SelectionDAG.cpp revision b5a55d979c84e100f383ce11c5c5e3b616b9dcc4
17839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger//===-- SelectionDAG.cpp - Implement the SelectionDAG data structures -----===// 27839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger// 37839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger// The LLVM Compiler Infrastructure 47839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger// 57839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger// This file is distributed under the University of Illinois Open Source 67839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger// License. See LICENSE.TXT for details. 77839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger// 87839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger//===----------------------------------------------------------------------===// 97839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger// 107839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger// This implements the SelectionDAG class. 117839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger// 127839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger//===----------------------------------------------------------------------===// 137839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 147839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/CodeGen/SelectionDAG.h" 157839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "SDNodeOrdering.h" 167839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "SDNodeDbgValue.h" 177839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Constants.h" 187839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Analysis/DebugInfo.h" 197839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Analysis/ValueTracking.h" 207839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Function.h" 217839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/GlobalAlias.h" 227839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/GlobalVariable.h" 237839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Intrinsics.h" 247839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/DerivedTypes.h" 257839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Assembly/Writer.h" 267839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/CallingConv.h" 277839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/CodeGen/MachineBasicBlock.h" 287839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/CodeGen/MachineConstantPool.h" 297839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/CodeGen/MachineFrameInfo.h" 307839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/CodeGen/MachineModuleInfo.h" 317839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/CodeGen/PseudoSourceValue.h" 327839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Target/TargetRegisterInfo.h" 337839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Target/TargetData.h" 347839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Target/TargetLowering.h" 357839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Target/TargetSelectionDAGInfo.h" 367839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Target/TargetOptions.h" 377839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Target/TargetInstrInfo.h" 387839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Target/TargetIntrinsicInfo.h" 397839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Target/TargetMachine.h" 407839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Support/CommandLine.h" 417839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Support/Debug.h" 427839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Support/ErrorHandling.h" 437839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Support/ManagedStatic.h" 447839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Support/MathExtras.h" 457839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Support/raw_ostream.h" 467839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/Support/Mutex.h" 477839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/ADT/SetVector.h" 487839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/ADT/SmallPtrSet.h" 497839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/ADT/SmallSet.h" 507839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/ADT/SmallVector.h" 517839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include "llvm/ADT/StringExtras.h" 527839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include <algorithm> 537839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger#include <cmath> 547839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenbergerusing namespace llvm; 557839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 567839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger/// makeVTList - Return an instance of the SDVTList struct initialized with the 577839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger/// specified members. 587839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenbergerstatic SDVTList makeVTList(const EVT *VTs, unsigned NumVTs) { 597839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger SDVTList Res = {VTs, NumVTs}; 607839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger return Res; 617839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger} 627839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 637839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenbergerstatic const fltSemantics *EVTToAPFloatSemantics(EVT VT) { 647839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger switch (VT.getSimpleVT().SimpleTy) { 657839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger default: llvm_unreachable("Unknown FP format"); 667839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger case MVT::f32: return &APFloat::IEEEsingle; 677839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger case MVT::f64: return &APFloat::IEEEdouble; 687839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger case MVT::f80: return &APFloat::x87DoubleExtended; 697839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger case MVT::f128: return &APFloat::IEEEquad; 707839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger case MVT::ppcf128: return &APFloat::PPCDoubleDouble; 717839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger } 727839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger} 737839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 747839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek SollenbergerSelectionDAG::DAGUpdateListener::~DAGUpdateListener() {} 757839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 767839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger//===----------------------------------------------------------------------===// 777839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger// ConstantFPSDNode Class 787839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger//===----------------------------------------------------------------------===// 797839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 807839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger/// isExactlyValue - We don't rely on operator== working on double values, as 817839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger/// it returns true for things that are clearly not equal, like -0.0 and 0.0. 827839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger/// As such, this method can be used to do an exact bit-for-bit comparison of 837839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger/// two floating point values. 847839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenbergerbool ConstantFPSDNode::isExactlyValue(const APFloat& V) const { 857839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger return getValueAPF().bitwiseIsEqual(V); 867839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger} 877839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 887839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenbergerbool ConstantFPSDNode::isValueValidForType(EVT VT, 897839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger const APFloat& Val) { 907839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger assert(VT.isFloatingPoint() && "Can only convert between FP types"); 917839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 927839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger // PPC long double cannot be converted to any other type. 930a657bbc2c6fc9daf699942e023050536d5ec95fDerek Sollenberger if (VT == MVT::ppcf128 || 947839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger &Val.getSemantics() == &APFloat::PPCDoubleDouble) 957839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger return false; 967839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 977839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger // convert modifies in place, so make a copy. 987839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger APFloat Val2 = APFloat(Val); 997839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger bool losesInfo; 1007839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger (void) Val2.convert(*EVTToAPFloatSemantics(VT), APFloat::rmNearestTiesToEven, 1017839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger &losesInfo); 1027839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger return !losesInfo; 1037839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger} 1047839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1057839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger//===----------------------------------------------------------------------===// 1067839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger// ISD Namespace 1077839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger//===----------------------------------------------------------------------===// 1087839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1097839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger/// isBuildVectorAllOnes - Return true if the specified node is a 1107839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger/// BUILD_VECTOR where all of the elements are ~0 or undef. 1117839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenbergerbool ISD::isBuildVectorAllOnes(const SDNode *N) { 1127839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger // Look through a bit convert. 1137839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger if (N->getOpcode() == ISD::BITCAST) 1147839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger N = N->getOperand(0).getNode(); 1157839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1167839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger if (N->getOpcode() != ISD::BUILD_VECTOR) return false; 1177839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1187839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger unsigned i = 0, e = N->getNumOperands(); 1197839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1207839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger // Skip over all of the undef values. 1217839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF) 1227839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger ++i; 1237839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1247839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger // Do not accept an all-undef vector. 1257839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger if (i == e) return false; 1267839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1277839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger // Do not accept build_vectors that aren't all constants or which have non-~0 1287839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger // elements. 1297839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger SDValue NotZero = N->getOperand(i); 1307839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger if (isa<ConstantSDNode>(NotZero)) { 1317839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger if (!cast<ConstantSDNode>(NotZero)->isAllOnesValue()) 1327839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger return false; 1337839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger } else if (isa<ConstantFPSDNode>(NotZero)) { 1347839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger if (!cast<ConstantFPSDNode>(NotZero)->getValueAPF(). 1357839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger bitcastToAPInt().isAllOnesValue()) 1367839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger return false; 1377839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger } else 1387839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger return false; 1397839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1407839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger // Okay, we have at least one ~0 value, check to see if the rest match or are 1417839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger // undefs. 1427839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger for (++i; i != e; ++i) 1437839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger if (N->getOperand(i) != NotZero && 1447839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger N->getOperand(i).getOpcode() != ISD::UNDEF) 1457839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger return false; 1467839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger return true; 1477839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger} 1487839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1497839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1507839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger/// isBuildVectorAllZeros - Return true if the specified node is a 1517839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger/// BUILD_VECTOR where all of the elements are 0 or undef. 1527839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenbergerbool ISD::isBuildVectorAllZeros(const SDNode *N) { 1537839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger // Look through a bit convert. 1547839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger if (N->getOpcode() == ISD::BITCAST) 1557839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger N = N->getOperand(0).getNode(); 1567839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1577839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger if (N->getOpcode() != ISD::BUILD_VECTOR) return false; 1587839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1597839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger unsigned i = 0, e = N->getNumOperands(); 1607839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1617839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger // Skip over all of the undef values. 1627839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF) 1637839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger ++i; 1647839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger 1657839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger // Do not accept an all-undef vector. 1667839ce1af63bf12fe7b3caa866970bbbb3afb13dDerek Sollenberger if (i == e) return false; 167 168 // Do not accept build_vectors that aren't all constants or which have non-0 169 // elements. 170 SDValue Zero = N->getOperand(i); 171 if (isa<ConstantSDNode>(Zero)) { 172 if (!cast<ConstantSDNode>(Zero)->isNullValue()) 173 return false; 174 } else if (isa<ConstantFPSDNode>(Zero)) { 175 if (!cast<ConstantFPSDNode>(Zero)->getValueAPF().isPosZero()) 176 return false; 177 } else 178 return false; 179 180 // Okay, we have at least one 0 value, check to see if the rest match or are 181 // undefs. 182 for (++i; i != e; ++i) 183 if (N->getOperand(i) != Zero && 184 N->getOperand(i).getOpcode() != ISD::UNDEF) 185 return false; 186 return true; 187} 188 189/// isScalarToVector - Return true if the specified node is a 190/// ISD::SCALAR_TO_VECTOR node or a BUILD_VECTOR node where only the low 191/// element is not an undef. 192bool ISD::isScalarToVector(const SDNode *N) { 193 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) 194 return true; 195 196 if (N->getOpcode() != ISD::BUILD_VECTOR) 197 return false; 198 if (N->getOperand(0).getOpcode() == ISD::UNDEF) 199 return false; 200 unsigned NumElems = N->getNumOperands(); 201 if (NumElems == 1) 202 return false; 203 for (unsigned i = 1; i < NumElems; ++i) { 204 SDValue V = N->getOperand(i); 205 if (V.getOpcode() != ISD::UNDEF) 206 return false; 207 } 208 return true; 209} 210 211/// getSetCCSwappedOperands - Return the operation corresponding to (Y op X) 212/// when given the operation for (X op Y). 213ISD::CondCode ISD::getSetCCSwappedOperands(ISD::CondCode Operation) { 214 // To perform this operation, we just need to swap the L and G bits of the 215 // operation. 216 unsigned OldL = (Operation >> 2) & 1; 217 unsigned OldG = (Operation >> 1) & 1; 218 return ISD::CondCode((Operation & ~6) | // Keep the N, U, E bits 219 (OldL << 1) | // New G bit 220 (OldG << 2)); // New L bit. 221} 222 223/// getSetCCInverse - Return the operation corresponding to !(X op Y), where 224/// 'op' is a valid SetCC operation. 225ISD::CondCode ISD::getSetCCInverse(ISD::CondCode Op, bool isInteger) { 226 unsigned Operation = Op; 227 if (isInteger) 228 Operation ^= 7; // Flip L, G, E bits, but not U. 229 else 230 Operation ^= 15; // Flip all of the condition bits. 231 232 if (Operation > ISD::SETTRUE2) 233 Operation &= ~8; // Don't let N and U bits get set. 234 235 return ISD::CondCode(Operation); 236} 237 238 239/// isSignedOp - For an integer comparison, return 1 if the comparison is a 240/// signed operation and 2 if the result is an unsigned comparison. Return zero 241/// if the operation does not depend on the sign of the input (setne and seteq). 242static int isSignedOp(ISD::CondCode Opcode) { 243 switch (Opcode) { 244 default: llvm_unreachable("Illegal integer setcc operation!"); 245 case ISD::SETEQ: 246 case ISD::SETNE: return 0; 247 case ISD::SETLT: 248 case ISD::SETLE: 249 case ISD::SETGT: 250 case ISD::SETGE: return 1; 251 case ISD::SETULT: 252 case ISD::SETULE: 253 case ISD::SETUGT: 254 case ISD::SETUGE: return 2; 255 } 256} 257 258/// getSetCCOrOperation - Return the result of a logical OR between different 259/// comparisons of identical values: ((X op1 Y) | (X op2 Y)). This function 260/// returns SETCC_INVALID if it is not possible to represent the resultant 261/// comparison. 262ISD::CondCode ISD::getSetCCOrOperation(ISD::CondCode Op1, ISD::CondCode Op2, 263 bool isInteger) { 264 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3) 265 // Cannot fold a signed integer setcc with an unsigned integer setcc. 266 return ISD::SETCC_INVALID; 267 268 unsigned Op = Op1 | Op2; // Combine all of the condition bits. 269 270 // If the N and U bits get set then the resultant comparison DOES suddenly 271 // care about orderedness, and is true when ordered. 272 if (Op > ISD::SETTRUE2) 273 Op &= ~16; // Clear the U bit if the N bit is set. 274 275 // Canonicalize illegal integer setcc's. 276 if (isInteger && Op == ISD::SETUNE) // e.g. SETUGT | SETULT 277 Op = ISD::SETNE; 278 279 return ISD::CondCode(Op); 280} 281 282/// getSetCCAndOperation - Return the result of a logical AND between different 283/// comparisons of identical values: ((X op1 Y) & (X op2 Y)). This 284/// function returns zero if it is not possible to represent the resultant 285/// comparison. 286ISD::CondCode ISD::getSetCCAndOperation(ISD::CondCode Op1, ISD::CondCode Op2, 287 bool isInteger) { 288 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3) 289 // Cannot fold a signed setcc with an unsigned setcc. 290 return ISD::SETCC_INVALID; 291 292 // Combine all of the condition bits. 293 ISD::CondCode Result = ISD::CondCode(Op1 & Op2); 294 295 // Canonicalize illegal integer setcc's. 296 if (isInteger) { 297 switch (Result) { 298 default: break; 299 case ISD::SETUO : Result = ISD::SETFALSE; break; // SETUGT & SETULT 300 case ISD::SETOEQ: // SETEQ & SETU[LG]E 301 case ISD::SETUEQ: Result = ISD::SETEQ ; break; // SETUGE & SETULE 302 case ISD::SETOLT: Result = ISD::SETULT ; break; // SETULT & SETNE 303 case ISD::SETOGT: Result = ISD::SETUGT ; break; // SETUGT & SETNE 304 } 305 } 306 307 return Result; 308} 309 310//===----------------------------------------------------------------------===// 311// SDNode Profile Support 312//===----------------------------------------------------------------------===// 313 314/// AddNodeIDOpcode - Add the node opcode to the NodeID data. 315/// 316static void AddNodeIDOpcode(FoldingSetNodeID &ID, unsigned OpC) { 317 ID.AddInteger(OpC); 318} 319 320/// AddNodeIDValueTypes - Value type lists are intern'd so we can represent them 321/// solely with their pointer. 322static void AddNodeIDValueTypes(FoldingSetNodeID &ID, SDVTList VTList) { 323 ID.AddPointer(VTList.VTs); 324} 325 326/// AddNodeIDOperands - Various routines for adding operands to the NodeID data. 327/// 328static void AddNodeIDOperands(FoldingSetNodeID &ID, 329 const SDValue *Ops, unsigned NumOps) { 330 for (; NumOps; --NumOps, ++Ops) { 331 ID.AddPointer(Ops->getNode()); 332 ID.AddInteger(Ops->getResNo()); 333 } 334} 335 336/// AddNodeIDOperands - Various routines for adding operands to the NodeID data. 337/// 338static void AddNodeIDOperands(FoldingSetNodeID &ID, 339 const SDUse *Ops, unsigned NumOps) { 340 for (; NumOps; --NumOps, ++Ops) { 341 ID.AddPointer(Ops->getNode()); 342 ID.AddInteger(Ops->getResNo()); 343 } 344} 345 346static void AddNodeIDNode(FoldingSetNodeID &ID, 347 unsigned short OpC, SDVTList VTList, 348 const SDValue *OpList, unsigned N) { 349 AddNodeIDOpcode(ID, OpC); 350 AddNodeIDValueTypes(ID, VTList); 351 AddNodeIDOperands(ID, OpList, N); 352} 353 354/// AddNodeIDCustom - If this is an SDNode with special info, add this info to 355/// the NodeID data. 356static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) { 357 switch (N->getOpcode()) { 358 case ISD::TargetExternalSymbol: 359 case ISD::ExternalSymbol: 360 llvm_unreachable("Should only be used on nodes with operands"); 361 default: break; // Normal nodes don't need extra info. 362 case ISD::TargetConstant: 363 case ISD::Constant: 364 ID.AddPointer(cast<ConstantSDNode>(N)->getConstantIntValue()); 365 break; 366 case ISD::TargetConstantFP: 367 case ISD::ConstantFP: { 368 ID.AddPointer(cast<ConstantFPSDNode>(N)->getConstantFPValue()); 369 break; 370 } 371 case ISD::TargetGlobalAddress: 372 case ISD::GlobalAddress: 373 case ISD::TargetGlobalTLSAddress: 374 case ISD::GlobalTLSAddress: { 375 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N); 376 ID.AddPointer(GA->getGlobal()); 377 ID.AddInteger(GA->getOffset()); 378 ID.AddInteger(GA->getTargetFlags()); 379 break; 380 } 381 case ISD::BasicBlock: 382 ID.AddPointer(cast<BasicBlockSDNode>(N)->getBasicBlock()); 383 break; 384 case ISD::Register: 385 ID.AddInteger(cast<RegisterSDNode>(N)->getReg()); 386 break; 387 388 case ISD::SRCVALUE: 389 ID.AddPointer(cast<SrcValueSDNode>(N)->getValue()); 390 break; 391 case ISD::FrameIndex: 392 case ISD::TargetFrameIndex: 393 ID.AddInteger(cast<FrameIndexSDNode>(N)->getIndex()); 394 break; 395 case ISD::JumpTable: 396 case ISD::TargetJumpTable: 397 ID.AddInteger(cast<JumpTableSDNode>(N)->getIndex()); 398 ID.AddInteger(cast<JumpTableSDNode>(N)->getTargetFlags()); 399 break; 400 case ISD::ConstantPool: 401 case ISD::TargetConstantPool: { 402 const ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N); 403 ID.AddInteger(CP->getAlignment()); 404 ID.AddInteger(CP->getOffset()); 405 if (CP->isMachineConstantPoolEntry()) 406 CP->getMachineCPVal()->AddSelectionDAGCSEId(ID); 407 else 408 ID.AddPointer(CP->getConstVal()); 409 ID.AddInteger(CP->getTargetFlags()); 410 break; 411 } 412 case ISD::LOAD: { 413 const LoadSDNode *LD = cast<LoadSDNode>(N); 414 ID.AddInteger(LD->getMemoryVT().getRawBits()); 415 ID.AddInteger(LD->getRawSubclassData()); 416 break; 417 } 418 case ISD::STORE: { 419 const StoreSDNode *ST = cast<StoreSDNode>(N); 420 ID.AddInteger(ST->getMemoryVT().getRawBits()); 421 ID.AddInteger(ST->getRawSubclassData()); 422 break; 423 } 424 case ISD::ATOMIC_CMP_SWAP: 425 case ISD::ATOMIC_SWAP: 426 case ISD::ATOMIC_LOAD_ADD: 427 case ISD::ATOMIC_LOAD_SUB: 428 case ISD::ATOMIC_LOAD_AND: 429 case ISD::ATOMIC_LOAD_OR: 430 case ISD::ATOMIC_LOAD_XOR: 431 case ISD::ATOMIC_LOAD_NAND: 432 case ISD::ATOMIC_LOAD_MIN: 433 case ISD::ATOMIC_LOAD_MAX: 434 case ISD::ATOMIC_LOAD_UMIN: 435 case ISD::ATOMIC_LOAD_UMAX: { 436 const AtomicSDNode *AT = cast<AtomicSDNode>(N); 437 ID.AddInteger(AT->getMemoryVT().getRawBits()); 438 ID.AddInteger(AT->getRawSubclassData()); 439 break; 440 } 441 case ISD::VECTOR_SHUFFLE: { 442 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 443 for (unsigned i = 0, e = N->getValueType(0).getVectorNumElements(); 444 i != e; ++i) 445 ID.AddInteger(SVN->getMaskElt(i)); 446 break; 447 } 448 case ISD::TargetBlockAddress: 449 case ISD::BlockAddress: { 450 ID.AddPointer(cast<BlockAddressSDNode>(N)->getBlockAddress()); 451 ID.AddInteger(cast<BlockAddressSDNode>(N)->getTargetFlags()); 452 break; 453 } 454 } // end switch (N->getOpcode()) 455} 456 457/// AddNodeIDNode - Generic routine for adding a nodes info to the NodeID 458/// data. 459static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) { 460 AddNodeIDOpcode(ID, N->getOpcode()); 461 // Add the return value info. 462 AddNodeIDValueTypes(ID, N->getVTList()); 463 // Add the operand info. 464 AddNodeIDOperands(ID, N->op_begin(), N->getNumOperands()); 465 466 // Handle SDNode leafs with special info. 467 AddNodeIDCustom(ID, N); 468} 469 470/// encodeMemSDNodeFlags - Generic routine for computing a value for use in 471/// the CSE map that carries volatility, temporalness, indexing mode, and 472/// extension/truncation information. 473/// 474static inline unsigned 475encodeMemSDNodeFlags(int ConvType, ISD::MemIndexedMode AM, bool isVolatile, 476 bool isNonTemporal) { 477 assert((ConvType & 3) == ConvType && 478 "ConvType may not require more than 2 bits!"); 479 assert((AM & 7) == AM && 480 "AM may not require more than 3 bits!"); 481 return ConvType | 482 (AM << 2) | 483 (isVolatile << 5) | 484 (isNonTemporal << 6); 485} 486 487//===----------------------------------------------------------------------===// 488// SelectionDAG Class 489//===----------------------------------------------------------------------===// 490 491/// doNotCSE - Return true if CSE should not be performed for this node. 492static bool doNotCSE(SDNode *N) { 493 if (N->getValueType(0) == MVT::Glue) 494 return true; // Never CSE anything that produces a flag. 495 496 switch (N->getOpcode()) { 497 default: break; 498 case ISD::HANDLENODE: 499 case ISD::EH_LABEL: 500 return true; // Never CSE these nodes. 501 } 502 503 // Check that remaining values produced are not flags. 504 for (unsigned i = 1, e = N->getNumValues(); i != e; ++i) 505 if (N->getValueType(i) == MVT::Glue) 506 return true; // Never CSE anything that produces a flag. 507 508 return false; 509} 510 511/// RemoveDeadNodes - This method deletes all unreachable nodes in the 512/// SelectionDAG. 513void SelectionDAG::RemoveDeadNodes() { 514 // Create a dummy node (which is not added to allnodes), that adds a reference 515 // to the root node, preventing it from being deleted. 516 HandleSDNode Dummy(getRoot()); 517 518 SmallVector<SDNode*, 128> DeadNodes; 519 520 // Add all obviously-dead nodes to the DeadNodes worklist. 521 for (allnodes_iterator I = allnodes_begin(), E = allnodes_end(); I != E; ++I) 522 if (I->use_empty()) 523 DeadNodes.push_back(I); 524 525 RemoveDeadNodes(DeadNodes); 526 527 // If the root changed (e.g. it was a dead load, update the root). 528 setRoot(Dummy.getValue()); 529} 530 531/// RemoveDeadNodes - This method deletes the unreachable nodes in the 532/// given list, and any nodes that become unreachable as a result. 533void SelectionDAG::RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes, 534 DAGUpdateListener *UpdateListener) { 535 536 // Process the worklist, deleting the nodes and adding their uses to the 537 // worklist. 538 while (!DeadNodes.empty()) { 539 SDNode *N = DeadNodes.pop_back_val(); 540 541 if (UpdateListener) 542 UpdateListener->NodeDeleted(N, 0); 543 544 // Take the node out of the appropriate CSE map. 545 RemoveNodeFromCSEMaps(N); 546 547 // Next, brutally remove the operand list. This is safe to do, as there are 548 // no cycles in the graph. 549 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) { 550 SDUse &Use = *I++; 551 SDNode *Operand = Use.getNode(); 552 Use.set(SDValue()); 553 554 // Now that we removed this operand, see if there are no uses of it left. 555 if (Operand->use_empty()) 556 DeadNodes.push_back(Operand); 557 } 558 559 DeallocateNode(N); 560 } 561} 562 563void SelectionDAG::RemoveDeadNode(SDNode *N, DAGUpdateListener *UpdateListener){ 564 SmallVector<SDNode*, 16> DeadNodes(1, N); 565 RemoveDeadNodes(DeadNodes, UpdateListener); 566} 567 568void SelectionDAG::DeleteNode(SDNode *N) { 569 // First take this out of the appropriate CSE map. 570 RemoveNodeFromCSEMaps(N); 571 572 // Finally, remove uses due to operands of this node, remove from the 573 // AllNodes list, and delete the node. 574 DeleteNodeNotInCSEMaps(N); 575} 576 577void SelectionDAG::DeleteNodeNotInCSEMaps(SDNode *N) { 578 assert(N != AllNodes.begin() && "Cannot delete the entry node!"); 579 assert(N->use_empty() && "Cannot delete a node that is not dead!"); 580 581 // Drop all of the operands and decrement used node's use counts. 582 N->DropOperands(); 583 584 DeallocateNode(N); 585} 586 587void SelectionDAG::DeallocateNode(SDNode *N) { 588 if (N->OperandsNeedDelete) 589 delete[] N->OperandList; 590 591 // Set the opcode to DELETED_NODE to help catch bugs when node 592 // memory is reallocated. 593 N->NodeType = ISD::DELETED_NODE; 594 595 NodeAllocator.Deallocate(AllNodes.remove(N)); 596 597 // Remove the ordering of this node. 598 Ordering->remove(N); 599 600 // If any of the SDDbgValue nodes refer to this SDNode, invalidate them. 601 SmallVector<SDDbgValue*, 2> &DbgVals = DbgInfo->getSDDbgValues(N); 602 for (unsigned i = 0, e = DbgVals.size(); i != e; ++i) 603 DbgVals[i]->setIsInvalidated(); 604} 605 606/// RemoveNodeFromCSEMaps - Take the specified node out of the CSE map that 607/// correspond to it. This is useful when we're about to delete or repurpose 608/// the node. We don't want future request for structurally identical nodes 609/// to return N anymore. 610bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) { 611 bool Erased = false; 612 switch (N->getOpcode()) { 613 case ISD::HANDLENODE: return false; // noop. 614 case ISD::CONDCODE: 615 assert(CondCodeNodes[cast<CondCodeSDNode>(N)->get()] && 616 "Cond code doesn't exist!"); 617 Erased = CondCodeNodes[cast<CondCodeSDNode>(N)->get()] != 0; 618 CondCodeNodes[cast<CondCodeSDNode>(N)->get()] = 0; 619 break; 620 case ISD::ExternalSymbol: 621 Erased = ExternalSymbols.erase(cast<ExternalSymbolSDNode>(N)->getSymbol()); 622 break; 623 case ISD::TargetExternalSymbol: { 624 ExternalSymbolSDNode *ESN = cast<ExternalSymbolSDNode>(N); 625 Erased = TargetExternalSymbols.erase( 626 std::pair<std::string,unsigned char>(ESN->getSymbol(), 627 ESN->getTargetFlags())); 628 break; 629 } 630 case ISD::VALUETYPE: { 631 EVT VT = cast<VTSDNode>(N)->getVT(); 632 if (VT.isExtended()) { 633 Erased = ExtendedValueTypeNodes.erase(VT); 634 } else { 635 Erased = ValueTypeNodes[VT.getSimpleVT().SimpleTy] != 0; 636 ValueTypeNodes[VT.getSimpleVT().SimpleTy] = 0; 637 } 638 break; 639 } 640 default: 641 // Remove it from the CSE Map. 642 assert(N->getOpcode() != ISD::DELETED_NODE && "DELETED_NODE in CSEMap!"); 643 assert(N->getOpcode() != ISD::EntryToken && "EntryToken in CSEMap!"); 644 Erased = CSEMap.RemoveNode(N); 645 break; 646 } 647#ifndef NDEBUG 648 // Verify that the node was actually in one of the CSE maps, unless it has a 649 // flag result (which cannot be CSE'd) or is one of the special cases that are 650 // not subject to CSE. 651 if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Glue && 652 !N->isMachineOpcode() && !doNotCSE(N)) { 653 N->dump(this); 654 dbgs() << "\n"; 655 llvm_unreachable("Node is not in map!"); 656 } 657#endif 658 return Erased; 659} 660 661/// AddModifiedNodeToCSEMaps - The specified node has been removed from the CSE 662/// maps and modified in place. Add it back to the CSE maps, unless an identical 663/// node already exists, in which case transfer all its users to the existing 664/// node. This transfer can potentially trigger recursive merging. 665/// 666void 667SelectionDAG::AddModifiedNodeToCSEMaps(SDNode *N, 668 DAGUpdateListener *UpdateListener) { 669 // For node types that aren't CSE'd, just act as if no identical node 670 // already exists. 671 if (!doNotCSE(N)) { 672 SDNode *Existing = CSEMap.GetOrInsertNode(N); 673 if (Existing != N) { 674 // If there was already an existing matching node, use ReplaceAllUsesWith 675 // to replace the dead one with the existing one. This can cause 676 // recursive merging of other unrelated nodes down the line. 677 ReplaceAllUsesWith(N, Existing, UpdateListener); 678 679 // N is now dead. Inform the listener if it exists and delete it. 680 if (UpdateListener) 681 UpdateListener->NodeDeleted(N, Existing); 682 DeleteNodeNotInCSEMaps(N); 683 return; 684 } 685 } 686 687 // If the node doesn't already exist, we updated it. Inform a listener if 688 // it exists. 689 if (UpdateListener) 690 UpdateListener->NodeUpdated(N); 691} 692 693/// FindModifiedNodeSlot - Find a slot for the specified node if its operands 694/// were replaced with those specified. If this node is never memoized, 695/// return null, otherwise return a pointer to the slot it would take. If a 696/// node already exists with these operands, the slot will be non-null. 697SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, SDValue Op, 698 void *&InsertPos) { 699 if (doNotCSE(N)) 700 return 0; 701 702 SDValue Ops[] = { Op }; 703 FoldingSetNodeID ID; 704 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 1); 705 AddNodeIDCustom(ID, N); 706 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos); 707 return Node; 708} 709 710/// FindModifiedNodeSlot - Find a slot for the specified node if its operands 711/// were replaced with those specified. If this node is never memoized, 712/// return null, otherwise return a pointer to the slot it would take. If a 713/// node already exists with these operands, the slot will be non-null. 714SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, 715 SDValue Op1, SDValue Op2, 716 void *&InsertPos) { 717 if (doNotCSE(N)) 718 return 0; 719 720 SDValue Ops[] = { Op1, Op2 }; 721 FoldingSetNodeID ID; 722 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 2); 723 AddNodeIDCustom(ID, N); 724 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos); 725 return Node; 726} 727 728 729/// FindModifiedNodeSlot - Find a slot for the specified node if its operands 730/// were replaced with those specified. If this node is never memoized, 731/// return null, otherwise return a pointer to the slot it would take. If a 732/// node already exists with these operands, the slot will be non-null. 733SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, 734 const SDValue *Ops,unsigned NumOps, 735 void *&InsertPos) { 736 if (doNotCSE(N)) 737 return 0; 738 739 FoldingSetNodeID ID; 740 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, NumOps); 741 AddNodeIDCustom(ID, N); 742 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos); 743 return Node; 744} 745 746#ifndef NDEBUG 747/// VerifyNodeCommon - Sanity check the given node. Aborts if it is invalid. 748static void VerifyNodeCommon(SDNode *N) { 749 switch (N->getOpcode()) { 750 default: 751 break; 752 case ISD::BUILD_PAIR: { 753 EVT VT = N->getValueType(0); 754 assert(N->getNumValues() == 1 && "Too many results!"); 755 assert(!VT.isVector() && (VT.isInteger() || VT.isFloatingPoint()) && 756 "Wrong return type!"); 757 assert(N->getNumOperands() == 2 && "Wrong number of operands!"); 758 assert(N->getOperand(0).getValueType() == N->getOperand(1).getValueType() && 759 "Mismatched operand types!"); 760 assert(N->getOperand(0).getValueType().isInteger() == VT.isInteger() && 761 "Wrong operand type!"); 762 assert(VT.getSizeInBits() == 2 * N->getOperand(0).getValueSizeInBits() && 763 "Wrong return type size"); 764 break; 765 } 766 case ISD::BUILD_VECTOR: { 767 assert(N->getNumValues() == 1 && "Too many results!"); 768 assert(N->getValueType(0).isVector() && "Wrong return type!"); 769 assert(N->getNumOperands() == N->getValueType(0).getVectorNumElements() && 770 "Wrong number of operands!"); 771 EVT EltVT = N->getValueType(0).getVectorElementType(); 772 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I) 773 assert((I->getValueType() == EltVT || 774 (EltVT.isInteger() && I->getValueType().isInteger() && 775 EltVT.bitsLE(I->getValueType()))) && 776 "Wrong operand type!"); 777 break; 778 } 779 } 780} 781 782/// VerifySDNode - Sanity check the given SDNode. Aborts if it is invalid. 783static void VerifySDNode(SDNode *N) { 784 // The SDNode allocators cannot be used to allocate nodes with fields that are 785 // not present in an SDNode! 786 assert(!isa<MemSDNode>(N) && "Bad MemSDNode!"); 787 assert(!isa<ShuffleVectorSDNode>(N) && "Bad ShuffleVectorSDNode!"); 788 assert(!isa<ConstantSDNode>(N) && "Bad ConstantSDNode!"); 789 assert(!isa<ConstantFPSDNode>(N) && "Bad ConstantFPSDNode!"); 790 assert(!isa<GlobalAddressSDNode>(N) && "Bad GlobalAddressSDNode!"); 791 assert(!isa<FrameIndexSDNode>(N) && "Bad FrameIndexSDNode!"); 792 assert(!isa<JumpTableSDNode>(N) && "Bad JumpTableSDNode!"); 793 assert(!isa<ConstantPoolSDNode>(N) && "Bad ConstantPoolSDNode!"); 794 assert(!isa<BasicBlockSDNode>(N) && "Bad BasicBlockSDNode!"); 795 assert(!isa<SrcValueSDNode>(N) && "Bad SrcValueSDNode!"); 796 assert(!isa<MDNodeSDNode>(N) && "Bad MDNodeSDNode!"); 797 assert(!isa<RegisterSDNode>(N) && "Bad RegisterSDNode!"); 798 assert(!isa<BlockAddressSDNode>(N) && "Bad BlockAddressSDNode!"); 799 assert(!isa<EHLabelSDNode>(N) && "Bad EHLabelSDNode!"); 800 assert(!isa<ExternalSymbolSDNode>(N) && "Bad ExternalSymbolSDNode!"); 801 assert(!isa<CondCodeSDNode>(N) && "Bad CondCodeSDNode!"); 802 assert(!isa<CvtRndSatSDNode>(N) && "Bad CvtRndSatSDNode!"); 803 assert(!isa<VTSDNode>(N) && "Bad VTSDNode!"); 804 assert(!isa<MachineSDNode>(N) && "Bad MachineSDNode!"); 805 806 VerifyNodeCommon(N); 807} 808 809/// VerifyMachineNode - Sanity check the given MachineNode. Aborts if it is 810/// invalid. 811static void VerifyMachineNode(SDNode *N) { 812 // The MachineNode allocators cannot be used to allocate nodes with fields 813 // that are not present in a MachineNode! 814 // Currently there are no such nodes. 815 816 VerifyNodeCommon(N); 817} 818#endif // NDEBUG 819 820/// getEVTAlignment - Compute the default alignment value for the 821/// given type. 822/// 823unsigned SelectionDAG::getEVTAlignment(EVT VT) const { 824 const Type *Ty = VT == MVT::iPTR ? 825 PointerType::get(Type::getInt8Ty(*getContext()), 0) : 826 VT.getTypeForEVT(*getContext()); 827 828 return TLI.getTargetData()->getABITypeAlignment(Ty); 829} 830 831// EntryNode could meaningfully have debug info if we can find it... 832SelectionDAG::SelectionDAG(const TargetMachine &tm) 833 : TM(tm), TLI(*tm.getTargetLowering()), TSI(*tm.getSelectionDAGInfo()), 834 EntryNode(ISD::EntryToken, DebugLoc(), getVTList(MVT::Other)), 835 Root(getEntryNode()), Ordering(0) { 836 AllNodes.push_back(&EntryNode); 837 Ordering = new SDNodeOrdering(); 838 DbgInfo = new SDDbgInfo(); 839} 840 841void SelectionDAG::init(MachineFunction &mf) { 842 MF = &mf; 843 Context = &mf.getFunction()->getContext(); 844} 845 846SelectionDAG::~SelectionDAG() { 847 allnodes_clear(); 848 delete Ordering; 849 delete DbgInfo; 850} 851 852void SelectionDAG::allnodes_clear() { 853 assert(&*AllNodes.begin() == &EntryNode); 854 AllNodes.remove(AllNodes.begin()); 855 while (!AllNodes.empty()) 856 DeallocateNode(AllNodes.begin()); 857} 858 859void SelectionDAG::clear() { 860 allnodes_clear(); 861 OperandAllocator.Reset(); 862 CSEMap.clear(); 863 864 ExtendedValueTypeNodes.clear(); 865 ExternalSymbols.clear(); 866 TargetExternalSymbols.clear(); 867 std::fill(CondCodeNodes.begin(), CondCodeNodes.end(), 868 static_cast<CondCodeSDNode*>(0)); 869 std::fill(ValueTypeNodes.begin(), ValueTypeNodes.end(), 870 static_cast<SDNode*>(0)); 871 872 EntryNode.UseList = 0; 873 AllNodes.push_back(&EntryNode); 874 Root = getEntryNode(); 875 Ordering->clear(); 876 DbgInfo->clear(); 877} 878 879SDValue SelectionDAG::getSExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) { 880 return VT.bitsGT(Op.getValueType()) ? 881 getNode(ISD::SIGN_EXTEND, DL, VT, Op) : 882 getNode(ISD::TRUNCATE, DL, VT, Op); 883} 884 885SDValue SelectionDAG::getZExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) { 886 return VT.bitsGT(Op.getValueType()) ? 887 getNode(ISD::ZERO_EXTEND, DL, VT, Op) : 888 getNode(ISD::TRUNCATE, DL, VT, Op); 889} 890 891SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, DebugLoc DL, EVT VT) { 892 assert(!VT.isVector() && 893 "getZeroExtendInReg should use the vector element type instead of " 894 "the vector type!"); 895 if (Op.getValueType() == VT) return Op; 896 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits(); 897 APInt Imm = APInt::getLowBitsSet(BitWidth, 898 VT.getSizeInBits()); 899 return getNode(ISD::AND, DL, Op.getValueType(), Op, 900 getConstant(Imm, Op.getValueType())); 901} 902 903/// getNOT - Create a bitwise NOT operation as (XOR Val, -1). 904/// 905SDValue SelectionDAG::getNOT(DebugLoc DL, SDValue Val, EVT VT) { 906 EVT EltVT = VT.getScalarType(); 907 SDValue NegOne = 908 getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT); 909 return getNode(ISD::XOR, DL, VT, Val, NegOne); 910} 911 912SDValue SelectionDAG::getConstant(uint64_t Val, EVT VT, bool isT) { 913 EVT EltVT = VT.getScalarType(); 914 assert((EltVT.getSizeInBits() >= 64 || 915 (uint64_t)((int64_t)Val >> EltVT.getSizeInBits()) + 1 < 2) && 916 "getConstant with a uint64_t value that doesn't fit in the type!"); 917 return getConstant(APInt(EltVT.getSizeInBits(), Val), VT, isT); 918} 919 920SDValue SelectionDAG::getConstant(const APInt &Val, EVT VT, bool isT) { 921 return getConstant(*ConstantInt::get(*Context, Val), VT, isT); 922} 923 924SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT) { 925 assert(VT.isInteger() && "Cannot create FP integer constant!"); 926 927 EVT EltVT = VT.getScalarType(); 928 assert(Val.getBitWidth() == EltVT.getSizeInBits() && 929 "APInt size does not match type size!"); 930 931 unsigned Opc = isT ? ISD::TargetConstant : ISD::Constant; 932 FoldingSetNodeID ID; 933 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0); 934 ID.AddPointer(&Val); 935 void *IP = 0; 936 SDNode *N = NULL; 937 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP))) 938 if (!VT.isVector()) 939 return SDValue(N, 0); 940 941 if (!N) { 942 N = new (NodeAllocator) ConstantSDNode(isT, &Val, EltVT); 943 CSEMap.InsertNode(N, IP); 944 AllNodes.push_back(N); 945 } 946 947 SDValue Result(N, 0); 948 if (VT.isVector()) { 949 SmallVector<SDValue, 8> Ops; 950 Ops.assign(VT.getVectorNumElements(), Result); 951 Result = getNode(ISD::BUILD_VECTOR, DebugLoc(), VT, &Ops[0], Ops.size()); 952 } 953 return Result; 954} 955 956SDValue SelectionDAG::getIntPtrConstant(uint64_t Val, bool isTarget) { 957 return getConstant(Val, TLI.getPointerTy(), isTarget); 958} 959 960 961SDValue SelectionDAG::getConstantFP(const APFloat& V, EVT VT, bool isTarget) { 962 return getConstantFP(*ConstantFP::get(*getContext(), V), VT, isTarget); 963} 964 965SDValue SelectionDAG::getConstantFP(const ConstantFP& V, EVT VT, bool isTarget){ 966 assert(VT.isFloatingPoint() && "Cannot create integer FP constant!"); 967 968 EVT EltVT = VT.getScalarType(); 969 970 // Do the map lookup using the actual bit pattern for the floating point 971 // value, so that we don't have problems with 0.0 comparing equal to -0.0, and 972 // we don't have issues with SNANs. 973 unsigned Opc = isTarget ? ISD::TargetConstantFP : ISD::ConstantFP; 974 FoldingSetNodeID ID; 975 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0); 976 ID.AddPointer(&V); 977 void *IP = 0; 978 SDNode *N = NULL; 979 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP))) 980 if (!VT.isVector()) 981 return SDValue(N, 0); 982 983 if (!N) { 984 N = new (NodeAllocator) ConstantFPSDNode(isTarget, &V, EltVT); 985 CSEMap.InsertNode(N, IP); 986 AllNodes.push_back(N); 987 } 988 989 SDValue Result(N, 0); 990 if (VT.isVector()) { 991 SmallVector<SDValue, 8> Ops; 992 Ops.assign(VT.getVectorNumElements(), Result); 993 // FIXME DebugLoc info might be appropriate here 994 Result = getNode(ISD::BUILD_VECTOR, DebugLoc(), VT, &Ops[0], Ops.size()); 995 } 996 return Result; 997} 998 999SDValue SelectionDAG::getConstantFP(double Val, EVT VT, bool isTarget) { 1000 EVT EltVT = VT.getScalarType(); 1001 if (EltVT==MVT::f32) 1002 return getConstantFP(APFloat((float)Val), VT, isTarget); 1003 else if (EltVT==MVT::f64) 1004 return getConstantFP(APFloat(Val), VT, isTarget); 1005 else if (EltVT==MVT::f80 || EltVT==MVT::f128) { 1006 bool ignored; 1007 APFloat apf = APFloat(Val); 1008 apf.convert(*EVTToAPFloatSemantics(EltVT), APFloat::rmNearestTiesToEven, 1009 &ignored); 1010 return getConstantFP(apf, VT, isTarget); 1011 } else { 1012 assert(0 && "Unsupported type in getConstantFP"); 1013 return SDValue(); 1014 } 1015} 1016 1017SDValue SelectionDAG::getGlobalAddress(const GlobalValue *GV, DebugLoc DL, 1018 EVT VT, int64_t Offset, 1019 bool isTargetGA, 1020 unsigned char TargetFlags) { 1021 assert((TargetFlags == 0 || isTargetGA) && 1022 "Cannot set target flags on target-independent globals"); 1023 1024 // Truncate (with sign-extension) the offset value to the pointer size. 1025 EVT PTy = TLI.getPointerTy(); 1026 unsigned BitWidth = PTy.getSizeInBits(); 1027 if (BitWidth < 64) 1028 Offset = (Offset << (64 - BitWidth) >> (64 - BitWidth)); 1029 1030 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV); 1031 if (!GVar) { 1032 // If GV is an alias then use the aliasee for determining thread-localness. 1033 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV)) 1034 GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal(false)); 1035 } 1036 1037 unsigned Opc; 1038 if (GVar && GVar->isThreadLocal()) 1039 Opc = isTargetGA ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress; 1040 else 1041 Opc = isTargetGA ? ISD::TargetGlobalAddress : ISD::GlobalAddress; 1042 1043 FoldingSetNodeID ID; 1044 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1045 ID.AddPointer(GV); 1046 ID.AddInteger(Offset); 1047 ID.AddInteger(TargetFlags); 1048 void *IP = 0; 1049 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1050 return SDValue(E, 0); 1051 1052 SDNode *N = new (NodeAllocator) GlobalAddressSDNode(Opc, DL, GV, VT, 1053 Offset, TargetFlags); 1054 CSEMap.InsertNode(N, IP); 1055 AllNodes.push_back(N); 1056 return SDValue(N, 0); 1057} 1058 1059SDValue SelectionDAG::getFrameIndex(int FI, EVT VT, bool isTarget) { 1060 unsigned Opc = isTarget ? ISD::TargetFrameIndex : ISD::FrameIndex; 1061 FoldingSetNodeID ID; 1062 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1063 ID.AddInteger(FI); 1064 void *IP = 0; 1065 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1066 return SDValue(E, 0); 1067 1068 SDNode *N = new (NodeAllocator) FrameIndexSDNode(FI, VT, isTarget); 1069 CSEMap.InsertNode(N, IP); 1070 AllNodes.push_back(N); 1071 return SDValue(N, 0); 1072} 1073 1074SDValue SelectionDAG::getJumpTable(int JTI, EVT VT, bool isTarget, 1075 unsigned char TargetFlags) { 1076 assert((TargetFlags == 0 || isTarget) && 1077 "Cannot set target flags on target-independent jump tables"); 1078 unsigned Opc = isTarget ? ISD::TargetJumpTable : ISD::JumpTable; 1079 FoldingSetNodeID ID; 1080 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1081 ID.AddInteger(JTI); 1082 ID.AddInteger(TargetFlags); 1083 void *IP = 0; 1084 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1085 return SDValue(E, 0); 1086 1087 SDNode *N = new (NodeAllocator) JumpTableSDNode(JTI, VT, isTarget, 1088 TargetFlags); 1089 CSEMap.InsertNode(N, IP); 1090 AllNodes.push_back(N); 1091 return SDValue(N, 0); 1092} 1093 1094SDValue SelectionDAG::getConstantPool(const Constant *C, EVT VT, 1095 unsigned Alignment, int Offset, 1096 bool isTarget, 1097 unsigned char TargetFlags) { 1098 assert((TargetFlags == 0 || isTarget) && 1099 "Cannot set target flags on target-independent globals"); 1100 if (Alignment == 0) 1101 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType()); 1102 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool; 1103 FoldingSetNodeID ID; 1104 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1105 ID.AddInteger(Alignment); 1106 ID.AddInteger(Offset); 1107 ID.AddPointer(C); 1108 ID.AddInteger(TargetFlags); 1109 void *IP = 0; 1110 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1111 return SDValue(E, 0); 1112 1113 SDNode *N = new (NodeAllocator) ConstantPoolSDNode(isTarget, C, VT, Offset, 1114 Alignment, TargetFlags); 1115 CSEMap.InsertNode(N, IP); 1116 AllNodes.push_back(N); 1117 return SDValue(N, 0); 1118} 1119 1120 1121SDValue SelectionDAG::getConstantPool(MachineConstantPoolValue *C, EVT VT, 1122 unsigned Alignment, int Offset, 1123 bool isTarget, 1124 unsigned char TargetFlags) { 1125 assert((TargetFlags == 0 || isTarget) && 1126 "Cannot set target flags on target-independent globals"); 1127 if (Alignment == 0) 1128 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType()); 1129 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool; 1130 FoldingSetNodeID ID; 1131 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1132 ID.AddInteger(Alignment); 1133 ID.AddInteger(Offset); 1134 C->AddSelectionDAGCSEId(ID); 1135 ID.AddInteger(TargetFlags); 1136 void *IP = 0; 1137 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1138 return SDValue(E, 0); 1139 1140 SDNode *N = new (NodeAllocator) ConstantPoolSDNode(isTarget, C, VT, Offset, 1141 Alignment, TargetFlags); 1142 CSEMap.InsertNode(N, IP); 1143 AllNodes.push_back(N); 1144 return SDValue(N, 0); 1145} 1146 1147SDValue SelectionDAG::getBasicBlock(MachineBasicBlock *MBB) { 1148 FoldingSetNodeID ID; 1149 AddNodeIDNode(ID, ISD::BasicBlock, getVTList(MVT::Other), 0, 0); 1150 ID.AddPointer(MBB); 1151 void *IP = 0; 1152 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1153 return SDValue(E, 0); 1154 1155 SDNode *N = new (NodeAllocator) BasicBlockSDNode(MBB); 1156 CSEMap.InsertNode(N, IP); 1157 AllNodes.push_back(N); 1158 return SDValue(N, 0); 1159} 1160 1161SDValue SelectionDAG::getValueType(EVT VT) { 1162 if (VT.isSimple() && (unsigned)VT.getSimpleVT().SimpleTy >= 1163 ValueTypeNodes.size()) 1164 ValueTypeNodes.resize(VT.getSimpleVT().SimpleTy+1); 1165 1166 SDNode *&N = VT.isExtended() ? 1167 ExtendedValueTypeNodes[VT] : ValueTypeNodes[VT.getSimpleVT().SimpleTy]; 1168 1169 if (N) return SDValue(N, 0); 1170 N = new (NodeAllocator) VTSDNode(VT); 1171 AllNodes.push_back(N); 1172 return SDValue(N, 0); 1173} 1174 1175SDValue SelectionDAG::getExternalSymbol(const char *Sym, EVT VT) { 1176 SDNode *&N = ExternalSymbols[Sym]; 1177 if (N) return SDValue(N, 0); 1178 N = new (NodeAllocator) ExternalSymbolSDNode(false, Sym, 0, VT); 1179 AllNodes.push_back(N); 1180 return SDValue(N, 0); 1181} 1182 1183SDValue SelectionDAG::getTargetExternalSymbol(const char *Sym, EVT VT, 1184 unsigned char TargetFlags) { 1185 SDNode *&N = 1186 TargetExternalSymbols[std::pair<std::string,unsigned char>(Sym, 1187 TargetFlags)]; 1188 if (N) return SDValue(N, 0); 1189 N = new (NodeAllocator) ExternalSymbolSDNode(true, Sym, TargetFlags, VT); 1190 AllNodes.push_back(N); 1191 return SDValue(N, 0); 1192} 1193 1194SDValue SelectionDAG::getCondCode(ISD::CondCode Cond) { 1195 if ((unsigned)Cond >= CondCodeNodes.size()) 1196 CondCodeNodes.resize(Cond+1); 1197 1198 if (CondCodeNodes[Cond] == 0) { 1199 CondCodeSDNode *N = new (NodeAllocator) CondCodeSDNode(Cond); 1200 CondCodeNodes[Cond] = N; 1201 AllNodes.push_back(N); 1202 } 1203 1204 return SDValue(CondCodeNodes[Cond], 0); 1205} 1206 1207// commuteShuffle - swaps the values of N1 and N2, and swaps all indices in 1208// the shuffle mask M that point at N1 to point at N2, and indices that point 1209// N2 to point at N1. 1210static void commuteShuffle(SDValue &N1, SDValue &N2, SmallVectorImpl<int> &M) { 1211 std::swap(N1, N2); 1212 int NElts = M.size(); 1213 for (int i = 0; i != NElts; ++i) { 1214 if (M[i] >= NElts) 1215 M[i] -= NElts; 1216 else if (M[i] >= 0) 1217 M[i] += NElts; 1218 } 1219} 1220 1221SDValue SelectionDAG::getVectorShuffle(EVT VT, DebugLoc dl, SDValue N1, 1222 SDValue N2, const int *Mask) { 1223 assert(N1.getValueType() == N2.getValueType() && "Invalid VECTOR_SHUFFLE"); 1224 assert(VT.isVector() && N1.getValueType().isVector() && 1225 "Vector Shuffle VTs must be a vectors"); 1226 assert(VT.getVectorElementType() == N1.getValueType().getVectorElementType() 1227 && "Vector Shuffle VTs must have same element type"); 1228 1229 // Canonicalize shuffle undef, undef -> undef 1230 if (N1.getOpcode() == ISD::UNDEF && N2.getOpcode() == ISD::UNDEF) 1231 return getUNDEF(VT); 1232 1233 // Validate that all indices in Mask are within the range of the elements 1234 // input to the shuffle. 1235 unsigned NElts = VT.getVectorNumElements(); 1236 SmallVector<int, 8> MaskVec; 1237 for (unsigned i = 0; i != NElts; ++i) { 1238 assert(Mask[i] < (int)(NElts * 2) && "Index out of range"); 1239 MaskVec.push_back(Mask[i]); 1240 } 1241 1242 // Canonicalize shuffle v, v -> v, undef 1243 if (N1 == N2) { 1244 N2 = getUNDEF(VT); 1245 for (unsigned i = 0; i != NElts; ++i) 1246 if (MaskVec[i] >= (int)NElts) MaskVec[i] -= NElts; 1247 } 1248 1249 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask. 1250 if (N1.getOpcode() == ISD::UNDEF) 1251 commuteShuffle(N1, N2, MaskVec); 1252 1253 // Canonicalize all index into lhs, -> shuffle lhs, undef 1254 // Canonicalize all index into rhs, -> shuffle rhs, undef 1255 bool AllLHS = true, AllRHS = true; 1256 bool N2Undef = N2.getOpcode() == ISD::UNDEF; 1257 for (unsigned i = 0; i != NElts; ++i) { 1258 if (MaskVec[i] >= (int)NElts) { 1259 if (N2Undef) 1260 MaskVec[i] = -1; 1261 else 1262 AllLHS = false; 1263 } else if (MaskVec[i] >= 0) { 1264 AllRHS = false; 1265 } 1266 } 1267 if (AllLHS && AllRHS) 1268 return getUNDEF(VT); 1269 if (AllLHS && !N2Undef) 1270 N2 = getUNDEF(VT); 1271 if (AllRHS) { 1272 N1 = getUNDEF(VT); 1273 commuteShuffle(N1, N2, MaskVec); 1274 } 1275 1276 // If Identity shuffle, or all shuffle in to undef, return that node. 1277 bool AllUndef = true; 1278 bool Identity = true; 1279 for (unsigned i = 0; i != NElts; ++i) { 1280 if (MaskVec[i] >= 0 && MaskVec[i] != (int)i) Identity = false; 1281 if (MaskVec[i] >= 0) AllUndef = false; 1282 } 1283 if (Identity && NElts == N1.getValueType().getVectorNumElements()) 1284 return N1; 1285 if (AllUndef) 1286 return getUNDEF(VT); 1287 1288 FoldingSetNodeID ID; 1289 SDValue Ops[2] = { N1, N2 }; 1290 AddNodeIDNode(ID, ISD::VECTOR_SHUFFLE, getVTList(VT), Ops, 2); 1291 for (unsigned i = 0; i != NElts; ++i) 1292 ID.AddInteger(MaskVec[i]); 1293 1294 void* IP = 0; 1295 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1296 return SDValue(E, 0); 1297 1298 // Allocate the mask array for the node out of the BumpPtrAllocator, since 1299 // SDNode doesn't have access to it. This memory will be "leaked" when 1300 // the node is deallocated, but recovered when the NodeAllocator is released. 1301 int *MaskAlloc = OperandAllocator.Allocate<int>(NElts); 1302 memcpy(MaskAlloc, &MaskVec[0], NElts * sizeof(int)); 1303 1304 ShuffleVectorSDNode *N = 1305 new (NodeAllocator) ShuffleVectorSDNode(VT, dl, N1, N2, MaskAlloc); 1306 CSEMap.InsertNode(N, IP); 1307 AllNodes.push_back(N); 1308 return SDValue(N, 0); 1309} 1310 1311SDValue SelectionDAG::getConvertRndSat(EVT VT, DebugLoc dl, 1312 SDValue Val, SDValue DTy, 1313 SDValue STy, SDValue Rnd, SDValue Sat, 1314 ISD::CvtCode Code) { 1315 // If the src and dest types are the same and the conversion is between 1316 // integer types of the same sign or two floats, no conversion is necessary. 1317 if (DTy == STy && 1318 (Code == ISD::CVT_UU || Code == ISD::CVT_SS || Code == ISD::CVT_FF)) 1319 return Val; 1320 1321 FoldingSetNodeID ID; 1322 SDValue Ops[] = { Val, DTy, STy, Rnd, Sat }; 1323 AddNodeIDNode(ID, ISD::CONVERT_RNDSAT, getVTList(VT), &Ops[0], 5); 1324 void* IP = 0; 1325 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1326 return SDValue(E, 0); 1327 1328 CvtRndSatSDNode *N = new (NodeAllocator) CvtRndSatSDNode(VT, dl, Ops, 5, 1329 Code); 1330 CSEMap.InsertNode(N, IP); 1331 AllNodes.push_back(N); 1332 return SDValue(N, 0); 1333} 1334 1335SDValue SelectionDAG::getRegister(unsigned RegNo, EVT VT) { 1336 FoldingSetNodeID ID; 1337 AddNodeIDNode(ID, ISD::Register, getVTList(VT), 0, 0); 1338 ID.AddInteger(RegNo); 1339 void *IP = 0; 1340 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1341 return SDValue(E, 0); 1342 1343 SDNode *N = new (NodeAllocator) RegisterSDNode(RegNo, VT); 1344 CSEMap.InsertNode(N, IP); 1345 AllNodes.push_back(N); 1346 return SDValue(N, 0); 1347} 1348 1349SDValue SelectionDAG::getEHLabel(DebugLoc dl, SDValue Root, MCSymbol *Label) { 1350 FoldingSetNodeID ID; 1351 SDValue Ops[] = { Root }; 1352 AddNodeIDNode(ID, ISD::EH_LABEL, getVTList(MVT::Other), &Ops[0], 1); 1353 ID.AddPointer(Label); 1354 void *IP = 0; 1355 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1356 return SDValue(E, 0); 1357 1358 SDNode *N = new (NodeAllocator) EHLabelSDNode(dl, Root, Label); 1359 CSEMap.InsertNode(N, IP); 1360 AllNodes.push_back(N); 1361 return SDValue(N, 0); 1362} 1363 1364 1365SDValue SelectionDAG::getBlockAddress(const BlockAddress *BA, EVT VT, 1366 bool isTarget, 1367 unsigned char TargetFlags) { 1368 unsigned Opc = isTarget ? ISD::TargetBlockAddress : ISD::BlockAddress; 1369 1370 FoldingSetNodeID ID; 1371 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1372 ID.AddPointer(BA); 1373 ID.AddInteger(TargetFlags); 1374 void *IP = 0; 1375 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1376 return SDValue(E, 0); 1377 1378 SDNode *N = new (NodeAllocator) BlockAddressSDNode(Opc, VT, BA, TargetFlags); 1379 CSEMap.InsertNode(N, IP); 1380 AllNodes.push_back(N); 1381 return SDValue(N, 0); 1382} 1383 1384SDValue SelectionDAG::getSrcValue(const Value *V) { 1385 assert((!V || V->getType()->isPointerTy()) && 1386 "SrcValue is not a pointer?"); 1387 1388 FoldingSetNodeID ID; 1389 AddNodeIDNode(ID, ISD::SRCVALUE, getVTList(MVT::Other), 0, 0); 1390 ID.AddPointer(V); 1391 1392 void *IP = 0; 1393 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1394 return SDValue(E, 0); 1395 1396 SDNode *N = new (NodeAllocator) SrcValueSDNode(V); 1397 CSEMap.InsertNode(N, IP); 1398 AllNodes.push_back(N); 1399 return SDValue(N, 0); 1400} 1401 1402/// getMDNode - Return an MDNodeSDNode which holds an MDNode. 1403SDValue SelectionDAG::getMDNode(const MDNode *MD) { 1404 FoldingSetNodeID ID; 1405 AddNodeIDNode(ID, ISD::MDNODE_SDNODE, getVTList(MVT::Other), 0, 0); 1406 ID.AddPointer(MD); 1407 1408 void *IP = 0; 1409 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1410 return SDValue(E, 0); 1411 1412 SDNode *N = new (NodeAllocator) MDNodeSDNode(MD); 1413 CSEMap.InsertNode(N, IP); 1414 AllNodes.push_back(N); 1415 return SDValue(N, 0); 1416} 1417 1418 1419/// getShiftAmountOperand - Return the specified value casted to 1420/// the target's desired shift amount type. 1421SDValue SelectionDAG::getShiftAmountOperand(EVT LHSTy, SDValue Op) { 1422 EVT OpTy = Op.getValueType(); 1423 MVT ShTy = TLI.getShiftAmountTy(LHSTy); 1424 if (OpTy == ShTy || OpTy.isVector()) return Op; 1425 1426 ISD::NodeType Opcode = OpTy.bitsGT(ShTy) ? ISD::TRUNCATE : ISD::ZERO_EXTEND; 1427 return getNode(Opcode, Op.getDebugLoc(), ShTy, Op); 1428} 1429 1430/// CreateStackTemporary - Create a stack temporary, suitable for holding the 1431/// specified value type. 1432SDValue SelectionDAG::CreateStackTemporary(EVT VT, unsigned minAlign) { 1433 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo(); 1434 unsigned ByteSize = VT.getStoreSize(); 1435 const Type *Ty = VT.getTypeForEVT(*getContext()); 1436 unsigned StackAlign = 1437 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), minAlign); 1438 1439 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false); 1440 return getFrameIndex(FrameIdx, TLI.getPointerTy()); 1441} 1442 1443/// CreateStackTemporary - Create a stack temporary suitable for holding 1444/// either of the specified value types. 1445SDValue SelectionDAG::CreateStackTemporary(EVT VT1, EVT VT2) { 1446 unsigned Bytes = std::max(VT1.getStoreSizeInBits(), 1447 VT2.getStoreSizeInBits())/8; 1448 const Type *Ty1 = VT1.getTypeForEVT(*getContext()); 1449 const Type *Ty2 = VT2.getTypeForEVT(*getContext()); 1450 const TargetData *TD = TLI.getTargetData(); 1451 unsigned Align = std::max(TD->getPrefTypeAlignment(Ty1), 1452 TD->getPrefTypeAlignment(Ty2)); 1453 1454 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo(); 1455 int FrameIdx = FrameInfo->CreateStackObject(Bytes, Align, false); 1456 return getFrameIndex(FrameIdx, TLI.getPointerTy()); 1457} 1458 1459SDValue SelectionDAG::FoldSetCC(EVT VT, SDValue N1, 1460 SDValue N2, ISD::CondCode Cond, DebugLoc dl) { 1461 // These setcc operations always fold. 1462 switch (Cond) { 1463 default: break; 1464 case ISD::SETFALSE: 1465 case ISD::SETFALSE2: return getConstant(0, VT); 1466 case ISD::SETTRUE: 1467 case ISD::SETTRUE2: return getConstant(1, VT); 1468 1469 case ISD::SETOEQ: 1470 case ISD::SETOGT: 1471 case ISD::SETOGE: 1472 case ISD::SETOLT: 1473 case ISD::SETOLE: 1474 case ISD::SETONE: 1475 case ISD::SETO: 1476 case ISD::SETUO: 1477 case ISD::SETUEQ: 1478 case ISD::SETUNE: 1479 assert(!N1.getValueType().isInteger() && "Illegal setcc for integer!"); 1480 break; 1481 } 1482 1483 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode())) { 1484 const APInt &C2 = N2C->getAPIntValue(); 1485 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { 1486 const APInt &C1 = N1C->getAPIntValue(); 1487 1488 switch (Cond) { 1489 default: llvm_unreachable("Unknown integer setcc!"); 1490 case ISD::SETEQ: return getConstant(C1 == C2, VT); 1491 case ISD::SETNE: return getConstant(C1 != C2, VT); 1492 case ISD::SETULT: return getConstant(C1.ult(C2), VT); 1493 case ISD::SETUGT: return getConstant(C1.ugt(C2), VT); 1494 case ISD::SETULE: return getConstant(C1.ule(C2), VT); 1495 case ISD::SETUGE: return getConstant(C1.uge(C2), VT); 1496 case ISD::SETLT: return getConstant(C1.slt(C2), VT); 1497 case ISD::SETGT: return getConstant(C1.sgt(C2), VT); 1498 case ISD::SETLE: return getConstant(C1.sle(C2), VT); 1499 case ISD::SETGE: return getConstant(C1.sge(C2), VT); 1500 } 1501 } 1502 } 1503 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.getNode())) { 1504 if (ConstantFPSDNode *N2C = dyn_cast<ConstantFPSDNode>(N2.getNode())) { 1505 // No compile time operations on this type yet. 1506 if (N1C->getValueType(0) == MVT::ppcf128) 1507 return SDValue(); 1508 1509 APFloat::cmpResult R = N1C->getValueAPF().compare(N2C->getValueAPF()); 1510 switch (Cond) { 1511 default: break; 1512 case ISD::SETEQ: if (R==APFloat::cmpUnordered) 1513 return getUNDEF(VT); 1514 // fall through 1515 case ISD::SETOEQ: return getConstant(R==APFloat::cmpEqual, VT); 1516 case ISD::SETNE: if (R==APFloat::cmpUnordered) 1517 return getUNDEF(VT); 1518 // fall through 1519 case ISD::SETONE: return getConstant(R==APFloat::cmpGreaterThan || 1520 R==APFloat::cmpLessThan, VT); 1521 case ISD::SETLT: if (R==APFloat::cmpUnordered) 1522 return getUNDEF(VT); 1523 // fall through 1524 case ISD::SETOLT: return getConstant(R==APFloat::cmpLessThan, VT); 1525 case ISD::SETGT: if (R==APFloat::cmpUnordered) 1526 return getUNDEF(VT); 1527 // fall through 1528 case ISD::SETOGT: return getConstant(R==APFloat::cmpGreaterThan, VT); 1529 case ISD::SETLE: if (R==APFloat::cmpUnordered) 1530 return getUNDEF(VT); 1531 // fall through 1532 case ISD::SETOLE: return getConstant(R==APFloat::cmpLessThan || 1533 R==APFloat::cmpEqual, VT); 1534 case ISD::SETGE: if (R==APFloat::cmpUnordered) 1535 return getUNDEF(VT); 1536 // fall through 1537 case ISD::SETOGE: return getConstant(R==APFloat::cmpGreaterThan || 1538 R==APFloat::cmpEqual, VT); 1539 case ISD::SETO: return getConstant(R!=APFloat::cmpUnordered, VT); 1540 case ISD::SETUO: return getConstant(R==APFloat::cmpUnordered, VT); 1541 case ISD::SETUEQ: return getConstant(R==APFloat::cmpUnordered || 1542 R==APFloat::cmpEqual, VT); 1543 case ISD::SETUNE: return getConstant(R!=APFloat::cmpEqual, VT); 1544 case ISD::SETULT: return getConstant(R==APFloat::cmpUnordered || 1545 R==APFloat::cmpLessThan, VT); 1546 case ISD::SETUGT: return getConstant(R==APFloat::cmpGreaterThan || 1547 R==APFloat::cmpUnordered, VT); 1548 case ISD::SETULE: return getConstant(R!=APFloat::cmpGreaterThan, VT); 1549 case ISD::SETUGE: return getConstant(R!=APFloat::cmpLessThan, VT); 1550 } 1551 } else { 1552 // Ensure that the constant occurs on the RHS. 1553 return getSetCC(dl, VT, N2, N1, ISD::getSetCCSwappedOperands(Cond)); 1554 } 1555 } 1556 1557 // Could not fold it. 1558 return SDValue(); 1559} 1560 1561/// SignBitIsZero - Return true if the sign bit of Op is known to be zero. We 1562/// use this predicate to simplify operations downstream. 1563bool SelectionDAG::SignBitIsZero(SDValue Op, unsigned Depth) const { 1564 // This predicate is not safe for vector operations. 1565 if (Op.getValueType().isVector()) 1566 return false; 1567 1568 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits(); 1569 return MaskedValueIsZero(Op, APInt::getSignBit(BitWidth), Depth); 1570} 1571 1572/// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use 1573/// this predicate to simplify operations downstream. Mask is known to be zero 1574/// for bits that V cannot have. 1575bool SelectionDAG::MaskedValueIsZero(SDValue Op, const APInt &Mask, 1576 unsigned Depth) const { 1577 APInt KnownZero, KnownOne; 1578 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth); 1579 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1580 return (KnownZero & Mask) == Mask; 1581} 1582 1583/// ComputeMaskedBits - Determine which of the bits specified in Mask are 1584/// known to be either zero or one and return them in the KnownZero/KnownOne 1585/// bitsets. This code only analyzes bits in Mask, in order to short-circuit 1586/// processing. 1587void SelectionDAG::ComputeMaskedBits(SDValue Op, const APInt &Mask, 1588 APInt &KnownZero, APInt &KnownOne, 1589 unsigned Depth) const { 1590 unsigned BitWidth = Mask.getBitWidth(); 1591 assert(BitWidth == Op.getValueType().getScalarType().getSizeInBits() && 1592 "Mask size mismatches value type size!"); 1593 1594 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything. 1595 if (Depth == 6 || Mask == 0) 1596 return; // Limit search depth. 1597 1598 APInt KnownZero2, KnownOne2; 1599 1600 switch (Op.getOpcode()) { 1601 case ISD::Constant: 1602 // We know all of the bits for a constant! 1603 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & Mask; 1604 KnownZero = ~KnownOne & Mask; 1605 return; 1606 case ISD::AND: 1607 // If either the LHS or the RHS are Zero, the result is zero. 1608 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1); 1609 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownZero, 1610 KnownZero2, KnownOne2, Depth+1); 1611 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1612 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1613 1614 // Output known-1 bits are only known if set in both the LHS & RHS. 1615 KnownOne &= KnownOne2; 1616 // Output known-0 are known to be clear if zero in either the LHS | RHS. 1617 KnownZero |= KnownZero2; 1618 return; 1619 case ISD::OR: 1620 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1); 1621 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownOne, 1622 KnownZero2, KnownOne2, Depth+1); 1623 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1624 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1625 1626 // Output known-0 bits are only known if clear in both the LHS & RHS. 1627 KnownZero &= KnownZero2; 1628 // Output known-1 are known to be set if set in either the LHS | RHS. 1629 KnownOne |= KnownOne2; 1630 return; 1631 case ISD::XOR: { 1632 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1); 1633 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1); 1634 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1635 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1636 1637 // Output known-0 bits are known if clear or set in both the LHS & RHS. 1638 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 1639 // Output known-1 are known to be set if set in only one of the LHS, RHS. 1640 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 1641 KnownZero = KnownZeroOut; 1642 return; 1643 } 1644 case ISD::MUL: { 1645 APInt Mask2 = APInt::getAllOnesValue(BitWidth); 1646 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero, KnownOne, Depth+1); 1647 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1); 1648 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1649 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1650 1651 // If low bits are zero in either operand, output low known-0 bits. 1652 // Also compute a conserative estimate for high known-0 bits. 1653 // More trickiness is possible, but this is sufficient for the 1654 // interesting case of alignment computation. 1655 KnownOne.clearAllBits(); 1656 unsigned TrailZ = KnownZero.countTrailingOnes() + 1657 KnownZero2.countTrailingOnes(); 1658 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() + 1659 KnownZero2.countLeadingOnes(), 1660 BitWidth) - BitWidth; 1661 1662 TrailZ = std::min(TrailZ, BitWidth); 1663 LeadZ = std::min(LeadZ, BitWidth); 1664 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) | 1665 APInt::getHighBitsSet(BitWidth, LeadZ); 1666 KnownZero &= Mask; 1667 return; 1668 } 1669 case ISD::UDIV: { 1670 // For the purposes of computing leading zeros we can conservatively 1671 // treat a udiv as a logical right shift by the power of 2 known to 1672 // be less than the denominator. 1673 APInt AllOnes = APInt::getAllOnesValue(BitWidth); 1674 ComputeMaskedBits(Op.getOperand(0), 1675 AllOnes, KnownZero2, KnownOne2, Depth+1); 1676 unsigned LeadZ = KnownZero2.countLeadingOnes(); 1677 1678 KnownOne2.clearAllBits(); 1679 KnownZero2.clearAllBits(); 1680 ComputeMaskedBits(Op.getOperand(1), 1681 AllOnes, KnownZero2, KnownOne2, Depth+1); 1682 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros(); 1683 if (RHSUnknownLeadingOnes != BitWidth) 1684 LeadZ = std::min(BitWidth, 1685 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1); 1686 1687 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask; 1688 return; 1689 } 1690 case ISD::SELECT: 1691 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero, KnownOne, Depth+1); 1692 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1); 1693 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1694 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1695 1696 // Only known if known in both the LHS and RHS. 1697 KnownOne &= KnownOne2; 1698 KnownZero &= KnownZero2; 1699 return; 1700 case ISD::SELECT_CC: 1701 ComputeMaskedBits(Op.getOperand(3), Mask, KnownZero, KnownOne, Depth+1); 1702 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero2, KnownOne2, Depth+1); 1703 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1704 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1705 1706 // Only known if known in both the LHS and RHS. 1707 KnownOne &= KnownOne2; 1708 KnownZero &= KnownZero2; 1709 return; 1710 case ISD::SADDO: 1711 case ISD::UADDO: 1712 case ISD::SSUBO: 1713 case ISD::USUBO: 1714 case ISD::SMULO: 1715 case ISD::UMULO: 1716 if (Op.getResNo() != 1) 1717 return; 1718 // The boolean result conforms to getBooleanContents. Fall through. 1719 case ISD::SETCC: 1720 // If we know the result of a setcc has the top bits zero, use this info. 1721 if (TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent && 1722 BitWidth > 1) 1723 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1); 1724 return; 1725 case ISD::SHL: 1726 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0 1727 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1728 unsigned ShAmt = SA->getZExtValue(); 1729 1730 // If the shift count is an invalid immediate, don't do anything. 1731 if (ShAmt >= BitWidth) 1732 return; 1733 1734 ComputeMaskedBits(Op.getOperand(0), Mask.lshr(ShAmt), 1735 KnownZero, KnownOne, Depth+1); 1736 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1737 KnownZero <<= ShAmt; 1738 KnownOne <<= ShAmt; 1739 // low bits known zero. 1740 KnownZero |= APInt::getLowBitsSet(BitWidth, ShAmt); 1741 } 1742 return; 1743 case ISD::SRL: 1744 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0 1745 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1746 unsigned ShAmt = SA->getZExtValue(); 1747 1748 // If the shift count is an invalid immediate, don't do anything. 1749 if (ShAmt >= BitWidth) 1750 return; 1751 1752 ComputeMaskedBits(Op.getOperand(0), (Mask << ShAmt), 1753 KnownZero, KnownOne, Depth+1); 1754 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1755 KnownZero = KnownZero.lshr(ShAmt); 1756 KnownOne = KnownOne.lshr(ShAmt); 1757 1758 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask; 1759 KnownZero |= HighBits; // High bits known zero. 1760 } 1761 return; 1762 case ISD::SRA: 1763 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1764 unsigned ShAmt = SA->getZExtValue(); 1765 1766 // If the shift count is an invalid immediate, don't do anything. 1767 if (ShAmt >= BitWidth) 1768 return; 1769 1770 APInt InDemandedMask = (Mask << ShAmt); 1771 // If any of the demanded bits are produced by the sign extension, we also 1772 // demand the input sign bit. 1773 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask; 1774 if (HighBits.getBoolValue()) 1775 InDemandedMask |= APInt::getSignBit(BitWidth); 1776 1777 ComputeMaskedBits(Op.getOperand(0), InDemandedMask, KnownZero, KnownOne, 1778 Depth+1); 1779 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1780 KnownZero = KnownZero.lshr(ShAmt); 1781 KnownOne = KnownOne.lshr(ShAmt); 1782 1783 // Handle the sign bits. 1784 APInt SignBit = APInt::getSignBit(BitWidth); 1785 SignBit = SignBit.lshr(ShAmt); // Adjust to where it is now in the mask. 1786 1787 if (KnownZero.intersects(SignBit)) { 1788 KnownZero |= HighBits; // New bits are known zero. 1789 } else if (KnownOne.intersects(SignBit)) { 1790 KnownOne |= HighBits; // New bits are known one. 1791 } 1792 } 1793 return; 1794 case ISD::SIGN_EXTEND_INREG: { 1795 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1796 unsigned EBits = EVT.getScalarType().getSizeInBits(); 1797 1798 // Sign extension. Compute the demanded bits in the result that are not 1799 // present in the input. 1800 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - EBits) & Mask; 1801 1802 APInt InSignBit = APInt::getSignBit(EBits); 1803 APInt InputDemandedBits = Mask & APInt::getLowBitsSet(BitWidth, EBits); 1804 1805 // If the sign extended bits are demanded, we know that the sign 1806 // bit is demanded. 1807 InSignBit = InSignBit.zext(BitWidth); 1808 if (NewBits.getBoolValue()) 1809 InputDemandedBits |= InSignBit; 1810 1811 ComputeMaskedBits(Op.getOperand(0), InputDemandedBits, 1812 KnownZero, KnownOne, Depth+1); 1813 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1814 1815 // If the sign bit of the input is known set or clear, then we know the 1816 // top bits of the result. 1817 if (KnownZero.intersects(InSignBit)) { // Input sign bit known clear 1818 KnownZero |= NewBits; 1819 KnownOne &= ~NewBits; 1820 } else if (KnownOne.intersects(InSignBit)) { // Input sign bit known set 1821 KnownOne |= NewBits; 1822 KnownZero &= ~NewBits; 1823 } else { // Input sign bit unknown 1824 KnownZero &= ~NewBits; 1825 KnownOne &= ~NewBits; 1826 } 1827 return; 1828 } 1829 case ISD::CTTZ: 1830 case ISD::CTLZ: 1831 case ISD::CTPOP: { 1832 unsigned LowBits = Log2_32(BitWidth)+1; 1833 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits); 1834 KnownOne.clearAllBits(); 1835 return; 1836 } 1837 case ISD::LOAD: { 1838 if (ISD::isZEXTLoad(Op.getNode())) { 1839 LoadSDNode *LD = cast<LoadSDNode>(Op); 1840 EVT VT = LD->getMemoryVT(); 1841 unsigned MemBits = VT.getScalarType().getSizeInBits(); 1842 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits) & Mask; 1843 } 1844 return; 1845 } 1846 case ISD::ZERO_EXTEND: { 1847 EVT InVT = Op.getOperand(0).getValueType(); 1848 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1849 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask; 1850 APInt InMask = Mask.trunc(InBits); 1851 KnownZero = KnownZero.trunc(InBits); 1852 KnownOne = KnownOne.trunc(InBits); 1853 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1); 1854 KnownZero = KnownZero.zext(BitWidth); 1855 KnownOne = KnownOne.zext(BitWidth); 1856 KnownZero |= NewBits; 1857 return; 1858 } 1859 case ISD::SIGN_EXTEND: { 1860 EVT InVT = Op.getOperand(0).getValueType(); 1861 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1862 APInt InSignBit = APInt::getSignBit(InBits); 1863 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask; 1864 APInt InMask = Mask.trunc(InBits); 1865 1866 // If any of the sign extended bits are demanded, we know that the sign 1867 // bit is demanded. Temporarily set this bit in the mask for our callee. 1868 if (NewBits.getBoolValue()) 1869 InMask |= InSignBit; 1870 1871 KnownZero = KnownZero.trunc(InBits); 1872 KnownOne = KnownOne.trunc(InBits); 1873 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1); 1874 1875 // Note if the sign bit is known to be zero or one. 1876 bool SignBitKnownZero = KnownZero.isNegative(); 1877 bool SignBitKnownOne = KnownOne.isNegative(); 1878 assert(!(SignBitKnownZero && SignBitKnownOne) && 1879 "Sign bit can't be known to be both zero and one!"); 1880 1881 // If the sign bit wasn't actually demanded by our caller, we don't 1882 // want it set in the KnownZero and KnownOne result values. Reset the 1883 // mask and reapply it to the result values. 1884 InMask = Mask.trunc(InBits); 1885 KnownZero &= InMask; 1886 KnownOne &= InMask; 1887 1888 KnownZero = KnownZero.zext(BitWidth); 1889 KnownOne = KnownOne.zext(BitWidth); 1890 1891 // If the sign bit is known zero or one, the top bits match. 1892 if (SignBitKnownZero) 1893 KnownZero |= NewBits; 1894 else if (SignBitKnownOne) 1895 KnownOne |= NewBits; 1896 return; 1897 } 1898 case ISD::ANY_EXTEND: { 1899 EVT InVT = Op.getOperand(0).getValueType(); 1900 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1901 APInt InMask = Mask.trunc(InBits); 1902 KnownZero = KnownZero.trunc(InBits); 1903 KnownOne = KnownOne.trunc(InBits); 1904 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1); 1905 KnownZero = KnownZero.zext(BitWidth); 1906 KnownOne = KnownOne.zext(BitWidth); 1907 return; 1908 } 1909 case ISD::TRUNCATE: { 1910 EVT InVT = Op.getOperand(0).getValueType(); 1911 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1912 APInt InMask = Mask.zext(InBits); 1913 KnownZero = KnownZero.zext(InBits); 1914 KnownOne = KnownOne.zext(InBits); 1915 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1); 1916 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1917 KnownZero = KnownZero.trunc(BitWidth); 1918 KnownOne = KnownOne.trunc(BitWidth); 1919 break; 1920 } 1921 case ISD::AssertZext: { 1922 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1923 APInt InMask = APInt::getLowBitsSet(BitWidth, VT.getSizeInBits()); 1924 ComputeMaskedBits(Op.getOperand(0), Mask & InMask, KnownZero, 1925 KnownOne, Depth+1); 1926 KnownZero |= (~InMask) & Mask; 1927 return; 1928 } 1929 case ISD::FGETSIGN: 1930 // All bits are zero except the low bit. 1931 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - 1); 1932 return; 1933 1934 case ISD::SUB: { 1935 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) { 1936 // We know that the top bits of C-X are clear if X contains less bits 1937 // than C (i.e. no wrap-around can happen). For example, 20-X is 1938 // positive if we can prove that X is >= 0 and < 16. 1939 if (CLHS->getAPIntValue().isNonNegative()) { 1940 unsigned NLZ = (CLHS->getAPIntValue()+1).countLeadingZeros(); 1941 // NLZ can't be BitWidth with no sign bit 1942 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1); 1943 ComputeMaskedBits(Op.getOperand(1), MaskV, KnownZero2, KnownOne2, 1944 Depth+1); 1945 1946 // If all of the MaskV bits are known to be zero, then we know the 1947 // output top bits are zero, because we now know that the output is 1948 // from [0-C]. 1949 if ((KnownZero2 & MaskV) == MaskV) { 1950 unsigned NLZ2 = CLHS->getAPIntValue().countLeadingZeros(); 1951 // Top bits known zero. 1952 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask; 1953 } 1954 } 1955 } 1956 } 1957 // fall through 1958 case ISD::ADD: 1959 case ISD::ADDE: { 1960 // Output known-0 bits are known if clear or set in both the low clear bits 1961 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the 1962 // low 3 bits clear. 1963 APInt Mask2 = APInt::getLowBitsSet(BitWidth, 1964 BitWidth - Mask.countLeadingZeros()); 1965 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1); 1966 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1967 unsigned KnownZeroOut = KnownZero2.countTrailingOnes(); 1968 1969 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero2, KnownOne2, Depth+1); 1970 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1971 KnownZeroOut = std::min(KnownZeroOut, 1972 KnownZero2.countTrailingOnes()); 1973 1974 if (Op.getOpcode() == ISD::ADD) { 1975 KnownZero |= APInt::getLowBitsSet(BitWidth, KnownZeroOut); 1976 return; 1977 } 1978 1979 // With ADDE, a carry bit may be added in, so we can only use this 1980 // information if we know (at least) that the low two bits are clear. We 1981 // then return to the caller that the low bit is unknown but that other bits 1982 // are known zero. 1983 if (KnownZeroOut >= 2) // ADDE 1984 KnownZero |= APInt::getBitsSet(BitWidth, 1, KnownZeroOut); 1985 return; 1986 } 1987 case ISD::SREM: 1988 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1989 const APInt &RA = Rem->getAPIntValue().abs(); 1990 if (RA.isPowerOf2()) { 1991 APInt LowBits = RA - 1; 1992 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth); 1993 ComputeMaskedBits(Op.getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1); 1994 1995 // The low bits of the first operand are unchanged by the srem. 1996 KnownZero = KnownZero2 & LowBits; 1997 KnownOne = KnownOne2 & LowBits; 1998 1999 // If the first operand is non-negative or has all low bits zero, then 2000 // the upper bits are all zero. 2001 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits)) 2002 KnownZero |= ~LowBits; 2003 2004 // If the first operand is negative and not all low bits are zero, then 2005 // the upper bits are all one. 2006 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0)) 2007 KnownOne |= ~LowBits; 2008 2009 KnownZero &= Mask; 2010 KnownOne &= Mask; 2011 2012 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?"); 2013 } 2014 } 2015 return; 2016 case ISD::UREM: { 2017 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 2018 const APInt &RA = Rem->getAPIntValue(); 2019 if (RA.isPowerOf2()) { 2020 APInt LowBits = (RA - 1); 2021 APInt Mask2 = LowBits & Mask; 2022 KnownZero |= ~LowBits & Mask; 2023 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero, KnownOne,Depth+1); 2024 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?"); 2025 break; 2026 } 2027 } 2028 2029 // Since the result is less than or equal to either operand, any leading 2030 // zero bits in either operand must also exist in the result. 2031 APInt AllOnes = APInt::getAllOnesValue(BitWidth); 2032 ComputeMaskedBits(Op.getOperand(0), AllOnes, KnownZero, KnownOne, 2033 Depth+1); 2034 ComputeMaskedBits(Op.getOperand(1), AllOnes, KnownZero2, KnownOne2, 2035 Depth+1); 2036 2037 uint32_t Leaders = std::max(KnownZero.countLeadingOnes(), 2038 KnownZero2.countLeadingOnes()); 2039 KnownOne.clearAllBits(); 2040 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask; 2041 return; 2042 } 2043 case ISD::FrameIndex: 2044 case ISD::TargetFrameIndex: 2045 if (unsigned Align = InferPtrAlignment(Op)) { 2046 // The low bits are known zero if the pointer is aligned. 2047 KnownZero = APInt::getLowBitsSet(BitWidth, Log2_32(Align)); 2048 return; 2049 } 2050 break; 2051 2052 default: 2053 if (Op.getOpcode() < ISD::BUILTIN_OP_END) 2054 break; 2055 // Fallthrough 2056 case ISD::INTRINSIC_WO_CHAIN: 2057 case ISD::INTRINSIC_W_CHAIN: 2058 case ISD::INTRINSIC_VOID: 2059 // Allow the target to implement this method for its nodes. 2060 TLI.computeMaskedBitsForTargetNode(Op, Mask, KnownZero, KnownOne, *this, 2061 Depth); 2062 return; 2063 } 2064} 2065 2066/// ComputeNumSignBits - Return the number of times the sign bit of the 2067/// register is replicated into the other bits. We know that at least 1 bit 2068/// is always equal to the sign bit (itself), but other cases can give us 2069/// information. For example, immediately after an "SRA X, 2", we know that 2070/// the top 3 bits are all equal to each other, so we return 3. 2071unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const{ 2072 EVT VT = Op.getValueType(); 2073 assert(VT.isInteger() && "Invalid VT!"); 2074 unsigned VTBits = VT.getScalarType().getSizeInBits(); 2075 unsigned Tmp, Tmp2; 2076 unsigned FirstAnswer = 1; 2077 2078 if (Depth == 6) 2079 return 1; // Limit search depth. 2080 2081 switch (Op.getOpcode()) { 2082 default: break; 2083 case ISD::AssertSext: 2084 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits(); 2085 return VTBits-Tmp+1; 2086 case ISD::AssertZext: 2087 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits(); 2088 return VTBits-Tmp; 2089 2090 case ISD::Constant: { 2091 const APInt &Val = cast<ConstantSDNode>(Op)->getAPIntValue(); 2092 return Val.getNumSignBits(); 2093 } 2094 2095 case ISD::SIGN_EXTEND: 2096 Tmp = VTBits-Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 2097 return ComputeNumSignBits(Op.getOperand(0), Depth+1) + Tmp; 2098 2099 case ISD::SIGN_EXTEND_INREG: 2100 // Max of the input and what this extends. 2101 Tmp = 2102 cast<VTSDNode>(Op.getOperand(1))->getVT().getScalarType().getSizeInBits(); 2103 Tmp = VTBits-Tmp+1; 2104 2105 Tmp2 = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2106 return std::max(Tmp, Tmp2); 2107 2108 case ISD::SRA: 2109 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2110 // SRA X, C -> adds C sign bits. 2111 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 2112 Tmp += C->getZExtValue(); 2113 if (Tmp > VTBits) Tmp = VTBits; 2114 } 2115 return Tmp; 2116 case ISD::SHL: 2117 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 2118 // shl destroys sign bits. 2119 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2120 if (C->getZExtValue() >= VTBits || // Bad shift. 2121 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out. 2122 return Tmp - C->getZExtValue(); 2123 } 2124 break; 2125 case ISD::AND: 2126 case ISD::OR: 2127 case ISD::XOR: // NOT is handled here. 2128 // Logical binary ops preserve the number of sign bits at the worst. 2129 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2130 if (Tmp != 1) { 2131 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1); 2132 FirstAnswer = std::min(Tmp, Tmp2); 2133 // We computed what we know about the sign bits as our first 2134 // answer. Now proceed to the generic code that uses 2135 // ComputeMaskedBits, and pick whichever answer is better. 2136 } 2137 break; 2138 2139 case ISD::SELECT: 2140 Tmp = ComputeNumSignBits(Op.getOperand(1), Depth+1); 2141 if (Tmp == 1) return 1; // Early out. 2142 Tmp2 = ComputeNumSignBits(Op.getOperand(2), Depth+1); 2143 return std::min(Tmp, Tmp2); 2144 2145 case ISD::SADDO: 2146 case ISD::UADDO: 2147 case ISD::SSUBO: 2148 case ISD::USUBO: 2149 case ISD::SMULO: 2150 case ISD::UMULO: 2151 if (Op.getResNo() != 1) 2152 break; 2153 // The boolean result conforms to getBooleanContents. Fall through. 2154 case ISD::SETCC: 2155 // If setcc returns 0/-1, all bits are sign bits. 2156 if (TLI.getBooleanContents() == 2157 TargetLowering::ZeroOrNegativeOneBooleanContent) 2158 return VTBits; 2159 break; 2160 case ISD::ROTL: 2161 case ISD::ROTR: 2162 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 2163 unsigned RotAmt = C->getZExtValue() & (VTBits-1); 2164 2165 // Handle rotate right by N like a rotate left by 32-N. 2166 if (Op.getOpcode() == ISD::ROTR) 2167 RotAmt = (VTBits-RotAmt) & (VTBits-1); 2168 2169 // If we aren't rotating out all of the known-in sign bits, return the 2170 // number that are left. This handles rotl(sext(x), 1) for example. 2171 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2172 if (Tmp > RotAmt+1) return Tmp-RotAmt; 2173 } 2174 break; 2175 case ISD::ADD: 2176 // Add can have at most one carry bit. Thus we know that the output 2177 // is, at worst, one more bit than the inputs. 2178 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2179 if (Tmp == 1) return 1; // Early out. 2180 2181 // Special case decrementing a value (ADD X, -1): 2182 if (ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1))) 2183 if (CRHS->isAllOnesValue()) { 2184 APInt KnownZero, KnownOne; 2185 APInt Mask = APInt::getAllOnesValue(VTBits); 2186 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1); 2187 2188 // If the input is known to be 0 or 1, the output is 0/-1, which is all 2189 // sign bits set. 2190 if ((KnownZero | APInt(VTBits, 1)) == Mask) 2191 return VTBits; 2192 2193 // If we are subtracting one from a positive number, there is no carry 2194 // out of the result. 2195 if (KnownZero.isNegative()) 2196 return Tmp; 2197 } 2198 2199 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1); 2200 if (Tmp2 == 1) return 1; 2201 return std::min(Tmp, Tmp2)-1; 2202 break; 2203 2204 case ISD::SUB: 2205 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1); 2206 if (Tmp2 == 1) return 1; 2207 2208 // Handle NEG. 2209 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) 2210 if (CLHS->isNullValue()) { 2211 APInt KnownZero, KnownOne; 2212 APInt Mask = APInt::getAllOnesValue(VTBits); 2213 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1); 2214 // If the input is known to be 0 or 1, the output is 0/-1, which is all 2215 // sign bits set. 2216 if ((KnownZero | APInt(VTBits, 1)) == Mask) 2217 return VTBits; 2218 2219 // If the input is known to be positive (the sign bit is known clear), 2220 // the output of the NEG has the same number of sign bits as the input. 2221 if (KnownZero.isNegative()) 2222 return Tmp2; 2223 2224 // Otherwise, we treat this like a SUB. 2225 } 2226 2227 // Sub can have at most one carry bit. Thus we know that the output 2228 // is, at worst, one more bit than the inputs. 2229 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2230 if (Tmp == 1) return 1; // Early out. 2231 return std::min(Tmp, Tmp2)-1; 2232 break; 2233 case ISD::TRUNCATE: 2234 // FIXME: it's tricky to do anything useful for this, but it is an important 2235 // case for targets like X86. 2236 break; 2237 } 2238 2239 // Handle LOADX separately here. EXTLOAD case will fallthrough. 2240 if (Op.getOpcode() == ISD::LOAD) { 2241 LoadSDNode *LD = cast<LoadSDNode>(Op); 2242 unsigned ExtType = LD->getExtensionType(); 2243 switch (ExtType) { 2244 default: break; 2245 case ISD::SEXTLOAD: // '17' bits known 2246 Tmp = LD->getMemoryVT().getScalarType().getSizeInBits(); 2247 return VTBits-Tmp+1; 2248 case ISD::ZEXTLOAD: // '16' bits known 2249 Tmp = LD->getMemoryVT().getScalarType().getSizeInBits(); 2250 return VTBits-Tmp; 2251 } 2252 } 2253 2254 // Allow the target to implement this method for its nodes. 2255 if (Op.getOpcode() >= ISD::BUILTIN_OP_END || 2256 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 2257 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 2258 Op.getOpcode() == ISD::INTRINSIC_VOID) { 2259 unsigned NumBits = TLI.ComputeNumSignBitsForTargetNode(Op, Depth); 2260 if (NumBits > 1) FirstAnswer = std::max(FirstAnswer, NumBits); 2261 } 2262 2263 // Finally, if we can prove that the top bits of the result are 0's or 1's, 2264 // use this information. 2265 APInt KnownZero, KnownOne; 2266 APInt Mask = APInt::getAllOnesValue(VTBits); 2267 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth); 2268 2269 if (KnownZero.isNegative()) { // sign bit is 0 2270 Mask = KnownZero; 2271 } else if (KnownOne.isNegative()) { // sign bit is 1; 2272 Mask = KnownOne; 2273 } else { 2274 // Nothing known. 2275 return FirstAnswer; 2276 } 2277 2278 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine 2279 // the number of identical bits in the top of the input value. 2280 Mask = ~Mask; 2281 Mask <<= Mask.getBitWidth()-VTBits; 2282 // Return # leading zeros. We use 'min' here in case Val was zero before 2283 // shifting. We don't want to return '64' as for an i32 "0". 2284 return std::max(FirstAnswer, std::min(VTBits, Mask.countLeadingZeros())); 2285} 2286 2287/// isBaseWithConstantOffset - Return true if the specified operand is an 2288/// ISD::ADD with a ConstantSDNode on the right-hand side, or if it is an 2289/// ISD::OR with a ConstantSDNode that is guaranteed to have the same 2290/// semantics as an ADD. This handles the equivalence: 2291/// X|Cst == X+Cst iff X&Cst = 0. 2292bool SelectionDAG::isBaseWithConstantOffset(SDValue Op) const { 2293 if ((Op.getOpcode() != ISD::ADD && Op.getOpcode() != ISD::OR) || 2294 !isa<ConstantSDNode>(Op.getOperand(1))) 2295 return false; 2296 2297 if (Op.getOpcode() == ISD::OR && 2298 !MaskedValueIsZero(Op.getOperand(0), 2299 cast<ConstantSDNode>(Op.getOperand(1))->getAPIntValue())) 2300 return false; 2301 2302 return true; 2303} 2304 2305 2306bool SelectionDAG::isKnownNeverNaN(SDValue Op) const { 2307 // If we're told that NaNs won't happen, assume they won't. 2308 if (NoNaNsFPMath) 2309 return true; 2310 2311 // If the value is a constant, we can obviously see if it is a NaN or not. 2312 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op)) 2313 return !C->getValueAPF().isNaN(); 2314 2315 // TODO: Recognize more cases here. 2316 2317 return false; 2318} 2319 2320bool SelectionDAG::isKnownNeverZero(SDValue Op) const { 2321 // If the value is a constant, we can obviously see if it is a zero or not. 2322 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op)) 2323 return !C->isZero(); 2324 2325 // TODO: Recognize more cases here. 2326 switch (Op.getOpcode()) { 2327 default: break; 2328 case ISD::OR: 2329 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) 2330 return !C->isNullValue(); 2331 break; 2332 } 2333 2334 return false; 2335} 2336 2337bool SelectionDAG::isEqualTo(SDValue A, SDValue B) const { 2338 // Check the obvious case. 2339 if (A == B) return true; 2340 2341 // For for negative and positive zero. 2342 if (const ConstantFPSDNode *CA = dyn_cast<ConstantFPSDNode>(A)) 2343 if (const ConstantFPSDNode *CB = dyn_cast<ConstantFPSDNode>(B)) 2344 if (CA->isZero() && CB->isZero()) return true; 2345 2346 // Otherwise they may not be equal. 2347 return false; 2348} 2349 2350bool SelectionDAG::isVerifiedDebugInfoDesc(SDValue Op) const { 2351 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op); 2352 if (!GA) return false; 2353 if (GA->getOffset() != 0) return false; 2354 const GlobalVariable *GV = dyn_cast<GlobalVariable>(GA->getGlobal()); 2355 if (!GV) return false; 2356 return MF->getMMI().hasDebugInfo(); 2357} 2358 2359 2360/// getNode - Gets or creates the specified node. 2361/// 2362SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT) { 2363 FoldingSetNodeID ID; 2364 AddNodeIDNode(ID, Opcode, getVTList(VT), 0, 0); 2365 void *IP = 0; 2366 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 2367 return SDValue(E, 0); 2368 2369 SDNode *N = new (NodeAllocator) SDNode(Opcode, DL, getVTList(VT)); 2370 CSEMap.InsertNode(N, IP); 2371 2372 AllNodes.push_back(N); 2373#ifndef NDEBUG 2374 VerifySDNode(N); 2375#endif 2376 return SDValue(N, 0); 2377} 2378 2379SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, 2380 EVT VT, SDValue Operand) { 2381 // Constant fold unary operations with an integer constant operand. 2382 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Operand.getNode())) { 2383 const APInt &Val = C->getAPIntValue(); 2384 switch (Opcode) { 2385 default: break; 2386 case ISD::SIGN_EXTEND: 2387 return getConstant(Val.sextOrTrunc(VT.getSizeInBits()), VT); 2388 case ISD::ANY_EXTEND: 2389 case ISD::ZERO_EXTEND: 2390 case ISD::TRUNCATE: 2391 return getConstant(Val.zextOrTrunc(VT.getSizeInBits()), VT); 2392 case ISD::UINT_TO_FP: 2393 case ISD::SINT_TO_FP: { 2394 // No compile time operations on ppcf128. 2395 if (VT == MVT::ppcf128) break; 2396 APFloat apf(APInt::getNullValue(VT.getSizeInBits())); 2397 (void)apf.convertFromAPInt(Val, 2398 Opcode==ISD::SINT_TO_FP, 2399 APFloat::rmNearestTiesToEven); 2400 return getConstantFP(apf, VT); 2401 } 2402 case ISD::BITCAST: 2403 if (VT == MVT::f32 && C->getValueType(0) == MVT::i32) 2404 return getConstantFP(Val.bitsToFloat(), VT); 2405 else if (VT == MVT::f64 && C->getValueType(0) == MVT::i64) 2406 return getConstantFP(Val.bitsToDouble(), VT); 2407 break; 2408 case ISD::BSWAP: 2409 return getConstant(Val.byteSwap(), VT); 2410 case ISD::CTPOP: 2411 return getConstant(Val.countPopulation(), VT); 2412 case ISD::CTLZ: 2413 return getConstant(Val.countLeadingZeros(), VT); 2414 case ISD::CTTZ: 2415 return getConstant(Val.countTrailingZeros(), VT); 2416 } 2417 } 2418 2419 // Constant fold unary operations with a floating point constant operand. 2420 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Operand.getNode())) { 2421 APFloat V = C->getValueAPF(); // make copy 2422 if (VT != MVT::ppcf128 && Operand.getValueType() != MVT::ppcf128) { 2423 switch (Opcode) { 2424 case ISD::FNEG: 2425 V.changeSign(); 2426 return getConstantFP(V, VT); 2427 case ISD::FABS: 2428 V.clearSign(); 2429 return getConstantFP(V, VT); 2430 case ISD::FP_ROUND: 2431 case ISD::FP_EXTEND: { 2432 bool ignored; 2433 // This can return overflow, underflow, or inexact; we don't care. 2434 // FIXME need to be more flexible about rounding mode. 2435 (void)V.convert(*EVTToAPFloatSemantics(VT), 2436 APFloat::rmNearestTiesToEven, &ignored); 2437 return getConstantFP(V, VT); 2438 } 2439 case ISD::FP_TO_SINT: 2440 case ISD::FP_TO_UINT: { 2441 integerPart x[2]; 2442 bool ignored; 2443 assert(integerPartWidth >= 64); 2444 // FIXME need to be more flexible about rounding mode. 2445 APFloat::opStatus s = V.convertToInteger(x, VT.getSizeInBits(), 2446 Opcode==ISD::FP_TO_SINT, 2447 APFloat::rmTowardZero, &ignored); 2448 if (s==APFloat::opInvalidOp) // inexact is OK, in fact usual 2449 break; 2450 APInt api(VT.getSizeInBits(), 2, x); 2451 return getConstant(api, VT); 2452 } 2453 case ISD::BITCAST: 2454 if (VT == MVT::i32 && C->getValueType(0) == MVT::f32) 2455 return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), VT); 2456 else if (VT == MVT::i64 && C->getValueType(0) == MVT::f64) 2457 return getConstant(V.bitcastToAPInt().getZExtValue(), VT); 2458 break; 2459 } 2460 } 2461 } 2462 2463 unsigned OpOpcode = Operand.getNode()->getOpcode(); 2464 switch (Opcode) { 2465 case ISD::TokenFactor: 2466 case ISD::MERGE_VALUES: 2467 case ISD::CONCAT_VECTORS: 2468 return Operand; // Factor, merge or concat of one node? No need. 2469 case ISD::FP_ROUND: llvm_unreachable("Invalid method to make FP_ROUND node"); 2470 case ISD::FP_EXTEND: 2471 assert(VT.isFloatingPoint() && 2472 Operand.getValueType().isFloatingPoint() && "Invalid FP cast!"); 2473 if (Operand.getValueType() == VT) return Operand; // noop conversion. 2474 assert((!VT.isVector() || 2475 VT.getVectorNumElements() == 2476 Operand.getValueType().getVectorNumElements()) && 2477 "Vector element count mismatch!"); 2478 if (Operand.getOpcode() == ISD::UNDEF) 2479 return getUNDEF(VT); 2480 break; 2481 case ISD::SIGN_EXTEND: 2482 assert(VT.isInteger() && Operand.getValueType().isInteger() && 2483 "Invalid SIGN_EXTEND!"); 2484 if (Operand.getValueType() == VT) return Operand; // noop extension 2485 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) && 2486 "Invalid sext node, dst < src!"); 2487 assert((!VT.isVector() || 2488 VT.getVectorNumElements() == 2489 Operand.getValueType().getVectorNumElements()) && 2490 "Vector element count mismatch!"); 2491 if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND) 2492 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0)); 2493 else if (OpOpcode == ISD::UNDEF) 2494 // sext(undef) = 0, because the top bits will all be the same. 2495 return getConstant(0, VT); 2496 break; 2497 case ISD::ZERO_EXTEND: 2498 assert(VT.isInteger() && Operand.getValueType().isInteger() && 2499 "Invalid ZERO_EXTEND!"); 2500 if (Operand.getValueType() == VT) return Operand; // noop extension 2501 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) && 2502 "Invalid zext node, dst < src!"); 2503 assert((!VT.isVector() || 2504 VT.getVectorNumElements() == 2505 Operand.getValueType().getVectorNumElements()) && 2506 "Vector element count mismatch!"); 2507 if (OpOpcode == ISD::ZERO_EXTEND) // (zext (zext x)) -> (zext x) 2508 return getNode(ISD::ZERO_EXTEND, DL, VT, 2509 Operand.getNode()->getOperand(0)); 2510 else if (OpOpcode == ISD::UNDEF) 2511 // zext(undef) = 0, because the top bits will be zero. 2512 return getConstant(0, VT); 2513 break; 2514 case ISD::ANY_EXTEND: 2515 assert(VT.isInteger() && Operand.getValueType().isInteger() && 2516 "Invalid ANY_EXTEND!"); 2517 if (Operand.getValueType() == VT) return Operand; // noop extension 2518 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) && 2519 "Invalid anyext node, dst < src!"); 2520 assert((!VT.isVector() || 2521 VT.getVectorNumElements() == 2522 Operand.getValueType().getVectorNumElements()) && 2523 "Vector element count mismatch!"); 2524 2525 if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND || 2526 OpOpcode == ISD::ANY_EXTEND) 2527 // (ext (zext x)) -> (zext x) and (ext (sext x)) -> (sext x) 2528 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0)); 2529 else if (OpOpcode == ISD::UNDEF) 2530 return getUNDEF(VT); 2531 2532 // (ext (trunx x)) -> x 2533 if (OpOpcode == ISD::TRUNCATE) { 2534 SDValue OpOp = Operand.getNode()->getOperand(0); 2535 if (OpOp.getValueType() == VT) 2536 return OpOp; 2537 } 2538 break; 2539 case ISD::TRUNCATE: 2540 assert(VT.isInteger() && Operand.getValueType().isInteger() && 2541 "Invalid TRUNCATE!"); 2542 if (Operand.getValueType() == VT) return Operand; // noop truncate 2543 assert(Operand.getValueType().getScalarType().bitsGT(VT.getScalarType()) && 2544 "Invalid truncate node, src < dst!"); 2545 assert((!VT.isVector() || 2546 VT.getVectorNumElements() == 2547 Operand.getValueType().getVectorNumElements()) && 2548 "Vector element count mismatch!"); 2549 if (OpOpcode == ISD::TRUNCATE) 2550 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0)); 2551 else if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND || 2552 OpOpcode == ISD::ANY_EXTEND) { 2553 // If the source is smaller than the dest, we still need an extend. 2554 if (Operand.getNode()->getOperand(0).getValueType().getScalarType() 2555 .bitsLT(VT.getScalarType())) 2556 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0)); 2557 else if (Operand.getNode()->getOperand(0).getValueType().bitsGT(VT)) 2558 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0)); 2559 else 2560 return Operand.getNode()->getOperand(0); 2561 } 2562 break; 2563 case ISD::BITCAST: 2564 // Basic sanity checking. 2565 assert(VT.getSizeInBits() == Operand.getValueType().getSizeInBits() 2566 && "Cannot BITCAST between types of different sizes!"); 2567 if (VT == Operand.getValueType()) return Operand; // noop conversion. 2568 if (OpOpcode == ISD::BITCAST) // bitconv(bitconv(x)) -> bitconv(x) 2569 return getNode(ISD::BITCAST, DL, VT, Operand.getOperand(0)); 2570 if (OpOpcode == ISD::UNDEF) 2571 return getUNDEF(VT); 2572 break; 2573 case ISD::SCALAR_TO_VECTOR: 2574 assert(VT.isVector() && !Operand.getValueType().isVector() && 2575 (VT.getVectorElementType() == Operand.getValueType() || 2576 (VT.getVectorElementType().isInteger() && 2577 Operand.getValueType().isInteger() && 2578 VT.getVectorElementType().bitsLE(Operand.getValueType()))) && 2579 "Illegal SCALAR_TO_VECTOR node!"); 2580 if (OpOpcode == ISD::UNDEF) 2581 return getUNDEF(VT); 2582 // scalar_to_vector(extract_vector_elt V, 0) -> V, top bits are undefined. 2583 if (OpOpcode == ISD::EXTRACT_VECTOR_ELT && 2584 isa<ConstantSDNode>(Operand.getOperand(1)) && 2585 Operand.getConstantOperandVal(1) == 0 && 2586 Operand.getOperand(0).getValueType() == VT) 2587 return Operand.getOperand(0); 2588 break; 2589 case ISD::FNEG: 2590 // -(X-Y) -> (Y-X) is unsafe because when X==Y, -0.0 != +0.0 2591 if (UnsafeFPMath && OpOpcode == ISD::FSUB) 2592 return getNode(ISD::FSUB, DL, VT, Operand.getNode()->getOperand(1), 2593 Operand.getNode()->getOperand(0)); 2594 if (OpOpcode == ISD::FNEG) // --X -> X 2595 return Operand.getNode()->getOperand(0); 2596 break; 2597 case ISD::FABS: 2598 if (OpOpcode == ISD::FNEG) // abs(-X) -> abs(X) 2599 return getNode(ISD::FABS, DL, VT, Operand.getNode()->getOperand(0)); 2600 break; 2601 } 2602 2603 SDNode *N; 2604 SDVTList VTs = getVTList(VT); 2605 if (VT != MVT::Glue) { // Don't CSE flag producing nodes 2606 FoldingSetNodeID ID; 2607 SDValue Ops[1] = { Operand }; 2608 AddNodeIDNode(ID, Opcode, VTs, Ops, 1); 2609 void *IP = 0; 2610 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 2611 return SDValue(E, 0); 2612 2613 N = new (NodeAllocator) UnarySDNode(Opcode, DL, VTs, Operand); 2614 CSEMap.InsertNode(N, IP); 2615 } else { 2616 N = new (NodeAllocator) UnarySDNode(Opcode, DL, VTs, Operand); 2617 } 2618 2619 AllNodes.push_back(N); 2620#ifndef NDEBUG 2621 VerifySDNode(N); 2622#endif 2623 return SDValue(N, 0); 2624} 2625 2626SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode, 2627 EVT VT, 2628 ConstantSDNode *Cst1, 2629 ConstantSDNode *Cst2) { 2630 const APInt &C1 = Cst1->getAPIntValue(), &C2 = Cst2->getAPIntValue(); 2631 2632 switch (Opcode) { 2633 case ISD::ADD: return getConstant(C1 + C2, VT); 2634 case ISD::SUB: return getConstant(C1 - C2, VT); 2635 case ISD::MUL: return getConstant(C1 * C2, VT); 2636 case ISD::UDIV: 2637 if (C2.getBoolValue()) return getConstant(C1.udiv(C2), VT); 2638 break; 2639 case ISD::UREM: 2640 if (C2.getBoolValue()) return getConstant(C1.urem(C2), VT); 2641 break; 2642 case ISD::SDIV: 2643 if (C2.getBoolValue()) return getConstant(C1.sdiv(C2), VT); 2644 break; 2645 case ISD::SREM: 2646 if (C2.getBoolValue()) return getConstant(C1.srem(C2), VT); 2647 break; 2648 case ISD::AND: return getConstant(C1 & C2, VT); 2649 case ISD::OR: return getConstant(C1 | C2, VT); 2650 case ISD::XOR: return getConstant(C1 ^ C2, VT); 2651 case ISD::SHL: return getConstant(C1 << C2, VT); 2652 case ISD::SRL: return getConstant(C1.lshr(C2), VT); 2653 case ISD::SRA: return getConstant(C1.ashr(C2), VT); 2654 case ISD::ROTL: return getConstant(C1.rotl(C2), VT); 2655 case ISD::ROTR: return getConstant(C1.rotr(C2), VT); 2656 default: break; 2657 } 2658 2659 return SDValue(); 2660} 2661 2662SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 2663 SDValue N1, SDValue N2) { 2664 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 2665 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode()); 2666 switch (Opcode) { 2667 default: break; 2668 case ISD::TokenFactor: 2669 assert(VT == MVT::Other && N1.getValueType() == MVT::Other && 2670 N2.getValueType() == MVT::Other && "Invalid token factor!"); 2671 // Fold trivial token factors. 2672 if (N1.getOpcode() == ISD::EntryToken) return N2; 2673 if (N2.getOpcode() == ISD::EntryToken) return N1; 2674 if (N1 == N2) return N1; 2675 break; 2676 case ISD::CONCAT_VECTORS: 2677 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to 2678 // one big BUILD_VECTOR. 2679 if (N1.getOpcode() == ISD::BUILD_VECTOR && 2680 N2.getOpcode() == ISD::BUILD_VECTOR) { 2681 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), 2682 N1.getNode()->op_end()); 2683 Elts.append(N2.getNode()->op_begin(), N2.getNode()->op_end()); 2684 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size()); 2685 } 2686 break; 2687 case ISD::AND: 2688 assert(VT.isInteger() && "This operator does not apply to FP types!"); 2689 assert(N1.getValueType() == N2.getValueType() && 2690 N1.getValueType() == VT && "Binary operator types must match!"); 2691 // (X & 0) -> 0. This commonly occurs when legalizing i64 values, so it's 2692 // worth handling here. 2693 if (N2C && N2C->isNullValue()) 2694 return N2; 2695 if (N2C && N2C->isAllOnesValue()) // X & -1 -> X 2696 return N1; 2697 break; 2698 case ISD::OR: 2699 case ISD::XOR: 2700 case ISD::ADD: 2701 case ISD::SUB: 2702 assert(VT.isInteger() && "This operator does not apply to FP types!"); 2703 assert(N1.getValueType() == N2.getValueType() && 2704 N1.getValueType() == VT && "Binary operator types must match!"); 2705 // (X ^|+- 0) -> X. This commonly occurs when legalizing i64 values, so 2706 // it's worth handling here. 2707 if (N2C && N2C->isNullValue()) 2708 return N1; 2709 break; 2710 case ISD::UDIV: 2711 case ISD::UREM: 2712 case ISD::MULHU: 2713 case ISD::MULHS: 2714 case ISD::MUL: 2715 case ISD::SDIV: 2716 case ISD::SREM: 2717 assert(VT.isInteger() && "This operator does not apply to FP types!"); 2718 assert(N1.getValueType() == N2.getValueType() && 2719 N1.getValueType() == VT && "Binary operator types must match!"); 2720 break; 2721 case ISD::FADD: 2722 case ISD::FSUB: 2723 case ISD::FMUL: 2724 case ISD::FDIV: 2725 case ISD::FREM: 2726 if (UnsafeFPMath) { 2727 if (Opcode == ISD::FADD) { 2728 // 0+x --> x 2729 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) 2730 if (CFP->getValueAPF().isZero()) 2731 return N2; 2732 // x+0 --> x 2733 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2)) 2734 if (CFP->getValueAPF().isZero()) 2735 return N1; 2736 } else if (Opcode == ISD::FSUB) { 2737 // x-0 --> x 2738 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2)) 2739 if (CFP->getValueAPF().isZero()) 2740 return N1; 2741 } 2742 } 2743 assert(VT.isFloatingPoint() && "This operator only applies to FP types!"); 2744 assert(N1.getValueType() == N2.getValueType() && 2745 N1.getValueType() == VT && "Binary operator types must match!"); 2746 break; 2747 case ISD::FCOPYSIGN: // N1 and result must match. N1/N2 need not match. 2748 assert(N1.getValueType() == VT && 2749 N1.getValueType().isFloatingPoint() && 2750 N2.getValueType().isFloatingPoint() && 2751 "Invalid FCOPYSIGN!"); 2752 break; 2753 case ISD::SHL: 2754 case ISD::SRA: 2755 case ISD::SRL: 2756 case ISD::ROTL: 2757 case ISD::ROTR: 2758 assert(VT == N1.getValueType() && 2759 "Shift operators return type must be the same as their first arg"); 2760 assert(VT.isInteger() && N2.getValueType().isInteger() && 2761 "Shifts only work on integers"); 2762 // Verify that the shift amount VT is bit enough to hold valid shift 2763 // amounts. This catches things like trying to shift an i1024 value by an 2764 // i8, which is easy to fall into in generic code that uses 2765 // TLI.getShiftAmount(). 2766 assert(N2.getValueType().getSizeInBits() >= 2767 Log2_32_Ceil(N1.getValueType().getSizeInBits()) && 2768 "Invalid use of small shift amount with oversized value!"); 2769 2770 // Always fold shifts of i1 values so the code generator doesn't need to 2771 // handle them. Since we know the size of the shift has to be less than the 2772 // size of the value, the shift/rotate count is guaranteed to be zero. 2773 if (VT == MVT::i1) 2774 return N1; 2775 if (N2C && N2C->isNullValue()) 2776 return N1; 2777 break; 2778 case ISD::FP_ROUND_INREG: { 2779 EVT EVT = cast<VTSDNode>(N2)->getVT(); 2780 assert(VT == N1.getValueType() && "Not an inreg round!"); 2781 assert(VT.isFloatingPoint() && EVT.isFloatingPoint() && 2782 "Cannot FP_ROUND_INREG integer types"); 2783 assert(EVT.isVector() == VT.isVector() && 2784 "FP_ROUND_INREG type should be vector iff the operand " 2785 "type is vector!"); 2786 assert((!EVT.isVector() || 2787 EVT.getVectorNumElements() == VT.getVectorNumElements()) && 2788 "Vector element counts must match in FP_ROUND_INREG"); 2789 assert(EVT.bitsLE(VT) && "Not rounding down!"); 2790 if (cast<VTSDNode>(N2)->getVT() == VT) return N1; // Not actually rounding. 2791 break; 2792 } 2793 case ISD::FP_ROUND: 2794 assert(VT.isFloatingPoint() && 2795 N1.getValueType().isFloatingPoint() && 2796 VT.bitsLE(N1.getValueType()) && 2797 isa<ConstantSDNode>(N2) && "Invalid FP_ROUND!"); 2798 if (N1.getValueType() == VT) return N1; // noop conversion. 2799 break; 2800 case ISD::AssertSext: 2801 case ISD::AssertZext: { 2802 EVT EVT = cast<VTSDNode>(N2)->getVT(); 2803 assert(VT == N1.getValueType() && "Not an inreg extend!"); 2804 assert(VT.isInteger() && EVT.isInteger() && 2805 "Cannot *_EXTEND_INREG FP types"); 2806 assert(!EVT.isVector() && 2807 "AssertSExt/AssertZExt type should be the vector element type " 2808 "rather than the vector type!"); 2809 assert(EVT.bitsLE(VT) && "Not extending!"); 2810 if (VT == EVT) return N1; // noop assertion. 2811 break; 2812 } 2813 case ISD::SIGN_EXTEND_INREG: { 2814 EVT EVT = cast<VTSDNode>(N2)->getVT(); 2815 assert(VT == N1.getValueType() && "Not an inreg extend!"); 2816 assert(VT.isInteger() && EVT.isInteger() && 2817 "Cannot *_EXTEND_INREG FP types"); 2818 assert(EVT.isVector() == VT.isVector() && 2819 "SIGN_EXTEND_INREG type should be vector iff the operand " 2820 "type is vector!"); 2821 assert((!EVT.isVector() || 2822 EVT.getVectorNumElements() == VT.getVectorNumElements()) && 2823 "Vector element counts must match in SIGN_EXTEND_INREG"); 2824 assert(EVT.bitsLE(VT) && "Not extending!"); 2825 if (EVT == VT) return N1; // Not actually extending 2826 2827 if (N1C) { 2828 APInt Val = N1C->getAPIntValue(); 2829 unsigned FromBits = EVT.getScalarType().getSizeInBits(); 2830 Val <<= Val.getBitWidth()-FromBits; 2831 Val = Val.ashr(Val.getBitWidth()-FromBits); 2832 return getConstant(Val, VT); 2833 } 2834 break; 2835 } 2836 case ISD::EXTRACT_VECTOR_ELT: 2837 // EXTRACT_VECTOR_ELT of an UNDEF is an UNDEF. 2838 if (N1.getOpcode() == ISD::UNDEF) 2839 return getUNDEF(VT); 2840 2841 // EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is 2842 // expanding copies of large vectors from registers. 2843 if (N2C && 2844 N1.getOpcode() == ISD::CONCAT_VECTORS && 2845 N1.getNumOperands() > 0) { 2846 unsigned Factor = 2847 N1.getOperand(0).getValueType().getVectorNumElements(); 2848 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, 2849 N1.getOperand(N2C->getZExtValue() / Factor), 2850 getConstant(N2C->getZExtValue() % Factor, 2851 N2.getValueType())); 2852 } 2853 2854 // EXTRACT_VECTOR_ELT of BUILD_VECTOR is often formed while lowering is 2855 // expanding large vector constants. 2856 if (N2C && N1.getOpcode() == ISD::BUILD_VECTOR) { 2857 SDValue Elt = N1.getOperand(N2C->getZExtValue()); 2858 EVT VEltTy = N1.getValueType().getVectorElementType(); 2859 if (Elt.getValueType() != VEltTy) { 2860 // If the vector element type is not legal, the BUILD_VECTOR operands 2861 // are promoted and implicitly truncated. Make that explicit here. 2862 Elt = getNode(ISD::TRUNCATE, DL, VEltTy, Elt); 2863 } 2864 if (VT != VEltTy) { 2865 // If the vector element type is not legal, the EXTRACT_VECTOR_ELT 2866 // result is implicitly extended. 2867 Elt = getNode(ISD::ANY_EXTEND, DL, VT, Elt); 2868 } 2869 return Elt; 2870 } 2871 2872 // EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector 2873 // operations are lowered to scalars. 2874 if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) { 2875 // If the indices are the same, return the inserted element else 2876 // if the indices are known different, extract the element from 2877 // the original vector. 2878 SDValue N1Op2 = N1.getOperand(2); 2879 ConstantSDNode *N1Op2C = dyn_cast<ConstantSDNode>(N1Op2.getNode()); 2880 2881 if (N1Op2C && N2C) { 2882 if (N1Op2C->getZExtValue() == N2C->getZExtValue()) { 2883 if (VT == N1.getOperand(1).getValueType()) 2884 return N1.getOperand(1); 2885 else 2886 return getSExtOrTrunc(N1.getOperand(1), DL, VT); 2887 } 2888 2889 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, N1.getOperand(0), N2); 2890 } 2891 } 2892 break; 2893 case ISD::EXTRACT_ELEMENT: 2894 assert(N2C && (unsigned)N2C->getZExtValue() < 2 && "Bad EXTRACT_ELEMENT!"); 2895 assert(!N1.getValueType().isVector() && !VT.isVector() && 2896 (N1.getValueType().isInteger() == VT.isInteger()) && 2897 "Wrong types for EXTRACT_ELEMENT!"); 2898 2899 // EXTRACT_ELEMENT of BUILD_PAIR is often formed while legalize is expanding 2900 // 64-bit integers into 32-bit parts. Instead of building the extract of 2901 // the BUILD_PAIR, only to have legalize rip it apart, just do it now. 2902 if (N1.getOpcode() == ISD::BUILD_PAIR) 2903 return N1.getOperand(N2C->getZExtValue()); 2904 2905 // EXTRACT_ELEMENT of a constant int is also very common. 2906 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) { 2907 unsigned ElementSize = VT.getSizeInBits(); 2908 unsigned Shift = ElementSize * N2C->getZExtValue(); 2909 APInt ShiftedVal = C->getAPIntValue().lshr(Shift); 2910 return getConstant(ShiftedVal.trunc(ElementSize), VT); 2911 } 2912 break; 2913 case ISD::EXTRACT_SUBVECTOR: { 2914 SDValue Index = N2; 2915 if (VT.isSimple() && N1.getValueType().isSimple()) { 2916 assert(VT.isVector() && N1.getValueType().isVector() && 2917 "Extract subvector VTs must be a vectors!"); 2918 assert(VT.getVectorElementType() == N1.getValueType().getVectorElementType() && 2919 "Extract subvector VTs must have the same element type!"); 2920 assert(VT.getSimpleVT() <= N1.getValueType().getSimpleVT() && 2921 "Extract subvector must be from larger vector to smaller vector!"); 2922 2923 if (isa<ConstantSDNode>(Index.getNode())) { 2924 assert((VT.getVectorNumElements() + 2925 cast<ConstantSDNode>(Index.getNode())->getZExtValue() 2926 <= N1.getValueType().getVectorNumElements()) 2927 && "Extract subvector overflow!"); 2928 } 2929 2930 // Trivial extraction. 2931 if (VT.getSimpleVT() == N1.getValueType().getSimpleVT()) 2932 return N1; 2933 } 2934 break; 2935 } 2936 } 2937 2938 if (N1C) { 2939 if (N2C) { 2940 SDValue SV = FoldConstantArithmetic(Opcode, VT, N1C, N2C); 2941 if (SV.getNode()) return SV; 2942 } else { // Cannonicalize constant to RHS if commutative 2943 if (isCommutativeBinOp(Opcode)) { 2944 std::swap(N1C, N2C); 2945 std::swap(N1, N2); 2946 } 2947 } 2948 } 2949 2950 // Constant fold FP operations. 2951 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode()); 2952 ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode()); 2953 if (N1CFP) { 2954 if (!N2CFP && isCommutativeBinOp(Opcode)) { 2955 // Cannonicalize constant to RHS if commutative 2956 std::swap(N1CFP, N2CFP); 2957 std::swap(N1, N2); 2958 } else if (N2CFP && VT != MVT::ppcf128) { 2959 APFloat V1 = N1CFP->getValueAPF(), V2 = N2CFP->getValueAPF(); 2960 APFloat::opStatus s; 2961 switch (Opcode) { 2962 case ISD::FADD: 2963 s = V1.add(V2, APFloat::rmNearestTiesToEven); 2964 if (s != APFloat::opInvalidOp) 2965 return getConstantFP(V1, VT); 2966 break; 2967 case ISD::FSUB: 2968 s = V1.subtract(V2, APFloat::rmNearestTiesToEven); 2969 if (s!=APFloat::opInvalidOp) 2970 return getConstantFP(V1, VT); 2971 break; 2972 case ISD::FMUL: 2973 s = V1.multiply(V2, APFloat::rmNearestTiesToEven); 2974 if (s!=APFloat::opInvalidOp) 2975 return getConstantFP(V1, VT); 2976 break; 2977 case ISD::FDIV: 2978 s = V1.divide(V2, APFloat::rmNearestTiesToEven); 2979 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero) 2980 return getConstantFP(V1, VT); 2981 break; 2982 case ISD::FREM : 2983 s = V1.mod(V2, APFloat::rmNearestTiesToEven); 2984 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero) 2985 return getConstantFP(V1, VT); 2986 break; 2987 case ISD::FCOPYSIGN: 2988 V1.copySign(V2); 2989 return getConstantFP(V1, VT); 2990 default: break; 2991 } 2992 } 2993 } 2994 2995 // Canonicalize an UNDEF to the RHS, even over a constant. 2996 if (N1.getOpcode() == ISD::UNDEF) { 2997 if (isCommutativeBinOp(Opcode)) { 2998 std::swap(N1, N2); 2999 } else { 3000 switch (Opcode) { 3001 case ISD::FP_ROUND_INREG: 3002 case ISD::SIGN_EXTEND_INREG: 3003 case ISD::SUB: 3004 case ISD::FSUB: 3005 case ISD::FDIV: 3006 case ISD::FREM: 3007 case ISD::SRA: 3008 return N1; // fold op(undef, arg2) -> undef 3009 case ISD::UDIV: 3010 case ISD::SDIV: 3011 case ISD::UREM: 3012 case ISD::SREM: 3013 case ISD::SRL: 3014 case ISD::SHL: 3015 if (!VT.isVector()) 3016 return getConstant(0, VT); // fold op(undef, arg2) -> 0 3017 // For vectors, we can't easily build an all zero vector, just return 3018 // the LHS. 3019 return N2; 3020 } 3021 } 3022 } 3023 3024 // Fold a bunch of operators when the RHS is undef. 3025 if (N2.getOpcode() == ISD::UNDEF) { 3026 switch (Opcode) { 3027 case ISD::XOR: 3028 if (N1.getOpcode() == ISD::UNDEF) 3029 // Handle undef ^ undef -> 0 special case. This is a common 3030 // idiom (misuse). 3031 return getConstant(0, VT); 3032 // fallthrough 3033 case ISD::ADD: 3034 case ISD::ADDC: 3035 case ISD::ADDE: 3036 case ISD::SUB: 3037 case ISD::UDIV: 3038 case ISD::SDIV: 3039 case ISD::UREM: 3040 case ISD::SREM: 3041 return N2; // fold op(arg1, undef) -> undef 3042 case ISD::FADD: 3043 case ISD::FSUB: 3044 case ISD::FMUL: 3045 case ISD::FDIV: 3046 case ISD::FREM: 3047 if (UnsafeFPMath) 3048 return N2; 3049 break; 3050 case ISD::MUL: 3051 case ISD::AND: 3052 case ISD::SRL: 3053 case ISD::SHL: 3054 if (!VT.isVector()) 3055 return getConstant(0, VT); // fold op(arg1, undef) -> 0 3056 // For vectors, we can't easily build an all zero vector, just return 3057 // the LHS. 3058 return N1; 3059 case ISD::OR: 3060 if (!VT.isVector()) 3061 return getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT); 3062 // For vectors, we can't easily build an all one vector, just return 3063 // the LHS. 3064 return N1; 3065 case ISD::SRA: 3066 return N1; 3067 } 3068 } 3069 3070 // Memoize this node if possible. 3071 SDNode *N; 3072 SDVTList VTs = getVTList(VT); 3073 if (VT != MVT::Glue) { 3074 SDValue Ops[] = { N1, N2 }; 3075 FoldingSetNodeID ID; 3076 AddNodeIDNode(ID, Opcode, VTs, Ops, 2); 3077 void *IP = 0; 3078 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 3079 return SDValue(E, 0); 3080 3081 N = new (NodeAllocator) BinarySDNode(Opcode, DL, VTs, N1, N2); 3082 CSEMap.InsertNode(N, IP); 3083 } else { 3084 N = new (NodeAllocator) BinarySDNode(Opcode, DL, VTs, N1, N2); 3085 } 3086 3087 AllNodes.push_back(N); 3088#ifndef NDEBUG 3089 VerifySDNode(N); 3090#endif 3091 return SDValue(N, 0); 3092} 3093 3094SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 3095 SDValue N1, SDValue N2, SDValue N3) { 3096 // Perform various simplifications. 3097 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 3098 switch (Opcode) { 3099 case ISD::CONCAT_VECTORS: 3100 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to 3101 // one big BUILD_VECTOR. 3102 if (N1.getOpcode() == ISD::BUILD_VECTOR && 3103 N2.getOpcode() == ISD::BUILD_VECTOR && 3104 N3.getOpcode() == ISD::BUILD_VECTOR) { 3105 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), 3106 N1.getNode()->op_end()); 3107 Elts.append(N2.getNode()->op_begin(), N2.getNode()->op_end()); 3108 Elts.append(N3.getNode()->op_begin(), N3.getNode()->op_end()); 3109 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size()); 3110 } 3111 break; 3112 case ISD::SETCC: { 3113 // Use FoldSetCC to simplify SETCC's. 3114 SDValue Simp = FoldSetCC(VT, N1, N2, cast<CondCodeSDNode>(N3)->get(), DL); 3115 if (Simp.getNode()) return Simp; 3116 break; 3117 } 3118 case ISD::SELECT: 3119 if (N1C) { 3120 if (N1C->getZExtValue()) 3121 return N2; // select true, X, Y -> X 3122 else 3123 return N3; // select false, X, Y -> Y 3124 } 3125 3126 if (N2 == N3) return N2; // select C, X, X -> X 3127 break; 3128 case ISD::VECTOR_SHUFFLE: 3129 llvm_unreachable("should use getVectorShuffle constructor!"); 3130 break; 3131 case ISD::INSERT_SUBVECTOR: { 3132 SDValue Index = N3; 3133 if (VT.isSimple() && N1.getValueType().isSimple() 3134 && N2.getValueType().isSimple()) { 3135 assert(VT.isVector() && N1.getValueType().isVector() && 3136 N2.getValueType().isVector() && 3137 "Insert subvector VTs must be a vectors"); 3138 assert(VT == N1.getValueType() && 3139 "Dest and insert subvector source types must match!"); 3140 assert(N2.getValueType().getSimpleVT() <= N1.getValueType().getSimpleVT() && 3141 "Insert subvector must be from smaller vector to larger vector!"); 3142 if (isa<ConstantSDNode>(Index.getNode())) { 3143 assert((N2.getValueType().getVectorNumElements() + 3144 cast<ConstantSDNode>(Index.getNode())->getZExtValue() 3145 <= VT.getVectorNumElements()) 3146 && "Insert subvector overflow!"); 3147 } 3148 3149 // Trivial insertion. 3150 if (VT.getSimpleVT() == N2.getValueType().getSimpleVT()) 3151 return N2; 3152 } 3153 break; 3154 } 3155 case ISD::BITCAST: 3156 // Fold bit_convert nodes from a type to themselves. 3157 if (N1.getValueType() == VT) 3158 return N1; 3159 break; 3160 } 3161 3162 // Memoize node if it doesn't produce a flag. 3163 SDNode *N; 3164 SDVTList VTs = getVTList(VT); 3165 if (VT != MVT::Glue) { 3166 SDValue Ops[] = { N1, N2, N3 }; 3167 FoldingSetNodeID ID; 3168 AddNodeIDNode(ID, Opcode, VTs, Ops, 3); 3169 void *IP = 0; 3170 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 3171 return SDValue(E, 0); 3172 3173 N = new (NodeAllocator) TernarySDNode(Opcode, DL, VTs, N1, N2, N3); 3174 CSEMap.InsertNode(N, IP); 3175 } else { 3176 N = new (NodeAllocator) TernarySDNode(Opcode, DL, VTs, N1, N2, N3); 3177 } 3178 3179 AllNodes.push_back(N); 3180#ifndef NDEBUG 3181 VerifySDNode(N); 3182#endif 3183 return SDValue(N, 0); 3184} 3185 3186SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 3187 SDValue N1, SDValue N2, SDValue N3, 3188 SDValue N4) { 3189 SDValue Ops[] = { N1, N2, N3, N4 }; 3190 return getNode(Opcode, DL, VT, Ops, 4); 3191} 3192 3193SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 3194 SDValue N1, SDValue N2, SDValue N3, 3195 SDValue N4, SDValue N5) { 3196 SDValue Ops[] = { N1, N2, N3, N4, N5 }; 3197 return getNode(Opcode, DL, VT, Ops, 5); 3198} 3199 3200/// getStackArgumentTokenFactor - Compute a TokenFactor to force all 3201/// the incoming stack arguments to be loaded from the stack. 3202SDValue SelectionDAG::getStackArgumentTokenFactor(SDValue Chain) { 3203 SmallVector<SDValue, 8> ArgChains; 3204 3205 // Include the original chain at the beginning of the list. When this is 3206 // used by target LowerCall hooks, this helps legalize find the 3207 // CALLSEQ_BEGIN node. 3208 ArgChains.push_back(Chain); 3209 3210 // Add a chain value for each stack argument. 3211 for (SDNode::use_iterator U = getEntryNode().getNode()->use_begin(), 3212 UE = getEntryNode().getNode()->use_end(); U != UE; ++U) 3213 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U)) 3214 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr())) 3215 if (FI->getIndex() < 0) 3216 ArgChains.push_back(SDValue(L, 1)); 3217 3218 // Build a tokenfactor for all the chains. 3219 return getNode(ISD::TokenFactor, Chain.getDebugLoc(), MVT::Other, 3220 &ArgChains[0], ArgChains.size()); 3221} 3222 3223/// SplatByte - Distribute ByteVal over NumBits bits. 3224static APInt SplatByte(unsigned NumBits, uint8_t ByteVal) { 3225 APInt Val = APInt(NumBits, ByteVal); 3226 unsigned Shift = 8; 3227 for (unsigned i = NumBits; i > 8; i >>= 1) { 3228 Val = (Val << Shift) | Val; 3229 Shift <<= 1; 3230 } 3231 return Val; 3232} 3233 3234/// getMemsetValue - Vectorized representation of the memset value 3235/// operand. 3236static SDValue getMemsetValue(SDValue Value, EVT VT, SelectionDAG &DAG, 3237 DebugLoc dl) { 3238 assert(Value.getOpcode() != ISD::UNDEF); 3239 3240 unsigned NumBits = VT.getScalarType().getSizeInBits(); 3241 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) { 3242 APInt Val = SplatByte(NumBits, C->getZExtValue() & 255); 3243 if (VT.isInteger()) 3244 return DAG.getConstant(Val, VT); 3245 return DAG.getConstantFP(APFloat(Val), VT); 3246 } 3247 3248 Value = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Value); 3249 if (NumBits > 8) { 3250 // Use a multiplication with 0x010101... to extend the input to the 3251 // required length. 3252 APInt Magic = SplatByte(NumBits, 0x01); 3253 Value = DAG.getNode(ISD::MUL, dl, VT, Value, DAG.getConstant(Magic, VT)); 3254 } 3255 3256 return Value; 3257} 3258 3259/// getMemsetStringVal - Similar to getMemsetValue. Except this is only 3260/// used when a memcpy is turned into a memset when the source is a constant 3261/// string ptr. 3262static SDValue getMemsetStringVal(EVT VT, DebugLoc dl, SelectionDAG &DAG, 3263 const TargetLowering &TLI, 3264 std::string &Str, unsigned Offset) { 3265 // Handle vector with all elements zero. 3266 if (Str.empty()) { 3267 if (VT.isInteger()) 3268 return DAG.getConstant(0, VT); 3269 else if (VT == MVT::f32 || VT == MVT::f64) 3270 return DAG.getConstantFP(0.0, VT); 3271 else if (VT.isVector()) { 3272 unsigned NumElts = VT.getVectorNumElements(); 3273 MVT EltVT = (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64; 3274 return DAG.getNode(ISD::BITCAST, dl, VT, 3275 DAG.getConstant(0, EVT::getVectorVT(*DAG.getContext(), 3276 EltVT, NumElts))); 3277 } else 3278 llvm_unreachable("Expected type!"); 3279 } 3280 3281 assert(!VT.isVector() && "Can't handle vector type here!"); 3282 unsigned NumBits = VT.getSizeInBits(); 3283 unsigned MSB = NumBits / 8; 3284 uint64_t Val = 0; 3285 if (TLI.isLittleEndian()) 3286 Offset = Offset + MSB - 1; 3287 for (unsigned i = 0; i != MSB; ++i) { 3288 Val = (Val << 8) | (unsigned char)Str[Offset]; 3289 Offset += TLI.isLittleEndian() ? -1 : 1; 3290 } 3291 return DAG.getConstant(Val, VT); 3292} 3293 3294/// getMemBasePlusOffset - Returns base and offset node for the 3295/// 3296static SDValue getMemBasePlusOffset(SDValue Base, unsigned Offset, 3297 SelectionDAG &DAG) { 3298 EVT VT = Base.getValueType(); 3299 return DAG.getNode(ISD::ADD, Base.getDebugLoc(), 3300 VT, Base, DAG.getConstant(Offset, VT)); 3301} 3302 3303/// isMemSrcFromString - Returns true if memcpy source is a string constant. 3304/// 3305static bool isMemSrcFromString(SDValue Src, std::string &Str) { 3306 unsigned SrcDelta = 0; 3307 GlobalAddressSDNode *G = NULL; 3308 if (Src.getOpcode() == ISD::GlobalAddress) 3309 G = cast<GlobalAddressSDNode>(Src); 3310 else if (Src.getOpcode() == ISD::ADD && 3311 Src.getOperand(0).getOpcode() == ISD::GlobalAddress && 3312 Src.getOperand(1).getOpcode() == ISD::Constant) { 3313 G = cast<GlobalAddressSDNode>(Src.getOperand(0)); 3314 SrcDelta = cast<ConstantSDNode>(Src.getOperand(1))->getZExtValue(); 3315 } 3316 if (!G) 3317 return false; 3318 3319 const GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal()); 3320 if (GV && GetConstantStringInfo(GV, Str, SrcDelta, false)) 3321 return true; 3322 3323 return false; 3324} 3325 3326/// FindOptimalMemOpLowering - Determines the optimial series memory ops 3327/// to replace the memset / memcpy. Return true if the number of memory ops 3328/// is below the threshold. It returns the types of the sequence of 3329/// memory ops to perform memset / memcpy by reference. 3330static bool FindOptimalMemOpLowering(std::vector<EVT> &MemOps, 3331 unsigned Limit, uint64_t Size, 3332 unsigned DstAlign, unsigned SrcAlign, 3333 bool NonScalarIntSafe, 3334 bool MemcpyStrSrc, 3335 SelectionDAG &DAG, 3336 const TargetLowering &TLI) { 3337 assert((SrcAlign == 0 || SrcAlign >= DstAlign) && 3338 "Expecting memcpy / memset source to meet alignment requirement!"); 3339 // If 'SrcAlign' is zero, that means the memory operation does not need load 3340 // the value, i.e. memset or memcpy from constant string. Otherwise, it's 3341 // the inferred alignment of the source. 'DstAlign', on the other hand, is the 3342 // specified alignment of the memory operation. If it is zero, that means 3343 // it's possible to change the alignment of the destination. 'MemcpyStrSrc' 3344 // indicates whether the memcpy source is constant so it does not need to be 3345 // loaded. 3346 EVT VT = TLI.getOptimalMemOpType(Size, DstAlign, SrcAlign, 3347 NonScalarIntSafe, MemcpyStrSrc, 3348 DAG.getMachineFunction()); 3349 3350 if (VT == MVT::Other) { 3351 if (DstAlign >= TLI.getTargetData()->getPointerPrefAlignment() || 3352 TLI.allowsUnalignedMemoryAccesses(VT)) { 3353 VT = TLI.getPointerTy(); 3354 } else { 3355 switch (DstAlign & 7) { 3356 case 0: VT = MVT::i64; break; 3357 case 4: VT = MVT::i32; break; 3358 case 2: VT = MVT::i16; break; 3359 default: VT = MVT::i8; break; 3360 } 3361 } 3362 3363 MVT LVT = MVT::i64; 3364 while (!TLI.isTypeLegal(LVT)) 3365 LVT = (MVT::SimpleValueType)(LVT.SimpleTy - 1); 3366 assert(LVT.isInteger()); 3367 3368 if (VT.bitsGT(LVT)) 3369 VT = LVT; 3370 } 3371 3372 unsigned NumMemOps = 0; 3373 while (Size != 0) { 3374 unsigned VTSize = VT.getSizeInBits() / 8; 3375 while (VTSize > Size) { 3376 // For now, only use non-vector load / store's for the left-over pieces. 3377 if (VT.isVector() || VT.isFloatingPoint()) { 3378 VT = MVT::i64; 3379 while (!TLI.isTypeLegal(VT)) 3380 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1); 3381 VTSize = VT.getSizeInBits() / 8; 3382 } else { 3383 // This can result in a type that is not legal on the target, e.g. 3384 // 1 or 2 bytes on PPC. 3385 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1); 3386 VTSize >>= 1; 3387 } 3388 } 3389 3390 if (++NumMemOps > Limit) 3391 return false; 3392 MemOps.push_back(VT); 3393 Size -= VTSize; 3394 } 3395 3396 return true; 3397} 3398 3399static SDValue getMemcpyLoadsAndStores(SelectionDAG &DAG, DebugLoc dl, 3400 SDValue Chain, SDValue Dst, 3401 SDValue Src, uint64_t Size, 3402 unsigned Align, bool isVol, 3403 bool AlwaysInline, 3404 MachinePointerInfo DstPtrInfo, 3405 MachinePointerInfo SrcPtrInfo) { 3406 // Turn a memcpy of undef to nop. 3407 if (Src.getOpcode() == ISD::UNDEF) 3408 return Chain; 3409 3410 // Expand memcpy to a series of load and store ops if the size operand falls 3411 // below a certain threshold. 3412 // TODO: In the AlwaysInline case, if the size is big then generate a loop 3413 // rather than maybe a humongous number of loads and stores. 3414 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3415 std::vector<EVT> MemOps; 3416 bool DstAlignCanChange = false; 3417 MachineFunction &MF = DAG.getMachineFunction(); 3418 MachineFrameInfo *MFI = MF.getFrameInfo(); 3419 bool OptSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize); 3420 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst); 3421 if (FI && !MFI->isFixedObjectIndex(FI->getIndex())) 3422 DstAlignCanChange = true; 3423 unsigned SrcAlign = DAG.InferPtrAlignment(Src); 3424 if (Align > SrcAlign) 3425 SrcAlign = Align; 3426 std::string Str; 3427 bool CopyFromStr = isMemSrcFromString(Src, Str); 3428 bool isZeroStr = CopyFromStr && Str.empty(); 3429 unsigned Limit = AlwaysInline ? ~0U : TLI.getMaxStoresPerMemcpy(OptSize); 3430 3431 if (!FindOptimalMemOpLowering(MemOps, Limit, Size, 3432 (DstAlignCanChange ? 0 : Align), 3433 (isZeroStr ? 0 : SrcAlign), 3434 true, CopyFromStr, DAG, TLI)) 3435 return SDValue(); 3436 3437 if (DstAlignCanChange) { 3438 const Type *Ty = MemOps[0].getTypeForEVT(*DAG.getContext()); 3439 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty); 3440 if (NewAlign > Align) { 3441 // Give the stack frame object a larger alignment if needed. 3442 if (MFI->getObjectAlignment(FI->getIndex()) < NewAlign) 3443 MFI->setObjectAlignment(FI->getIndex(), NewAlign); 3444 Align = NewAlign; 3445 } 3446 } 3447 3448 SmallVector<SDValue, 8> OutChains; 3449 unsigned NumMemOps = MemOps.size(); 3450 uint64_t SrcOff = 0, DstOff = 0; 3451 for (unsigned i = 0; i != NumMemOps; ++i) { 3452 EVT VT = MemOps[i]; 3453 unsigned VTSize = VT.getSizeInBits() / 8; 3454 SDValue Value, Store; 3455 3456 if (CopyFromStr && 3457 (isZeroStr || (VT.isInteger() && !VT.isVector()))) { 3458 // It's unlikely a store of a vector immediate can be done in a single 3459 // instruction. It would require a load from a constantpool first. 3460 // We only handle zero vectors here. 3461 // FIXME: Handle other cases where store of vector immediate is done in 3462 // a single instruction. 3463 Value = getMemsetStringVal(VT, dl, DAG, TLI, Str, SrcOff); 3464 Store = DAG.getStore(Chain, dl, Value, 3465 getMemBasePlusOffset(Dst, DstOff, DAG), 3466 DstPtrInfo.getWithOffset(DstOff), isVol, 3467 false, Align); 3468 } else { 3469 // The type might not be legal for the target. This should only happen 3470 // if the type is smaller than a legal type, as on PPC, so the right 3471 // thing to do is generate a LoadExt/StoreTrunc pair. These simplify 3472 // to Load/Store if NVT==VT. 3473 // FIXME does the case above also need this? 3474 EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT); 3475 assert(NVT.bitsGE(VT)); 3476 Value = DAG.getExtLoad(ISD::EXTLOAD, dl, NVT, Chain, 3477 getMemBasePlusOffset(Src, SrcOff, DAG), 3478 SrcPtrInfo.getWithOffset(SrcOff), VT, isVol, false, 3479 MinAlign(SrcAlign, SrcOff)); 3480 Store = DAG.getTruncStore(Chain, dl, Value, 3481 getMemBasePlusOffset(Dst, DstOff, DAG), 3482 DstPtrInfo.getWithOffset(DstOff), VT, isVol, 3483 false, Align); 3484 } 3485 OutChains.push_back(Store); 3486 SrcOff += VTSize; 3487 DstOff += VTSize; 3488 } 3489 3490 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 3491 &OutChains[0], OutChains.size()); 3492} 3493 3494static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, DebugLoc dl, 3495 SDValue Chain, SDValue Dst, 3496 SDValue Src, uint64_t Size, 3497 unsigned Align, bool isVol, 3498 bool AlwaysInline, 3499 MachinePointerInfo DstPtrInfo, 3500 MachinePointerInfo SrcPtrInfo) { 3501 // Turn a memmove of undef to nop. 3502 if (Src.getOpcode() == ISD::UNDEF) 3503 return Chain; 3504 3505 // Expand memmove to a series of load and store ops if the size operand falls 3506 // below a certain threshold. 3507 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3508 std::vector<EVT> MemOps; 3509 bool DstAlignCanChange = false; 3510 MachineFunction &MF = DAG.getMachineFunction(); 3511 MachineFrameInfo *MFI = MF.getFrameInfo(); 3512 bool OptSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize); 3513 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst); 3514 if (FI && !MFI->isFixedObjectIndex(FI->getIndex())) 3515 DstAlignCanChange = true; 3516 unsigned SrcAlign = DAG.InferPtrAlignment(Src); 3517 if (Align > SrcAlign) 3518 SrcAlign = Align; 3519 unsigned Limit = AlwaysInline ? ~0U : TLI.getMaxStoresPerMemmove(OptSize); 3520 3521 if (!FindOptimalMemOpLowering(MemOps, Limit, Size, 3522 (DstAlignCanChange ? 0 : Align), 3523 SrcAlign, true, false, DAG, TLI)) 3524 return SDValue(); 3525 3526 if (DstAlignCanChange) { 3527 const Type *Ty = MemOps[0].getTypeForEVT(*DAG.getContext()); 3528 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty); 3529 if (NewAlign > Align) { 3530 // Give the stack frame object a larger alignment if needed. 3531 if (MFI->getObjectAlignment(FI->getIndex()) < NewAlign) 3532 MFI->setObjectAlignment(FI->getIndex(), NewAlign); 3533 Align = NewAlign; 3534 } 3535 } 3536 3537 uint64_t SrcOff = 0, DstOff = 0; 3538 SmallVector<SDValue, 8> LoadValues; 3539 SmallVector<SDValue, 8> LoadChains; 3540 SmallVector<SDValue, 8> OutChains; 3541 unsigned NumMemOps = MemOps.size(); 3542 for (unsigned i = 0; i < NumMemOps; i++) { 3543 EVT VT = MemOps[i]; 3544 unsigned VTSize = VT.getSizeInBits() / 8; 3545 SDValue Value, Store; 3546 3547 Value = DAG.getLoad(VT, dl, Chain, 3548 getMemBasePlusOffset(Src, SrcOff, DAG), 3549 SrcPtrInfo.getWithOffset(SrcOff), isVol, 3550 false, SrcAlign); 3551 LoadValues.push_back(Value); 3552 LoadChains.push_back(Value.getValue(1)); 3553 SrcOff += VTSize; 3554 } 3555 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 3556 &LoadChains[0], LoadChains.size()); 3557 OutChains.clear(); 3558 for (unsigned i = 0; i < NumMemOps; i++) { 3559 EVT VT = MemOps[i]; 3560 unsigned VTSize = VT.getSizeInBits() / 8; 3561 SDValue Value, Store; 3562 3563 Store = DAG.getStore(Chain, dl, LoadValues[i], 3564 getMemBasePlusOffset(Dst, DstOff, DAG), 3565 DstPtrInfo.getWithOffset(DstOff), isVol, false, Align); 3566 OutChains.push_back(Store); 3567 DstOff += VTSize; 3568 } 3569 3570 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 3571 &OutChains[0], OutChains.size()); 3572} 3573 3574static SDValue getMemsetStores(SelectionDAG &DAG, DebugLoc dl, 3575 SDValue Chain, SDValue Dst, 3576 SDValue Src, uint64_t Size, 3577 unsigned Align, bool isVol, 3578 MachinePointerInfo DstPtrInfo) { 3579 // Turn a memset of undef to nop. 3580 if (Src.getOpcode() == ISD::UNDEF) 3581 return Chain; 3582 3583 // Expand memset to a series of load/store ops if the size operand 3584 // falls below a certain threshold. 3585 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3586 std::vector<EVT> MemOps; 3587 bool DstAlignCanChange = false; 3588 MachineFunction &MF = DAG.getMachineFunction(); 3589 MachineFrameInfo *MFI = MF.getFrameInfo(); 3590 bool OptSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize); 3591 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Dst); 3592 if (FI && !MFI->isFixedObjectIndex(FI->getIndex())) 3593 DstAlignCanChange = true; 3594 bool NonScalarIntSafe = 3595 isa<ConstantSDNode>(Src) && cast<ConstantSDNode>(Src)->isNullValue(); 3596 if (!FindOptimalMemOpLowering(MemOps, TLI.getMaxStoresPerMemset(OptSize), 3597 Size, (DstAlignCanChange ? 0 : Align), 0, 3598 NonScalarIntSafe, false, DAG, TLI)) 3599 return SDValue(); 3600 3601 if (DstAlignCanChange) { 3602 const Type *Ty = MemOps[0].getTypeForEVT(*DAG.getContext()); 3603 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty); 3604 if (NewAlign > Align) { 3605 // Give the stack frame object a larger alignment if needed. 3606 if (MFI->getObjectAlignment(FI->getIndex()) < NewAlign) 3607 MFI->setObjectAlignment(FI->getIndex(), NewAlign); 3608 Align = NewAlign; 3609 } 3610 } 3611 3612 SmallVector<SDValue, 8> OutChains; 3613 uint64_t DstOff = 0; 3614 unsigned NumMemOps = MemOps.size(); 3615 3616 // Find the largest store and generate the bit pattern for it. 3617 EVT LargestVT = MemOps[0]; 3618 for (unsigned i = 1; i < NumMemOps; i++) 3619 if (MemOps[i].bitsGT(LargestVT)) 3620 LargestVT = MemOps[i]; 3621 SDValue MemSetValue = getMemsetValue(Src, LargestVT, DAG, dl); 3622 3623 for (unsigned i = 0; i < NumMemOps; i++) { 3624 EVT VT = MemOps[i]; 3625 3626 // If this store is smaller than the largest store see whether we can get 3627 // the smaller value for free with a truncate. 3628 SDValue Value = MemSetValue; 3629 if (VT.bitsLT(LargestVT)) { 3630 if (!LargestVT.isVector() && !VT.isVector() && 3631 TLI.isTruncateFree(LargestVT, VT)) 3632 Value = DAG.getNode(ISD::TRUNCATE, dl, VT, MemSetValue); 3633 else 3634 Value = getMemsetValue(Src, VT, DAG, dl); 3635 } 3636 assert(Value.getValueType() == VT && "Value with wrong type."); 3637 SDValue Store = DAG.getStore(Chain, dl, Value, 3638 getMemBasePlusOffset(Dst, DstOff, DAG), 3639 DstPtrInfo.getWithOffset(DstOff), 3640 isVol, false, Align); 3641 OutChains.push_back(Store); 3642 DstOff += VT.getSizeInBits() / 8; 3643 } 3644 3645 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 3646 &OutChains[0], OutChains.size()); 3647} 3648 3649SDValue SelectionDAG::getMemcpy(SDValue Chain, DebugLoc dl, SDValue Dst, 3650 SDValue Src, SDValue Size, 3651 unsigned Align, bool isVol, bool AlwaysInline, 3652 MachinePointerInfo DstPtrInfo, 3653 MachinePointerInfo SrcPtrInfo) { 3654 3655 // Check to see if we should lower the memcpy to loads and stores first. 3656 // For cases within the target-specified limits, this is the best choice. 3657 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); 3658 if (ConstantSize) { 3659 // Memcpy with size zero? Just return the original chain. 3660 if (ConstantSize->isNullValue()) 3661 return Chain; 3662 3663 SDValue Result = getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src, 3664 ConstantSize->getZExtValue(),Align, 3665 isVol, false, DstPtrInfo, SrcPtrInfo); 3666 if (Result.getNode()) 3667 return Result; 3668 } 3669 3670 // Then check to see if we should lower the memcpy with target-specific 3671 // code. If the target chooses to do this, this is the next best. 3672 SDValue Result = 3673 TSI.EmitTargetCodeForMemcpy(*this, dl, Chain, Dst, Src, Size, Align, 3674 isVol, AlwaysInline, 3675 DstPtrInfo, SrcPtrInfo); 3676 if (Result.getNode()) 3677 return Result; 3678 3679 // If we really need inline code and the target declined to provide it, 3680 // use a (potentially long) sequence of loads and stores. 3681 if (AlwaysInline) { 3682 assert(ConstantSize && "AlwaysInline requires a constant size!"); 3683 return getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src, 3684 ConstantSize->getZExtValue(), Align, isVol, 3685 true, DstPtrInfo, SrcPtrInfo); 3686 } 3687 3688 // FIXME: If the memcpy is volatile (isVol), lowering it to a plain libc 3689 // memcpy is not guaranteed to be safe. libc memcpys aren't required to 3690 // respect volatile, so they may do things like read or write memory 3691 // beyond the given memory regions. But fixing this isn't easy, and most 3692 // people don't care. 3693 3694 // Emit a library call. 3695 TargetLowering::ArgListTy Args; 3696 TargetLowering::ArgListEntry Entry; 3697 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext()); 3698 Entry.Node = Dst; Args.push_back(Entry); 3699 Entry.Node = Src; Args.push_back(Entry); 3700 Entry.Node = Size; Args.push_back(Entry); 3701 // FIXME: pass in DebugLoc 3702 std::pair<SDValue,SDValue> CallResult = 3703 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()), 3704 false, false, false, false, 0, 3705 TLI.getLibcallCallingConv(RTLIB::MEMCPY), false, 3706 /*isReturnValueUsed=*/false, 3707 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMCPY), 3708 TLI.getPointerTy()), 3709 Args, *this, dl); 3710 return CallResult.second; 3711} 3712 3713SDValue SelectionDAG::getMemmove(SDValue Chain, DebugLoc dl, SDValue Dst, 3714 SDValue Src, SDValue Size, 3715 unsigned Align, bool isVol, 3716 MachinePointerInfo DstPtrInfo, 3717 MachinePointerInfo SrcPtrInfo) { 3718 3719 // Check to see if we should lower the memmove to loads and stores first. 3720 // For cases within the target-specified limits, this is the best choice. 3721 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); 3722 if (ConstantSize) { 3723 // Memmove with size zero? Just return the original chain. 3724 if (ConstantSize->isNullValue()) 3725 return Chain; 3726 3727 SDValue Result = 3728 getMemmoveLoadsAndStores(*this, dl, Chain, Dst, Src, 3729 ConstantSize->getZExtValue(), Align, isVol, 3730 false, DstPtrInfo, SrcPtrInfo); 3731 if (Result.getNode()) 3732 return Result; 3733 } 3734 3735 // Then check to see if we should lower the memmove with target-specific 3736 // code. If the target chooses to do this, this is the next best. 3737 SDValue Result = 3738 TSI.EmitTargetCodeForMemmove(*this, dl, Chain, Dst, Src, Size, Align, isVol, 3739 DstPtrInfo, SrcPtrInfo); 3740 if (Result.getNode()) 3741 return Result; 3742 3743 // FIXME: If the memmove is volatile, lowering it to plain libc memmove may 3744 // not be safe. See memcpy above for more details. 3745 3746 // Emit a library call. 3747 TargetLowering::ArgListTy Args; 3748 TargetLowering::ArgListEntry Entry; 3749 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext()); 3750 Entry.Node = Dst; Args.push_back(Entry); 3751 Entry.Node = Src; Args.push_back(Entry); 3752 Entry.Node = Size; Args.push_back(Entry); 3753 // FIXME: pass in DebugLoc 3754 std::pair<SDValue,SDValue> CallResult = 3755 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()), 3756 false, false, false, false, 0, 3757 TLI.getLibcallCallingConv(RTLIB::MEMMOVE), false, 3758 /*isReturnValueUsed=*/false, 3759 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMMOVE), 3760 TLI.getPointerTy()), 3761 Args, *this, dl); 3762 return CallResult.second; 3763} 3764 3765SDValue SelectionDAG::getMemset(SDValue Chain, DebugLoc dl, SDValue Dst, 3766 SDValue Src, SDValue Size, 3767 unsigned Align, bool isVol, 3768 MachinePointerInfo DstPtrInfo) { 3769 3770 // Check to see if we should lower the memset to stores first. 3771 // For cases within the target-specified limits, this is the best choice. 3772 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); 3773 if (ConstantSize) { 3774 // Memset with size zero? Just return the original chain. 3775 if (ConstantSize->isNullValue()) 3776 return Chain; 3777 3778 SDValue Result = 3779 getMemsetStores(*this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(), 3780 Align, isVol, DstPtrInfo); 3781 3782 if (Result.getNode()) 3783 return Result; 3784 } 3785 3786 // Then check to see if we should lower the memset with target-specific 3787 // code. If the target chooses to do this, this is the next best. 3788 SDValue Result = 3789 TSI.EmitTargetCodeForMemset(*this, dl, Chain, Dst, Src, Size, Align, isVol, 3790 DstPtrInfo); 3791 if (Result.getNode()) 3792 return Result; 3793 3794 // Emit a library call. 3795 const Type *IntPtrTy = TLI.getTargetData()->getIntPtrType(*getContext()); 3796 TargetLowering::ArgListTy Args; 3797 TargetLowering::ArgListEntry Entry; 3798 Entry.Node = Dst; Entry.Ty = IntPtrTy; 3799 Args.push_back(Entry); 3800 // Extend or truncate the argument to be an i32 value for the call. 3801 if (Src.getValueType().bitsGT(MVT::i32)) 3802 Src = getNode(ISD::TRUNCATE, dl, MVT::i32, Src); 3803 else 3804 Src = getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Src); 3805 Entry.Node = Src; 3806 Entry.Ty = Type::getInt32Ty(*getContext()); 3807 Entry.isSExt = true; 3808 Args.push_back(Entry); 3809 Entry.Node = Size; 3810 Entry.Ty = IntPtrTy; 3811 Entry.isSExt = false; 3812 Args.push_back(Entry); 3813 // FIXME: pass in DebugLoc 3814 std::pair<SDValue,SDValue> CallResult = 3815 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()), 3816 false, false, false, false, 0, 3817 TLI.getLibcallCallingConv(RTLIB::MEMSET), false, 3818 /*isReturnValueUsed=*/false, 3819 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMSET), 3820 TLI.getPointerTy()), 3821 Args, *this, dl); 3822 return CallResult.second; 3823} 3824 3825SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT, 3826 SDValue Chain, SDValue Ptr, SDValue Cmp, 3827 SDValue Swp, MachinePointerInfo PtrInfo, 3828 unsigned Alignment) { 3829 if (Alignment == 0) // Ensure that codegen never sees alignment 0 3830 Alignment = getEVTAlignment(MemVT); 3831 3832 MachineFunction &MF = getMachineFunction(); 3833 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore; 3834 3835 // For now, atomics are considered to be volatile always. 3836 Flags |= MachineMemOperand::MOVolatile; 3837 3838 MachineMemOperand *MMO = 3839 MF.getMachineMemOperand(PtrInfo, Flags, MemVT.getStoreSize(), Alignment); 3840 3841 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Cmp, Swp, MMO); 3842} 3843 3844SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT, 3845 SDValue Chain, 3846 SDValue Ptr, SDValue Cmp, 3847 SDValue Swp, MachineMemOperand *MMO) { 3848 assert(Opcode == ISD::ATOMIC_CMP_SWAP && "Invalid Atomic Op"); 3849 assert(Cmp.getValueType() == Swp.getValueType() && "Invalid Atomic Op Types"); 3850 3851 EVT VT = Cmp.getValueType(); 3852 3853 SDVTList VTs = getVTList(VT, MVT::Other); 3854 FoldingSetNodeID ID; 3855 ID.AddInteger(MemVT.getRawBits()); 3856 SDValue Ops[] = {Chain, Ptr, Cmp, Swp}; 3857 AddNodeIDNode(ID, Opcode, VTs, Ops, 4); 3858 void* IP = 0; 3859 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3860 cast<AtomicSDNode>(E)->refineAlignment(MMO); 3861 return SDValue(E, 0); 3862 } 3863 SDNode *N = new (NodeAllocator) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain, 3864 Ptr, Cmp, Swp, MMO); 3865 CSEMap.InsertNode(N, IP); 3866 AllNodes.push_back(N); 3867 return SDValue(N, 0); 3868} 3869 3870SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT, 3871 SDValue Chain, 3872 SDValue Ptr, SDValue Val, 3873 const Value* PtrVal, 3874 unsigned Alignment) { 3875 if (Alignment == 0) // Ensure that codegen never sees alignment 0 3876 Alignment = getEVTAlignment(MemVT); 3877 3878 MachineFunction &MF = getMachineFunction(); 3879 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore; 3880 3881 // For now, atomics are considered to be volatile always. 3882 Flags |= MachineMemOperand::MOVolatile; 3883 3884 MachineMemOperand *MMO = 3885 MF.getMachineMemOperand(MachinePointerInfo(PtrVal), Flags, 3886 MemVT.getStoreSize(), Alignment); 3887 3888 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Val, MMO); 3889} 3890 3891SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT, 3892 SDValue Chain, 3893 SDValue Ptr, SDValue Val, 3894 MachineMemOperand *MMO) { 3895 assert((Opcode == ISD::ATOMIC_LOAD_ADD || 3896 Opcode == ISD::ATOMIC_LOAD_SUB || 3897 Opcode == ISD::ATOMIC_LOAD_AND || 3898 Opcode == ISD::ATOMIC_LOAD_OR || 3899 Opcode == ISD::ATOMIC_LOAD_XOR || 3900 Opcode == ISD::ATOMIC_LOAD_NAND || 3901 Opcode == ISD::ATOMIC_LOAD_MIN || 3902 Opcode == ISD::ATOMIC_LOAD_MAX || 3903 Opcode == ISD::ATOMIC_LOAD_UMIN || 3904 Opcode == ISD::ATOMIC_LOAD_UMAX || 3905 Opcode == ISD::ATOMIC_SWAP) && 3906 "Invalid Atomic Op"); 3907 3908 EVT VT = Val.getValueType(); 3909 3910 SDVTList VTs = getVTList(VT, MVT::Other); 3911 FoldingSetNodeID ID; 3912 ID.AddInteger(MemVT.getRawBits()); 3913 SDValue Ops[] = {Chain, Ptr, Val}; 3914 AddNodeIDNode(ID, Opcode, VTs, Ops, 3); 3915 void* IP = 0; 3916 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3917 cast<AtomicSDNode>(E)->refineAlignment(MMO); 3918 return SDValue(E, 0); 3919 } 3920 SDNode *N = new (NodeAllocator) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain, 3921 Ptr, Val, MMO); 3922 CSEMap.InsertNode(N, IP); 3923 AllNodes.push_back(N); 3924 return SDValue(N, 0); 3925} 3926 3927/// getMergeValues - Create a MERGE_VALUES node from the given operands. 3928SDValue SelectionDAG::getMergeValues(const SDValue *Ops, unsigned NumOps, 3929 DebugLoc dl) { 3930 if (NumOps == 1) 3931 return Ops[0]; 3932 3933 SmallVector<EVT, 4> VTs; 3934 VTs.reserve(NumOps); 3935 for (unsigned i = 0; i < NumOps; ++i) 3936 VTs.push_back(Ops[i].getValueType()); 3937 return getNode(ISD::MERGE_VALUES, dl, getVTList(&VTs[0], NumOps), 3938 Ops, NumOps); 3939} 3940 3941SDValue 3942SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, 3943 const EVT *VTs, unsigned NumVTs, 3944 const SDValue *Ops, unsigned NumOps, 3945 EVT MemVT, MachinePointerInfo PtrInfo, 3946 unsigned Align, bool Vol, 3947 bool ReadMem, bool WriteMem) { 3948 return getMemIntrinsicNode(Opcode, dl, makeVTList(VTs, NumVTs), Ops, NumOps, 3949 MemVT, PtrInfo, Align, Vol, 3950 ReadMem, WriteMem); 3951} 3952 3953SDValue 3954SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList, 3955 const SDValue *Ops, unsigned NumOps, 3956 EVT MemVT, MachinePointerInfo PtrInfo, 3957 unsigned Align, bool Vol, 3958 bool ReadMem, bool WriteMem) { 3959 if (Align == 0) // Ensure that codegen never sees alignment 0 3960 Align = getEVTAlignment(MemVT); 3961 3962 MachineFunction &MF = getMachineFunction(); 3963 unsigned Flags = 0; 3964 if (WriteMem) 3965 Flags |= MachineMemOperand::MOStore; 3966 if (ReadMem) 3967 Flags |= MachineMemOperand::MOLoad; 3968 if (Vol) 3969 Flags |= MachineMemOperand::MOVolatile; 3970 MachineMemOperand *MMO = 3971 MF.getMachineMemOperand(PtrInfo, Flags, MemVT.getStoreSize(), Align); 3972 3973 return getMemIntrinsicNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO); 3974} 3975 3976SDValue 3977SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList, 3978 const SDValue *Ops, unsigned NumOps, 3979 EVT MemVT, MachineMemOperand *MMO) { 3980 assert((Opcode == ISD::INTRINSIC_VOID || 3981 Opcode == ISD::INTRINSIC_W_CHAIN || 3982 Opcode == ISD::PREFETCH || 3983 (Opcode <= INT_MAX && 3984 (int)Opcode >= ISD::FIRST_TARGET_MEMORY_OPCODE)) && 3985 "Opcode is not a memory-accessing opcode!"); 3986 3987 // Memoize the node unless it returns a flag. 3988 MemIntrinsicSDNode *N; 3989 if (VTList.VTs[VTList.NumVTs-1] != MVT::Glue) { 3990 FoldingSetNodeID ID; 3991 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps); 3992 void *IP = 0; 3993 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3994 cast<MemIntrinsicSDNode>(E)->refineAlignment(MMO); 3995 return SDValue(E, 0); 3996 } 3997 3998 N = new (NodeAllocator) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, 3999 MemVT, MMO); 4000 CSEMap.InsertNode(N, IP); 4001 } else { 4002 N = new (NodeAllocator) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, 4003 MemVT, MMO); 4004 } 4005 AllNodes.push_back(N); 4006 return SDValue(N, 0); 4007} 4008 4009/// InferPointerInfo - If the specified ptr/offset is a frame index, infer a 4010/// MachinePointerInfo record from it. This is particularly useful because the 4011/// code generator has many cases where it doesn't bother passing in a 4012/// MachinePointerInfo to getLoad or getStore when it has "FI+Cst". 4013static MachinePointerInfo InferPointerInfo(SDValue Ptr, int64_t Offset = 0) { 4014 // If this is FI+Offset, we can model it. 4015 if (const FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr)) 4016 return MachinePointerInfo::getFixedStack(FI->getIndex(), Offset); 4017 4018 // If this is (FI+Offset1)+Offset2, we can model it. 4019 if (Ptr.getOpcode() != ISD::ADD || 4020 !isa<ConstantSDNode>(Ptr.getOperand(1)) || 4021 !isa<FrameIndexSDNode>(Ptr.getOperand(0))) 4022 return MachinePointerInfo(); 4023 4024 int FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex(); 4025 return MachinePointerInfo::getFixedStack(FI, Offset+ 4026 cast<ConstantSDNode>(Ptr.getOperand(1))->getSExtValue()); 4027} 4028 4029/// InferPointerInfo - If the specified ptr/offset is a frame index, infer a 4030/// MachinePointerInfo record from it. This is particularly useful because the 4031/// code generator has many cases where it doesn't bother passing in a 4032/// MachinePointerInfo to getLoad or getStore when it has "FI+Cst". 4033static MachinePointerInfo InferPointerInfo(SDValue Ptr, SDValue OffsetOp) { 4034 // If the 'Offset' value isn't a constant, we can't handle this. 4035 if (ConstantSDNode *OffsetNode = dyn_cast<ConstantSDNode>(OffsetOp)) 4036 return InferPointerInfo(Ptr, OffsetNode->getSExtValue()); 4037 if (OffsetOp.getOpcode() == ISD::UNDEF) 4038 return InferPointerInfo(Ptr); 4039 return MachinePointerInfo(); 4040} 4041 4042 4043SDValue 4044SelectionDAG::getLoad(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, 4045 EVT VT, DebugLoc dl, SDValue Chain, 4046 SDValue Ptr, SDValue Offset, 4047 MachinePointerInfo PtrInfo, EVT MemVT, 4048 bool isVolatile, bool isNonTemporal, 4049 unsigned Alignment, const MDNode *TBAAInfo) { 4050 if (Alignment == 0) // Ensure that codegen never sees alignment 0 4051 Alignment = getEVTAlignment(VT); 4052 4053 unsigned Flags = MachineMemOperand::MOLoad; 4054 if (isVolatile) 4055 Flags |= MachineMemOperand::MOVolatile; 4056 if (isNonTemporal) 4057 Flags |= MachineMemOperand::MONonTemporal; 4058 4059 // If we don't have a PtrInfo, infer the trivial frame index case to simplify 4060 // clients. 4061 if (PtrInfo.V == 0) 4062 PtrInfo = InferPointerInfo(Ptr, Offset); 4063 4064 MachineFunction &MF = getMachineFunction(); 4065 MachineMemOperand *MMO = 4066 MF.getMachineMemOperand(PtrInfo, Flags, MemVT.getStoreSize(), Alignment, 4067 TBAAInfo); 4068 return getLoad(AM, ExtType, VT, dl, Chain, Ptr, Offset, MemVT, MMO); 4069} 4070 4071SDValue 4072SelectionDAG::getLoad(ISD::MemIndexedMode AM, ISD::LoadExtType ExtType, 4073 EVT VT, DebugLoc dl, SDValue Chain, 4074 SDValue Ptr, SDValue Offset, EVT MemVT, 4075 MachineMemOperand *MMO) { 4076 if (VT == MemVT) { 4077 ExtType = ISD::NON_EXTLOAD; 4078 } else if (ExtType == ISD::NON_EXTLOAD) { 4079 assert(VT == MemVT && "Non-extending load from different memory type!"); 4080 } else { 4081 // Extending load. 4082 assert(MemVT.getScalarType().bitsLT(VT.getScalarType()) && 4083 "Should only be an extending load, not truncating!"); 4084 assert(VT.isInteger() == MemVT.isInteger() && 4085 "Cannot convert from FP to Int or Int -> FP!"); 4086 assert(VT.isVector() == MemVT.isVector() && 4087 "Cannot use trunc store to convert to or from a vector!"); 4088 assert((!VT.isVector() || 4089 VT.getVectorNumElements() == MemVT.getVectorNumElements()) && 4090 "Cannot use trunc store to change the number of vector elements!"); 4091 } 4092 4093 bool Indexed = AM != ISD::UNINDEXED; 4094 assert((Indexed || Offset.getOpcode() == ISD::UNDEF) && 4095 "Unindexed load with an offset!"); 4096 4097 SDVTList VTs = Indexed ? 4098 getVTList(VT, Ptr.getValueType(), MVT::Other) : getVTList(VT, MVT::Other); 4099 SDValue Ops[] = { Chain, Ptr, Offset }; 4100 FoldingSetNodeID ID; 4101 AddNodeIDNode(ID, ISD::LOAD, VTs, Ops, 3); 4102 ID.AddInteger(MemVT.getRawBits()); 4103 ID.AddInteger(encodeMemSDNodeFlags(ExtType, AM, MMO->isVolatile(), 4104 MMO->isNonTemporal())); 4105 void *IP = 0; 4106 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 4107 cast<LoadSDNode>(E)->refineAlignment(MMO); 4108 return SDValue(E, 0); 4109 } 4110 SDNode *N = new (NodeAllocator) LoadSDNode(Ops, dl, VTs, AM, ExtType, 4111 MemVT, MMO); 4112 CSEMap.InsertNode(N, IP); 4113 AllNodes.push_back(N); 4114 return SDValue(N, 0); 4115} 4116 4117SDValue SelectionDAG::getLoad(EVT VT, DebugLoc dl, 4118 SDValue Chain, SDValue Ptr, 4119 MachinePointerInfo PtrInfo, 4120 bool isVolatile, bool isNonTemporal, 4121 unsigned Alignment, const MDNode *TBAAInfo) { 4122 SDValue Undef = getUNDEF(Ptr.getValueType()); 4123 return getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD, VT, dl, Chain, Ptr, Undef, 4124 PtrInfo, VT, isVolatile, isNonTemporal, Alignment, TBAAInfo); 4125} 4126 4127SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, DebugLoc dl, EVT VT, 4128 SDValue Chain, SDValue Ptr, 4129 MachinePointerInfo PtrInfo, EVT MemVT, 4130 bool isVolatile, bool isNonTemporal, 4131 unsigned Alignment, const MDNode *TBAAInfo) { 4132 SDValue Undef = getUNDEF(Ptr.getValueType()); 4133 return getLoad(ISD::UNINDEXED, ExtType, VT, dl, Chain, Ptr, Undef, 4134 PtrInfo, MemVT, isVolatile, isNonTemporal, Alignment, 4135 TBAAInfo); 4136} 4137 4138 4139SDValue 4140SelectionDAG::getIndexedLoad(SDValue OrigLoad, DebugLoc dl, SDValue Base, 4141 SDValue Offset, ISD::MemIndexedMode AM) { 4142 LoadSDNode *LD = cast<LoadSDNode>(OrigLoad); 4143 assert(LD->getOffset().getOpcode() == ISD::UNDEF && 4144 "Load is already a indexed load!"); 4145 return getLoad(AM, LD->getExtensionType(), OrigLoad.getValueType(), dl, 4146 LD->getChain(), Base, Offset, LD->getPointerInfo(), 4147 LD->getMemoryVT(), 4148 LD->isVolatile(), LD->isNonTemporal(), LD->getAlignment()); 4149} 4150 4151SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val, 4152 SDValue Ptr, MachinePointerInfo PtrInfo, 4153 bool isVolatile, bool isNonTemporal, 4154 unsigned Alignment, const MDNode *TBAAInfo) { 4155 if (Alignment == 0) // Ensure that codegen never sees alignment 0 4156 Alignment = getEVTAlignment(Val.getValueType()); 4157 4158 unsigned Flags = MachineMemOperand::MOStore; 4159 if (isVolatile) 4160 Flags |= MachineMemOperand::MOVolatile; 4161 if (isNonTemporal) 4162 Flags |= MachineMemOperand::MONonTemporal; 4163 4164 if (PtrInfo.V == 0) 4165 PtrInfo = InferPointerInfo(Ptr); 4166 4167 MachineFunction &MF = getMachineFunction(); 4168 MachineMemOperand *MMO = 4169 MF.getMachineMemOperand(PtrInfo, Flags, 4170 Val.getValueType().getStoreSize(), Alignment, 4171 TBAAInfo); 4172 4173 return getStore(Chain, dl, Val, Ptr, MMO); 4174} 4175 4176SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val, 4177 SDValue Ptr, MachineMemOperand *MMO) { 4178 EVT VT = Val.getValueType(); 4179 SDVTList VTs = getVTList(MVT::Other); 4180 SDValue Undef = getUNDEF(Ptr.getValueType()); 4181 SDValue Ops[] = { Chain, Val, Ptr, Undef }; 4182 FoldingSetNodeID ID; 4183 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4); 4184 ID.AddInteger(VT.getRawBits()); 4185 ID.AddInteger(encodeMemSDNodeFlags(false, ISD::UNINDEXED, MMO->isVolatile(), 4186 MMO->isNonTemporal())); 4187 void *IP = 0; 4188 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 4189 cast<StoreSDNode>(E)->refineAlignment(MMO); 4190 return SDValue(E, 0); 4191 } 4192 SDNode *N = new (NodeAllocator) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, 4193 false, VT, MMO); 4194 CSEMap.InsertNode(N, IP); 4195 AllNodes.push_back(N); 4196 return SDValue(N, 0); 4197} 4198 4199SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val, 4200 SDValue Ptr, MachinePointerInfo PtrInfo, 4201 EVT SVT,bool isVolatile, bool isNonTemporal, 4202 unsigned Alignment, 4203 const MDNode *TBAAInfo) { 4204 if (Alignment == 0) // Ensure that codegen never sees alignment 0 4205 Alignment = getEVTAlignment(SVT); 4206 4207 unsigned Flags = MachineMemOperand::MOStore; 4208 if (isVolatile) 4209 Flags |= MachineMemOperand::MOVolatile; 4210 if (isNonTemporal) 4211 Flags |= MachineMemOperand::MONonTemporal; 4212 4213 if (PtrInfo.V == 0) 4214 PtrInfo = InferPointerInfo(Ptr); 4215 4216 MachineFunction &MF = getMachineFunction(); 4217 MachineMemOperand *MMO = 4218 MF.getMachineMemOperand(PtrInfo, Flags, SVT.getStoreSize(), Alignment, 4219 TBAAInfo); 4220 4221 return getTruncStore(Chain, dl, Val, Ptr, SVT, MMO); 4222} 4223 4224SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val, 4225 SDValue Ptr, EVT SVT, 4226 MachineMemOperand *MMO) { 4227 EVT VT = Val.getValueType(); 4228 4229 if (VT == SVT) 4230 return getStore(Chain, dl, Val, Ptr, MMO); 4231 4232 assert(SVT.getScalarType().bitsLT(VT.getScalarType()) && 4233 "Should only be a truncating store, not extending!"); 4234 assert(VT.isInteger() == SVT.isInteger() && 4235 "Can't do FP-INT conversion!"); 4236 assert(VT.isVector() == SVT.isVector() && 4237 "Cannot use trunc store to convert to or from a vector!"); 4238 assert((!VT.isVector() || 4239 VT.getVectorNumElements() == SVT.getVectorNumElements()) && 4240 "Cannot use trunc store to change the number of vector elements!"); 4241 4242 SDVTList VTs = getVTList(MVT::Other); 4243 SDValue Undef = getUNDEF(Ptr.getValueType()); 4244 SDValue Ops[] = { Chain, Val, Ptr, Undef }; 4245 FoldingSetNodeID ID; 4246 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4); 4247 ID.AddInteger(SVT.getRawBits()); 4248 ID.AddInteger(encodeMemSDNodeFlags(true, ISD::UNINDEXED, MMO->isVolatile(), 4249 MMO->isNonTemporal())); 4250 void *IP = 0; 4251 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 4252 cast<StoreSDNode>(E)->refineAlignment(MMO); 4253 return SDValue(E, 0); 4254 } 4255 SDNode *N = new (NodeAllocator) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, 4256 true, SVT, MMO); 4257 CSEMap.InsertNode(N, IP); 4258 AllNodes.push_back(N); 4259 return SDValue(N, 0); 4260} 4261 4262SDValue 4263SelectionDAG::getIndexedStore(SDValue OrigStore, DebugLoc dl, SDValue Base, 4264 SDValue Offset, ISD::MemIndexedMode AM) { 4265 StoreSDNode *ST = cast<StoreSDNode>(OrigStore); 4266 assert(ST->getOffset().getOpcode() == ISD::UNDEF && 4267 "Store is already a indexed store!"); 4268 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other); 4269 SDValue Ops[] = { ST->getChain(), ST->getValue(), Base, Offset }; 4270 FoldingSetNodeID ID; 4271 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4); 4272 ID.AddInteger(ST->getMemoryVT().getRawBits()); 4273 ID.AddInteger(ST->getRawSubclassData()); 4274 void *IP = 0; 4275 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 4276 return SDValue(E, 0); 4277 4278 SDNode *N = new (NodeAllocator) StoreSDNode(Ops, dl, VTs, AM, 4279 ST->isTruncatingStore(), 4280 ST->getMemoryVT(), 4281 ST->getMemOperand()); 4282 CSEMap.InsertNode(N, IP); 4283 AllNodes.push_back(N); 4284 return SDValue(N, 0); 4285} 4286 4287SDValue SelectionDAG::getVAArg(EVT VT, DebugLoc dl, 4288 SDValue Chain, SDValue Ptr, 4289 SDValue SV, 4290 unsigned Align) { 4291 SDValue Ops[] = { Chain, Ptr, SV, getTargetConstant(Align, MVT::i32) }; 4292 return getNode(ISD::VAARG, dl, getVTList(VT, MVT::Other), Ops, 4); 4293} 4294 4295SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 4296 const SDUse *Ops, unsigned NumOps) { 4297 switch (NumOps) { 4298 case 0: return getNode(Opcode, DL, VT); 4299 case 1: return getNode(Opcode, DL, VT, Ops[0]); 4300 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]); 4301 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]); 4302 default: break; 4303 } 4304 4305 // Copy from an SDUse array into an SDValue array for use with 4306 // the regular getNode logic. 4307 SmallVector<SDValue, 8> NewOps(Ops, Ops + NumOps); 4308 return getNode(Opcode, DL, VT, &NewOps[0], NumOps); 4309} 4310 4311SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 4312 const SDValue *Ops, unsigned NumOps) { 4313 switch (NumOps) { 4314 case 0: return getNode(Opcode, DL, VT); 4315 case 1: return getNode(Opcode, DL, VT, Ops[0]); 4316 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]); 4317 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]); 4318 default: break; 4319 } 4320 4321 switch (Opcode) { 4322 default: break; 4323 case ISD::SELECT_CC: { 4324 assert(NumOps == 5 && "SELECT_CC takes 5 operands!"); 4325 assert(Ops[0].getValueType() == Ops[1].getValueType() && 4326 "LHS and RHS of condition must have same type!"); 4327 assert(Ops[2].getValueType() == Ops[3].getValueType() && 4328 "True and False arms of SelectCC must have same type!"); 4329 assert(Ops[2].getValueType() == VT && 4330 "select_cc node must be of same type as true and false value!"); 4331 break; 4332 } 4333 case ISD::BR_CC: { 4334 assert(NumOps == 5 && "BR_CC takes 5 operands!"); 4335 assert(Ops[2].getValueType() == Ops[3].getValueType() && 4336 "LHS/RHS of comparison should match types!"); 4337 break; 4338 } 4339 } 4340 4341 // Memoize nodes. 4342 SDNode *N; 4343 SDVTList VTs = getVTList(VT); 4344 4345 if (VT != MVT::Glue) { 4346 FoldingSetNodeID ID; 4347 AddNodeIDNode(ID, Opcode, VTs, Ops, NumOps); 4348 void *IP = 0; 4349 4350 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 4351 return SDValue(E, 0); 4352 4353 N = new (NodeAllocator) SDNode(Opcode, DL, VTs, Ops, NumOps); 4354 CSEMap.InsertNode(N, IP); 4355 } else { 4356 N = new (NodeAllocator) SDNode(Opcode, DL, VTs, Ops, NumOps); 4357 } 4358 4359 AllNodes.push_back(N); 4360#ifndef NDEBUG 4361 VerifySDNode(N); 4362#endif 4363 return SDValue(N, 0); 4364} 4365 4366SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, 4367 const std::vector<EVT> &ResultTys, 4368 const SDValue *Ops, unsigned NumOps) { 4369 return getNode(Opcode, DL, getVTList(&ResultTys[0], ResultTys.size()), 4370 Ops, NumOps); 4371} 4372 4373SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, 4374 const EVT *VTs, unsigned NumVTs, 4375 const SDValue *Ops, unsigned NumOps) { 4376 if (NumVTs == 1) 4377 return getNode(Opcode, DL, VTs[0], Ops, NumOps); 4378 return getNode(Opcode, DL, makeVTList(VTs, NumVTs), Ops, NumOps); 4379} 4380 4381SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4382 const SDValue *Ops, unsigned NumOps) { 4383 if (VTList.NumVTs == 1) 4384 return getNode(Opcode, DL, VTList.VTs[0], Ops, NumOps); 4385 4386#if 0 4387 switch (Opcode) { 4388 // FIXME: figure out how to safely handle things like 4389 // int foo(int x) { return 1 << (x & 255); } 4390 // int bar() { return foo(256); } 4391 case ISD::SRA_PARTS: 4392 case ISD::SRL_PARTS: 4393 case ISD::SHL_PARTS: 4394 if (N3.getOpcode() == ISD::SIGN_EXTEND_INREG && 4395 cast<VTSDNode>(N3.getOperand(1))->getVT() != MVT::i1) 4396 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0)); 4397 else if (N3.getOpcode() == ISD::AND) 4398 if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(N3.getOperand(1))) { 4399 // If the and is only masking out bits that cannot effect the shift, 4400 // eliminate the and. 4401 unsigned NumBits = VT.getScalarType().getSizeInBits()*2; 4402 if ((AndRHS->getValue() & (NumBits-1)) == NumBits-1) 4403 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0)); 4404 } 4405 break; 4406 } 4407#endif 4408 4409 // Memoize the node unless it returns a flag. 4410 SDNode *N; 4411 if (VTList.VTs[VTList.NumVTs-1] != MVT::Glue) { 4412 FoldingSetNodeID ID; 4413 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps); 4414 void *IP = 0; 4415 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 4416 return SDValue(E, 0); 4417 4418 if (NumOps == 1) { 4419 N = new (NodeAllocator) UnarySDNode(Opcode, DL, VTList, Ops[0]); 4420 } else if (NumOps == 2) { 4421 N = new (NodeAllocator) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]); 4422 } else if (NumOps == 3) { 4423 N = new (NodeAllocator) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], 4424 Ops[2]); 4425 } else { 4426 N = new (NodeAllocator) SDNode(Opcode, DL, VTList, Ops, NumOps); 4427 } 4428 CSEMap.InsertNode(N, IP); 4429 } else { 4430 if (NumOps == 1) { 4431 N = new (NodeAllocator) UnarySDNode(Opcode, DL, VTList, Ops[0]); 4432 } else if (NumOps == 2) { 4433 N = new (NodeAllocator) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]); 4434 } else if (NumOps == 3) { 4435 N = new (NodeAllocator) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], 4436 Ops[2]); 4437 } else { 4438 N = new (NodeAllocator) SDNode(Opcode, DL, VTList, Ops, NumOps); 4439 } 4440 } 4441 AllNodes.push_back(N); 4442#ifndef NDEBUG 4443 VerifySDNode(N); 4444#endif 4445 return SDValue(N, 0); 4446} 4447 4448SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList) { 4449 return getNode(Opcode, DL, VTList, 0, 0); 4450} 4451 4452SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4453 SDValue N1) { 4454 SDValue Ops[] = { N1 }; 4455 return getNode(Opcode, DL, VTList, Ops, 1); 4456} 4457 4458SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4459 SDValue N1, SDValue N2) { 4460 SDValue Ops[] = { N1, N2 }; 4461 return getNode(Opcode, DL, VTList, Ops, 2); 4462} 4463 4464SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4465 SDValue N1, SDValue N2, SDValue N3) { 4466 SDValue Ops[] = { N1, N2, N3 }; 4467 return getNode(Opcode, DL, VTList, Ops, 3); 4468} 4469 4470SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4471 SDValue N1, SDValue N2, SDValue N3, 4472 SDValue N4) { 4473 SDValue Ops[] = { N1, N2, N3, N4 }; 4474 return getNode(Opcode, DL, VTList, Ops, 4); 4475} 4476 4477SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4478 SDValue N1, SDValue N2, SDValue N3, 4479 SDValue N4, SDValue N5) { 4480 SDValue Ops[] = { N1, N2, N3, N4, N5 }; 4481 return getNode(Opcode, DL, VTList, Ops, 5); 4482} 4483 4484SDVTList SelectionDAG::getVTList(EVT VT) { 4485 return makeVTList(SDNode::getValueTypeList(VT), 1); 4486} 4487 4488SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2) { 4489 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(), 4490 E = VTList.rend(); I != E; ++I) 4491 if (I->NumVTs == 2 && I->VTs[0] == VT1 && I->VTs[1] == VT2) 4492 return *I; 4493 4494 EVT *Array = Allocator.Allocate<EVT>(2); 4495 Array[0] = VT1; 4496 Array[1] = VT2; 4497 SDVTList Result = makeVTList(Array, 2); 4498 VTList.push_back(Result); 4499 return Result; 4500} 4501 4502SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3) { 4503 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(), 4504 E = VTList.rend(); I != E; ++I) 4505 if (I->NumVTs == 3 && I->VTs[0] == VT1 && I->VTs[1] == VT2 && 4506 I->VTs[2] == VT3) 4507 return *I; 4508 4509 EVT *Array = Allocator.Allocate<EVT>(3); 4510 Array[0] = VT1; 4511 Array[1] = VT2; 4512 Array[2] = VT3; 4513 SDVTList Result = makeVTList(Array, 3); 4514 VTList.push_back(Result); 4515 return Result; 4516} 4517 4518SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3, EVT VT4) { 4519 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(), 4520 E = VTList.rend(); I != E; ++I) 4521 if (I->NumVTs == 4 && I->VTs[0] == VT1 && I->VTs[1] == VT2 && 4522 I->VTs[2] == VT3 && I->VTs[3] == VT4) 4523 return *I; 4524 4525 EVT *Array = Allocator.Allocate<EVT>(4); 4526 Array[0] = VT1; 4527 Array[1] = VT2; 4528 Array[2] = VT3; 4529 Array[3] = VT4; 4530 SDVTList Result = makeVTList(Array, 4); 4531 VTList.push_back(Result); 4532 return Result; 4533} 4534 4535SDVTList SelectionDAG::getVTList(const EVT *VTs, unsigned NumVTs) { 4536 switch (NumVTs) { 4537 case 0: llvm_unreachable("Cannot have nodes without results!"); 4538 case 1: return getVTList(VTs[0]); 4539 case 2: return getVTList(VTs[0], VTs[1]); 4540 case 3: return getVTList(VTs[0], VTs[1], VTs[2]); 4541 case 4: return getVTList(VTs[0], VTs[1], VTs[2], VTs[3]); 4542 default: break; 4543 } 4544 4545 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(), 4546 E = VTList.rend(); I != E; ++I) { 4547 if (I->NumVTs != NumVTs || VTs[0] != I->VTs[0] || VTs[1] != I->VTs[1]) 4548 continue; 4549 4550 bool NoMatch = false; 4551 for (unsigned i = 2; i != NumVTs; ++i) 4552 if (VTs[i] != I->VTs[i]) { 4553 NoMatch = true; 4554 break; 4555 } 4556 if (!NoMatch) 4557 return *I; 4558 } 4559 4560 EVT *Array = Allocator.Allocate<EVT>(NumVTs); 4561 std::copy(VTs, VTs+NumVTs, Array); 4562 SDVTList Result = makeVTList(Array, NumVTs); 4563 VTList.push_back(Result); 4564 return Result; 4565} 4566 4567 4568/// UpdateNodeOperands - *Mutate* the specified node in-place to have the 4569/// specified operands. If the resultant node already exists in the DAG, 4570/// this does not modify the specified node, instead it returns the node that 4571/// already exists. If the resultant node does not exist in the DAG, the 4572/// input node is returned. As a degenerate case, if you specify the same 4573/// input operands as the node already has, the input node is returned. 4574SDNode *SelectionDAG::UpdateNodeOperands(SDNode *N, SDValue Op) { 4575 assert(N->getNumOperands() == 1 && "Update with wrong number of operands"); 4576 4577 // Check to see if there is no change. 4578 if (Op == N->getOperand(0)) return N; 4579 4580 // See if the modified node already exists. 4581 void *InsertPos = 0; 4582 if (SDNode *Existing = FindModifiedNodeSlot(N, Op, InsertPos)) 4583 return Existing; 4584 4585 // Nope it doesn't. Remove the node from its current place in the maps. 4586 if (InsertPos) 4587 if (!RemoveNodeFromCSEMaps(N)) 4588 InsertPos = 0; 4589 4590 // Now we update the operands. 4591 N->OperandList[0].set(Op); 4592 4593 // If this gets put into a CSE map, add it. 4594 if (InsertPos) CSEMap.InsertNode(N, InsertPos); 4595 return N; 4596} 4597 4598SDNode *SelectionDAG::UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2) { 4599 assert(N->getNumOperands() == 2 && "Update with wrong number of operands"); 4600 4601 // Check to see if there is no change. 4602 if (Op1 == N->getOperand(0) && Op2 == N->getOperand(1)) 4603 return N; // No operands changed, just return the input node. 4604 4605 // See if the modified node already exists. 4606 void *InsertPos = 0; 4607 if (SDNode *Existing = FindModifiedNodeSlot(N, Op1, Op2, InsertPos)) 4608 return Existing; 4609 4610 // Nope it doesn't. Remove the node from its current place in the maps. 4611 if (InsertPos) 4612 if (!RemoveNodeFromCSEMaps(N)) 4613 InsertPos = 0; 4614 4615 // Now we update the operands. 4616 if (N->OperandList[0] != Op1) 4617 N->OperandList[0].set(Op1); 4618 if (N->OperandList[1] != Op2) 4619 N->OperandList[1].set(Op2); 4620 4621 // If this gets put into a CSE map, add it. 4622 if (InsertPos) CSEMap.InsertNode(N, InsertPos); 4623 return N; 4624} 4625 4626SDNode *SelectionDAG:: 4627UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2, SDValue Op3) { 4628 SDValue Ops[] = { Op1, Op2, Op3 }; 4629 return UpdateNodeOperands(N, Ops, 3); 4630} 4631 4632SDNode *SelectionDAG:: 4633UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2, 4634 SDValue Op3, SDValue Op4) { 4635 SDValue Ops[] = { Op1, Op2, Op3, Op4 }; 4636 return UpdateNodeOperands(N, Ops, 4); 4637} 4638 4639SDNode *SelectionDAG:: 4640UpdateNodeOperands(SDNode *N, SDValue Op1, SDValue Op2, 4641 SDValue Op3, SDValue Op4, SDValue Op5) { 4642 SDValue Ops[] = { Op1, Op2, Op3, Op4, Op5 }; 4643 return UpdateNodeOperands(N, Ops, 5); 4644} 4645 4646SDNode *SelectionDAG:: 4647UpdateNodeOperands(SDNode *N, const SDValue *Ops, unsigned NumOps) { 4648 assert(N->getNumOperands() == NumOps && 4649 "Update with wrong number of operands"); 4650 4651 // Check to see if there is no change. 4652 bool AnyChange = false; 4653 for (unsigned i = 0; i != NumOps; ++i) { 4654 if (Ops[i] != N->getOperand(i)) { 4655 AnyChange = true; 4656 break; 4657 } 4658 } 4659 4660 // No operands changed, just return the input node. 4661 if (!AnyChange) return N; 4662 4663 // See if the modified node already exists. 4664 void *InsertPos = 0; 4665 if (SDNode *Existing = FindModifiedNodeSlot(N, Ops, NumOps, InsertPos)) 4666 return Existing; 4667 4668 // Nope it doesn't. Remove the node from its current place in the maps. 4669 if (InsertPos) 4670 if (!RemoveNodeFromCSEMaps(N)) 4671 InsertPos = 0; 4672 4673 // Now we update the operands. 4674 for (unsigned i = 0; i != NumOps; ++i) 4675 if (N->OperandList[i] != Ops[i]) 4676 N->OperandList[i].set(Ops[i]); 4677 4678 // If this gets put into a CSE map, add it. 4679 if (InsertPos) CSEMap.InsertNode(N, InsertPos); 4680 return N; 4681} 4682 4683/// DropOperands - Release the operands and set this node to have 4684/// zero operands. 4685void SDNode::DropOperands() { 4686 // Unlike the code in MorphNodeTo that does this, we don't need to 4687 // watch for dead nodes here. 4688 for (op_iterator I = op_begin(), E = op_end(); I != E; ) { 4689 SDUse &Use = *I++; 4690 Use.set(SDValue()); 4691 } 4692} 4693 4694/// SelectNodeTo - These are wrappers around MorphNodeTo that accept a 4695/// machine opcode. 4696/// 4697SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4698 EVT VT) { 4699 SDVTList VTs = getVTList(VT); 4700 return SelectNodeTo(N, MachineOpc, VTs, 0, 0); 4701} 4702 4703SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4704 EVT VT, SDValue Op1) { 4705 SDVTList VTs = getVTList(VT); 4706 SDValue Ops[] = { Op1 }; 4707 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1); 4708} 4709 4710SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4711 EVT VT, SDValue Op1, 4712 SDValue Op2) { 4713 SDVTList VTs = getVTList(VT); 4714 SDValue Ops[] = { Op1, Op2 }; 4715 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2); 4716} 4717 4718SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4719 EVT VT, SDValue Op1, 4720 SDValue Op2, SDValue Op3) { 4721 SDVTList VTs = getVTList(VT); 4722 SDValue Ops[] = { Op1, Op2, Op3 }; 4723 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3); 4724} 4725 4726SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4727 EVT VT, const SDValue *Ops, 4728 unsigned NumOps) { 4729 SDVTList VTs = getVTList(VT); 4730 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps); 4731} 4732 4733SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4734 EVT VT1, EVT VT2, const SDValue *Ops, 4735 unsigned NumOps) { 4736 SDVTList VTs = getVTList(VT1, VT2); 4737 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps); 4738} 4739 4740SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4741 EVT VT1, EVT VT2) { 4742 SDVTList VTs = getVTList(VT1, VT2); 4743 return SelectNodeTo(N, MachineOpc, VTs, (SDValue *)0, 0); 4744} 4745 4746SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4747 EVT VT1, EVT VT2, EVT VT3, 4748 const SDValue *Ops, unsigned NumOps) { 4749 SDVTList VTs = getVTList(VT1, VT2, VT3); 4750 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps); 4751} 4752 4753SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4754 EVT VT1, EVT VT2, EVT VT3, EVT VT4, 4755 const SDValue *Ops, unsigned NumOps) { 4756 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4); 4757 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps); 4758} 4759 4760SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4761 EVT VT1, EVT VT2, 4762 SDValue Op1) { 4763 SDVTList VTs = getVTList(VT1, VT2); 4764 SDValue Ops[] = { Op1 }; 4765 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1); 4766} 4767 4768SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4769 EVT VT1, EVT VT2, 4770 SDValue Op1, SDValue Op2) { 4771 SDVTList VTs = getVTList(VT1, VT2); 4772 SDValue Ops[] = { Op1, Op2 }; 4773 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2); 4774} 4775 4776SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4777 EVT VT1, EVT VT2, 4778 SDValue Op1, SDValue Op2, 4779 SDValue Op3) { 4780 SDVTList VTs = getVTList(VT1, VT2); 4781 SDValue Ops[] = { Op1, Op2, Op3 }; 4782 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3); 4783} 4784 4785SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4786 EVT VT1, EVT VT2, EVT VT3, 4787 SDValue Op1, SDValue Op2, 4788 SDValue Op3) { 4789 SDVTList VTs = getVTList(VT1, VT2, VT3); 4790 SDValue Ops[] = { Op1, Op2, Op3 }; 4791 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3); 4792} 4793 4794SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4795 SDVTList VTs, const SDValue *Ops, 4796 unsigned NumOps) { 4797 N = MorphNodeTo(N, ~MachineOpc, VTs, Ops, NumOps); 4798 // Reset the NodeID to -1. 4799 N->setNodeId(-1); 4800 return N; 4801} 4802 4803/// MorphNodeTo - This *mutates* the specified node to have the specified 4804/// return type, opcode, and operands. 4805/// 4806/// Note that MorphNodeTo returns the resultant node. If there is already a 4807/// node of the specified opcode and operands, it returns that node instead of 4808/// the current one. Note that the DebugLoc need not be the same. 4809/// 4810/// Using MorphNodeTo is faster than creating a new node and swapping it in 4811/// with ReplaceAllUsesWith both because it often avoids allocating a new 4812/// node, and because it doesn't require CSE recalculation for any of 4813/// the node's users. 4814/// 4815SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4816 SDVTList VTs, const SDValue *Ops, 4817 unsigned NumOps) { 4818 // If an identical node already exists, use it. 4819 void *IP = 0; 4820 if (VTs.VTs[VTs.NumVTs-1] != MVT::Glue) { 4821 FoldingSetNodeID ID; 4822 AddNodeIDNode(ID, Opc, VTs, Ops, NumOps); 4823 if (SDNode *ON = CSEMap.FindNodeOrInsertPos(ID, IP)) 4824 return ON; 4825 } 4826 4827 if (!RemoveNodeFromCSEMaps(N)) 4828 IP = 0; 4829 4830 // Start the morphing. 4831 N->NodeType = Opc; 4832 N->ValueList = VTs.VTs; 4833 N->NumValues = VTs.NumVTs; 4834 4835 // Clear the operands list, updating used nodes to remove this from their 4836 // use list. Keep track of any operands that become dead as a result. 4837 SmallPtrSet<SDNode*, 16> DeadNodeSet; 4838 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) { 4839 SDUse &Use = *I++; 4840 SDNode *Used = Use.getNode(); 4841 Use.set(SDValue()); 4842 if (Used->use_empty()) 4843 DeadNodeSet.insert(Used); 4844 } 4845 4846 if (MachineSDNode *MN = dyn_cast<MachineSDNode>(N)) { 4847 // Initialize the memory references information. 4848 MN->setMemRefs(0, 0); 4849 // If NumOps is larger than the # of operands we can have in a 4850 // MachineSDNode, reallocate the operand list. 4851 if (NumOps > MN->NumOperands || !MN->OperandsNeedDelete) { 4852 if (MN->OperandsNeedDelete) 4853 delete[] MN->OperandList; 4854 if (NumOps > array_lengthof(MN->LocalOperands)) 4855 // We're creating a final node that will live unmorphed for the 4856 // remainder of the current SelectionDAG iteration, so we can allocate 4857 // the operands directly out of a pool with no recycling metadata. 4858 MN->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps), 4859 Ops, NumOps); 4860 else 4861 MN->InitOperands(MN->LocalOperands, Ops, NumOps); 4862 MN->OperandsNeedDelete = false; 4863 } else 4864 MN->InitOperands(MN->OperandList, Ops, NumOps); 4865 } else { 4866 // If NumOps is larger than the # of operands we currently have, reallocate 4867 // the operand list. 4868 if (NumOps > N->NumOperands) { 4869 if (N->OperandsNeedDelete) 4870 delete[] N->OperandList; 4871 N->InitOperands(new SDUse[NumOps], Ops, NumOps); 4872 N->OperandsNeedDelete = true; 4873 } else 4874 N->InitOperands(N->OperandList, Ops, NumOps); 4875 } 4876 4877 // Delete any nodes that are still dead after adding the uses for the 4878 // new operands. 4879 if (!DeadNodeSet.empty()) { 4880 SmallVector<SDNode *, 16> DeadNodes; 4881 for (SmallPtrSet<SDNode *, 16>::iterator I = DeadNodeSet.begin(), 4882 E = DeadNodeSet.end(); I != E; ++I) 4883 if ((*I)->use_empty()) 4884 DeadNodes.push_back(*I); 4885 RemoveDeadNodes(DeadNodes); 4886 } 4887 4888 if (IP) 4889 CSEMap.InsertNode(N, IP); // Memoize the new node. 4890 return N; 4891} 4892 4893 4894/// getMachineNode - These are used for target selectors to create a new node 4895/// with specified return type(s), MachineInstr opcode, and operands. 4896/// 4897/// Note that getMachineNode returns the resultant node. If there is already a 4898/// node of the specified opcode and operands, it returns that node instead of 4899/// the current one. 4900MachineSDNode * 4901SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT) { 4902 SDVTList VTs = getVTList(VT); 4903 return getMachineNode(Opcode, dl, VTs, 0, 0); 4904} 4905 4906MachineSDNode * 4907SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, SDValue Op1) { 4908 SDVTList VTs = getVTList(VT); 4909 SDValue Ops[] = { Op1 }; 4910 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4911} 4912 4913MachineSDNode * 4914SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, 4915 SDValue Op1, SDValue Op2) { 4916 SDVTList VTs = getVTList(VT); 4917 SDValue Ops[] = { Op1, Op2 }; 4918 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4919} 4920 4921MachineSDNode * 4922SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, 4923 SDValue Op1, SDValue Op2, SDValue Op3) { 4924 SDVTList VTs = getVTList(VT); 4925 SDValue Ops[] = { Op1, Op2, Op3 }; 4926 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4927} 4928 4929MachineSDNode * 4930SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, 4931 const SDValue *Ops, unsigned NumOps) { 4932 SDVTList VTs = getVTList(VT); 4933 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 4934} 4935 4936MachineSDNode * 4937SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1, EVT VT2) { 4938 SDVTList VTs = getVTList(VT1, VT2); 4939 return getMachineNode(Opcode, dl, VTs, 0, 0); 4940} 4941 4942MachineSDNode * 4943SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4944 EVT VT1, EVT VT2, SDValue Op1) { 4945 SDVTList VTs = getVTList(VT1, VT2); 4946 SDValue Ops[] = { Op1 }; 4947 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4948} 4949 4950MachineSDNode * 4951SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4952 EVT VT1, EVT VT2, SDValue Op1, SDValue Op2) { 4953 SDVTList VTs = getVTList(VT1, VT2); 4954 SDValue Ops[] = { Op1, Op2 }; 4955 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4956} 4957 4958MachineSDNode * 4959SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4960 EVT VT1, EVT VT2, SDValue Op1, 4961 SDValue Op2, SDValue Op3) { 4962 SDVTList VTs = getVTList(VT1, VT2); 4963 SDValue Ops[] = { Op1, Op2, Op3 }; 4964 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4965} 4966 4967MachineSDNode * 4968SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4969 EVT VT1, EVT VT2, 4970 const SDValue *Ops, unsigned NumOps) { 4971 SDVTList VTs = getVTList(VT1, VT2); 4972 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 4973} 4974 4975MachineSDNode * 4976SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4977 EVT VT1, EVT VT2, EVT VT3, 4978 SDValue Op1, SDValue Op2) { 4979 SDVTList VTs = getVTList(VT1, VT2, VT3); 4980 SDValue Ops[] = { Op1, Op2 }; 4981 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4982} 4983 4984MachineSDNode * 4985SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4986 EVT VT1, EVT VT2, EVT VT3, 4987 SDValue Op1, SDValue Op2, SDValue Op3) { 4988 SDVTList VTs = getVTList(VT1, VT2, VT3); 4989 SDValue Ops[] = { Op1, Op2, Op3 }; 4990 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4991} 4992 4993MachineSDNode * 4994SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4995 EVT VT1, EVT VT2, EVT VT3, 4996 const SDValue *Ops, unsigned NumOps) { 4997 SDVTList VTs = getVTList(VT1, VT2, VT3); 4998 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 4999} 5000 5001MachineSDNode * 5002SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1, 5003 EVT VT2, EVT VT3, EVT VT4, 5004 const SDValue *Ops, unsigned NumOps) { 5005 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4); 5006 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 5007} 5008 5009MachineSDNode * 5010SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 5011 const std::vector<EVT> &ResultTys, 5012 const SDValue *Ops, unsigned NumOps) { 5013 SDVTList VTs = getVTList(&ResultTys[0], ResultTys.size()); 5014 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 5015} 5016 5017MachineSDNode * 5018SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc DL, SDVTList VTs, 5019 const SDValue *Ops, unsigned NumOps) { 5020 bool DoCSE = VTs.VTs[VTs.NumVTs-1] != MVT::Glue; 5021 MachineSDNode *N; 5022 void *IP = 0; 5023 5024 if (DoCSE) { 5025 FoldingSetNodeID ID; 5026 AddNodeIDNode(ID, ~Opcode, VTs, Ops, NumOps); 5027 IP = 0; 5028 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 5029 return cast<MachineSDNode>(E); 5030 } 5031 5032 // Allocate a new MachineSDNode. 5033 N = new (NodeAllocator) MachineSDNode(~Opcode, DL, VTs); 5034 5035 // Initialize the operands list. 5036 if (NumOps > array_lengthof(N->LocalOperands)) 5037 // We're creating a final node that will live unmorphed for the 5038 // remainder of the current SelectionDAG iteration, so we can allocate 5039 // the operands directly out of a pool with no recycling metadata. 5040 N->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps), 5041 Ops, NumOps); 5042 else 5043 N->InitOperands(N->LocalOperands, Ops, NumOps); 5044 N->OperandsNeedDelete = false; 5045 5046 if (DoCSE) 5047 CSEMap.InsertNode(N, IP); 5048 5049 AllNodes.push_back(N); 5050#ifndef NDEBUG 5051 VerifyMachineNode(N); 5052#endif 5053 return N; 5054} 5055 5056/// getTargetExtractSubreg - A convenience function for creating 5057/// TargetOpcode::EXTRACT_SUBREG nodes. 5058SDValue 5059SelectionDAG::getTargetExtractSubreg(int SRIdx, DebugLoc DL, EVT VT, 5060 SDValue Operand) { 5061 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32); 5062 SDNode *Subreg = getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, 5063 VT, Operand, SRIdxVal); 5064 return SDValue(Subreg, 0); 5065} 5066 5067/// getTargetInsertSubreg - A convenience function for creating 5068/// TargetOpcode::INSERT_SUBREG nodes. 5069SDValue 5070SelectionDAG::getTargetInsertSubreg(int SRIdx, DebugLoc DL, EVT VT, 5071 SDValue Operand, SDValue Subreg) { 5072 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32); 5073 SDNode *Result = getMachineNode(TargetOpcode::INSERT_SUBREG, DL, 5074 VT, Operand, Subreg, SRIdxVal); 5075 return SDValue(Result, 0); 5076} 5077 5078/// getNodeIfExists - Get the specified node if it's already available, or 5079/// else return NULL. 5080SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList, 5081 const SDValue *Ops, unsigned NumOps) { 5082 if (VTList.VTs[VTList.NumVTs-1] != MVT::Glue) { 5083 FoldingSetNodeID ID; 5084 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps); 5085 void *IP = 0; 5086 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 5087 return E; 5088 } 5089 return NULL; 5090} 5091 5092/// getDbgValue - Creates a SDDbgValue node. 5093/// 5094SDDbgValue * 5095SelectionDAG::getDbgValue(MDNode *MDPtr, SDNode *N, unsigned R, uint64_t Off, 5096 DebugLoc DL, unsigned O) { 5097 return new (Allocator) SDDbgValue(MDPtr, N, R, Off, DL, O); 5098} 5099 5100SDDbgValue * 5101SelectionDAG::getDbgValue(MDNode *MDPtr, const Value *C, uint64_t Off, 5102 DebugLoc DL, unsigned O) { 5103 return new (Allocator) SDDbgValue(MDPtr, C, Off, DL, O); 5104} 5105 5106SDDbgValue * 5107SelectionDAG::getDbgValue(MDNode *MDPtr, unsigned FI, uint64_t Off, 5108 DebugLoc DL, unsigned O) { 5109 return new (Allocator) SDDbgValue(MDPtr, FI, Off, DL, O); 5110} 5111 5112namespace { 5113 5114/// RAUWUpdateListener - Helper for ReplaceAllUsesWith - When the node 5115/// pointed to by a use iterator is deleted, increment the use iterator 5116/// so that it doesn't dangle. 5117/// 5118/// This class also manages a "downlink" DAGUpdateListener, to forward 5119/// messages to ReplaceAllUsesWith's callers. 5120/// 5121class RAUWUpdateListener : public SelectionDAG::DAGUpdateListener { 5122 SelectionDAG::DAGUpdateListener *DownLink; 5123 SDNode::use_iterator &UI; 5124 SDNode::use_iterator &UE; 5125 5126 virtual void NodeDeleted(SDNode *N, SDNode *E) { 5127 // Increment the iterator as needed. 5128 while (UI != UE && N == *UI) 5129 ++UI; 5130 5131 // Then forward the message. 5132 if (DownLink) DownLink->NodeDeleted(N, E); 5133 } 5134 5135 virtual void NodeUpdated(SDNode *N) { 5136 // Just forward the message. 5137 if (DownLink) DownLink->NodeUpdated(N); 5138 } 5139 5140public: 5141 RAUWUpdateListener(SelectionDAG::DAGUpdateListener *dl, 5142 SDNode::use_iterator &ui, 5143 SDNode::use_iterator &ue) 5144 : DownLink(dl), UI(ui), UE(ue) {} 5145}; 5146 5147} 5148 5149/// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead. 5150/// This can cause recursive merging of nodes in the DAG. 5151/// 5152/// This version assumes From has a single result value. 5153/// 5154void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To, 5155 DAGUpdateListener *UpdateListener) { 5156 SDNode *From = FromN.getNode(); 5157 assert(From->getNumValues() == 1 && FromN.getResNo() == 0 && 5158 "Cannot replace with this method!"); 5159 assert(From != To.getNode() && "Cannot replace uses of with self"); 5160 5161 // Iterate over all the existing uses of From. New uses will be added 5162 // to the beginning of the use list, which we avoid visiting. 5163 // This specifically avoids visiting uses of From that arise while the 5164 // replacement is happening, because any such uses would be the result 5165 // of CSE: If an existing node looks like From after one of its operands 5166 // is replaced by To, we don't want to replace of all its users with To 5167 // too. See PR3018 for more info. 5168 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); 5169 RAUWUpdateListener Listener(UpdateListener, UI, UE); 5170 while (UI != UE) { 5171 SDNode *User = *UI; 5172 5173 // This node is about to morph, remove its old self from the CSE maps. 5174 RemoveNodeFromCSEMaps(User); 5175 5176 // A user can appear in a use list multiple times, and when this 5177 // happens the uses are usually next to each other in the list. 5178 // To help reduce the number of CSE recomputations, process all 5179 // the uses of this user that we can find this way. 5180 do { 5181 SDUse &Use = UI.getUse(); 5182 ++UI; 5183 Use.set(To); 5184 } while (UI != UE && *UI == User); 5185 5186 // Now that we have modified User, add it back to the CSE maps. If it 5187 // already exists there, recursively merge the results together. 5188 AddModifiedNodeToCSEMaps(User, &Listener); 5189 } 5190} 5191 5192/// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead. 5193/// This can cause recursive merging of nodes in the DAG. 5194/// 5195/// This version assumes that for each value of From, there is a 5196/// corresponding value in To in the same position with the same type. 5197/// 5198void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To, 5199 DAGUpdateListener *UpdateListener) { 5200#ifndef NDEBUG 5201 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i) 5202 assert((!From->hasAnyUseOfValue(i) || 5203 From->getValueType(i) == To->getValueType(i)) && 5204 "Cannot use this version of ReplaceAllUsesWith!"); 5205#endif 5206 5207 // Handle the trivial case. 5208 if (From == To) 5209 return; 5210 5211 // Iterate over just the existing users of From. See the comments in 5212 // the ReplaceAllUsesWith above. 5213 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); 5214 RAUWUpdateListener Listener(UpdateListener, UI, UE); 5215 while (UI != UE) { 5216 SDNode *User = *UI; 5217 5218 // This node is about to morph, remove its old self from the CSE maps. 5219 RemoveNodeFromCSEMaps(User); 5220 5221 // A user can appear in a use list multiple times, and when this 5222 // happens the uses are usually next to each other in the list. 5223 // To help reduce the number of CSE recomputations, process all 5224 // the uses of this user that we can find this way. 5225 do { 5226 SDUse &Use = UI.getUse(); 5227 ++UI; 5228 Use.setNode(To); 5229 } while (UI != UE && *UI == User); 5230 5231 // Now that we have modified User, add it back to the CSE maps. If it 5232 // already exists there, recursively merge the results together. 5233 AddModifiedNodeToCSEMaps(User, &Listener); 5234 } 5235} 5236 5237/// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead. 5238/// This can cause recursive merging of nodes in the DAG. 5239/// 5240/// This version can replace From with any result values. To must match the 5241/// number and types of values returned by From. 5242void SelectionDAG::ReplaceAllUsesWith(SDNode *From, 5243 const SDValue *To, 5244 DAGUpdateListener *UpdateListener) { 5245 if (From->getNumValues() == 1) // Handle the simple case efficiently. 5246 return ReplaceAllUsesWith(SDValue(From, 0), To[0], UpdateListener); 5247 5248 // Iterate over just the existing users of From. See the comments in 5249 // the ReplaceAllUsesWith above. 5250 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); 5251 RAUWUpdateListener Listener(UpdateListener, UI, UE); 5252 while (UI != UE) { 5253 SDNode *User = *UI; 5254 5255 // This node is about to morph, remove its old self from the CSE maps. 5256 RemoveNodeFromCSEMaps(User); 5257 5258 // A user can appear in a use list multiple times, and when this 5259 // happens the uses are usually next to each other in the list. 5260 // To help reduce the number of CSE recomputations, process all 5261 // the uses of this user that we can find this way. 5262 do { 5263 SDUse &Use = UI.getUse(); 5264 const SDValue &ToOp = To[Use.getResNo()]; 5265 ++UI; 5266 Use.set(ToOp); 5267 } while (UI != UE && *UI == User); 5268 5269 // Now that we have modified User, add it back to the CSE maps. If it 5270 // already exists there, recursively merge the results together. 5271 AddModifiedNodeToCSEMaps(User, &Listener); 5272 } 5273} 5274 5275/// ReplaceAllUsesOfValueWith - Replace any uses of From with To, leaving 5276/// uses of other values produced by From.getNode() alone. The Deleted 5277/// vector is handled the same way as for ReplaceAllUsesWith. 5278void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To, 5279 DAGUpdateListener *UpdateListener){ 5280 // Handle the really simple, really trivial case efficiently. 5281 if (From == To) return; 5282 5283 // Handle the simple, trivial, case efficiently. 5284 if (From.getNode()->getNumValues() == 1) { 5285 ReplaceAllUsesWith(From, To, UpdateListener); 5286 return; 5287 } 5288 5289 // Iterate over just the existing users of From. See the comments in 5290 // the ReplaceAllUsesWith above. 5291 SDNode::use_iterator UI = From.getNode()->use_begin(), 5292 UE = From.getNode()->use_end(); 5293 RAUWUpdateListener Listener(UpdateListener, UI, UE); 5294 while (UI != UE) { 5295 SDNode *User = *UI; 5296 bool UserRemovedFromCSEMaps = false; 5297 5298 // A user can appear in a use list multiple times, and when this 5299 // happens the uses are usually next to each other in the list. 5300 // To help reduce the number of CSE recomputations, process all 5301 // the uses of this user that we can find this way. 5302 do { 5303 SDUse &Use = UI.getUse(); 5304 5305 // Skip uses of different values from the same node. 5306 if (Use.getResNo() != From.getResNo()) { 5307 ++UI; 5308 continue; 5309 } 5310 5311 // If this node hasn't been modified yet, it's still in the CSE maps, 5312 // so remove its old self from the CSE maps. 5313 if (!UserRemovedFromCSEMaps) { 5314 RemoveNodeFromCSEMaps(User); 5315 UserRemovedFromCSEMaps = true; 5316 } 5317 5318 ++UI; 5319 Use.set(To); 5320 } while (UI != UE && *UI == User); 5321 5322 // We are iterating over all uses of the From node, so if a use 5323 // doesn't use the specific value, no changes are made. 5324 if (!UserRemovedFromCSEMaps) 5325 continue; 5326 5327 // Now that we have modified User, add it back to the CSE maps. If it 5328 // already exists there, recursively merge the results together. 5329 AddModifiedNodeToCSEMaps(User, &Listener); 5330 } 5331} 5332 5333namespace { 5334 /// UseMemo - This class is used by SelectionDAG::ReplaceAllUsesOfValuesWith 5335 /// to record information about a use. 5336 struct UseMemo { 5337 SDNode *User; 5338 unsigned Index; 5339 SDUse *Use; 5340 }; 5341 5342 /// operator< - Sort Memos by User. 5343 bool operator<(const UseMemo &L, const UseMemo &R) { 5344 return (intptr_t)L.User < (intptr_t)R.User; 5345 } 5346} 5347 5348/// ReplaceAllUsesOfValuesWith - Replace any uses of From with To, leaving 5349/// uses of other values produced by From.getNode() alone. The same value 5350/// may appear in both the From and To list. The Deleted vector is 5351/// handled the same way as for ReplaceAllUsesWith. 5352void SelectionDAG::ReplaceAllUsesOfValuesWith(const SDValue *From, 5353 const SDValue *To, 5354 unsigned Num, 5355 DAGUpdateListener *UpdateListener){ 5356 // Handle the simple, trivial case efficiently. 5357 if (Num == 1) 5358 return ReplaceAllUsesOfValueWith(*From, *To, UpdateListener); 5359 5360 // Read up all the uses and make records of them. This helps 5361 // processing new uses that are introduced during the 5362 // replacement process. 5363 SmallVector<UseMemo, 4> Uses; 5364 for (unsigned i = 0; i != Num; ++i) { 5365 unsigned FromResNo = From[i].getResNo(); 5366 SDNode *FromNode = From[i].getNode(); 5367 for (SDNode::use_iterator UI = FromNode->use_begin(), 5368 E = FromNode->use_end(); UI != E; ++UI) { 5369 SDUse &Use = UI.getUse(); 5370 if (Use.getResNo() == FromResNo) { 5371 UseMemo Memo = { *UI, i, &Use }; 5372 Uses.push_back(Memo); 5373 } 5374 } 5375 } 5376 5377 // Sort the uses, so that all the uses from a given User are together. 5378 std::sort(Uses.begin(), Uses.end()); 5379 5380 for (unsigned UseIndex = 0, UseIndexEnd = Uses.size(); 5381 UseIndex != UseIndexEnd; ) { 5382 // We know that this user uses some value of From. If it is the right 5383 // value, update it. 5384 SDNode *User = Uses[UseIndex].User; 5385 5386 // This node is about to morph, remove its old self from the CSE maps. 5387 RemoveNodeFromCSEMaps(User); 5388 5389 // The Uses array is sorted, so all the uses for a given User 5390 // are next to each other in the list. 5391 // To help reduce the number of CSE recomputations, process all 5392 // the uses of this user that we can find this way. 5393 do { 5394 unsigned i = Uses[UseIndex].Index; 5395 SDUse &Use = *Uses[UseIndex].Use; 5396 ++UseIndex; 5397 5398 Use.set(To[i]); 5399 } while (UseIndex != UseIndexEnd && Uses[UseIndex].User == User); 5400 5401 // Now that we have modified User, add it back to the CSE maps. If it 5402 // already exists there, recursively merge the results together. 5403 AddModifiedNodeToCSEMaps(User, UpdateListener); 5404 } 5405} 5406 5407/// AssignTopologicalOrder - Assign a unique node id for each node in the DAG 5408/// based on their topological order. It returns the maximum id and a vector 5409/// of the SDNodes* in assigned order by reference. 5410unsigned SelectionDAG::AssignTopologicalOrder() { 5411 5412 unsigned DAGSize = 0; 5413 5414 // SortedPos tracks the progress of the algorithm. Nodes before it are 5415 // sorted, nodes after it are unsorted. When the algorithm completes 5416 // it is at the end of the list. 5417 allnodes_iterator SortedPos = allnodes_begin(); 5418 5419 // Visit all the nodes. Move nodes with no operands to the front of 5420 // the list immediately. Annotate nodes that do have operands with their 5421 // operand count. Before we do this, the Node Id fields of the nodes 5422 // may contain arbitrary values. After, the Node Id fields for nodes 5423 // before SortedPos will contain the topological sort index, and the 5424 // Node Id fields for nodes At SortedPos and after will contain the 5425 // count of outstanding operands. 5426 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ) { 5427 SDNode *N = I++; 5428 checkForCycles(N); 5429 unsigned Degree = N->getNumOperands(); 5430 if (Degree == 0) { 5431 // A node with no uses, add it to the result array immediately. 5432 N->setNodeId(DAGSize++); 5433 allnodes_iterator Q = N; 5434 if (Q != SortedPos) 5435 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(Q)); 5436 assert(SortedPos != AllNodes.end() && "Overran node list"); 5437 ++SortedPos; 5438 } else { 5439 // Temporarily use the Node Id as scratch space for the degree count. 5440 N->setNodeId(Degree); 5441 } 5442 } 5443 5444 // Visit all the nodes. As we iterate, moves nodes into sorted order, 5445 // such that by the time the end is reached all nodes will be sorted. 5446 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ++I) { 5447 SDNode *N = I; 5448 checkForCycles(N); 5449 // N is in sorted position, so all its uses have one less operand 5450 // that needs to be sorted. 5451 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 5452 UI != UE; ++UI) { 5453 SDNode *P = *UI; 5454 unsigned Degree = P->getNodeId(); 5455 assert(Degree != 0 && "Invalid node degree"); 5456 --Degree; 5457 if (Degree == 0) { 5458 // All of P's operands are sorted, so P may sorted now. 5459 P->setNodeId(DAGSize++); 5460 if (P != SortedPos) 5461 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(P)); 5462 assert(SortedPos != AllNodes.end() && "Overran node list"); 5463 ++SortedPos; 5464 } else { 5465 // Update P's outstanding operand count. 5466 P->setNodeId(Degree); 5467 } 5468 } 5469 if (I == SortedPos) { 5470#ifndef NDEBUG 5471 SDNode *S = ++I; 5472 dbgs() << "Overran sorted position:\n"; 5473 S->dumprFull(); 5474#endif 5475 llvm_unreachable(0); 5476 } 5477 } 5478 5479 assert(SortedPos == AllNodes.end() && 5480 "Topological sort incomplete!"); 5481 assert(AllNodes.front().getOpcode() == ISD::EntryToken && 5482 "First node in topological sort is not the entry token!"); 5483 assert(AllNodes.front().getNodeId() == 0 && 5484 "First node in topological sort has non-zero id!"); 5485 assert(AllNodes.front().getNumOperands() == 0 && 5486 "First node in topological sort has operands!"); 5487 assert(AllNodes.back().getNodeId() == (int)DAGSize-1 && 5488 "Last node in topologic sort has unexpected id!"); 5489 assert(AllNodes.back().use_empty() && 5490 "Last node in topologic sort has users!"); 5491 assert(DAGSize == allnodes_size() && "Node count mismatch!"); 5492 return DAGSize; 5493} 5494 5495/// AssignOrdering - Assign an order to the SDNode. 5496void SelectionDAG::AssignOrdering(const SDNode *SD, unsigned Order) { 5497 assert(SD && "Trying to assign an order to a null node!"); 5498 Ordering->add(SD, Order); 5499} 5500 5501/// GetOrdering - Get the order for the SDNode. 5502unsigned SelectionDAG::GetOrdering(const SDNode *SD) const { 5503 assert(SD && "Trying to get the order of a null node!"); 5504 return Ordering->getOrder(SD); 5505} 5506 5507/// AddDbgValue - Add a dbg_value SDNode. If SD is non-null that means the 5508/// value is produced by SD. 5509void SelectionDAG::AddDbgValue(SDDbgValue *DB, SDNode *SD, bool isParameter) { 5510 DbgInfo->add(DB, SD, isParameter); 5511 if (SD) 5512 SD->setHasDebugValue(true); 5513} 5514 5515/// TransferDbgValues - Transfer SDDbgValues. 5516void SelectionDAG::TransferDbgValues(SDValue From, SDValue To) { 5517 if (From == To || !From.getNode()->getHasDebugValue()) 5518 return; 5519 SDNode *FromNode = From.getNode(); 5520 SDNode *ToNode = To.getNode(); 5521 SmallVector<SDDbgValue *, 2> &DVs = GetDbgValues(FromNode); 5522 SmallVector<SDDbgValue *, 2> ClonedDVs; 5523 for (SmallVector<SDDbgValue *, 2>::iterator I = DVs.begin(), E = DVs.end(); 5524 I != E; ++I) { 5525 SDDbgValue *Dbg = *I; 5526 if (Dbg->getKind() == SDDbgValue::SDNODE) { 5527 SDDbgValue *Clone = getDbgValue(Dbg->getMDPtr(), ToNode, To.getResNo(), 5528 Dbg->getOffset(), Dbg->getDebugLoc(), 5529 Dbg->getOrder()); 5530 ClonedDVs.push_back(Clone); 5531 } 5532 } 5533 for (SmallVector<SDDbgValue *, 2>::iterator I = ClonedDVs.begin(), 5534 E = ClonedDVs.end(); I != E; ++I) 5535 AddDbgValue(*I, ToNode, false); 5536} 5537 5538//===----------------------------------------------------------------------===// 5539// SDNode Class 5540//===----------------------------------------------------------------------===// 5541 5542HandleSDNode::~HandleSDNode() { 5543 DropOperands(); 5544} 5545 5546GlobalAddressSDNode::GlobalAddressSDNode(unsigned Opc, DebugLoc DL, 5547 const GlobalValue *GA, 5548 EVT VT, int64_t o, unsigned char TF) 5549 : SDNode(Opc, DL, getSDVTList(VT)), Offset(o), TargetFlags(TF) { 5550 TheGlobal = GA; 5551} 5552 5553MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, EVT memvt, 5554 MachineMemOperand *mmo) 5555 : SDNode(Opc, dl, VTs), MemoryVT(memvt), MMO(mmo) { 5556 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile(), 5557 MMO->isNonTemporal()); 5558 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!"); 5559 assert(isNonTemporal() == MMO->isNonTemporal() && 5560 "Non-temporal encoding error!"); 5561 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!"); 5562} 5563 5564MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, 5565 const SDValue *Ops, unsigned NumOps, EVT memvt, 5566 MachineMemOperand *mmo) 5567 : SDNode(Opc, dl, VTs, Ops, NumOps), 5568 MemoryVT(memvt), MMO(mmo) { 5569 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile(), 5570 MMO->isNonTemporal()); 5571 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!"); 5572 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!"); 5573} 5574 5575/// Profile - Gather unique data for the node. 5576/// 5577void SDNode::Profile(FoldingSetNodeID &ID) const { 5578 AddNodeIDNode(ID, this); 5579} 5580 5581namespace { 5582 struct EVTArray { 5583 std::vector<EVT> VTs; 5584 5585 EVTArray() { 5586 VTs.reserve(MVT::LAST_VALUETYPE); 5587 for (unsigned i = 0; i < MVT::LAST_VALUETYPE; ++i) 5588 VTs.push_back(MVT((MVT::SimpleValueType)i)); 5589 } 5590 }; 5591} 5592 5593static ManagedStatic<std::set<EVT, EVT::compareRawBits> > EVTs; 5594static ManagedStatic<EVTArray> SimpleVTArray; 5595static ManagedStatic<sys::SmartMutex<true> > VTMutex; 5596 5597/// getValueTypeList - Return a pointer to the specified value type. 5598/// 5599const EVT *SDNode::getValueTypeList(EVT VT) { 5600 if (VT.isExtended()) { 5601 sys::SmartScopedLock<true> Lock(*VTMutex); 5602 return &(*EVTs->insert(VT).first); 5603 } else { 5604 assert(VT.getSimpleVT() < MVT::LAST_VALUETYPE && 5605 "Value type out of range!"); 5606 return &SimpleVTArray->VTs[VT.getSimpleVT().SimpleTy]; 5607 } 5608} 5609 5610/// hasNUsesOfValue - Return true if there are exactly NUSES uses of the 5611/// indicated value. This method ignores uses of other values defined by this 5612/// operation. 5613bool SDNode::hasNUsesOfValue(unsigned NUses, unsigned Value) const { 5614 assert(Value < getNumValues() && "Bad value!"); 5615 5616 // TODO: Only iterate over uses of a given value of the node 5617 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) { 5618 if (UI.getUse().getResNo() == Value) { 5619 if (NUses == 0) 5620 return false; 5621 --NUses; 5622 } 5623 } 5624 5625 // Found exactly the right number of uses? 5626 return NUses == 0; 5627} 5628 5629 5630/// hasAnyUseOfValue - Return true if there are any use of the indicated 5631/// value. This method ignores uses of other values defined by this operation. 5632bool SDNode::hasAnyUseOfValue(unsigned Value) const { 5633 assert(Value < getNumValues() && "Bad value!"); 5634 5635 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) 5636 if (UI.getUse().getResNo() == Value) 5637 return true; 5638 5639 return false; 5640} 5641 5642 5643/// isOnlyUserOf - Return true if this node is the only use of N. 5644/// 5645bool SDNode::isOnlyUserOf(SDNode *N) const { 5646 bool Seen = false; 5647 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) { 5648 SDNode *User = *I; 5649 if (User == this) 5650 Seen = true; 5651 else 5652 return false; 5653 } 5654 5655 return Seen; 5656} 5657 5658/// isOperand - Return true if this node is an operand of N. 5659/// 5660bool SDValue::isOperandOf(SDNode *N) const { 5661 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 5662 if (*this == N->getOperand(i)) 5663 return true; 5664 return false; 5665} 5666 5667bool SDNode::isOperandOf(SDNode *N) const { 5668 for (unsigned i = 0, e = N->NumOperands; i != e; ++i) 5669 if (this == N->OperandList[i].getNode()) 5670 return true; 5671 return false; 5672} 5673 5674/// reachesChainWithoutSideEffects - Return true if this operand (which must 5675/// be a chain) reaches the specified operand without crossing any 5676/// side-effecting instructions on any chain path. In practice, this looks 5677/// through token factors and non-volatile loads. In order to remain efficient, 5678/// this only looks a couple of nodes in, it does not do an exhaustive search. 5679bool SDValue::reachesChainWithoutSideEffects(SDValue Dest, 5680 unsigned Depth) const { 5681 if (*this == Dest) return true; 5682 5683 // Don't search too deeply, we just want to be able to see through 5684 // TokenFactor's etc. 5685 if (Depth == 0) return false; 5686 5687 // If this is a token factor, all inputs to the TF happen in parallel. If any 5688 // of the operands of the TF does not reach dest, then we cannot do the xform. 5689 if (getOpcode() == ISD::TokenFactor) { 5690 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) 5691 if (!getOperand(i).reachesChainWithoutSideEffects(Dest, Depth-1)) 5692 return false; 5693 return true; 5694 } 5695 5696 // Loads don't have side effects, look through them. 5697 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(*this)) { 5698 if (!Ld->isVolatile()) 5699 return Ld->getChain().reachesChainWithoutSideEffects(Dest, Depth-1); 5700 } 5701 return false; 5702} 5703 5704/// isPredecessorOf - Return true if this node is a predecessor of N. This node 5705/// is either an operand of N or it can be reached by traversing up the operands. 5706/// NOTE: this is an expensive method. Use it carefully. 5707bool SDNode::isPredecessorOf(SDNode *N) const { 5708 SmallPtrSet<SDNode *, 32> Visited; 5709 SmallVector<SDNode *, 16> Worklist; 5710 Worklist.push_back(N); 5711 5712 do { 5713 N = Worklist.pop_back_val(); 5714 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 5715 SDNode *Op = N->getOperand(i).getNode(); 5716 if (Op == this) 5717 return true; 5718 if (Visited.insert(Op)) 5719 Worklist.push_back(Op); 5720 } 5721 } while (!Worklist.empty()); 5722 5723 return false; 5724} 5725 5726uint64_t SDNode::getConstantOperandVal(unsigned Num) const { 5727 assert(Num < NumOperands && "Invalid child # of SDNode!"); 5728 return cast<ConstantSDNode>(OperandList[Num])->getZExtValue(); 5729} 5730 5731std::string SDNode::getOperationName(const SelectionDAG *G) const { 5732 switch (getOpcode()) { 5733 default: 5734 if (getOpcode() < ISD::BUILTIN_OP_END) 5735 return "<<Unknown DAG Node>>"; 5736 if (isMachineOpcode()) { 5737 if (G) 5738 if (const TargetInstrInfo *TII = G->getTarget().getInstrInfo()) 5739 if (getMachineOpcode() < TII->getNumOpcodes()) 5740 return TII->get(getMachineOpcode()).getName(); 5741 return "<<Unknown Machine Node #" + utostr(getOpcode()) + ">>"; 5742 } 5743 if (G) { 5744 const TargetLowering &TLI = G->getTargetLoweringInfo(); 5745 const char *Name = TLI.getTargetNodeName(getOpcode()); 5746 if (Name) return Name; 5747 return "<<Unknown Target Node #" + utostr(getOpcode()) + ">>"; 5748 } 5749 return "<<Unknown Node #" + utostr(getOpcode()) + ">>"; 5750 5751#ifndef NDEBUG 5752 case ISD::DELETED_NODE: 5753 return "<<Deleted Node!>>"; 5754#endif 5755 case ISD::PREFETCH: return "Prefetch"; 5756 case ISD::MEMBARRIER: return "MemBarrier"; 5757 case ISD::ATOMIC_CMP_SWAP: return "AtomicCmpSwap"; 5758 case ISD::ATOMIC_SWAP: return "AtomicSwap"; 5759 case ISD::ATOMIC_LOAD_ADD: return "AtomicLoadAdd"; 5760 case ISD::ATOMIC_LOAD_SUB: return "AtomicLoadSub"; 5761 case ISD::ATOMIC_LOAD_AND: return "AtomicLoadAnd"; 5762 case ISD::ATOMIC_LOAD_OR: return "AtomicLoadOr"; 5763 case ISD::ATOMIC_LOAD_XOR: return "AtomicLoadXor"; 5764 case ISD::ATOMIC_LOAD_NAND: return "AtomicLoadNand"; 5765 case ISD::ATOMIC_LOAD_MIN: return "AtomicLoadMin"; 5766 case ISD::ATOMIC_LOAD_MAX: return "AtomicLoadMax"; 5767 case ISD::ATOMIC_LOAD_UMIN: return "AtomicLoadUMin"; 5768 case ISD::ATOMIC_LOAD_UMAX: return "AtomicLoadUMax"; 5769 case ISD::PCMARKER: return "PCMarker"; 5770 case ISD::READCYCLECOUNTER: return "ReadCycleCounter"; 5771 case ISD::SRCVALUE: return "SrcValue"; 5772 case ISD::MDNODE_SDNODE: return "MDNode"; 5773 case ISD::EntryToken: return "EntryToken"; 5774 case ISD::TokenFactor: return "TokenFactor"; 5775 case ISD::AssertSext: return "AssertSext"; 5776 case ISD::AssertZext: return "AssertZext"; 5777 5778 case ISD::BasicBlock: return "BasicBlock"; 5779 case ISD::VALUETYPE: return "ValueType"; 5780 case ISD::Register: return "Register"; 5781 5782 case ISD::Constant: return "Constant"; 5783 case ISD::ConstantFP: return "ConstantFP"; 5784 case ISD::GlobalAddress: return "GlobalAddress"; 5785 case ISD::GlobalTLSAddress: return "GlobalTLSAddress"; 5786 case ISD::FrameIndex: return "FrameIndex"; 5787 case ISD::JumpTable: return "JumpTable"; 5788 case ISD::GLOBAL_OFFSET_TABLE: return "GLOBAL_OFFSET_TABLE"; 5789 case ISD::RETURNADDR: return "RETURNADDR"; 5790 case ISD::FRAMEADDR: return "FRAMEADDR"; 5791 case ISD::FRAME_TO_ARGS_OFFSET: return "FRAME_TO_ARGS_OFFSET"; 5792 case ISD::EXCEPTIONADDR: return "EXCEPTIONADDR"; 5793 case ISD::LSDAADDR: return "LSDAADDR"; 5794 case ISD::EHSELECTION: return "EHSELECTION"; 5795 case ISD::EH_RETURN: return "EH_RETURN"; 5796 case ISD::EH_SJLJ_SETJMP: return "EH_SJLJ_SETJMP"; 5797 case ISD::EH_SJLJ_LONGJMP: return "EH_SJLJ_LONGJMP"; 5798 case ISD::EH_SJLJ_DISPATCHSETUP: return "EH_SJLJ_DISPATCHSETUP"; 5799 case ISD::ConstantPool: return "ConstantPool"; 5800 case ISD::ExternalSymbol: return "ExternalSymbol"; 5801 case ISD::BlockAddress: return "BlockAddress"; 5802 case ISD::INTRINSIC_WO_CHAIN: 5803 case ISD::INTRINSIC_VOID: 5804 case ISD::INTRINSIC_W_CHAIN: { 5805 unsigned OpNo = getOpcode() == ISD::INTRINSIC_WO_CHAIN ? 0 : 1; 5806 unsigned IID = cast<ConstantSDNode>(getOperand(OpNo))->getZExtValue(); 5807 if (IID < Intrinsic::num_intrinsics) 5808 return Intrinsic::getName((Intrinsic::ID)IID); 5809 else if (const TargetIntrinsicInfo *TII = G->getTarget().getIntrinsicInfo()) 5810 return TII->getName(IID); 5811 llvm_unreachable("Invalid intrinsic ID"); 5812 } 5813 5814 case ISD::BUILD_VECTOR: return "BUILD_VECTOR"; 5815 case ISD::TargetConstant: return "TargetConstant"; 5816 case ISD::TargetConstantFP:return "TargetConstantFP"; 5817 case ISD::TargetGlobalAddress: return "TargetGlobalAddress"; 5818 case ISD::TargetGlobalTLSAddress: return "TargetGlobalTLSAddress"; 5819 case ISD::TargetFrameIndex: return "TargetFrameIndex"; 5820 case ISD::TargetJumpTable: return "TargetJumpTable"; 5821 case ISD::TargetConstantPool: return "TargetConstantPool"; 5822 case ISD::TargetExternalSymbol: return "TargetExternalSymbol"; 5823 case ISD::TargetBlockAddress: return "TargetBlockAddress"; 5824 5825 case ISD::CopyToReg: return "CopyToReg"; 5826 case ISD::CopyFromReg: return "CopyFromReg"; 5827 case ISD::UNDEF: return "undef"; 5828 case ISD::MERGE_VALUES: return "merge_values"; 5829 case ISD::INLINEASM: return "inlineasm"; 5830 case ISD::EH_LABEL: return "eh_label"; 5831 case ISD::HANDLENODE: return "handlenode"; 5832 5833 // Unary operators 5834 case ISD::FABS: return "fabs"; 5835 case ISD::FNEG: return "fneg"; 5836 case ISD::FSQRT: return "fsqrt"; 5837 case ISD::FSIN: return "fsin"; 5838 case ISD::FCOS: return "fcos"; 5839 case ISD::FTRUNC: return "ftrunc"; 5840 case ISD::FFLOOR: return "ffloor"; 5841 case ISD::FCEIL: return "fceil"; 5842 case ISD::FRINT: return "frint"; 5843 case ISD::FNEARBYINT: return "fnearbyint"; 5844 case ISD::FEXP: return "fexp"; 5845 case ISD::FEXP2: return "fexp2"; 5846 case ISD::FLOG: return "flog"; 5847 case ISD::FLOG2: return "flog2"; 5848 case ISD::FLOG10: return "flog10"; 5849 5850 // Binary operators 5851 case ISD::ADD: return "add"; 5852 case ISD::SUB: return "sub"; 5853 case ISD::MUL: return "mul"; 5854 case ISD::MULHU: return "mulhu"; 5855 case ISD::MULHS: return "mulhs"; 5856 case ISD::SDIV: return "sdiv"; 5857 case ISD::UDIV: return "udiv"; 5858 case ISD::SREM: return "srem"; 5859 case ISD::UREM: return "urem"; 5860 case ISD::SMUL_LOHI: return "smul_lohi"; 5861 case ISD::UMUL_LOHI: return "umul_lohi"; 5862 case ISD::SDIVREM: return "sdivrem"; 5863 case ISD::UDIVREM: return "udivrem"; 5864 case ISD::AND: return "and"; 5865 case ISD::OR: return "or"; 5866 case ISD::XOR: return "xor"; 5867 case ISD::SHL: return "shl"; 5868 case ISD::SRA: return "sra"; 5869 case ISD::SRL: return "srl"; 5870 case ISD::ROTL: return "rotl"; 5871 case ISD::ROTR: return "rotr"; 5872 case ISD::FADD: return "fadd"; 5873 case ISD::FSUB: return "fsub"; 5874 case ISD::FMUL: return "fmul"; 5875 case ISD::FDIV: return "fdiv"; 5876 case ISD::FREM: return "frem"; 5877 case ISD::FCOPYSIGN: return "fcopysign"; 5878 case ISD::FGETSIGN: return "fgetsign"; 5879 case ISD::FPOW: return "fpow"; 5880 5881 case ISD::FPOWI: return "fpowi"; 5882 case ISD::SETCC: return "setcc"; 5883 case ISD::VSETCC: return "vsetcc"; 5884 case ISD::SELECT: return "select"; 5885 case ISD::SELECT_CC: return "select_cc"; 5886 case ISD::INSERT_VECTOR_ELT: return "insert_vector_elt"; 5887 case ISD::EXTRACT_VECTOR_ELT: return "extract_vector_elt"; 5888 case ISD::CONCAT_VECTORS: return "concat_vectors"; 5889 case ISD::INSERT_SUBVECTOR: return "insert_subvector"; 5890 case ISD::EXTRACT_SUBVECTOR: return "extract_subvector"; 5891 case ISD::SCALAR_TO_VECTOR: return "scalar_to_vector"; 5892 case ISD::VECTOR_SHUFFLE: return "vector_shuffle"; 5893 case ISD::CARRY_FALSE: return "carry_false"; 5894 case ISD::ADDC: return "addc"; 5895 case ISD::ADDE: return "adde"; 5896 case ISD::SADDO: return "saddo"; 5897 case ISD::UADDO: return "uaddo"; 5898 case ISD::SSUBO: return "ssubo"; 5899 case ISD::USUBO: return "usubo"; 5900 case ISD::SMULO: return "smulo"; 5901 case ISD::UMULO: return "umulo"; 5902 case ISD::SUBC: return "subc"; 5903 case ISD::SUBE: return "sube"; 5904 case ISD::SHL_PARTS: return "shl_parts"; 5905 case ISD::SRA_PARTS: return "sra_parts"; 5906 case ISD::SRL_PARTS: return "srl_parts"; 5907 5908 // Conversion operators. 5909 case ISD::SIGN_EXTEND: return "sign_extend"; 5910 case ISD::ZERO_EXTEND: return "zero_extend"; 5911 case ISD::ANY_EXTEND: return "any_extend"; 5912 case ISD::SIGN_EXTEND_INREG: return "sign_extend_inreg"; 5913 case ISD::TRUNCATE: return "truncate"; 5914 case ISD::FP_ROUND: return "fp_round"; 5915 case ISD::FLT_ROUNDS_: return "flt_rounds"; 5916 case ISD::FP_ROUND_INREG: return "fp_round_inreg"; 5917 case ISD::FP_EXTEND: return "fp_extend"; 5918 5919 case ISD::SINT_TO_FP: return "sint_to_fp"; 5920 case ISD::UINT_TO_FP: return "uint_to_fp"; 5921 case ISD::FP_TO_SINT: return "fp_to_sint"; 5922 case ISD::FP_TO_UINT: return "fp_to_uint"; 5923 case ISD::BITCAST: return "bitcast"; 5924 case ISD::FP16_TO_FP32: return "fp16_to_fp32"; 5925 case ISD::FP32_TO_FP16: return "fp32_to_fp16"; 5926 5927 case ISD::CONVERT_RNDSAT: { 5928 switch (cast<CvtRndSatSDNode>(this)->getCvtCode()) { 5929 default: llvm_unreachable("Unknown cvt code!"); 5930 case ISD::CVT_FF: return "cvt_ff"; 5931 case ISD::CVT_FS: return "cvt_fs"; 5932 case ISD::CVT_FU: return "cvt_fu"; 5933 case ISD::CVT_SF: return "cvt_sf"; 5934 case ISD::CVT_UF: return "cvt_uf"; 5935 case ISD::CVT_SS: return "cvt_ss"; 5936 case ISD::CVT_SU: return "cvt_su"; 5937 case ISD::CVT_US: return "cvt_us"; 5938 case ISD::CVT_UU: return "cvt_uu"; 5939 } 5940 } 5941 5942 // Control flow instructions 5943 case ISD::BR: return "br"; 5944 case ISD::BRIND: return "brind"; 5945 case ISD::BR_JT: return "br_jt"; 5946 case ISD::BRCOND: return "brcond"; 5947 case ISD::BR_CC: return "br_cc"; 5948 case ISD::CALLSEQ_START: return "callseq_start"; 5949 case ISD::CALLSEQ_END: return "callseq_end"; 5950 5951 // Other operators 5952 case ISD::LOAD: return "load"; 5953 case ISD::STORE: return "store"; 5954 case ISD::VAARG: return "vaarg"; 5955 case ISD::VACOPY: return "vacopy"; 5956 case ISD::VAEND: return "vaend"; 5957 case ISD::VASTART: return "vastart"; 5958 case ISD::DYNAMIC_STACKALLOC: return "dynamic_stackalloc"; 5959 case ISD::EXTRACT_ELEMENT: return "extract_element"; 5960 case ISD::BUILD_PAIR: return "build_pair"; 5961 case ISD::STACKSAVE: return "stacksave"; 5962 case ISD::STACKRESTORE: return "stackrestore"; 5963 case ISD::TRAP: return "trap"; 5964 5965 // Bit manipulation 5966 case ISD::BSWAP: return "bswap"; 5967 case ISD::CTPOP: return "ctpop"; 5968 case ISD::CTTZ: return "cttz"; 5969 case ISD::CTLZ: return "ctlz"; 5970 5971 // Trampolines 5972 case ISD::TRAMPOLINE: return "trampoline"; 5973 5974 case ISD::CONDCODE: 5975 switch (cast<CondCodeSDNode>(this)->get()) { 5976 default: llvm_unreachable("Unknown setcc condition!"); 5977 case ISD::SETOEQ: return "setoeq"; 5978 case ISD::SETOGT: return "setogt"; 5979 case ISD::SETOGE: return "setoge"; 5980 case ISD::SETOLT: return "setolt"; 5981 case ISD::SETOLE: return "setole"; 5982 case ISD::SETONE: return "setone"; 5983 5984 case ISD::SETO: return "seto"; 5985 case ISD::SETUO: return "setuo"; 5986 case ISD::SETUEQ: return "setue"; 5987 case ISD::SETUGT: return "setugt"; 5988 case ISD::SETUGE: return "setuge"; 5989 case ISD::SETULT: return "setult"; 5990 case ISD::SETULE: return "setule"; 5991 case ISD::SETUNE: return "setune"; 5992 5993 case ISD::SETEQ: return "seteq"; 5994 case ISD::SETGT: return "setgt"; 5995 case ISD::SETGE: return "setge"; 5996 case ISD::SETLT: return "setlt"; 5997 case ISD::SETLE: return "setle"; 5998 case ISD::SETNE: return "setne"; 5999 } 6000 } 6001} 6002 6003const char *SDNode::getIndexedModeName(ISD::MemIndexedMode AM) { 6004 switch (AM) { 6005 default: 6006 return ""; 6007 case ISD::PRE_INC: 6008 return "<pre-inc>"; 6009 case ISD::PRE_DEC: 6010 return "<pre-dec>"; 6011 case ISD::POST_INC: 6012 return "<post-inc>"; 6013 case ISD::POST_DEC: 6014 return "<post-dec>"; 6015 } 6016} 6017 6018std::string ISD::ArgFlagsTy::getArgFlagsString() { 6019 std::string S = "< "; 6020 6021 if (isZExt()) 6022 S += "zext "; 6023 if (isSExt()) 6024 S += "sext "; 6025 if (isInReg()) 6026 S += "inreg "; 6027 if (isSRet()) 6028 S += "sret "; 6029 if (isByVal()) 6030 S += "byval "; 6031 if (isNest()) 6032 S += "nest "; 6033 if (getByValAlign()) 6034 S += "byval-align:" + utostr(getByValAlign()) + " "; 6035 if (getOrigAlign()) 6036 S += "orig-align:" + utostr(getOrigAlign()) + " "; 6037 if (getByValSize()) 6038 S += "byval-size:" + utostr(getByValSize()) + " "; 6039 return S + ">"; 6040} 6041 6042void SDNode::dump() const { dump(0); } 6043void SDNode::dump(const SelectionDAG *G) const { 6044 print(dbgs(), G); 6045 dbgs() << '\n'; 6046} 6047 6048void SDNode::print_types(raw_ostream &OS, const SelectionDAG *G) const { 6049 OS << (void*)this << ": "; 6050 6051 for (unsigned i = 0, e = getNumValues(); i != e; ++i) { 6052 if (i) OS << ","; 6053 if (getValueType(i) == MVT::Other) 6054 OS << "ch"; 6055 else 6056 OS << getValueType(i).getEVTString(); 6057 } 6058 OS << " = " << getOperationName(G); 6059} 6060 6061void SDNode::print_details(raw_ostream &OS, const SelectionDAG *G) const { 6062 if (const MachineSDNode *MN = dyn_cast<MachineSDNode>(this)) { 6063 if (!MN->memoperands_empty()) { 6064 OS << "<"; 6065 OS << "Mem:"; 6066 for (MachineSDNode::mmo_iterator i = MN->memoperands_begin(), 6067 e = MN->memoperands_end(); i != e; ++i) { 6068 OS << **i; 6069 if (llvm::next(i) != e) 6070 OS << " "; 6071 } 6072 OS << ">"; 6073 } 6074 } else if (const ShuffleVectorSDNode *SVN = 6075 dyn_cast<ShuffleVectorSDNode>(this)) { 6076 OS << "<"; 6077 for (unsigned i = 0, e = ValueList[0].getVectorNumElements(); i != e; ++i) { 6078 int Idx = SVN->getMaskElt(i); 6079 if (i) OS << ","; 6080 if (Idx < 0) 6081 OS << "u"; 6082 else 6083 OS << Idx; 6084 } 6085 OS << ">"; 6086 } else if (const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(this)) { 6087 OS << '<' << CSDN->getAPIntValue() << '>'; 6088 } else if (const ConstantFPSDNode *CSDN = dyn_cast<ConstantFPSDNode>(this)) { 6089 if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEsingle) 6090 OS << '<' << CSDN->getValueAPF().convertToFloat() << '>'; 6091 else if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEdouble) 6092 OS << '<' << CSDN->getValueAPF().convertToDouble() << '>'; 6093 else { 6094 OS << "<APFloat("; 6095 CSDN->getValueAPF().bitcastToAPInt().dump(); 6096 OS << ")>"; 6097 } 6098 } else if (const GlobalAddressSDNode *GADN = 6099 dyn_cast<GlobalAddressSDNode>(this)) { 6100 int64_t offset = GADN->getOffset(); 6101 OS << '<'; 6102 WriteAsOperand(OS, GADN->getGlobal()); 6103 OS << '>'; 6104 if (offset > 0) 6105 OS << " + " << offset; 6106 else 6107 OS << " " << offset; 6108 if (unsigned int TF = GADN->getTargetFlags()) 6109 OS << " [TF=" << TF << ']'; 6110 } else if (const FrameIndexSDNode *FIDN = dyn_cast<FrameIndexSDNode>(this)) { 6111 OS << "<" << FIDN->getIndex() << ">"; 6112 } else if (const JumpTableSDNode *JTDN = dyn_cast<JumpTableSDNode>(this)) { 6113 OS << "<" << JTDN->getIndex() << ">"; 6114 if (unsigned int TF = JTDN->getTargetFlags()) 6115 OS << " [TF=" << TF << ']'; 6116 } else if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(this)){ 6117 int offset = CP->getOffset(); 6118 if (CP->isMachineConstantPoolEntry()) 6119 OS << "<" << *CP->getMachineCPVal() << ">"; 6120 else 6121 OS << "<" << *CP->getConstVal() << ">"; 6122 if (offset > 0) 6123 OS << " + " << offset; 6124 else 6125 OS << " " << offset; 6126 if (unsigned int TF = CP->getTargetFlags()) 6127 OS << " [TF=" << TF << ']'; 6128 } else if (const BasicBlockSDNode *BBDN = dyn_cast<BasicBlockSDNode>(this)) { 6129 OS << "<"; 6130 const Value *LBB = (const Value*)BBDN->getBasicBlock()->getBasicBlock(); 6131 if (LBB) 6132 OS << LBB->getName() << " "; 6133 OS << (const void*)BBDN->getBasicBlock() << ">"; 6134 } else if (const RegisterSDNode *R = dyn_cast<RegisterSDNode>(this)) { 6135 OS << ' ' << PrintReg(R->getReg(), G ? G->getTarget().getRegisterInfo() :0); 6136 } else if (const ExternalSymbolSDNode *ES = 6137 dyn_cast<ExternalSymbolSDNode>(this)) { 6138 OS << "'" << ES->getSymbol() << "'"; 6139 if (unsigned int TF = ES->getTargetFlags()) 6140 OS << " [TF=" << TF << ']'; 6141 } else if (const SrcValueSDNode *M = dyn_cast<SrcValueSDNode>(this)) { 6142 if (M->getValue()) 6143 OS << "<" << M->getValue() << ">"; 6144 else 6145 OS << "<null>"; 6146 } else if (const MDNodeSDNode *MD = dyn_cast<MDNodeSDNode>(this)) { 6147 if (MD->getMD()) 6148 OS << "<" << MD->getMD() << ">"; 6149 else 6150 OS << "<null>"; 6151 } else if (const VTSDNode *N = dyn_cast<VTSDNode>(this)) { 6152 OS << ":" << N->getVT().getEVTString(); 6153 } 6154 else if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(this)) { 6155 OS << "<" << *LD->getMemOperand(); 6156 6157 bool doExt = true; 6158 switch (LD->getExtensionType()) { 6159 default: doExt = false; break; 6160 case ISD::EXTLOAD: OS << ", anyext"; break; 6161 case ISD::SEXTLOAD: OS << ", sext"; break; 6162 case ISD::ZEXTLOAD: OS << ", zext"; break; 6163 } 6164 if (doExt) 6165 OS << " from " << LD->getMemoryVT().getEVTString(); 6166 6167 const char *AM = getIndexedModeName(LD->getAddressingMode()); 6168 if (*AM) 6169 OS << ", " << AM; 6170 6171 OS << ">"; 6172 } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(this)) { 6173 OS << "<" << *ST->getMemOperand(); 6174 6175 if (ST->isTruncatingStore()) 6176 OS << ", trunc to " << ST->getMemoryVT().getEVTString(); 6177 6178 const char *AM = getIndexedModeName(ST->getAddressingMode()); 6179 if (*AM) 6180 OS << ", " << AM; 6181 6182 OS << ">"; 6183 } else if (const MemSDNode* M = dyn_cast<MemSDNode>(this)) { 6184 OS << "<" << *M->getMemOperand() << ">"; 6185 } else if (const BlockAddressSDNode *BA = 6186 dyn_cast<BlockAddressSDNode>(this)) { 6187 OS << "<"; 6188 WriteAsOperand(OS, BA->getBlockAddress()->getFunction(), false); 6189 OS << ", "; 6190 WriteAsOperand(OS, BA->getBlockAddress()->getBasicBlock(), false); 6191 OS << ">"; 6192 if (unsigned int TF = BA->getTargetFlags()) 6193 OS << " [TF=" << TF << ']'; 6194 } 6195 6196 if (G) 6197 if (unsigned Order = G->GetOrdering(this)) 6198 OS << " [ORD=" << Order << ']'; 6199 6200 if (getNodeId() != -1) 6201 OS << " [ID=" << getNodeId() << ']'; 6202 6203 DebugLoc dl = getDebugLoc(); 6204 if (G && !dl.isUnknown()) { 6205 DIScope 6206 Scope(dl.getScope(G->getMachineFunction().getFunction()->getContext())); 6207 OS << " dbg:"; 6208 // Omit the directory, since it's usually long and uninteresting. 6209 if (Scope.Verify()) 6210 OS << Scope.getFilename(); 6211 else 6212 OS << "<unknown>"; 6213 OS << ':' << dl.getLine(); 6214 if (dl.getCol() != 0) 6215 OS << ':' << dl.getCol(); 6216 } 6217} 6218 6219void SDNode::print(raw_ostream &OS, const SelectionDAG *G) const { 6220 print_types(OS, G); 6221 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) { 6222 if (i) OS << ", "; else OS << " "; 6223 OS << (void*)getOperand(i).getNode(); 6224 if (unsigned RN = getOperand(i).getResNo()) 6225 OS << ":" << RN; 6226 } 6227 print_details(OS, G); 6228} 6229 6230static void printrWithDepthHelper(raw_ostream &OS, const SDNode *N, 6231 const SelectionDAG *G, unsigned depth, 6232 unsigned indent) 6233{ 6234 if (depth == 0) 6235 return; 6236 6237 OS.indent(indent); 6238 6239 N->print(OS, G); 6240 6241 if (depth < 1) 6242 return; 6243 6244 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 6245 // Don't follow chain operands. 6246 if (N->getOperand(i).getValueType() == MVT::Other) 6247 continue; 6248 OS << '\n'; 6249 printrWithDepthHelper(OS, N->getOperand(i).getNode(), G, depth-1, indent+2); 6250 } 6251} 6252 6253void SDNode::printrWithDepth(raw_ostream &OS, const SelectionDAG *G, 6254 unsigned depth) const { 6255 printrWithDepthHelper(OS, this, G, depth, 0); 6256} 6257 6258void SDNode::printrFull(raw_ostream &OS, const SelectionDAG *G) const { 6259 // Don't print impossibly deep things. 6260 printrWithDepth(OS, G, 10); 6261} 6262 6263void SDNode::dumprWithDepth(const SelectionDAG *G, unsigned depth) const { 6264 printrWithDepth(dbgs(), G, depth); 6265} 6266 6267void SDNode::dumprFull(const SelectionDAG *G) const { 6268 // Don't print impossibly deep things. 6269 dumprWithDepth(G, 10); 6270} 6271 6272static void DumpNodes(const SDNode *N, unsigned indent, const SelectionDAG *G) { 6273 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 6274 if (N->getOperand(i).getNode()->hasOneUse()) 6275 DumpNodes(N->getOperand(i).getNode(), indent+2, G); 6276 else 6277 dbgs() << "\n" << std::string(indent+2, ' ') 6278 << (void*)N->getOperand(i).getNode() << ": <multiple use>"; 6279 6280 6281 dbgs() << "\n"; 6282 dbgs().indent(indent); 6283 N->dump(G); 6284} 6285 6286SDValue SelectionDAG::UnrollVectorOp(SDNode *N, unsigned ResNE) { 6287 assert(N->getNumValues() == 1 && 6288 "Can't unroll a vector with multiple results!"); 6289 6290 EVT VT = N->getValueType(0); 6291 unsigned NE = VT.getVectorNumElements(); 6292 EVT EltVT = VT.getVectorElementType(); 6293 DebugLoc dl = N->getDebugLoc(); 6294 6295 SmallVector<SDValue, 8> Scalars; 6296 SmallVector<SDValue, 4> Operands(N->getNumOperands()); 6297 6298 // If ResNE is 0, fully unroll the vector op. 6299 if (ResNE == 0) 6300 ResNE = NE; 6301 else if (NE > ResNE) 6302 NE = ResNE; 6303 6304 unsigned i; 6305 for (i= 0; i != NE; ++i) { 6306 for (unsigned j = 0, e = N->getNumOperands(); j != e; ++j) { 6307 SDValue Operand = N->getOperand(j); 6308 EVT OperandVT = Operand.getValueType(); 6309 if (OperandVT.isVector()) { 6310 // A vector operand; extract a single element. 6311 EVT OperandEltVT = OperandVT.getVectorElementType(); 6312 Operands[j] = getNode(ISD::EXTRACT_VECTOR_ELT, dl, 6313 OperandEltVT, 6314 Operand, 6315 getConstant(i, TLI.getPointerTy())); 6316 } else { 6317 // A scalar operand; just use it as is. 6318 Operands[j] = Operand; 6319 } 6320 } 6321 6322 switch (N->getOpcode()) { 6323 default: 6324 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, 6325 &Operands[0], Operands.size())); 6326 break; 6327 case ISD::SHL: 6328 case ISD::SRA: 6329 case ISD::SRL: 6330 case ISD::ROTL: 6331 case ISD::ROTR: 6332 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, Operands[0], 6333 getShiftAmountOperand(Operands[0].getValueType(), 6334 Operands[1]))); 6335 break; 6336 case ISD::SIGN_EXTEND_INREG: 6337 case ISD::FP_ROUND_INREG: { 6338 EVT ExtVT = cast<VTSDNode>(Operands[1])->getVT().getVectorElementType(); 6339 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, 6340 Operands[0], 6341 getValueType(ExtVT))); 6342 } 6343 } 6344 } 6345 6346 for (; i < ResNE; ++i) 6347 Scalars.push_back(getUNDEF(EltVT)); 6348 6349 return getNode(ISD::BUILD_VECTOR, dl, 6350 EVT::getVectorVT(*getContext(), EltVT, ResNE), 6351 &Scalars[0], Scalars.size()); 6352} 6353 6354 6355/// isConsecutiveLoad - Return true if LD is loading 'Bytes' bytes from a 6356/// location that is 'Dist' units away from the location that the 'Base' load 6357/// is loading from. 6358bool SelectionDAG::isConsecutiveLoad(LoadSDNode *LD, LoadSDNode *Base, 6359 unsigned Bytes, int Dist) const { 6360 if (LD->getChain() != Base->getChain()) 6361 return false; 6362 EVT VT = LD->getValueType(0); 6363 if (VT.getSizeInBits() / 8 != Bytes) 6364 return false; 6365 6366 SDValue Loc = LD->getOperand(1); 6367 SDValue BaseLoc = Base->getOperand(1); 6368 if (Loc.getOpcode() == ISD::FrameIndex) { 6369 if (BaseLoc.getOpcode() != ISD::FrameIndex) 6370 return false; 6371 const MachineFrameInfo *MFI = getMachineFunction().getFrameInfo(); 6372 int FI = cast<FrameIndexSDNode>(Loc)->getIndex(); 6373 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex(); 6374 int FS = MFI->getObjectSize(FI); 6375 int BFS = MFI->getObjectSize(BFI); 6376 if (FS != BFS || FS != (int)Bytes) return false; 6377 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes); 6378 } 6379 6380 // Handle X+C 6381 if (isBaseWithConstantOffset(Loc) && Loc.getOperand(0) == BaseLoc && 6382 cast<ConstantSDNode>(Loc.getOperand(1))->getSExtValue() == Dist*Bytes) 6383 return true; 6384 6385 const GlobalValue *GV1 = NULL; 6386 const GlobalValue *GV2 = NULL; 6387 int64_t Offset1 = 0; 6388 int64_t Offset2 = 0; 6389 bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1); 6390 bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2); 6391 if (isGA1 && isGA2 && GV1 == GV2) 6392 return Offset1 == (Offset2 + Dist*Bytes); 6393 return false; 6394} 6395 6396 6397/// InferPtrAlignment - Infer alignment of a load / store address. Return 0 if 6398/// it cannot be inferred. 6399unsigned SelectionDAG::InferPtrAlignment(SDValue Ptr) const { 6400 // If this is a GlobalAddress + cst, return the alignment. 6401 const GlobalValue *GV; 6402 int64_t GVOffset = 0; 6403 if (TLI.isGAPlusOffset(Ptr.getNode(), GV, GVOffset)) { 6404 // If GV has specified alignment, then use it. Otherwise, use the preferred 6405 // alignment. 6406 unsigned Align = GV->getAlignment(); 6407 if (!Align) { 6408 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) { 6409 if (GVar->hasInitializer()) { 6410 const TargetData *TD = TLI.getTargetData(); 6411 Align = TD->getPreferredAlignment(GVar); 6412 } 6413 } 6414 } 6415 return MinAlign(Align, GVOffset); 6416 } 6417 6418 // If this is a direct reference to a stack slot, use information about the 6419 // stack slot's alignment. 6420 int FrameIdx = 1 << 31; 6421 int64_t FrameOffset = 0; 6422 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr)) { 6423 FrameIdx = FI->getIndex(); 6424 } else if (isBaseWithConstantOffset(Ptr) && 6425 isa<FrameIndexSDNode>(Ptr.getOperand(0))) { 6426 // Handle FI+Cst 6427 FrameIdx = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex(); 6428 FrameOffset = Ptr.getConstantOperandVal(1); 6429 } 6430 6431 if (FrameIdx != (1 << 31)) { 6432 const MachineFrameInfo &MFI = *getMachineFunction().getFrameInfo(); 6433 unsigned FIInfoAlign = MinAlign(MFI.getObjectAlignment(FrameIdx), 6434 FrameOffset); 6435 return FIInfoAlign; 6436 } 6437 6438 return 0; 6439} 6440 6441void SelectionDAG::dump() const { 6442 dbgs() << "SelectionDAG has " << AllNodes.size() << " nodes:"; 6443 6444 for (allnodes_const_iterator I = allnodes_begin(), E = allnodes_end(); 6445 I != E; ++I) { 6446 const SDNode *N = I; 6447 if (!N->hasOneUse() && N != getRoot().getNode()) 6448 DumpNodes(N, 2, this); 6449 } 6450 6451 if (getRoot().getNode()) DumpNodes(getRoot().getNode(), 2, this); 6452 6453 dbgs() << "\n\n"; 6454} 6455 6456void SDNode::printr(raw_ostream &OS, const SelectionDAG *G) const { 6457 print_types(OS, G); 6458 print_details(OS, G); 6459} 6460 6461typedef SmallPtrSet<const SDNode *, 128> VisitedSDNodeSet; 6462static void DumpNodesr(raw_ostream &OS, const SDNode *N, unsigned indent, 6463 const SelectionDAG *G, VisitedSDNodeSet &once) { 6464 if (!once.insert(N)) // If we've been here before, return now. 6465 return; 6466 6467 // Dump the current SDNode, but don't end the line yet. 6468 OS << std::string(indent, ' '); 6469 N->printr(OS, G); 6470 6471 // Having printed this SDNode, walk the children: 6472 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 6473 const SDNode *child = N->getOperand(i).getNode(); 6474 6475 if (i) OS << ","; 6476 OS << " "; 6477 6478 if (child->getNumOperands() == 0) { 6479 // This child has no grandchildren; print it inline right here. 6480 child->printr(OS, G); 6481 once.insert(child); 6482 } else { // Just the address. FIXME: also print the child's opcode. 6483 OS << (void*)child; 6484 if (unsigned RN = N->getOperand(i).getResNo()) 6485 OS << ":" << RN; 6486 } 6487 } 6488 6489 OS << "\n"; 6490 6491 // Dump children that have grandchildren on their own line(s). 6492 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 6493 const SDNode *child = N->getOperand(i).getNode(); 6494 DumpNodesr(OS, child, indent+2, G, once); 6495 } 6496} 6497 6498void SDNode::dumpr() const { 6499 VisitedSDNodeSet once; 6500 DumpNodesr(dbgs(), this, 0, 0, once); 6501} 6502 6503void SDNode::dumpr(const SelectionDAG *G) const { 6504 VisitedSDNodeSet once; 6505 DumpNodesr(dbgs(), this, 0, G, once); 6506} 6507 6508 6509// getAddressSpace - Return the address space this GlobalAddress belongs to. 6510unsigned GlobalAddressSDNode::getAddressSpace() const { 6511 return getGlobal()->getType()->getAddressSpace(); 6512} 6513 6514 6515const Type *ConstantPoolSDNode::getType() const { 6516 if (isMachineConstantPoolEntry()) 6517 return Val.MachineCPVal->getType(); 6518 return Val.ConstVal->getType(); 6519} 6520 6521bool BuildVectorSDNode::isConstantSplat(APInt &SplatValue, 6522 APInt &SplatUndef, 6523 unsigned &SplatBitSize, 6524 bool &HasAnyUndefs, 6525 unsigned MinSplatBits, 6526 bool isBigEndian) { 6527 EVT VT = getValueType(0); 6528 assert(VT.isVector() && "Expected a vector type"); 6529 unsigned sz = VT.getSizeInBits(); 6530 if (MinSplatBits > sz) 6531 return false; 6532 6533 SplatValue = APInt(sz, 0); 6534 SplatUndef = APInt(sz, 0); 6535 6536 // Get the bits. Bits with undefined values (when the corresponding element 6537 // of the vector is an ISD::UNDEF value) are set in SplatUndef and cleared 6538 // in SplatValue. If any of the values are not constant, give up and return 6539 // false. 6540 unsigned int nOps = getNumOperands(); 6541 assert(nOps > 0 && "isConstantSplat has 0-size build vector"); 6542 unsigned EltBitSize = VT.getVectorElementType().getSizeInBits(); 6543 6544 for (unsigned j = 0; j < nOps; ++j) { 6545 unsigned i = isBigEndian ? nOps-1-j : j; 6546 SDValue OpVal = getOperand(i); 6547 unsigned BitPos = j * EltBitSize; 6548 6549 if (OpVal.getOpcode() == ISD::UNDEF) 6550 SplatUndef |= APInt::getBitsSet(sz, BitPos, BitPos + EltBitSize); 6551 else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) 6552 SplatValue |= CN->getAPIntValue().zextOrTrunc(EltBitSize). 6553 zextOrTrunc(sz) << BitPos; 6554 else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) 6555 SplatValue |= CN->getValueAPF().bitcastToAPInt().zextOrTrunc(sz) <<BitPos; 6556 else 6557 return false; 6558 } 6559 6560 // The build_vector is all constants or undefs. Find the smallest element 6561 // size that splats the vector. 6562 6563 HasAnyUndefs = (SplatUndef != 0); 6564 while (sz > 8) { 6565 6566 unsigned HalfSize = sz / 2; 6567 APInt HighValue = SplatValue.lshr(HalfSize).trunc(HalfSize); 6568 APInt LowValue = SplatValue.trunc(HalfSize); 6569 APInt HighUndef = SplatUndef.lshr(HalfSize).trunc(HalfSize); 6570 APInt LowUndef = SplatUndef.trunc(HalfSize); 6571 6572 // If the two halves do not match (ignoring undef bits), stop here. 6573 if ((HighValue & ~LowUndef) != (LowValue & ~HighUndef) || 6574 MinSplatBits > HalfSize) 6575 break; 6576 6577 SplatValue = HighValue | LowValue; 6578 SplatUndef = HighUndef & LowUndef; 6579 6580 sz = HalfSize; 6581 } 6582 6583 SplatBitSize = sz; 6584 return true; 6585} 6586 6587bool ShuffleVectorSDNode::isSplatMask(const int *Mask, EVT VT) { 6588 // Find the first non-undef value in the shuffle mask. 6589 unsigned i, e; 6590 for (i = 0, e = VT.getVectorNumElements(); i != e && Mask[i] < 0; ++i) 6591 /* search */; 6592 6593 assert(i != e && "VECTOR_SHUFFLE node with all undef indices!"); 6594 6595 // Make sure all remaining elements are either undef or the same as the first 6596 // non-undef value. 6597 for (int Idx = Mask[i]; i != e; ++i) 6598 if (Mask[i] >= 0 && Mask[i] != Idx) 6599 return false; 6600 return true; 6601} 6602 6603#ifdef XDEBUG 6604static void checkForCyclesHelper(const SDNode *N, 6605 SmallPtrSet<const SDNode*, 32> &Visited, 6606 SmallPtrSet<const SDNode*, 32> &Checked) { 6607 // If this node has already been checked, don't check it again. 6608 if (Checked.count(N)) 6609 return; 6610 6611 // If a node has already been visited on this depth-first walk, reject it as 6612 // a cycle. 6613 if (!Visited.insert(N)) { 6614 dbgs() << "Offending node:\n"; 6615 N->dumprFull(); 6616 errs() << "Detected cycle in SelectionDAG\n"; 6617 abort(); 6618 } 6619 6620 for(unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 6621 checkForCyclesHelper(N->getOperand(i).getNode(), Visited, Checked); 6622 6623 Checked.insert(N); 6624 Visited.erase(N); 6625} 6626#endif 6627 6628void llvm::checkForCycles(const llvm::SDNode *N) { 6629#ifdef XDEBUG 6630 assert(N && "Checking nonexistant SDNode"); 6631 SmallPtrSet<const SDNode*, 32> visited; 6632 SmallPtrSet<const SDNode*, 32> checked; 6633 checkForCyclesHelper(N, visited, checked); 6634#endif 6635} 6636 6637void llvm::checkForCycles(const llvm::SelectionDAG *DAG) { 6638 checkForCycles(DAG->getRoot().getNode()); 6639} 6640