1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file defines the interfaces that ARM uses to lower LLVM code into a 11 // selection DAG. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #define DEBUG_TYPE "arm-isel" 16 #include "ARMISelLowering.h" 17 #include "ARM.h" 18 #include "ARMCallingConv.h" 19 #include "ARMConstantPoolValue.h" 20 #include "ARMMachineFunctionInfo.h" 21 #include "ARMPerfectShuffle.h" 22 #include "ARMSubtarget.h" 23 #include "ARMTargetMachine.h" 24 #include "ARMTargetObjectFile.h" 25 #include "MCTargetDesc/ARMAddressingModes.h" 26 #include "llvm/CallingConv.h" 27 #include "llvm/Constants.h" 28 #include "llvm/Function.h" 29 #include "llvm/GlobalValue.h" 30 #include "llvm/Instruction.h" 31 #include "llvm/Instructions.h" 32 #include "llvm/Intrinsics.h" 33 #include "llvm/Type.h" 34 #include "llvm/CodeGen/CallingConvLower.h" 35 #include "llvm/CodeGen/IntrinsicLowering.h" 36 #include "llvm/CodeGen/MachineBasicBlock.h" 37 #include "llvm/CodeGen/MachineFrameInfo.h" 38 #include "llvm/CodeGen/MachineFunction.h" 39 #include "llvm/CodeGen/MachineInstrBuilder.h" 40 #include "llvm/CodeGen/MachineModuleInfo.h" 41 #include "llvm/CodeGen/MachineRegisterInfo.h" 42 #include "llvm/CodeGen/SelectionDAG.h" 43 #include "llvm/MC/MCSectionMachO.h" 44 #include "llvm/Target/TargetOptions.h" 45 #include "llvm/ADT/StringExtras.h" 46 #include "llvm/ADT/Statistic.h" 47 #include "llvm/Support/CommandLine.h" 48 #include "llvm/Support/ErrorHandling.h" 49 #include "llvm/Support/MathExtras.h" 50 #include "llvm/Support/raw_ostream.h" 51 using namespace llvm; 52 53 STATISTIC(NumTailCalls, "Number of tail calls"); 54 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt"); 55 56 // This option should go away when tail calls fully work. 57 static cl::opt<bool> 58 EnableARMTailCalls("arm-tail-calls", cl::Hidden, 59 cl::desc("Generate tail calls (TEMPORARY OPTION)."), 60 cl::init(false)); 61 62 cl::opt<bool> 63 EnableARMLongCalls("arm-long-calls", cl::Hidden, 64 cl::desc("Generate calls via indirect call instructions"), 65 cl::init(false)); 66 67 static cl::opt<bool> 68 ARMInterworking("arm-interworking", cl::Hidden, 69 cl::desc("Enable / disable ARM interworking (for debugging only)"), 70 cl::init(true)); 71 72 namespace { 73 class ARMCCState : public CCState { 74 public: 75 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF, 76 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs, 77 LLVMContext &C, ParmContext PC) 78 : CCState(CC, isVarArg, MF, TM, locs, C) { 79 assert(((PC == Call) || (PC == Prologue)) && 80 "ARMCCState users must specify whether their context is call" 81 "or prologue generation."); 82 CallOrPrologue = PC; 83 } 84 }; 85 } 86 87 // The APCS parameter registers. 88 static const uint16_t GPRArgRegs[] = { 89 ARM::R0, ARM::R1, ARM::R2, ARM::R3 90 }; 91 92 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT, 93 EVT PromotedBitwiseVT) { 94 if (VT != PromotedLdStVT) { 95 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote); 96 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(), 97 PromotedLdStVT.getSimpleVT()); 98 99 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote); 100 AddPromotedToType (ISD::STORE, VT.getSimpleVT(), 101 PromotedLdStVT.getSimpleVT()); 102 } 103 104 EVT ElemTy = VT.getVectorElementType(); 105 if (ElemTy != MVT::i64 && ElemTy != MVT::f64) 106 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom); 107 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom); 108 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom); 109 if (ElemTy == MVT::i32) { 110 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Custom); 111 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Custom); 112 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom); 113 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom); 114 } else { 115 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand); 116 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand); 117 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand); 118 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand); 119 } 120 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom); 121 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom); 122 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal); 123 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal); 124 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand); 125 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand); 126 setOperationAction(ISD::SIGN_EXTEND_INREG, VT.getSimpleVT(), Expand); 127 if (VT.isInteger()) { 128 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom); 129 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom); 130 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom); 131 } 132 133 // Promote all bit-wise operations. 134 if (VT.isInteger() && VT != PromotedBitwiseVT) { 135 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote); 136 AddPromotedToType (ISD::AND, VT.getSimpleVT(), 137 PromotedBitwiseVT.getSimpleVT()); 138 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote); 139 AddPromotedToType (ISD::OR, VT.getSimpleVT(), 140 PromotedBitwiseVT.getSimpleVT()); 141 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote); 142 AddPromotedToType (ISD::XOR, VT.getSimpleVT(), 143 PromotedBitwiseVT.getSimpleVT()); 144 } 145 146 // Neon does not support vector divide/remainder operations. 147 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand); 148 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand); 149 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand); 150 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand); 151 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand); 152 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand); 153 } 154 155 void ARMTargetLowering::addDRTypeForNEON(EVT VT) { 156 addRegisterClass(VT, ARM::DPRRegisterClass); 157 addTypeForNEON(VT, MVT::f64, MVT::v2i32); 158 } 159 160 void ARMTargetLowering::addQRTypeForNEON(EVT VT) { 161 addRegisterClass(VT, ARM::QPRRegisterClass); 162 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32); 163 } 164 165 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) { 166 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin()) 167 return new TargetLoweringObjectFileMachO(); 168 169 return new ARMElfTargetObjectFile(); 170 } 171 172 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) 173 : TargetLowering(TM, createTLOF(TM)) { 174 Subtarget = &TM.getSubtarget<ARMSubtarget>(); 175 RegInfo = TM.getRegisterInfo(); 176 Itins = TM.getInstrItineraryData(); 177 178 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); 179 180 if (Subtarget->isTargetDarwin()) { 181 // Uses VFP for Thumb libfuncs if available. 182 if (Subtarget->isThumb() && Subtarget->hasVFP2()) { 183 // Single-precision floating-point arithmetic. 184 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp"); 185 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp"); 186 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp"); 187 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp"); 188 189 // Double-precision floating-point arithmetic. 190 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp"); 191 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp"); 192 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp"); 193 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp"); 194 195 // Single-precision comparisons. 196 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp"); 197 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp"); 198 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp"); 199 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp"); 200 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp"); 201 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp"); 202 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp"); 203 setLibcallName(RTLIB::O_F32, "__unordsf2vfp"); 204 205 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE); 206 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE); 207 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE); 208 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE); 209 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE); 210 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE); 211 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE); 212 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ); 213 214 // Double-precision comparisons. 215 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp"); 216 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp"); 217 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp"); 218 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp"); 219 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp"); 220 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp"); 221 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp"); 222 setLibcallName(RTLIB::O_F64, "__unorddf2vfp"); 223 224 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE); 225 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE); 226 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE); 227 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE); 228 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE); 229 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE); 230 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE); 231 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ); 232 233 // Floating-point to integer conversions. 234 // i64 conversions are done via library routines even when generating VFP 235 // instructions, so use the same ones. 236 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp"); 237 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp"); 238 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp"); 239 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp"); 240 241 // Conversions between floating types. 242 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp"); 243 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp"); 244 245 // Integer to floating-point conversions. 246 // i64 conversions are done via library routines even when generating VFP 247 // instructions, so use the same ones. 248 // FIXME: There appears to be some naming inconsistency in ARM libgcc: 249 // e.g., __floatunsidf vs. __floatunssidfvfp. 250 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp"); 251 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp"); 252 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp"); 253 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp"); 254 } 255 } 256 257 // These libcalls are not available in 32-bit. 258 setLibcallName(RTLIB::SHL_I128, 0); 259 setLibcallName(RTLIB::SRL_I128, 0); 260 setLibcallName(RTLIB::SRA_I128, 0); 261 262 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) { 263 // Double-precision floating-point arithmetic helper functions 264 // RTABI chapter 4.1.2, Table 2 265 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd"); 266 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv"); 267 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul"); 268 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub"); 269 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS); 270 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS); 271 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS); 272 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS); 273 274 // Double-precision floating-point comparison helper functions 275 // RTABI chapter 4.1.2, Table 3 276 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq"); 277 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE); 278 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq"); 279 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ); 280 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt"); 281 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE); 282 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple"); 283 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE); 284 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge"); 285 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE); 286 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt"); 287 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE); 288 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun"); 289 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE); 290 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun"); 291 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ); 292 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS); 293 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS); 294 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS); 295 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS); 296 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS); 297 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS); 298 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS); 299 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS); 300 301 // Single-precision floating-point arithmetic helper functions 302 // RTABI chapter 4.1.2, Table 4 303 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd"); 304 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv"); 305 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul"); 306 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub"); 307 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS); 308 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS); 309 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS); 310 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS); 311 312 // Single-precision floating-point comparison helper functions 313 // RTABI chapter 4.1.2, Table 5 314 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq"); 315 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE); 316 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq"); 317 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ); 318 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt"); 319 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE); 320 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple"); 321 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE); 322 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge"); 323 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE); 324 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt"); 325 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE); 326 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun"); 327 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE); 328 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun"); 329 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ); 330 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS); 331 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS); 332 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS); 333 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS); 334 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS); 335 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS); 336 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS); 337 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS); 338 339 // Floating-point to integer conversions. 340 // RTABI chapter 4.1.2, Table 6 341 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz"); 342 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz"); 343 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz"); 344 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz"); 345 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz"); 346 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz"); 347 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz"); 348 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz"); 349 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS); 350 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS); 351 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS); 352 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS); 353 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS); 354 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS); 355 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS); 356 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS); 357 358 // Conversions between floating types. 359 // RTABI chapter 4.1.2, Table 7 360 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f"); 361 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d"); 362 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS); 363 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS); 364 365 // Integer to floating-point conversions. 366 // RTABI chapter 4.1.2, Table 8 367 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d"); 368 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d"); 369 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d"); 370 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d"); 371 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f"); 372 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f"); 373 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f"); 374 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f"); 375 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS); 376 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS); 377 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS); 378 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS); 379 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS); 380 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS); 381 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS); 382 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS); 383 384 // Long long helper functions 385 // RTABI chapter 4.2, Table 9 386 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul"); 387 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl"); 388 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr"); 389 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr"); 390 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS); 391 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS); 392 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS); 393 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS); 394 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS); 395 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS); 396 397 // Integer division functions 398 // RTABI chapter 4.3.1 399 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv"); 400 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv"); 401 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv"); 402 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod"); 403 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv"); 404 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv"); 405 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv"); 406 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod"); 407 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS); 408 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS); 409 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS); 410 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS); 411 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS); 412 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS); 413 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS); 414 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS); 415 416 // Memory operations 417 // RTABI chapter 4.3.4 418 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy"); 419 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove"); 420 setLibcallName(RTLIB::MEMSET, "__aeabi_memset"); 421 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS); 422 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS); 423 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS); 424 } 425 426 // Use divmod compiler-rt calls for iOS 5.0 and later. 427 if (Subtarget->getTargetTriple().getOS() == Triple::IOS && 428 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) { 429 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4"); 430 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4"); 431 } 432 433 if (Subtarget->isThumb1Only()) 434 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass); 435 else 436 addRegisterClass(MVT::i32, ARM::GPRRegisterClass); 437 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() && 438 !Subtarget->isThumb1Only()) { 439 addRegisterClass(MVT::f32, ARM::SPRRegisterClass); 440 if (!Subtarget->isFPOnlySP()) 441 addRegisterClass(MVT::f64, ARM::DPRRegisterClass); 442 443 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 444 } 445 446 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE; 447 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) { 448 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE; 449 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT) 450 setTruncStoreAction((MVT::SimpleValueType)VT, 451 (MVT::SimpleValueType)InnerVT, Expand); 452 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand); 453 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand); 454 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand); 455 } 456 457 setOperationAction(ISD::ConstantFP, MVT::f32, Custom); 458 459 if (Subtarget->hasNEON()) { 460 addDRTypeForNEON(MVT::v2f32); 461 addDRTypeForNEON(MVT::v8i8); 462 addDRTypeForNEON(MVT::v4i16); 463 addDRTypeForNEON(MVT::v2i32); 464 addDRTypeForNEON(MVT::v1i64); 465 466 addQRTypeForNEON(MVT::v4f32); 467 addQRTypeForNEON(MVT::v2f64); 468 addQRTypeForNEON(MVT::v16i8); 469 addQRTypeForNEON(MVT::v8i16); 470 addQRTypeForNEON(MVT::v4i32); 471 addQRTypeForNEON(MVT::v2i64); 472 473 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but 474 // neither Neon nor VFP support any arithmetic operations on it. 475 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively 476 // supported for v4f32. 477 setOperationAction(ISD::FADD, MVT::v2f64, Expand); 478 setOperationAction(ISD::FSUB, MVT::v2f64, Expand); 479 setOperationAction(ISD::FMUL, MVT::v2f64, Expand); 480 // FIXME: Code duplication: FDIV and FREM are expanded always, see 481 // ARMTargetLowering::addTypeForNEON method for details. 482 setOperationAction(ISD::FDIV, MVT::v2f64, Expand); 483 setOperationAction(ISD::FREM, MVT::v2f64, Expand); 484 // FIXME: Create unittest. 485 // In another words, find a way when "copysign" appears in DAG with vector 486 // operands. 487 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand); 488 // FIXME: Code duplication: SETCC has custom operation action, see 489 // ARMTargetLowering::addTypeForNEON method for details. 490 setOperationAction(ISD::SETCC, MVT::v2f64, Expand); 491 // FIXME: Create unittest for FNEG and for FABS. 492 setOperationAction(ISD::FNEG, MVT::v2f64, Expand); 493 setOperationAction(ISD::FABS, MVT::v2f64, Expand); 494 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand); 495 setOperationAction(ISD::FSIN, MVT::v2f64, Expand); 496 setOperationAction(ISD::FCOS, MVT::v2f64, Expand); 497 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand); 498 setOperationAction(ISD::FPOW, MVT::v2f64, Expand); 499 setOperationAction(ISD::FLOG, MVT::v2f64, Expand); 500 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand); 501 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand); 502 setOperationAction(ISD::FEXP, MVT::v2f64, Expand); 503 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand); 504 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR. 505 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand); 506 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand); 507 setOperationAction(ISD::FRINT, MVT::v2f64, Expand); 508 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand); 509 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand); 510 511 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand); 512 setOperationAction(ISD::FSIN, MVT::v4f32, Expand); 513 setOperationAction(ISD::FCOS, MVT::v4f32, Expand); 514 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand); 515 setOperationAction(ISD::FPOW, MVT::v4f32, Expand); 516 setOperationAction(ISD::FLOG, MVT::v4f32, Expand); 517 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand); 518 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand); 519 setOperationAction(ISD::FEXP, MVT::v4f32, Expand); 520 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand); 521 522 // Neon does not support some operations on v1i64 and v2i64 types. 523 setOperationAction(ISD::MUL, MVT::v1i64, Expand); 524 // Custom handling for some quad-vector types to detect VMULL. 525 setOperationAction(ISD::MUL, MVT::v8i16, Custom); 526 setOperationAction(ISD::MUL, MVT::v4i32, Custom); 527 setOperationAction(ISD::MUL, MVT::v2i64, Custom); 528 // Custom handling for some vector types to avoid expensive expansions 529 setOperationAction(ISD::SDIV, MVT::v4i16, Custom); 530 setOperationAction(ISD::SDIV, MVT::v8i8, Custom); 531 setOperationAction(ISD::UDIV, MVT::v4i16, Custom); 532 setOperationAction(ISD::UDIV, MVT::v8i8, Custom); 533 setOperationAction(ISD::SETCC, MVT::v1i64, Expand); 534 setOperationAction(ISD::SETCC, MVT::v2i64, Expand); 535 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with 536 // a destination type that is wider than the source, and nor does 537 // it have a FP_TO_[SU]INT instruction with a narrower destination than 538 // source. 539 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom); 540 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom); 541 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom); 542 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom); 543 544 setTargetDAGCombine(ISD::INTRINSIC_VOID); 545 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN); 546 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN); 547 setTargetDAGCombine(ISD::SHL); 548 setTargetDAGCombine(ISD::SRL); 549 setTargetDAGCombine(ISD::SRA); 550 setTargetDAGCombine(ISD::SIGN_EXTEND); 551 setTargetDAGCombine(ISD::ZERO_EXTEND); 552 setTargetDAGCombine(ISD::ANY_EXTEND); 553 setTargetDAGCombine(ISD::SELECT_CC); 554 setTargetDAGCombine(ISD::BUILD_VECTOR); 555 setTargetDAGCombine(ISD::VECTOR_SHUFFLE); 556 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT); 557 setTargetDAGCombine(ISD::STORE); 558 setTargetDAGCombine(ISD::FP_TO_SINT); 559 setTargetDAGCombine(ISD::FP_TO_UINT); 560 setTargetDAGCombine(ISD::FDIV); 561 562 // It is legal to extload from v4i8 to v4i16 or v4i32. 563 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8, 564 MVT::v4i16, MVT::v2i16, 565 MVT::v2i32}; 566 for (unsigned i = 0; i < 6; ++i) { 567 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal); 568 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal); 569 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal); 570 } 571 } 572 573 computeRegisterProperties(); 574 575 // ARM does not have f32 extending load. 576 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand); 577 578 // ARM does not have i1 sign extending load. 579 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 580 581 // ARM supports all 4 flavors of integer indexed load / store. 582 if (!Subtarget->isThumb1Only()) { 583 for (unsigned im = (unsigned)ISD::PRE_INC; 584 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) { 585 setIndexedLoadAction(im, MVT::i1, Legal); 586 setIndexedLoadAction(im, MVT::i8, Legal); 587 setIndexedLoadAction(im, MVT::i16, Legal); 588 setIndexedLoadAction(im, MVT::i32, Legal); 589 setIndexedStoreAction(im, MVT::i1, Legal); 590 setIndexedStoreAction(im, MVT::i8, Legal); 591 setIndexedStoreAction(im, MVT::i16, Legal); 592 setIndexedStoreAction(im, MVT::i32, Legal); 593 } 594 } 595 596 // i64 operation support. 597 setOperationAction(ISD::MUL, MVT::i64, Expand); 598 setOperationAction(ISD::MULHU, MVT::i32, Expand); 599 if (Subtarget->isThumb1Only()) { 600 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand); 601 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand); 602 } 603 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops() 604 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP())) 605 setOperationAction(ISD::MULHS, MVT::i32, Expand); 606 607 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); 608 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); 609 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); 610 setOperationAction(ISD::SRL, MVT::i64, Custom); 611 setOperationAction(ISD::SRA, MVT::i64, Custom); 612 613 if (!Subtarget->isThumb1Only()) { 614 // FIXME: We should do this for Thumb1 as well. 615 setOperationAction(ISD::ADDC, MVT::i32, Custom); 616 setOperationAction(ISD::ADDE, MVT::i32, Custom); 617 setOperationAction(ISD::SUBC, MVT::i32, Custom); 618 setOperationAction(ISD::SUBE, MVT::i32, Custom); 619 } 620 621 // ARM does not have ROTL. 622 setOperationAction(ISD::ROTL, MVT::i32, Expand); 623 setOperationAction(ISD::CTTZ, MVT::i32, Custom); 624 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 625 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only()) 626 setOperationAction(ISD::CTLZ, MVT::i32, Expand); 627 628 // These just redirect to CTTZ and CTLZ on ARM. 629 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand); 630 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand); 631 632 // Only ARMv6 has BSWAP. 633 if (!Subtarget->hasV6Ops()) 634 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 635 636 // These are expanded into libcalls. 637 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) { 638 // v7M has a hardware divider 639 setOperationAction(ISD::SDIV, MVT::i32, Expand); 640 setOperationAction(ISD::UDIV, MVT::i32, Expand); 641 } 642 setOperationAction(ISD::SREM, MVT::i32, Expand); 643 setOperationAction(ISD::UREM, MVT::i32, Expand); 644 setOperationAction(ISD::SDIVREM, MVT::i32, Expand); 645 setOperationAction(ISD::UDIVREM, MVT::i32, Expand); 646 647 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 648 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 649 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom); 650 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 651 setOperationAction(ISD::BlockAddress, MVT::i32, Custom); 652 653 setOperationAction(ISD::TRAP, MVT::Other, Legal); 654 655 // Use the default implementation. 656 setOperationAction(ISD::VASTART, MVT::Other, Custom); 657 setOperationAction(ISD::VAARG, MVT::Other, Expand); 658 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 659 setOperationAction(ISD::VAEND, MVT::Other, Expand); 660 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 661 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 662 663 if (!Subtarget->isTargetDarwin()) { 664 // Non-Darwin platforms may return values in these registers via the 665 // personality function. 666 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand); 667 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand); 668 setExceptionPointerRegister(ARM::R0); 669 setExceptionSelectorRegister(ARM::R1); 670 } 671 672 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 673 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use 674 // the default expansion. 675 // FIXME: This should be checking for v6k, not just v6. 676 if (Subtarget->hasDataBarrier() || 677 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) { 678 // membarrier needs custom lowering; the rest are legal and handled 679 // normally. 680 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom); 681 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom); 682 // Custom lowering for 64-bit ops 683 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom); 684 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom); 685 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom); 686 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom); 687 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom); 688 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom); 689 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom); 690 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc. 691 setInsertFencesForAtomic(true); 692 } else { 693 // Set them all for expansion, which will force libcalls. 694 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand); 695 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand); 696 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand); 697 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand); 698 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand); 699 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand); 700 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand); 701 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand); 702 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand); 703 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand); 704 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand); 705 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand); 706 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand); 707 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand); 708 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the 709 // Unordered/Monotonic case. 710 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom); 711 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom); 712 // Since the libcalls include locking, fold in the fences 713 setShouldFoldAtomicFences(true); 714 } 715 716 setOperationAction(ISD::PREFETCH, MVT::Other, Custom); 717 718 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes. 719 if (!Subtarget->hasV6Ops()) { 720 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 721 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); 722 } 723 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 724 725 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() && 726 !Subtarget->isThumb1Only()) { 727 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR 728 // iff target supports vfp2. 729 setOperationAction(ISD::BITCAST, MVT::i64, Custom); 730 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom); 731 } 732 733 // We want to custom lower some of our intrinsics. 734 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); 735 if (Subtarget->isTargetDarwin()) { 736 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom); 737 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom); 738 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume"); 739 } 740 741 setOperationAction(ISD::SETCC, MVT::i32, Expand); 742 setOperationAction(ISD::SETCC, MVT::f32, Expand); 743 setOperationAction(ISD::SETCC, MVT::f64, Expand); 744 setOperationAction(ISD::SELECT, MVT::i32, Custom); 745 setOperationAction(ISD::SELECT, MVT::f32, Custom); 746 setOperationAction(ISD::SELECT, MVT::f64, Custom); 747 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom); 748 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 749 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 750 751 setOperationAction(ISD::BRCOND, MVT::Other, Expand); 752 setOperationAction(ISD::BR_CC, MVT::i32, Custom); 753 setOperationAction(ISD::BR_CC, MVT::f32, Custom); 754 setOperationAction(ISD::BR_CC, MVT::f64, Custom); 755 setOperationAction(ISD::BR_JT, MVT::Other, Custom); 756 757 // We don't support sin/cos/fmod/copysign/pow 758 setOperationAction(ISD::FSIN, MVT::f64, Expand); 759 setOperationAction(ISD::FSIN, MVT::f32, Expand); 760 setOperationAction(ISD::FCOS, MVT::f32, Expand); 761 setOperationAction(ISD::FCOS, MVT::f64, Expand); 762 setOperationAction(ISD::FREM, MVT::f64, Expand); 763 setOperationAction(ISD::FREM, MVT::f32, Expand); 764 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() && 765 !Subtarget->isThumb1Only()) { 766 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 767 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 768 } 769 setOperationAction(ISD::FPOW, MVT::f64, Expand); 770 setOperationAction(ISD::FPOW, MVT::f32, Expand); 771 772 if (!Subtarget->hasVFP4()) { 773 setOperationAction(ISD::FMA, MVT::f64, Expand); 774 setOperationAction(ISD::FMA, MVT::f32, Expand); 775 } 776 777 // Various VFP goodness 778 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) { 779 // int <-> fp are custom expanded into bit_convert + ARMISD ops. 780 if (Subtarget->hasVFP2()) { 781 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom); 782 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom); 783 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); 784 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 785 } 786 // Special handling for half-precision FP. 787 if (!Subtarget->hasFP16()) { 788 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand); 789 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand); 790 } 791 } 792 793 // We have target-specific dag combine patterns for the following nodes: 794 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine 795 setTargetDAGCombine(ISD::ADD); 796 setTargetDAGCombine(ISD::SUB); 797 setTargetDAGCombine(ISD::MUL); 798 799 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON()) { 800 setTargetDAGCombine(ISD::AND); 801 setTargetDAGCombine(ISD::OR); 802 setTargetDAGCombine(ISD::XOR); 803 } 804 805 if (Subtarget->hasV6Ops()) 806 setTargetDAGCombine(ISD::SRL); 807 808 setStackPointerRegisterToSaveRestore(ARM::SP); 809 810 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() || 811 !Subtarget->hasVFP2()) 812 setSchedulingPreference(Sched::RegPressure); 813 else 814 setSchedulingPreference(Sched::Hybrid); 815 816 //// temporary - rewrite interface to use type 817 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1; 818 maxStoresPerMemset = 16; 819 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4; 820 821 // On ARM arguments smaller than 4 bytes are extended, so all arguments 822 // are at least 4 bytes aligned. 823 setMinStackArgumentAlignment(4); 824 825 benefitFromCodePlacementOpt = true; 826 827 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2); 828 } 829 830 // FIXME: It might make sense to define the representative register class as the 831 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is 832 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently, 833 // SPR's representative would be DPR_VFP2. This should work well if register 834 // pressure tracking were modified such that a register use would increment the 835 // pressure of the register class's representative and all of it's super 836 // classes' representatives transitively. We have not implemented this because 837 // of the difficulty prior to coalescing of modeling operand register classes 838 // due to the common occurrence of cross class copies and subregister insertions 839 // and extractions. 840 std::pair<const TargetRegisterClass*, uint8_t> 841 ARMTargetLowering::findRepresentativeClass(EVT VT) const{ 842 const TargetRegisterClass *RRC = 0; 843 uint8_t Cost = 1; 844 switch (VT.getSimpleVT().SimpleTy) { 845 default: 846 return TargetLowering::findRepresentativeClass(VT); 847 // Use DPR as representative register class for all floating point 848 // and vector types. Since there are 32 SPR registers and 32 DPR registers so 849 // the cost is 1 for both f32 and f64. 850 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16: 851 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32: 852 RRC = ARM::DPRRegisterClass; 853 // When NEON is used for SP, only half of the register file is available 854 // because operations that define both SP and DP results will be constrained 855 // to the VFP2 class (D0-D15). We currently model this constraint prior to 856 // coalescing by double-counting the SP regs. See the FIXME above. 857 if (Subtarget->useNEONForSinglePrecisionFP()) 858 Cost = 2; 859 break; 860 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64: 861 case MVT::v4f32: case MVT::v2f64: 862 RRC = ARM::DPRRegisterClass; 863 Cost = 2; 864 break; 865 case MVT::v4i64: 866 RRC = ARM::DPRRegisterClass; 867 Cost = 4; 868 break; 869 case MVT::v8i64: 870 RRC = ARM::DPRRegisterClass; 871 Cost = 8; 872 break; 873 } 874 return std::make_pair(RRC, Cost); 875 } 876 877 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const { 878 switch (Opcode) { 879 default: return 0; 880 case ARMISD::Wrapper: return "ARMISD::Wrapper"; 881 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN"; 882 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC"; 883 case ARMISD::WrapperJT: return "ARMISD::WrapperJT"; 884 case ARMISD::CALL: return "ARMISD::CALL"; 885 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED"; 886 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK"; 887 case ARMISD::tCALL: return "ARMISD::tCALL"; 888 case ARMISD::BRCOND: return "ARMISD::BRCOND"; 889 case ARMISD::BR_JT: return "ARMISD::BR_JT"; 890 case ARMISD::BR2_JT: return "ARMISD::BR2_JT"; 891 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG"; 892 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD"; 893 case ARMISD::CMP: return "ARMISD::CMP"; 894 case ARMISD::CMPZ: return "ARMISD::CMPZ"; 895 case ARMISD::CMPFP: return "ARMISD::CMPFP"; 896 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0"; 897 case ARMISD::BCC_i64: return "ARMISD::BCC_i64"; 898 case ARMISD::FMSTAT: return "ARMISD::FMSTAT"; 899 900 case ARMISD::CMOV: return "ARMISD::CMOV"; 901 case ARMISD::CAND: return "ARMISD::CAND"; 902 case ARMISD::COR: return "ARMISD::COR"; 903 case ARMISD::CXOR: return "ARMISD::CXOR"; 904 905 case ARMISD::RBIT: return "ARMISD::RBIT"; 906 907 case ARMISD::FTOSI: return "ARMISD::FTOSI"; 908 case ARMISD::FTOUI: return "ARMISD::FTOUI"; 909 case ARMISD::SITOF: return "ARMISD::SITOF"; 910 case ARMISD::UITOF: return "ARMISD::UITOF"; 911 912 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG"; 913 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG"; 914 case ARMISD::RRX: return "ARMISD::RRX"; 915 916 case ARMISD::ADDC: return "ARMISD::ADDC"; 917 case ARMISD::ADDE: return "ARMISD::ADDE"; 918 case ARMISD::SUBC: return "ARMISD::SUBC"; 919 case ARMISD::SUBE: return "ARMISD::SUBE"; 920 921 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD"; 922 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR"; 923 924 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP"; 925 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP"; 926 927 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN"; 928 929 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER"; 930 931 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC"; 932 933 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER"; 934 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR"; 935 936 case ARMISD::PRELOAD: return "ARMISD::PRELOAD"; 937 938 case ARMISD::VCEQ: return "ARMISD::VCEQ"; 939 case ARMISD::VCEQZ: return "ARMISD::VCEQZ"; 940 case ARMISD::VCGE: return "ARMISD::VCGE"; 941 case ARMISD::VCGEZ: return "ARMISD::VCGEZ"; 942 case ARMISD::VCLEZ: return "ARMISD::VCLEZ"; 943 case ARMISD::VCGEU: return "ARMISD::VCGEU"; 944 case ARMISD::VCGT: return "ARMISD::VCGT"; 945 case ARMISD::VCGTZ: return "ARMISD::VCGTZ"; 946 case ARMISD::VCLTZ: return "ARMISD::VCLTZ"; 947 case ARMISD::VCGTU: return "ARMISD::VCGTU"; 948 case ARMISD::VTST: return "ARMISD::VTST"; 949 950 case ARMISD::VSHL: return "ARMISD::VSHL"; 951 case ARMISD::VSHRs: return "ARMISD::VSHRs"; 952 case ARMISD::VSHRu: return "ARMISD::VSHRu"; 953 case ARMISD::VSHLLs: return "ARMISD::VSHLLs"; 954 case ARMISD::VSHLLu: return "ARMISD::VSHLLu"; 955 case ARMISD::VSHLLi: return "ARMISD::VSHLLi"; 956 case ARMISD::VSHRN: return "ARMISD::VSHRN"; 957 case ARMISD::VRSHRs: return "ARMISD::VRSHRs"; 958 case ARMISD::VRSHRu: return "ARMISD::VRSHRu"; 959 case ARMISD::VRSHRN: return "ARMISD::VRSHRN"; 960 case ARMISD::VQSHLs: return "ARMISD::VQSHLs"; 961 case ARMISD::VQSHLu: return "ARMISD::VQSHLu"; 962 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu"; 963 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs"; 964 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu"; 965 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu"; 966 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs"; 967 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu"; 968 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu"; 969 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu"; 970 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs"; 971 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM"; 972 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM"; 973 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM"; 974 case ARMISD::VDUP: return "ARMISD::VDUP"; 975 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE"; 976 case ARMISD::VEXT: return "ARMISD::VEXT"; 977 case ARMISD::VREV64: return "ARMISD::VREV64"; 978 case ARMISD::VREV32: return "ARMISD::VREV32"; 979 case ARMISD::VREV16: return "ARMISD::VREV16"; 980 case ARMISD::VZIP: return "ARMISD::VZIP"; 981 case ARMISD::VUZP: return "ARMISD::VUZP"; 982 case ARMISD::VTRN: return "ARMISD::VTRN"; 983 case ARMISD::VTBL1: return "ARMISD::VTBL1"; 984 case ARMISD::VTBL2: return "ARMISD::VTBL2"; 985 case ARMISD::VMULLs: return "ARMISD::VMULLs"; 986 case ARMISD::VMULLu: return "ARMISD::VMULLu"; 987 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR"; 988 case ARMISD::FMAX: return "ARMISD::FMAX"; 989 case ARMISD::FMIN: return "ARMISD::FMIN"; 990 case ARMISD::BFI: return "ARMISD::BFI"; 991 case ARMISD::VORRIMM: return "ARMISD::VORRIMM"; 992 case ARMISD::VBICIMM: return "ARMISD::VBICIMM"; 993 case ARMISD::VBSL: return "ARMISD::VBSL"; 994 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP"; 995 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP"; 996 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP"; 997 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD"; 998 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD"; 999 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD"; 1000 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD"; 1001 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD"; 1002 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD"; 1003 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD"; 1004 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD"; 1005 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD"; 1006 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD"; 1007 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD"; 1008 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD"; 1009 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD"; 1010 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD"; 1011 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD"; 1012 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD"; 1013 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD"; 1014 } 1015 } 1016 1017 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const { 1018 if (!VT.isVector()) return getPointerTy(); 1019 return VT.changeVectorElementTypeToInteger(); 1020 } 1021 1022 /// getRegClassFor - Return the register class that should be used for the 1023 /// specified value type. 1024 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const { 1025 // Map v4i64 to QQ registers but do not make the type legal. Similarly map 1026 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to 1027 // load / store 4 to 8 consecutive D registers. 1028 if (Subtarget->hasNEON()) { 1029 if (VT == MVT::v4i64) 1030 return ARM::QQPRRegisterClass; 1031 else if (VT == MVT::v8i64) 1032 return ARM::QQQQPRRegisterClass; 1033 } 1034 return TargetLowering::getRegClassFor(VT); 1035 } 1036 1037 // Create a fast isel object. 1038 FastISel * 1039 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const { 1040 return ARM::createFastISel(funcInfo); 1041 } 1042 1043 /// getMaximalGlobalOffset - Returns the maximal possible offset which can 1044 /// be used for loads / stores from the global. 1045 unsigned ARMTargetLowering::getMaximalGlobalOffset() const { 1046 return (Subtarget->isThumb1Only() ? 127 : 4095); 1047 } 1048 1049 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const { 1050 unsigned NumVals = N->getNumValues(); 1051 if (!NumVals) 1052 return Sched::RegPressure; 1053 1054 for (unsigned i = 0; i != NumVals; ++i) { 1055 EVT VT = N->getValueType(i); 1056 if (VT == MVT::Glue || VT == MVT::Other) 1057 continue; 1058 if (VT.isFloatingPoint() || VT.isVector()) 1059 return Sched::ILP; 1060 } 1061 1062 if (!N->isMachineOpcode()) 1063 return Sched::RegPressure; 1064 1065 // Load are scheduled for latency even if there instruction itinerary 1066 // is not available. 1067 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1068 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode()); 1069 1070 if (MCID.getNumDefs() == 0) 1071 return Sched::RegPressure; 1072 if (!Itins->isEmpty() && 1073 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2) 1074 return Sched::ILP; 1075 1076 return Sched::RegPressure; 1077 } 1078 1079 //===----------------------------------------------------------------------===// 1080 // Lowering Code 1081 //===----------------------------------------------------------------------===// 1082 1083 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC 1084 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) { 1085 switch (CC) { 1086 default: llvm_unreachable("Unknown condition code!"); 1087 case ISD::SETNE: return ARMCC::NE; 1088 case ISD::SETEQ: return ARMCC::EQ; 1089 case ISD::SETGT: return ARMCC::GT; 1090 case ISD::SETGE: return ARMCC::GE; 1091 case ISD::SETLT: return ARMCC::LT; 1092 case ISD::SETLE: return ARMCC::LE; 1093 case ISD::SETUGT: return ARMCC::HI; 1094 case ISD::SETUGE: return ARMCC::HS; 1095 case ISD::SETULT: return ARMCC::LO; 1096 case ISD::SETULE: return ARMCC::LS; 1097 } 1098 } 1099 1100 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. 1101 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode, 1102 ARMCC::CondCodes &CondCode2) { 1103 CondCode2 = ARMCC::AL; 1104 switch (CC) { 1105 default: llvm_unreachable("Unknown FP condition!"); 1106 case ISD::SETEQ: 1107 case ISD::SETOEQ: CondCode = ARMCC::EQ; break; 1108 case ISD::SETGT: 1109 case ISD::SETOGT: CondCode = ARMCC::GT; break; 1110 case ISD::SETGE: 1111 case ISD::SETOGE: CondCode = ARMCC::GE; break; 1112 case ISD::SETOLT: CondCode = ARMCC::MI; break; 1113 case ISD::SETOLE: CondCode = ARMCC::LS; break; 1114 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break; 1115 case ISD::SETO: CondCode = ARMCC::VC; break; 1116 case ISD::SETUO: CondCode = ARMCC::VS; break; 1117 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break; 1118 case ISD::SETUGT: CondCode = ARMCC::HI; break; 1119 case ISD::SETUGE: CondCode = ARMCC::PL; break; 1120 case ISD::SETLT: 1121 case ISD::SETULT: CondCode = ARMCC::LT; break; 1122 case ISD::SETLE: 1123 case ISD::SETULE: CondCode = ARMCC::LE; break; 1124 case ISD::SETNE: 1125 case ISD::SETUNE: CondCode = ARMCC::NE; break; 1126 } 1127 } 1128 1129 //===----------------------------------------------------------------------===// 1130 // Calling Convention Implementation 1131 //===----------------------------------------------------------------------===// 1132 1133 #include "ARMGenCallingConv.inc" 1134 1135 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the 1136 /// given CallingConvention value. 1137 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC, 1138 bool Return, 1139 bool isVarArg) const { 1140 switch (CC) { 1141 default: 1142 llvm_unreachable("Unsupported calling convention"); 1143 case CallingConv::Fast: 1144 if (Subtarget->hasVFP2() && !isVarArg) { 1145 if (!Subtarget->isAAPCS_ABI()) 1146 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS); 1147 // For AAPCS ABI targets, just use VFP variant of the calling convention. 1148 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP); 1149 } 1150 // Fallthrough 1151 case CallingConv::C: { 1152 // Use target triple & subtarget features to do actual dispatch. 1153 if (!Subtarget->isAAPCS_ABI()) 1154 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS); 1155 else if (Subtarget->hasVFP2() && 1156 getTargetMachine().Options.FloatABIType == FloatABI::Hard && 1157 !isVarArg) 1158 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP); 1159 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS); 1160 } 1161 case CallingConv::ARM_AAPCS_VFP: 1162 if (!isVarArg) 1163 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP); 1164 // Fallthrough 1165 case CallingConv::ARM_AAPCS: 1166 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS); 1167 case CallingConv::ARM_APCS: 1168 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS); 1169 } 1170 } 1171 1172 /// LowerCallResult - Lower the result values of a call into the 1173 /// appropriate copies out of appropriate physical registers. 1174 SDValue 1175 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, 1176 CallingConv::ID CallConv, bool isVarArg, 1177 const SmallVectorImpl<ISD::InputArg> &Ins, 1178 DebugLoc dl, SelectionDAG &DAG, 1179 SmallVectorImpl<SDValue> &InVals) const { 1180 1181 // Assign locations to each value returned by this call. 1182 SmallVector<CCValAssign, 16> RVLocs; 1183 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1184 getTargetMachine(), RVLocs, *DAG.getContext(), Call); 1185 CCInfo.AnalyzeCallResult(Ins, 1186 CCAssignFnForNode(CallConv, /* Return*/ true, 1187 isVarArg)); 1188 1189 // Copy all of the result registers out of their specified physreg. 1190 for (unsigned i = 0; i != RVLocs.size(); ++i) { 1191 CCValAssign VA = RVLocs[i]; 1192 1193 SDValue Val; 1194 if (VA.needsCustom()) { 1195 // Handle f64 or half of a v2f64. 1196 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, 1197 InFlag); 1198 Chain = Lo.getValue(1); 1199 InFlag = Lo.getValue(2); 1200 VA = RVLocs[++i]; // skip ahead to next loc 1201 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, 1202 InFlag); 1203 Chain = Hi.getValue(1); 1204 InFlag = Hi.getValue(2); 1205 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi); 1206 1207 if (VA.getLocVT() == MVT::v2f64) { 1208 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64); 1209 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val, 1210 DAG.getConstant(0, MVT::i32)); 1211 1212 VA = RVLocs[++i]; // skip ahead to next loc 1213 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag); 1214 Chain = Lo.getValue(1); 1215 InFlag = Lo.getValue(2); 1216 VA = RVLocs[++i]; // skip ahead to next loc 1217 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag); 1218 Chain = Hi.getValue(1); 1219 InFlag = Hi.getValue(2); 1220 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi); 1221 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val, 1222 DAG.getConstant(1, MVT::i32)); 1223 } 1224 } else { 1225 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(), 1226 InFlag); 1227 Chain = Val.getValue(1); 1228 InFlag = Val.getValue(2); 1229 } 1230 1231 switch (VA.getLocInfo()) { 1232 default: llvm_unreachable("Unknown loc info!"); 1233 case CCValAssign::Full: break; 1234 case CCValAssign::BCvt: 1235 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val); 1236 break; 1237 } 1238 1239 InVals.push_back(Val); 1240 } 1241 1242 return Chain; 1243 } 1244 1245 /// LowerMemOpCallTo - Store the argument to the stack. 1246 SDValue 1247 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain, 1248 SDValue StackPtr, SDValue Arg, 1249 DebugLoc dl, SelectionDAG &DAG, 1250 const CCValAssign &VA, 1251 ISD::ArgFlagsTy Flags) const { 1252 unsigned LocMemOffset = VA.getLocMemOffset(); 1253 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset); 1254 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff); 1255 return DAG.getStore(Chain, dl, Arg, PtrOff, 1256 MachinePointerInfo::getStack(LocMemOffset), 1257 false, false, 0); 1258 } 1259 1260 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG, 1261 SDValue Chain, SDValue &Arg, 1262 RegsToPassVector &RegsToPass, 1263 CCValAssign &VA, CCValAssign &NextVA, 1264 SDValue &StackPtr, 1265 SmallVector<SDValue, 8> &MemOpChains, 1266 ISD::ArgFlagsTy Flags) const { 1267 1268 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl, 1269 DAG.getVTList(MVT::i32, MVT::i32), Arg); 1270 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd)); 1271 1272 if (NextVA.isRegLoc()) 1273 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1))); 1274 else { 1275 assert(NextVA.isMemLoc()); 1276 if (StackPtr.getNode() == 0) 1277 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 1278 1279 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1), 1280 dl, DAG, NextVA, 1281 Flags)); 1282 } 1283 } 1284 1285 /// LowerCall - Lowering a call into a callseq_start <- 1286 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter 1287 /// nodes. 1288 SDValue 1289 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee, 1290 CallingConv::ID CallConv, bool isVarArg, 1291 bool doesNotRet, bool &isTailCall, 1292 const SmallVectorImpl<ISD::OutputArg> &Outs, 1293 const SmallVectorImpl<SDValue> &OutVals, 1294 const SmallVectorImpl<ISD::InputArg> &Ins, 1295 DebugLoc dl, SelectionDAG &DAG, 1296 SmallVectorImpl<SDValue> &InVals) const { 1297 MachineFunction &MF = DAG.getMachineFunction(); 1298 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet(); 1299 bool IsSibCall = false; 1300 // Disable tail calls if they're not supported. 1301 if (!EnableARMTailCalls && !Subtarget->supportsTailCall()) 1302 isTailCall = false; 1303 if (isTailCall) { 1304 // Check if it's really possible to do a tail call. 1305 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, 1306 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(), 1307 Outs, OutVals, Ins, DAG); 1308 // We don't support GuaranteedTailCallOpt for ARM, only automatically 1309 // detected sibcalls. 1310 if (isTailCall) { 1311 ++NumTailCalls; 1312 IsSibCall = true; 1313 } 1314 } 1315 1316 // Analyze operands of the call, assigning locations to each operand. 1317 SmallVector<CCValAssign, 16> ArgLocs; 1318 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1319 getTargetMachine(), ArgLocs, *DAG.getContext(), Call); 1320 CCInfo.AnalyzeCallOperands(Outs, 1321 CCAssignFnForNode(CallConv, /* Return*/ false, 1322 isVarArg)); 1323 1324 // Get a count of how many bytes are to be pushed on the stack. 1325 unsigned NumBytes = CCInfo.getNextStackOffset(); 1326 1327 // For tail calls, memory operands are available in our caller's stack. 1328 if (IsSibCall) 1329 NumBytes = 0; 1330 1331 // Adjust the stack pointer for the new arguments... 1332 // These operations are automatically eliminated by the prolog/epilog pass 1333 if (!IsSibCall) 1334 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true)); 1335 1336 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 1337 1338 RegsToPassVector RegsToPass; 1339 SmallVector<SDValue, 8> MemOpChains; 1340 1341 // Walk the register/memloc assignments, inserting copies/loads. In the case 1342 // of tail call optimization, arguments are handled later. 1343 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); 1344 i != e; 1345 ++i, ++realArgIdx) { 1346 CCValAssign &VA = ArgLocs[i]; 1347 SDValue Arg = OutVals[realArgIdx]; 1348 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags; 1349 bool isByVal = Flags.isByVal(); 1350 1351 // Promote the value if needed. 1352 switch (VA.getLocInfo()) { 1353 default: llvm_unreachable("Unknown loc info!"); 1354 case CCValAssign::Full: break; 1355 case CCValAssign::SExt: 1356 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); 1357 break; 1358 case CCValAssign::ZExt: 1359 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); 1360 break; 1361 case CCValAssign::AExt: 1362 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg); 1363 break; 1364 case CCValAssign::BCvt: 1365 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg); 1366 break; 1367 } 1368 1369 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces 1370 if (VA.needsCustom()) { 1371 if (VA.getLocVT() == MVT::v2f64) { 1372 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1373 DAG.getConstant(0, MVT::i32)); 1374 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1375 DAG.getConstant(1, MVT::i32)); 1376 1377 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass, 1378 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags); 1379 1380 VA = ArgLocs[++i]; // skip ahead to next loc 1381 if (VA.isRegLoc()) { 1382 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass, 1383 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags); 1384 } else { 1385 assert(VA.isMemLoc()); 1386 1387 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1, 1388 dl, DAG, VA, Flags)); 1389 } 1390 } else { 1391 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i], 1392 StackPtr, MemOpChains, Flags); 1393 } 1394 } else if (VA.isRegLoc()) { 1395 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 1396 } else if (isByVal) { 1397 assert(VA.isMemLoc()); 1398 unsigned offset = 0; 1399 1400 // True if this byval aggregate will be split between registers 1401 // and memory. 1402 if (CCInfo.isFirstByValRegValid()) { 1403 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 1404 unsigned int i, j; 1405 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) { 1406 SDValue Const = DAG.getConstant(4*i, MVT::i32); 1407 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const); 1408 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg, 1409 MachinePointerInfo(), 1410 false, false, false, 0); 1411 MemOpChains.push_back(Load.getValue(1)); 1412 RegsToPass.push_back(std::make_pair(j, Load)); 1413 } 1414 offset = ARM::R4 - CCInfo.getFirstByValReg(); 1415 CCInfo.clearFirstByValReg(); 1416 } 1417 1418 unsigned LocMemOffset = VA.getLocMemOffset(); 1419 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset); 1420 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, 1421 StkPtrOff); 1422 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset); 1423 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset); 1424 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, 1425 MVT::i32); 1426 MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, 1427 Flags.getByValAlign(), 1428 /*isVolatile=*/false, 1429 /*AlwaysInline=*/false, 1430 MachinePointerInfo(0), 1431 MachinePointerInfo(0))); 1432 1433 } else if (!IsSibCall) { 1434 assert(VA.isMemLoc()); 1435 1436 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg, 1437 dl, DAG, VA, Flags)); 1438 } 1439 } 1440 1441 if (!MemOpChains.empty()) 1442 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 1443 &MemOpChains[0], MemOpChains.size()); 1444 1445 // Build a sequence of copy-to-reg nodes chained together with token chain 1446 // and flag operands which copy the outgoing args into the appropriate regs. 1447 SDValue InFlag; 1448 // Tail call byval lowering might overwrite argument registers so in case of 1449 // tail call optimization the copies to registers are lowered later. 1450 if (!isTailCall) 1451 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 1452 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, 1453 RegsToPass[i].second, InFlag); 1454 InFlag = Chain.getValue(1); 1455 } 1456 1457 // For tail calls lower the arguments to the 'real' stack slot. 1458 if (isTailCall) { 1459 // Force all the incoming stack arguments to be loaded from the stack 1460 // before any new outgoing arguments are stored to the stack, because the 1461 // outgoing stack slots may alias the incoming argument stack slots, and 1462 // the alias isn't otherwise explicit. This is slightly more conservative 1463 // than necessary, because it means that each store effectively depends 1464 // on every argument instead of just those arguments it would clobber. 1465 1466 // Do not flag preceding copytoreg stuff together with the following stuff. 1467 InFlag = SDValue(); 1468 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 1469 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, 1470 RegsToPass[i].second, InFlag); 1471 InFlag = Chain.getValue(1); 1472 } 1473 InFlag =SDValue(); 1474 } 1475 1476 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 1477 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 1478 // node so that legalize doesn't hack it. 1479 bool isDirect = false; 1480 bool isARMFunc = false; 1481 bool isLocalARMFunc = false; 1482 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1483 1484 if (EnableARMLongCalls) { 1485 assert (getTargetMachine().getRelocationModel() == Reloc::Static 1486 && "long-calls with non-static relocation model!"); 1487 // Handle a global address or an external symbol. If it's not one of 1488 // those, the target's already in a register, so we don't need to do 1489 // anything extra. 1490 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 1491 const GlobalValue *GV = G->getGlobal(); 1492 // Create a constant pool entry for the callee address 1493 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1494 ARMConstantPoolValue *CPV = 1495 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0); 1496 1497 // Get the address of the callee into a register 1498 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1499 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1500 Callee = DAG.getLoad(getPointerTy(), dl, 1501 DAG.getEntryNode(), CPAddr, 1502 MachinePointerInfo::getConstantPool(), 1503 false, false, false, 0); 1504 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) { 1505 const char *Sym = S->getSymbol(); 1506 1507 // Create a constant pool entry for the callee address 1508 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1509 ARMConstantPoolValue *CPV = 1510 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym, 1511 ARMPCLabelIndex, 0); 1512 // Get the address of the callee into a register 1513 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1514 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1515 Callee = DAG.getLoad(getPointerTy(), dl, 1516 DAG.getEntryNode(), CPAddr, 1517 MachinePointerInfo::getConstantPool(), 1518 false, false, false, 0); 1519 } 1520 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 1521 const GlobalValue *GV = G->getGlobal(); 1522 isDirect = true; 1523 bool isExt = GV->isDeclaration() || GV->isWeakForLinker(); 1524 bool isStub = (isExt && Subtarget->isTargetDarwin()) && 1525 getTargetMachine().getRelocationModel() != Reloc::Static; 1526 isARMFunc = !Subtarget->isThumb() || isStub; 1527 // ARM call to a local ARM function is predicable. 1528 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking); 1529 // tBX takes a register source operand. 1530 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) { 1531 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1532 ARMConstantPoolValue *CPV = 1533 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4); 1534 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1535 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1536 Callee = DAG.getLoad(getPointerTy(), dl, 1537 DAG.getEntryNode(), CPAddr, 1538 MachinePointerInfo::getConstantPool(), 1539 false, false, false, 0); 1540 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 1541 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, 1542 getPointerTy(), Callee, PICLabel); 1543 } else { 1544 // On ELF targets for PIC code, direct calls should go through the PLT 1545 unsigned OpFlags = 0; 1546 if (Subtarget->isTargetELF() && 1547 getTargetMachine().getRelocationModel() == Reloc::PIC_) 1548 OpFlags = ARMII::MO_PLT; 1549 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags); 1550 } 1551 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 1552 isDirect = true; 1553 bool isStub = Subtarget->isTargetDarwin() && 1554 getTargetMachine().getRelocationModel() != Reloc::Static; 1555 isARMFunc = !Subtarget->isThumb() || isStub; 1556 // tBX takes a register source operand. 1557 const char *Sym = S->getSymbol(); 1558 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) { 1559 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1560 ARMConstantPoolValue *CPV = 1561 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym, 1562 ARMPCLabelIndex, 4); 1563 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1564 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1565 Callee = DAG.getLoad(getPointerTy(), dl, 1566 DAG.getEntryNode(), CPAddr, 1567 MachinePointerInfo::getConstantPool(), 1568 false, false, false, 0); 1569 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 1570 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, 1571 getPointerTy(), Callee, PICLabel); 1572 } else { 1573 unsigned OpFlags = 0; 1574 // On ELF targets for PIC code, direct calls should go through the PLT 1575 if (Subtarget->isTargetELF() && 1576 getTargetMachine().getRelocationModel() == Reloc::PIC_) 1577 OpFlags = ARMII::MO_PLT; 1578 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags); 1579 } 1580 } 1581 1582 // FIXME: handle tail calls differently. 1583 unsigned CallOpc; 1584 if (Subtarget->isThumb()) { 1585 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps()) 1586 CallOpc = ARMISD::CALL_NOLINK; 1587 else if (doesNotRet && isDirect && !isARMFunc && 1588 Subtarget->hasRAS() && !Subtarget->isThumb1Only()) 1589 // "mov lr, pc; b _foo" to avoid confusing the RSP 1590 CallOpc = ARMISD::CALL_NOLINK; 1591 else 1592 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL; 1593 } else { 1594 if (!isDirect && !Subtarget->hasV5TOps()) { 1595 CallOpc = ARMISD::CALL_NOLINK; 1596 } else if (doesNotRet && isDirect && Subtarget->hasRAS()) 1597 // "mov lr, pc; b _foo" to avoid confusing the RSP 1598 CallOpc = ARMISD::CALL_NOLINK; 1599 else 1600 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL; 1601 } 1602 1603 std::vector<SDValue> Ops; 1604 Ops.push_back(Chain); 1605 Ops.push_back(Callee); 1606 1607 // Add argument registers to the end of the list so that they are known live 1608 // into the call. 1609 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 1610 Ops.push_back(DAG.getRegister(RegsToPass[i].first, 1611 RegsToPass[i].second.getValueType())); 1612 1613 // Add a register mask operand representing the call-preserved registers. 1614 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo(); 1615 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv); 1616 assert(Mask && "Missing call preserved mask for calling convention"); 1617 Ops.push_back(DAG.getRegisterMask(Mask)); 1618 1619 if (InFlag.getNode()) 1620 Ops.push_back(InFlag); 1621 1622 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 1623 if (isTailCall) 1624 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size()); 1625 1626 // Returns a chain and a flag for retval copy to use. 1627 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size()); 1628 InFlag = Chain.getValue(1); 1629 1630 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true), 1631 DAG.getIntPtrConstant(0, true), InFlag); 1632 if (!Ins.empty()) 1633 InFlag = Chain.getValue(1); 1634 1635 // Handle result values, copying them out of physregs into vregs that we 1636 // return. 1637 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, 1638 dl, DAG, InVals); 1639 } 1640 1641 /// HandleByVal - Every parameter *after* a byval parameter is passed 1642 /// on the stack. Remember the next parameter register to allocate, 1643 /// and then confiscate the rest of the parameter registers to insure 1644 /// this. 1645 void 1646 ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const { 1647 unsigned reg = State->AllocateReg(GPRArgRegs, 4); 1648 assert((State->getCallOrPrologue() == Prologue || 1649 State->getCallOrPrologue() == Call) && 1650 "unhandled ParmContext"); 1651 if ((!State->isFirstByValRegValid()) && 1652 (ARM::R0 <= reg) && (reg <= ARM::R3)) { 1653 State->setFirstByValReg(reg); 1654 // At a call site, a byval parameter that is split between 1655 // registers and memory needs its size truncated here. In a 1656 // function prologue, such byval parameters are reassembled in 1657 // memory, and are not truncated. 1658 if (State->getCallOrPrologue() == Call) { 1659 unsigned excess = 4 * (ARM::R4 - reg); 1660 assert(size >= excess && "expected larger existing stack allocation"); 1661 size -= excess; 1662 } 1663 } 1664 // Confiscate any remaining parameter registers to preclude their 1665 // assignment to subsequent parameters. 1666 while (State->AllocateReg(GPRArgRegs, 4)) 1667 ; 1668 } 1669 1670 /// MatchingStackOffset - Return true if the given stack call argument is 1671 /// already available in the same position (relatively) of the caller's 1672 /// incoming argument stack. 1673 static 1674 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags, 1675 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI, 1676 const TargetInstrInfo *TII) { 1677 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8; 1678 int FI = INT_MAX; 1679 if (Arg.getOpcode() == ISD::CopyFromReg) { 1680 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg(); 1681 if (!TargetRegisterInfo::isVirtualRegister(VR)) 1682 return false; 1683 MachineInstr *Def = MRI->getVRegDef(VR); 1684 if (!Def) 1685 return false; 1686 if (!Flags.isByVal()) { 1687 if (!TII->isLoadFromStackSlot(Def, FI)) 1688 return false; 1689 } else { 1690 return false; 1691 } 1692 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) { 1693 if (Flags.isByVal()) 1694 // ByVal argument is passed in as a pointer but it's now being 1695 // dereferenced. e.g. 1696 // define @foo(%struct.X* %A) { 1697 // tail call @bar(%struct.X* byval %A) 1698 // } 1699 return false; 1700 SDValue Ptr = Ld->getBasePtr(); 1701 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr); 1702 if (!FINode) 1703 return false; 1704 FI = FINode->getIndex(); 1705 } else 1706 return false; 1707 1708 assert(FI != INT_MAX); 1709 if (!MFI->isFixedObjectIndex(FI)) 1710 return false; 1711 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI); 1712 } 1713 1714 /// IsEligibleForTailCallOptimization - Check whether the call is eligible 1715 /// for tail call optimization. Targets which want to do tail call 1716 /// optimization should implement this function. 1717 bool 1718 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee, 1719 CallingConv::ID CalleeCC, 1720 bool isVarArg, 1721 bool isCalleeStructRet, 1722 bool isCallerStructRet, 1723 const SmallVectorImpl<ISD::OutputArg> &Outs, 1724 const SmallVectorImpl<SDValue> &OutVals, 1725 const SmallVectorImpl<ISD::InputArg> &Ins, 1726 SelectionDAG& DAG) const { 1727 const Function *CallerF = DAG.getMachineFunction().getFunction(); 1728 CallingConv::ID CallerCC = CallerF->getCallingConv(); 1729 bool CCMatch = CallerCC == CalleeCC; 1730 1731 // Look for obvious safe cases to perform tail call optimization that do not 1732 // require ABI changes. This is what gcc calls sibcall. 1733 1734 // Do not sibcall optimize vararg calls unless the call site is not passing 1735 // any arguments. 1736 if (isVarArg && !Outs.empty()) 1737 return false; 1738 1739 // Also avoid sibcall optimization if either caller or callee uses struct 1740 // return semantics. 1741 if (isCalleeStructRet || isCallerStructRet) 1742 return false; 1743 1744 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo:: 1745 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as 1746 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation 1747 // support in the assembler and linker to be used. This would need to be 1748 // fixed to fully support tail calls in Thumb1. 1749 // 1750 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take 1751 // LR. This means if we need to reload LR, it takes an extra instructions, 1752 // which outweighs the value of the tail call; but here we don't know yet 1753 // whether LR is going to be used. Probably the right approach is to 1754 // generate the tail call here and turn it back into CALL/RET in 1755 // emitEpilogue if LR is used. 1756 1757 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls, 1758 // but we need to make sure there are enough registers; the only valid 1759 // registers are the 4 used for parameters. We don't currently do this 1760 // case. 1761 if (Subtarget->isThumb1Only()) 1762 return false; 1763 1764 // If the calling conventions do not match, then we'd better make sure the 1765 // results are returned in the same way as what the caller expects. 1766 if (!CCMatch) { 1767 SmallVector<CCValAssign, 16> RVLocs1; 1768 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), 1769 getTargetMachine(), RVLocs1, *DAG.getContext(), Call); 1770 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg)); 1771 1772 SmallVector<CCValAssign, 16> RVLocs2; 1773 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), 1774 getTargetMachine(), RVLocs2, *DAG.getContext(), Call); 1775 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg)); 1776 1777 if (RVLocs1.size() != RVLocs2.size()) 1778 return false; 1779 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) { 1780 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc()) 1781 return false; 1782 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo()) 1783 return false; 1784 if (RVLocs1[i].isRegLoc()) { 1785 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg()) 1786 return false; 1787 } else { 1788 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset()) 1789 return false; 1790 } 1791 } 1792 } 1793 1794 // If the callee takes no arguments then go on to check the results of the 1795 // call. 1796 if (!Outs.empty()) { 1797 // Check if stack adjustment is needed. For now, do not do this if any 1798 // argument is passed on the stack. 1799 SmallVector<CCValAssign, 16> ArgLocs; 1800 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), 1801 getTargetMachine(), ArgLocs, *DAG.getContext(), Call); 1802 CCInfo.AnalyzeCallOperands(Outs, 1803 CCAssignFnForNode(CalleeCC, false, isVarArg)); 1804 if (CCInfo.getNextStackOffset()) { 1805 MachineFunction &MF = DAG.getMachineFunction(); 1806 1807 // Check if the arguments are already laid out in the right way as 1808 // the caller's fixed stack objects. 1809 MachineFrameInfo *MFI = MF.getFrameInfo(); 1810 const MachineRegisterInfo *MRI = &MF.getRegInfo(); 1811 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1812 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); 1813 i != e; 1814 ++i, ++realArgIdx) { 1815 CCValAssign &VA = ArgLocs[i]; 1816 EVT RegVT = VA.getLocVT(); 1817 SDValue Arg = OutVals[realArgIdx]; 1818 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags; 1819 if (VA.getLocInfo() == CCValAssign::Indirect) 1820 return false; 1821 if (VA.needsCustom()) { 1822 // f64 and vector types are split into multiple registers or 1823 // register/stack-slot combinations. The types will not match 1824 // the registers; give up on memory f64 refs until we figure 1825 // out what to do about this. 1826 if (!VA.isRegLoc()) 1827 return false; 1828 if (!ArgLocs[++i].isRegLoc()) 1829 return false; 1830 if (RegVT == MVT::v2f64) { 1831 if (!ArgLocs[++i].isRegLoc()) 1832 return false; 1833 if (!ArgLocs[++i].isRegLoc()) 1834 return false; 1835 } 1836 } else if (!VA.isRegLoc()) { 1837 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags, 1838 MFI, MRI, TII)) 1839 return false; 1840 } 1841 } 1842 } 1843 } 1844 1845 return true; 1846 } 1847 1848 SDValue 1849 ARMTargetLowering::LowerReturn(SDValue Chain, 1850 CallingConv::ID CallConv, bool isVarArg, 1851 const SmallVectorImpl<ISD::OutputArg> &Outs, 1852 const SmallVectorImpl<SDValue> &OutVals, 1853 DebugLoc dl, SelectionDAG &DAG) const { 1854 1855 // CCValAssign - represent the assignment of the return value to a location. 1856 SmallVector<CCValAssign, 16> RVLocs; 1857 1858 // CCState - Info about the registers and stack slots. 1859 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1860 getTargetMachine(), RVLocs, *DAG.getContext(), Call); 1861 1862 // Analyze outgoing return values. 1863 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true, 1864 isVarArg)); 1865 1866 // If this is the first return lowered for this function, add 1867 // the regs to the liveout set for the function. 1868 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) { 1869 for (unsigned i = 0; i != RVLocs.size(); ++i) 1870 if (RVLocs[i].isRegLoc()) 1871 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg()); 1872 } 1873 1874 SDValue Flag; 1875 1876 // Copy the result values into the output registers. 1877 for (unsigned i = 0, realRVLocIdx = 0; 1878 i != RVLocs.size(); 1879 ++i, ++realRVLocIdx) { 1880 CCValAssign &VA = RVLocs[i]; 1881 assert(VA.isRegLoc() && "Can only return in registers!"); 1882 1883 SDValue Arg = OutVals[realRVLocIdx]; 1884 1885 switch (VA.getLocInfo()) { 1886 default: llvm_unreachable("Unknown loc info!"); 1887 case CCValAssign::Full: break; 1888 case CCValAssign::BCvt: 1889 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg); 1890 break; 1891 } 1892 1893 if (VA.needsCustom()) { 1894 if (VA.getLocVT() == MVT::v2f64) { 1895 // Extract the first half and return it in two registers. 1896 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1897 DAG.getConstant(0, MVT::i32)); 1898 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl, 1899 DAG.getVTList(MVT::i32, MVT::i32), Half); 1900 1901 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag); 1902 Flag = Chain.getValue(1); 1903 VA = RVLocs[++i]; // skip ahead to next loc 1904 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), 1905 HalfGPRs.getValue(1), Flag); 1906 Flag = Chain.getValue(1); 1907 VA = RVLocs[++i]; // skip ahead to next loc 1908 1909 // Extract the 2nd half and fall through to handle it as an f64 value. 1910 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1911 DAG.getConstant(1, MVT::i32)); 1912 } 1913 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is 1914 // available. 1915 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl, 1916 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1); 1917 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag); 1918 Flag = Chain.getValue(1); 1919 VA = RVLocs[++i]; // skip ahead to next loc 1920 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1), 1921 Flag); 1922 } else 1923 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag); 1924 1925 // Guarantee that all emitted copies are 1926 // stuck together, avoiding something bad. 1927 Flag = Chain.getValue(1); 1928 } 1929 1930 SDValue result; 1931 if (Flag.getNode()) 1932 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag); 1933 else // Return Void 1934 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain); 1935 1936 return result; 1937 } 1938 1939 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const { 1940 if (N->getNumValues() != 1) 1941 return false; 1942 if (!N->hasNUsesOfValue(1, 0)) 1943 return false; 1944 1945 SDValue TCChain = Chain; 1946 SDNode *Copy = *N->use_begin(); 1947 if (Copy->getOpcode() == ISD::CopyToReg) { 1948 // If the copy has a glue operand, we conservatively assume it isn't safe to 1949 // perform a tail call. 1950 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue) 1951 return false; 1952 TCChain = Copy->getOperand(0); 1953 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) { 1954 SDNode *VMov = Copy; 1955 // f64 returned in a pair of GPRs. 1956 SmallPtrSet<SDNode*, 2> Copies; 1957 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end(); 1958 UI != UE; ++UI) { 1959 if (UI->getOpcode() != ISD::CopyToReg) 1960 return false; 1961 Copies.insert(*UI); 1962 } 1963 if (Copies.size() > 2) 1964 return false; 1965 1966 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end(); 1967 UI != UE; ++UI) { 1968 SDValue UseChain = UI->getOperand(0); 1969 if (Copies.count(UseChain.getNode())) 1970 // Second CopyToReg 1971 Copy = *UI; 1972 else 1973 // First CopyToReg 1974 TCChain = UseChain; 1975 } 1976 } else if (Copy->getOpcode() == ISD::BITCAST) { 1977 // f32 returned in a single GPR. 1978 if (!Copy->hasOneUse()) 1979 return false; 1980 Copy = *Copy->use_begin(); 1981 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0)) 1982 return false; 1983 Chain = Copy->getOperand(0); 1984 } else { 1985 return false; 1986 } 1987 1988 bool HasRet = false; 1989 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end(); 1990 UI != UE; ++UI) { 1991 if (UI->getOpcode() != ARMISD::RET_FLAG) 1992 return false; 1993 HasRet = true; 1994 } 1995 1996 if (!HasRet) 1997 return false; 1998 1999 Chain = TCChain; 2000 return true; 2001 } 2002 2003 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const { 2004 if (!EnableARMTailCalls && !Subtarget->supportsTailCall()) 2005 return false; 2006 2007 if (!CI->isTailCall()) 2008 return false; 2009 2010 return !Subtarget->isThumb1Only(); 2011 } 2012 2013 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as 2014 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is 2015 // one of the above mentioned nodes. It has to be wrapped because otherwise 2016 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only 2017 // be used to form addressing mode. These wrapped nodes will be selected 2018 // into MOVi. 2019 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) { 2020 EVT PtrVT = Op.getValueType(); 2021 // FIXME there is no actual debug info here 2022 DebugLoc dl = Op.getDebugLoc(); 2023 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); 2024 SDValue Res; 2025 if (CP->isMachineConstantPoolEntry()) 2026 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT, 2027 CP->getAlignment()); 2028 else 2029 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, 2030 CP->getAlignment()); 2031 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res); 2032 } 2033 2034 unsigned ARMTargetLowering::getJumpTableEncoding() const { 2035 return MachineJumpTableInfo::EK_Inline; 2036 } 2037 2038 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, 2039 SelectionDAG &DAG) const { 2040 MachineFunction &MF = DAG.getMachineFunction(); 2041 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2042 unsigned ARMPCLabelIndex = 0; 2043 DebugLoc DL = Op.getDebugLoc(); 2044 EVT PtrVT = getPointerTy(); 2045 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress(); 2046 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2047 SDValue CPAddr; 2048 if (RelocM == Reloc::Static) { 2049 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4); 2050 } else { 2051 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8; 2052 ARMPCLabelIndex = AFI->createPICLabelUId(); 2053 ARMConstantPoolValue *CPV = 2054 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex, 2055 ARMCP::CPBlockAddress, PCAdj); 2056 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2057 } 2058 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr); 2059 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr, 2060 MachinePointerInfo::getConstantPool(), 2061 false, false, false, 0); 2062 if (RelocM == Reloc::Static) 2063 return Result; 2064 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2065 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel); 2066 } 2067 2068 // Lower ISD::GlobalTLSAddress using the "general dynamic" model 2069 SDValue 2070 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, 2071 SelectionDAG &DAG) const { 2072 DebugLoc dl = GA->getDebugLoc(); 2073 EVT PtrVT = getPointerTy(); 2074 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8; 2075 MachineFunction &MF = DAG.getMachineFunction(); 2076 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2077 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2078 ARMConstantPoolValue *CPV = 2079 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex, 2080 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true); 2081 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2082 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument); 2083 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, 2084 MachinePointerInfo::getConstantPool(), 2085 false, false, false, 0); 2086 SDValue Chain = Argument.getValue(1); 2087 2088 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2089 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel); 2090 2091 // call __tls_get_addr. 2092 ArgListTy Args; 2093 ArgListEntry Entry; 2094 Entry.Node = Argument; 2095 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext()); 2096 Args.push_back(Entry); 2097 // FIXME: is there useful debug info available here? 2098 std::pair<SDValue, SDValue> CallResult = 2099 LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()), 2100 false, false, false, false, 2101 0, CallingConv::C, /*isTailCall=*/false, 2102 /*doesNotRet=*/false, /*isReturnValueUsed=*/true, 2103 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl); 2104 return CallResult.first; 2105 } 2106 2107 // Lower ISD::GlobalTLSAddress using the "initial exec" or 2108 // "local exec" model. 2109 SDValue 2110 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA, 2111 SelectionDAG &DAG) const { 2112 const GlobalValue *GV = GA->getGlobal(); 2113 DebugLoc dl = GA->getDebugLoc(); 2114 SDValue Offset; 2115 SDValue Chain = DAG.getEntryNode(); 2116 EVT PtrVT = getPointerTy(); 2117 // Get the Thread Pointer 2118 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT); 2119 2120 if (GV->isDeclaration()) { 2121 MachineFunction &MF = DAG.getMachineFunction(); 2122 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2123 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2124 // Initial exec model. 2125 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8; 2126 ARMConstantPoolValue *CPV = 2127 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex, 2128 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, 2129 true); 2130 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2131 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset); 2132 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, 2133 MachinePointerInfo::getConstantPool(), 2134 false, false, false, 0); 2135 Chain = Offset.getValue(1); 2136 2137 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2138 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel); 2139 2140 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, 2141 MachinePointerInfo::getConstantPool(), 2142 false, false, false, 0); 2143 } else { 2144 // local exec model 2145 ARMConstantPoolValue *CPV = 2146 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF); 2147 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2148 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset); 2149 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, 2150 MachinePointerInfo::getConstantPool(), 2151 false, false, false, 0); 2152 } 2153 2154 // The address of the thread local variable is the add of the thread 2155 // pointer with the offset of the variable. 2156 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset); 2157 } 2158 2159 SDValue 2160 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const { 2161 // TODO: implement the "local dynamic" model 2162 assert(Subtarget->isTargetELF() && 2163 "TLS not implemented for non-ELF targets"); 2164 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 2165 // If the relocation model is PIC, use the "General Dynamic" TLS Model, 2166 // otherwise use the "Local Exec" TLS Model 2167 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) 2168 return LowerToTLSGeneralDynamicModel(GA, DAG); 2169 else 2170 return LowerToTLSExecModels(GA, DAG); 2171 } 2172 2173 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op, 2174 SelectionDAG &DAG) const { 2175 EVT PtrVT = getPointerTy(); 2176 DebugLoc dl = Op.getDebugLoc(); 2177 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 2178 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2179 if (RelocM == Reloc::PIC_) { 2180 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility(); 2181 ARMConstantPoolValue *CPV = 2182 ARMConstantPoolConstant::Create(GV, 2183 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT); 2184 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2185 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2186 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), 2187 CPAddr, 2188 MachinePointerInfo::getConstantPool(), 2189 false, false, false, 0); 2190 SDValue Chain = Result.getValue(1); 2191 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT); 2192 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT); 2193 if (!UseGOTOFF) 2194 Result = DAG.getLoad(PtrVT, dl, Chain, Result, 2195 MachinePointerInfo::getGOT(), 2196 false, false, false, 0); 2197 return Result; 2198 } 2199 2200 // If we have T2 ops, we can materialize the address directly via movt/movw 2201 // pair. This is always cheaper. 2202 if (Subtarget->useMovt()) { 2203 ++NumMovwMovt; 2204 // FIXME: Once remat is capable of dealing with instructions with register 2205 // operands, expand this into two nodes. 2206 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT, 2207 DAG.getTargetGlobalAddress(GV, dl, PtrVT)); 2208 } else { 2209 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4); 2210 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2211 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2212 MachinePointerInfo::getConstantPool(), 2213 false, false, false, 0); 2214 } 2215 } 2216 2217 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op, 2218 SelectionDAG &DAG) const { 2219 EVT PtrVT = getPointerTy(); 2220 DebugLoc dl = Op.getDebugLoc(); 2221 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 2222 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2223 MachineFunction &MF = DAG.getMachineFunction(); 2224 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2225 2226 // FIXME: Enable this for static codegen when tool issues are fixed. Also 2227 // update ARMFastISel::ARMMaterializeGV. 2228 if (Subtarget->useMovt() && RelocM != Reloc::Static) { 2229 ++NumMovwMovt; 2230 // FIXME: Once remat is capable of dealing with instructions with register 2231 // operands, expand this into two nodes. 2232 if (RelocM == Reloc::Static) 2233 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT, 2234 DAG.getTargetGlobalAddress(GV, dl, PtrVT)); 2235 2236 unsigned Wrapper = (RelocM == Reloc::PIC_) 2237 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN; 2238 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, 2239 DAG.getTargetGlobalAddress(GV, dl, PtrVT)); 2240 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) 2241 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result, 2242 MachinePointerInfo::getGOT(), 2243 false, false, false, 0); 2244 return Result; 2245 } 2246 2247 unsigned ARMPCLabelIndex = 0; 2248 SDValue CPAddr; 2249 if (RelocM == Reloc::Static) { 2250 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4); 2251 } else { 2252 ARMPCLabelIndex = AFI->createPICLabelUId(); 2253 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8); 2254 ARMConstantPoolValue *CPV = 2255 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 2256 PCAdj); 2257 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2258 } 2259 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2260 2261 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2262 MachinePointerInfo::getConstantPool(), 2263 false, false, false, 0); 2264 SDValue Chain = Result.getValue(1); 2265 2266 if (RelocM == Reloc::PIC_) { 2267 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2268 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 2269 } 2270 2271 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) 2272 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(), 2273 false, false, false, 0); 2274 2275 return Result; 2276 } 2277 2278 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op, 2279 SelectionDAG &DAG) const { 2280 assert(Subtarget->isTargetELF() && 2281 "GLOBAL OFFSET TABLE not implemented for non-ELF targets"); 2282 MachineFunction &MF = DAG.getMachineFunction(); 2283 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2284 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2285 EVT PtrVT = getPointerTy(); 2286 DebugLoc dl = Op.getDebugLoc(); 2287 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8; 2288 ARMConstantPoolValue *CPV = 2289 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_", 2290 ARMPCLabelIndex, PCAdj); 2291 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2292 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2293 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2294 MachinePointerInfo::getConstantPool(), 2295 false, false, false, 0); 2296 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2297 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 2298 } 2299 2300 SDValue 2301 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const { 2302 DebugLoc dl = Op.getDebugLoc(); 2303 SDValue Val = DAG.getConstant(0, MVT::i32); 2304 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, 2305 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0), 2306 Op.getOperand(1), Val); 2307 } 2308 2309 SDValue 2310 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const { 2311 DebugLoc dl = Op.getDebugLoc(); 2312 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0), 2313 Op.getOperand(1), DAG.getConstant(0, MVT::i32)); 2314 } 2315 2316 SDValue 2317 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG, 2318 const ARMSubtarget *Subtarget) const { 2319 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 2320 DebugLoc dl = Op.getDebugLoc(); 2321 switch (IntNo) { 2322 default: return SDValue(); // Don't custom lower most intrinsics. 2323 case Intrinsic::arm_thread_pointer: { 2324 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 2325 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT); 2326 } 2327 case Intrinsic::eh_sjlj_lsda: { 2328 MachineFunction &MF = DAG.getMachineFunction(); 2329 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2330 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2331 EVT PtrVT = getPointerTy(); 2332 DebugLoc dl = Op.getDebugLoc(); 2333 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2334 SDValue CPAddr; 2335 unsigned PCAdj = (RelocM != Reloc::PIC_) 2336 ? 0 : (Subtarget->isThumb() ? 4 : 8); 2337 ARMConstantPoolValue *CPV = 2338 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex, 2339 ARMCP::CPLSDA, PCAdj); 2340 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2341 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2342 SDValue Result = 2343 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2344 MachinePointerInfo::getConstantPool(), 2345 false, false, false, 0); 2346 2347 if (RelocM == Reloc::PIC_) { 2348 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2349 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 2350 } 2351 return Result; 2352 } 2353 case Intrinsic::arm_neon_vmulls: 2354 case Intrinsic::arm_neon_vmullu: { 2355 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls) 2356 ? ARMISD::VMULLs : ARMISD::VMULLu; 2357 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(), 2358 Op.getOperand(1), Op.getOperand(2)); 2359 } 2360 } 2361 } 2362 2363 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG, 2364 const ARMSubtarget *Subtarget) { 2365 DebugLoc dl = Op.getDebugLoc(); 2366 if (!Subtarget->hasDataBarrier()) { 2367 // Some ARMv6 cpus can support data barriers with an mcr instruction. 2368 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get 2369 // here. 2370 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() && 2371 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!"); 2372 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0), 2373 DAG.getConstant(0, MVT::i32)); 2374 } 2375 2376 SDValue Op5 = Op.getOperand(5); 2377 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0; 2378 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); 2379 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue(); 2380 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0); 2381 2382 ARM_MB::MemBOpt DMBOpt; 2383 if (isDeviceBarrier) 2384 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY; 2385 else 2386 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH; 2387 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0), 2388 DAG.getConstant(DMBOpt, MVT::i32)); 2389 } 2390 2391 2392 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG, 2393 const ARMSubtarget *Subtarget) { 2394 // FIXME: handle "fence singlethread" more efficiently. 2395 DebugLoc dl = Op.getDebugLoc(); 2396 if (!Subtarget->hasDataBarrier()) { 2397 // Some ARMv6 cpus can support data barriers with an mcr instruction. 2398 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get 2399 // here. 2400 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() && 2401 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!"); 2402 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0), 2403 DAG.getConstant(0, MVT::i32)); 2404 } 2405 2406 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0), 2407 DAG.getConstant(ARM_MB::ISH, MVT::i32)); 2408 } 2409 2410 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG, 2411 const ARMSubtarget *Subtarget) { 2412 // ARM pre v5TE and Thumb1 does not have preload instructions. 2413 if (!(Subtarget->isThumb2() || 2414 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps()))) 2415 // Just preserve the chain. 2416 return Op.getOperand(0); 2417 2418 DebugLoc dl = Op.getDebugLoc(); 2419 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1; 2420 if (!isRead && 2421 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension())) 2422 // ARMv7 with MP extension has PLDW. 2423 return Op.getOperand(0); 2424 2425 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue(); 2426 if (Subtarget->isThumb()) { 2427 // Invert the bits. 2428 isRead = ~isRead & 1; 2429 isData = ~isData & 1; 2430 } 2431 2432 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0), 2433 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32), 2434 DAG.getConstant(isData, MVT::i32)); 2435 } 2436 2437 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) { 2438 MachineFunction &MF = DAG.getMachineFunction(); 2439 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>(); 2440 2441 // vastart just stores the address of the VarArgsFrameIndex slot into the 2442 // memory location argument. 2443 DebugLoc dl = Op.getDebugLoc(); 2444 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 2445 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); 2446 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 2447 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), 2448 MachinePointerInfo(SV), false, false, 0); 2449 } 2450 2451 SDValue 2452 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA, 2453 SDValue &Root, SelectionDAG &DAG, 2454 DebugLoc dl) const { 2455 MachineFunction &MF = DAG.getMachineFunction(); 2456 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2457 2458 const TargetRegisterClass *RC; 2459 if (AFI->isThumb1OnlyFunction()) 2460 RC = ARM::tGPRRegisterClass; 2461 else 2462 RC = ARM::GPRRegisterClass; 2463 2464 // Transform the arguments stored in physical registers into virtual ones. 2465 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); 2466 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32); 2467 2468 SDValue ArgValue2; 2469 if (NextVA.isMemLoc()) { 2470 MachineFrameInfo *MFI = MF.getFrameInfo(); 2471 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true); 2472 2473 // Create load node to retrieve arguments from the stack. 2474 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2475 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, 2476 MachinePointerInfo::getFixedStack(FI), 2477 false, false, false, 0); 2478 } else { 2479 Reg = MF.addLiveIn(NextVA.getLocReg(), RC); 2480 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32); 2481 } 2482 2483 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2); 2484 } 2485 2486 void 2487 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF, 2488 unsigned &VARegSize, unsigned &VARegSaveSize) 2489 const { 2490 unsigned NumGPRs; 2491 if (CCInfo.isFirstByValRegValid()) 2492 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg(); 2493 else { 2494 unsigned int firstUnalloced; 2495 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs, 2496 sizeof(GPRArgRegs) / 2497 sizeof(GPRArgRegs[0])); 2498 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0; 2499 } 2500 2501 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment(); 2502 VARegSize = NumGPRs * 4; 2503 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1); 2504 } 2505 2506 // The remaining GPRs hold either the beginning of variable-argument 2507 // data, or the beginning of an aggregate passed by value (usuall 2508 // byval). Either way, we allocate stack slots adjacent to the data 2509 // provided by our caller, and store the unallocated registers there. 2510 // If this is a variadic function, the va_list pointer will begin with 2511 // these values; otherwise, this reassembles a (byval) structure that 2512 // was split between registers and memory. 2513 void 2514 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG, 2515 DebugLoc dl, SDValue &Chain, 2516 unsigned ArgOffset) const { 2517 MachineFunction &MF = DAG.getMachineFunction(); 2518 MachineFrameInfo *MFI = MF.getFrameInfo(); 2519 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2520 unsigned firstRegToSaveIndex; 2521 if (CCInfo.isFirstByValRegValid()) 2522 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0; 2523 else { 2524 firstRegToSaveIndex = CCInfo.getFirstUnallocated 2525 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0])); 2526 } 2527 2528 unsigned VARegSize, VARegSaveSize; 2529 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize); 2530 if (VARegSaveSize) { 2531 // If this function is vararg, store any remaining integer argument regs 2532 // to their spots on the stack so that they may be loaded by deferencing 2533 // the result of va_next. 2534 AFI->setVarArgsRegSaveSize(VARegSaveSize); 2535 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize, 2536 ArgOffset + VARegSaveSize 2537 - VARegSize, 2538 false)); 2539 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(), 2540 getPointerTy()); 2541 2542 SmallVector<SDValue, 4> MemOps; 2543 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) { 2544 const TargetRegisterClass *RC; 2545 if (AFI->isThumb1OnlyFunction()) 2546 RC = ARM::tGPRRegisterClass; 2547 else 2548 RC = ARM::GPRRegisterClass; 2549 2550 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC); 2551 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32); 2552 SDValue Store = 2553 DAG.getStore(Val.getValue(1), dl, Val, FIN, 2554 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()), 2555 false, false, 0); 2556 MemOps.push_back(Store); 2557 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN, 2558 DAG.getConstant(4, getPointerTy())); 2559 } 2560 if (!MemOps.empty()) 2561 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 2562 &MemOps[0], MemOps.size()); 2563 } else 2564 // This will point to the next argument passed via stack. 2565 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true)); 2566 } 2567 2568 SDValue 2569 ARMTargetLowering::LowerFormalArguments(SDValue Chain, 2570 CallingConv::ID CallConv, bool isVarArg, 2571 const SmallVectorImpl<ISD::InputArg> 2572 &Ins, 2573 DebugLoc dl, SelectionDAG &DAG, 2574 SmallVectorImpl<SDValue> &InVals) 2575 const { 2576 MachineFunction &MF = DAG.getMachineFunction(); 2577 MachineFrameInfo *MFI = MF.getFrameInfo(); 2578 2579 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2580 2581 // Assign locations to all of the incoming arguments. 2582 SmallVector<CCValAssign, 16> ArgLocs; 2583 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 2584 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue); 2585 CCInfo.AnalyzeFormalArguments(Ins, 2586 CCAssignFnForNode(CallConv, /* Return*/ false, 2587 isVarArg)); 2588 2589 SmallVector<SDValue, 16> ArgValues; 2590 int lastInsIndex = -1; 2591 2592 SDValue ArgValue; 2593 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 2594 CCValAssign &VA = ArgLocs[i]; 2595 2596 // Arguments stored in registers. 2597 if (VA.isRegLoc()) { 2598 EVT RegVT = VA.getLocVT(); 2599 2600 if (VA.needsCustom()) { 2601 // f64 and vector types are split up into multiple registers or 2602 // combinations of registers and stack slots. 2603 if (VA.getLocVT() == MVT::v2f64) { 2604 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i], 2605 Chain, DAG, dl); 2606 VA = ArgLocs[++i]; // skip ahead to next loc 2607 SDValue ArgValue2; 2608 if (VA.isMemLoc()) { 2609 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true); 2610 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2611 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN, 2612 MachinePointerInfo::getFixedStack(FI), 2613 false, false, false, 0); 2614 } else { 2615 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i], 2616 Chain, DAG, dl); 2617 } 2618 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64); 2619 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, 2620 ArgValue, ArgValue1, DAG.getIntPtrConstant(0)); 2621 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, 2622 ArgValue, ArgValue2, DAG.getIntPtrConstant(1)); 2623 } else 2624 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl); 2625 2626 } else { 2627 const TargetRegisterClass *RC; 2628 2629 if (RegVT == MVT::f32) 2630 RC = ARM::SPRRegisterClass; 2631 else if (RegVT == MVT::f64) 2632 RC = ARM::DPRRegisterClass; 2633 else if (RegVT == MVT::v2f64) 2634 RC = ARM::QPRRegisterClass; 2635 else if (RegVT == MVT::i32) 2636 RC = (AFI->isThumb1OnlyFunction() ? 2637 ARM::tGPRRegisterClass : ARM::GPRRegisterClass); 2638 else 2639 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering"); 2640 2641 // Transform the arguments in physical registers into virtual ones. 2642 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); 2643 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT); 2644 } 2645 2646 // If this is an 8 or 16-bit value, it is really passed promoted 2647 // to 32 bits. Insert an assert[sz]ext to capture this, then 2648 // truncate to the right size. 2649 switch (VA.getLocInfo()) { 2650 default: llvm_unreachable("Unknown loc info!"); 2651 case CCValAssign::Full: break; 2652 case CCValAssign::BCvt: 2653 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue); 2654 break; 2655 case CCValAssign::SExt: 2656 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue, 2657 DAG.getValueType(VA.getValVT())); 2658 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 2659 break; 2660 case CCValAssign::ZExt: 2661 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue, 2662 DAG.getValueType(VA.getValVT())); 2663 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 2664 break; 2665 } 2666 2667 InVals.push_back(ArgValue); 2668 2669 } else { // VA.isRegLoc() 2670 2671 // sanity check 2672 assert(VA.isMemLoc()); 2673 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered"); 2674 2675 int index = ArgLocs[i].getValNo(); 2676 2677 // Some Ins[] entries become multiple ArgLoc[] entries. 2678 // Process them only once. 2679 if (index != lastInsIndex) 2680 { 2681 ISD::ArgFlagsTy Flags = Ins[index].Flags; 2682 // FIXME: For now, all byval parameter objects are marked mutable. 2683 // This can be changed with more analysis. 2684 // In case of tail call optimization mark all arguments mutable. 2685 // Since they could be overwritten by lowering of arguments in case of 2686 // a tail call. 2687 if (Flags.isByVal()) { 2688 unsigned VARegSize, VARegSaveSize; 2689 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize); 2690 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0); 2691 unsigned Bytes = Flags.getByValSize() - VARegSize; 2692 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects. 2693 int FI = MFI->CreateFixedObject(Bytes, 2694 VA.getLocMemOffset(), false); 2695 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy())); 2696 } else { 2697 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8, 2698 VA.getLocMemOffset(), true); 2699 2700 // Create load nodes to retrieve arguments from the stack. 2701 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2702 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, 2703 MachinePointerInfo::getFixedStack(FI), 2704 false, false, false, 0)); 2705 } 2706 lastInsIndex = index; 2707 } 2708 } 2709 } 2710 2711 // varargs 2712 if (isVarArg) 2713 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset()); 2714 2715 return Chain; 2716 } 2717 2718 /// isFloatingPointZero - Return true if this is +0.0. 2719 static bool isFloatingPointZero(SDValue Op) { 2720 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op)) 2721 return CFP->getValueAPF().isPosZero(); 2722 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) { 2723 // Maybe this has already been legalized into the constant pool? 2724 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) { 2725 SDValue WrapperOp = Op.getOperand(1).getOperand(0); 2726 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp)) 2727 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal())) 2728 return CFP->getValueAPF().isPosZero(); 2729 } 2730 } 2731 return false; 2732 } 2733 2734 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for 2735 /// the given operands. 2736 SDValue 2737 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC, 2738 SDValue &ARMcc, SelectionDAG &DAG, 2739 DebugLoc dl) const { 2740 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) { 2741 unsigned C = RHSC->getZExtValue(); 2742 if (!isLegalICmpImmediate(C)) { 2743 // Constant does not fit, try adjusting it by one? 2744 switch (CC) { 2745 default: break; 2746 case ISD::SETLT: 2747 case ISD::SETGE: 2748 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) { 2749 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT; 2750 RHS = DAG.getConstant(C-1, MVT::i32); 2751 } 2752 break; 2753 case ISD::SETULT: 2754 case ISD::SETUGE: 2755 if (C != 0 && isLegalICmpImmediate(C-1)) { 2756 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT; 2757 RHS = DAG.getConstant(C-1, MVT::i32); 2758 } 2759 break; 2760 case ISD::SETLE: 2761 case ISD::SETGT: 2762 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) { 2763 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE; 2764 RHS = DAG.getConstant(C+1, MVT::i32); 2765 } 2766 break; 2767 case ISD::SETULE: 2768 case ISD::SETUGT: 2769 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) { 2770 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; 2771 RHS = DAG.getConstant(C+1, MVT::i32); 2772 } 2773 break; 2774 } 2775 } 2776 } 2777 2778 ARMCC::CondCodes CondCode = IntCCToARMCC(CC); 2779 ARMISD::NodeType CompareType; 2780 switch (CondCode) { 2781 default: 2782 CompareType = ARMISD::CMP; 2783 break; 2784 case ARMCC::EQ: 2785 case ARMCC::NE: 2786 // Uses only Z Flag 2787 CompareType = ARMISD::CMPZ; 2788 break; 2789 } 2790 ARMcc = DAG.getConstant(CondCode, MVT::i32); 2791 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS); 2792 } 2793 2794 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands. 2795 SDValue 2796 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG, 2797 DebugLoc dl) const { 2798 SDValue Cmp; 2799 if (!isFloatingPointZero(RHS)) 2800 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS); 2801 else 2802 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS); 2803 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp); 2804 } 2805 2806 /// duplicateCmp - Glue values can have only one use, so this function 2807 /// duplicates a comparison node. 2808 SDValue 2809 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const { 2810 unsigned Opc = Cmp.getOpcode(); 2811 DebugLoc DL = Cmp.getDebugLoc(); 2812 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ) 2813 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1)); 2814 2815 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation"); 2816 Cmp = Cmp.getOperand(0); 2817 Opc = Cmp.getOpcode(); 2818 if (Opc == ARMISD::CMPFP) 2819 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1)); 2820 else { 2821 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT"); 2822 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0)); 2823 } 2824 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp); 2825 } 2826 2827 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const { 2828 SDValue Cond = Op.getOperand(0); 2829 SDValue SelectTrue = Op.getOperand(1); 2830 SDValue SelectFalse = Op.getOperand(2); 2831 DebugLoc dl = Op.getDebugLoc(); 2832 2833 // Convert: 2834 // 2835 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond) 2836 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond) 2837 // 2838 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) { 2839 const ConstantSDNode *CMOVTrue = 2840 dyn_cast<ConstantSDNode>(Cond.getOperand(0)); 2841 const ConstantSDNode *CMOVFalse = 2842 dyn_cast<ConstantSDNode>(Cond.getOperand(1)); 2843 2844 if (CMOVTrue && CMOVFalse) { 2845 unsigned CMOVTrueVal = CMOVTrue->getZExtValue(); 2846 unsigned CMOVFalseVal = CMOVFalse->getZExtValue(); 2847 2848 SDValue True; 2849 SDValue False; 2850 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) { 2851 True = SelectTrue; 2852 False = SelectFalse; 2853 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) { 2854 True = SelectFalse; 2855 False = SelectTrue; 2856 } 2857 2858 if (True.getNode() && False.getNode()) { 2859 EVT VT = Op.getValueType(); 2860 SDValue ARMcc = Cond.getOperand(2); 2861 SDValue CCR = Cond.getOperand(3); 2862 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG); 2863 assert(True.getValueType() == VT); 2864 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp); 2865 } 2866 } 2867 } 2868 2869 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the 2870 // undefined bits before doing a full-word comparison with zero. 2871 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond, 2872 DAG.getConstant(1, Cond.getValueType())); 2873 2874 return DAG.getSelectCC(dl, Cond, 2875 DAG.getConstant(0, Cond.getValueType()), 2876 SelectTrue, SelectFalse, ISD::SETNE); 2877 } 2878 2879 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const { 2880 EVT VT = Op.getValueType(); 2881 SDValue LHS = Op.getOperand(0); 2882 SDValue RHS = Op.getOperand(1); 2883 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get(); 2884 SDValue TrueVal = Op.getOperand(2); 2885 SDValue FalseVal = Op.getOperand(3); 2886 DebugLoc dl = Op.getDebugLoc(); 2887 2888 if (LHS.getValueType() == MVT::i32) { 2889 SDValue ARMcc; 2890 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 2891 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl); 2892 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp); 2893 } 2894 2895 ARMCC::CondCodes CondCode, CondCode2; 2896 FPCCToARMCC(CC, CondCode, CondCode2); 2897 2898 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32); 2899 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl); 2900 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 2901 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, 2902 ARMcc, CCR, Cmp); 2903 if (CondCode2 != ARMCC::AL) { 2904 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32); 2905 // FIXME: Needs another CMP because flag can have but one use. 2906 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl); 2907 Result = DAG.getNode(ARMISD::CMOV, dl, VT, 2908 Result, TrueVal, ARMcc2, CCR, Cmp2); 2909 } 2910 return Result; 2911 } 2912 2913 /// canChangeToInt - Given the fp compare operand, return true if it is suitable 2914 /// to morph to an integer compare sequence. 2915 static bool canChangeToInt(SDValue Op, bool &SeenZero, 2916 const ARMSubtarget *Subtarget) { 2917 SDNode *N = Op.getNode(); 2918 if (!N->hasOneUse()) 2919 // Otherwise it requires moving the value from fp to integer registers. 2920 return false; 2921 if (!N->getNumValues()) 2922 return false; 2923 EVT VT = Op.getValueType(); 2924 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow()) 2925 // f32 case is generally profitable. f64 case only makes sense when vcmpe + 2926 // vmrs are very slow, e.g. cortex-a8. 2927 return false; 2928 2929 if (isFloatingPointZero(Op)) { 2930 SeenZero = true; 2931 return true; 2932 } 2933 return ISD::isNormalLoad(N); 2934 } 2935 2936 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) { 2937 if (isFloatingPointZero(Op)) 2938 return DAG.getConstant(0, MVT::i32); 2939 2940 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) 2941 return DAG.getLoad(MVT::i32, Op.getDebugLoc(), 2942 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(), 2943 Ld->isVolatile(), Ld->isNonTemporal(), 2944 Ld->isInvariant(), Ld->getAlignment()); 2945 2946 llvm_unreachable("Unknown VFP cmp argument!"); 2947 } 2948 2949 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG, 2950 SDValue &RetVal1, SDValue &RetVal2) { 2951 if (isFloatingPointZero(Op)) { 2952 RetVal1 = DAG.getConstant(0, MVT::i32); 2953 RetVal2 = DAG.getConstant(0, MVT::i32); 2954 return; 2955 } 2956 2957 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) { 2958 SDValue Ptr = Ld->getBasePtr(); 2959 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(), 2960 Ld->getChain(), Ptr, 2961 Ld->getPointerInfo(), 2962 Ld->isVolatile(), Ld->isNonTemporal(), 2963 Ld->isInvariant(), Ld->getAlignment()); 2964 2965 EVT PtrType = Ptr.getValueType(); 2966 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4); 2967 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(), 2968 PtrType, Ptr, DAG.getConstant(4, PtrType)); 2969 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(), 2970 Ld->getChain(), NewPtr, 2971 Ld->getPointerInfo().getWithOffset(4), 2972 Ld->isVolatile(), Ld->isNonTemporal(), 2973 Ld->isInvariant(), NewAlign); 2974 return; 2975 } 2976 2977 llvm_unreachable("Unknown VFP cmp argument!"); 2978 } 2979 2980 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some 2981 /// f32 and even f64 comparisons to integer ones. 2982 SDValue 2983 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const { 2984 SDValue Chain = Op.getOperand(0); 2985 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 2986 SDValue LHS = Op.getOperand(2); 2987 SDValue RHS = Op.getOperand(3); 2988 SDValue Dest = Op.getOperand(4); 2989 DebugLoc dl = Op.getDebugLoc(); 2990 2991 bool LHSSeenZero = false; 2992 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget); 2993 bool RHSSeenZero = false; 2994 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget); 2995 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) { 2996 // If unsafe fp math optimization is enabled and there are no other uses of 2997 // the CMP operands, and the condition code is EQ or NE, we can optimize it 2998 // to an integer comparison. 2999 if (CC == ISD::SETOEQ) 3000 CC = ISD::SETEQ; 3001 else if (CC == ISD::SETUNE) 3002 CC = ISD::SETNE; 3003 3004 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32); 3005 SDValue ARMcc; 3006 if (LHS.getValueType() == MVT::f32) { 3007 LHS = DAG.getNode(ISD::AND, dl, MVT::i32, 3008 bitcastf32Toi32(LHS, DAG), Mask); 3009 RHS = DAG.getNode(ISD::AND, dl, MVT::i32, 3010 bitcastf32Toi32(RHS, DAG), Mask); 3011 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl); 3012 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3013 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other, 3014 Chain, Dest, ARMcc, CCR, Cmp); 3015 } 3016 3017 SDValue LHS1, LHS2; 3018 SDValue RHS1, RHS2; 3019 expandf64Toi32(LHS, DAG, LHS1, LHS2); 3020 expandf64Toi32(RHS, DAG, RHS1, RHS2); 3021 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask); 3022 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask); 3023 ARMCC::CondCodes CondCode = IntCCToARMCC(CC); 3024 ARMcc = DAG.getConstant(CondCode, MVT::i32); 3025 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue); 3026 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest }; 3027 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7); 3028 } 3029 3030 return SDValue(); 3031 } 3032 3033 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const { 3034 SDValue Chain = Op.getOperand(0); 3035 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 3036 SDValue LHS = Op.getOperand(2); 3037 SDValue RHS = Op.getOperand(3); 3038 SDValue Dest = Op.getOperand(4); 3039 DebugLoc dl = Op.getDebugLoc(); 3040 3041 if (LHS.getValueType() == MVT::i32) { 3042 SDValue ARMcc; 3043 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl); 3044 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3045 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other, 3046 Chain, Dest, ARMcc, CCR, Cmp); 3047 } 3048 3049 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64); 3050 3051 if (getTargetMachine().Options.UnsafeFPMath && 3052 (CC == ISD::SETEQ || CC == ISD::SETOEQ || 3053 CC == ISD::SETNE || CC == ISD::SETUNE)) { 3054 SDValue Result = OptimizeVFPBrcond(Op, DAG); 3055 if (Result.getNode()) 3056 return Result; 3057 } 3058 3059 ARMCC::CondCodes CondCode, CondCode2; 3060 FPCCToARMCC(CC, CondCode, CondCode2); 3061 3062 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32); 3063 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl); 3064 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3065 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue); 3066 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp }; 3067 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5); 3068 if (CondCode2 != ARMCC::AL) { 3069 ARMcc = DAG.getConstant(CondCode2, MVT::i32); 3070 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) }; 3071 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5); 3072 } 3073 return Res; 3074 } 3075 3076 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const { 3077 SDValue Chain = Op.getOperand(0); 3078 SDValue Table = Op.getOperand(1); 3079 SDValue Index = Op.getOperand(2); 3080 DebugLoc dl = Op.getDebugLoc(); 3081 3082 EVT PTy = getPointerTy(); 3083 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table); 3084 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>(); 3085 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy); 3086 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy); 3087 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId); 3088 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy)); 3089 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table); 3090 if (Subtarget->isThumb2()) { 3091 // Thumb2 uses a two-level jump. That is, it jumps into the jump table 3092 // which does another jump to the destination. This also makes it easier 3093 // to translate it to TBB / TBH later. 3094 // FIXME: This might not work if the function is extremely large. 3095 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain, 3096 Addr, Op.getOperand(2), JTI, UId); 3097 } 3098 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) { 3099 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, 3100 MachinePointerInfo::getJumpTable(), 3101 false, false, false, 0); 3102 Chain = Addr.getValue(1); 3103 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table); 3104 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId); 3105 } else { 3106 Addr = DAG.getLoad(PTy, dl, Chain, Addr, 3107 MachinePointerInfo::getJumpTable(), 3108 false, false, false, 0); 3109 Chain = Addr.getValue(1); 3110 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId); 3111 } 3112 } 3113 3114 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) { 3115 EVT VT = Op.getValueType(); 3116 DebugLoc dl = Op.getDebugLoc(); 3117 3118 if (Op.getValueType().getVectorElementType() == MVT::i32) { 3119 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32) 3120 return Op; 3121 return DAG.UnrollVectorOp(Op.getNode()); 3122 } 3123 3124 assert(Op.getOperand(0).getValueType() == MVT::v4f32 && 3125 "Invalid type for custom lowering!"); 3126 if (VT != MVT::v4i16) 3127 return DAG.UnrollVectorOp(Op.getNode()); 3128 3129 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0)); 3130 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op); 3131 } 3132 3133 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) { 3134 EVT VT = Op.getValueType(); 3135 if (VT.isVector()) 3136 return LowerVectorFP_TO_INT(Op, DAG); 3137 3138 DebugLoc dl = Op.getDebugLoc(); 3139 unsigned Opc; 3140 3141 switch (Op.getOpcode()) { 3142 default: llvm_unreachable("Invalid opcode!"); 3143 case ISD::FP_TO_SINT: 3144 Opc = ARMISD::FTOSI; 3145 break; 3146 case ISD::FP_TO_UINT: 3147 Opc = ARMISD::FTOUI; 3148 break; 3149 } 3150 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0)); 3151 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); 3152 } 3153 3154 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) { 3155 EVT VT = Op.getValueType(); 3156 DebugLoc dl = Op.getDebugLoc(); 3157 3158 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) { 3159 if (VT.getVectorElementType() == MVT::f32) 3160 return Op; 3161 return DAG.UnrollVectorOp(Op.getNode()); 3162 } 3163 3164 assert(Op.getOperand(0).getValueType() == MVT::v4i16 && 3165 "Invalid type for custom lowering!"); 3166 if (VT != MVT::v4f32) 3167 return DAG.UnrollVectorOp(Op.getNode()); 3168 3169 unsigned CastOpc; 3170 unsigned Opc; 3171 switch (Op.getOpcode()) { 3172 default: llvm_unreachable("Invalid opcode!"); 3173 case ISD::SINT_TO_FP: 3174 CastOpc = ISD::SIGN_EXTEND; 3175 Opc = ISD::SINT_TO_FP; 3176 break; 3177 case ISD::UINT_TO_FP: 3178 CastOpc = ISD::ZERO_EXTEND; 3179 Opc = ISD::UINT_TO_FP; 3180 break; 3181 } 3182 3183 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0)); 3184 return DAG.getNode(Opc, dl, VT, Op); 3185 } 3186 3187 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) { 3188 EVT VT = Op.getValueType(); 3189 if (VT.isVector()) 3190 return LowerVectorINT_TO_FP(Op, DAG); 3191 3192 DebugLoc dl = Op.getDebugLoc(); 3193 unsigned Opc; 3194 3195 switch (Op.getOpcode()) { 3196 default: llvm_unreachable("Invalid opcode!"); 3197 case ISD::SINT_TO_FP: 3198 Opc = ARMISD::SITOF; 3199 break; 3200 case ISD::UINT_TO_FP: 3201 Opc = ARMISD::UITOF; 3202 break; 3203 } 3204 3205 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0)); 3206 return DAG.getNode(Opc, dl, VT, Op); 3207 } 3208 3209 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { 3210 // Implement fcopysign with a fabs and a conditional fneg. 3211 SDValue Tmp0 = Op.getOperand(0); 3212 SDValue Tmp1 = Op.getOperand(1); 3213 DebugLoc dl = Op.getDebugLoc(); 3214 EVT VT = Op.getValueType(); 3215 EVT SrcVT = Tmp1.getValueType(); 3216 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST || 3217 Tmp0.getOpcode() == ARMISD::VMOVDRR; 3218 bool UseNEON = !InGPR && Subtarget->hasNEON(); 3219 3220 if (UseNEON) { 3221 // Use VBSL to copy the sign bit. 3222 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80); 3223 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32, 3224 DAG.getTargetConstant(EncodedVal, MVT::i32)); 3225 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64; 3226 if (VT == MVT::f64) 3227 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT, 3228 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask), 3229 DAG.getConstant(32, MVT::i32)); 3230 else /*if (VT == MVT::f32)*/ 3231 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0); 3232 if (SrcVT == MVT::f32) { 3233 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1); 3234 if (VT == MVT::f64) 3235 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT, 3236 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1), 3237 DAG.getConstant(32, MVT::i32)); 3238 } else if (VT == MVT::f32) 3239 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64, 3240 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1), 3241 DAG.getConstant(32, MVT::i32)); 3242 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0); 3243 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1); 3244 3245 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff), 3246 MVT::i32); 3247 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes); 3248 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask, 3249 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes)); 3250 3251 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT, 3252 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask), 3253 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot)); 3254 if (VT == MVT::f32) { 3255 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res); 3256 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res, 3257 DAG.getConstant(0, MVT::i32)); 3258 } else { 3259 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res); 3260 } 3261 3262 return Res; 3263 } 3264 3265 // Bitcast operand 1 to i32. 3266 if (SrcVT == MVT::f64) 3267 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32), 3268 &Tmp1, 1).getValue(1); 3269 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1); 3270 3271 // Or in the signbit with integer operations. 3272 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32); 3273 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32); 3274 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1); 3275 if (VT == MVT::f32) { 3276 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32, 3277 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2); 3278 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, 3279 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1)); 3280 } 3281 3282 // f64: Or the high part with signbit and then combine two parts. 3283 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32), 3284 &Tmp0, 1); 3285 SDValue Lo = Tmp0.getValue(0); 3286 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2); 3287 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1); 3288 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi); 3289 } 3290 3291 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{ 3292 MachineFunction &MF = DAG.getMachineFunction(); 3293 MachineFrameInfo *MFI = MF.getFrameInfo(); 3294 MFI->setReturnAddressIsTaken(true); 3295 3296 EVT VT = Op.getValueType(); 3297 DebugLoc dl = Op.getDebugLoc(); 3298 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 3299 if (Depth) { 3300 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG); 3301 SDValue Offset = DAG.getConstant(4, MVT::i32); 3302 return DAG.getLoad(VT, dl, DAG.getEntryNode(), 3303 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset), 3304 MachinePointerInfo(), false, false, false, 0); 3305 } 3306 3307 // Return LR, which contains the return address. Mark it an implicit live-in. 3308 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32)); 3309 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT); 3310 } 3311 3312 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { 3313 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 3314 MFI->setFrameAddressIsTaken(true); 3315 3316 EVT VT = Op.getValueType(); 3317 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful 3318 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 3319 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin()) 3320 ? ARM::R7 : ARM::R11; 3321 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT); 3322 while (Depth--) 3323 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, 3324 MachinePointerInfo(), 3325 false, false, false, 0); 3326 return FrameAddr; 3327 } 3328 3329 /// ExpandBITCAST - If the target supports VFP, this function is called to 3330 /// expand a bit convert where either the source or destination type is i64 to 3331 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64 3332 /// operand type is illegal (e.g., v2f32 for a target that doesn't support 3333 /// vectors), since the legalizer won't know what to do with that. 3334 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) { 3335 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3336 DebugLoc dl = N->getDebugLoc(); 3337 SDValue Op = N->getOperand(0); 3338 3339 // This function is only supposed to be called for i64 types, either as the 3340 // source or destination of the bit convert. 3341 EVT SrcVT = Op.getValueType(); 3342 EVT DstVT = N->getValueType(0); 3343 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) && 3344 "ExpandBITCAST called for non-i64 type"); 3345 3346 // Turn i64->f64 into VMOVDRR. 3347 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) { 3348 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op, 3349 DAG.getConstant(0, MVT::i32)); 3350 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op, 3351 DAG.getConstant(1, MVT::i32)); 3352 return DAG.getNode(ISD::BITCAST, dl, DstVT, 3353 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi)); 3354 } 3355 3356 // Turn f64->i64 into VMOVRRD. 3357 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) { 3358 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl, 3359 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1); 3360 // Merge the pieces into a single i64 value. 3361 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1)); 3362 } 3363 3364 return SDValue(); 3365 } 3366 3367 /// getZeroVector - Returns a vector of specified type with all zero elements. 3368 /// Zero vectors are used to represent vector negation and in those cases 3369 /// will be implemented with the NEON VNEG instruction. However, VNEG does 3370 /// not support i64 elements, so sometimes the zero vectors will need to be 3371 /// explicitly constructed. Regardless, use a canonical VMOV to create the 3372 /// zero vector. 3373 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) { 3374 assert(VT.isVector() && "Expected a vector type"); 3375 // The canonical modified immediate encoding of a zero vector is....0! 3376 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32); 3377 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32; 3378 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal); 3379 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov); 3380 } 3381 3382 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two 3383 /// i32 values and take a 2 x i32 value to shift plus a shift amount. 3384 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op, 3385 SelectionDAG &DAG) const { 3386 assert(Op.getNumOperands() == 3 && "Not a double-shift!"); 3387 EVT VT = Op.getValueType(); 3388 unsigned VTBits = VT.getSizeInBits(); 3389 DebugLoc dl = Op.getDebugLoc(); 3390 SDValue ShOpLo = Op.getOperand(0); 3391 SDValue ShOpHi = Op.getOperand(1); 3392 SDValue ShAmt = Op.getOperand(2); 3393 SDValue ARMcc; 3394 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL; 3395 3396 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS); 3397 3398 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, 3399 DAG.getConstant(VTBits, MVT::i32), ShAmt); 3400 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt); 3401 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt, 3402 DAG.getConstant(VTBits, MVT::i32)); 3403 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt); 3404 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2); 3405 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt); 3406 3407 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3408 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE, 3409 ARMcc, DAG, dl); 3410 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt); 3411 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, 3412 CCR, Cmp); 3413 3414 SDValue Ops[2] = { Lo, Hi }; 3415 return DAG.getMergeValues(Ops, 2, dl); 3416 } 3417 3418 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two 3419 /// i32 values and take a 2 x i32 value to shift plus a shift amount. 3420 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op, 3421 SelectionDAG &DAG) const { 3422 assert(Op.getNumOperands() == 3 && "Not a double-shift!"); 3423 EVT VT = Op.getValueType(); 3424 unsigned VTBits = VT.getSizeInBits(); 3425 DebugLoc dl = Op.getDebugLoc(); 3426 SDValue ShOpLo = Op.getOperand(0); 3427 SDValue ShOpHi = Op.getOperand(1); 3428 SDValue ShAmt = Op.getOperand(2); 3429 SDValue ARMcc; 3430 3431 assert(Op.getOpcode() == ISD::SHL_PARTS); 3432 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, 3433 DAG.getConstant(VTBits, MVT::i32), ShAmt); 3434 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt); 3435 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt, 3436 DAG.getConstant(VTBits, MVT::i32)); 3437 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt); 3438 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt); 3439 3440 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2); 3441 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3442 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE, 3443 ARMcc, DAG, dl); 3444 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt); 3445 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc, 3446 CCR, Cmp); 3447 3448 SDValue Ops[2] = { Lo, Hi }; 3449 return DAG.getMergeValues(Ops, 2, dl); 3450 } 3451 3452 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op, 3453 SelectionDAG &DAG) const { 3454 // The rounding mode is in bits 23:22 of the FPSCR. 3455 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0 3456 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3) 3457 // so that the shift + and get folded into a bitfield extract. 3458 DebugLoc dl = Op.getDebugLoc(); 3459 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32, 3460 DAG.getConstant(Intrinsic::arm_get_fpscr, 3461 MVT::i32)); 3462 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR, 3463 DAG.getConstant(1U << 22, MVT::i32)); 3464 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds, 3465 DAG.getConstant(22, MVT::i32)); 3466 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE, 3467 DAG.getConstant(3, MVT::i32)); 3468 } 3469 3470 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG, 3471 const ARMSubtarget *ST) { 3472 EVT VT = N->getValueType(0); 3473 DebugLoc dl = N->getDebugLoc(); 3474 3475 if (!ST->hasV6T2Ops()) 3476 return SDValue(); 3477 3478 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0)); 3479 return DAG.getNode(ISD::CTLZ, dl, VT, rbit); 3480 } 3481 3482 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG, 3483 const ARMSubtarget *ST) { 3484 EVT VT = N->getValueType(0); 3485 DebugLoc dl = N->getDebugLoc(); 3486 3487 if (!VT.isVector()) 3488 return SDValue(); 3489 3490 // Lower vector shifts on NEON to use VSHL. 3491 assert(ST->hasNEON() && "unexpected vector shift"); 3492 3493 // Left shifts translate directly to the vshiftu intrinsic. 3494 if (N->getOpcode() == ISD::SHL) 3495 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, 3496 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32), 3497 N->getOperand(0), N->getOperand(1)); 3498 3499 assert((N->getOpcode() == ISD::SRA || 3500 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode"); 3501 3502 // NEON uses the same intrinsics for both left and right shifts. For 3503 // right shifts, the shift amounts are negative, so negate the vector of 3504 // shift amounts. 3505 EVT ShiftVT = N->getOperand(1).getValueType(); 3506 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT, 3507 getZeroVector(ShiftVT, DAG, dl), 3508 N->getOperand(1)); 3509 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ? 3510 Intrinsic::arm_neon_vshifts : 3511 Intrinsic::arm_neon_vshiftu); 3512 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, 3513 DAG.getConstant(vshiftInt, MVT::i32), 3514 N->getOperand(0), NegatedCount); 3515 } 3516 3517 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG, 3518 const ARMSubtarget *ST) { 3519 EVT VT = N->getValueType(0); 3520 DebugLoc dl = N->getDebugLoc(); 3521 3522 // We can get here for a node like i32 = ISD::SHL i32, i64 3523 if (VT != MVT::i64) 3524 return SDValue(); 3525 3526 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) && 3527 "Unknown shift to lower!"); 3528 3529 // We only lower SRA, SRL of 1 here, all others use generic lowering. 3530 if (!isa<ConstantSDNode>(N->getOperand(1)) || 3531 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1) 3532 return SDValue(); 3533 3534 // If we are in thumb mode, we don't have RRX. 3535 if (ST->isThumb1Only()) return SDValue(); 3536 3537 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr. 3538 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0), 3539 DAG.getConstant(0, MVT::i32)); 3540 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0), 3541 DAG.getConstant(1, MVT::i32)); 3542 3543 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and 3544 // captures the result into a carry flag. 3545 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG; 3546 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1); 3547 3548 // The low part is an ARMISD::RRX operand, which shifts the carry in. 3549 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1)); 3550 3551 // Merge the pieces into a single i64 value. 3552 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi); 3553 } 3554 3555 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) { 3556 SDValue TmpOp0, TmpOp1; 3557 bool Invert = false; 3558 bool Swap = false; 3559 unsigned Opc = 0; 3560 3561 SDValue Op0 = Op.getOperand(0); 3562 SDValue Op1 = Op.getOperand(1); 3563 SDValue CC = Op.getOperand(2); 3564 EVT VT = Op.getValueType(); 3565 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get(); 3566 DebugLoc dl = Op.getDebugLoc(); 3567 3568 if (Op.getOperand(1).getValueType().isFloatingPoint()) { 3569 switch (SetCCOpcode) { 3570 default: llvm_unreachable("Illegal FP comparison"); 3571 case ISD::SETUNE: 3572 case ISD::SETNE: Invert = true; // Fallthrough 3573 case ISD::SETOEQ: 3574 case ISD::SETEQ: Opc = ARMISD::VCEQ; break; 3575 case ISD::SETOLT: 3576 case ISD::SETLT: Swap = true; // Fallthrough 3577 case ISD::SETOGT: 3578 case ISD::SETGT: Opc = ARMISD::VCGT; break; 3579 case ISD::SETOLE: 3580 case ISD::SETLE: Swap = true; // Fallthrough 3581 case ISD::SETOGE: 3582 case ISD::SETGE: Opc = ARMISD::VCGE; break; 3583 case ISD::SETUGE: Swap = true; // Fallthrough 3584 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break; 3585 case ISD::SETUGT: Swap = true; // Fallthrough 3586 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break; 3587 case ISD::SETUEQ: Invert = true; // Fallthrough 3588 case ISD::SETONE: 3589 // Expand this to (OLT | OGT). 3590 TmpOp0 = Op0; 3591 TmpOp1 = Op1; 3592 Opc = ISD::OR; 3593 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0); 3594 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1); 3595 break; 3596 case ISD::SETUO: Invert = true; // Fallthrough 3597 case ISD::SETO: 3598 // Expand this to (OLT | OGE). 3599 TmpOp0 = Op0; 3600 TmpOp1 = Op1; 3601 Opc = ISD::OR; 3602 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0); 3603 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1); 3604 break; 3605 } 3606 } else { 3607 // Integer comparisons. 3608 switch (SetCCOpcode) { 3609 default: llvm_unreachable("Illegal integer comparison"); 3610 case ISD::SETNE: Invert = true; 3611 case ISD::SETEQ: Opc = ARMISD::VCEQ; break; 3612 case ISD::SETLT: Swap = true; 3613 case ISD::SETGT: Opc = ARMISD::VCGT; break; 3614 case ISD::SETLE: Swap = true; 3615 case ISD::SETGE: Opc = ARMISD::VCGE; break; 3616 case ISD::SETULT: Swap = true; 3617 case ISD::SETUGT: Opc = ARMISD::VCGTU; break; 3618 case ISD::SETULE: Swap = true; 3619 case ISD::SETUGE: Opc = ARMISD::VCGEU; break; 3620 } 3621 3622 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero). 3623 if (Opc == ARMISD::VCEQ) { 3624 3625 SDValue AndOp; 3626 if (ISD::isBuildVectorAllZeros(Op1.getNode())) 3627 AndOp = Op0; 3628 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) 3629 AndOp = Op1; 3630 3631 // Ignore bitconvert. 3632 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST) 3633 AndOp = AndOp.getOperand(0); 3634 3635 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) { 3636 Opc = ARMISD::VTST; 3637 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0)); 3638 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1)); 3639 Invert = !Invert; 3640 } 3641 } 3642 } 3643 3644 if (Swap) 3645 std::swap(Op0, Op1); 3646 3647 // If one of the operands is a constant vector zero, attempt to fold the 3648 // comparison to a specialized compare-against-zero form. 3649 SDValue SingleOp; 3650 if (ISD::isBuildVectorAllZeros(Op1.getNode())) 3651 SingleOp = Op0; 3652 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) { 3653 if (Opc == ARMISD::VCGE) 3654 Opc = ARMISD::VCLEZ; 3655 else if (Opc == ARMISD::VCGT) 3656 Opc = ARMISD::VCLTZ; 3657 SingleOp = Op1; 3658 } 3659 3660 SDValue Result; 3661 if (SingleOp.getNode()) { 3662 switch (Opc) { 3663 case ARMISD::VCEQ: 3664 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break; 3665 case ARMISD::VCGE: 3666 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break; 3667 case ARMISD::VCLEZ: 3668 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break; 3669 case ARMISD::VCGT: 3670 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break; 3671 case ARMISD::VCLTZ: 3672 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break; 3673 default: 3674 Result = DAG.getNode(Opc, dl, VT, Op0, Op1); 3675 } 3676 } else { 3677 Result = DAG.getNode(Opc, dl, VT, Op0, Op1); 3678 } 3679 3680 if (Invert) 3681 Result = DAG.getNOT(dl, Result, VT); 3682 3683 return Result; 3684 } 3685 3686 /// isNEONModifiedImm - Check if the specified splat value corresponds to a 3687 /// valid vector constant for a NEON instruction with a "modified immediate" 3688 /// operand (e.g., VMOV). If so, return the encoded value. 3689 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef, 3690 unsigned SplatBitSize, SelectionDAG &DAG, 3691 EVT &VT, bool is128Bits, NEONModImmType type) { 3692 unsigned OpCmode, Imm; 3693 3694 // SplatBitSize is set to the smallest size that splats the vector, so a 3695 // zero vector will always have SplatBitSize == 8. However, NEON modified 3696 // immediate instructions others than VMOV do not support the 8-bit encoding 3697 // of a zero vector, and the default encoding of zero is supposed to be the 3698 // 32-bit version. 3699 if (SplatBits == 0) 3700 SplatBitSize = 32; 3701 3702 switch (SplatBitSize) { 3703 case 8: 3704 if (type != VMOVModImm) 3705 return SDValue(); 3706 // Any 1-byte value is OK. Op=0, Cmode=1110. 3707 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big"); 3708 OpCmode = 0xe; 3709 Imm = SplatBits; 3710 VT = is128Bits ? MVT::v16i8 : MVT::v8i8; 3711 break; 3712 3713 case 16: 3714 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero. 3715 VT = is128Bits ? MVT::v8i16 : MVT::v4i16; 3716 if ((SplatBits & ~0xff) == 0) { 3717 // Value = 0x00nn: Op=x, Cmode=100x. 3718 OpCmode = 0x8; 3719 Imm = SplatBits; 3720 break; 3721 } 3722 if ((SplatBits & ~0xff00) == 0) { 3723 // Value = 0xnn00: Op=x, Cmode=101x. 3724 OpCmode = 0xa; 3725 Imm = SplatBits >> 8; 3726 break; 3727 } 3728 return SDValue(); 3729 3730 case 32: 3731 // NEON's 32-bit VMOV supports splat values where: 3732 // * only one byte is nonzero, or 3733 // * the least significant byte is 0xff and the second byte is nonzero, or 3734 // * the least significant 2 bytes are 0xff and the third is nonzero. 3735 VT = is128Bits ? MVT::v4i32 : MVT::v2i32; 3736 if ((SplatBits & ~0xff) == 0) { 3737 // Value = 0x000000nn: Op=x, Cmode=000x. 3738 OpCmode = 0; 3739 Imm = SplatBits; 3740 break; 3741 } 3742 if ((SplatBits & ~0xff00) == 0) { 3743 // Value = 0x0000nn00: Op=x, Cmode=001x. 3744 OpCmode = 0x2; 3745 Imm = SplatBits >> 8; 3746 break; 3747 } 3748 if ((SplatBits & ~0xff0000) == 0) { 3749 // Value = 0x00nn0000: Op=x, Cmode=010x. 3750 OpCmode = 0x4; 3751 Imm = SplatBits >> 16; 3752 break; 3753 } 3754 if ((SplatBits & ~0xff000000) == 0) { 3755 // Value = 0xnn000000: Op=x, Cmode=011x. 3756 OpCmode = 0x6; 3757 Imm = SplatBits >> 24; 3758 break; 3759 } 3760 3761 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC 3762 if (type == OtherModImm) return SDValue(); 3763 3764 if ((SplatBits & ~0xffff) == 0 && 3765 ((SplatBits | SplatUndef) & 0xff) == 0xff) { 3766 // Value = 0x0000nnff: Op=x, Cmode=1100. 3767 OpCmode = 0xc; 3768 Imm = SplatBits >> 8; 3769 SplatBits |= 0xff; 3770 break; 3771 } 3772 3773 if ((SplatBits & ~0xffffff) == 0 && 3774 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) { 3775 // Value = 0x00nnffff: Op=x, Cmode=1101. 3776 OpCmode = 0xd; 3777 Imm = SplatBits >> 16; 3778 SplatBits |= 0xffff; 3779 break; 3780 } 3781 3782 // Note: there are a few 32-bit splat values (specifically: 00ffff00, 3783 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not 3784 // VMOV.I32. A (very) minor optimization would be to replicate the value 3785 // and fall through here to test for a valid 64-bit splat. But, then the 3786 // caller would also need to check and handle the change in size. 3787 return SDValue(); 3788 3789 case 64: { 3790 if (type != VMOVModImm) 3791 return SDValue(); 3792 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff. 3793 uint64_t BitMask = 0xff; 3794 uint64_t Val = 0; 3795 unsigned ImmMask = 1; 3796 Imm = 0; 3797 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) { 3798 if (((SplatBits | SplatUndef) & BitMask) == BitMask) { 3799 Val |= BitMask; 3800 Imm |= ImmMask; 3801 } else if ((SplatBits & BitMask) != 0) { 3802 return SDValue(); 3803 } 3804 BitMask <<= 8; 3805 ImmMask <<= 1; 3806 } 3807 // Op=1, Cmode=1110. 3808 OpCmode = 0x1e; 3809 SplatBits = Val; 3810 VT = is128Bits ? MVT::v2i64 : MVT::v1i64; 3811 break; 3812 } 3813 3814 default: 3815 llvm_unreachable("unexpected size for isNEONModifiedImm"); 3816 } 3817 3818 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm); 3819 return DAG.getTargetConstant(EncodedVal, MVT::i32); 3820 } 3821 3822 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG, 3823 const ARMSubtarget *ST) const { 3824 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16()) 3825 return SDValue(); 3826 3827 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op); 3828 assert(Op.getValueType() == MVT::f32 && 3829 "ConstantFP custom lowering should only occur for f32."); 3830 3831 // Try splatting with a VMOV.f32... 3832 APFloat FPVal = CFP->getValueAPF(); 3833 int ImmVal = ARM_AM::getFP32Imm(FPVal); 3834 if (ImmVal != -1) { 3835 DebugLoc DL = Op.getDebugLoc(); 3836 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32); 3837 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32, 3838 NewVal); 3839 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant, 3840 DAG.getConstant(0, MVT::i32)); 3841 } 3842 3843 // If that fails, try a VMOV.i32 3844 EVT VMovVT; 3845 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue(); 3846 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false, 3847 VMOVModImm); 3848 if (NewVal != SDValue()) { 3849 DebugLoc DL = Op.getDebugLoc(); 3850 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT, 3851 NewVal); 3852 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32, 3853 VecConstant); 3854 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant, 3855 DAG.getConstant(0, MVT::i32)); 3856 } 3857 3858 // Finally, try a VMVN.i32 3859 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false, 3860 VMVNModImm); 3861 if (NewVal != SDValue()) { 3862 DebugLoc DL = Op.getDebugLoc(); 3863 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal); 3864 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32, 3865 VecConstant); 3866 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant, 3867 DAG.getConstant(0, MVT::i32)); 3868 } 3869 3870 return SDValue(); 3871 } 3872 3873 3874 static bool isVEXTMask(ArrayRef<int> M, EVT VT, 3875 bool &ReverseVEXT, unsigned &Imm) { 3876 unsigned NumElts = VT.getVectorNumElements(); 3877 ReverseVEXT = false; 3878 3879 // Assume that the first shuffle index is not UNDEF. Fail if it is. 3880 if (M[0] < 0) 3881 return false; 3882 3883 Imm = M[0]; 3884 3885 // If this is a VEXT shuffle, the immediate value is the index of the first 3886 // element. The other shuffle indices must be the successive elements after 3887 // the first one. 3888 unsigned ExpectedElt = Imm; 3889 for (unsigned i = 1; i < NumElts; ++i) { 3890 // Increment the expected index. If it wraps around, it may still be 3891 // a VEXT but the source vectors must be swapped. 3892 ExpectedElt += 1; 3893 if (ExpectedElt == NumElts * 2) { 3894 ExpectedElt = 0; 3895 ReverseVEXT = true; 3896 } 3897 3898 if (M[i] < 0) continue; // ignore UNDEF indices 3899 if (ExpectedElt != static_cast<unsigned>(M[i])) 3900 return false; 3901 } 3902 3903 // Adjust the index value if the source operands will be swapped. 3904 if (ReverseVEXT) 3905 Imm -= NumElts; 3906 3907 return true; 3908 } 3909 3910 /// isVREVMask - Check if a vector shuffle corresponds to a VREV 3911 /// instruction with the specified blocksize. (The order of the elements 3912 /// within each block of the vector is reversed.) 3913 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) { 3914 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) && 3915 "Only possible block sizes for VREV are: 16, 32, 64"); 3916 3917 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 3918 if (EltSz == 64) 3919 return false; 3920 3921 unsigned NumElts = VT.getVectorNumElements(); 3922 unsigned BlockElts = M[0] + 1; 3923 // If the first shuffle index is UNDEF, be optimistic. 3924 if (M[0] < 0) 3925 BlockElts = BlockSize / EltSz; 3926 3927 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz) 3928 return false; 3929 3930 for (unsigned i = 0; i < NumElts; ++i) { 3931 if (M[i] < 0) continue; // ignore UNDEF indices 3932 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts)) 3933 return false; 3934 } 3935 3936 return true; 3937 } 3938 3939 static bool isVTBLMask(ArrayRef<int> M, EVT VT) { 3940 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of 3941 // range, then 0 is placed into the resulting vector. So pretty much any mask 3942 // of 8 elements can work here. 3943 return VT == MVT::v8i8 && M.size() == 8; 3944 } 3945 3946 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 3947 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 3948 if (EltSz == 64) 3949 return false; 3950 3951 unsigned NumElts = VT.getVectorNumElements(); 3952 WhichResult = (M[0] == 0 ? 0 : 1); 3953 for (unsigned i = 0; i < NumElts; i += 2) { 3954 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) || 3955 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult)) 3956 return false; 3957 } 3958 return true; 3959 } 3960 3961 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of 3962 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 3963 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>. 3964 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){ 3965 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 3966 if (EltSz == 64) 3967 return false; 3968 3969 unsigned NumElts = VT.getVectorNumElements(); 3970 WhichResult = (M[0] == 0 ? 0 : 1); 3971 for (unsigned i = 0; i < NumElts; i += 2) { 3972 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) || 3973 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult)) 3974 return false; 3975 } 3976 return true; 3977 } 3978 3979 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 3980 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 3981 if (EltSz == 64) 3982 return false; 3983 3984 unsigned NumElts = VT.getVectorNumElements(); 3985 WhichResult = (M[0] == 0 ? 0 : 1); 3986 for (unsigned i = 0; i != NumElts; ++i) { 3987 if (M[i] < 0) continue; // ignore UNDEF indices 3988 if ((unsigned) M[i] != 2 * i + WhichResult) 3989 return false; 3990 } 3991 3992 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 3993 if (VT.is64BitVector() && EltSz == 32) 3994 return false; 3995 3996 return true; 3997 } 3998 3999 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of 4000 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 4001 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>, 4002 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){ 4003 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4004 if (EltSz == 64) 4005 return false; 4006 4007 unsigned Half = VT.getVectorNumElements() / 2; 4008 WhichResult = (M[0] == 0 ? 0 : 1); 4009 for (unsigned j = 0; j != 2; ++j) { 4010 unsigned Idx = WhichResult; 4011 for (unsigned i = 0; i != Half; ++i) { 4012 int MIdx = M[i + j * Half]; 4013 if (MIdx >= 0 && (unsigned) MIdx != Idx) 4014 return false; 4015 Idx += 2; 4016 } 4017 } 4018 4019 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4020 if (VT.is64BitVector() && EltSz == 32) 4021 return false; 4022 4023 return true; 4024 } 4025 4026 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 4027 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4028 if (EltSz == 64) 4029 return false; 4030 4031 unsigned NumElts = VT.getVectorNumElements(); 4032 WhichResult = (M[0] == 0 ? 0 : 1); 4033 unsigned Idx = WhichResult * NumElts / 2; 4034 for (unsigned i = 0; i != NumElts; i += 2) { 4035 if ((M[i] >= 0 && (unsigned) M[i] != Idx) || 4036 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts)) 4037 return false; 4038 Idx += 1; 4039 } 4040 4041 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4042 if (VT.is64BitVector() && EltSz == 32) 4043 return false; 4044 4045 return true; 4046 } 4047 4048 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of 4049 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 4050 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>. 4051 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){ 4052 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4053 if (EltSz == 64) 4054 return false; 4055 4056 unsigned NumElts = VT.getVectorNumElements(); 4057 WhichResult = (M[0] == 0 ? 0 : 1); 4058 unsigned Idx = WhichResult * NumElts / 2; 4059 for (unsigned i = 0; i != NumElts; i += 2) { 4060 if ((M[i] >= 0 && (unsigned) M[i] != Idx) || 4061 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx)) 4062 return false; 4063 Idx += 1; 4064 } 4065 4066 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4067 if (VT.is64BitVector() && EltSz == 32) 4068 return false; 4069 4070 return true; 4071 } 4072 4073 // If N is an integer constant that can be moved into a register in one 4074 // instruction, return an SDValue of such a constant (will become a MOV 4075 // instruction). Otherwise return null. 4076 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG, 4077 const ARMSubtarget *ST, DebugLoc dl) { 4078 uint64_t Val; 4079 if (!isa<ConstantSDNode>(N)) 4080 return SDValue(); 4081 Val = cast<ConstantSDNode>(N)->getZExtValue(); 4082 4083 if (ST->isThumb1Only()) { 4084 if (Val <= 255 || ~Val <= 255) 4085 return DAG.getConstant(Val, MVT::i32); 4086 } else { 4087 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1) 4088 return DAG.getConstant(Val, MVT::i32); 4089 } 4090 return SDValue(); 4091 } 4092 4093 // If this is a case we can't handle, return null and let the default 4094 // expansion code take care of it. 4095 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG, 4096 const ARMSubtarget *ST) const { 4097 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode()); 4098 DebugLoc dl = Op.getDebugLoc(); 4099 EVT VT = Op.getValueType(); 4100 4101 APInt SplatBits, SplatUndef; 4102 unsigned SplatBitSize; 4103 bool HasAnyUndefs; 4104 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) { 4105 if (SplatBitSize <= 64) { 4106 // Check if an immediate VMOV works. 4107 EVT VmovVT; 4108 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(), 4109 SplatUndef.getZExtValue(), SplatBitSize, 4110 DAG, VmovVT, VT.is128BitVector(), 4111 VMOVModImm); 4112 if (Val.getNode()) { 4113 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val); 4114 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov); 4115 } 4116 4117 // Try an immediate VMVN. 4118 uint64_t NegatedImm = (~SplatBits).getZExtValue(); 4119 Val = isNEONModifiedImm(NegatedImm, 4120 SplatUndef.getZExtValue(), SplatBitSize, 4121 DAG, VmovVT, VT.is128BitVector(), 4122 VMVNModImm); 4123 if (Val.getNode()) { 4124 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val); 4125 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov); 4126 } 4127 4128 // Use vmov.f32 to materialize other v2f32 and v4f32 splats. 4129 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) { 4130 int ImmVal = ARM_AM::getFP32Imm(SplatBits); 4131 if (ImmVal != -1) { 4132 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32); 4133 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val); 4134 } 4135 } 4136 } 4137 } 4138 4139 // Scan through the operands to see if only one value is used. 4140 unsigned NumElts = VT.getVectorNumElements(); 4141 bool isOnlyLowElement = true; 4142 bool usesOnlyOneValue = true; 4143 bool isConstant = true; 4144 SDValue Value; 4145 for (unsigned i = 0; i < NumElts; ++i) { 4146 SDValue V = Op.getOperand(i); 4147 if (V.getOpcode() == ISD::UNDEF) 4148 continue; 4149 if (i > 0) 4150 isOnlyLowElement = false; 4151 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V)) 4152 isConstant = false; 4153 4154 if (!Value.getNode()) 4155 Value = V; 4156 else if (V != Value) 4157 usesOnlyOneValue = false; 4158 } 4159 4160 if (!Value.getNode()) 4161 return DAG.getUNDEF(VT); 4162 4163 if (isOnlyLowElement) 4164 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value); 4165 4166 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 4167 4168 // Use VDUP for non-constant splats. For f32 constant splats, reduce to 4169 // i32 and try again. 4170 if (usesOnlyOneValue && EltSize <= 32) { 4171 if (!isConstant) 4172 return DAG.getNode(ARMISD::VDUP, dl, VT, Value); 4173 if (VT.getVectorElementType().isFloatingPoint()) { 4174 SmallVector<SDValue, 8> Ops; 4175 for (unsigned i = 0; i < NumElts; ++i) 4176 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32, 4177 Op.getOperand(i))); 4178 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts); 4179 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts); 4180 Val = LowerBUILD_VECTOR(Val, DAG, ST); 4181 if (Val.getNode()) 4182 return DAG.getNode(ISD::BITCAST, dl, VT, Val); 4183 } 4184 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl); 4185 if (Val.getNode()) 4186 return DAG.getNode(ARMISD::VDUP, dl, VT, Val); 4187 } 4188 4189 // If all elements are constants and the case above didn't get hit, fall back 4190 // to the default expansion, which will generate a load from the constant 4191 // pool. 4192 if (isConstant) 4193 return SDValue(); 4194 4195 // Empirical tests suggest this is rarely worth it for vectors of length <= 2. 4196 if (NumElts >= 4) { 4197 SDValue shuffle = ReconstructShuffle(Op, DAG); 4198 if (shuffle != SDValue()) 4199 return shuffle; 4200 } 4201 4202 // Vectors with 32- or 64-bit elements can be built by directly assigning 4203 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands 4204 // will be legalized. 4205 if (EltSize >= 32) { 4206 // Do the expansion with floating-point types, since that is what the VFP 4207 // registers are defined to use, and since i64 is not legal. 4208 EVT EltVT = EVT::getFloatingPointVT(EltSize); 4209 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts); 4210 SmallVector<SDValue, 8> Ops; 4211 for (unsigned i = 0; i < NumElts; ++i) 4212 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i))); 4213 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts); 4214 return DAG.getNode(ISD::BITCAST, dl, VT, Val); 4215 } 4216 4217 return SDValue(); 4218 } 4219 4220 // Gather data to see if the operation can be modelled as a 4221 // shuffle in combination with VEXTs. 4222 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op, 4223 SelectionDAG &DAG) const { 4224 DebugLoc dl = Op.getDebugLoc(); 4225 EVT VT = Op.getValueType(); 4226 unsigned NumElts = VT.getVectorNumElements(); 4227 4228 SmallVector<SDValue, 2> SourceVecs; 4229 SmallVector<unsigned, 2> MinElts; 4230 SmallVector<unsigned, 2> MaxElts; 4231 4232 for (unsigned i = 0; i < NumElts; ++i) { 4233 SDValue V = Op.getOperand(i); 4234 if (V.getOpcode() == ISD::UNDEF) 4235 continue; 4236 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) { 4237 // A shuffle can only come from building a vector from various 4238 // elements of other vectors. 4239 return SDValue(); 4240 } else if (V.getOperand(0).getValueType().getVectorElementType() != 4241 VT.getVectorElementType()) { 4242 // This code doesn't know how to handle shuffles where the vector 4243 // element types do not match (this happens because type legalization 4244 // promotes the return type of EXTRACT_VECTOR_ELT). 4245 // FIXME: It might be appropriate to extend this code to handle 4246 // mismatched types. 4247 return SDValue(); 4248 } 4249 4250 // Record this extraction against the appropriate vector if possible... 4251 SDValue SourceVec = V.getOperand(0); 4252 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue(); 4253 bool FoundSource = false; 4254 for (unsigned j = 0; j < SourceVecs.size(); ++j) { 4255 if (SourceVecs[j] == SourceVec) { 4256 if (MinElts[j] > EltNo) 4257 MinElts[j] = EltNo; 4258 if (MaxElts[j] < EltNo) 4259 MaxElts[j] = EltNo; 4260 FoundSource = true; 4261 break; 4262 } 4263 } 4264 4265 // Or record a new source if not... 4266 if (!FoundSource) { 4267 SourceVecs.push_back(SourceVec); 4268 MinElts.push_back(EltNo); 4269 MaxElts.push_back(EltNo); 4270 } 4271 } 4272 4273 // Currently only do something sane when at most two source vectors 4274 // involved. 4275 if (SourceVecs.size() > 2) 4276 return SDValue(); 4277 4278 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) }; 4279 int VEXTOffsets[2] = {0, 0}; 4280 4281 // This loop extracts the usage patterns of the source vectors 4282 // and prepares appropriate SDValues for a shuffle if possible. 4283 for (unsigned i = 0; i < SourceVecs.size(); ++i) { 4284 if (SourceVecs[i].getValueType() == VT) { 4285 // No VEXT necessary 4286 ShuffleSrcs[i] = SourceVecs[i]; 4287 VEXTOffsets[i] = 0; 4288 continue; 4289 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) { 4290 // It probably isn't worth padding out a smaller vector just to 4291 // break it down again in a shuffle. 4292 return SDValue(); 4293 } 4294 4295 // Since only 64-bit and 128-bit vectors are legal on ARM and 4296 // we've eliminated the other cases... 4297 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts && 4298 "unexpected vector sizes in ReconstructShuffle"); 4299 4300 if (MaxElts[i] - MinElts[i] >= NumElts) { 4301 // Span too large for a VEXT to cope 4302 return SDValue(); 4303 } 4304 4305 if (MinElts[i] >= NumElts) { 4306 // The extraction can just take the second half 4307 VEXTOffsets[i] = NumElts; 4308 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4309 SourceVecs[i], 4310 DAG.getIntPtrConstant(NumElts)); 4311 } else if (MaxElts[i] < NumElts) { 4312 // The extraction can just take the first half 4313 VEXTOffsets[i] = 0; 4314 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4315 SourceVecs[i], 4316 DAG.getIntPtrConstant(0)); 4317 } else { 4318 // An actual VEXT is needed 4319 VEXTOffsets[i] = MinElts[i]; 4320 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4321 SourceVecs[i], 4322 DAG.getIntPtrConstant(0)); 4323 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4324 SourceVecs[i], 4325 DAG.getIntPtrConstant(NumElts)); 4326 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2, 4327 DAG.getConstant(VEXTOffsets[i], MVT::i32)); 4328 } 4329 } 4330 4331 SmallVector<int, 8> Mask; 4332 4333 for (unsigned i = 0; i < NumElts; ++i) { 4334 SDValue Entry = Op.getOperand(i); 4335 if (Entry.getOpcode() == ISD::UNDEF) { 4336 Mask.push_back(-1); 4337 continue; 4338 } 4339 4340 SDValue ExtractVec = Entry.getOperand(0); 4341 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i) 4342 .getOperand(1))->getSExtValue(); 4343 if (ExtractVec == SourceVecs[0]) { 4344 Mask.push_back(ExtractElt - VEXTOffsets[0]); 4345 } else { 4346 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]); 4347 } 4348 } 4349 4350 // Final check before we try to produce nonsense... 4351 if (isShuffleMaskLegal(Mask, VT)) 4352 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1], 4353 &Mask[0]); 4354 4355 return SDValue(); 4356 } 4357 4358 /// isShuffleMaskLegal - Targets can use this to indicate that they only 4359 /// support *some* VECTOR_SHUFFLE operations, those with specific masks. 4360 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values 4361 /// are assumed to be legal. 4362 bool 4363 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M, 4364 EVT VT) const { 4365 if (VT.getVectorNumElements() == 4 && 4366 (VT.is128BitVector() || VT.is64BitVector())) { 4367 unsigned PFIndexes[4]; 4368 for (unsigned i = 0; i != 4; ++i) { 4369 if (M[i] < 0) 4370 PFIndexes[i] = 8; 4371 else 4372 PFIndexes[i] = M[i]; 4373 } 4374 4375 // Compute the index in the perfect shuffle table. 4376 unsigned PFTableIndex = 4377 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3]; 4378 unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; 4379 unsigned Cost = (PFEntry >> 30); 4380 4381 if (Cost <= 4) 4382 return true; 4383 } 4384 4385 bool ReverseVEXT; 4386 unsigned Imm, WhichResult; 4387 4388 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 4389 return (EltSize >= 32 || 4390 ShuffleVectorSDNode::isSplatMask(&M[0], VT) || 4391 isVREVMask(M, VT, 64) || 4392 isVREVMask(M, VT, 32) || 4393 isVREVMask(M, VT, 16) || 4394 isVEXTMask(M, VT, ReverseVEXT, Imm) || 4395 isVTBLMask(M, VT) || 4396 isVTRNMask(M, VT, WhichResult) || 4397 isVUZPMask(M, VT, WhichResult) || 4398 isVZIPMask(M, VT, WhichResult) || 4399 isVTRN_v_undef_Mask(M, VT, WhichResult) || 4400 isVUZP_v_undef_Mask(M, VT, WhichResult) || 4401 isVZIP_v_undef_Mask(M, VT, WhichResult)); 4402 } 4403 4404 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit 4405 /// the specified operations to build the shuffle. 4406 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS, 4407 SDValue RHS, SelectionDAG &DAG, 4408 DebugLoc dl) { 4409 unsigned OpNum = (PFEntry >> 26) & 0x0F; 4410 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1); 4411 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1); 4412 4413 enum { 4414 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3> 4415 OP_VREV, 4416 OP_VDUP0, 4417 OP_VDUP1, 4418 OP_VDUP2, 4419 OP_VDUP3, 4420 OP_VEXT1, 4421 OP_VEXT2, 4422 OP_VEXT3, 4423 OP_VUZPL, // VUZP, left result 4424 OP_VUZPR, // VUZP, right result 4425 OP_VZIPL, // VZIP, left result 4426 OP_VZIPR, // VZIP, right result 4427 OP_VTRNL, // VTRN, left result 4428 OP_VTRNR // VTRN, right result 4429 }; 4430 4431 if (OpNum == OP_COPY) { 4432 if (LHSID == (1*9+2)*9+3) return LHS; 4433 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!"); 4434 return RHS; 4435 } 4436 4437 SDValue OpLHS, OpRHS; 4438 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl); 4439 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl); 4440 EVT VT = OpLHS.getValueType(); 4441 4442 switch (OpNum) { 4443 default: llvm_unreachable("Unknown shuffle opcode!"); 4444 case OP_VREV: 4445 // VREV divides the vector in half and swaps within the half. 4446 if (VT.getVectorElementType() == MVT::i32 || 4447 VT.getVectorElementType() == MVT::f32) 4448 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS); 4449 // vrev <4 x i16> -> VREV32 4450 if (VT.getVectorElementType() == MVT::i16) 4451 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS); 4452 // vrev <4 x i8> -> VREV16 4453 assert(VT.getVectorElementType() == MVT::i8); 4454 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS); 4455 case OP_VDUP0: 4456 case OP_VDUP1: 4457 case OP_VDUP2: 4458 case OP_VDUP3: 4459 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, 4460 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32)); 4461 case OP_VEXT1: 4462 case OP_VEXT2: 4463 case OP_VEXT3: 4464 return DAG.getNode(ARMISD::VEXT, dl, VT, 4465 OpLHS, OpRHS, 4466 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32)); 4467 case OP_VUZPL: 4468 case OP_VUZPR: 4469 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 4470 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL); 4471 case OP_VZIPL: 4472 case OP_VZIPR: 4473 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 4474 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL); 4475 case OP_VTRNL: 4476 case OP_VTRNR: 4477 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 4478 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL); 4479 } 4480 } 4481 4482 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op, 4483 ArrayRef<int> ShuffleMask, 4484 SelectionDAG &DAG) { 4485 // Check to see if we can use the VTBL instruction. 4486 SDValue V1 = Op.getOperand(0); 4487 SDValue V2 = Op.getOperand(1); 4488 DebugLoc DL = Op.getDebugLoc(); 4489 4490 SmallVector<SDValue, 8> VTBLMask; 4491 for (ArrayRef<int>::iterator 4492 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I) 4493 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32)); 4494 4495 if (V2.getNode()->getOpcode() == ISD::UNDEF) 4496 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1, 4497 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, 4498 &VTBLMask[0], 8)); 4499 4500 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2, 4501 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, 4502 &VTBLMask[0], 8)); 4503 } 4504 4505 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) { 4506 SDValue V1 = Op.getOperand(0); 4507 SDValue V2 = Op.getOperand(1); 4508 DebugLoc dl = Op.getDebugLoc(); 4509 EVT VT = Op.getValueType(); 4510 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode()); 4511 4512 // Convert shuffles that are directly supported on NEON to target-specific 4513 // DAG nodes, instead of keeping them as shuffles and matching them again 4514 // during code selection. This is more efficient and avoids the possibility 4515 // of inconsistencies between legalization and selection. 4516 // FIXME: floating-point vectors should be canonicalized to integer vectors 4517 // of the same time so that they get CSEd properly. 4518 ArrayRef<int> ShuffleMask = SVN->getMask(); 4519 4520 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 4521 if (EltSize <= 32) { 4522 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) { 4523 int Lane = SVN->getSplatIndex(); 4524 // If this is undef splat, generate it via "just" vdup, if possible. 4525 if (Lane == -1) Lane = 0; 4526 4527 // Test if V1 is a SCALAR_TO_VECTOR. 4528 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) { 4529 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0)); 4530 } 4531 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR 4532 // (and probably will turn into a SCALAR_TO_VECTOR once legalization 4533 // reaches it). 4534 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR && 4535 !isa<ConstantSDNode>(V1.getOperand(0))) { 4536 bool IsScalarToVector = true; 4537 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i) 4538 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) { 4539 IsScalarToVector = false; 4540 break; 4541 } 4542 if (IsScalarToVector) 4543 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0)); 4544 } 4545 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1, 4546 DAG.getConstant(Lane, MVT::i32)); 4547 } 4548 4549 bool ReverseVEXT; 4550 unsigned Imm; 4551 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) { 4552 if (ReverseVEXT) 4553 std::swap(V1, V2); 4554 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2, 4555 DAG.getConstant(Imm, MVT::i32)); 4556 } 4557 4558 if (isVREVMask(ShuffleMask, VT, 64)) 4559 return DAG.getNode(ARMISD::VREV64, dl, VT, V1); 4560 if (isVREVMask(ShuffleMask, VT, 32)) 4561 return DAG.getNode(ARMISD::VREV32, dl, VT, V1); 4562 if (isVREVMask(ShuffleMask, VT, 16)) 4563 return DAG.getNode(ARMISD::VREV16, dl, VT, V1); 4564 4565 // Check for Neon shuffles that modify both input vectors in place. 4566 // If both results are used, i.e., if there are two shuffles with the same 4567 // source operands and with masks corresponding to both results of one of 4568 // these operations, DAG memoization will ensure that a single node is 4569 // used for both shuffles. 4570 unsigned WhichResult; 4571 if (isVTRNMask(ShuffleMask, VT, WhichResult)) 4572 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 4573 V1, V2).getValue(WhichResult); 4574 if (isVUZPMask(ShuffleMask, VT, WhichResult)) 4575 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 4576 V1, V2).getValue(WhichResult); 4577 if (isVZIPMask(ShuffleMask, VT, WhichResult)) 4578 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 4579 V1, V2).getValue(WhichResult); 4580 4581 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) 4582 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 4583 V1, V1).getValue(WhichResult); 4584 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) 4585 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 4586 V1, V1).getValue(WhichResult); 4587 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) 4588 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 4589 V1, V1).getValue(WhichResult); 4590 } 4591 4592 // If the shuffle is not directly supported and it has 4 elements, use 4593 // the PerfectShuffle-generated table to synthesize it from other shuffles. 4594 unsigned NumElts = VT.getVectorNumElements(); 4595 if (NumElts == 4) { 4596 unsigned PFIndexes[4]; 4597 for (unsigned i = 0; i != 4; ++i) { 4598 if (ShuffleMask[i] < 0) 4599 PFIndexes[i] = 8; 4600 else 4601 PFIndexes[i] = ShuffleMask[i]; 4602 } 4603 4604 // Compute the index in the perfect shuffle table. 4605 unsigned PFTableIndex = 4606 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3]; 4607 unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; 4608 unsigned Cost = (PFEntry >> 30); 4609 4610 if (Cost <= 4) 4611 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl); 4612 } 4613 4614 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs. 4615 if (EltSize >= 32) { 4616 // Do the expansion with floating-point types, since that is what the VFP 4617 // registers are defined to use, and since i64 is not legal. 4618 EVT EltVT = EVT::getFloatingPointVT(EltSize); 4619 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts); 4620 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1); 4621 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2); 4622 SmallVector<SDValue, 8> Ops; 4623 for (unsigned i = 0; i < NumElts; ++i) { 4624 if (ShuffleMask[i] < 0) 4625 Ops.push_back(DAG.getUNDEF(EltVT)); 4626 else 4627 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, 4628 ShuffleMask[i] < (int)NumElts ? V1 : V2, 4629 DAG.getConstant(ShuffleMask[i] & (NumElts-1), 4630 MVT::i32))); 4631 } 4632 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts); 4633 return DAG.getNode(ISD::BITCAST, dl, VT, Val); 4634 } 4635 4636 if (VT == MVT::v8i8) { 4637 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG); 4638 if (NewOp.getNode()) 4639 return NewOp; 4640 } 4641 4642 return SDValue(); 4643 } 4644 4645 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) { 4646 // INSERT_VECTOR_ELT is legal only for immediate indexes. 4647 SDValue Lane = Op.getOperand(2); 4648 if (!isa<ConstantSDNode>(Lane)) 4649 return SDValue(); 4650 4651 return Op; 4652 } 4653 4654 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) { 4655 // EXTRACT_VECTOR_ELT is legal only for immediate indexes. 4656 SDValue Lane = Op.getOperand(1); 4657 if (!isa<ConstantSDNode>(Lane)) 4658 return SDValue(); 4659 4660 SDValue Vec = Op.getOperand(0); 4661 if (Op.getValueType() == MVT::i32 && 4662 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) { 4663 DebugLoc dl = Op.getDebugLoc(); 4664 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane); 4665 } 4666 4667 return Op; 4668 } 4669 4670 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) { 4671 // The only time a CONCAT_VECTORS operation can have legal types is when 4672 // two 64-bit vectors are concatenated to a 128-bit vector. 4673 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 && 4674 "unexpected CONCAT_VECTORS"); 4675 DebugLoc dl = Op.getDebugLoc(); 4676 SDValue Val = DAG.getUNDEF(MVT::v2f64); 4677 SDValue Op0 = Op.getOperand(0); 4678 SDValue Op1 = Op.getOperand(1); 4679 if (Op0.getOpcode() != ISD::UNDEF) 4680 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val, 4681 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0), 4682 DAG.getIntPtrConstant(0)); 4683 if (Op1.getOpcode() != ISD::UNDEF) 4684 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val, 4685 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1), 4686 DAG.getIntPtrConstant(1)); 4687 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val); 4688 } 4689 4690 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each 4691 /// element has been zero/sign-extended, depending on the isSigned parameter, 4692 /// from an integer type half its size. 4693 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG, 4694 bool isSigned) { 4695 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32. 4696 EVT VT = N->getValueType(0); 4697 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) { 4698 SDNode *BVN = N->getOperand(0).getNode(); 4699 if (BVN->getValueType(0) != MVT::v4i32 || 4700 BVN->getOpcode() != ISD::BUILD_VECTOR) 4701 return false; 4702 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0; 4703 unsigned HiElt = 1 - LoElt; 4704 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt)); 4705 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt)); 4706 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2)); 4707 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2)); 4708 if (!Lo0 || !Hi0 || !Lo1 || !Hi1) 4709 return false; 4710 if (isSigned) { 4711 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 && 4712 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32) 4713 return true; 4714 } else { 4715 if (Hi0->isNullValue() && Hi1->isNullValue()) 4716 return true; 4717 } 4718 return false; 4719 } 4720 4721 if (N->getOpcode() != ISD::BUILD_VECTOR) 4722 return false; 4723 4724 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 4725 SDNode *Elt = N->getOperand(i).getNode(); 4726 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) { 4727 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 4728 unsigned HalfSize = EltSize / 2; 4729 if (isSigned) { 4730 if (!isIntN(HalfSize, C->getSExtValue())) 4731 return false; 4732 } else { 4733 if (!isUIntN(HalfSize, C->getZExtValue())) 4734 return false; 4735 } 4736 continue; 4737 } 4738 return false; 4739 } 4740 4741 return true; 4742 } 4743 4744 /// isSignExtended - Check if a node is a vector value that is sign-extended 4745 /// or a constant BUILD_VECTOR with sign-extended elements. 4746 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) { 4747 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N)) 4748 return true; 4749 if (isExtendedBUILD_VECTOR(N, DAG, true)) 4750 return true; 4751 return false; 4752 } 4753 4754 /// isZeroExtended - Check if a node is a vector value that is zero-extended 4755 /// or a constant BUILD_VECTOR with zero-extended elements. 4756 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) { 4757 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N)) 4758 return true; 4759 if (isExtendedBUILD_VECTOR(N, DAG, false)) 4760 return true; 4761 return false; 4762 } 4763 4764 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending 4765 /// load, or BUILD_VECTOR with extended elements, return the unextended value. 4766 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) { 4767 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND) 4768 return N->getOperand(0); 4769 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) 4770 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(), 4771 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(), 4772 LD->isNonTemporal(), LD->isInvariant(), 4773 LD->getAlignment()); 4774 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will 4775 // have been legalized as a BITCAST from v4i32. 4776 if (N->getOpcode() == ISD::BITCAST) { 4777 SDNode *BVN = N->getOperand(0).getNode(); 4778 assert(BVN->getOpcode() == ISD::BUILD_VECTOR && 4779 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR"); 4780 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0; 4781 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32, 4782 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2)); 4783 } 4784 // Construct a new BUILD_VECTOR with elements truncated to half the size. 4785 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR"); 4786 EVT VT = N->getValueType(0); 4787 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2; 4788 unsigned NumElts = VT.getVectorNumElements(); 4789 MVT TruncVT = MVT::getIntegerVT(EltSize); 4790 SmallVector<SDValue, 8> Ops; 4791 for (unsigned i = 0; i != NumElts; ++i) { 4792 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i)); 4793 const APInt &CInt = C->getAPIntValue(); 4794 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT)); 4795 } 4796 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), 4797 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts); 4798 } 4799 4800 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) { 4801 unsigned Opcode = N->getOpcode(); 4802 if (Opcode == ISD::ADD || Opcode == ISD::SUB) { 4803 SDNode *N0 = N->getOperand(0).getNode(); 4804 SDNode *N1 = N->getOperand(1).getNode(); 4805 return N0->hasOneUse() && N1->hasOneUse() && 4806 isSignExtended(N0, DAG) && isSignExtended(N1, DAG); 4807 } 4808 return false; 4809 } 4810 4811 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) { 4812 unsigned Opcode = N->getOpcode(); 4813 if (Opcode == ISD::ADD || Opcode == ISD::SUB) { 4814 SDNode *N0 = N->getOperand(0).getNode(); 4815 SDNode *N1 = N->getOperand(1).getNode(); 4816 return N0->hasOneUse() && N1->hasOneUse() && 4817 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG); 4818 } 4819 return false; 4820 } 4821 4822 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) { 4823 // Multiplications are only custom-lowered for 128-bit vectors so that 4824 // VMULL can be detected. Otherwise v2i64 multiplications are not legal. 4825 EVT VT = Op.getValueType(); 4826 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL"); 4827 SDNode *N0 = Op.getOperand(0).getNode(); 4828 SDNode *N1 = Op.getOperand(1).getNode(); 4829 unsigned NewOpc = 0; 4830 bool isMLA = false; 4831 bool isN0SExt = isSignExtended(N0, DAG); 4832 bool isN1SExt = isSignExtended(N1, DAG); 4833 if (isN0SExt && isN1SExt) 4834 NewOpc = ARMISD::VMULLs; 4835 else { 4836 bool isN0ZExt = isZeroExtended(N0, DAG); 4837 bool isN1ZExt = isZeroExtended(N1, DAG); 4838 if (isN0ZExt && isN1ZExt) 4839 NewOpc = ARMISD::VMULLu; 4840 else if (isN1SExt || isN1ZExt) { 4841 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these 4842 // into (s/zext A * s/zext C) + (s/zext B * s/zext C) 4843 if (isN1SExt && isAddSubSExt(N0, DAG)) { 4844 NewOpc = ARMISD::VMULLs; 4845 isMLA = true; 4846 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) { 4847 NewOpc = ARMISD::VMULLu; 4848 isMLA = true; 4849 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) { 4850 std::swap(N0, N1); 4851 NewOpc = ARMISD::VMULLu; 4852 isMLA = true; 4853 } 4854 } 4855 4856 if (!NewOpc) { 4857 if (VT == MVT::v2i64) 4858 // Fall through to expand this. It is not legal. 4859 return SDValue(); 4860 else 4861 // Other vector multiplications are legal. 4862 return Op; 4863 } 4864 } 4865 4866 // Legalize to a VMULL instruction. 4867 DebugLoc DL = Op.getDebugLoc(); 4868 SDValue Op0; 4869 SDValue Op1 = SkipExtension(N1, DAG); 4870 if (!isMLA) { 4871 Op0 = SkipExtension(N0, DAG); 4872 assert(Op0.getValueType().is64BitVector() && 4873 Op1.getValueType().is64BitVector() && 4874 "unexpected types for extended operands to VMULL"); 4875 return DAG.getNode(NewOpc, DL, VT, Op0, Op1); 4876 } 4877 4878 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during 4879 // isel lowering to take advantage of no-stall back to back vmul + vmla. 4880 // vmull q0, d4, d6 4881 // vmlal q0, d5, d6 4882 // is faster than 4883 // vaddl q0, d4, d5 4884 // vmovl q1, d6 4885 // vmul q0, q0, q1 4886 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG); 4887 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG); 4888 EVT Op1VT = Op1.getValueType(); 4889 return DAG.getNode(N0->getOpcode(), DL, VT, 4890 DAG.getNode(NewOpc, DL, VT, 4891 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1), 4892 DAG.getNode(NewOpc, DL, VT, 4893 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1)); 4894 } 4895 4896 static SDValue 4897 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) { 4898 // Convert to float 4899 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo)); 4900 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo)); 4901 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X); 4902 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y); 4903 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X); 4904 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y); 4905 // Get reciprocal estimate. 4906 // float4 recip = vrecpeq_f32(yf); 4907 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 4908 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y); 4909 // Because char has a smaller range than uchar, we can actually get away 4910 // without any newton steps. This requires that we use a weird bias 4911 // of 0xb000, however (again, this has been exhaustively tested). 4912 // float4 result = as_float4(as_int4(xf*recip) + 0xb000); 4913 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y); 4914 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X); 4915 Y = DAG.getConstant(0xb000, MVT::i32); 4916 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y); 4917 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y); 4918 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X); 4919 // Convert back to short. 4920 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X); 4921 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X); 4922 return X; 4923 } 4924 4925 static SDValue 4926 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) { 4927 SDValue N2; 4928 // Convert to float. 4929 // float4 yf = vcvt_f32_s32(vmovl_s16(y)); 4930 // float4 xf = vcvt_f32_s32(vmovl_s16(x)); 4931 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0); 4932 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1); 4933 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0); 4934 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1); 4935 4936 // Use reciprocal estimate and one refinement step. 4937 // float4 recip = vrecpeq_f32(yf); 4938 // recip *= vrecpsq_f32(yf, recip); 4939 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 4940 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1); 4941 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 4942 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), 4943 N1, N2); 4944 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); 4945 // Because short has a smaller range than ushort, we can actually get away 4946 // with only a single newton step. This requires that we use a weird bias 4947 // of 89, however (again, this has been exhaustively tested). 4948 // float4 result = as_float4(as_int4(xf*recip) + 0x89); 4949 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2); 4950 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0); 4951 N1 = DAG.getConstant(0x89, MVT::i32); 4952 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1); 4953 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1); 4954 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0); 4955 // Convert back to integer and return. 4956 // return vmovn_s32(vcvt_s32_f32(result)); 4957 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0); 4958 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0); 4959 return N0; 4960 } 4961 4962 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) { 4963 EVT VT = Op.getValueType(); 4964 assert((VT == MVT::v4i16 || VT == MVT::v8i8) && 4965 "unexpected type for custom-lowering ISD::SDIV"); 4966 4967 DebugLoc dl = Op.getDebugLoc(); 4968 SDValue N0 = Op.getOperand(0); 4969 SDValue N1 = Op.getOperand(1); 4970 SDValue N2, N3; 4971 4972 if (VT == MVT::v8i8) { 4973 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0); 4974 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1); 4975 4976 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 4977 DAG.getIntPtrConstant(4)); 4978 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 4979 DAG.getIntPtrConstant(4)); 4980 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 4981 DAG.getIntPtrConstant(0)); 4982 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 4983 DAG.getIntPtrConstant(0)); 4984 4985 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16 4986 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16 4987 4988 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2); 4989 N0 = LowerCONCAT_VECTORS(N0, DAG); 4990 4991 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0); 4992 return N0; 4993 } 4994 return LowerSDIV_v4i16(N0, N1, dl, DAG); 4995 } 4996 4997 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) { 4998 EVT VT = Op.getValueType(); 4999 assert((VT == MVT::v4i16 || VT == MVT::v8i8) && 5000 "unexpected type for custom-lowering ISD::UDIV"); 5001 5002 DebugLoc dl = Op.getDebugLoc(); 5003 SDValue N0 = Op.getOperand(0); 5004 SDValue N1 = Op.getOperand(1); 5005 SDValue N2, N3; 5006 5007 if (VT == MVT::v8i8) { 5008 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0); 5009 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1); 5010 5011 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5012 DAG.getIntPtrConstant(4)); 5013 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5014 DAG.getIntPtrConstant(4)); 5015 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5016 DAG.getIntPtrConstant(0)); 5017 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5018 DAG.getIntPtrConstant(0)); 5019 5020 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16 5021 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16 5022 5023 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2); 5024 N0 = LowerCONCAT_VECTORS(N0, DAG); 5025 5026 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8, 5027 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32), 5028 N0); 5029 return N0; 5030 } 5031 5032 // v4i16 sdiv ... Convert to float. 5033 // float4 yf = vcvt_f32_s32(vmovl_u16(y)); 5034 // float4 xf = vcvt_f32_s32(vmovl_u16(x)); 5035 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0); 5036 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1); 5037 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0); 5038 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1); 5039 5040 // Use reciprocal estimate and two refinement steps. 5041 // float4 recip = vrecpeq_f32(yf); 5042 // recip *= vrecpsq_f32(yf, recip); 5043 // recip *= vrecpsq_f32(yf, recip); 5044 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5045 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1); 5046 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5047 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), 5048 BN1, N2); 5049 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); 5050 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5051 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), 5052 BN1, N2); 5053 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); 5054 // Simply multiplying by the reciprocal estimate can leave us a few ulps 5055 // too low, so we add 2 ulps (exhaustive testing shows that this is enough, 5056 // and that it will never cause us to return an answer too large). 5057 // float4 result = as_float4(as_int4(xf*recip) + 2); 5058 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2); 5059 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0); 5060 N1 = DAG.getConstant(2, MVT::i32); 5061 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1); 5062 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1); 5063 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0); 5064 // Convert back to integer and return. 5065 // return vmovn_u32(vcvt_s32_f32(result)); 5066 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0); 5067 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0); 5068 return N0; 5069 } 5070 5071 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) { 5072 EVT VT = Op.getNode()->getValueType(0); 5073 SDVTList VTs = DAG.getVTList(VT, MVT::i32); 5074 5075 unsigned Opc; 5076 bool ExtraOp = false; 5077 switch (Op.getOpcode()) { 5078 default: llvm_unreachable("Invalid code"); 5079 case ISD::ADDC: Opc = ARMISD::ADDC; break; 5080 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break; 5081 case ISD::SUBC: Opc = ARMISD::SUBC; break; 5082 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break; 5083 } 5084 5085 if (!ExtraOp) 5086 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0), 5087 Op.getOperand(1)); 5088 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0), 5089 Op.getOperand(1), Op.getOperand(2)); 5090 } 5091 5092 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) { 5093 // Monotonic load/store is legal for all targets 5094 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic) 5095 return Op; 5096 5097 // Aquire/Release load/store is not legal for targets without a 5098 // dmb or equivalent available. 5099 return SDValue(); 5100 } 5101 5102 5103 static void 5104 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results, 5105 SelectionDAG &DAG, unsigned NewOp) { 5106 DebugLoc dl = Node->getDebugLoc(); 5107 assert (Node->getValueType(0) == MVT::i64 && 5108 "Only know how to expand i64 atomics"); 5109 5110 SmallVector<SDValue, 6> Ops; 5111 Ops.push_back(Node->getOperand(0)); // Chain 5112 Ops.push_back(Node->getOperand(1)); // Ptr 5113 // Low part of Val1 5114 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5115 Node->getOperand(2), DAG.getIntPtrConstant(0))); 5116 // High part of Val1 5117 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5118 Node->getOperand(2), DAG.getIntPtrConstant(1))); 5119 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) { 5120 // High part of Val1 5121 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5122 Node->getOperand(3), DAG.getIntPtrConstant(0))); 5123 // High part of Val2 5124 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5125 Node->getOperand(3), DAG.getIntPtrConstant(1))); 5126 } 5127 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other); 5128 SDValue Result = 5129 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64, 5130 cast<MemSDNode>(Node)->getMemOperand()); 5131 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) }; 5132 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2)); 5133 Results.push_back(Result.getValue(2)); 5134 } 5135 5136 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { 5137 switch (Op.getOpcode()) { 5138 default: llvm_unreachable("Don't know how to custom lower this!"); 5139 case ISD::ConstantPool: return LowerConstantPool(Op, DAG); 5140 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); 5141 case ISD::GlobalAddress: 5142 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) : 5143 LowerGlobalAddressELF(Op, DAG); 5144 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); 5145 case ISD::SELECT: return LowerSELECT(Op, DAG); 5146 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG); 5147 case ISD::BR_CC: return LowerBR_CC(Op, DAG); 5148 case ISD::BR_JT: return LowerBR_JT(Op, DAG); 5149 case ISD::VASTART: return LowerVASTART(Op, DAG); 5150 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget); 5151 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget); 5152 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget); 5153 case ISD::SINT_TO_FP: 5154 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG); 5155 case ISD::FP_TO_SINT: 5156 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG); 5157 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG); 5158 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG); 5159 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); 5160 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG); 5161 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG); 5162 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG); 5163 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG, 5164 Subtarget); 5165 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG); 5166 case ISD::SHL: 5167 case ISD::SRL: 5168 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget); 5169 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG); 5170 case ISD::SRL_PARTS: 5171 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG); 5172 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget); 5173 case ISD::SETCC: return LowerVSETCC(Op, DAG); 5174 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget); 5175 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget); 5176 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG); 5177 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG); 5178 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG); 5179 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG); 5180 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG); 5181 case ISD::MUL: return LowerMUL(Op, DAG); 5182 case ISD::SDIV: return LowerSDIV(Op, DAG); 5183 case ISD::UDIV: return LowerUDIV(Op, DAG); 5184 case ISD::ADDC: 5185 case ISD::ADDE: 5186 case ISD::SUBC: 5187 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG); 5188 case ISD::ATOMIC_LOAD: 5189 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG); 5190 } 5191 } 5192 5193 /// ReplaceNodeResults - Replace the results of node with an illegal result 5194 /// type with new values built out of custom code. 5195 void ARMTargetLowering::ReplaceNodeResults(SDNode *N, 5196 SmallVectorImpl<SDValue>&Results, 5197 SelectionDAG &DAG) const { 5198 SDValue Res; 5199 switch (N->getOpcode()) { 5200 default: 5201 llvm_unreachable("Don't know how to custom expand this!"); 5202 case ISD::BITCAST: 5203 Res = ExpandBITCAST(N, DAG); 5204 break; 5205 case ISD::SRL: 5206 case ISD::SRA: 5207 Res = Expand64BitShift(N, DAG, Subtarget); 5208 break; 5209 case ISD::ATOMIC_LOAD_ADD: 5210 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG); 5211 return; 5212 case ISD::ATOMIC_LOAD_AND: 5213 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG); 5214 return; 5215 case ISD::ATOMIC_LOAD_NAND: 5216 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG); 5217 return; 5218 case ISD::ATOMIC_LOAD_OR: 5219 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG); 5220 return; 5221 case ISD::ATOMIC_LOAD_SUB: 5222 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG); 5223 return; 5224 case ISD::ATOMIC_LOAD_XOR: 5225 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG); 5226 return; 5227 case ISD::ATOMIC_SWAP: 5228 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG); 5229 return; 5230 case ISD::ATOMIC_CMP_SWAP: 5231 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG); 5232 return; 5233 } 5234 if (Res.getNode()) 5235 Results.push_back(Res); 5236 } 5237 5238 //===----------------------------------------------------------------------===// 5239 // ARM Scheduler Hooks 5240 //===----------------------------------------------------------------------===// 5241 5242 MachineBasicBlock * 5243 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI, 5244 MachineBasicBlock *BB, 5245 unsigned Size) const { 5246 unsigned dest = MI->getOperand(0).getReg(); 5247 unsigned ptr = MI->getOperand(1).getReg(); 5248 unsigned oldval = MI->getOperand(2).getReg(); 5249 unsigned newval = MI->getOperand(3).getReg(); 5250 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5251 DebugLoc dl = MI->getDebugLoc(); 5252 bool isThumb2 = Subtarget->isThumb2(); 5253 5254 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 5255 unsigned scratch = 5256 MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass 5257 : ARM::GPRRegisterClass); 5258 5259 if (isThumb2) { 5260 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass); 5261 MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass); 5262 MRI.constrainRegClass(newval, ARM::rGPRRegisterClass); 5263 } 5264 5265 unsigned ldrOpc, strOpc; 5266 switch (Size) { 5267 default: llvm_unreachable("unsupported size for AtomicCmpSwap!"); 5268 case 1: 5269 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB; 5270 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB; 5271 break; 5272 case 2: 5273 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH; 5274 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH; 5275 break; 5276 case 4: 5277 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX; 5278 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX; 5279 break; 5280 } 5281 5282 MachineFunction *MF = BB->getParent(); 5283 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 5284 MachineFunction::iterator It = BB; 5285 ++It; // insert the new blocks after the current block 5286 5287 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 5288 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 5289 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5290 MF->insert(It, loop1MBB); 5291 MF->insert(It, loop2MBB); 5292 MF->insert(It, exitMBB); 5293 5294 // Transfer the remainder of BB and its successor edges to exitMBB. 5295 exitMBB->splice(exitMBB->begin(), BB, 5296 llvm::next(MachineBasicBlock::iterator(MI)), 5297 BB->end()); 5298 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 5299 5300 // thisMBB: 5301 // ... 5302 // fallthrough --> loop1MBB 5303 BB->addSuccessor(loop1MBB); 5304 5305 // loop1MBB: 5306 // ldrex dest, [ptr] 5307 // cmp dest, oldval 5308 // bne exitMBB 5309 BB = loop1MBB; 5310 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr); 5311 if (ldrOpc == ARM::t2LDREX) 5312 MIB.addImm(0); 5313 AddDefaultPred(MIB); 5314 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) 5315 .addReg(dest).addReg(oldval)); 5316 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5317 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5318 BB->addSuccessor(loop2MBB); 5319 BB->addSuccessor(exitMBB); 5320 5321 // loop2MBB: 5322 // strex scratch, newval, [ptr] 5323 // cmp scratch, #0 5324 // bne loop1MBB 5325 BB = loop2MBB; 5326 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr); 5327 if (strOpc == ARM::t2STREX) 5328 MIB.addImm(0); 5329 AddDefaultPred(MIB); 5330 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 5331 .addReg(scratch).addImm(0)); 5332 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5333 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5334 BB->addSuccessor(loop1MBB); 5335 BB->addSuccessor(exitMBB); 5336 5337 // exitMBB: 5338 // ... 5339 BB = exitMBB; 5340 5341 MI->eraseFromParent(); // The instruction is gone now. 5342 5343 return BB; 5344 } 5345 5346 MachineBasicBlock * 5347 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB, 5348 unsigned Size, unsigned BinOpcode) const { 5349 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0. 5350 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5351 5352 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 5353 MachineFunction *MF = BB->getParent(); 5354 MachineFunction::iterator It = BB; 5355 ++It; 5356 5357 unsigned dest = MI->getOperand(0).getReg(); 5358 unsigned ptr = MI->getOperand(1).getReg(); 5359 unsigned incr = MI->getOperand(2).getReg(); 5360 DebugLoc dl = MI->getDebugLoc(); 5361 bool isThumb2 = Subtarget->isThumb2(); 5362 5363 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 5364 if (isThumb2) { 5365 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass); 5366 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass); 5367 } 5368 5369 unsigned ldrOpc, strOpc; 5370 switch (Size) { 5371 default: llvm_unreachable("unsupported size for AtomicCmpSwap!"); 5372 case 1: 5373 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB; 5374 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB; 5375 break; 5376 case 2: 5377 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH; 5378 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH; 5379 break; 5380 case 4: 5381 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX; 5382 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX; 5383 break; 5384 } 5385 5386 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5387 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5388 MF->insert(It, loopMBB); 5389 MF->insert(It, exitMBB); 5390 5391 // Transfer the remainder of BB and its successor edges to exitMBB. 5392 exitMBB->splice(exitMBB->begin(), BB, 5393 llvm::next(MachineBasicBlock::iterator(MI)), 5394 BB->end()); 5395 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 5396 5397 const TargetRegisterClass *TRC = 5398 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass; 5399 unsigned scratch = MRI.createVirtualRegister(TRC); 5400 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC); 5401 5402 // thisMBB: 5403 // ... 5404 // fallthrough --> loopMBB 5405 BB->addSuccessor(loopMBB); 5406 5407 // loopMBB: 5408 // ldrex dest, ptr 5409 // <binop> scratch2, dest, incr 5410 // strex scratch, scratch2, ptr 5411 // cmp scratch, #0 5412 // bne- loopMBB 5413 // fallthrough --> exitMBB 5414 BB = loopMBB; 5415 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr); 5416 if (ldrOpc == ARM::t2LDREX) 5417 MIB.addImm(0); 5418 AddDefaultPred(MIB); 5419 if (BinOpcode) { 5420 // operand order needs to go the other way for NAND 5421 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr) 5422 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2). 5423 addReg(incr).addReg(dest)).addReg(0); 5424 else 5425 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2). 5426 addReg(dest).addReg(incr)).addReg(0); 5427 } 5428 5429 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr); 5430 if (strOpc == ARM::t2STREX) 5431 MIB.addImm(0); 5432 AddDefaultPred(MIB); 5433 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 5434 .addReg(scratch).addImm(0)); 5435 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5436 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5437 5438 BB->addSuccessor(loopMBB); 5439 BB->addSuccessor(exitMBB); 5440 5441 // exitMBB: 5442 // ... 5443 BB = exitMBB; 5444 5445 MI->eraseFromParent(); // The instruction is gone now. 5446 5447 return BB; 5448 } 5449 5450 MachineBasicBlock * 5451 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI, 5452 MachineBasicBlock *BB, 5453 unsigned Size, 5454 bool signExtend, 5455 ARMCC::CondCodes Cond) const { 5456 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5457 5458 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 5459 MachineFunction *MF = BB->getParent(); 5460 MachineFunction::iterator It = BB; 5461 ++It; 5462 5463 unsigned dest = MI->getOperand(0).getReg(); 5464 unsigned ptr = MI->getOperand(1).getReg(); 5465 unsigned incr = MI->getOperand(2).getReg(); 5466 unsigned oldval = dest; 5467 DebugLoc dl = MI->getDebugLoc(); 5468 bool isThumb2 = Subtarget->isThumb2(); 5469 5470 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 5471 if (isThumb2) { 5472 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass); 5473 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass); 5474 } 5475 5476 unsigned ldrOpc, strOpc, extendOpc; 5477 switch (Size) { 5478 default: llvm_unreachable("unsupported size for AtomicCmpSwap!"); 5479 case 1: 5480 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB; 5481 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB; 5482 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB; 5483 break; 5484 case 2: 5485 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH; 5486 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH; 5487 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH; 5488 break; 5489 case 4: 5490 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX; 5491 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX; 5492 extendOpc = 0; 5493 break; 5494 } 5495 5496 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5497 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5498 MF->insert(It, loopMBB); 5499 MF->insert(It, exitMBB); 5500 5501 // Transfer the remainder of BB and its successor edges to exitMBB. 5502 exitMBB->splice(exitMBB->begin(), BB, 5503 llvm::next(MachineBasicBlock::iterator(MI)), 5504 BB->end()); 5505 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 5506 5507 const TargetRegisterClass *TRC = 5508 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass; 5509 unsigned scratch = MRI.createVirtualRegister(TRC); 5510 unsigned scratch2 = MRI.createVirtualRegister(TRC); 5511 5512 // thisMBB: 5513 // ... 5514 // fallthrough --> loopMBB 5515 BB->addSuccessor(loopMBB); 5516 5517 // loopMBB: 5518 // ldrex dest, ptr 5519 // (sign extend dest, if required) 5520 // cmp dest, incr 5521 // cmov.cond scratch2, dest, incr 5522 // strex scratch, scratch2, ptr 5523 // cmp scratch, #0 5524 // bne- loopMBB 5525 // fallthrough --> exitMBB 5526 BB = loopMBB; 5527 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr); 5528 if (ldrOpc == ARM::t2LDREX) 5529 MIB.addImm(0); 5530 AddDefaultPred(MIB); 5531 5532 // Sign extend the value, if necessary. 5533 if (signExtend && extendOpc) { 5534 oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass); 5535 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval) 5536 .addReg(dest) 5537 .addImm(0)); 5538 } 5539 5540 // Build compare and cmov instructions. 5541 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) 5542 .addReg(oldval).addReg(incr)); 5543 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2) 5544 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR); 5545 5546 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr); 5547 if (strOpc == ARM::t2STREX) 5548 MIB.addImm(0); 5549 AddDefaultPred(MIB); 5550 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 5551 .addReg(scratch).addImm(0)); 5552 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5553 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5554 5555 BB->addSuccessor(loopMBB); 5556 BB->addSuccessor(exitMBB); 5557 5558 // exitMBB: 5559 // ... 5560 BB = exitMBB; 5561 5562 MI->eraseFromParent(); // The instruction is gone now. 5563 5564 return BB; 5565 } 5566 5567 MachineBasicBlock * 5568 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB, 5569 unsigned Op1, unsigned Op2, 5570 bool NeedsCarry, bool IsCmpxchg) const { 5571 // This also handles ATOMIC_SWAP, indicated by Op1==0. 5572 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5573 5574 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 5575 MachineFunction *MF = BB->getParent(); 5576 MachineFunction::iterator It = BB; 5577 ++It; 5578 5579 unsigned destlo = MI->getOperand(0).getReg(); 5580 unsigned desthi = MI->getOperand(1).getReg(); 5581 unsigned ptr = MI->getOperand(2).getReg(); 5582 unsigned vallo = MI->getOperand(3).getReg(); 5583 unsigned valhi = MI->getOperand(4).getReg(); 5584 DebugLoc dl = MI->getDebugLoc(); 5585 bool isThumb2 = Subtarget->isThumb2(); 5586 5587 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 5588 if (isThumb2) { 5589 MRI.constrainRegClass(destlo, ARM::rGPRRegisterClass); 5590 MRI.constrainRegClass(desthi, ARM::rGPRRegisterClass); 5591 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass); 5592 } 5593 5594 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD; 5595 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD; 5596 5597 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5598 MachineBasicBlock *contBB = 0, *cont2BB = 0; 5599 if (IsCmpxchg) { 5600 contBB = MF->CreateMachineBasicBlock(LLVM_BB); 5601 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB); 5602 } 5603 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5604 MF->insert(It, loopMBB); 5605 if (IsCmpxchg) { 5606 MF->insert(It, contBB); 5607 MF->insert(It, cont2BB); 5608 } 5609 MF->insert(It, exitMBB); 5610 5611 // Transfer the remainder of BB and its successor edges to exitMBB. 5612 exitMBB->splice(exitMBB->begin(), BB, 5613 llvm::next(MachineBasicBlock::iterator(MI)), 5614 BB->end()); 5615 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 5616 5617 const TargetRegisterClass *TRC = 5618 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass; 5619 unsigned storesuccess = MRI.createVirtualRegister(TRC); 5620 5621 // thisMBB: 5622 // ... 5623 // fallthrough --> loopMBB 5624 BB->addSuccessor(loopMBB); 5625 5626 // loopMBB: 5627 // ldrexd r2, r3, ptr 5628 // <binopa> r0, r2, incr 5629 // <binopb> r1, r3, incr 5630 // strexd storesuccess, r0, r1, ptr 5631 // cmp storesuccess, #0 5632 // bne- loopMBB 5633 // fallthrough --> exitMBB 5634 // 5635 // Note that the registers are explicitly specified because there is not any 5636 // way to force the register allocator to allocate a register pair. 5637 // 5638 // FIXME: The hardcoded registers are not necessary for Thumb2, but we 5639 // need to properly enforce the restriction that the two output registers 5640 // for ldrexd must be different. 5641 BB = loopMBB; 5642 // Load 5643 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc)) 5644 .addReg(ARM::R2, RegState::Define) 5645 .addReg(ARM::R3, RegState::Define).addReg(ptr)); 5646 // Copy r2/r3 into dest. (This copy will normally be coalesced.) 5647 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2); 5648 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3); 5649 5650 if (IsCmpxchg) { 5651 // Add early exit 5652 for (unsigned i = 0; i < 2; i++) { 5653 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : 5654 ARM::CMPrr)) 5655 .addReg(i == 0 ? destlo : desthi) 5656 .addReg(i == 0 ? vallo : valhi)); 5657 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5658 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5659 BB->addSuccessor(exitMBB); 5660 BB->addSuccessor(i == 0 ? contBB : cont2BB); 5661 BB = (i == 0 ? contBB : cont2BB); 5662 } 5663 5664 // Copy to physregs for strexd 5665 unsigned setlo = MI->getOperand(5).getReg(); 5666 unsigned sethi = MI->getOperand(6).getReg(); 5667 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo); 5668 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi); 5669 } else if (Op1) { 5670 // Perform binary operation 5671 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0) 5672 .addReg(destlo).addReg(vallo)) 5673 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry)); 5674 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1) 5675 .addReg(desthi).addReg(valhi)).addReg(0); 5676 } else { 5677 // Copy to physregs for strexd 5678 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo); 5679 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi); 5680 } 5681 5682 // Store 5683 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess) 5684 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr)); 5685 // Cmp+jump 5686 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 5687 .addReg(storesuccess).addImm(0)); 5688 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5689 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5690 5691 BB->addSuccessor(loopMBB); 5692 BB->addSuccessor(exitMBB); 5693 5694 // exitMBB: 5695 // ... 5696 BB = exitMBB; 5697 5698 MI->eraseFromParent(); // The instruction is gone now. 5699 5700 return BB; 5701 } 5702 5703 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and 5704 /// registers the function context. 5705 void ARMTargetLowering:: 5706 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB, 5707 MachineBasicBlock *DispatchBB, int FI) const { 5708 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5709 DebugLoc dl = MI->getDebugLoc(); 5710 MachineFunction *MF = MBB->getParent(); 5711 MachineRegisterInfo *MRI = &MF->getRegInfo(); 5712 MachineConstantPool *MCP = MF->getConstantPool(); 5713 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>(); 5714 const Function *F = MF->getFunction(); 5715 5716 bool isThumb = Subtarget->isThumb(); 5717 bool isThumb2 = Subtarget->isThumb2(); 5718 5719 unsigned PCLabelId = AFI->createPICLabelUId(); 5720 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8; 5721 ARMConstantPoolValue *CPV = 5722 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj); 5723 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4); 5724 5725 const TargetRegisterClass *TRC = 5726 isThumb ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass; 5727 5728 // Grab constant pool and fixed stack memory operands. 5729 MachineMemOperand *CPMMO = 5730 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(), 5731 MachineMemOperand::MOLoad, 4, 4); 5732 5733 MachineMemOperand *FIMMOSt = 5734 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI), 5735 MachineMemOperand::MOStore, 4, 4); 5736 5737 // Load the address of the dispatch MBB into the jump buffer. 5738 if (isThumb2) { 5739 // Incoming value: jbuf 5740 // ldr.n r5, LCPI1_1 5741 // orr r5, r5, #1 5742 // add r5, pc 5743 // str r5, [$jbuf, #+4] ; &jbuf[1] 5744 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 5745 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1) 5746 .addConstantPoolIndex(CPI) 5747 .addMemOperand(CPMMO)); 5748 // Set the low bit because of thumb mode. 5749 unsigned NewVReg2 = MRI->createVirtualRegister(TRC); 5750 AddDefaultCC( 5751 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2) 5752 .addReg(NewVReg1, RegState::Kill) 5753 .addImm(0x01))); 5754 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 5755 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3) 5756 .addReg(NewVReg2, RegState::Kill) 5757 .addImm(PCLabelId); 5758 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12)) 5759 .addReg(NewVReg3, RegState::Kill) 5760 .addFrameIndex(FI) 5761 .addImm(36) // &jbuf[1] :: pc 5762 .addMemOperand(FIMMOSt)); 5763 } else if (isThumb) { 5764 // Incoming value: jbuf 5765 // ldr.n r1, LCPI1_4 5766 // add r1, pc 5767 // mov r2, #1 5768 // orrs r1, r2 5769 // add r2, $jbuf, #+4 ; &jbuf[1] 5770 // str r1, [r2] 5771 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 5772 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1) 5773 .addConstantPoolIndex(CPI) 5774 .addMemOperand(CPMMO)); 5775 unsigned NewVReg2 = MRI->createVirtualRegister(TRC); 5776 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2) 5777 .addReg(NewVReg1, RegState::Kill) 5778 .addImm(PCLabelId); 5779 // Set the low bit because of thumb mode. 5780 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 5781 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3) 5782 .addReg(ARM::CPSR, RegState::Define) 5783 .addImm(1)); 5784 unsigned NewVReg4 = MRI->createVirtualRegister(TRC); 5785 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4) 5786 .addReg(ARM::CPSR, RegState::Define) 5787 .addReg(NewVReg2, RegState::Kill) 5788 .addReg(NewVReg3, RegState::Kill)); 5789 unsigned NewVReg5 = MRI->createVirtualRegister(TRC); 5790 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5) 5791 .addFrameIndex(FI) 5792 .addImm(36)); // &jbuf[1] :: pc 5793 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi)) 5794 .addReg(NewVReg4, RegState::Kill) 5795 .addReg(NewVReg5, RegState::Kill) 5796 .addImm(0) 5797 .addMemOperand(FIMMOSt)); 5798 } else { 5799 // Incoming value: jbuf 5800 // ldr r1, LCPI1_1 5801 // add r1, pc, r1 5802 // str r1, [$jbuf, #+4] ; &jbuf[1] 5803 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 5804 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1) 5805 .addConstantPoolIndex(CPI) 5806 .addImm(0) 5807 .addMemOperand(CPMMO)); 5808 unsigned NewVReg2 = MRI->createVirtualRegister(TRC); 5809 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2) 5810 .addReg(NewVReg1, RegState::Kill) 5811 .addImm(PCLabelId)); 5812 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12)) 5813 .addReg(NewVReg2, RegState::Kill) 5814 .addFrameIndex(FI) 5815 .addImm(36) // &jbuf[1] :: pc 5816 .addMemOperand(FIMMOSt)); 5817 } 5818 } 5819 5820 MachineBasicBlock *ARMTargetLowering:: 5821 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const { 5822 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5823 DebugLoc dl = MI->getDebugLoc(); 5824 MachineFunction *MF = MBB->getParent(); 5825 MachineRegisterInfo *MRI = &MF->getRegInfo(); 5826 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>(); 5827 MachineFrameInfo *MFI = MF->getFrameInfo(); 5828 int FI = MFI->getFunctionContextIndex(); 5829 5830 const TargetRegisterClass *TRC = 5831 Subtarget->isThumb() ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass; 5832 5833 // Get a mapping of the call site numbers to all of the landing pads they're 5834 // associated with. 5835 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad; 5836 unsigned MaxCSNum = 0; 5837 MachineModuleInfo &MMI = MF->getMMI(); 5838 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; 5839 ++BB) { 5840 if (!BB->isLandingPad()) continue; 5841 5842 // FIXME: We should assert that the EH_LABEL is the first MI in the landing 5843 // pad. 5844 for (MachineBasicBlock::iterator 5845 II = BB->begin(), IE = BB->end(); II != IE; ++II) { 5846 if (!II->isEHLabel()) continue; 5847 5848 MCSymbol *Sym = II->getOperand(0).getMCSymbol(); 5849 if (!MMI.hasCallSiteLandingPad(Sym)) continue; 5850 5851 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym); 5852 for (SmallVectorImpl<unsigned>::iterator 5853 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end(); 5854 CSI != CSE; ++CSI) { 5855 CallSiteNumToLPad[*CSI].push_back(BB); 5856 MaxCSNum = std::max(MaxCSNum, *CSI); 5857 } 5858 break; 5859 } 5860 } 5861 5862 // Get an ordered list of the machine basic blocks for the jump table. 5863 std::vector<MachineBasicBlock*> LPadList; 5864 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs; 5865 LPadList.reserve(CallSiteNumToLPad.size()); 5866 for (unsigned I = 1; I <= MaxCSNum; ++I) { 5867 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I]; 5868 for (SmallVectorImpl<MachineBasicBlock*>::iterator 5869 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) { 5870 LPadList.push_back(*II); 5871 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end()); 5872 } 5873 } 5874 5875 assert(!LPadList.empty() && 5876 "No landing pad destinations for the dispatch jump table!"); 5877 5878 // Create the jump table and associated information. 5879 MachineJumpTableInfo *JTI = 5880 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline); 5881 unsigned MJTI = JTI->createJumpTableIndex(LPadList); 5882 unsigned UId = AFI->createJumpTableUId(); 5883 5884 // Create the MBBs for the dispatch code. 5885 5886 // Shove the dispatch's address into the return slot in the function context. 5887 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock(); 5888 DispatchBB->setIsLandingPad(); 5889 5890 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock(); 5891 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP)); 5892 DispatchBB->addSuccessor(TrapBB); 5893 5894 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock(); 5895 DispatchBB->addSuccessor(DispContBB); 5896 5897 // Insert and MBBs. 5898 MF->insert(MF->end(), DispatchBB); 5899 MF->insert(MF->end(), DispContBB); 5900 MF->insert(MF->end(), TrapBB); 5901 5902 // Insert code into the entry block that creates and registers the function 5903 // context. 5904 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI); 5905 5906 MachineMemOperand *FIMMOLd = 5907 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI), 5908 MachineMemOperand::MOLoad | 5909 MachineMemOperand::MOVolatile, 4, 4); 5910 5911 if (AFI->isThumb1OnlyFunction()) 5912 BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup)); 5913 else if (!Subtarget->hasVFP2()) 5914 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp)); 5915 else 5916 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup)); 5917 5918 unsigned NumLPads = LPadList.size(); 5919 if (Subtarget->isThumb2()) { 5920 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 5921 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1) 5922 .addFrameIndex(FI) 5923 .addImm(4) 5924 .addMemOperand(FIMMOLd)); 5925 5926 if (NumLPads < 256) { 5927 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri)) 5928 .addReg(NewVReg1) 5929 .addImm(LPadList.size())); 5930 } else { 5931 unsigned VReg1 = MRI->createVirtualRegister(TRC); 5932 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1) 5933 .addImm(NumLPads & 0xFFFF)); 5934 5935 unsigned VReg2 = VReg1; 5936 if ((NumLPads & 0xFFFF0000) != 0) { 5937 VReg2 = MRI->createVirtualRegister(TRC); 5938 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2) 5939 .addReg(VReg1) 5940 .addImm(NumLPads >> 16)); 5941 } 5942 5943 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr)) 5944 .addReg(NewVReg1) 5945 .addReg(VReg2)); 5946 } 5947 5948 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc)) 5949 .addMBB(TrapBB) 5950 .addImm(ARMCC::HI) 5951 .addReg(ARM::CPSR); 5952 5953 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 5954 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3) 5955 .addJumpTableIndex(MJTI) 5956 .addImm(UId)); 5957 5958 unsigned NewVReg4 = MRI->createVirtualRegister(TRC); 5959 AddDefaultCC( 5960 AddDefaultPred( 5961 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4) 5962 .addReg(NewVReg3, RegState::Kill) 5963 .addReg(NewVReg1) 5964 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2)))); 5965 5966 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT)) 5967 .addReg(NewVReg4, RegState::Kill) 5968 .addReg(NewVReg1) 5969 .addJumpTableIndex(MJTI) 5970 .addImm(UId); 5971 } else if (Subtarget->isThumb()) { 5972 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 5973 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1) 5974 .addFrameIndex(FI) 5975 .addImm(1) 5976 .addMemOperand(FIMMOLd)); 5977 5978 if (NumLPads < 256) { 5979 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8)) 5980 .addReg(NewVReg1) 5981 .addImm(NumLPads)); 5982 } else { 5983 MachineConstantPool *ConstantPool = MF->getConstantPool(); 5984 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext()); 5985 const Constant *C = ConstantInt::get(Int32Ty, NumLPads); 5986 5987 // MachineConstantPool wants an explicit alignment. 5988 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty); 5989 if (Align == 0) 5990 Align = getTargetData()->getTypeAllocSize(C->getType()); 5991 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align); 5992 5993 unsigned VReg1 = MRI->createVirtualRegister(TRC); 5994 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci)) 5995 .addReg(VReg1, RegState::Define) 5996 .addConstantPoolIndex(Idx)); 5997 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr)) 5998 .addReg(NewVReg1) 5999 .addReg(VReg1)); 6000 } 6001 6002 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc)) 6003 .addMBB(TrapBB) 6004 .addImm(ARMCC::HI) 6005 .addReg(ARM::CPSR); 6006 6007 unsigned NewVReg2 = MRI->createVirtualRegister(TRC); 6008 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2) 6009 .addReg(ARM::CPSR, RegState::Define) 6010 .addReg(NewVReg1) 6011 .addImm(2)); 6012 6013 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 6014 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3) 6015 .addJumpTableIndex(MJTI) 6016 .addImm(UId)); 6017 6018 unsigned NewVReg4 = MRI->createVirtualRegister(TRC); 6019 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4) 6020 .addReg(ARM::CPSR, RegState::Define) 6021 .addReg(NewVReg2, RegState::Kill) 6022 .addReg(NewVReg3)); 6023 6024 MachineMemOperand *JTMMOLd = 6025 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(), 6026 MachineMemOperand::MOLoad, 4, 4); 6027 6028 unsigned NewVReg5 = MRI->createVirtualRegister(TRC); 6029 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5) 6030 .addReg(NewVReg4, RegState::Kill) 6031 .addImm(0) 6032 .addMemOperand(JTMMOLd)); 6033 6034 unsigned NewVReg6 = MRI->createVirtualRegister(TRC); 6035 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6) 6036 .addReg(ARM::CPSR, RegState::Define) 6037 .addReg(NewVReg5, RegState::Kill) 6038 .addReg(NewVReg3)); 6039 6040 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr)) 6041 .addReg(NewVReg6, RegState::Kill) 6042 .addJumpTableIndex(MJTI) 6043 .addImm(UId); 6044 } else { 6045 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 6046 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1) 6047 .addFrameIndex(FI) 6048 .addImm(4) 6049 .addMemOperand(FIMMOLd)); 6050 6051 if (NumLPads < 256) { 6052 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri)) 6053 .addReg(NewVReg1) 6054 .addImm(NumLPads)); 6055 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) { 6056 unsigned VReg1 = MRI->createVirtualRegister(TRC); 6057 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1) 6058 .addImm(NumLPads & 0xFFFF)); 6059 6060 unsigned VReg2 = VReg1; 6061 if ((NumLPads & 0xFFFF0000) != 0) { 6062 VReg2 = MRI->createVirtualRegister(TRC); 6063 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2) 6064 .addReg(VReg1) 6065 .addImm(NumLPads >> 16)); 6066 } 6067 6068 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr)) 6069 .addReg(NewVReg1) 6070 .addReg(VReg2)); 6071 } else { 6072 MachineConstantPool *ConstantPool = MF->getConstantPool(); 6073 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext()); 6074 const Constant *C = ConstantInt::get(Int32Ty, NumLPads); 6075 6076 // MachineConstantPool wants an explicit alignment. 6077 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty); 6078 if (Align == 0) 6079 Align = getTargetData()->getTypeAllocSize(C->getType()); 6080 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align); 6081 6082 unsigned VReg1 = MRI->createVirtualRegister(TRC); 6083 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp)) 6084 .addReg(VReg1, RegState::Define) 6085 .addConstantPoolIndex(Idx) 6086 .addImm(0)); 6087 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr)) 6088 .addReg(NewVReg1) 6089 .addReg(VReg1, RegState::Kill)); 6090 } 6091 6092 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc)) 6093 .addMBB(TrapBB) 6094 .addImm(ARMCC::HI) 6095 .addReg(ARM::CPSR); 6096 6097 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 6098 AddDefaultCC( 6099 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3) 6100 .addReg(NewVReg1) 6101 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2)))); 6102 unsigned NewVReg4 = MRI->createVirtualRegister(TRC); 6103 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4) 6104 .addJumpTableIndex(MJTI) 6105 .addImm(UId)); 6106 6107 MachineMemOperand *JTMMOLd = 6108 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(), 6109 MachineMemOperand::MOLoad, 4, 4); 6110 unsigned NewVReg5 = MRI->createVirtualRegister(TRC); 6111 AddDefaultPred( 6112 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5) 6113 .addReg(NewVReg3, RegState::Kill) 6114 .addReg(NewVReg4) 6115 .addImm(0) 6116 .addMemOperand(JTMMOLd)); 6117 6118 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd)) 6119 .addReg(NewVReg5, RegState::Kill) 6120 .addReg(NewVReg4) 6121 .addJumpTableIndex(MJTI) 6122 .addImm(UId); 6123 } 6124 6125 // Add the jump table entries as successors to the MBB. 6126 MachineBasicBlock *PrevMBB = 0; 6127 for (std::vector<MachineBasicBlock*>::iterator 6128 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) { 6129 MachineBasicBlock *CurMBB = *I; 6130 if (PrevMBB != CurMBB) 6131 DispContBB->addSuccessor(CurMBB); 6132 PrevMBB = CurMBB; 6133 } 6134 6135 // N.B. the order the invoke BBs are processed in doesn't matter here. 6136 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII); 6137 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo(); 6138 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF); 6139 SmallVector<MachineBasicBlock*, 64> MBBLPads; 6140 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator 6141 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) { 6142 MachineBasicBlock *BB = *I; 6143 6144 // Remove the landing pad successor from the invoke block and replace it 6145 // with the new dispatch block. 6146 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(), 6147 BB->succ_end()); 6148 while (!Successors.empty()) { 6149 MachineBasicBlock *SMBB = Successors.pop_back_val(); 6150 if (SMBB->isLandingPad()) { 6151 BB->removeSuccessor(SMBB); 6152 MBBLPads.push_back(SMBB); 6153 } 6154 } 6155 6156 BB->addSuccessor(DispatchBB); 6157 6158 // Find the invoke call and mark all of the callee-saved registers as 6159 // 'implicit defined' so that they're spilled. This prevents code from 6160 // moving instructions to before the EH block, where they will never be 6161 // executed. 6162 for (MachineBasicBlock::reverse_iterator 6163 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) { 6164 if (!II->isCall()) continue; 6165 6166 DenseMap<unsigned, bool> DefRegs; 6167 for (MachineInstr::mop_iterator 6168 OI = II->operands_begin(), OE = II->operands_end(); 6169 OI != OE; ++OI) { 6170 if (!OI->isReg()) continue; 6171 DefRegs[OI->getReg()] = true; 6172 } 6173 6174 MachineInstrBuilder MIB(&*II); 6175 6176 for (unsigned i = 0; SavedRegs[i] != 0; ++i) { 6177 unsigned Reg = SavedRegs[i]; 6178 if (Subtarget->isThumb2() && 6179 !ARM::tGPRRegisterClass->contains(Reg) && 6180 !ARM::hGPRRegisterClass->contains(Reg)) 6181 continue; 6182 else if (Subtarget->isThumb1Only() && 6183 !ARM::tGPRRegisterClass->contains(Reg)) 6184 continue; 6185 else if (!Subtarget->isThumb() && 6186 !ARM::GPRRegisterClass->contains(Reg)) 6187 continue; 6188 if (!DefRegs[Reg]) 6189 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead); 6190 } 6191 6192 break; 6193 } 6194 } 6195 6196 // Mark all former landing pads as non-landing pads. The dispatch is the only 6197 // landing pad now. 6198 for (SmallVectorImpl<MachineBasicBlock*>::iterator 6199 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I) 6200 (*I)->setIsLandingPad(false); 6201 6202 // The instruction is gone now. 6203 MI->eraseFromParent(); 6204 6205 return MBB; 6206 } 6207 6208 static 6209 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) { 6210 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(), 6211 E = MBB->succ_end(); I != E; ++I) 6212 if (*I != Succ) 6213 return *I; 6214 llvm_unreachable("Expecting a BB with two successors!"); 6215 } 6216 6217 MachineBasicBlock * 6218 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 6219 MachineBasicBlock *BB) const { 6220 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 6221 DebugLoc dl = MI->getDebugLoc(); 6222 bool isThumb2 = Subtarget->isThumb2(); 6223 switch (MI->getOpcode()) { 6224 default: { 6225 MI->dump(); 6226 llvm_unreachable("Unexpected instr type to insert"); 6227 } 6228 // The Thumb2 pre-indexed stores have the same MI operands, they just 6229 // define them differently in the .td files from the isel patterns, so 6230 // they need pseudos. 6231 case ARM::t2STR_preidx: 6232 MI->setDesc(TII->get(ARM::t2STR_PRE)); 6233 return BB; 6234 case ARM::t2STRB_preidx: 6235 MI->setDesc(TII->get(ARM::t2STRB_PRE)); 6236 return BB; 6237 case ARM::t2STRH_preidx: 6238 MI->setDesc(TII->get(ARM::t2STRH_PRE)); 6239 return BB; 6240 6241 case ARM::STRi_preidx: 6242 case ARM::STRBi_preidx: { 6243 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ? 6244 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM; 6245 // Decode the offset. 6246 unsigned Offset = MI->getOperand(4).getImm(); 6247 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub; 6248 Offset = ARM_AM::getAM2Offset(Offset); 6249 if (isSub) 6250 Offset = -Offset; 6251 6252 MachineMemOperand *MMO = *MI->memoperands_begin(); 6253 BuildMI(*BB, MI, dl, TII->get(NewOpc)) 6254 .addOperand(MI->getOperand(0)) // Rn_wb 6255 .addOperand(MI->getOperand(1)) // Rt 6256 .addOperand(MI->getOperand(2)) // Rn 6257 .addImm(Offset) // offset (skip GPR==zero_reg) 6258 .addOperand(MI->getOperand(5)) // pred 6259 .addOperand(MI->getOperand(6)) 6260 .addMemOperand(MMO); 6261 MI->eraseFromParent(); 6262 return BB; 6263 } 6264 case ARM::STRr_preidx: 6265 case ARM::STRBr_preidx: 6266 case ARM::STRH_preidx: { 6267 unsigned NewOpc; 6268 switch (MI->getOpcode()) { 6269 default: llvm_unreachable("unexpected opcode!"); 6270 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break; 6271 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break; 6272 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break; 6273 } 6274 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc)); 6275 for (unsigned i = 0; i < MI->getNumOperands(); ++i) 6276 MIB.addOperand(MI->getOperand(i)); 6277 MI->eraseFromParent(); 6278 return BB; 6279 } 6280 case ARM::ATOMIC_LOAD_ADD_I8: 6281 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr); 6282 case ARM::ATOMIC_LOAD_ADD_I16: 6283 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr); 6284 case ARM::ATOMIC_LOAD_ADD_I32: 6285 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr); 6286 6287 case ARM::ATOMIC_LOAD_AND_I8: 6288 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr); 6289 case ARM::ATOMIC_LOAD_AND_I16: 6290 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr); 6291 case ARM::ATOMIC_LOAD_AND_I32: 6292 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr); 6293 6294 case ARM::ATOMIC_LOAD_OR_I8: 6295 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr); 6296 case ARM::ATOMIC_LOAD_OR_I16: 6297 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr); 6298 case ARM::ATOMIC_LOAD_OR_I32: 6299 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr); 6300 6301 case ARM::ATOMIC_LOAD_XOR_I8: 6302 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr); 6303 case ARM::ATOMIC_LOAD_XOR_I16: 6304 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr); 6305 case ARM::ATOMIC_LOAD_XOR_I32: 6306 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr); 6307 6308 case ARM::ATOMIC_LOAD_NAND_I8: 6309 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr); 6310 case ARM::ATOMIC_LOAD_NAND_I16: 6311 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr); 6312 case ARM::ATOMIC_LOAD_NAND_I32: 6313 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr); 6314 6315 case ARM::ATOMIC_LOAD_SUB_I8: 6316 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr); 6317 case ARM::ATOMIC_LOAD_SUB_I16: 6318 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr); 6319 case ARM::ATOMIC_LOAD_SUB_I32: 6320 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr); 6321 6322 case ARM::ATOMIC_LOAD_MIN_I8: 6323 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT); 6324 case ARM::ATOMIC_LOAD_MIN_I16: 6325 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT); 6326 case ARM::ATOMIC_LOAD_MIN_I32: 6327 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT); 6328 6329 case ARM::ATOMIC_LOAD_MAX_I8: 6330 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT); 6331 case ARM::ATOMIC_LOAD_MAX_I16: 6332 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT); 6333 case ARM::ATOMIC_LOAD_MAX_I32: 6334 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT); 6335 6336 case ARM::ATOMIC_LOAD_UMIN_I8: 6337 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO); 6338 case ARM::ATOMIC_LOAD_UMIN_I16: 6339 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO); 6340 case ARM::ATOMIC_LOAD_UMIN_I32: 6341 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO); 6342 6343 case ARM::ATOMIC_LOAD_UMAX_I8: 6344 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI); 6345 case ARM::ATOMIC_LOAD_UMAX_I16: 6346 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI); 6347 case ARM::ATOMIC_LOAD_UMAX_I32: 6348 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI); 6349 6350 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0); 6351 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0); 6352 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0); 6353 6354 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1); 6355 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2); 6356 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4); 6357 6358 6359 case ARM::ATOMADD6432: 6360 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr, 6361 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr, 6362 /*NeedsCarry*/ true); 6363 case ARM::ATOMSUB6432: 6364 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr, 6365 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr, 6366 /*NeedsCarry*/ true); 6367 case ARM::ATOMOR6432: 6368 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr, 6369 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr); 6370 case ARM::ATOMXOR6432: 6371 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr, 6372 isThumb2 ? ARM::t2EORrr : ARM::EORrr); 6373 case ARM::ATOMAND6432: 6374 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr, 6375 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr); 6376 case ARM::ATOMSWAP6432: 6377 return EmitAtomicBinary64(MI, BB, 0, 0, false); 6378 case ARM::ATOMCMPXCHG6432: 6379 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr, 6380 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr, 6381 /*NeedsCarry*/ false, /*IsCmpxchg*/true); 6382 6383 case ARM::tMOVCCr_pseudo: { 6384 // To "insert" a SELECT_CC instruction, we actually have to insert the 6385 // diamond control-flow pattern. The incoming instruction knows the 6386 // destination vreg to set, the condition code register to branch on, the 6387 // true/false values to select between, and a branch opcode to use. 6388 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 6389 MachineFunction::iterator It = BB; 6390 ++It; 6391 6392 // thisMBB: 6393 // ... 6394 // TrueVal = ... 6395 // cmpTY ccX, r1, r2 6396 // bCC copy1MBB 6397 // fallthrough --> copy0MBB 6398 MachineBasicBlock *thisMBB = BB; 6399 MachineFunction *F = BB->getParent(); 6400 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 6401 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); 6402 F->insert(It, copy0MBB); 6403 F->insert(It, sinkMBB); 6404 6405 // Transfer the remainder of BB and its successor edges to sinkMBB. 6406 sinkMBB->splice(sinkMBB->begin(), BB, 6407 llvm::next(MachineBasicBlock::iterator(MI)), 6408 BB->end()); 6409 sinkMBB->transferSuccessorsAndUpdatePHIs(BB); 6410 6411 BB->addSuccessor(copy0MBB); 6412 BB->addSuccessor(sinkMBB); 6413 6414 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB) 6415 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg()); 6416 6417 // copy0MBB: 6418 // %FalseValue = ... 6419 // # fallthrough to sinkMBB 6420 BB = copy0MBB; 6421 6422 // Update machine-CFG edges 6423 BB->addSuccessor(sinkMBB); 6424 6425 // sinkMBB: 6426 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] 6427 // ... 6428 BB = sinkMBB; 6429 BuildMI(*BB, BB->begin(), dl, 6430 TII->get(ARM::PHI), MI->getOperand(0).getReg()) 6431 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB) 6432 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB); 6433 6434 MI->eraseFromParent(); // The pseudo instruction is gone now. 6435 return BB; 6436 } 6437 6438 case ARM::BCCi64: 6439 case ARM::BCCZi64: { 6440 // If there is an unconditional branch to the other successor, remove it. 6441 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end()); 6442 6443 // Compare both parts that make up the double comparison separately for 6444 // equality. 6445 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64; 6446 6447 unsigned LHS1 = MI->getOperand(1).getReg(); 6448 unsigned LHS2 = MI->getOperand(2).getReg(); 6449 if (RHSisZero) { 6450 AddDefaultPred(BuildMI(BB, dl, 6451 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 6452 .addReg(LHS1).addImm(0)); 6453 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 6454 .addReg(LHS2).addImm(0) 6455 .addImm(ARMCC::EQ).addReg(ARM::CPSR); 6456 } else { 6457 unsigned RHS1 = MI->getOperand(3).getReg(); 6458 unsigned RHS2 = MI->getOperand(4).getReg(); 6459 AddDefaultPred(BuildMI(BB, dl, 6460 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) 6461 .addReg(LHS1).addReg(RHS1)); 6462 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) 6463 .addReg(LHS2).addReg(RHS2) 6464 .addImm(ARMCC::EQ).addReg(ARM::CPSR); 6465 } 6466 6467 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB(); 6468 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB); 6469 if (MI->getOperand(0).getImm() == ARMCC::NE) 6470 std::swap(destMBB, exitMBB); 6471 6472 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 6473 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR); 6474 if (isThumb2) 6475 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB)); 6476 else 6477 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB); 6478 6479 MI->eraseFromParent(); // The pseudo instruction is gone now. 6480 return BB; 6481 } 6482 6483 case ARM::Int_eh_sjlj_setjmp: 6484 case ARM::Int_eh_sjlj_setjmp_nofp: 6485 case ARM::tInt_eh_sjlj_setjmp: 6486 case ARM::t2Int_eh_sjlj_setjmp: 6487 case ARM::t2Int_eh_sjlj_setjmp_nofp: 6488 EmitSjLjDispatchBlock(MI, BB); 6489 return BB; 6490 6491 case ARM::ABS: 6492 case ARM::t2ABS: { 6493 // To insert an ABS instruction, we have to insert the 6494 // diamond control-flow pattern. The incoming instruction knows the 6495 // source vreg to test against 0, the destination vreg to set, 6496 // the condition code register to branch on, the 6497 // true/false values to select between, and a branch opcode to use. 6498 // It transforms 6499 // V1 = ABS V0 6500 // into 6501 // V2 = MOVS V0 6502 // BCC (branch to SinkBB if V0 >= 0) 6503 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0) 6504 // SinkBB: V1 = PHI(V2, V3) 6505 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 6506 MachineFunction::iterator BBI = BB; 6507 ++BBI; 6508 MachineFunction *Fn = BB->getParent(); 6509 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB); 6510 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB); 6511 Fn->insert(BBI, RSBBB); 6512 Fn->insert(BBI, SinkBB); 6513 6514 unsigned int ABSSrcReg = MI->getOperand(1).getReg(); 6515 unsigned int ABSDstReg = MI->getOperand(0).getReg(); 6516 bool isThumb2 = Subtarget->isThumb2(); 6517 MachineRegisterInfo &MRI = Fn->getRegInfo(); 6518 // In Thumb mode S must not be specified if source register is the SP or 6519 // PC and if destination register is the SP, so restrict register class 6520 unsigned NewMovDstReg = MRI.createVirtualRegister( 6521 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass); 6522 unsigned NewRsbDstReg = MRI.createVirtualRegister( 6523 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass); 6524 6525 // Transfer the remainder of BB and its successor edges to sinkMBB. 6526 SinkBB->splice(SinkBB->begin(), BB, 6527 llvm::next(MachineBasicBlock::iterator(MI)), 6528 BB->end()); 6529 SinkBB->transferSuccessorsAndUpdatePHIs(BB); 6530 6531 BB->addSuccessor(RSBBB); 6532 BB->addSuccessor(SinkBB); 6533 6534 // fall through to SinkMBB 6535 RSBBB->addSuccessor(SinkBB); 6536 6537 // insert a movs at the end of BB 6538 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVr : ARM::MOVr), 6539 NewMovDstReg) 6540 .addReg(ABSSrcReg, RegState::Kill) 6541 .addImm((unsigned)ARMCC::AL).addReg(0) 6542 .addReg(ARM::CPSR, RegState::Define); 6543 6544 // insert a bcc with opposite CC to ARMCC::MI at the end of BB 6545 BuildMI(BB, dl, 6546 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB) 6547 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR); 6548 6549 // insert rsbri in RSBBB 6550 // Note: BCC and rsbri will be converted into predicated rsbmi 6551 // by if-conversion pass 6552 BuildMI(*RSBBB, RSBBB->begin(), dl, 6553 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg) 6554 .addReg(NewMovDstReg, RegState::Kill) 6555 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0); 6556 6557 // insert PHI in SinkBB, 6558 // reuse ABSDstReg to not change uses of ABS instruction 6559 BuildMI(*SinkBB, SinkBB->begin(), dl, 6560 TII->get(ARM::PHI), ABSDstReg) 6561 .addReg(NewRsbDstReg).addMBB(RSBBB) 6562 .addReg(NewMovDstReg).addMBB(BB); 6563 6564 // remove ABS instruction 6565 MI->eraseFromParent(); 6566 6567 // return last added BB 6568 return SinkBB; 6569 } 6570 } 6571 } 6572 6573 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI, 6574 SDNode *Node) const { 6575 if (!MI->hasPostISelHook()) { 6576 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) && 6577 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'"); 6578 return; 6579 } 6580 6581 const MCInstrDesc *MCID = &MI->getDesc(); 6582 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB, 6583 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional 6584 // operand is still set to noreg. If needed, set the optional operand's 6585 // register to CPSR, and remove the redundant implicit def. 6586 // 6587 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>). 6588 6589 // Rename pseudo opcodes. 6590 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode()); 6591 if (NewOpc) { 6592 const ARMBaseInstrInfo *TII = 6593 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo()); 6594 MCID = &TII->get(NewOpc); 6595 6596 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 && 6597 "converted opcode should be the same except for cc_out"); 6598 6599 MI->setDesc(*MCID); 6600 6601 // Add the optional cc_out operand 6602 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true)); 6603 } 6604 unsigned ccOutIdx = MCID->getNumOperands() - 1; 6605 6606 // Any ARM instruction that sets the 's' bit should specify an optional 6607 // "cc_out" operand in the last operand position. 6608 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) { 6609 assert(!NewOpc && "Optional cc_out operand required"); 6610 return; 6611 } 6612 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it 6613 // since we already have an optional CPSR def. 6614 bool definesCPSR = false; 6615 bool deadCPSR = false; 6616 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands(); 6617 i != e; ++i) { 6618 const MachineOperand &MO = MI->getOperand(i); 6619 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) { 6620 definesCPSR = true; 6621 if (MO.isDead()) 6622 deadCPSR = true; 6623 MI->RemoveOperand(i); 6624 break; 6625 } 6626 } 6627 if (!definesCPSR) { 6628 assert(!NewOpc && "Optional cc_out operand required"); 6629 return; 6630 } 6631 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag"); 6632 if (deadCPSR) { 6633 assert(!MI->getOperand(ccOutIdx).getReg() && 6634 "expect uninitialized optional cc_out operand"); 6635 return; 6636 } 6637 6638 // If this instruction was defined with an optional CPSR def and its dag node 6639 // had a live implicit CPSR def, then activate the optional CPSR def. 6640 MachineOperand &MO = MI->getOperand(ccOutIdx); 6641 MO.setReg(ARM::CPSR); 6642 MO.setIsDef(true); 6643 } 6644 6645 //===----------------------------------------------------------------------===// 6646 // ARM Optimization Hooks 6647 //===----------------------------------------------------------------------===// 6648 6649 static 6650 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp, 6651 TargetLowering::DAGCombinerInfo &DCI) { 6652 SelectionDAG &DAG = DCI.DAG; 6653 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 6654 EVT VT = N->getValueType(0); 6655 unsigned Opc = N->getOpcode(); 6656 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC; 6657 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1); 6658 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2); 6659 ISD::CondCode CC = ISD::SETCC_INVALID; 6660 6661 if (isSlctCC) { 6662 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get(); 6663 } else { 6664 SDValue CCOp = Slct.getOperand(0); 6665 if (CCOp.getOpcode() == ISD::SETCC) 6666 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get(); 6667 } 6668 6669 bool DoXform = false; 6670 bool InvCC = false; 6671 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) && 6672 "Bad input!"); 6673 6674 if (LHS.getOpcode() == ISD::Constant && 6675 cast<ConstantSDNode>(LHS)->isNullValue()) { 6676 DoXform = true; 6677 } else if (CC != ISD::SETCC_INVALID && 6678 RHS.getOpcode() == ISD::Constant && 6679 cast<ConstantSDNode>(RHS)->isNullValue()) { 6680 std::swap(LHS, RHS); 6681 SDValue Op0 = Slct.getOperand(0); 6682 EVT OpVT = isSlctCC ? Op0.getValueType() : 6683 Op0.getOperand(0).getValueType(); 6684 bool isInt = OpVT.isInteger(); 6685 CC = ISD::getSetCCInverse(CC, isInt); 6686 6687 if (!TLI.isCondCodeLegal(CC, OpVT)) 6688 return SDValue(); // Inverse operator isn't legal. 6689 6690 DoXform = true; 6691 InvCC = true; 6692 } 6693 6694 if (DoXform) { 6695 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS); 6696 if (isSlctCC) 6697 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result, 6698 Slct.getOperand(0), Slct.getOperand(1), CC); 6699 SDValue CCOp = Slct.getOperand(0); 6700 if (InvCC) 6701 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(), 6702 CCOp.getOperand(0), CCOp.getOperand(1), CC); 6703 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT, 6704 CCOp, OtherOp, Result); 6705 } 6706 return SDValue(); 6707 } 6708 6709 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction 6710 // (only after legalization). 6711 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1, 6712 TargetLowering::DAGCombinerInfo &DCI, 6713 const ARMSubtarget *Subtarget) { 6714 6715 // Only perform optimization if after legalize, and if NEON is available. We 6716 // also expected both operands to be BUILD_VECTORs. 6717 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON() 6718 || N0.getOpcode() != ISD::BUILD_VECTOR 6719 || N1.getOpcode() != ISD::BUILD_VECTOR) 6720 return SDValue(); 6721 6722 // Check output type since VPADDL operand elements can only be 8, 16, or 32. 6723 EVT VT = N->getValueType(0); 6724 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64) 6725 return SDValue(); 6726 6727 // Check that the vector operands are of the right form. 6728 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR 6729 // operands, where N is the size of the formed vector. 6730 // Each EXTRACT_VECTOR should have the same input vector and odd or even 6731 // index such that we have a pair wise add pattern. 6732 6733 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing. 6734 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT) 6735 return SDValue(); 6736 SDValue Vec = N0->getOperand(0)->getOperand(0); 6737 SDNode *V = Vec.getNode(); 6738 unsigned nextIndex = 0; 6739 6740 // For each operands to the ADD which are BUILD_VECTORs, 6741 // check to see if each of their operands are an EXTRACT_VECTOR with 6742 // the same vector and appropriate index. 6743 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) { 6744 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT 6745 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) { 6746 6747 SDValue ExtVec0 = N0->getOperand(i); 6748 SDValue ExtVec1 = N1->getOperand(i); 6749 6750 // First operand is the vector, verify its the same. 6751 if (V != ExtVec0->getOperand(0).getNode() || 6752 V != ExtVec1->getOperand(0).getNode()) 6753 return SDValue(); 6754 6755 // Second is the constant, verify its correct. 6756 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1)); 6757 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1)); 6758 6759 // For the constant, we want to see all the even or all the odd. 6760 if (!C0 || !C1 || C0->getZExtValue() != nextIndex 6761 || C1->getZExtValue() != nextIndex+1) 6762 return SDValue(); 6763 6764 // Increment index. 6765 nextIndex+=2; 6766 } else 6767 return SDValue(); 6768 } 6769 6770 // Create VPADDL node. 6771 SelectionDAG &DAG = DCI.DAG; 6772 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 6773 6774 // Build operand list. 6775 SmallVector<SDValue, 8> Ops; 6776 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, 6777 TLI.getPointerTy())); 6778 6779 // Input is the vector. 6780 Ops.push_back(Vec); 6781 6782 // Get widened type and narrowed type. 6783 MVT widenType; 6784 unsigned numElem = VT.getVectorNumElements(); 6785 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) { 6786 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break; 6787 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break; 6788 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break; 6789 default: 6790 llvm_unreachable("Invalid vector element type for padd optimization."); 6791 } 6792 6793 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(), 6794 widenType, &Ops[0], Ops.size()); 6795 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp); 6796 } 6797 6798 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with 6799 /// operands N0 and N1. This is a helper for PerformADDCombine that is 6800 /// called with the default operands, and if that fails, with commuted 6801 /// operands. 6802 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1, 6803 TargetLowering::DAGCombinerInfo &DCI, 6804 const ARMSubtarget *Subtarget){ 6805 6806 // Attempt to create vpaddl for this add. 6807 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget); 6808 if (Result.getNode()) 6809 return Result; 6810 6811 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c)) 6812 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) { 6813 SDValue Result = combineSelectAndUse(N, N0, N1, DCI); 6814 if (Result.getNode()) return Result; 6815 } 6816 return SDValue(); 6817 } 6818 6819 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD. 6820 /// 6821 static SDValue PerformADDCombine(SDNode *N, 6822 TargetLowering::DAGCombinerInfo &DCI, 6823 const ARMSubtarget *Subtarget) { 6824 SDValue N0 = N->getOperand(0); 6825 SDValue N1 = N->getOperand(1); 6826 6827 // First try with the default operand order. 6828 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget); 6829 if (Result.getNode()) 6830 return Result; 6831 6832 // If that didn't work, try again with the operands commuted. 6833 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget); 6834 } 6835 6836 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB. 6837 /// 6838 static SDValue PerformSUBCombine(SDNode *N, 6839 TargetLowering::DAGCombinerInfo &DCI) { 6840 SDValue N0 = N->getOperand(0); 6841 SDValue N1 = N->getOperand(1); 6842 6843 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c)) 6844 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) { 6845 SDValue Result = combineSelectAndUse(N, N1, N0, DCI); 6846 if (Result.getNode()) return Result; 6847 } 6848 6849 return SDValue(); 6850 } 6851 6852 /// PerformVMULCombine 6853 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the 6854 /// special multiplier accumulator forwarding. 6855 /// vmul d3, d0, d2 6856 /// vmla d3, d1, d2 6857 /// is faster than 6858 /// vadd d3, d0, d1 6859 /// vmul d3, d3, d2 6860 static SDValue PerformVMULCombine(SDNode *N, 6861 TargetLowering::DAGCombinerInfo &DCI, 6862 const ARMSubtarget *Subtarget) { 6863 if (!Subtarget->hasVMLxForwarding()) 6864 return SDValue(); 6865 6866 SelectionDAG &DAG = DCI.DAG; 6867 SDValue N0 = N->getOperand(0); 6868 SDValue N1 = N->getOperand(1); 6869 unsigned Opcode = N0.getOpcode(); 6870 if (Opcode != ISD::ADD && Opcode != ISD::SUB && 6871 Opcode != ISD::FADD && Opcode != ISD::FSUB) { 6872 Opcode = N1.getOpcode(); 6873 if (Opcode != ISD::ADD && Opcode != ISD::SUB && 6874 Opcode != ISD::FADD && Opcode != ISD::FSUB) 6875 return SDValue(); 6876 std::swap(N0, N1); 6877 } 6878 6879 EVT VT = N->getValueType(0); 6880 DebugLoc DL = N->getDebugLoc(); 6881 SDValue N00 = N0->getOperand(0); 6882 SDValue N01 = N0->getOperand(1); 6883 return DAG.getNode(Opcode, DL, VT, 6884 DAG.getNode(ISD::MUL, DL, VT, N00, N1), 6885 DAG.getNode(ISD::MUL, DL, VT, N01, N1)); 6886 } 6887 6888 static SDValue PerformMULCombine(SDNode *N, 6889 TargetLowering::DAGCombinerInfo &DCI, 6890 const ARMSubtarget *Subtarget) { 6891 SelectionDAG &DAG = DCI.DAG; 6892 6893 if (Subtarget->isThumb1Only()) 6894 return SDValue(); 6895 6896 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer()) 6897 return SDValue(); 6898 6899 EVT VT = N->getValueType(0); 6900 if (VT.is64BitVector() || VT.is128BitVector()) 6901 return PerformVMULCombine(N, DCI, Subtarget); 6902 if (VT != MVT::i32) 6903 return SDValue(); 6904 6905 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)); 6906 if (!C) 6907 return SDValue(); 6908 6909 int64_t MulAmt = C->getSExtValue(); 6910 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt); 6911 6912 ShiftAmt = ShiftAmt & (32 - 1); 6913 SDValue V = N->getOperand(0); 6914 DebugLoc DL = N->getDebugLoc(); 6915 6916 SDValue Res; 6917 MulAmt >>= ShiftAmt; 6918 6919 if (MulAmt >= 0) { 6920 if (isPowerOf2_32(MulAmt - 1)) { 6921 // (mul x, 2^N + 1) => (add (shl x, N), x) 6922 Res = DAG.getNode(ISD::ADD, DL, VT, 6923 V, 6924 DAG.getNode(ISD::SHL, DL, VT, 6925 V, 6926 DAG.getConstant(Log2_32(MulAmt - 1), 6927 MVT::i32))); 6928 } else if (isPowerOf2_32(MulAmt + 1)) { 6929 // (mul x, 2^N - 1) => (sub (shl x, N), x) 6930 Res = DAG.getNode(ISD::SUB, DL, VT, 6931 DAG.getNode(ISD::SHL, DL, VT, 6932 V, 6933 DAG.getConstant(Log2_32(MulAmt + 1), 6934 MVT::i32)), 6935 V); 6936 } else 6937 return SDValue(); 6938 } else { 6939 uint64_t MulAmtAbs = -MulAmt; 6940 if (isPowerOf2_32(MulAmtAbs + 1)) { 6941 // (mul x, -(2^N - 1)) => (sub x, (shl x, N)) 6942 Res = DAG.getNode(ISD::SUB, DL, VT, 6943 V, 6944 DAG.getNode(ISD::SHL, DL, VT, 6945 V, 6946 DAG.getConstant(Log2_32(MulAmtAbs + 1), 6947 MVT::i32))); 6948 } else if (isPowerOf2_32(MulAmtAbs - 1)) { 6949 // (mul x, -(2^N + 1)) => - (add (shl x, N), x) 6950 Res = DAG.getNode(ISD::ADD, DL, VT, 6951 V, 6952 DAG.getNode(ISD::SHL, DL, VT, 6953 V, 6954 DAG.getConstant(Log2_32(MulAmtAbs-1), 6955 MVT::i32))); 6956 Res = DAG.getNode(ISD::SUB, DL, VT, 6957 DAG.getConstant(0, MVT::i32),Res); 6958 6959 } else 6960 return SDValue(); 6961 } 6962 6963 if (ShiftAmt != 0) 6964 Res = DAG.getNode(ISD::SHL, DL, VT, 6965 Res, DAG.getConstant(ShiftAmt, MVT::i32)); 6966 6967 // Do not add new nodes to DAG combiner worklist. 6968 DCI.CombineTo(N, Res, false); 6969 return SDValue(); 6970 } 6971 6972 static bool isCMOVWithZeroOrAllOnesLHS(SDValue N, bool AllOnes) { 6973 if (N.getOpcode() != ARMISD::CMOV || !N.getNode()->hasOneUse()) 6974 return false; 6975 6976 SDValue FalseVal = N.getOperand(0); 6977 ConstantSDNode *C = dyn_cast<ConstantSDNode>(FalseVal); 6978 if (!C) 6979 return false; 6980 if (AllOnes) 6981 return C->isAllOnesValue(); 6982 return C->isNullValue(); 6983 } 6984 6985 /// formConditionalOp - Combine an operation with a conditional move operand 6986 /// to form a conditional op. e.g. (or x, (cmov 0, y, cond)) => (or.cond x, y) 6987 /// (and x, (cmov -1, y, cond)) => (and.cond, x, y) 6988 static SDValue formConditionalOp(SDNode *N, SelectionDAG &DAG, 6989 bool Commutable) { 6990 SDValue N0 = N->getOperand(0); 6991 SDValue N1 = N->getOperand(1); 6992 6993 bool isAND = N->getOpcode() == ISD::AND; 6994 bool isCand = isCMOVWithZeroOrAllOnesLHS(N1, isAND); 6995 if (!isCand && Commutable) { 6996 isCand = isCMOVWithZeroOrAllOnesLHS(N0, isAND); 6997 if (isCand) 6998 std::swap(N0, N1); 6999 } 7000 if (!isCand) 7001 return SDValue(); 7002 7003 unsigned Opc = 0; 7004 switch (N->getOpcode()) { 7005 default: llvm_unreachable("Unexpected node"); 7006 case ISD::AND: Opc = ARMISD::CAND; break; 7007 case ISD::OR: Opc = ARMISD::COR; break; 7008 case ISD::XOR: Opc = ARMISD::CXOR; break; 7009 } 7010 return DAG.getNode(Opc, N->getDebugLoc(), N->getValueType(0), N0, 7011 N1.getOperand(1), N1.getOperand(2), N1.getOperand(3), 7012 N1.getOperand(4)); 7013 } 7014 7015 static SDValue PerformANDCombine(SDNode *N, 7016 TargetLowering::DAGCombinerInfo &DCI, 7017 const ARMSubtarget *Subtarget) { 7018 7019 // Attempt to use immediate-form VBIC 7020 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1)); 7021 DebugLoc dl = N->getDebugLoc(); 7022 EVT VT = N->getValueType(0); 7023 SelectionDAG &DAG = DCI.DAG; 7024 7025 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT)) 7026 return SDValue(); 7027 7028 APInt SplatBits, SplatUndef; 7029 unsigned SplatBitSize; 7030 bool HasAnyUndefs; 7031 if (BVN && 7032 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) { 7033 if (SplatBitSize <= 64) { 7034 EVT VbicVT; 7035 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(), 7036 SplatUndef.getZExtValue(), SplatBitSize, 7037 DAG, VbicVT, VT.is128BitVector(), 7038 OtherModImm); 7039 if (Val.getNode()) { 7040 SDValue Input = 7041 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0)); 7042 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val); 7043 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic); 7044 } 7045 } 7046 } 7047 7048 if (!Subtarget->isThumb1Only()) { 7049 // (and x, (cmov -1, y, cond)) => (and.cond x, y) 7050 SDValue CAND = formConditionalOp(N, DAG, true); 7051 if (CAND.getNode()) 7052 return CAND; 7053 } 7054 7055 return SDValue(); 7056 } 7057 7058 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR 7059 static SDValue PerformORCombine(SDNode *N, 7060 TargetLowering::DAGCombinerInfo &DCI, 7061 const ARMSubtarget *Subtarget) { 7062 // Attempt to use immediate-form VORR 7063 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1)); 7064 DebugLoc dl = N->getDebugLoc(); 7065 EVT VT = N->getValueType(0); 7066 SelectionDAG &DAG = DCI.DAG; 7067 7068 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT)) 7069 return SDValue(); 7070 7071 APInt SplatBits, SplatUndef; 7072 unsigned SplatBitSize; 7073 bool HasAnyUndefs; 7074 if (BVN && Subtarget->hasNEON() && 7075 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) { 7076 if (SplatBitSize <= 64) { 7077 EVT VorrVT; 7078 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(), 7079 SplatUndef.getZExtValue(), SplatBitSize, 7080 DAG, VorrVT, VT.is128BitVector(), 7081 OtherModImm); 7082 if (Val.getNode()) { 7083 SDValue Input = 7084 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0)); 7085 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val); 7086 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr); 7087 } 7088 } 7089 } 7090 7091 if (!Subtarget->isThumb1Only()) { 7092 // (or x, (cmov 0, y, cond)) => (or.cond x, y) 7093 SDValue COR = formConditionalOp(N, DAG, true); 7094 if (COR.getNode()) 7095 return COR; 7096 } 7097 7098 SDValue N0 = N->getOperand(0); 7099 if (N0.getOpcode() != ISD::AND) 7100 return SDValue(); 7101 SDValue N1 = N->getOperand(1); 7102 7103 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant. 7104 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() && 7105 DAG.getTargetLoweringInfo().isTypeLegal(VT)) { 7106 APInt SplatUndef; 7107 unsigned SplatBitSize; 7108 bool HasAnyUndefs; 7109 7110 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1)); 7111 APInt SplatBits0; 7112 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize, 7113 HasAnyUndefs) && !HasAnyUndefs) { 7114 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1)); 7115 APInt SplatBits1; 7116 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize, 7117 HasAnyUndefs) && !HasAnyUndefs && 7118 SplatBits0 == ~SplatBits1) { 7119 // Canonicalize the vector type to make instruction selection simpler. 7120 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32; 7121 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT, 7122 N0->getOperand(1), N0->getOperand(0), 7123 N1->getOperand(0)); 7124 return DAG.getNode(ISD::BITCAST, dl, VT, Result); 7125 } 7126 } 7127 } 7128 7129 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when 7130 // reasonable. 7131 7132 // BFI is only available on V6T2+ 7133 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops()) 7134 return SDValue(); 7135 7136 DebugLoc DL = N->getDebugLoc(); 7137 // 1) or (and A, mask), val => ARMbfi A, val, mask 7138 // iff (val & mask) == val 7139 // 7140 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask 7141 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2) 7142 // && mask == ~mask2 7143 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2) 7144 // && ~mask == mask2 7145 // (i.e., copy a bitfield value into another bitfield of the same width) 7146 7147 if (VT != MVT::i32) 7148 return SDValue(); 7149 7150 SDValue N00 = N0.getOperand(0); 7151 7152 // The value and the mask need to be constants so we can verify this is 7153 // actually a bitfield set. If the mask is 0xffff, we can do better 7154 // via a movt instruction, so don't use BFI in that case. 7155 SDValue MaskOp = N0.getOperand(1); 7156 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp); 7157 if (!MaskC) 7158 return SDValue(); 7159 unsigned Mask = MaskC->getZExtValue(); 7160 if (Mask == 0xffff) 7161 return SDValue(); 7162 SDValue Res; 7163 // Case (1): or (and A, mask), val => ARMbfi A, val, mask 7164 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 7165 if (N1C) { 7166 unsigned Val = N1C->getZExtValue(); 7167 if ((Val & ~Mask) != Val) 7168 return SDValue(); 7169 7170 if (ARM::isBitFieldInvertedMask(Mask)) { 7171 Val >>= CountTrailingZeros_32(~Mask); 7172 7173 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, 7174 DAG.getConstant(Val, MVT::i32), 7175 DAG.getConstant(Mask, MVT::i32)); 7176 7177 // Do not add new nodes to DAG combiner worklist. 7178 DCI.CombineTo(N, Res, false); 7179 return SDValue(); 7180 } 7181 } else if (N1.getOpcode() == ISD::AND) { 7182 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask 7183 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); 7184 if (!N11C) 7185 return SDValue(); 7186 unsigned Mask2 = N11C->getZExtValue(); 7187 7188 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern 7189 // as is to match. 7190 if (ARM::isBitFieldInvertedMask(Mask) && 7191 (Mask == ~Mask2)) { 7192 // The pack halfword instruction works better for masks that fit it, 7193 // so use that when it's available. 7194 if (Subtarget->hasT2ExtractPack() && 7195 (Mask == 0xffff || Mask == 0xffff0000)) 7196 return SDValue(); 7197 // 2a 7198 unsigned amt = CountTrailingZeros_32(Mask2); 7199 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0), 7200 DAG.getConstant(amt, MVT::i32)); 7201 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res, 7202 DAG.getConstant(Mask, MVT::i32)); 7203 // Do not add new nodes to DAG combiner worklist. 7204 DCI.CombineTo(N, Res, false); 7205 return SDValue(); 7206 } else if (ARM::isBitFieldInvertedMask(~Mask) && 7207 (~Mask == Mask2)) { 7208 // The pack halfword instruction works better for masks that fit it, 7209 // so use that when it's available. 7210 if (Subtarget->hasT2ExtractPack() && 7211 (Mask2 == 0xffff || Mask2 == 0xffff0000)) 7212 return SDValue(); 7213 // 2b 7214 unsigned lsb = CountTrailingZeros_32(Mask); 7215 Res = DAG.getNode(ISD::SRL, DL, VT, N00, 7216 DAG.getConstant(lsb, MVT::i32)); 7217 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res, 7218 DAG.getConstant(Mask2, MVT::i32)); 7219 // Do not add new nodes to DAG combiner worklist. 7220 DCI.CombineTo(N, Res, false); 7221 return SDValue(); 7222 } 7223 } 7224 7225 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) && 7226 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) && 7227 ARM::isBitFieldInvertedMask(~Mask)) { 7228 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask 7229 // where lsb(mask) == #shamt and masked bits of B are known zero. 7230 SDValue ShAmt = N00.getOperand(1); 7231 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue(); 7232 unsigned LSB = CountTrailingZeros_32(Mask); 7233 if (ShAmtC != LSB) 7234 return SDValue(); 7235 7236 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0), 7237 DAG.getConstant(~Mask, MVT::i32)); 7238 7239 // Do not add new nodes to DAG combiner worklist. 7240 DCI.CombineTo(N, Res, false); 7241 } 7242 7243 return SDValue(); 7244 } 7245 7246 static SDValue PerformXORCombine(SDNode *N, 7247 TargetLowering::DAGCombinerInfo &DCI, 7248 const ARMSubtarget *Subtarget) { 7249 EVT VT = N->getValueType(0); 7250 SelectionDAG &DAG = DCI.DAG; 7251 7252 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT)) 7253 return SDValue(); 7254 7255 if (!Subtarget->isThumb1Only()) { 7256 // (xor x, (cmov 0, y, cond)) => (xor.cond x, y) 7257 SDValue CXOR = formConditionalOp(N, DAG, true); 7258 if (CXOR.getNode()) 7259 return CXOR; 7260 } 7261 7262 return SDValue(); 7263 } 7264 7265 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff 7266 /// the bits being cleared by the AND are not demanded by the BFI. 7267 static SDValue PerformBFICombine(SDNode *N, 7268 TargetLowering::DAGCombinerInfo &DCI) { 7269 SDValue N1 = N->getOperand(1); 7270 if (N1.getOpcode() == ISD::AND) { 7271 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); 7272 if (!N11C) 7273 return SDValue(); 7274 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue(); 7275 unsigned LSB = CountTrailingZeros_32(~InvMask); 7276 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB; 7277 unsigned Mask = (1 << Width)-1; 7278 unsigned Mask2 = N11C->getZExtValue(); 7279 if ((Mask & (~Mask2)) == 0) 7280 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0), 7281 N->getOperand(0), N1.getOperand(0), 7282 N->getOperand(2)); 7283 } 7284 return SDValue(); 7285 } 7286 7287 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for 7288 /// ARMISD::VMOVRRD. 7289 static SDValue PerformVMOVRRDCombine(SDNode *N, 7290 TargetLowering::DAGCombinerInfo &DCI) { 7291 // vmovrrd(vmovdrr x, y) -> x,y 7292 SDValue InDouble = N->getOperand(0); 7293 if (InDouble.getOpcode() == ARMISD::VMOVDRR) 7294 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1)); 7295 7296 // vmovrrd(load f64) -> (load i32), (load i32) 7297 SDNode *InNode = InDouble.getNode(); 7298 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() && 7299 InNode->getValueType(0) == MVT::f64 && 7300 InNode->getOperand(1).getOpcode() == ISD::FrameIndex && 7301 !cast<LoadSDNode>(InNode)->isVolatile()) { 7302 // TODO: Should this be done for non-FrameIndex operands? 7303 LoadSDNode *LD = cast<LoadSDNode>(InNode); 7304 7305 SelectionDAG &DAG = DCI.DAG; 7306 DebugLoc DL = LD->getDebugLoc(); 7307 SDValue BasePtr = LD->getBasePtr(); 7308 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr, 7309 LD->getPointerInfo(), LD->isVolatile(), 7310 LD->isNonTemporal(), LD->isInvariant(), 7311 LD->getAlignment()); 7312 7313 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr, 7314 DAG.getConstant(4, MVT::i32)); 7315 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr, 7316 LD->getPointerInfo(), LD->isVolatile(), 7317 LD->isNonTemporal(), LD->isInvariant(), 7318 std::min(4U, LD->getAlignment() / 2)); 7319 7320 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1)); 7321 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2); 7322 DCI.RemoveFromWorklist(LD); 7323 DAG.DeleteNode(LD); 7324 return Result; 7325 } 7326 7327 return SDValue(); 7328 } 7329 7330 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for 7331 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands. 7332 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) { 7333 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X) 7334 SDValue Op0 = N->getOperand(0); 7335 SDValue Op1 = N->getOperand(1); 7336 if (Op0.getOpcode() == ISD::BITCAST) 7337 Op0 = Op0.getOperand(0); 7338 if (Op1.getOpcode() == ISD::BITCAST) 7339 Op1 = Op1.getOperand(0); 7340 if (Op0.getOpcode() == ARMISD::VMOVRRD && 7341 Op0.getNode() == Op1.getNode() && 7342 Op0.getResNo() == 0 && Op1.getResNo() == 1) 7343 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), 7344 N->getValueType(0), Op0.getOperand(0)); 7345 return SDValue(); 7346 } 7347 7348 /// PerformSTORECombine - Target-specific dag combine xforms for 7349 /// ISD::STORE. 7350 static SDValue PerformSTORECombine(SDNode *N, 7351 TargetLowering::DAGCombinerInfo &DCI) { 7352 StoreSDNode *St = cast<StoreSDNode>(N); 7353 if (St->isVolatile()) 7354 return SDValue(); 7355 7356 // Optimize trunc store (of multiple scalars) to shuffle and store. First, 7357 // pack all of the elements in one place. Next, store to memory in fewer 7358 // chunks. 7359 SDValue StVal = St->getValue(); 7360 EVT VT = StVal.getValueType(); 7361 if (St->isTruncatingStore() && VT.isVector()) { 7362 SelectionDAG &DAG = DCI.DAG; 7363 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 7364 EVT StVT = St->getMemoryVT(); 7365 unsigned NumElems = VT.getVectorNumElements(); 7366 assert(StVT != VT && "Cannot truncate to the same type"); 7367 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits(); 7368 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits(); 7369 7370 // From, To sizes and ElemCount must be pow of two 7371 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue(); 7372 7373 // We are going to use the original vector elt for storing. 7374 // Accumulated smaller vector elements must be a multiple of the store size. 7375 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue(); 7376 7377 unsigned SizeRatio = FromEltSz / ToEltSz; 7378 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits()); 7379 7380 // Create a type on which we perform the shuffle. 7381 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(), 7382 NumElems*SizeRatio); 7383 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits()); 7384 7385 DebugLoc DL = St->getDebugLoc(); 7386 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal); 7387 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1); 7388 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio; 7389 7390 // Can't shuffle using an illegal type. 7391 if (!TLI.isTypeLegal(WideVecVT)) return SDValue(); 7392 7393 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec, 7394 DAG.getUNDEF(WideVec.getValueType()), 7395 ShuffleVec.data()); 7396 // At this point all of the data is stored at the bottom of the 7397 // register. We now need to save it to mem. 7398 7399 // Find the largest store unit 7400 MVT StoreType = MVT::i8; 7401 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE; 7402 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) { 7403 MVT Tp = (MVT::SimpleValueType)tp; 7404 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz) 7405 StoreType = Tp; 7406 } 7407 // Didn't find a legal store type. 7408 if (!TLI.isTypeLegal(StoreType)) 7409 return SDValue(); 7410 7411 // Bitcast the original vector into a vector of store-size units 7412 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(), 7413 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits()); 7414 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits()); 7415 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff); 7416 SmallVector<SDValue, 8> Chains; 7417 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8, 7418 TLI.getPointerTy()); 7419 SDValue BasePtr = St->getBasePtr(); 7420 7421 // Perform one or more big stores into memory. 7422 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits(); 7423 for (unsigned I = 0; I < E; I++) { 7424 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, 7425 StoreType, ShuffWide, 7426 DAG.getIntPtrConstant(I)); 7427 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr, 7428 St->getPointerInfo(), St->isVolatile(), 7429 St->isNonTemporal(), St->getAlignment()); 7430 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr, 7431 Increment); 7432 Chains.push_back(Ch); 7433 } 7434 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0], 7435 Chains.size()); 7436 } 7437 7438 if (!ISD::isNormalStore(St)) 7439 return SDValue(); 7440 7441 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and 7442 // ARM stores of arguments in the same cache line. 7443 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR && 7444 StVal.getNode()->hasOneUse()) { 7445 SelectionDAG &DAG = DCI.DAG; 7446 DebugLoc DL = St->getDebugLoc(); 7447 SDValue BasePtr = St->getBasePtr(); 7448 SDValue NewST1 = DAG.getStore(St->getChain(), DL, 7449 StVal.getNode()->getOperand(0), BasePtr, 7450 St->getPointerInfo(), St->isVolatile(), 7451 St->isNonTemporal(), St->getAlignment()); 7452 7453 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr, 7454 DAG.getConstant(4, MVT::i32)); 7455 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1), 7456 OffsetPtr, St->getPointerInfo(), St->isVolatile(), 7457 St->isNonTemporal(), 7458 std::min(4U, St->getAlignment() / 2)); 7459 } 7460 7461 if (StVal.getValueType() != MVT::i64 || 7462 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT) 7463 return SDValue(); 7464 7465 // Bitcast an i64 store extracted from a vector to f64. 7466 // Otherwise, the i64 value will be legalized to a pair of i32 values. 7467 SelectionDAG &DAG = DCI.DAG; 7468 DebugLoc dl = StVal.getDebugLoc(); 7469 SDValue IntVec = StVal.getOperand(0); 7470 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, 7471 IntVec.getValueType().getVectorNumElements()); 7472 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec); 7473 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, 7474 Vec, StVal.getOperand(1)); 7475 dl = N->getDebugLoc(); 7476 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt); 7477 // Make the DAGCombiner fold the bitcasts. 7478 DCI.AddToWorklist(Vec.getNode()); 7479 DCI.AddToWorklist(ExtElt.getNode()); 7480 DCI.AddToWorklist(V.getNode()); 7481 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(), 7482 St->getPointerInfo(), St->isVolatile(), 7483 St->isNonTemporal(), St->getAlignment(), 7484 St->getTBAAInfo()); 7485 } 7486 7487 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node 7488 /// are normal, non-volatile loads. If so, it is profitable to bitcast an 7489 /// i64 vector to have f64 elements, since the value can then be loaded 7490 /// directly into a VFP register. 7491 static bool hasNormalLoadOperand(SDNode *N) { 7492 unsigned NumElts = N->getValueType(0).getVectorNumElements(); 7493 for (unsigned i = 0; i < NumElts; ++i) { 7494 SDNode *Elt = N->getOperand(i).getNode(); 7495 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile()) 7496 return true; 7497 } 7498 return false; 7499 } 7500 7501 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for 7502 /// ISD::BUILD_VECTOR. 7503 static SDValue PerformBUILD_VECTORCombine(SDNode *N, 7504 TargetLowering::DAGCombinerInfo &DCI){ 7505 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X): 7506 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value 7507 // into a pair of GPRs, which is fine when the value is used as a scalar, 7508 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD. 7509 SelectionDAG &DAG = DCI.DAG; 7510 if (N->getNumOperands() == 2) { 7511 SDValue RV = PerformVMOVDRRCombine(N, DAG); 7512 if (RV.getNode()) 7513 return RV; 7514 } 7515 7516 // Load i64 elements as f64 values so that type legalization does not split 7517 // them up into i32 values. 7518 EVT VT = N->getValueType(0); 7519 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N)) 7520 return SDValue(); 7521 DebugLoc dl = N->getDebugLoc(); 7522 SmallVector<SDValue, 8> Ops; 7523 unsigned NumElts = VT.getVectorNumElements(); 7524 for (unsigned i = 0; i < NumElts; ++i) { 7525 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i)); 7526 Ops.push_back(V); 7527 // Make the DAGCombiner fold the bitcast. 7528 DCI.AddToWorklist(V.getNode()); 7529 } 7530 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts); 7531 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts); 7532 return DAG.getNode(ISD::BITCAST, dl, VT, BV); 7533 } 7534 7535 /// PerformInsertEltCombine - Target-specific dag combine xforms for 7536 /// ISD::INSERT_VECTOR_ELT. 7537 static SDValue PerformInsertEltCombine(SDNode *N, 7538 TargetLowering::DAGCombinerInfo &DCI) { 7539 // Bitcast an i64 load inserted into a vector to f64. 7540 // Otherwise, the i64 value will be legalized to a pair of i32 values. 7541 EVT VT = N->getValueType(0); 7542 SDNode *Elt = N->getOperand(1).getNode(); 7543 if (VT.getVectorElementType() != MVT::i64 || 7544 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile()) 7545 return SDValue(); 7546 7547 SelectionDAG &DAG = DCI.DAG; 7548 DebugLoc dl = N->getDebugLoc(); 7549 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, 7550 VT.getVectorNumElements()); 7551 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0)); 7552 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1)); 7553 // Make the DAGCombiner fold the bitcasts. 7554 DCI.AddToWorklist(Vec.getNode()); 7555 DCI.AddToWorklist(V.getNode()); 7556 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT, 7557 Vec, V, N->getOperand(2)); 7558 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt); 7559 } 7560 7561 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for 7562 /// ISD::VECTOR_SHUFFLE. 7563 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) { 7564 // The LLVM shufflevector instruction does not require the shuffle mask 7565 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does 7566 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the 7567 // operands do not match the mask length, they are extended by concatenating 7568 // them with undef vectors. That is probably the right thing for other 7569 // targets, but for NEON it is better to concatenate two double-register 7570 // size vector operands into a single quad-register size vector. Do that 7571 // transformation here: 7572 // shuffle(concat(v1, undef), concat(v2, undef)) -> 7573 // shuffle(concat(v1, v2), undef) 7574 SDValue Op0 = N->getOperand(0); 7575 SDValue Op1 = N->getOperand(1); 7576 if (Op0.getOpcode() != ISD::CONCAT_VECTORS || 7577 Op1.getOpcode() != ISD::CONCAT_VECTORS || 7578 Op0.getNumOperands() != 2 || 7579 Op1.getNumOperands() != 2) 7580 return SDValue(); 7581 SDValue Concat0Op1 = Op0.getOperand(1); 7582 SDValue Concat1Op1 = Op1.getOperand(1); 7583 if (Concat0Op1.getOpcode() != ISD::UNDEF || 7584 Concat1Op1.getOpcode() != ISD::UNDEF) 7585 return SDValue(); 7586 // Skip the transformation if any of the types are illegal. 7587 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 7588 EVT VT = N->getValueType(0); 7589 if (!TLI.isTypeLegal(VT) || 7590 !TLI.isTypeLegal(Concat0Op1.getValueType()) || 7591 !TLI.isTypeLegal(Concat1Op1.getValueType())) 7592 return SDValue(); 7593 7594 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT, 7595 Op0.getOperand(0), Op1.getOperand(0)); 7596 // Translate the shuffle mask. 7597 SmallVector<int, 16> NewMask; 7598 unsigned NumElts = VT.getVectorNumElements(); 7599 unsigned HalfElts = NumElts/2; 7600 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 7601 for (unsigned n = 0; n < NumElts; ++n) { 7602 int MaskElt = SVN->getMaskElt(n); 7603 int NewElt = -1; 7604 if (MaskElt < (int)HalfElts) 7605 NewElt = MaskElt; 7606 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts)) 7607 NewElt = HalfElts + MaskElt - NumElts; 7608 NewMask.push_back(NewElt); 7609 } 7610 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat, 7611 DAG.getUNDEF(VT), NewMask.data()); 7612 } 7613 7614 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and 7615 /// NEON load/store intrinsics to merge base address updates. 7616 static SDValue CombineBaseUpdate(SDNode *N, 7617 TargetLowering::DAGCombinerInfo &DCI) { 7618 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer()) 7619 return SDValue(); 7620 7621 SelectionDAG &DAG = DCI.DAG; 7622 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID || 7623 N->getOpcode() == ISD::INTRINSIC_W_CHAIN); 7624 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1); 7625 SDValue Addr = N->getOperand(AddrOpIdx); 7626 7627 // Search for a use of the address operand that is an increment. 7628 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), 7629 UE = Addr.getNode()->use_end(); UI != UE; ++UI) { 7630 SDNode *User = *UI; 7631 if (User->getOpcode() != ISD::ADD || 7632 UI.getUse().getResNo() != Addr.getResNo()) 7633 continue; 7634 7635 // Check that the add is independent of the load/store. Otherwise, folding 7636 // it would create a cycle. 7637 if (User->isPredecessorOf(N) || N->isPredecessorOf(User)) 7638 continue; 7639 7640 // Find the new opcode for the updating load/store. 7641 bool isLoad = true; 7642 bool isLaneOp = false; 7643 unsigned NewOpc = 0; 7644 unsigned NumVecs = 0; 7645 if (isIntrinsic) { 7646 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); 7647 switch (IntNo) { 7648 default: llvm_unreachable("unexpected intrinsic for Neon base update"); 7649 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD; 7650 NumVecs = 1; break; 7651 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD; 7652 NumVecs = 2; break; 7653 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD; 7654 NumVecs = 3; break; 7655 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD; 7656 NumVecs = 4; break; 7657 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD; 7658 NumVecs = 2; isLaneOp = true; break; 7659 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD; 7660 NumVecs = 3; isLaneOp = true; break; 7661 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD; 7662 NumVecs = 4; isLaneOp = true; break; 7663 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD; 7664 NumVecs = 1; isLoad = false; break; 7665 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD; 7666 NumVecs = 2; isLoad = false; break; 7667 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD; 7668 NumVecs = 3; isLoad = false; break; 7669 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD; 7670 NumVecs = 4; isLoad = false; break; 7671 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD; 7672 NumVecs = 2; isLoad = false; isLaneOp = true; break; 7673 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD; 7674 NumVecs = 3; isLoad = false; isLaneOp = true; break; 7675 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD; 7676 NumVecs = 4; isLoad = false; isLaneOp = true; break; 7677 } 7678 } else { 7679 isLaneOp = true; 7680 switch (N->getOpcode()) { 7681 default: llvm_unreachable("unexpected opcode for Neon base update"); 7682 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break; 7683 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break; 7684 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break; 7685 } 7686 } 7687 7688 // Find the size of memory referenced by the load/store. 7689 EVT VecTy; 7690 if (isLoad) 7691 VecTy = N->getValueType(0); 7692 else 7693 VecTy = N->getOperand(AddrOpIdx+1).getValueType(); 7694 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8; 7695 if (isLaneOp) 7696 NumBytes /= VecTy.getVectorNumElements(); 7697 7698 // If the increment is a constant, it must match the memory ref size. 7699 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0); 7700 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) { 7701 uint64_t IncVal = CInc->getZExtValue(); 7702 if (IncVal != NumBytes) 7703 continue; 7704 } else if (NumBytes >= 3 * 16) { 7705 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two 7706 // separate instructions that make it harder to use a non-constant update. 7707 continue; 7708 } 7709 7710 // Create the new updating load/store node. 7711 EVT Tys[6]; 7712 unsigned NumResultVecs = (isLoad ? NumVecs : 0); 7713 unsigned n; 7714 for (n = 0; n < NumResultVecs; ++n) 7715 Tys[n] = VecTy; 7716 Tys[n++] = MVT::i32; 7717 Tys[n] = MVT::Other; 7718 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2); 7719 SmallVector<SDValue, 8> Ops; 7720 Ops.push_back(N->getOperand(0)); // incoming chain 7721 Ops.push_back(N->getOperand(AddrOpIdx)); 7722 Ops.push_back(Inc); 7723 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) { 7724 Ops.push_back(N->getOperand(i)); 7725 } 7726 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N); 7727 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys, 7728 Ops.data(), Ops.size(), 7729 MemInt->getMemoryVT(), 7730 MemInt->getMemOperand()); 7731 7732 // Update the uses. 7733 std::vector<SDValue> NewResults; 7734 for (unsigned i = 0; i < NumResultVecs; ++i) { 7735 NewResults.push_back(SDValue(UpdN.getNode(), i)); 7736 } 7737 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain 7738 DCI.CombineTo(N, NewResults); 7739 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs)); 7740 7741 break; 7742 } 7743 return SDValue(); 7744 } 7745 7746 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a 7747 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic 7748 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and 7749 /// return true. 7750 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { 7751 SelectionDAG &DAG = DCI.DAG; 7752 EVT VT = N->getValueType(0); 7753 // vldN-dup instructions only support 64-bit vectors for N > 1. 7754 if (!VT.is64BitVector()) 7755 return false; 7756 7757 // Check if the VDUPLANE operand is a vldN-dup intrinsic. 7758 SDNode *VLD = N->getOperand(0).getNode(); 7759 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN) 7760 return false; 7761 unsigned NumVecs = 0; 7762 unsigned NewOpc = 0; 7763 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue(); 7764 if (IntNo == Intrinsic::arm_neon_vld2lane) { 7765 NumVecs = 2; 7766 NewOpc = ARMISD::VLD2DUP; 7767 } else if (IntNo == Intrinsic::arm_neon_vld3lane) { 7768 NumVecs = 3; 7769 NewOpc = ARMISD::VLD3DUP; 7770 } else if (IntNo == Intrinsic::arm_neon_vld4lane) { 7771 NumVecs = 4; 7772 NewOpc = ARMISD::VLD4DUP; 7773 } else { 7774 return false; 7775 } 7776 7777 // First check that all the vldN-lane uses are VDUPLANEs and that the lane 7778 // numbers match the load. 7779 unsigned VLDLaneNo = 7780 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue(); 7781 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end(); 7782 UI != UE; ++UI) { 7783 // Ignore uses of the chain result. 7784 if (UI.getUse().getResNo() == NumVecs) 7785 continue; 7786 SDNode *User = *UI; 7787 if (User->getOpcode() != ARMISD::VDUPLANE || 7788 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue()) 7789 return false; 7790 } 7791 7792 // Create the vldN-dup node. 7793 EVT Tys[5]; 7794 unsigned n; 7795 for (n = 0; n < NumVecs; ++n) 7796 Tys[n] = VT; 7797 Tys[n] = MVT::Other; 7798 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1); 7799 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) }; 7800 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD); 7801 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys, 7802 Ops, 2, VLDMemInt->getMemoryVT(), 7803 VLDMemInt->getMemOperand()); 7804 7805 // Update the uses. 7806 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end(); 7807 UI != UE; ++UI) { 7808 unsigned ResNo = UI.getUse().getResNo(); 7809 // Ignore uses of the chain result. 7810 if (ResNo == NumVecs) 7811 continue; 7812 SDNode *User = *UI; 7813 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo)); 7814 } 7815 7816 // Now the vldN-lane intrinsic is dead except for its chain result. 7817 // Update uses of the chain. 7818 std::vector<SDValue> VLDDupResults; 7819 for (unsigned n = 0; n < NumVecs; ++n) 7820 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n)); 7821 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs)); 7822 DCI.CombineTo(VLD, VLDDupResults); 7823 7824 return true; 7825 } 7826 7827 /// PerformVDUPLANECombine - Target-specific dag combine xforms for 7828 /// ARMISD::VDUPLANE. 7829 static SDValue PerformVDUPLANECombine(SDNode *N, 7830 TargetLowering::DAGCombinerInfo &DCI) { 7831 SDValue Op = N->getOperand(0); 7832 7833 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses 7834 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation. 7835 if (CombineVLDDUP(N, DCI)) 7836 return SDValue(N, 0); 7837 7838 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is 7839 // redundant. Ignore bit_converts for now; element sizes are checked below. 7840 while (Op.getOpcode() == ISD::BITCAST) 7841 Op = Op.getOperand(0); 7842 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM) 7843 return SDValue(); 7844 7845 // Make sure the VMOV element size is not bigger than the VDUPLANE elements. 7846 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits(); 7847 // The canonical VMOV for a zero vector uses a 32-bit element size. 7848 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 7849 unsigned EltBits; 7850 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0) 7851 EltSize = 8; 7852 EVT VT = N->getValueType(0); 7853 if (EltSize > VT.getVectorElementType().getSizeInBits()) 7854 return SDValue(); 7855 7856 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op); 7857 } 7858 7859 // isConstVecPow2 - Return true if each vector element is a power of 2, all 7860 // elements are the same constant, C, and Log2(C) ranges from 1 to 32. 7861 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C) 7862 { 7863 integerPart cN; 7864 integerPart c0 = 0; 7865 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements(); 7866 I != E; I++) { 7867 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I)); 7868 if (!C) 7869 return false; 7870 7871 bool isExact; 7872 APFloat APF = C->getValueAPF(); 7873 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact) 7874 != APFloat::opOK || !isExact) 7875 return false; 7876 7877 c0 = (I == 0) ? cN : c0; 7878 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32) 7879 return false; 7880 } 7881 C = c0; 7882 return true; 7883 } 7884 7885 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD) 7886 /// can replace combinations of VMUL and VCVT (floating-point to integer) 7887 /// when the VMUL has a constant operand that is a power of 2. 7888 /// 7889 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>): 7890 /// vmul.f32 d16, d17, d16 7891 /// vcvt.s32.f32 d16, d16 7892 /// becomes: 7893 /// vcvt.s32.f32 d16, d16, #3 7894 static SDValue PerformVCVTCombine(SDNode *N, 7895 TargetLowering::DAGCombinerInfo &DCI, 7896 const ARMSubtarget *Subtarget) { 7897 SelectionDAG &DAG = DCI.DAG; 7898 SDValue Op = N->getOperand(0); 7899 7900 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() || 7901 Op.getOpcode() != ISD::FMUL) 7902 return SDValue(); 7903 7904 uint64_t C; 7905 SDValue N0 = Op->getOperand(0); 7906 SDValue ConstVec = Op->getOperand(1); 7907 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT; 7908 7909 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR || 7910 !isConstVecPow2(ConstVec, isSigned, C)) 7911 return SDValue(); 7912 7913 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs : 7914 Intrinsic::arm_neon_vcvtfp2fxu; 7915 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(), 7916 N->getValueType(0), 7917 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0, 7918 DAG.getConstant(Log2_64(C), MVT::i32)); 7919 } 7920 7921 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD) 7922 /// can replace combinations of VCVT (integer to floating-point) and VDIV 7923 /// when the VDIV has a constant operand that is a power of 2. 7924 /// 7925 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>): 7926 /// vcvt.f32.s32 d16, d16 7927 /// vdiv.f32 d16, d17, d16 7928 /// becomes: 7929 /// vcvt.f32.s32 d16, d16, #3 7930 static SDValue PerformVDIVCombine(SDNode *N, 7931 TargetLowering::DAGCombinerInfo &DCI, 7932 const ARMSubtarget *Subtarget) { 7933 SelectionDAG &DAG = DCI.DAG; 7934 SDValue Op = N->getOperand(0); 7935 unsigned OpOpcode = Op.getNode()->getOpcode(); 7936 7937 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() || 7938 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP)) 7939 return SDValue(); 7940 7941 uint64_t C; 7942 SDValue ConstVec = N->getOperand(1); 7943 bool isSigned = OpOpcode == ISD::SINT_TO_FP; 7944 7945 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR || 7946 !isConstVecPow2(ConstVec, isSigned, C)) 7947 return SDValue(); 7948 7949 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp : 7950 Intrinsic::arm_neon_vcvtfxu2fp; 7951 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(), 7952 Op.getValueType(), 7953 DAG.getConstant(IntrinsicOpcode, MVT::i32), 7954 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32)); 7955 } 7956 7957 /// Getvshiftimm - Check if this is a valid build_vector for the immediate 7958 /// operand of a vector shift operation, where all the elements of the 7959 /// build_vector must have the same constant integer value. 7960 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) { 7961 // Ignore bit_converts. 7962 while (Op.getOpcode() == ISD::BITCAST) 7963 Op = Op.getOperand(0); 7964 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode()); 7965 APInt SplatBits, SplatUndef; 7966 unsigned SplatBitSize; 7967 bool HasAnyUndefs; 7968 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, 7969 HasAnyUndefs, ElementBits) || 7970 SplatBitSize > ElementBits) 7971 return false; 7972 Cnt = SplatBits.getSExtValue(); 7973 return true; 7974 } 7975 7976 /// isVShiftLImm - Check if this is a valid build_vector for the immediate 7977 /// operand of a vector shift left operation. That value must be in the range: 7978 /// 0 <= Value < ElementBits for a left shift; or 7979 /// 0 <= Value <= ElementBits for a long left shift. 7980 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) { 7981 assert(VT.isVector() && "vector shift count is not a vector type"); 7982 unsigned ElementBits = VT.getVectorElementType().getSizeInBits(); 7983 if (! getVShiftImm(Op, ElementBits, Cnt)) 7984 return false; 7985 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits); 7986 } 7987 7988 /// isVShiftRImm - Check if this is a valid build_vector for the immediate 7989 /// operand of a vector shift right operation. For a shift opcode, the value 7990 /// is positive, but for an intrinsic the value count must be negative. The 7991 /// absolute value must be in the range: 7992 /// 1 <= |Value| <= ElementBits for a right shift; or 7993 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift. 7994 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic, 7995 int64_t &Cnt) { 7996 assert(VT.isVector() && "vector shift count is not a vector type"); 7997 unsigned ElementBits = VT.getVectorElementType().getSizeInBits(); 7998 if (! getVShiftImm(Op, ElementBits, Cnt)) 7999 return false; 8000 if (isIntrinsic) 8001 Cnt = -Cnt; 8002 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits)); 8003 } 8004 8005 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics. 8006 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) { 8007 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); 8008 switch (IntNo) { 8009 default: 8010 // Don't do anything for most intrinsics. 8011 break; 8012 8013 // Vector shifts: check for immediate versions and lower them. 8014 // Note: This is done during DAG combining instead of DAG legalizing because 8015 // the build_vectors for 64-bit vector element shift counts are generally 8016 // not legal, and it is hard to see their values after they get legalized to 8017 // loads from a constant pool. 8018 case Intrinsic::arm_neon_vshifts: 8019 case Intrinsic::arm_neon_vshiftu: 8020 case Intrinsic::arm_neon_vshiftls: 8021 case Intrinsic::arm_neon_vshiftlu: 8022 case Intrinsic::arm_neon_vshiftn: 8023 case Intrinsic::arm_neon_vrshifts: 8024 case Intrinsic::arm_neon_vrshiftu: 8025 case Intrinsic::arm_neon_vrshiftn: 8026 case Intrinsic::arm_neon_vqshifts: 8027 case Intrinsic::arm_neon_vqshiftu: 8028 case Intrinsic::arm_neon_vqshiftsu: 8029 case Intrinsic::arm_neon_vqshiftns: 8030 case Intrinsic::arm_neon_vqshiftnu: 8031 case Intrinsic::arm_neon_vqshiftnsu: 8032 case Intrinsic::arm_neon_vqrshiftns: 8033 case Intrinsic::arm_neon_vqrshiftnu: 8034 case Intrinsic::arm_neon_vqrshiftnsu: { 8035 EVT VT = N->getOperand(1).getValueType(); 8036 int64_t Cnt; 8037 unsigned VShiftOpc = 0; 8038 8039 switch (IntNo) { 8040 case Intrinsic::arm_neon_vshifts: 8041 case Intrinsic::arm_neon_vshiftu: 8042 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) { 8043 VShiftOpc = ARMISD::VSHL; 8044 break; 8045 } 8046 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) { 8047 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ? 8048 ARMISD::VSHRs : ARMISD::VSHRu); 8049 break; 8050 } 8051 return SDValue(); 8052 8053 case Intrinsic::arm_neon_vshiftls: 8054 case Intrinsic::arm_neon_vshiftlu: 8055 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt)) 8056 break; 8057 llvm_unreachable("invalid shift count for vshll intrinsic"); 8058 8059 case Intrinsic::arm_neon_vrshifts: 8060 case Intrinsic::arm_neon_vrshiftu: 8061 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) 8062 break; 8063 return SDValue(); 8064 8065 case Intrinsic::arm_neon_vqshifts: 8066 case Intrinsic::arm_neon_vqshiftu: 8067 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) 8068 break; 8069 return SDValue(); 8070 8071 case Intrinsic::arm_neon_vqshiftsu: 8072 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) 8073 break; 8074 llvm_unreachable("invalid shift count for vqshlu intrinsic"); 8075 8076 case Intrinsic::arm_neon_vshiftn: 8077 case Intrinsic::arm_neon_vrshiftn: 8078 case Intrinsic::arm_neon_vqshiftns: 8079 case Intrinsic::arm_neon_vqshiftnu: 8080 case Intrinsic::arm_neon_vqshiftnsu: 8081 case Intrinsic::arm_neon_vqrshiftns: 8082 case Intrinsic::arm_neon_vqrshiftnu: 8083 case Intrinsic::arm_neon_vqrshiftnsu: 8084 // Narrowing shifts require an immediate right shift. 8085 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt)) 8086 break; 8087 llvm_unreachable("invalid shift count for narrowing vector shift " 8088 "intrinsic"); 8089 8090 default: 8091 llvm_unreachable("unhandled vector shift"); 8092 } 8093 8094 switch (IntNo) { 8095 case Intrinsic::arm_neon_vshifts: 8096 case Intrinsic::arm_neon_vshiftu: 8097 // Opcode already set above. 8098 break; 8099 case Intrinsic::arm_neon_vshiftls: 8100 case Intrinsic::arm_neon_vshiftlu: 8101 if (Cnt == VT.getVectorElementType().getSizeInBits()) 8102 VShiftOpc = ARMISD::VSHLLi; 8103 else 8104 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ? 8105 ARMISD::VSHLLs : ARMISD::VSHLLu); 8106 break; 8107 case Intrinsic::arm_neon_vshiftn: 8108 VShiftOpc = ARMISD::VSHRN; break; 8109 case Intrinsic::arm_neon_vrshifts: 8110 VShiftOpc = ARMISD::VRSHRs; break; 8111 case Intrinsic::arm_neon_vrshiftu: 8112 VShiftOpc = ARMISD::VRSHRu; break; 8113 case Intrinsic::arm_neon_vrshiftn: 8114 VShiftOpc = ARMISD::VRSHRN; break; 8115 case Intrinsic::arm_neon_vqshifts: 8116 VShiftOpc = ARMISD::VQSHLs; break; 8117 case Intrinsic::arm_neon_vqshiftu: 8118 VShiftOpc = ARMISD::VQSHLu; break; 8119 case Intrinsic::arm_neon_vqshiftsu: 8120 VShiftOpc = ARMISD::VQSHLsu; break; 8121 case Intrinsic::arm_neon_vqshiftns: 8122 VShiftOpc = ARMISD::VQSHRNs; break; 8123 case Intrinsic::arm_neon_vqshiftnu: 8124 VShiftOpc = ARMISD::VQSHRNu; break; 8125 case Intrinsic::arm_neon_vqshiftnsu: 8126 VShiftOpc = ARMISD::VQSHRNsu; break; 8127 case Intrinsic::arm_neon_vqrshiftns: 8128 VShiftOpc = ARMISD::VQRSHRNs; break; 8129 case Intrinsic::arm_neon_vqrshiftnu: 8130 VShiftOpc = ARMISD::VQRSHRNu; break; 8131 case Intrinsic::arm_neon_vqrshiftnsu: 8132 VShiftOpc = ARMISD::VQRSHRNsu; break; 8133 } 8134 8135 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0), 8136 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32)); 8137 } 8138 8139 case Intrinsic::arm_neon_vshiftins: { 8140 EVT VT = N->getOperand(1).getValueType(); 8141 int64_t Cnt; 8142 unsigned VShiftOpc = 0; 8143 8144 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt)) 8145 VShiftOpc = ARMISD::VSLI; 8146 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt)) 8147 VShiftOpc = ARMISD::VSRI; 8148 else { 8149 llvm_unreachable("invalid shift count for vsli/vsri intrinsic"); 8150 } 8151 8152 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0), 8153 N->getOperand(1), N->getOperand(2), 8154 DAG.getConstant(Cnt, MVT::i32)); 8155 } 8156 8157 case Intrinsic::arm_neon_vqrshifts: 8158 case Intrinsic::arm_neon_vqrshiftu: 8159 // No immediate versions of these to check for. 8160 break; 8161 } 8162 8163 return SDValue(); 8164 } 8165 8166 /// PerformShiftCombine - Checks for immediate versions of vector shifts and 8167 /// lowers them. As with the vector shift intrinsics, this is done during DAG 8168 /// combining instead of DAG legalizing because the build_vectors for 64-bit 8169 /// vector element shift counts are generally not legal, and it is hard to see 8170 /// their values after they get legalized to loads from a constant pool. 8171 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG, 8172 const ARMSubtarget *ST) { 8173 EVT VT = N->getValueType(0); 8174 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) { 8175 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high 8176 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16. 8177 SDValue N1 = N->getOperand(1); 8178 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) { 8179 SDValue N0 = N->getOperand(0); 8180 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP && 8181 DAG.MaskedValueIsZero(N0.getOperand(0), 8182 APInt::getHighBitsSet(32, 16))) 8183 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1); 8184 } 8185 } 8186 8187 // Nothing to be done for scalar shifts. 8188 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 8189 if (!VT.isVector() || !TLI.isTypeLegal(VT)) 8190 return SDValue(); 8191 8192 assert(ST->hasNEON() && "unexpected vector shift"); 8193 int64_t Cnt; 8194 8195 switch (N->getOpcode()) { 8196 default: llvm_unreachable("unexpected shift opcode"); 8197 8198 case ISD::SHL: 8199 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) 8200 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0), 8201 DAG.getConstant(Cnt, MVT::i32)); 8202 break; 8203 8204 case ISD::SRA: 8205 case ISD::SRL: 8206 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) { 8207 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ? 8208 ARMISD::VSHRs : ARMISD::VSHRu); 8209 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0), 8210 DAG.getConstant(Cnt, MVT::i32)); 8211 } 8212 } 8213 return SDValue(); 8214 } 8215 8216 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND, 8217 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND. 8218 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG, 8219 const ARMSubtarget *ST) { 8220 SDValue N0 = N->getOperand(0); 8221 8222 // Check for sign- and zero-extensions of vector extract operations of 8- 8223 // and 16-bit vector elements. NEON supports these directly. They are 8224 // handled during DAG combining because type legalization will promote them 8225 // to 32-bit types and it is messy to recognize the operations after that. 8226 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) { 8227 SDValue Vec = N0.getOperand(0); 8228 SDValue Lane = N0.getOperand(1); 8229 EVT VT = N->getValueType(0); 8230 EVT EltVT = N0.getValueType(); 8231 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 8232 8233 if (VT == MVT::i32 && 8234 (EltVT == MVT::i8 || EltVT == MVT::i16) && 8235 TLI.isTypeLegal(Vec.getValueType()) && 8236 isa<ConstantSDNode>(Lane)) { 8237 8238 unsigned Opc = 0; 8239 switch (N->getOpcode()) { 8240 default: llvm_unreachable("unexpected opcode"); 8241 case ISD::SIGN_EXTEND: 8242 Opc = ARMISD::VGETLANEs; 8243 break; 8244 case ISD::ZERO_EXTEND: 8245 case ISD::ANY_EXTEND: 8246 Opc = ARMISD::VGETLANEu; 8247 break; 8248 } 8249 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane); 8250 } 8251 } 8252 8253 return SDValue(); 8254 } 8255 8256 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC 8257 /// to match f32 max/min patterns to use NEON vmax/vmin instructions. 8258 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG, 8259 const ARMSubtarget *ST) { 8260 // If the target supports NEON, try to use vmax/vmin instructions for f32 8261 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set, 8262 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is 8263 // a NaN; only do the transformation when it matches that behavior. 8264 8265 // For now only do this when using NEON for FP operations; if using VFP, it 8266 // is not obvious that the benefit outweighs the cost of switching to the 8267 // NEON pipeline. 8268 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() || 8269 N->getValueType(0) != MVT::f32) 8270 return SDValue(); 8271 8272 SDValue CondLHS = N->getOperand(0); 8273 SDValue CondRHS = N->getOperand(1); 8274 SDValue LHS = N->getOperand(2); 8275 SDValue RHS = N->getOperand(3); 8276 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get(); 8277 8278 unsigned Opcode = 0; 8279 bool IsReversed; 8280 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) { 8281 IsReversed = false; // x CC y ? x : y 8282 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) { 8283 IsReversed = true ; // x CC y ? y : x 8284 } else { 8285 return SDValue(); 8286 } 8287 8288 bool IsUnordered; 8289 switch (CC) { 8290 default: break; 8291 case ISD::SETOLT: 8292 case ISD::SETOLE: 8293 case ISD::SETLT: 8294 case ISD::SETLE: 8295 case ISD::SETULT: 8296 case ISD::SETULE: 8297 // If LHS is NaN, an ordered comparison will be false and the result will 8298 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS 8299 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN. 8300 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE); 8301 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS)) 8302 break; 8303 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin 8304 // will return -0, so vmin can only be used for unsafe math or if one of 8305 // the operands is known to be nonzero. 8306 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) && 8307 !DAG.getTarget().Options.UnsafeFPMath && 8308 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) 8309 break; 8310 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN; 8311 break; 8312 8313 case ISD::SETOGT: 8314 case ISD::SETOGE: 8315 case ISD::SETGT: 8316 case ISD::SETGE: 8317 case ISD::SETUGT: 8318 case ISD::SETUGE: 8319 // If LHS is NaN, an ordered comparison will be false and the result will 8320 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS 8321 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN. 8322 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE); 8323 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS)) 8324 break; 8325 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax 8326 // will return +0, so vmax can only be used for unsafe math or if one of 8327 // the operands is known to be nonzero. 8328 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) && 8329 !DAG.getTarget().Options.UnsafeFPMath && 8330 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) 8331 break; 8332 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX; 8333 break; 8334 } 8335 8336 if (!Opcode) 8337 return SDValue(); 8338 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS); 8339 } 8340 8341 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV. 8342 SDValue 8343 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const { 8344 SDValue Cmp = N->getOperand(4); 8345 if (Cmp.getOpcode() != ARMISD::CMPZ) 8346 // Only looking at EQ and NE cases. 8347 return SDValue(); 8348 8349 EVT VT = N->getValueType(0); 8350 DebugLoc dl = N->getDebugLoc(); 8351 SDValue LHS = Cmp.getOperand(0); 8352 SDValue RHS = Cmp.getOperand(1); 8353 SDValue FalseVal = N->getOperand(0); 8354 SDValue TrueVal = N->getOperand(1); 8355 SDValue ARMcc = N->getOperand(2); 8356 ARMCC::CondCodes CC = 8357 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue(); 8358 8359 // Simplify 8360 // mov r1, r0 8361 // cmp r1, x 8362 // mov r0, y 8363 // moveq r0, x 8364 // to 8365 // cmp r0, x 8366 // movne r0, y 8367 // 8368 // mov r1, r0 8369 // cmp r1, x 8370 // mov r0, x 8371 // movne r0, y 8372 // to 8373 // cmp r0, x 8374 // movne r0, y 8375 /// FIXME: Turn this into a target neutral optimization? 8376 SDValue Res; 8377 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) { 8378 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc, 8379 N->getOperand(3), Cmp); 8380 } else if (CC == ARMCC::EQ && TrueVal == RHS) { 8381 SDValue ARMcc; 8382 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl); 8383 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc, 8384 N->getOperand(3), NewCmp); 8385 } 8386 8387 if (Res.getNode()) { 8388 APInt KnownZero, KnownOne; 8389 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne); 8390 // Capture demanded bits information that would be otherwise lost. 8391 if (KnownZero == 0xfffffffe) 8392 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res, 8393 DAG.getValueType(MVT::i1)); 8394 else if (KnownZero == 0xffffff00) 8395 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res, 8396 DAG.getValueType(MVT::i8)); 8397 else if (KnownZero == 0xffff0000) 8398 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res, 8399 DAG.getValueType(MVT::i16)); 8400 } 8401 8402 return Res; 8403 } 8404 8405 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N, 8406 DAGCombinerInfo &DCI) const { 8407 switch (N->getOpcode()) { 8408 default: break; 8409 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget); 8410 case ISD::SUB: return PerformSUBCombine(N, DCI); 8411 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget); 8412 case ISD::OR: return PerformORCombine(N, DCI, Subtarget); 8413 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget); 8414 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget); 8415 case ARMISD::BFI: return PerformBFICombine(N, DCI); 8416 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI); 8417 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG); 8418 case ISD::STORE: return PerformSTORECombine(N, DCI); 8419 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI); 8420 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI); 8421 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG); 8422 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI); 8423 case ISD::FP_TO_SINT: 8424 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget); 8425 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget); 8426 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG); 8427 case ISD::SHL: 8428 case ISD::SRA: 8429 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget); 8430 case ISD::SIGN_EXTEND: 8431 case ISD::ZERO_EXTEND: 8432 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget); 8433 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget); 8434 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG); 8435 case ARMISD::VLD2DUP: 8436 case ARMISD::VLD3DUP: 8437 case ARMISD::VLD4DUP: 8438 return CombineBaseUpdate(N, DCI); 8439 case ISD::INTRINSIC_VOID: 8440 case ISD::INTRINSIC_W_CHAIN: 8441 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) { 8442 case Intrinsic::arm_neon_vld1: 8443 case Intrinsic::arm_neon_vld2: 8444 case Intrinsic::arm_neon_vld3: 8445 case Intrinsic::arm_neon_vld4: 8446 case Intrinsic::arm_neon_vld2lane: 8447 case Intrinsic::arm_neon_vld3lane: 8448 case Intrinsic::arm_neon_vld4lane: 8449 case Intrinsic::arm_neon_vst1: 8450 case Intrinsic::arm_neon_vst2: 8451 case Intrinsic::arm_neon_vst3: 8452 case Intrinsic::arm_neon_vst4: 8453 case Intrinsic::arm_neon_vst2lane: 8454 case Intrinsic::arm_neon_vst3lane: 8455 case Intrinsic::arm_neon_vst4lane: 8456 return CombineBaseUpdate(N, DCI); 8457 default: break; 8458 } 8459 break; 8460 } 8461 return SDValue(); 8462 } 8463 8464 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc, 8465 EVT VT) const { 8466 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE); 8467 } 8468 8469 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const { 8470 if (!Subtarget->allowsUnalignedMem()) 8471 return false; 8472 8473 switch (VT.getSimpleVT().SimpleTy) { 8474 default: 8475 return false; 8476 case MVT::i8: 8477 case MVT::i16: 8478 case MVT::i32: 8479 return true; 8480 // FIXME: VLD1 etc with standard alignment is legal. 8481 } 8482 } 8483 8484 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign, 8485 unsigned AlignCheck) { 8486 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) && 8487 (DstAlign == 0 || DstAlign % AlignCheck == 0)); 8488 } 8489 8490 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size, 8491 unsigned DstAlign, unsigned SrcAlign, 8492 bool IsZeroVal, 8493 bool MemcpyStrSrc, 8494 MachineFunction &MF) const { 8495 const Function *F = MF.getFunction(); 8496 8497 // See if we can use NEON instructions for this... 8498 if (IsZeroVal && 8499 !F->hasFnAttr(Attribute::NoImplicitFloat) && 8500 Subtarget->hasNEON()) { 8501 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) { 8502 return MVT::v4i32; 8503 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) { 8504 return MVT::v2i32; 8505 } 8506 } 8507 8508 // Lowering to i32/i16 if the size permits. 8509 if (Size >= 4) { 8510 return MVT::i32; 8511 } else if (Size >= 2) { 8512 return MVT::i16; 8513 } 8514 8515 // Let the target-independent logic figure it out. 8516 return MVT::Other; 8517 } 8518 8519 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) { 8520 if (V < 0) 8521 return false; 8522 8523 unsigned Scale = 1; 8524 switch (VT.getSimpleVT().SimpleTy) { 8525 default: return false; 8526 case MVT::i1: 8527 case MVT::i8: 8528 // Scale == 1; 8529 break; 8530 case MVT::i16: 8531 // Scale == 2; 8532 Scale = 2; 8533 break; 8534 case MVT::i32: 8535 // Scale == 4; 8536 Scale = 4; 8537 break; 8538 } 8539 8540 if ((V & (Scale - 1)) != 0) 8541 return false; 8542 V /= Scale; 8543 return V == (V & ((1LL << 5) - 1)); 8544 } 8545 8546 static bool isLegalT2AddressImmediate(int64_t V, EVT VT, 8547 const ARMSubtarget *Subtarget) { 8548 bool isNeg = false; 8549 if (V < 0) { 8550 isNeg = true; 8551 V = - V; 8552 } 8553 8554 switch (VT.getSimpleVT().SimpleTy) { 8555 default: return false; 8556 case MVT::i1: 8557 case MVT::i8: 8558 case MVT::i16: 8559 case MVT::i32: 8560 // + imm12 or - imm8 8561 if (isNeg) 8562 return V == (V & ((1LL << 8) - 1)); 8563 return V == (V & ((1LL << 12) - 1)); 8564 case MVT::f32: 8565 case MVT::f64: 8566 // Same as ARM mode. FIXME: NEON? 8567 if (!Subtarget->hasVFP2()) 8568 return false; 8569 if ((V & 3) != 0) 8570 return false; 8571 V >>= 2; 8572 return V == (V & ((1LL << 8) - 1)); 8573 } 8574 } 8575 8576 /// isLegalAddressImmediate - Return true if the integer value can be used 8577 /// as the offset of the target addressing mode for load / store of the 8578 /// given type. 8579 static bool isLegalAddressImmediate(int64_t V, EVT VT, 8580 const ARMSubtarget *Subtarget) { 8581 if (V == 0) 8582 return true; 8583 8584 if (!VT.isSimple()) 8585 return false; 8586 8587 if (Subtarget->isThumb1Only()) 8588 return isLegalT1AddressImmediate(V, VT); 8589 else if (Subtarget->isThumb2()) 8590 return isLegalT2AddressImmediate(V, VT, Subtarget); 8591 8592 // ARM mode. 8593 if (V < 0) 8594 V = - V; 8595 switch (VT.getSimpleVT().SimpleTy) { 8596 default: return false; 8597 case MVT::i1: 8598 case MVT::i8: 8599 case MVT::i32: 8600 // +- imm12 8601 return V == (V & ((1LL << 12) - 1)); 8602 case MVT::i16: 8603 // +- imm8 8604 return V == (V & ((1LL << 8) - 1)); 8605 case MVT::f32: 8606 case MVT::f64: 8607 if (!Subtarget->hasVFP2()) // FIXME: NEON? 8608 return false; 8609 if ((V & 3) != 0) 8610 return false; 8611 V >>= 2; 8612 return V == (V & ((1LL << 8) - 1)); 8613 } 8614 } 8615 8616 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM, 8617 EVT VT) const { 8618 int Scale = AM.Scale; 8619 if (Scale < 0) 8620 return false; 8621 8622 switch (VT.getSimpleVT().SimpleTy) { 8623 default: return false; 8624 case MVT::i1: 8625 case MVT::i8: 8626 case MVT::i16: 8627 case MVT::i32: 8628 if (Scale == 1) 8629 return true; 8630 // r + r << imm 8631 Scale = Scale & ~1; 8632 return Scale == 2 || Scale == 4 || Scale == 8; 8633 case MVT::i64: 8634 // r + r 8635 if (((unsigned)AM.HasBaseReg + Scale) <= 2) 8636 return true; 8637 return false; 8638 case MVT::isVoid: 8639 // Note, we allow "void" uses (basically, uses that aren't loads or 8640 // stores), because arm allows folding a scale into many arithmetic 8641 // operations. This should be made more precise and revisited later. 8642 8643 // Allow r << imm, but the imm has to be a multiple of two. 8644 if (Scale & 1) return false; 8645 return isPowerOf2_32(Scale); 8646 } 8647 } 8648 8649 /// isLegalAddressingMode - Return true if the addressing mode represented 8650 /// by AM is legal for this target, for a load/store of the specified type. 8651 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM, 8652 Type *Ty) const { 8653 EVT VT = getValueType(Ty, true); 8654 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget)) 8655 return false; 8656 8657 // Can never fold addr of global into load/store. 8658 if (AM.BaseGV) 8659 return false; 8660 8661 switch (AM.Scale) { 8662 case 0: // no scale reg, must be "r+i" or "r", or "i". 8663 break; 8664 case 1: 8665 if (Subtarget->isThumb1Only()) 8666 return false; 8667 // FALL THROUGH. 8668 default: 8669 // ARM doesn't support any R+R*scale+imm addr modes. 8670 if (AM.BaseOffs) 8671 return false; 8672 8673 if (!VT.isSimple()) 8674 return false; 8675 8676 if (Subtarget->isThumb2()) 8677 return isLegalT2ScaledAddressingMode(AM, VT); 8678 8679 int Scale = AM.Scale; 8680 switch (VT.getSimpleVT().SimpleTy) { 8681 default: return false; 8682 case MVT::i1: 8683 case MVT::i8: 8684 case MVT::i32: 8685 if (Scale < 0) Scale = -Scale; 8686 if (Scale == 1) 8687 return true; 8688 // r + r << imm 8689 return isPowerOf2_32(Scale & ~1); 8690 case MVT::i16: 8691 case MVT::i64: 8692 // r + r 8693 if (((unsigned)AM.HasBaseReg + Scale) <= 2) 8694 return true; 8695 return false; 8696 8697 case MVT::isVoid: 8698 // Note, we allow "void" uses (basically, uses that aren't loads or 8699 // stores), because arm allows folding a scale into many arithmetic 8700 // operations. This should be made more precise and revisited later. 8701 8702 // Allow r << imm, but the imm has to be a multiple of two. 8703 if (Scale & 1) return false; 8704 return isPowerOf2_32(Scale); 8705 } 8706 } 8707 return true; 8708 } 8709 8710 /// isLegalICmpImmediate - Return true if the specified immediate is legal 8711 /// icmp immediate, that is the target has icmp instructions which can compare 8712 /// a register against the immediate without having to materialize the 8713 /// immediate into a register. 8714 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const { 8715 // Thumb2 and ARM modes can use cmn for negative immediates. 8716 if (!Subtarget->isThumb()) 8717 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1; 8718 if (Subtarget->isThumb2()) 8719 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1; 8720 // Thumb1 doesn't have cmn, and only 8-bit immediates. 8721 return Imm >= 0 && Imm <= 255; 8722 } 8723 8724 /// isLegalAddImmediate - Return true if the specified immediate is legal 8725 /// add immediate, that is the target has add instructions which can add 8726 /// a register with the immediate without having to materialize the 8727 /// immediate into a register. 8728 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const { 8729 return ARM_AM::getSOImmVal(Imm) != -1; 8730 } 8731 8732 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT, 8733 bool isSEXTLoad, SDValue &Base, 8734 SDValue &Offset, bool &isInc, 8735 SelectionDAG &DAG) { 8736 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB) 8737 return false; 8738 8739 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) { 8740 // AddressingMode 3 8741 Base = Ptr->getOperand(0); 8742 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 8743 int RHSC = (int)RHS->getZExtValue(); 8744 if (RHSC < 0 && RHSC > -256) { 8745 assert(Ptr->getOpcode() == ISD::ADD); 8746 isInc = false; 8747 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 8748 return true; 8749 } 8750 } 8751 isInc = (Ptr->getOpcode() == ISD::ADD); 8752 Offset = Ptr->getOperand(1); 8753 return true; 8754 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) { 8755 // AddressingMode 2 8756 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 8757 int RHSC = (int)RHS->getZExtValue(); 8758 if (RHSC < 0 && RHSC > -0x1000) { 8759 assert(Ptr->getOpcode() == ISD::ADD); 8760 isInc = false; 8761 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 8762 Base = Ptr->getOperand(0); 8763 return true; 8764 } 8765 } 8766 8767 if (Ptr->getOpcode() == ISD::ADD) { 8768 isInc = true; 8769 ARM_AM::ShiftOpc ShOpcVal= 8770 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode()); 8771 if (ShOpcVal != ARM_AM::no_shift) { 8772 Base = Ptr->getOperand(1); 8773 Offset = Ptr->getOperand(0); 8774 } else { 8775 Base = Ptr->getOperand(0); 8776 Offset = Ptr->getOperand(1); 8777 } 8778 return true; 8779 } 8780 8781 isInc = (Ptr->getOpcode() == ISD::ADD); 8782 Base = Ptr->getOperand(0); 8783 Offset = Ptr->getOperand(1); 8784 return true; 8785 } 8786 8787 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store. 8788 return false; 8789 } 8790 8791 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT, 8792 bool isSEXTLoad, SDValue &Base, 8793 SDValue &Offset, bool &isInc, 8794 SelectionDAG &DAG) { 8795 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB) 8796 return false; 8797 8798 Base = Ptr->getOperand(0); 8799 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 8800 int RHSC = (int)RHS->getZExtValue(); 8801 if (RHSC < 0 && RHSC > -0x100) { // 8 bits. 8802 assert(Ptr->getOpcode() == ISD::ADD); 8803 isInc = false; 8804 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 8805 return true; 8806 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero. 8807 isInc = Ptr->getOpcode() == ISD::ADD; 8808 Offset = DAG.getConstant(RHSC, RHS->getValueType(0)); 8809 return true; 8810 } 8811 } 8812 8813 return false; 8814 } 8815 8816 /// getPreIndexedAddressParts - returns true by value, base pointer and 8817 /// offset pointer and addressing mode by reference if the node's address 8818 /// can be legally represented as pre-indexed load / store address. 8819 bool 8820 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base, 8821 SDValue &Offset, 8822 ISD::MemIndexedMode &AM, 8823 SelectionDAG &DAG) const { 8824 if (Subtarget->isThumb1Only()) 8825 return false; 8826 8827 EVT VT; 8828 SDValue Ptr; 8829 bool isSEXTLoad = false; 8830 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 8831 Ptr = LD->getBasePtr(); 8832 VT = LD->getMemoryVT(); 8833 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD; 8834 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 8835 Ptr = ST->getBasePtr(); 8836 VT = ST->getMemoryVT(); 8837 } else 8838 return false; 8839 8840 bool isInc; 8841 bool isLegal = false; 8842 if (Subtarget->isThumb2()) 8843 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base, 8844 Offset, isInc, DAG); 8845 else 8846 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base, 8847 Offset, isInc, DAG); 8848 if (!isLegal) 8849 return false; 8850 8851 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC; 8852 return true; 8853 } 8854 8855 /// getPostIndexedAddressParts - returns true by value, base pointer and 8856 /// offset pointer and addressing mode by reference if this node can be 8857 /// combined with a load / store to form a post-indexed load / store. 8858 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op, 8859 SDValue &Base, 8860 SDValue &Offset, 8861 ISD::MemIndexedMode &AM, 8862 SelectionDAG &DAG) const { 8863 if (Subtarget->isThumb1Only()) 8864 return false; 8865 8866 EVT VT; 8867 SDValue Ptr; 8868 bool isSEXTLoad = false; 8869 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 8870 VT = LD->getMemoryVT(); 8871 Ptr = LD->getBasePtr(); 8872 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD; 8873 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 8874 VT = ST->getMemoryVT(); 8875 Ptr = ST->getBasePtr(); 8876 } else 8877 return false; 8878 8879 bool isInc; 8880 bool isLegal = false; 8881 if (Subtarget->isThumb2()) 8882 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset, 8883 isInc, DAG); 8884 else 8885 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset, 8886 isInc, DAG); 8887 if (!isLegal) 8888 return false; 8889 8890 if (Ptr != Base) { 8891 // Swap base ptr and offset to catch more post-index load / store when 8892 // it's legal. In Thumb2 mode, offset must be an immediate. 8893 if (Ptr == Offset && Op->getOpcode() == ISD::ADD && 8894 !Subtarget->isThumb2()) 8895 std::swap(Base, Offset); 8896 8897 // Post-indexed load / store update the base pointer. 8898 if (Ptr != Base) 8899 return false; 8900 } 8901 8902 AM = isInc ? ISD::POST_INC : ISD::POST_DEC; 8903 return true; 8904 } 8905 8906 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 8907 APInt &KnownZero, 8908 APInt &KnownOne, 8909 const SelectionDAG &DAG, 8910 unsigned Depth) const { 8911 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); 8912 switch (Op.getOpcode()) { 8913 default: break; 8914 case ARMISD::CMOV: { 8915 // Bits are known zero/one if known on the LHS and RHS. 8916 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1); 8917 if (KnownZero == 0 && KnownOne == 0) return; 8918 8919 APInt KnownZeroRHS, KnownOneRHS; 8920 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1); 8921 KnownZero &= KnownZeroRHS; 8922 KnownOne &= KnownOneRHS; 8923 return; 8924 } 8925 } 8926 } 8927 8928 //===----------------------------------------------------------------------===// 8929 // ARM Inline Assembly Support 8930 //===----------------------------------------------------------------------===// 8931 8932 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const { 8933 // Looking for "rev" which is V6+. 8934 if (!Subtarget->hasV6Ops()) 8935 return false; 8936 8937 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue()); 8938 std::string AsmStr = IA->getAsmString(); 8939 SmallVector<StringRef, 4> AsmPieces; 8940 SplitString(AsmStr, AsmPieces, ";\n"); 8941 8942 switch (AsmPieces.size()) { 8943 default: return false; 8944 case 1: 8945 AsmStr = AsmPieces[0]; 8946 AsmPieces.clear(); 8947 SplitString(AsmStr, AsmPieces, " \t,"); 8948 8949 // rev $0, $1 8950 if (AsmPieces.size() == 3 && 8951 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" && 8952 IA->getConstraintString().compare(0, 4, "=l,l") == 0) { 8953 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType()); 8954 if (Ty && Ty->getBitWidth() == 32) 8955 return IntrinsicLowering::LowerToByteSwap(CI); 8956 } 8957 break; 8958 } 8959 8960 return false; 8961 } 8962 8963 /// getConstraintType - Given a constraint letter, return the type of 8964 /// constraint it is for this target. 8965 ARMTargetLowering::ConstraintType 8966 ARMTargetLowering::getConstraintType(const std::string &Constraint) const { 8967 if (Constraint.size() == 1) { 8968 switch (Constraint[0]) { 8969 default: break; 8970 case 'l': return C_RegisterClass; 8971 case 'w': return C_RegisterClass; 8972 case 'h': return C_RegisterClass; 8973 case 'x': return C_RegisterClass; 8974 case 't': return C_RegisterClass; 8975 case 'j': return C_Other; // Constant for movw. 8976 // An address with a single base register. Due to the way we 8977 // currently handle addresses it is the same as an 'r' memory constraint. 8978 case 'Q': return C_Memory; 8979 } 8980 } else if (Constraint.size() == 2) { 8981 switch (Constraint[0]) { 8982 default: break; 8983 // All 'U+' constraints are addresses. 8984 case 'U': return C_Memory; 8985 } 8986 } 8987 return TargetLowering::getConstraintType(Constraint); 8988 } 8989 8990 /// Examine constraint type and operand type and determine a weight value. 8991 /// This object must already have been set up with the operand type 8992 /// and the current alternative constraint selected. 8993 TargetLowering::ConstraintWeight 8994 ARMTargetLowering::getSingleConstraintMatchWeight( 8995 AsmOperandInfo &info, const char *constraint) const { 8996 ConstraintWeight weight = CW_Invalid; 8997 Value *CallOperandVal = info.CallOperandVal; 8998 // If we don't have a value, we can't do a match, 8999 // but allow it at the lowest weight. 9000 if (CallOperandVal == NULL) 9001 return CW_Default; 9002 Type *type = CallOperandVal->getType(); 9003 // Look at the constraint type. 9004 switch (*constraint) { 9005 default: 9006 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); 9007 break; 9008 case 'l': 9009 if (type->isIntegerTy()) { 9010 if (Subtarget->isThumb()) 9011 weight = CW_SpecificReg; 9012 else 9013 weight = CW_Register; 9014 } 9015 break; 9016 case 'w': 9017 if (type->isFloatingPointTy()) 9018 weight = CW_Register; 9019 break; 9020 } 9021 return weight; 9022 } 9023 9024 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair; 9025 RCPair 9026 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint, 9027 EVT VT) const { 9028 if (Constraint.size() == 1) { 9029 // GCC ARM Constraint Letters 9030 switch (Constraint[0]) { 9031 case 'l': // Low regs or general regs. 9032 if (Subtarget->isThumb()) 9033 return RCPair(0U, ARM::tGPRRegisterClass); 9034 else 9035 return RCPair(0U, ARM::GPRRegisterClass); 9036 case 'h': // High regs or no regs. 9037 if (Subtarget->isThumb()) 9038 return RCPair(0U, ARM::hGPRRegisterClass); 9039 break; 9040 case 'r': 9041 return RCPair(0U, ARM::GPRRegisterClass); 9042 case 'w': 9043 if (VT == MVT::f32) 9044 return RCPair(0U, ARM::SPRRegisterClass); 9045 if (VT.getSizeInBits() == 64) 9046 return RCPair(0U, ARM::DPRRegisterClass); 9047 if (VT.getSizeInBits() == 128) 9048 return RCPair(0U, ARM::QPRRegisterClass); 9049 break; 9050 case 'x': 9051 if (VT == MVT::f32) 9052 return RCPair(0U, ARM::SPR_8RegisterClass); 9053 if (VT.getSizeInBits() == 64) 9054 return RCPair(0U, ARM::DPR_8RegisterClass); 9055 if (VT.getSizeInBits() == 128) 9056 return RCPair(0U, ARM::QPR_8RegisterClass); 9057 break; 9058 case 't': 9059 if (VT == MVT::f32) 9060 return RCPair(0U, ARM::SPRRegisterClass); 9061 break; 9062 } 9063 } 9064 if (StringRef("{cc}").equals_lower(Constraint)) 9065 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass); 9066 9067 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 9068 } 9069 9070 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 9071 /// vector. If it is invalid, don't add anything to Ops. 9072 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op, 9073 std::string &Constraint, 9074 std::vector<SDValue>&Ops, 9075 SelectionDAG &DAG) const { 9076 SDValue Result(0, 0); 9077 9078 // Currently only support length 1 constraints. 9079 if (Constraint.length() != 1) return; 9080 9081 char ConstraintLetter = Constraint[0]; 9082 switch (ConstraintLetter) { 9083 default: break; 9084 case 'j': 9085 case 'I': case 'J': case 'K': case 'L': 9086 case 'M': case 'N': case 'O': 9087 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 9088 if (!C) 9089 return; 9090 9091 int64_t CVal64 = C->getSExtValue(); 9092 int CVal = (int) CVal64; 9093 // None of these constraints allow values larger than 32 bits. Check 9094 // that the value fits in an int. 9095 if (CVal != CVal64) 9096 return; 9097 9098 switch (ConstraintLetter) { 9099 case 'j': 9100 // Constant suitable for movw, must be between 0 and 9101 // 65535. 9102 if (Subtarget->hasV6T2Ops()) 9103 if (CVal >= 0 && CVal <= 65535) 9104 break; 9105 return; 9106 case 'I': 9107 if (Subtarget->isThumb1Only()) { 9108 // This must be a constant between 0 and 255, for ADD 9109 // immediates. 9110 if (CVal >= 0 && CVal <= 255) 9111 break; 9112 } else if (Subtarget->isThumb2()) { 9113 // A constant that can be used as an immediate value in a 9114 // data-processing instruction. 9115 if (ARM_AM::getT2SOImmVal(CVal) != -1) 9116 break; 9117 } else { 9118 // A constant that can be used as an immediate value in a 9119 // data-processing instruction. 9120 if (ARM_AM::getSOImmVal(CVal) != -1) 9121 break; 9122 } 9123 return; 9124 9125 case 'J': 9126 if (Subtarget->isThumb()) { // FIXME thumb2 9127 // This must be a constant between -255 and -1, for negated ADD 9128 // immediates. This can be used in GCC with an "n" modifier that 9129 // prints the negated value, for use with SUB instructions. It is 9130 // not useful otherwise but is implemented for compatibility. 9131 if (CVal >= -255 && CVal <= -1) 9132 break; 9133 } else { 9134 // This must be a constant between -4095 and 4095. It is not clear 9135 // what this constraint is intended for. Implemented for 9136 // compatibility with GCC. 9137 if (CVal >= -4095 && CVal <= 4095) 9138 break; 9139 } 9140 return; 9141 9142 case 'K': 9143 if (Subtarget->isThumb1Only()) { 9144 // A 32-bit value where only one byte has a nonzero value. Exclude 9145 // zero to match GCC. This constraint is used by GCC internally for 9146 // constants that can be loaded with a move/shift combination. 9147 // It is not useful otherwise but is implemented for compatibility. 9148 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal)) 9149 break; 9150 } else if (Subtarget->isThumb2()) { 9151 // A constant whose bitwise inverse can be used as an immediate 9152 // value in a data-processing instruction. This can be used in GCC 9153 // with a "B" modifier that prints the inverted value, for use with 9154 // BIC and MVN instructions. It is not useful otherwise but is 9155 // implemented for compatibility. 9156 if (ARM_AM::getT2SOImmVal(~CVal) != -1) 9157 break; 9158 } else { 9159 // A constant whose bitwise inverse can be used as an immediate 9160 // value in a data-processing instruction. This can be used in GCC 9161 // with a "B" modifier that prints the inverted value, for use with 9162 // BIC and MVN instructions. It is not useful otherwise but is 9163 // implemented for compatibility. 9164 if (ARM_AM::getSOImmVal(~CVal) != -1) 9165 break; 9166 } 9167 return; 9168 9169 case 'L': 9170 if (Subtarget->isThumb1Only()) { 9171 // This must be a constant between -7 and 7, 9172 // for 3-operand ADD/SUB immediate instructions. 9173 if (CVal >= -7 && CVal < 7) 9174 break; 9175 } else if (Subtarget->isThumb2()) { 9176 // A constant whose negation can be used as an immediate value in a 9177 // data-processing instruction. This can be used in GCC with an "n" 9178 // modifier that prints the negated value, for use with SUB 9179 // instructions. It is not useful otherwise but is implemented for 9180 // compatibility. 9181 if (ARM_AM::getT2SOImmVal(-CVal) != -1) 9182 break; 9183 } else { 9184 // A constant whose negation can be used as an immediate value in a 9185 // data-processing instruction. This can be used in GCC with an "n" 9186 // modifier that prints the negated value, for use with SUB 9187 // instructions. It is not useful otherwise but is implemented for 9188 // compatibility. 9189 if (ARM_AM::getSOImmVal(-CVal) != -1) 9190 break; 9191 } 9192 return; 9193 9194 case 'M': 9195 if (Subtarget->isThumb()) { // FIXME thumb2 9196 // This must be a multiple of 4 between 0 and 1020, for 9197 // ADD sp + immediate. 9198 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0)) 9199 break; 9200 } else { 9201 // A power of two or a constant between 0 and 32. This is used in 9202 // GCC for the shift amount on shifted register operands, but it is 9203 // useful in general for any shift amounts. 9204 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0)) 9205 break; 9206 } 9207 return; 9208 9209 case 'N': 9210 if (Subtarget->isThumb()) { // FIXME thumb2 9211 // This must be a constant between 0 and 31, for shift amounts. 9212 if (CVal >= 0 && CVal <= 31) 9213 break; 9214 } 9215 return; 9216 9217 case 'O': 9218 if (Subtarget->isThumb()) { // FIXME thumb2 9219 // This must be a multiple of 4 between -508 and 508, for 9220 // ADD/SUB sp = sp + immediate. 9221 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0)) 9222 break; 9223 } 9224 return; 9225 } 9226 Result = DAG.getTargetConstant(CVal, Op.getValueType()); 9227 break; 9228 } 9229 9230 if (Result.getNode()) { 9231 Ops.push_back(Result); 9232 return; 9233 } 9234 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); 9235 } 9236 9237 bool 9238 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 9239 // The ARM target isn't yet aware of offsets. 9240 return false; 9241 } 9242 9243 bool ARM::isBitFieldInvertedMask(unsigned v) { 9244 if (v == 0xffffffff) 9245 return 0; 9246 // there can be 1's on either or both "outsides", all the "inside" 9247 // bits must be 0's 9248 unsigned int lsb = 0, msb = 31; 9249 while (v & (1 << msb)) --msb; 9250 while (v & (1 << lsb)) ++lsb; 9251 for (unsigned int i = lsb; i <= msb; ++i) { 9252 if (v & (1 << i)) 9253 return 0; 9254 } 9255 return 1; 9256 } 9257 9258 /// isFPImmLegal - Returns true if the target can instruction select the 9259 /// specified FP immediate natively. If false, the legalizer will 9260 /// materialize the FP immediate as a load from a constant pool. 9261 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const { 9262 if (!Subtarget->hasVFP3()) 9263 return false; 9264 if (VT == MVT::f32) 9265 return ARM_AM::getFP32Imm(Imm) != -1; 9266 if (VT == MVT::f64) 9267 return ARM_AM::getFP64Imm(Imm) != -1; 9268 return false; 9269 } 9270 9271 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as 9272 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment 9273 /// specified in the intrinsic calls. 9274 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, 9275 const CallInst &I, 9276 unsigned Intrinsic) const { 9277 switch (Intrinsic) { 9278 case Intrinsic::arm_neon_vld1: 9279 case Intrinsic::arm_neon_vld2: 9280 case Intrinsic::arm_neon_vld3: 9281 case Intrinsic::arm_neon_vld4: 9282 case Intrinsic::arm_neon_vld2lane: 9283 case Intrinsic::arm_neon_vld3lane: 9284 case Intrinsic::arm_neon_vld4lane: { 9285 Info.opc = ISD::INTRINSIC_W_CHAIN; 9286 // Conservatively set memVT to the entire set of vectors loaded. 9287 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8; 9288 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts); 9289 Info.ptrVal = I.getArgOperand(0); 9290 Info.offset = 0; 9291 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1); 9292 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue(); 9293 Info.vol = false; // volatile loads with NEON intrinsics not supported 9294 Info.readMem = true; 9295 Info.writeMem = false; 9296 return true; 9297 } 9298 case Intrinsic::arm_neon_vst1: 9299 case Intrinsic::arm_neon_vst2: 9300 case Intrinsic::arm_neon_vst3: 9301 case Intrinsic::arm_neon_vst4: 9302 case Intrinsic::arm_neon_vst2lane: 9303 case Intrinsic::arm_neon_vst3lane: 9304 case Intrinsic::arm_neon_vst4lane: { 9305 Info.opc = ISD::INTRINSIC_VOID; 9306 // Conservatively set memVT to the entire set of vectors stored. 9307 unsigned NumElts = 0; 9308 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) { 9309 Type *ArgTy = I.getArgOperand(ArgI)->getType(); 9310 if (!ArgTy->isVectorTy()) 9311 break; 9312 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8; 9313 } 9314 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts); 9315 Info.ptrVal = I.getArgOperand(0); 9316 Info.offset = 0; 9317 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1); 9318 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue(); 9319 Info.vol = false; // volatile stores with NEON intrinsics not supported 9320 Info.readMem = false; 9321 Info.writeMem = true; 9322 return true; 9323 } 9324 case Intrinsic::arm_strexd: { 9325 Info.opc = ISD::INTRINSIC_W_CHAIN; 9326 Info.memVT = MVT::i64; 9327 Info.ptrVal = I.getArgOperand(2); 9328 Info.offset = 0; 9329 Info.align = 8; 9330 Info.vol = true; 9331 Info.readMem = false; 9332 Info.writeMem = true; 9333 return true; 9334 } 9335 case Intrinsic::arm_ldrexd: { 9336 Info.opc = ISD::INTRINSIC_W_CHAIN; 9337 Info.memVT = MVT::i64; 9338 Info.ptrVal = I.getArgOperand(0); 9339 Info.offset = 0; 9340 Info.align = 8; 9341 Info.vol = true; 9342 Info.readMem = true; 9343 Info.writeMem = false; 9344 return true; 9345 } 9346 default: 9347 break; 9348 } 9349 9350 return false; 9351 } 9352