 6256 6257 // Second pass over the constraints: compute which constraint option to use 6258 // and assign registers to constraints that want a specific physreg. 6259 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { 6260 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; 6261 6262 // If this is an output operand with a matching input operand, look up the 6263 // matching input. If their types mismatch, e.g. one is an integer, the 6264 // other is floating point, or their sizes are different, flag it as an 6265 // error. 6266 if (OpInfo.hasMatchingInput()) { 6267 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; 6268 6269 if (OpInfo.ConstraintVT != Input.ConstraintVT) { 6270 std::pair<unsigned, const TargetRegisterClass*> MatchRC = 6271 TLI->getRegForInlineAsmConstraint(OpInfo.ConstraintCode, 6272 OpInfo.ConstraintVT); 6273 std::pair<unsigned, const TargetRegisterClass*> InputRC = 6274 TLI->getRegForInlineAsmConstraint(Input.ConstraintCode, 6275 Input.ConstraintVT); 6276 if ((OpInfo.ConstraintVT.isInteger() != 6277 Input.ConstraintVT.isInteger()) || 6278 (MatchRC.second != InputRC.second)) { 6279 report_fatal_error("Unsupported asm: input constraint" 6280 " with a matching output constraint of" 6281 " incompatible type!"); 6282 } 6283 Input.ConstraintVT = OpInfo.ConstraintVT; 6284 } 6285 } 6286 6287 // Compute the constraint code and ConstraintType to use. 6288 TLI->ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG); 6289 6290 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 6291 OpInfo.Type == InlineAsm::isClobber) 6292 continue; 6293 6294 // If this is a memory input, and if the operand is not indirect, do what we 6295 // need to to provide an address for the memory input. 6296 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 6297 !OpInfo.isIndirect) { 6298 assert((OpInfo.isMultipleAlternative || 6299 (OpInfo.Type == InlineAsm::isInput)) && 6300 "Can only indirectify direct input operands!"); 6301 6302 // Memory operands really want the address of the value. If we don't have 6303 // an indirect input, put it in the constpool if we can, otherwise spill 6304 // it to a stack slot. 6305 // TODO: This isn't quite right. We need to handle these according to 6306 // the addressing mode that the constraint wants. Also, this may take 6307 // an additional register for the computation and we don't want that 6308 // either. 6309 6310 // If the operand is a float, integer, or vector constant, spill to a 6311 // constant pool entry to get its address. 6312 const Value *OpVal = OpInfo.CallOperandVal; 6313 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) || 6314 isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) { 6315 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal), 6316 TLI->getPointerTy()); 6317 } else { 6318 // Otherwise, create a stack slot and emit a store to it before the 6319 // asm. 6320 Type *Ty = OpVal->getType(); 6321 uint64_t TySize = TLI->getDataLayout()->getTypeAllocSize(Ty); 6322 unsigned Align = TLI->getDataLayout()->getPrefTypeAlignment(Ty); 6323 MachineFunction &MF = DAG.getMachineFunction(); 6324 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false); 6325 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI->getPointerTy()); 6326 Chain = DAG.getStore(Chain, getCurSDLoc(), 6327 OpInfo.CallOperand, StackSlot, 6328 MachinePointerInfo::getFixedStack(SSFI), 6329 false, false, 0); 6330 OpInfo.CallOperand = StackSlot; 6331 } 6332 6333 // There is no longer a Value* corresponding to this operand. 6334 OpInfo.CallOperandVal = 0; 6335 6336 // It is now an indirect operand. 6337 OpInfo.isIndirect = true; 6338 } 6339 6340 // If this constraint is for a specific register, allocate it before 6341 // anything else. 6342 if (OpInfo.ConstraintType == TargetLowering::C_Register) 6343 GetRegistersForValue(DAG, *TLI, getCurSDLoc(), OpInfo); 6344 } 6345 6346 // Second pass - Loop over all of the operands, assigning virtual or physregs 6347 // to register class operands. 6348 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { 6349 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; 6350 6351 // C_Register operands have already been allocated, Other/Memory don't need 6352 // to be. 6353 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass) 6354 GetRegistersForValue(DAG, *TLI, getCurSDLoc(), OpInfo); 6355 } 6356 6357 // AsmNodeOperands - The operands for the ISD::INLINEASM node. 6358 std::vector<SDValue> AsmNodeOperands; 6359 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain 6360 AsmNodeOperands.push_back( 6361 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(), 6362 TLI->getPointerTy())); 6363 6364 // If we have a !srcloc metadata node associated with it, we want to attach 6365 // this to the ultimately generated inline asm machineinstr. To do this, we 6366 // pass in the third operand as this (potentially null) inline asm MDNode. 6367 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc"); 6368 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc)); 6369 6370 // Remember the HasSideEffect, AlignStack, AsmDialect, MayLoad and MayStore 6371 // bits as operand 3. 6372 unsigned ExtraInfo = 0; 6373 if (IA->hasSideEffects()) 6374 ExtraInfo |= InlineAsm::Extra_HasSideEffects; 6375 if (IA->isAlignStack()) 6376 ExtraInfo |= InlineAsm::Extra_IsAlignStack; 6377 // Set the asm dialect. 6378 ExtraInfo |= IA->getDialect() * InlineAsm::Extra_AsmDialect; 6379 6380 // Determine if this InlineAsm MayLoad or MayStore based on the constraints. 6381 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { 6382 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i]; 6383 6384 // Compute the constraint code and ConstraintType to use. 6385 TLI->ComputeConstraintToUse(OpInfo, SDValue()); 6386 6387 // Ideally, we would only check against memory constraints. However, the 6388 // meaning of an other constraint can be target-specific and we can't easily 6389 // reason about it. Therefore, be conservative and set MayLoad/MayStore 6390 // for other constriants as well. 6391 if (OpInfo.ConstraintType == TargetLowering::C_Memory || 6392 OpInfo.ConstraintType == TargetLowering::C_Other) { 6393 if (OpInfo.Type == InlineAsm::isInput) 6394 ExtraInfo |= InlineAsm::Extra_MayLoad; 6395 else if (OpInfo.Type == InlineAsm::isOutput) 6396 ExtraInfo |= InlineAsm::Extra_MayStore; 6397 else if (OpInfo.Type == InlineAsm::isClobber) 6398 ExtraInfo |= (InlineAsm::Extra_MayLoad | InlineAsm::Extra_MayStore); 6399 } 6400 } 6401 6402 AsmNodeOperands.push_back(DAG.getTargetConstant(ExtraInfo, 6403 TLI->getPointerTy())); 6404 6405 // Loop over all of the inputs, copying the operand values into the 6406 // appropriate registers and processing the output regs. 6407 RegsForValue RetValRegs; 6408 6409 // IndirectStoresToEmit - The set of stores to emit after the inline asm node. 6410 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit; 6411 6412 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { 6413 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i]; 6414 6415 switch (OpInfo.Type) { 6416 case InlineAsm::isOutput: { 6417 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass && 6418 OpInfo.ConstraintType != TargetLowering::C_Register) { 6419 // Memory output, or 'other' output (e.g. 'X' constraint). 6420 assert(OpInfo.isIndirect && "Memory output must be indirect operand"); 6421 6422 // Add information to the INLINEASM node to know about this output. 6423 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1); 6424 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags, 6425 TLI->getPointerTy())); 6426 AsmNodeOperands.push_back(OpInfo.CallOperand); 6427 break; 6428 } 6429 6430 // Otherwise, this is a register or register class output. 6431 6432 // Copy the output from the appropriate register. Find a register that 6433 // we can use. 6434 if (OpInfo.AssignedRegs.Regs.empty()) { 6435 LLVMContext &Ctx = *DAG.getContext(); 6436 Ctx.emitError(CS.getInstruction(), 6437 "couldn't allocate output register for constraint '" + 6438 Twine(OpInfo.ConstraintCode) + "'"); 6439 return; 6440 } 6441 6442 // If this is an indirect operand, store through the pointer after the 6443 // asm. 6444 if (OpInfo.isIndirect) { 6445 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs, 6446 OpInfo.CallOperandVal)); 6447 } else { 6448 // This is the result value of the call. 6449 assert(!CS.getType()->isVoidTy() && "Bad inline asm!"); 6450 // Concatenate this output onto the outputs list. 6451 RetValRegs.append(OpInfo.AssignedRegs); 6452 } 6453 6454 // Add information to the INLINEASM node to know that this register is 6455 // set. 6456 OpInfo.AssignedRegs 6457 .AddInlineAsmOperands(OpInfo.isEarlyClobber 6458 ? InlineAsm::Kind_RegDefEarlyClobber 6459 : InlineAsm::Kind_RegDef, 6460 false, 0, DAG, AsmNodeOperands); 6461 break; 6462 } 6463 case InlineAsm::isInput: { 6464 SDValue InOperandVal = OpInfo.CallOperand; 6465 6466 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint? 6467 // If this is required to match an output register we have already set, 6468 // just use its register. 6469 unsigned OperandNo = OpInfo.getMatchedOperand(); 6470 6471 // Scan until we find the definition we already emitted of this operand. 6472 // When we find it, create a RegsForValue operand. 6473 unsigned CurOp = InlineAsm::Op_FirstOperand; 6474 for (; OperandNo; --OperandNo) { 6475 // Advance to the next operand. 6476 unsigned OpFlag = 6477 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue(); 6478 assert((InlineAsm::isRegDefKind(OpFlag) || 6479 InlineAsm::isRegDefEarlyClobberKind(OpFlag) || 6480 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?"); 6481 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1; 6482 } 6483 6484 unsigned OpFlag = 6485 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue(); 6486 if (InlineAsm::isRegDefKind(OpFlag) || 6487 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) { 6488 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs. 6489 if (OpInfo.isIndirect) { 6490 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c 6491 LLVMContext &Ctx = *DAG.getContext(); 6492 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:" 6493 " don't know how to handle tied " 6494 "indirect register inputs"); 6495 return; 6496 } 6497 6498 RegsForValue MatchedRegs; 6499 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType()); 6500 MVT RegVT = AsmNodeOperands[CurOp+1].getSimpleValueType(); 6501 MatchedRegs.RegVTs.push_back(RegVT); 6502 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo(); 6503 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag); 6504 i != e; ++i) { 6505 if (const TargetRegisterClass *RC = TLI->getRegClassFor(RegVT)) 6506 MatchedRegs.Regs.push_back(RegInfo.createVirtualRegister(RC)); 6507 else { 6508 LLVMContext &Ctx = *DAG.getContext(); 6509 Ctx.emitError(CS.getInstruction(), 6510 "inline asm error: This value" 6511 " type register class is not natively supported!"); 6512 return; 6513 } 6514 } 6515 // Use the produced MatchedRegs object to 6516 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(), 6517 Chain, &Flag, CS.getInstruction()); 6518 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, 6519 true, OpInfo.getMatchedOperand(), 6520 DAG, AsmNodeOperands); 6521 break; 6522 } 6523 6524 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!"); 6525 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 && 6526 "Unexpected number of operands"); 6527 // Add information to the INLINEASM node to know about this input. 6528 // See InlineAsm.h isUseOperandTiedToDef. 6529 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag, 6530 OpInfo.getMatchedOperand()); 6531 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag, 6532 TLI->getPointerTy())); 6533 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]); 6534 break; 6535 } 6536 6537 // Treat indirect 'X' constraint as memory. 6538 if (OpInfo.ConstraintType == TargetLowering::C_Other && 6539 OpInfo.isIndirect) 6540 OpInfo.ConstraintType = TargetLowering::C_Memory; 6541 6542 if (OpInfo.ConstraintType == TargetLowering::C_Other) { 6543 std::vector<SDValue> Ops; 6544 TLI->LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode, 6545 Ops, DAG); 6546 if (Ops.empty()) { 6547 LLVMContext &Ctx = *DAG.getContext(); 6548 Ctx.emitError(CS.getInstruction(), 6549 "invalid operand for inline asm constraint '" + 6550 Twine(OpInfo.ConstraintCode) + "'"); 6551 return; 6552 } 6553 6554 // Add information to the INLINEASM node to know about this input. 6555 unsigned ResOpType = 6556 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size()); 6557 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType, 6558 TLI->getPointerTy())); 6559 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end()); 6560 break; 6561 } 6562 6563 if (OpInfo.ConstraintType == TargetLowering::C_Memory) { 6564 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!"); 6565 assert(InOperandVal.getValueType() == TLI->getPointerTy() && 6566 "Memory operands expect pointer values"); 6567 6568 // Add information to the INLINEASM node to know about this input. 6569 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1); 6570 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType, 6571 TLI->getPointerTy())); 6572 AsmNodeOperands.push_back(InOperandVal); 6573 break; 6574 } 6575 6576 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass || 6577 OpInfo.ConstraintType == TargetLowering::C_Register) && 6578 "Unknown constraint type!"); 6579 6580 // TODO: Support this. 6581 if (OpInfo.isIndirect) { 6582 LLVMContext &Ctx = *DAG.getContext(); 6583 Ctx.emitError(CS.getInstruction(), 6584 "Don't know how to handle indirect register inputs yet " 6585 "for constraint '" + 6586 Twine(OpInfo.ConstraintCode) + "'"); 6587 return; 6588 } 6589 6590 // Copy the input into the appropriate registers. 6591 if (OpInfo.AssignedRegs.Regs.empty()) { 6592 LLVMContext &Ctx = *DAG.getContext(); 6593 Ctx.emitError(CS.getInstruction(), 6594 "couldn't allocate input reg for constraint '" + 6595 Twine(OpInfo.ConstraintCode) + "'"); 6596 return; 6597 } 6598 6599 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(), 6600 Chain, &Flag, CS.getInstruction()); 6601 6602 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0, 6603 DAG, AsmNodeOperands); 6604 break; 6605 } 6606 case InlineAsm::isClobber: { 6607 // Add the clobbered value to the operand list, so that the register 6608 // allocator is aware that the physreg got clobbered. 6609 if (!OpInfo.AssignedRegs.Regs.empty()) 6610 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber, 6611 false, 0, DAG, 6612 AsmNodeOperands); 6613 break; 6614 } 6615 } 6616 } 6617 6618 // Finish up input operands. Set the input chain and add the flag last. 6619 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain; 6620 if (Flag.getNode()) AsmNodeOperands.push_back(Flag); 6621 6622 Chain = DAG.getNode(ISD::INLINEASM, getCurSDLoc(), 6623 DAG.getVTList(MVT::Other, MVT::Glue), 6624 &AsmNodeOperands[0], AsmNodeOperands.size()); 6625 Flag = Chain.getValue(1); 6626 6627 // If this asm returns a register value, copy the result from that register 6628 // and set it as the value of the call. 6629 if (!RetValRegs.Regs.empty()) { 6630 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), 6631 Chain, &Flag, CS.getInstruction()); 6632 6633 // FIXME: Why don't we do this for inline asms with MRVs? 6634 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) { 6635 EVT ResultType = TLI->getValueType(CS.getType()); 6636 6637 // If any of the results of the inline asm is a vector, it may have the 6638 // wrong width/num elts. This can happen for register classes that can 6639 // contain multiple different value types. The preg or vreg allocated may 6640 // not have the same VT as was expected. Convert it to the right type 6641 // with bit_convert. 6642 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) { 6643 Val = DAG.getNode(ISD::BITCAST, getCurSDLoc(), 6644 ResultType, Val); 6645 6646 } else if (ResultType != Val.getValueType() && 6647 ResultType.isInteger() && Val.getValueType().isInteger()) { 6648 // If a result value was tied to an input value, the computed result may 6649 // have a wider width than the expected result. Extract the relevant 6650 // portion. 6651 Val = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), ResultType, Val); 6652 } 6653 6654 assert(ResultType == Val.getValueType() && "Asm result value mismatch!"); 6655 } 6656 6657 setValue(CS.getInstruction(), Val); 6658 // Don't need to use this as a chain in this case. 6659 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty()) 6660 return; 6661 } 6662 6663 std::vector<std::pair<SDValue, const Value *> > StoresToEmit; 6664 6665 // Process indirect outputs, first output all of the flagged copies out of 6666 // physregs. 6667 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) { 6668 RegsForValue &OutRegs = IndirectStoresToEmit[i].first; 6669 const Value *Ptr = IndirectStoresToEmit[i].second; 6670 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), 6671 Chain, &Flag, IA); 6672 StoresToEmit.push_back(std::make_pair(OutVal, Ptr)); 6673 } 6674 6675 // Emit the non-flagged stores from the physregs. 6676 SmallVector<SDValue, 8> OutChains; 6677 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) { 6678 SDValue Val = DAG.getStore(Chain, getCurSDLoc(), 6679 StoresToEmit[i].first, 6680 getValue(StoresToEmit[i].second), 6681 MachinePointerInfo(StoresToEmit[i].second), 6682 false, false, 0); 6683 OutChains.push_back(Val); 6684 } 6685 6686 if (!OutChains.empty()) 6687 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, 6688 &OutChains[0], OutChains.size()); 6689 6690 DAG.setRoot(Chain); 6691} 6692 6693void SelectionDAGBuilder::visitVAStart(const CallInst &I) { 6694 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurSDLoc(), 6695 MVT::Other, getRoot(), 6696 getValue(I.getArgOperand(0)), 6697 DAG.getSrcValue(I.getArgOperand(0)))); 6698} 6699 6700void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) { 6701 const TargetLowering *TLI = TM.getTargetLowering(); 6702 const DataLayout &TD = *TLI->getDataLayout(); 6703 SDValue V = DAG.getVAArg(TLI->getValueType(I.getType()), getCurSDLoc(), 6704 getRoot(), getValue(I.getOperand(0)), 6705 DAG.getSrcValue(I.getOperand(0)), 6706 TD.getABITypeAlignment(I.getType())); 6707 setValue(&I, V); 6708 DAG.setRoot(V.getValue(1)); 6709} 6710 6711void SelectionDAGBuilder::visitVAEnd(const CallInst &I) { 6712 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurSDLoc(), 6713 MVT::Other, getRoot(), 6714 getValue(I.getArgOperand(0)), 6715 DAG.getSrcValue(I.getArgOperand(0)))); 6716} 6717 6718void SelectionDAGBuilder::visitVACopy(const CallInst &I) { 6719 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurSDLoc(), 6720 MVT::Other, getRoot(), 6721 getValue(I.getArgOperand(0)), 6722 getValue(I.getArgOperand(1)), 6723 DAG.getSrcValue(I.getArgOperand(0)), 6724 DAG.getSrcValue(I.getArgOperand(1)))); 6725} 6726 6727/// \brief Lower an argument list according to the target calling convention. 6728/// 6729/// \return A tuple of <return-value, token-chain> 6730/// 6731/// This is a helper for lowering intrinsics that follow a target calling 6732/// convention or require stack pointer adjustment. Only a subset of the 6733/// intrinsic's operands need to participate in the calling convention. 6734std::pair<SDValue, SDValue> 6735SelectionDAGBuilder::LowerCallOperands(const CallInst &CI, unsigned ArgIdx, 6736 unsigned NumArgs, SDValue Callee, 6737 bool useVoidTy) { 6738 TargetLowering::ArgListTy Args; 6739 Args.reserve(NumArgs); 6740 6741 // Populate the argument list. 6742 // Attributes for args start at offset 1, after the return attribute. 6743 ImmutableCallSite CS(&CI); 6744 for (unsigned ArgI = ArgIdx, ArgE = ArgIdx + NumArgs, AttrI = ArgIdx + 1; 6745 ArgI != ArgE; ++ArgI) { 6746 const Value *V = CI.getOperand(ArgI); 6747 6748 assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic."); 6749 6750 TargetLowering::ArgListEntry Entry; 6751 Entry.Node = getValue(V); 6752 Entry.Ty = V->getType(); 6753 Entry.setAttributes(&CS, AttrI); 6754 Args.push_back(Entry); 6755 } 6756 6757 Type *retTy = useVoidTy ? Type::getVoidTy(*DAG.getContext()) : CI.getType(); 6758 TargetLowering::CallLoweringInfo CLI(getRoot(), retTy, /*retSExt*/ false, 6759 /*retZExt*/ false, /*isVarArg*/ false, /*isInReg*/ false, NumArgs, 6760 CI.getCallingConv(), /*isTailCall*/ false, /*doesNotReturn*/ false, 6761 /*isReturnValueUsed*/ CI.use_empty(), Callee, Args, DAG, getCurSDLoc()); 6762 6763 const TargetLowering *TLI = TM.getTargetLowering(); 6764 return TLI->LowerCallTo(CLI); 6765} 6766 6767/// \brief Lower llvm.experimental.stackmap directly to its target opcode. 6768void SelectionDAGBuilder::visitStackmap(const CallInst &CI) { 6769 // void @llvm.experimental.stackmap(i32 <id>, i32 <numShadowBytes>, 6770 // [live variables...]) 6771 6772 assert(CI.getType()->isVoidTy() && "Stackmap cannot return a value."); 6773 6774 SDValue Callee = getValue(CI.getCalledValue()); 6775 6776 // Lower into a call sequence with no args and no return value. 6777 std::pair<SDValue, SDValue> Result = LowerCallOperands(CI, 0, 0, Callee); 6778 // Set the root to the target-lowered call chain. 6779 SDValue Chain = Result.second; 6780 DAG.setRoot(Chain); 6781 6782 /// Get a call instruction from the call sequence chain. 6783 /// Tail calls are not allowed. 6784 SDNode *CallEnd = Chain.getNode(); 6785 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && 6786 "Expected a callseq node."); 6787 SDNode *Call = CallEnd->getOperand(0).getNode(); 6788 bool hasGlue = Call->getGluedNode(); 6789 6790 // Replace the target specific call node with the stackmap intrinsic. 6791 SmallVector<SDValue, 8> Ops; 6792 6793 // Add the <id> and <numShadowBytes> constants. 6794 for (unsigned i = 0; i < 2; ++i) { 6795 SDValue tmp = getValue(CI.getOperand(i)); 6796 Ops.push_back(DAG.getTargetConstant( 6797 cast<ConstantSDNode>(tmp)->getZExtValue(), MVT::i32)); 6798 } 6799 // Push live variables for the stack map. 6800 for (unsigned i = 2, e = CI.getNumArgOperands(); i != e; ++i) 6801 Ops.push_back(getValue(CI.getArgOperand(i))); 6802 6803 // Push the chain (this is originally the first operand of the call, but 6804 // becomes now the last or second to last operand). 6805 Ops.push_back(*(Call->op_begin())); 6806 6807 // Push the glue flag (last operand). 6808 if (hasGlue) 6809 Ops.push_back(*(Call->op_end()-1)); 6810 6811 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 6812 6813 // Replace the target specific call node with a STACKMAP node. 6814 MachineSDNode *MN = DAG.getMachineNode(TargetOpcode::STACKMAP, getCurSDLoc(), 6815 NodeTys, Ops); 6816 6817 // StackMap generates no value, so nothing goes in the NodeMap. 6818 6819 // Fixup the consumers of the intrinsic. The chain and glue may be used in the 6820 // call sequence. 6821 DAG.ReplaceAllUsesWith(Call, MN); 6822 6823 DAG.DeleteNode(Call); 6824} 6825 6826/// \brief Lower llvm.experimental.patchpoint directly to its target opcode. 6827void SelectionDAGBuilder::visitPatchpoint(const CallInst &CI) { 6828 // void|i64 @llvm.experimental.patchpoint.void|i64(i32 <id>, 6829 // i32 <numBytes>, 6830 // i8* <target>, 6831 // i32 <numArgs>, 6832 // [Args...], 6833 // [live variables...]) 6834 6835 CallingConv::ID CC = CI.getCallingConv(); 6836 bool isAnyRegCC = CC == CallingConv::AnyReg; 6837 bool hasDef = !CI.getType()->isVoidTy(); 6838 SDValue Callee = getValue(CI.getOperand(2)); // <target> 6839 6840 // Get the real number of arguments participating in the call <numArgs> 6841 unsigned NumArgs = 6842 cast<ConstantSDNode>(getValue(CI.getArgOperand(3)))->getZExtValue(); 6843 6844 // Skip the four meta args: <id>, <numNopBytes>, <target>, <numArgs> 6845 assert(CI.getNumArgOperands() >= NumArgs + 4 && 6846 "Not enough arguments provided to the patchpoint intrinsic"); 6847 6848 // For AnyRegCC the arguments are lowered later on manually. 6849 unsigned NumCallArgs = isAnyRegCC ? 0 : NumArgs; 6850 std::pair<SDValue, SDValue> Result = 6851 LowerCallOperands(CI, 4, NumCallArgs, Callee, isAnyRegCC); 6852 6853 // Set the root to the target-lowered call chain. 6854 SDValue Chain = Result.second; 6855 DAG.setRoot(Chain); 6856 6857 SDNode *CallEnd = Chain.getNode(); 6858 if (hasDef && (CallEnd->getOpcode() == ISD::CopyFromReg)) 6859 CallEnd = CallEnd->getOperand(0).getNode(); 6860 6861 /// Get a call instruction from the call sequence chain. 6862 /// Tail calls are not allowed. 6863 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END && 6864 "Expected a callseq node."); 6865 SDNode *Call = CallEnd->getOperand(0).getNode(); 6866 bool hasGlue = Call->getGluedNode(); 6867 6868 // Replace the target specific call node with the patchable intrinsic. 6869 SmallVector<SDValue, 8> Ops; 6870 6871 // Add the <id> and <numNopBytes> constants. 6872 for (unsigned i = 0; i < 2; ++i) { 6873 SDValue tmp = getValue(CI.getOperand(i)); 6874 Ops.push_back(DAG.getTargetConstant( 6875 cast<ConstantSDNode>(tmp)->getZExtValue(), MVT::i32)); 6876 } 6877 // Assume that the Callee is a constant address. 6878 Ops.push_back( 6879 DAG.getIntPtrConstant(cast<ConstantSDNode>(Callee)->getZExtValue(), 6880 /*isTarget=*/true)); 6881 6882 // Adjust <numArgs> to account for any arguments that have been passed on the 6883 // stack instead. 6884 // Call Node: Chain, Target, {Args}, RegMask, [Glue] 6885 unsigned NumCallRegArgs = Call->getNumOperands() - (hasGlue ? 4 : 3); 6886 NumCallRegArgs = isAnyRegCC ? NumArgs : NumCallRegArgs; 6887 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, MVT::i32)); 6888 6889 // Add the calling convention 6890 Ops.push_back(DAG.getTargetConstant((unsigned)CC, MVT::i32)); 6891 6892 // Add the arguments we omitted previously. The register allocator should 6893 // place these in any free register. 6894 if (isAnyRegCC) 6895 for (unsigned i = 4, e = NumArgs + 4; i != e; ++i) 6896 Ops.push_back(getValue(CI.getArgOperand(i))); 6897 6898 // Push the arguments from the call instruction. 6899 SDNode::op_iterator e = hasGlue ? Call->op_end()-2 : Call->op_end()-1; 6900 for (SDNode::op_iterator i = Call->op_begin()+2; i != e; ++i) 6901 Ops.push_back(*i); 6902 6903 // Push live variables for the stack map. 6904 for (unsigned i = NumArgs + 4, e = CI.getNumArgOperands(); i != e; ++i) { 6905 SDValue OpVal = getValue(CI.getArgOperand(i)); 6906 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(OpVal)) { 6907 Ops.push_back( 6908 DAG.getTargetConstant(StackMaps::ConstantOp, MVT::i64)); 6909 Ops.push_back( 6910 DAG.getTargetConstant(C->getSExtValue(), MVT::i64)); 6911 } else 6912 Ops.push_back(OpVal); 6913 } 6914 6915 // Push the register mask info. 6916 if (hasGlue) 6917 Ops.push_back(*(Call->op_end()-2)); 6918 else 6919 Ops.push_back(*(Call->op_end()-1)); 6920 6921 // Push the chain (this is originally the first operand of the call, but 6922 // becomes now the last or second to last operand). 6923 Ops.push_back(*(Call->op_begin())); 6924 6925 // Push the glue flag (last operand). 6926 if (hasGlue) 6927 Ops.push_back(*(Call->op_end()-1)); 6928 6929 SDVTList NodeTys; 6930 if (isAnyRegCC && hasDef) { 6931 // Create the return types based on the intrinsic definition 6932 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 6933 SmallVector<EVT, 3> ValueVTs; 6934 ComputeValueVTs(TLI, CI.getType(), ValueVTs); 6935 assert(ValueVTs.size() == 1 && "Expected only one return value type."); 6936 6937 // There is always a chain and a glue type at the end 6938 ValueVTs.push_back(MVT::Other); 6939 ValueVTs.push_back(MVT::Glue); 6940 NodeTys = DAG.getVTList(ValueVTs.data(), ValueVTs.size()); 6941 } else 6942 NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 6943 6944 // Replace the target specific call node with a PATCHPOINT node. 6945 MachineSDNode *MN = DAG.getMachineNode(TargetOpcode::PATCHPOINT, 6946 getCurSDLoc(), NodeTys, Ops); 6947 6948 // Update the NodeMap. 6949 if (hasDef) { 6950 if (isAnyRegCC) 6951 setValue(&CI, SDValue(MN, 0)); 6952 else 6953 setValue(&CI, Result.first); 6954 } 6955 6956 // Fixup the consumers of the intrinsic. The chain and glue may be used in the 6957 // call sequence. Furthermore the location of the chain and glue can change 6958 // when the AnyReg calling convention is used and the intrinsic returns a 6959 // value. 6960 if (isAnyRegCC && hasDef) { 6961 SDValue From[] = {SDValue(Call, 0), SDValue(Call, 1)}; 6962 SDValue To[] = {SDValue(MN, 1), SDValue(MN, 2)}; 6963 DAG.ReplaceAllUsesOfValuesWith(From, To, 2); 6964 } else 6965 DAG.ReplaceAllUsesWith(Call, MN); 6966 DAG.DeleteNode(Call); 6967} 6968 6969/// TargetLowering::LowerCallTo - This is the default LowerCallTo 6970/// implementation, which just calls LowerCall. 6971/// FIXME: When all targets are 6972/// migrated to using LowerCall, this hook should be integrated into SDISel. 6973std::pair<SDValue, SDValue> 6974TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const { 6975 // Handle the incoming return values from the call. 6976 CLI.Ins.clear(); 6977 SmallVector<EVT, 4> RetTys; 6978 ComputeValueVTs(*this, CLI.RetTy, RetTys); 6979 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) { 6980 EVT VT = RetTys[I]; 6981 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT); 6982 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT); 6983 for (unsigned i = 0; i != NumRegs; ++i) { 6984 ISD::InputArg MyFlags; 6985 MyFlags.VT = RegisterVT; 6986 MyFlags.ArgVT = VT; 6987 MyFlags.Used = CLI.IsReturnValueUsed; 6988 if (CLI.RetSExt) 6989 MyFlags.Flags.setSExt(); 6990 if (CLI.RetZExt) 6991 MyFlags.Flags.setZExt(); 6992 if (CLI.IsInReg) 6993 MyFlags.Flags.setInReg(); 6994 CLI.Ins.push_back(MyFlags); 6995 } 6996 } 6997 6998 // Handle all of the outgoing arguments. 6999 CLI.Outs.clear(); 7000 CLI.OutVals.clear(); 7001 ArgListTy &Args = CLI.Args; 7002 for (unsigned i = 0, e = Args.size(); i != e; ++i) { 7003 SmallVector<EVT, 4> ValueVTs; 7004 ComputeValueVTs(*this, Args[i].Ty, ValueVTs); 7005 for (unsigned Value = 0, NumValues = ValueVTs.size(); 7006 Value != NumValues; ++Value) { 7007 EVT VT = ValueVTs[Value]; 7008 Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext()); 7009 SDValue Op = SDValue(Args[i].Node.getNode(), 7010 Args[i].Node.getResNo() + Value); 7011 ISD::ArgFlagsTy Flags; 7012 unsigned OriginalAlignment = 7013 getDataLayout()->getABITypeAlignment(ArgTy); 7014 7015 if (Args[i].isZExt) 7016 Flags.setZExt(); 7017 if (Args[i].isSExt) 7018 Flags.setSExt(); 7019 if (Args[i].isInReg) 7020 Flags.setInReg(); 7021 if (Args[i].isSRet) 7022 Flags.setSRet(); 7023 if (Args[i].isByVal) { 7024 Flags.setByVal(); 7025 PointerType *Ty = cast<PointerType>(Args[i].Ty); 7026 Type *ElementTy = Ty->getElementType(); 7027 Flags.setByValSize(getDataLayout()->getTypeAllocSize(ElementTy)); 7028 // For ByVal, alignment should come from FE. BE will guess if this 7029 // info is not there but there are cases it cannot get right. 7030 unsigned FrameAlign; 7031 if (Args[i].Alignment) 7032 FrameAlign = Args[i].Alignment; 7033 else 7034 FrameAlign = getByValTypeAlignment(ElementTy); 7035 Flags.setByValAlign(FrameAlign); 7036 } 7037 if (Args[i].isNest) 7038 Flags.setNest(); 7039 Flags.setOrigAlign(OriginalAlignment); 7040 7041 MVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT); 7042 unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT); 7043 SmallVector<SDValue, 4> Parts(NumParts); 7044 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 7045 7046 if (Args[i].isSExt) 7047 ExtendKind = ISD::SIGN_EXTEND; 7048 else if (Args[i].isZExt) 7049 ExtendKind = ISD::ZERO_EXTEND; 7050 7051 // Conservatively only handle 'returned' on non-vectors for now 7052 if (Args[i].isReturned && !Op.getValueType().isVector()) { 7053 assert(CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues && 7054 "unexpected use of 'returned'"); 7055 // Before passing 'returned' to the target lowering code, ensure that 7056 // either the register MVT and the actual EVT are the same size or that 7057 // the return value and argument are extended in the same way; in these 7058 // cases it's safe to pass the argument register value unchanged as the 7059 // return register value (although it's at the target's option whether 7060 // to do so) 7061 // TODO: allow code generation to take advantage of partially preserved 7062 // registers rather than clobbering the entire register when the 7063 // parameter extension method is not compatible with the return 7064 // extension method 7065 if ((NumParts * PartVT.getSizeInBits() == VT.getSizeInBits()) || 7066 (ExtendKind != ISD::ANY_EXTEND && 7067 CLI.RetSExt == Args[i].isSExt && CLI.RetZExt == Args[i].isZExt)) 7068 Flags.setReturned(); 7069 } 7070 7071 getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts, 7072 PartVT, CLI.CS ? CLI.CS->getInstruction() : 0, ExtendKind); 7073 7074 for (unsigned j = 0; j != NumParts; ++j) { 7075 // if it isn't first piece, alignment must be 1 7076 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(), VT, 7077 i < CLI.NumFixedArgs, 7078 i, j*Parts[j].getValueType().getStoreSize()); 7079 if (NumParts > 1 && j == 0) 7080 MyFlags.Flags.setSplit(); 7081 else if (j != 0) 7082 MyFlags.Flags.setOrigAlign(1); 7083 7084 CLI.Outs.push_back(MyFlags); 7085 CLI.OutVals.push_back(Parts[j]); 7086 } 7087 } 7088 } 7089 7090 SmallVector<SDValue, 4> InVals; 7091 CLI.Chain = LowerCall(CLI, InVals); 7092 7093 // Verify that the target's LowerCall behaved as expected. 7094 assert(CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other && 7095 "LowerCall didn't return a valid chain!"); 7096 assert((!CLI.IsTailCall || InVals.empty()) && 7097 "LowerCall emitted a return value for a tail call!"); 7098 assert((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) && 7099 "LowerCall didn't emit the correct number of values!"); 7100 7101 // For a tail call, the return value is merely live-out and there aren't 7102 // any nodes in the DAG representing it. Return a special value to 7103 // indicate that a tail call has been emitted and no more Instructions 7104 // should be processed in the current block. 7105 if (CLI.IsTailCall) { 7106 CLI.DAG.setRoot(CLI.Chain); 7107 return std::make_pair(SDValue(), SDValue()); 7108 } 7109 7110 DEBUG(for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++i) { 7111 assert(InVals[i].getNode() && 7112 "LowerCall emitted a null value!"); 7113 assert(EVT(CLI.Ins[i].VT) == InVals[i].getValueType() && 7114 "LowerCall emitted a value with the wrong type!"); 7115 }); 7116 7117 // Collect the legal value parts into potentially illegal values 7118 // that correspond to the original function's return values. 7119 ISD::NodeType AssertOp = ISD::DELETED_NODE; 7120 if (CLI.RetSExt) 7121 AssertOp = ISD::AssertSext; 7122 else if (CLI.RetZExt) 7123 AssertOp = ISD::AssertZext; 7124 SmallVector<SDValue, 4> ReturnValues; 7125 unsigned CurReg = 0; 7126 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) { 7127 EVT VT = RetTys[I]; 7128 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT); 7129 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT); 7130 7131 ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg], 7132 NumRegs, RegisterVT, VT, NULL, 7133 AssertOp)); 7134 CurReg += NumRegs; 7135 } 7136 7137 // For a function returning void, there is no return value. We can't create 7138 // such a node, so we just return a null return value in that case. In 7139 // that case, nothing will actually look at the value. 7140 if (ReturnValues.empty()) 7141 return std::make_pair(SDValue(), CLI.Chain); 7142 7143 SDValue Res = CLI.DAG.getNode(ISD::MERGE_VALUES, CLI.DL, 7144 CLI.DAG.getVTList(&RetTys[0], RetTys.size()), 7145 &ReturnValues[0], ReturnValues.size()); 7146 return std::make_pair(Res, CLI.Chain); 7147} 7148 7149void TargetLowering::LowerOperationWrapper(SDNode *N, 7150 SmallVectorImpl<SDValue> &Results, 7151 SelectionDAG &DAG) const { 7152 SDValue Res = LowerOperation(SDValue(N, 0), DAG); 7153 if (Res.getNode()) 7154 Results.push_back(Res); 7155} 7156 7157SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { 7158 llvm_unreachable("LowerOperation not implemented for this target!"); 7159} 7160 7161void 7162SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) { 7163 SDValue Op = getNonRegisterValue(V); 7164 assert((Op.getOpcode() != ISD::CopyFromReg || 7165 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) && 7166 "Copy from a reg to the same reg!"); 7167 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg"); 7168 7169 const TargetLowering *TLI = TM.getTargetLowering(); 7170 RegsForValue RFV(V->getContext(), *TLI, Reg, V->getType()); 7171 SDValue Chain = DAG.getEntryNode(); 7172 RFV.getCopyToRegs(Op, DAG, getCurSDLoc(), Chain, 0, V); 7173 PendingExports.push_back(Chain); 7174} 7175 7176#include "llvm/CodeGen/SelectionDAGISel.h" 7177 7178/// isOnlyUsedInEntryBlock - If the specified argument is only used in the 7179/// entry block, return true. This includes arguments used by switches, since 7180/// the switch may expand into multiple basic blocks. 7181static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) { 7182 // With FastISel active, we may be splitting blocks, so force creation 7183 // of virtual registers for all non-dead arguments. 7184 if (FastISel) 7185 return A->use_empty(); 7186 7187 const BasicBlock *Entry = A->getParent()->begin(); 7188 for (Value::const_use_iterator UI = A->use_begin(), E = A->use_end(); 7189 UI != E; ++UI) { 7190 const User *U = *UI; 7191 if (cast<Instruction>(U)->getParent() != Entry || isa<SwitchInst>(U)) 7192 return false; // Use not in entry block. 7193 } 7194 return true; 7195} 7196 7197void SelectionDAGISel::LowerArguments(const Function &F) { 7198 SelectionDAG &DAG = SDB->DAG; 7199 SDLoc dl = SDB->getCurSDLoc(); 7200 const TargetLowering *TLI = getTargetLowering(); 7201 const DataLayout *TD = TLI->getDataLayout(); 7202 SmallVector<ISD::InputArg, 16> Ins; 7203 7204 if (!FuncInfo->CanLowerReturn) { 7205 // Put in an sret pointer parameter before all the other parameters. 7206 SmallVector<EVT, 1> ValueVTs; 7207 ComputeValueVTs(*getTargetLowering(), 7208 PointerType::getUnqual(F.getReturnType()), ValueVTs); 7209 7210 // NOTE: Assuming that a pointer will never break down to more than one VT 7211 // or one register. 7212 ISD::ArgFlagsTy Flags; 7213 Flags.setSRet(); 7214 MVT RegisterVT = TLI->getRegisterType(*DAG.getContext(), ValueVTs[0]); 7215 ISD::InputArg RetArg(Flags, RegisterVT, ValueVTs[0], true, 0, 0); 7216 Ins.push_back(RetArg); 7217 } 7218 7219 // Set up the incoming argument description vector. 7220 unsigned Idx = 1; 7221 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); 7222 I != E; ++I, ++Idx) { 7223 SmallVector<EVT, 4> ValueVTs; 7224 ComputeValueVTs(*TLI, I->getType(), ValueVTs); 7225 bool isArgValueUsed = !I->use_empty(); 7226 unsigned PartBase = 0; 7227 for (unsigned Value = 0, NumValues = ValueVTs.size(); 7228 Value != NumValues; ++Value) { 7229 EVT VT = ValueVTs[Value]; 7230 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext()); 7231 ISD::ArgFlagsTy Flags; 7232 unsigned OriginalAlignment = 7233 TD->getABITypeAlignment(ArgTy); 7234 7235 if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt)) 7236 Flags.setZExt(); 7237 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt)) 7238 Flags.setSExt(); 7239 if (F.getAttributes().hasAttribute(Idx, Attribute::InReg)) 7240 Flags.setInReg(); 7241 if (F.getAttributes().hasAttribute(Idx, Attribute::StructRet)) 7242 Flags.setSRet(); 7243 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal)) { 7244 Flags.setByVal(); 7245 PointerType *Ty = cast<PointerType>(I->getType()); 7246 Type *ElementTy = Ty->getElementType(); 7247 Flags.setByValSize(TD->getTypeAllocSize(ElementTy)); 7248 // For ByVal, alignment should be passed from FE. BE will guess if 7249 // this info is not there but there are cases it cannot get right. 7250 unsigned FrameAlign; 7251 if (F.getParamAlignment(Idx)) 7252 FrameAlign = F.getParamAlignment(Idx); 7253 else 7254 FrameAlign = TLI->getByValTypeAlignment(ElementTy); 7255 Flags.setByValAlign(FrameAlign); 7256 } 7257 if (F.getAttributes().hasAttribute(Idx, Attribute::Nest)) 7258 Flags.setNest(); 7259 Flags.setOrigAlign(OriginalAlignment); 7260 7261 MVT RegisterVT = TLI->getRegisterType(*CurDAG->getContext(), VT); 7262 unsigned NumRegs = TLI->getNumRegisters(*CurDAG->getContext(), VT); 7263 for (unsigned i = 0; i != NumRegs; ++i) { 7264 ISD::InputArg MyFlags(Flags, RegisterVT, VT, isArgValueUsed, 7265 Idx-1, PartBase+i*RegisterVT.getStoreSize()); 7266 if (NumRegs > 1 && i == 0) 7267 MyFlags.Flags.setSplit(); 7268 // if it isn't first piece, alignment must be 1 7269 else if (i > 0) 7270 MyFlags.Flags.setOrigAlign(1); 7271 Ins.push_back(MyFlags); 7272 } 7273 PartBase += VT.getStoreSize(); 7274 } 7275 } 7276 7277 // Call the target to set up the argument values. 7278 SmallVector<SDValue, 8> InVals; 7279 SDValue NewRoot = TLI->LowerFormalArguments(DAG.getRoot(), F.getCallingConv(), 7280 F.isVarArg(), Ins, 7281 dl, DAG, InVals); 7282 7283 // Verify that the target's LowerFormalArguments behaved as expected. 7284 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other && 7285 "LowerFormalArguments didn't return a valid chain!"); 7286 assert(InVals.size() == Ins.size() && 7287 "LowerFormalArguments didn't emit the correct number of values!"); 7288 DEBUG({ 7289 for (unsigned i = 0, e = Ins.size(); i != e; ++i) { 7290 assert(InVals[i].getNode() && 7291 "LowerFormalArguments emitted a null value!"); 7292 assert(EVT(Ins[i].VT) == InVals[i].getValueType() && 7293 "LowerFormalArguments emitted a value with the wrong type!"); 7294 } 7295 }); 7296 7297 // Update the DAG with the new chain value resulting from argument lowering. 7298 DAG.setRoot(NewRoot); 7299 7300 // Set up the argument values. 7301 unsigned i = 0; 7302 Idx = 1; 7303 if (!FuncInfo->CanLowerReturn) { 7304 // Create a virtual register for the sret pointer, and put in a copy 7305 // from the sret argument into it. 7306 SmallVector<EVT, 1> ValueVTs; 7307 ComputeValueVTs(*TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs); 7308 MVT VT = ValueVTs[0].getSimpleVT(); 7309 MVT RegVT = TLI->getRegisterType(*CurDAG->getContext(), VT); 7310 ISD::NodeType AssertOp = ISD::DELETED_NODE; 7311 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1, 7312 RegVT, VT, NULL, AssertOp); 7313 7314 MachineFunction& MF = SDB->DAG.getMachineFunction(); 7315 MachineRegisterInfo& RegInfo = MF.getRegInfo(); 7316 unsigned SRetReg = RegInfo.createVirtualRegister(TLI->getRegClassFor(RegVT)); 7317 FuncInfo->DemoteRegister = SRetReg; 7318 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurSDLoc(), 7319 SRetReg, ArgValue); 7320 DAG.setRoot(NewRoot); 7321 7322 // i indexes lowered arguments. Bump it past the hidden sret argument. 7323 // Idx indexes LLVM arguments. Don't touch it. 7324 ++i; 7325 } 7326 7327 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; 7328 ++I, ++Idx) { 7329 SmallVector<SDValue, 4> ArgValues; 7330 SmallVector<EVT, 4> ValueVTs; 7331 ComputeValueVTs(*TLI, I->getType(), ValueVTs); 7332 unsigned NumValues = ValueVTs.size(); 7333 7334 // If this argument is unused then remember its value. It is used to generate 7335 // debugging information. 7336 if (I->use_empty() && NumValues) { 7337 SDB->setUnusedArgValue(I, InVals[i]); 7338 7339 // Also remember any frame index for use in FastISel. 7340 if (FrameIndexSDNode *FI = 7341 dyn_cast<FrameIndexSDNode>(InVals[i].getNode())) 7342 FuncInfo->setArgumentFrameIndex(I, FI->getIndex()); 7343 } 7344 7345 for (unsigned Val = 0; Val != NumValues; ++Val) { 7346 EVT VT = ValueVTs[Val]; 7347 MVT PartVT = TLI->getRegisterType(*CurDAG->getContext(), VT); 7348 unsigned NumParts = TLI->getNumRegisters(*CurDAG->getContext(), VT); 7349 7350 if (!I->use_empty()) { 7351 ISD::NodeType AssertOp = ISD::DELETED_NODE; 7352 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt)) 7353 AssertOp = ISD::AssertSext; 7354 else if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt)) 7355 AssertOp = ISD::AssertZext; 7356 7357 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i], 7358 NumParts, PartVT, VT, 7359 NULL, AssertOp)); 7360 } 7361 7362 i += NumParts; 7363 } 7364 7365 // We don't need to do anything else for unused arguments. 7366 if (ArgValues.empty()) 7367 continue; 7368 7369 // Note down frame index. 7370 if (FrameIndexSDNode *FI = 7371 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode())) 7372 FuncInfo->setArgumentFrameIndex(I, FI->getIndex()); 7373 7374 SDValue Res = DAG.getMergeValues(&ArgValues[0], NumValues, 7375 SDB->getCurSDLoc()); 7376 7377 SDB->setValue(I, Res); 7378 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) { 7379 if (LoadSDNode *LNode = 7380 dyn_cast<LoadSDNode>(Res.getOperand(0).getNode())) 7381 if (FrameIndexSDNode *FI = 7382 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode())) 7383 FuncInfo->setArgumentFrameIndex(I, FI->getIndex()); 7384 } 7385 7386 // If this argument is live outside of the entry block, insert a copy from 7387 // wherever we got it to the vreg that other BB's will reference it as. 7388 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) { 7389 // If we can, though, try to skip creating an unnecessary vreg. 7390 // FIXME: This isn't very clean... it would be nice to make this more 7391 // general. It's also subtly incompatible with the hacks FastISel 7392 // uses with vregs. 7393 unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg(); 7394 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 7395 FuncInfo->ValueMap[I] = Reg; 7396 continue; 7397 } 7398 } 7399 if (!isOnlyUsedInEntryBlock(I, TM.Options.EnableFastISel)) { 7400 FuncInfo->InitializeRegForValue(I); 7401 SDB->CopyToExportRegsIfNeeded(I); 7402 } 7403 } 7404 7405 assert(i == InVals.size() && "Argument register count mismatch!"); 7406 7407 // Finally, if the target has anything special to do, allow it to do so. 7408 // FIXME: this should insert code into the DAG! 7409 EmitFunctionEntryCode(); 7410} 7411 7412/// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to 7413/// ensure constants are generated when needed. Remember the virtual registers 7414/// that need to be added to the Machine PHI nodes as input. We cannot just 7415/// directly add them, because expansion might result in multiple MBB's for one 7416/// BB. As such, the start of the BB might correspond to a different MBB than 7417/// the end. 7418/// 7419void 7420SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) { 7421 const TerminatorInst *TI = LLVMBB->getTerminator(); 7422 7423 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled; 7424 7425 // Check successor nodes' PHI nodes that expect a constant to be available 7426 // from this block. 7427 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) { 7428 const BasicBlock *SuccBB = TI->getSuccessor(succ); 7429 if (!isa<PHINode>(SuccBB->begin())) continue; 7430 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB]; 7431 7432 // If this terminator has multiple identical successors (common for 7433 // switches), only handle each succ once. 7434 if (!SuccsHandled.insert(SuccMBB)) continue; 7435 7436 MachineBasicBlock::iterator MBBI = SuccMBB->begin(); 7437 7438 // At this point we know that there is a 1-1 correspondence between LLVM PHI 7439 // nodes and Machine PHI nodes, but the incoming operands have not been 7440 // emitted yet. 7441 for (BasicBlock::const_iterator I = SuccBB->begin(); 7442 const PHINode *PN = dyn_cast<PHINode>(I); ++I) { 7443 // Ignore dead phi's. 7444 if (PN->use_empty()) continue; 7445 7446 // Skip empty types 7447 if (PN->getType()->isEmptyTy()) 7448 continue; 7449 7450 unsigned Reg; 7451 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB); 7452 7453 if (const Constant *C = dyn_cast<Constant>(PHIOp)) { 7454 unsigned &RegOut = ConstantsOut[C]; 7455 if (RegOut == 0) { 7456 RegOut = FuncInfo.CreateRegs(C->getType()); 7457 CopyValueToVirtualRegister(C, RegOut); 7458 } 7459 Reg = RegOut; 7460 } else { 7461 DenseMap<const Value *, unsigned>::iterator I = 7462 FuncInfo.ValueMap.find(PHIOp); 7463 if (I != FuncInfo.ValueMap.end()) 7464 Reg = I->second; 7465 else { 7466 assert(isa<AllocaInst>(PHIOp) && 7467 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) && 7468 "Didn't codegen value into a register!??"); 7469 Reg = FuncInfo.CreateRegs(PHIOp->getType()); 7470 CopyValueToVirtualRegister(PHIOp, Reg); 7471 } 7472 } 7473 7474 // Remember that this register needs to added to the machine PHI node as 7475 // the input for this MBB. 7476 SmallVector<EVT, 4> ValueVTs; 7477 const TargetLowering *TLI = TM.getTargetLowering(); 7478 ComputeValueVTs(*TLI, PN->getType(), ValueVTs); 7479 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) { 7480 EVT VT = ValueVTs[vti]; 7481 unsigned NumRegisters = TLI->getNumRegisters(*DAG.getContext(), VT); 7482 for (unsigned i = 0, e = NumRegisters; i != e; ++i) 7483 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i)); 7484 Reg += NumRegisters; 7485 } 7486 } 7487 } 7488 7489 ConstantsOut.clear(); 7490} 7491 7492/// Add a successor MBB to ParentMBB< creating a new MachineBB for BB if SuccMBB 7493/// is 0. 7494MachineBasicBlock * 7495SelectionDAGBuilder::StackProtectorDescriptor:: 7496AddSuccessorMBB(const BasicBlock *BB, 7497 MachineBasicBlock *ParentMBB, 7498 MachineBasicBlock *SuccMBB) { 7499 // If SuccBB has not been created yet, create it. 7500 if (!SuccMBB) { 7501 MachineFunction *MF = ParentMBB->getParent(); 7502 MachineFunction::iterator BBI = ParentMBB; 7503 SuccMBB = MF->CreateMachineBasicBlock(BB); 7504 MF->insert(++BBI, SuccMBB); 7505 } 7506 // Add it as a successor of ParentMBB. 7507 ParentMBB->addSuccessor(SuccMBB); 7508 return SuccMBB; 7509} 7510