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/ADT/Statistic.h" 27 #include "llvm/ADT/StringExtras.h" 28 #include "llvm/CodeGen/CallingConvLower.h" 29 #include "llvm/CodeGen/IntrinsicLowering.h" 30 #include "llvm/CodeGen/MachineBasicBlock.h" 31 #include "llvm/CodeGen/MachineFrameInfo.h" 32 #include "llvm/CodeGen/MachineFunction.h" 33 #include "llvm/CodeGen/MachineInstrBuilder.h" 34 #include "llvm/CodeGen/MachineModuleInfo.h" 35 #include "llvm/CodeGen/MachineRegisterInfo.h" 36 #include "llvm/CodeGen/SelectionDAG.h" 37 #include "llvm/IR/CallingConv.h" 38 #include "llvm/IR/Constants.h" 39 #include "llvm/IR/Function.h" 40 #include "llvm/IR/GlobalValue.h" 41 #include "llvm/IR/Instruction.h" 42 #include "llvm/IR/Instructions.h" 43 #include "llvm/IR/Intrinsics.h" 44 #include "llvm/IR/Type.h" 45 #include "llvm/MC/MCSectionMachO.h" 46 #include "llvm/Support/CommandLine.h" 47 #include "llvm/Support/ErrorHandling.h" 48 #include "llvm/Support/MathExtras.h" 49 #include "llvm/Support/raw_ostream.h" 50 #include "llvm/Target/TargetOptions.h" 51 using namespace llvm; 52 53 STATISTIC(NumTailCalls, "Number of tail calls"); 54 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt"); 55 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments"); 56 57 // This option should go away when tail calls fully work. 58 static cl::opt<bool> 59 EnableARMTailCalls("arm-tail-calls", cl::Hidden, 60 cl::desc("Generate tail calls (TEMPORARY OPTION)."), 61 cl::init(false)); 62 63 cl::opt<bool> 64 EnableARMLongCalls("arm-long-calls", cl::Hidden, 65 cl::desc("Generate calls via indirect call instructions"), 66 cl::init(false)); 67 68 static cl::opt<bool> 69 ARMInterworking("arm-interworking", cl::Hidden, 70 cl::desc("Enable / disable ARM interworking (for debugging only)"), 71 cl::init(true)); 72 73 namespace { 74 class ARMCCState : public CCState { 75 public: 76 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF, 77 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs, 78 LLVMContext &C, ParmContext PC) 79 : CCState(CC, isVarArg, MF, TM, locs, C) { 80 assert(((PC == Call) || (PC == Prologue)) && 81 "ARMCCState users must specify whether their context is call" 82 "or prologue generation."); 83 CallOrPrologue = PC; 84 } 85 }; 86 } 87 88 // The APCS parameter registers. 89 static const uint16_t GPRArgRegs[] = { 90 ARM::R0, ARM::R1, ARM::R2, ARM::R3 91 }; 92 93 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT, 94 MVT PromotedBitwiseVT) { 95 if (VT != PromotedLdStVT) { 96 setOperationAction(ISD::LOAD, VT, Promote); 97 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT); 98 99 setOperationAction(ISD::STORE, VT, Promote); 100 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT); 101 } 102 103 MVT ElemTy = VT.getVectorElementType(); 104 if (ElemTy != MVT::i64 && ElemTy != MVT::f64) 105 setOperationAction(ISD::SETCC, VT, Custom); 106 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); 107 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); 108 if (ElemTy == MVT::i32) { 109 setOperationAction(ISD::SINT_TO_FP, VT, Custom); 110 setOperationAction(ISD::UINT_TO_FP, VT, Custom); 111 setOperationAction(ISD::FP_TO_SINT, VT, Custom); 112 setOperationAction(ISD::FP_TO_UINT, VT, Custom); 113 } else { 114 setOperationAction(ISD::SINT_TO_FP, VT, Expand); 115 setOperationAction(ISD::UINT_TO_FP, VT, Expand); 116 setOperationAction(ISD::FP_TO_SINT, VT, Expand); 117 setOperationAction(ISD::FP_TO_UINT, VT, Expand); 118 } 119 setOperationAction(ISD::BUILD_VECTOR, VT, Custom); 120 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); 121 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal); 122 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal); 123 setOperationAction(ISD::SELECT, VT, Expand); 124 setOperationAction(ISD::SELECT_CC, VT, Expand); 125 setOperationAction(ISD::VSELECT, VT, Expand); 126 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand); 127 if (VT.isInteger()) { 128 setOperationAction(ISD::SHL, VT, Custom); 129 setOperationAction(ISD::SRA, VT, Custom); 130 setOperationAction(ISD::SRL, VT, Custom); 131 } 132 133 // Promote all bit-wise operations. 134 if (VT.isInteger() && VT != PromotedBitwiseVT) { 135 setOperationAction(ISD::AND, VT, Promote); 136 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT); 137 setOperationAction(ISD::OR, VT, Promote); 138 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT); 139 setOperationAction(ISD::XOR, VT, Promote); 140 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT); 141 } 142 143 // Neon does not support vector divide/remainder operations. 144 setOperationAction(ISD::SDIV, VT, Expand); 145 setOperationAction(ISD::UDIV, VT, Expand); 146 setOperationAction(ISD::FDIV, VT, Expand); 147 setOperationAction(ISD::SREM, VT, Expand); 148 setOperationAction(ISD::UREM, VT, Expand); 149 setOperationAction(ISD::FREM, VT, Expand); 150 } 151 152 void ARMTargetLowering::addDRTypeForNEON(MVT VT) { 153 addRegisterClass(VT, &ARM::DPRRegClass); 154 addTypeForNEON(VT, MVT::f64, MVT::v2i32); 155 } 156 157 void ARMTargetLowering::addQRTypeForNEON(MVT VT) { 158 addRegisterClass(VT, &ARM::QPRRegClass); 159 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32); 160 } 161 162 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) { 163 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin()) 164 return new TargetLoweringObjectFileMachO(); 165 166 return new ARMElfTargetObjectFile(); 167 } 168 169 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) 170 : TargetLowering(TM, createTLOF(TM)) { 171 Subtarget = &TM.getSubtarget<ARMSubtarget>(); 172 RegInfo = TM.getRegisterInfo(); 173 Itins = TM.getInstrItineraryData(); 174 175 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); 176 177 if (Subtarget->isTargetDarwin()) { 178 // Uses VFP for Thumb libfuncs if available. 179 if (Subtarget->isThumb() && Subtarget->hasVFP2()) { 180 // Single-precision floating-point arithmetic. 181 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp"); 182 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp"); 183 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp"); 184 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp"); 185 186 // Double-precision floating-point arithmetic. 187 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp"); 188 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp"); 189 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp"); 190 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp"); 191 192 // Single-precision comparisons. 193 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp"); 194 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp"); 195 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp"); 196 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp"); 197 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp"); 198 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp"); 199 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp"); 200 setLibcallName(RTLIB::O_F32, "__unordsf2vfp"); 201 202 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE); 203 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE); 204 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE); 205 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE); 206 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE); 207 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE); 208 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE); 209 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ); 210 211 // Double-precision comparisons. 212 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp"); 213 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp"); 214 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp"); 215 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp"); 216 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp"); 217 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp"); 218 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp"); 219 setLibcallName(RTLIB::O_F64, "__unorddf2vfp"); 220 221 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE); 222 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE); 223 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE); 224 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE); 225 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE); 226 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE); 227 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE); 228 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ); 229 230 // Floating-point to integer conversions. 231 // i64 conversions are done via library routines even when generating VFP 232 // instructions, so use the same ones. 233 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp"); 234 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp"); 235 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp"); 236 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp"); 237 238 // Conversions between floating types. 239 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp"); 240 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp"); 241 242 // Integer to floating-point conversions. 243 // i64 conversions are done via library routines even when generating VFP 244 // instructions, so use the same ones. 245 // FIXME: There appears to be some naming inconsistency in ARM libgcc: 246 // e.g., __floatunsidf vs. __floatunssidfvfp. 247 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp"); 248 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp"); 249 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp"); 250 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp"); 251 } 252 } 253 254 // These libcalls are not available in 32-bit. 255 setLibcallName(RTLIB::SHL_I128, 0); 256 setLibcallName(RTLIB::SRL_I128, 0); 257 setLibcallName(RTLIB::SRA_I128, 0); 258 259 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) { 260 // Double-precision floating-point arithmetic helper functions 261 // RTABI chapter 4.1.2, Table 2 262 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd"); 263 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv"); 264 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul"); 265 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub"); 266 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS); 267 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS); 268 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS); 269 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS); 270 271 // Double-precision floating-point comparison helper functions 272 // RTABI chapter 4.1.2, Table 3 273 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq"); 274 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE); 275 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq"); 276 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ); 277 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt"); 278 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE); 279 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple"); 280 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE); 281 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge"); 282 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE); 283 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt"); 284 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE); 285 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun"); 286 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE); 287 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun"); 288 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ); 289 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS); 290 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS); 291 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS); 292 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS); 293 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS); 294 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS); 295 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS); 296 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS); 297 298 // Single-precision floating-point arithmetic helper functions 299 // RTABI chapter 4.1.2, Table 4 300 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd"); 301 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv"); 302 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul"); 303 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub"); 304 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS); 305 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS); 306 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS); 307 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS); 308 309 // Single-precision floating-point comparison helper functions 310 // RTABI chapter 4.1.2, Table 5 311 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq"); 312 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE); 313 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq"); 314 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ); 315 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt"); 316 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE); 317 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple"); 318 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE); 319 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge"); 320 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE); 321 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt"); 322 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE); 323 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun"); 324 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE); 325 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun"); 326 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ); 327 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS); 328 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS); 329 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS); 330 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS); 331 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS); 332 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS); 333 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS); 334 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS); 335 336 // Floating-point to integer conversions. 337 // RTABI chapter 4.1.2, Table 6 338 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz"); 339 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz"); 340 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz"); 341 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz"); 342 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz"); 343 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz"); 344 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz"); 345 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz"); 346 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS); 347 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS); 348 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS); 349 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS); 350 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS); 351 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS); 352 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS); 353 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS); 354 355 // Conversions between floating types. 356 // RTABI chapter 4.1.2, Table 7 357 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f"); 358 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d"); 359 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS); 360 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS); 361 362 // Integer to floating-point conversions. 363 // RTABI chapter 4.1.2, Table 8 364 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d"); 365 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d"); 366 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d"); 367 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d"); 368 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f"); 369 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f"); 370 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f"); 371 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f"); 372 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS); 373 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS); 374 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS); 375 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS); 376 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS); 377 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS); 378 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS); 379 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS); 380 381 // Long long helper functions 382 // RTABI chapter 4.2, Table 9 383 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul"); 384 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl"); 385 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr"); 386 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr"); 387 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS); 388 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS); 389 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS); 390 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS); 391 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS); 392 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS); 393 394 // Integer division functions 395 // RTABI chapter 4.3.1 396 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv"); 397 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv"); 398 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv"); 399 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod"); 400 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv"); 401 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv"); 402 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv"); 403 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod"); 404 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS); 405 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS); 406 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS); 407 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS); 408 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS); 409 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS); 410 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS); 411 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS); 412 413 // Memory operations 414 // RTABI chapter 4.3.4 415 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy"); 416 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove"); 417 setLibcallName(RTLIB::MEMSET, "__aeabi_memset"); 418 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS); 419 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS); 420 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS); 421 } 422 423 // Use divmod compiler-rt calls for iOS 5.0 and later. 424 if (Subtarget->getTargetTriple().getOS() == Triple::IOS && 425 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) { 426 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4"); 427 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4"); 428 } 429 430 if (Subtarget->isThumb1Only()) 431 addRegisterClass(MVT::i32, &ARM::tGPRRegClass); 432 else 433 addRegisterClass(MVT::i32, &ARM::GPRRegClass); 434 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() && 435 !Subtarget->isThumb1Only()) { 436 addRegisterClass(MVT::f32, &ARM::SPRRegClass); 437 if (!Subtarget->isFPOnlySP()) 438 addRegisterClass(MVT::f64, &ARM::DPRRegClass); 439 440 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 441 } 442 443 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE; 444 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) { 445 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE; 446 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT) 447 setTruncStoreAction((MVT::SimpleValueType)VT, 448 (MVT::SimpleValueType)InnerVT, Expand); 449 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand); 450 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand); 451 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand); 452 } 453 454 setOperationAction(ISD::ConstantFP, MVT::f32, Custom); 455 456 if (Subtarget->hasNEON()) { 457 addDRTypeForNEON(MVT::v2f32); 458 addDRTypeForNEON(MVT::v8i8); 459 addDRTypeForNEON(MVT::v4i16); 460 addDRTypeForNEON(MVT::v2i32); 461 addDRTypeForNEON(MVT::v1i64); 462 463 addQRTypeForNEON(MVT::v4f32); 464 addQRTypeForNEON(MVT::v2f64); 465 addQRTypeForNEON(MVT::v16i8); 466 addQRTypeForNEON(MVT::v8i16); 467 addQRTypeForNEON(MVT::v4i32); 468 addQRTypeForNEON(MVT::v2i64); 469 470 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but 471 // neither Neon nor VFP support any arithmetic operations on it. 472 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively 473 // supported for v4f32. 474 setOperationAction(ISD::FADD, MVT::v2f64, Expand); 475 setOperationAction(ISD::FSUB, MVT::v2f64, Expand); 476 setOperationAction(ISD::FMUL, MVT::v2f64, Expand); 477 // FIXME: Code duplication: FDIV and FREM are expanded always, see 478 // ARMTargetLowering::addTypeForNEON method for details. 479 setOperationAction(ISD::FDIV, MVT::v2f64, Expand); 480 setOperationAction(ISD::FREM, MVT::v2f64, Expand); 481 // FIXME: Create unittest. 482 // In another words, find a way when "copysign" appears in DAG with vector 483 // operands. 484 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand); 485 // FIXME: Code duplication: SETCC has custom operation action, see 486 // ARMTargetLowering::addTypeForNEON method for details. 487 setOperationAction(ISD::SETCC, MVT::v2f64, Expand); 488 // FIXME: Create unittest for FNEG and for FABS. 489 setOperationAction(ISD::FNEG, MVT::v2f64, Expand); 490 setOperationAction(ISD::FABS, MVT::v2f64, Expand); 491 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand); 492 setOperationAction(ISD::FSIN, MVT::v2f64, Expand); 493 setOperationAction(ISD::FCOS, MVT::v2f64, Expand); 494 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand); 495 setOperationAction(ISD::FPOW, MVT::v2f64, Expand); 496 setOperationAction(ISD::FLOG, MVT::v2f64, Expand); 497 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand); 498 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand); 499 setOperationAction(ISD::FEXP, MVT::v2f64, Expand); 500 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand); 501 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR. 502 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand); 503 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand); 504 setOperationAction(ISD::FRINT, MVT::v2f64, Expand); 505 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand); 506 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand); 507 setOperationAction(ISD::FMA, MVT::v2f64, Expand); 508 509 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand); 510 setOperationAction(ISD::FSIN, MVT::v4f32, Expand); 511 setOperationAction(ISD::FCOS, MVT::v4f32, Expand); 512 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand); 513 setOperationAction(ISD::FPOW, MVT::v4f32, Expand); 514 setOperationAction(ISD::FLOG, MVT::v4f32, Expand); 515 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand); 516 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand); 517 setOperationAction(ISD::FEXP, MVT::v4f32, Expand); 518 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand); 519 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand); 520 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand); 521 setOperationAction(ISD::FRINT, MVT::v4f32, Expand); 522 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand); 523 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand); 524 525 // Mark v2f32 intrinsics. 526 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand); 527 setOperationAction(ISD::FSIN, MVT::v2f32, Expand); 528 setOperationAction(ISD::FCOS, MVT::v2f32, Expand); 529 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand); 530 setOperationAction(ISD::FPOW, MVT::v2f32, Expand); 531 setOperationAction(ISD::FLOG, MVT::v2f32, Expand); 532 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand); 533 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand); 534 setOperationAction(ISD::FEXP, MVT::v2f32, Expand); 535 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand); 536 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand); 537 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand); 538 setOperationAction(ISD::FRINT, MVT::v2f32, Expand); 539 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand); 540 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand); 541 542 // Neon does not support some operations on v1i64 and v2i64 types. 543 setOperationAction(ISD::MUL, MVT::v1i64, Expand); 544 // Custom handling for some quad-vector types to detect VMULL. 545 setOperationAction(ISD::MUL, MVT::v8i16, Custom); 546 setOperationAction(ISD::MUL, MVT::v4i32, Custom); 547 setOperationAction(ISD::MUL, MVT::v2i64, Custom); 548 // Custom handling for some vector types to avoid expensive expansions 549 setOperationAction(ISD::SDIV, MVT::v4i16, Custom); 550 setOperationAction(ISD::SDIV, MVT::v8i8, Custom); 551 setOperationAction(ISD::UDIV, MVT::v4i16, Custom); 552 setOperationAction(ISD::UDIV, MVT::v8i8, Custom); 553 setOperationAction(ISD::SETCC, MVT::v1i64, Expand); 554 setOperationAction(ISD::SETCC, MVT::v2i64, Expand); 555 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with 556 // a destination type that is wider than the source, and nor does 557 // it have a FP_TO_[SU]INT instruction with a narrower destination than 558 // source. 559 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom); 560 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom); 561 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom); 562 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom); 563 564 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand); 565 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand); 566 567 // NEON does not have single instruction CTPOP for vectors with element 568 // types wider than 8-bits. However, custom lowering can leverage the 569 // v8i8/v16i8 vcnt instruction. 570 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom); 571 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom); 572 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom); 573 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom); 574 575 // NEON only has FMA instructions as of VFP4. 576 if (!Subtarget->hasVFP4()) { 577 setOperationAction(ISD::FMA, MVT::v2f32, Expand); 578 setOperationAction(ISD::FMA, MVT::v4f32, Expand); 579 } 580 581 setTargetDAGCombine(ISD::INTRINSIC_VOID); 582 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN); 583 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN); 584 setTargetDAGCombine(ISD::SHL); 585 setTargetDAGCombine(ISD::SRL); 586 setTargetDAGCombine(ISD::SRA); 587 setTargetDAGCombine(ISD::SIGN_EXTEND); 588 setTargetDAGCombine(ISD::ZERO_EXTEND); 589 setTargetDAGCombine(ISD::ANY_EXTEND); 590 setTargetDAGCombine(ISD::SELECT_CC); 591 setTargetDAGCombine(ISD::BUILD_VECTOR); 592 setTargetDAGCombine(ISD::VECTOR_SHUFFLE); 593 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT); 594 setTargetDAGCombine(ISD::STORE); 595 setTargetDAGCombine(ISD::FP_TO_SINT); 596 setTargetDAGCombine(ISD::FP_TO_UINT); 597 setTargetDAGCombine(ISD::FDIV); 598 599 // It is legal to extload from v4i8 to v4i16 or v4i32. 600 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8, 601 MVT::v4i16, MVT::v2i16, 602 MVT::v2i32}; 603 for (unsigned i = 0; i < 6; ++i) { 604 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal); 605 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal); 606 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal); 607 } 608 } 609 610 // ARM and Thumb2 support UMLAL/SMLAL. 611 if (!Subtarget->isThumb1Only()) 612 setTargetDAGCombine(ISD::ADDC); 613 614 615 computeRegisterProperties(); 616 617 // ARM does not have f32 extending load. 618 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand); 619 620 // ARM does not have i1 sign extending load. 621 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 622 623 // ARM supports all 4 flavors of integer indexed load / store. 624 if (!Subtarget->isThumb1Only()) { 625 for (unsigned im = (unsigned)ISD::PRE_INC; 626 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) { 627 setIndexedLoadAction(im, MVT::i1, Legal); 628 setIndexedLoadAction(im, MVT::i8, Legal); 629 setIndexedLoadAction(im, MVT::i16, Legal); 630 setIndexedLoadAction(im, MVT::i32, Legal); 631 setIndexedStoreAction(im, MVT::i1, Legal); 632 setIndexedStoreAction(im, MVT::i8, Legal); 633 setIndexedStoreAction(im, MVT::i16, Legal); 634 setIndexedStoreAction(im, MVT::i32, Legal); 635 } 636 } 637 638 // i64 operation support. 639 setOperationAction(ISD::MUL, MVT::i64, Expand); 640 setOperationAction(ISD::MULHU, MVT::i32, Expand); 641 if (Subtarget->isThumb1Only()) { 642 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand); 643 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand); 644 } 645 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops() 646 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP())) 647 setOperationAction(ISD::MULHS, MVT::i32, Expand); 648 649 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); 650 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); 651 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); 652 setOperationAction(ISD::SRL, MVT::i64, Custom); 653 setOperationAction(ISD::SRA, MVT::i64, Custom); 654 655 if (!Subtarget->isThumb1Only()) { 656 // FIXME: We should do this for Thumb1 as well. 657 setOperationAction(ISD::ADDC, MVT::i32, Custom); 658 setOperationAction(ISD::ADDE, MVT::i32, Custom); 659 setOperationAction(ISD::SUBC, MVT::i32, Custom); 660 setOperationAction(ISD::SUBE, MVT::i32, Custom); 661 } 662 663 // ARM does not have ROTL. 664 setOperationAction(ISD::ROTL, MVT::i32, Expand); 665 setOperationAction(ISD::CTTZ, MVT::i32, Custom); 666 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 667 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only()) 668 setOperationAction(ISD::CTLZ, MVT::i32, Expand); 669 670 // These just redirect to CTTZ and CTLZ on ARM. 671 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand); 672 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand); 673 674 // Only ARMv6 has BSWAP. 675 if (!Subtarget->hasV6Ops()) 676 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 677 678 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) && 679 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) { 680 // These are expanded into libcalls if the cpu doesn't have HW divider. 681 setOperationAction(ISD::SDIV, MVT::i32, Expand); 682 setOperationAction(ISD::UDIV, MVT::i32, Expand); 683 } 684 setOperationAction(ISD::SREM, MVT::i32, Expand); 685 setOperationAction(ISD::UREM, MVT::i32, Expand); 686 setOperationAction(ISD::SDIVREM, MVT::i32, Expand); 687 setOperationAction(ISD::UDIVREM, MVT::i32, Expand); 688 689 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 690 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 691 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom); 692 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 693 setOperationAction(ISD::BlockAddress, MVT::i32, Custom); 694 695 setOperationAction(ISD::TRAP, MVT::Other, Legal); 696 697 // Use the default implementation. 698 setOperationAction(ISD::VASTART, MVT::Other, Custom); 699 setOperationAction(ISD::VAARG, MVT::Other, Expand); 700 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 701 setOperationAction(ISD::VAEND, MVT::Other, Expand); 702 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 703 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 704 705 if (!Subtarget->isTargetDarwin()) { 706 // Non-Darwin platforms may return values in these registers via the 707 // personality function. 708 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand); 709 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand); 710 setExceptionPointerRegister(ARM::R0); 711 setExceptionSelectorRegister(ARM::R1); 712 } 713 714 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 715 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use 716 // the default expansion. 717 // FIXME: This should be checking for v6k, not just v6. 718 if (Subtarget->hasDataBarrier() || 719 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) { 720 // membarrier needs custom lowering; the rest are legal and handled 721 // normally. 722 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom); 723 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom); 724 // Custom lowering for 64-bit ops 725 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom); 726 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom); 727 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom); 728 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom); 729 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom); 730 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom); 731 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i64, Custom); 732 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i64, Custom); 733 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom); 734 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom); 735 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom); 736 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc. 737 setInsertFencesForAtomic(true); 738 } else { 739 // Set them all for expansion, which will force libcalls. 740 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand); 741 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand); 742 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand); 743 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand); 744 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand); 745 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand); 746 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand); 747 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand); 748 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand); 749 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand); 750 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand); 751 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand); 752 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand); 753 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand); 754 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the 755 // Unordered/Monotonic case. 756 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom); 757 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom); 758 // Since the libcalls include locking, fold in the fences 759 setShouldFoldAtomicFences(true); 760 } 761 762 setOperationAction(ISD::PREFETCH, MVT::Other, Custom); 763 764 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes. 765 if (!Subtarget->hasV6Ops()) { 766 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 767 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); 768 } 769 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 770 771 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() && 772 !Subtarget->isThumb1Only()) { 773 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR 774 // iff target supports vfp2. 775 setOperationAction(ISD::BITCAST, MVT::i64, Custom); 776 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom); 777 } 778 779 // We want to custom lower some of our intrinsics. 780 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); 781 if (Subtarget->isTargetDarwin()) { 782 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom); 783 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom); 784 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume"); 785 } 786 787 setOperationAction(ISD::SETCC, MVT::i32, Expand); 788 setOperationAction(ISD::SETCC, MVT::f32, Expand); 789 setOperationAction(ISD::SETCC, MVT::f64, Expand); 790 setOperationAction(ISD::SELECT, MVT::i32, Custom); 791 setOperationAction(ISD::SELECT, MVT::f32, Custom); 792 setOperationAction(ISD::SELECT, MVT::f64, Custom); 793 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom); 794 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 795 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 796 797 setOperationAction(ISD::BRCOND, MVT::Other, Expand); 798 setOperationAction(ISD::BR_CC, MVT::i32, Custom); 799 setOperationAction(ISD::BR_CC, MVT::f32, Custom); 800 setOperationAction(ISD::BR_CC, MVT::f64, Custom); 801 setOperationAction(ISD::BR_JT, MVT::Other, Custom); 802 803 // We don't support sin/cos/fmod/copysign/pow 804 setOperationAction(ISD::FSIN, MVT::f64, Expand); 805 setOperationAction(ISD::FSIN, MVT::f32, Expand); 806 setOperationAction(ISD::FCOS, MVT::f32, Expand); 807 setOperationAction(ISD::FCOS, MVT::f64, Expand); 808 setOperationAction(ISD::FSINCOS, MVT::f64, Expand); 809 setOperationAction(ISD::FSINCOS, MVT::f32, Expand); 810 setOperationAction(ISD::FREM, MVT::f64, Expand); 811 setOperationAction(ISD::FREM, MVT::f32, Expand); 812 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() && 813 !Subtarget->isThumb1Only()) { 814 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 815 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 816 } 817 setOperationAction(ISD::FPOW, MVT::f64, Expand); 818 setOperationAction(ISD::FPOW, MVT::f32, Expand); 819 820 if (!Subtarget->hasVFP4()) { 821 setOperationAction(ISD::FMA, MVT::f64, Expand); 822 setOperationAction(ISD::FMA, MVT::f32, Expand); 823 } 824 825 // Various VFP goodness 826 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) { 827 // int <-> fp are custom expanded into bit_convert + ARMISD ops. 828 if (Subtarget->hasVFP2()) { 829 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom); 830 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom); 831 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); 832 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 833 } 834 // Special handling for half-precision FP. 835 if (!Subtarget->hasFP16()) { 836 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand); 837 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand); 838 } 839 } 840 841 // We have target-specific dag combine patterns for the following nodes: 842 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine 843 setTargetDAGCombine(ISD::ADD); 844 setTargetDAGCombine(ISD::SUB); 845 setTargetDAGCombine(ISD::MUL); 846 setTargetDAGCombine(ISD::AND); 847 setTargetDAGCombine(ISD::OR); 848 setTargetDAGCombine(ISD::XOR); 849 850 if (Subtarget->hasV6Ops()) 851 setTargetDAGCombine(ISD::SRL); 852 853 setStackPointerRegisterToSaveRestore(ARM::SP); 854 855 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() || 856 !Subtarget->hasVFP2()) 857 setSchedulingPreference(Sched::RegPressure); 858 else 859 setSchedulingPreference(Sched::Hybrid); 860 861 //// temporary - rewrite interface to use type 862 MaxStoresPerMemset = 8; 863 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4; 864 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores 865 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2; 866 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores 867 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2; 868 869 // On ARM arguments smaller than 4 bytes are extended, so all arguments 870 // are at least 4 bytes aligned. 871 setMinStackArgumentAlignment(4); 872 873 BenefitFromCodePlacementOpt = true; 874 875 // Prefer likely predicted branches to selects on out-of-order cores. 876 PredictableSelectIsExpensive = Subtarget->isLikeA9(); 877 878 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2); 879 } 880 881 // FIXME: It might make sense to define the representative register class as the 882 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is 883 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently, 884 // SPR's representative would be DPR_VFP2. This should work well if register 885 // pressure tracking were modified such that a register use would increment the 886 // pressure of the register class's representative and all of it's super 887 // classes' representatives transitively. We have not implemented this because 888 // of the difficulty prior to coalescing of modeling operand register classes 889 // due to the common occurrence of cross class copies and subregister insertions 890 // and extractions. 891 std::pair<const TargetRegisterClass*, uint8_t> 892 ARMTargetLowering::findRepresentativeClass(MVT VT) const{ 893 const TargetRegisterClass *RRC = 0; 894 uint8_t Cost = 1; 895 switch (VT.SimpleTy) { 896 default: 897 return TargetLowering::findRepresentativeClass(VT); 898 // Use DPR as representative register class for all floating point 899 // and vector types. Since there are 32 SPR registers and 32 DPR registers so 900 // the cost is 1 for both f32 and f64. 901 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16: 902 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32: 903 RRC = &ARM::DPRRegClass; 904 // When NEON is used for SP, only half of the register file is available 905 // because operations that define both SP and DP results will be constrained 906 // to the VFP2 class (D0-D15). We currently model this constraint prior to 907 // coalescing by double-counting the SP regs. See the FIXME above. 908 if (Subtarget->useNEONForSinglePrecisionFP()) 909 Cost = 2; 910 break; 911 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64: 912 case MVT::v4f32: case MVT::v2f64: 913 RRC = &ARM::DPRRegClass; 914 Cost = 2; 915 break; 916 case MVT::v4i64: 917 RRC = &ARM::DPRRegClass; 918 Cost = 4; 919 break; 920 case MVT::v8i64: 921 RRC = &ARM::DPRRegClass; 922 Cost = 8; 923 break; 924 } 925 return std::make_pair(RRC, Cost); 926 } 927 928 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const { 929 switch (Opcode) { 930 default: return 0; 931 case ARMISD::Wrapper: return "ARMISD::Wrapper"; 932 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN"; 933 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC"; 934 case ARMISD::WrapperJT: return "ARMISD::WrapperJT"; 935 case ARMISD::CALL: return "ARMISD::CALL"; 936 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED"; 937 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK"; 938 case ARMISD::tCALL: return "ARMISD::tCALL"; 939 case ARMISD::BRCOND: return "ARMISD::BRCOND"; 940 case ARMISD::BR_JT: return "ARMISD::BR_JT"; 941 case ARMISD::BR2_JT: return "ARMISD::BR2_JT"; 942 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG"; 943 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD"; 944 case ARMISD::CMP: return "ARMISD::CMP"; 945 case ARMISD::CMN: return "ARMISD::CMN"; 946 case ARMISD::CMPZ: return "ARMISD::CMPZ"; 947 case ARMISD::CMPFP: return "ARMISD::CMPFP"; 948 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0"; 949 case ARMISD::BCC_i64: return "ARMISD::BCC_i64"; 950 case ARMISD::FMSTAT: return "ARMISD::FMSTAT"; 951 952 case ARMISD::CMOV: return "ARMISD::CMOV"; 953 954 case ARMISD::RBIT: return "ARMISD::RBIT"; 955 956 case ARMISD::FTOSI: return "ARMISD::FTOSI"; 957 case ARMISD::FTOUI: return "ARMISD::FTOUI"; 958 case ARMISD::SITOF: return "ARMISD::SITOF"; 959 case ARMISD::UITOF: return "ARMISD::UITOF"; 960 961 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG"; 962 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG"; 963 case ARMISD::RRX: return "ARMISD::RRX"; 964 965 case ARMISD::ADDC: return "ARMISD::ADDC"; 966 case ARMISD::ADDE: return "ARMISD::ADDE"; 967 case ARMISD::SUBC: return "ARMISD::SUBC"; 968 case ARMISD::SUBE: return "ARMISD::SUBE"; 969 970 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD"; 971 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR"; 972 973 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP"; 974 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP"; 975 976 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN"; 977 978 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER"; 979 980 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC"; 981 982 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER"; 983 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR"; 984 985 case ARMISD::PRELOAD: return "ARMISD::PRELOAD"; 986 987 case ARMISD::VCEQ: return "ARMISD::VCEQ"; 988 case ARMISD::VCEQZ: return "ARMISD::VCEQZ"; 989 case ARMISD::VCGE: return "ARMISD::VCGE"; 990 case ARMISD::VCGEZ: return "ARMISD::VCGEZ"; 991 case ARMISD::VCLEZ: return "ARMISD::VCLEZ"; 992 case ARMISD::VCGEU: return "ARMISD::VCGEU"; 993 case ARMISD::VCGT: return "ARMISD::VCGT"; 994 case ARMISD::VCGTZ: return "ARMISD::VCGTZ"; 995 case ARMISD::VCLTZ: return "ARMISD::VCLTZ"; 996 case ARMISD::VCGTU: return "ARMISD::VCGTU"; 997 case ARMISD::VTST: return "ARMISD::VTST"; 998 999 case ARMISD::VSHL: return "ARMISD::VSHL"; 1000 case ARMISD::VSHRs: return "ARMISD::VSHRs"; 1001 case ARMISD::VSHRu: return "ARMISD::VSHRu"; 1002 case ARMISD::VSHLLs: return "ARMISD::VSHLLs"; 1003 case ARMISD::VSHLLu: return "ARMISD::VSHLLu"; 1004 case ARMISD::VSHLLi: return "ARMISD::VSHLLi"; 1005 case ARMISD::VSHRN: return "ARMISD::VSHRN"; 1006 case ARMISD::VRSHRs: return "ARMISD::VRSHRs"; 1007 case ARMISD::VRSHRu: return "ARMISD::VRSHRu"; 1008 case ARMISD::VRSHRN: return "ARMISD::VRSHRN"; 1009 case ARMISD::VQSHLs: return "ARMISD::VQSHLs"; 1010 case ARMISD::VQSHLu: return "ARMISD::VQSHLu"; 1011 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu"; 1012 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs"; 1013 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu"; 1014 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu"; 1015 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs"; 1016 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu"; 1017 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu"; 1018 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu"; 1019 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs"; 1020 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM"; 1021 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM"; 1022 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM"; 1023 case ARMISD::VDUP: return "ARMISD::VDUP"; 1024 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE"; 1025 case ARMISD::VEXT: return "ARMISD::VEXT"; 1026 case ARMISD::VREV64: return "ARMISD::VREV64"; 1027 case ARMISD::VREV32: return "ARMISD::VREV32"; 1028 case ARMISD::VREV16: return "ARMISD::VREV16"; 1029 case ARMISD::VZIP: return "ARMISD::VZIP"; 1030 case ARMISD::VUZP: return "ARMISD::VUZP"; 1031 case ARMISD::VTRN: return "ARMISD::VTRN"; 1032 case ARMISD::VTBL1: return "ARMISD::VTBL1"; 1033 case ARMISD::VTBL2: return "ARMISD::VTBL2"; 1034 case ARMISD::VMULLs: return "ARMISD::VMULLs"; 1035 case ARMISD::VMULLu: return "ARMISD::VMULLu"; 1036 case ARMISD::UMLAL: return "ARMISD::UMLAL"; 1037 case ARMISD::SMLAL: return "ARMISD::SMLAL"; 1038 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR"; 1039 case ARMISD::FMAX: return "ARMISD::FMAX"; 1040 case ARMISD::FMIN: return "ARMISD::FMIN"; 1041 case ARMISD::BFI: return "ARMISD::BFI"; 1042 case ARMISD::VORRIMM: return "ARMISD::VORRIMM"; 1043 case ARMISD::VBICIMM: return "ARMISD::VBICIMM"; 1044 case ARMISD::VBSL: return "ARMISD::VBSL"; 1045 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP"; 1046 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP"; 1047 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP"; 1048 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD"; 1049 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD"; 1050 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD"; 1051 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD"; 1052 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD"; 1053 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD"; 1054 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD"; 1055 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD"; 1056 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD"; 1057 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD"; 1058 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD"; 1059 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD"; 1060 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD"; 1061 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD"; 1062 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD"; 1063 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD"; 1064 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD"; 1065 } 1066 } 1067 1068 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const { 1069 if (!VT.isVector()) return getPointerTy(); 1070 return VT.changeVectorElementTypeToInteger(); 1071 } 1072 1073 /// getRegClassFor - Return the register class that should be used for the 1074 /// specified value type. 1075 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const { 1076 // Map v4i64 to QQ registers but do not make the type legal. Similarly map 1077 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to 1078 // load / store 4 to 8 consecutive D registers. 1079 if (Subtarget->hasNEON()) { 1080 if (VT == MVT::v4i64) 1081 return &ARM::QQPRRegClass; 1082 if (VT == MVT::v8i64) 1083 return &ARM::QQQQPRRegClass; 1084 } 1085 return TargetLowering::getRegClassFor(VT); 1086 } 1087 1088 // Create a fast isel object. 1089 FastISel * 1090 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo, 1091 const TargetLibraryInfo *libInfo) const { 1092 return ARM::createFastISel(funcInfo, libInfo); 1093 } 1094 1095 /// getMaximalGlobalOffset - Returns the maximal possible offset which can 1096 /// be used for loads / stores from the global. 1097 unsigned ARMTargetLowering::getMaximalGlobalOffset() const { 1098 return (Subtarget->isThumb1Only() ? 127 : 4095); 1099 } 1100 1101 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const { 1102 unsigned NumVals = N->getNumValues(); 1103 if (!NumVals) 1104 return Sched::RegPressure; 1105 1106 for (unsigned i = 0; i != NumVals; ++i) { 1107 EVT VT = N->getValueType(i); 1108 if (VT == MVT::Glue || VT == MVT::Other) 1109 continue; 1110 if (VT.isFloatingPoint() || VT.isVector()) 1111 return Sched::ILP; 1112 } 1113 1114 if (!N->isMachineOpcode()) 1115 return Sched::RegPressure; 1116 1117 // Load are scheduled for latency even if there instruction itinerary 1118 // is not available. 1119 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1120 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode()); 1121 1122 if (MCID.getNumDefs() == 0) 1123 return Sched::RegPressure; 1124 if (!Itins->isEmpty() && 1125 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2) 1126 return Sched::ILP; 1127 1128 return Sched::RegPressure; 1129 } 1130 1131 //===----------------------------------------------------------------------===// 1132 // Lowering Code 1133 //===----------------------------------------------------------------------===// 1134 1135 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC 1136 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) { 1137 switch (CC) { 1138 default: llvm_unreachable("Unknown condition code!"); 1139 case ISD::SETNE: return ARMCC::NE; 1140 case ISD::SETEQ: return ARMCC::EQ; 1141 case ISD::SETGT: return ARMCC::GT; 1142 case ISD::SETGE: return ARMCC::GE; 1143 case ISD::SETLT: return ARMCC::LT; 1144 case ISD::SETLE: return ARMCC::LE; 1145 case ISD::SETUGT: return ARMCC::HI; 1146 case ISD::SETUGE: return ARMCC::HS; 1147 case ISD::SETULT: return ARMCC::LO; 1148 case ISD::SETULE: return ARMCC::LS; 1149 } 1150 } 1151 1152 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. 1153 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode, 1154 ARMCC::CondCodes &CondCode2) { 1155 CondCode2 = ARMCC::AL; 1156 switch (CC) { 1157 default: llvm_unreachable("Unknown FP condition!"); 1158 case ISD::SETEQ: 1159 case ISD::SETOEQ: CondCode = ARMCC::EQ; break; 1160 case ISD::SETGT: 1161 case ISD::SETOGT: CondCode = ARMCC::GT; break; 1162 case ISD::SETGE: 1163 case ISD::SETOGE: CondCode = ARMCC::GE; break; 1164 case ISD::SETOLT: CondCode = ARMCC::MI; break; 1165 case ISD::SETOLE: CondCode = ARMCC::LS; break; 1166 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break; 1167 case ISD::SETO: CondCode = ARMCC::VC; break; 1168 case ISD::SETUO: CondCode = ARMCC::VS; break; 1169 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break; 1170 case ISD::SETUGT: CondCode = ARMCC::HI; break; 1171 case ISD::SETUGE: CondCode = ARMCC::PL; break; 1172 case ISD::SETLT: 1173 case ISD::SETULT: CondCode = ARMCC::LT; break; 1174 case ISD::SETLE: 1175 case ISD::SETULE: CondCode = ARMCC::LE; break; 1176 case ISD::SETNE: 1177 case ISD::SETUNE: CondCode = ARMCC::NE; break; 1178 } 1179 } 1180 1181 //===----------------------------------------------------------------------===// 1182 // Calling Convention Implementation 1183 //===----------------------------------------------------------------------===// 1184 1185 #include "ARMGenCallingConv.inc" 1186 1187 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the 1188 /// given CallingConvention value. 1189 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC, 1190 bool Return, 1191 bool isVarArg) const { 1192 switch (CC) { 1193 default: 1194 llvm_unreachable("Unsupported calling convention"); 1195 case CallingConv::Fast: 1196 if (Subtarget->hasVFP2() && !isVarArg) { 1197 if (!Subtarget->isAAPCS_ABI()) 1198 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS); 1199 // For AAPCS ABI targets, just use VFP variant of the calling convention. 1200 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP); 1201 } 1202 // Fallthrough 1203 case CallingConv::C: { 1204 // Use target triple & subtarget features to do actual dispatch. 1205 if (!Subtarget->isAAPCS_ABI()) 1206 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS); 1207 else if (Subtarget->hasVFP2() && 1208 getTargetMachine().Options.FloatABIType == FloatABI::Hard && 1209 !isVarArg) 1210 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP); 1211 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS); 1212 } 1213 case CallingConv::ARM_AAPCS_VFP: 1214 if (!isVarArg) 1215 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP); 1216 // Fallthrough 1217 case CallingConv::ARM_AAPCS: 1218 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS); 1219 case CallingConv::ARM_APCS: 1220 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS); 1221 case CallingConv::GHC: 1222 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC); 1223 } 1224 } 1225 1226 /// LowerCallResult - Lower the result values of a call into the 1227 /// appropriate copies out of appropriate physical registers. 1228 SDValue 1229 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, 1230 CallingConv::ID CallConv, bool isVarArg, 1231 const SmallVectorImpl<ISD::InputArg> &Ins, 1232 DebugLoc dl, SelectionDAG &DAG, 1233 SmallVectorImpl<SDValue> &InVals) const { 1234 1235 // Assign locations to each value returned by this call. 1236 SmallVector<CCValAssign, 16> RVLocs; 1237 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1238 getTargetMachine(), RVLocs, *DAG.getContext(), Call); 1239 CCInfo.AnalyzeCallResult(Ins, 1240 CCAssignFnForNode(CallConv, /* Return*/ true, 1241 isVarArg)); 1242 1243 // Copy all of the result registers out of their specified physreg. 1244 for (unsigned i = 0; i != RVLocs.size(); ++i) { 1245 CCValAssign VA = RVLocs[i]; 1246 1247 SDValue Val; 1248 if (VA.needsCustom()) { 1249 // Handle f64 or half of a v2f64. 1250 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, 1251 InFlag); 1252 Chain = Lo.getValue(1); 1253 InFlag = Lo.getValue(2); 1254 VA = RVLocs[++i]; // skip ahead to next loc 1255 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, 1256 InFlag); 1257 Chain = Hi.getValue(1); 1258 InFlag = Hi.getValue(2); 1259 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi); 1260 1261 if (VA.getLocVT() == MVT::v2f64) { 1262 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64); 1263 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val, 1264 DAG.getConstant(0, MVT::i32)); 1265 1266 VA = RVLocs[++i]; // skip ahead to next loc 1267 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag); 1268 Chain = Lo.getValue(1); 1269 InFlag = Lo.getValue(2); 1270 VA = RVLocs[++i]; // skip ahead to next loc 1271 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag); 1272 Chain = Hi.getValue(1); 1273 InFlag = Hi.getValue(2); 1274 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi); 1275 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val, 1276 DAG.getConstant(1, MVT::i32)); 1277 } 1278 } else { 1279 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(), 1280 InFlag); 1281 Chain = Val.getValue(1); 1282 InFlag = Val.getValue(2); 1283 } 1284 1285 switch (VA.getLocInfo()) { 1286 default: llvm_unreachable("Unknown loc info!"); 1287 case CCValAssign::Full: break; 1288 case CCValAssign::BCvt: 1289 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val); 1290 break; 1291 } 1292 1293 InVals.push_back(Val); 1294 } 1295 1296 return Chain; 1297 } 1298 1299 /// LowerMemOpCallTo - Store the argument to the stack. 1300 SDValue 1301 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain, 1302 SDValue StackPtr, SDValue Arg, 1303 DebugLoc dl, SelectionDAG &DAG, 1304 const CCValAssign &VA, 1305 ISD::ArgFlagsTy Flags) const { 1306 unsigned LocMemOffset = VA.getLocMemOffset(); 1307 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset); 1308 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff); 1309 return DAG.getStore(Chain, dl, Arg, PtrOff, 1310 MachinePointerInfo::getStack(LocMemOffset), 1311 false, false, 0); 1312 } 1313 1314 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG, 1315 SDValue Chain, SDValue &Arg, 1316 RegsToPassVector &RegsToPass, 1317 CCValAssign &VA, CCValAssign &NextVA, 1318 SDValue &StackPtr, 1319 SmallVector<SDValue, 8> &MemOpChains, 1320 ISD::ArgFlagsTy Flags) const { 1321 1322 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl, 1323 DAG.getVTList(MVT::i32, MVT::i32), Arg); 1324 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd)); 1325 1326 if (NextVA.isRegLoc()) 1327 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1))); 1328 else { 1329 assert(NextVA.isMemLoc()); 1330 if (StackPtr.getNode() == 0) 1331 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 1332 1333 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1), 1334 dl, DAG, NextVA, 1335 Flags)); 1336 } 1337 } 1338 1339 /// LowerCall - Lowering a call into a callseq_start <- 1340 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter 1341 /// nodes. 1342 SDValue 1343 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, 1344 SmallVectorImpl<SDValue> &InVals) const { 1345 SelectionDAG &DAG = CLI.DAG; 1346 DebugLoc &dl = CLI.DL; 1347 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs; 1348 SmallVector<SDValue, 32> &OutVals = CLI.OutVals; 1349 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins; 1350 SDValue Chain = CLI.Chain; 1351 SDValue Callee = CLI.Callee; 1352 bool &isTailCall = CLI.IsTailCall; 1353 CallingConv::ID CallConv = CLI.CallConv; 1354 bool doesNotRet = CLI.DoesNotReturn; 1355 bool isVarArg = CLI.IsVarArg; 1356 1357 MachineFunction &MF = DAG.getMachineFunction(); 1358 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet(); 1359 bool IsSibCall = false; 1360 // Disable tail calls if they're not supported. 1361 if (!EnableARMTailCalls && !Subtarget->supportsTailCall()) 1362 isTailCall = false; 1363 if (isTailCall) { 1364 // Check if it's really possible to do a tail call. 1365 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, 1366 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(), 1367 Outs, OutVals, Ins, DAG); 1368 // We don't support GuaranteedTailCallOpt for ARM, only automatically 1369 // detected sibcalls. 1370 if (isTailCall) { 1371 ++NumTailCalls; 1372 IsSibCall = true; 1373 } 1374 } 1375 1376 // Analyze operands of the call, assigning locations to each operand. 1377 SmallVector<CCValAssign, 16> ArgLocs; 1378 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1379 getTargetMachine(), ArgLocs, *DAG.getContext(), Call); 1380 CCInfo.AnalyzeCallOperands(Outs, 1381 CCAssignFnForNode(CallConv, /* Return*/ false, 1382 isVarArg)); 1383 1384 // Get a count of how many bytes are to be pushed on the stack. 1385 unsigned NumBytes = CCInfo.getNextStackOffset(); 1386 1387 // For tail calls, memory operands are available in our caller's stack. 1388 if (IsSibCall) 1389 NumBytes = 0; 1390 1391 // Adjust the stack pointer for the new arguments... 1392 // These operations are automatically eliminated by the prolog/epilog pass 1393 if (!IsSibCall) 1394 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true)); 1395 1396 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 1397 1398 RegsToPassVector RegsToPass; 1399 SmallVector<SDValue, 8> MemOpChains; 1400 1401 // Walk the register/memloc assignments, inserting copies/loads. In the case 1402 // of tail call optimization, arguments are handled later. 1403 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); 1404 i != e; 1405 ++i, ++realArgIdx) { 1406 CCValAssign &VA = ArgLocs[i]; 1407 SDValue Arg = OutVals[realArgIdx]; 1408 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags; 1409 bool isByVal = Flags.isByVal(); 1410 1411 // Promote the value if needed. 1412 switch (VA.getLocInfo()) { 1413 default: llvm_unreachable("Unknown loc info!"); 1414 case CCValAssign::Full: break; 1415 case CCValAssign::SExt: 1416 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); 1417 break; 1418 case CCValAssign::ZExt: 1419 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); 1420 break; 1421 case CCValAssign::AExt: 1422 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg); 1423 break; 1424 case CCValAssign::BCvt: 1425 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg); 1426 break; 1427 } 1428 1429 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces 1430 if (VA.needsCustom()) { 1431 if (VA.getLocVT() == MVT::v2f64) { 1432 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1433 DAG.getConstant(0, MVT::i32)); 1434 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1435 DAG.getConstant(1, MVT::i32)); 1436 1437 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass, 1438 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags); 1439 1440 VA = ArgLocs[++i]; // skip ahead to next loc 1441 if (VA.isRegLoc()) { 1442 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass, 1443 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags); 1444 } else { 1445 assert(VA.isMemLoc()); 1446 1447 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1, 1448 dl, DAG, VA, Flags)); 1449 } 1450 } else { 1451 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i], 1452 StackPtr, MemOpChains, Flags); 1453 } 1454 } else if (VA.isRegLoc()) { 1455 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 1456 } else if (isByVal) { 1457 assert(VA.isMemLoc()); 1458 unsigned offset = 0; 1459 1460 // True if this byval aggregate will be split between registers 1461 // and memory. 1462 if (CCInfo.isFirstByValRegValid()) { 1463 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 1464 unsigned int i, j; 1465 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) { 1466 SDValue Const = DAG.getConstant(4*i, MVT::i32); 1467 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const); 1468 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg, 1469 MachinePointerInfo(), 1470 false, false, false, 0); 1471 MemOpChains.push_back(Load.getValue(1)); 1472 RegsToPass.push_back(std::make_pair(j, Load)); 1473 } 1474 offset = ARM::R4 - CCInfo.getFirstByValReg(); 1475 CCInfo.clearFirstByValReg(); 1476 } 1477 1478 if (Flags.getByValSize() - 4*offset > 0) { 1479 unsigned LocMemOffset = VA.getLocMemOffset(); 1480 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset); 1481 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, 1482 StkPtrOff); 1483 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset); 1484 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset); 1485 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, 1486 MVT::i32); 1487 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32); 1488 1489 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue); 1490 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode}; 1491 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs, 1492 Ops, array_lengthof(Ops))); 1493 } 1494 } else if (!IsSibCall) { 1495 assert(VA.isMemLoc()); 1496 1497 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg, 1498 dl, DAG, VA, Flags)); 1499 } 1500 } 1501 1502 if (!MemOpChains.empty()) 1503 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 1504 &MemOpChains[0], MemOpChains.size()); 1505 1506 // Build a sequence of copy-to-reg nodes chained together with token chain 1507 // and flag operands which copy the outgoing args into the appropriate regs. 1508 SDValue InFlag; 1509 // Tail call byval lowering might overwrite argument registers so in case of 1510 // tail call optimization the copies to registers are lowered later. 1511 if (!isTailCall) 1512 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 1513 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, 1514 RegsToPass[i].second, InFlag); 1515 InFlag = Chain.getValue(1); 1516 } 1517 1518 // For tail calls lower the arguments to the 'real' stack slot. 1519 if (isTailCall) { 1520 // Force all the incoming stack arguments to be loaded from the stack 1521 // before any new outgoing arguments are stored to the stack, because the 1522 // outgoing stack slots may alias the incoming argument stack slots, and 1523 // the alias isn't otherwise explicit. This is slightly more conservative 1524 // than necessary, because it means that each store effectively depends 1525 // on every argument instead of just those arguments it would clobber. 1526 1527 // Do not flag preceding copytoreg stuff together with the following stuff. 1528 InFlag = SDValue(); 1529 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 1530 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, 1531 RegsToPass[i].second, InFlag); 1532 InFlag = Chain.getValue(1); 1533 } 1534 InFlag =SDValue(); 1535 } 1536 1537 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 1538 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 1539 // node so that legalize doesn't hack it. 1540 bool isDirect = false; 1541 bool isARMFunc = false; 1542 bool isLocalARMFunc = false; 1543 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1544 1545 if (EnableARMLongCalls) { 1546 assert (getTargetMachine().getRelocationModel() == Reloc::Static 1547 && "long-calls with non-static relocation model!"); 1548 // Handle a global address or an external symbol. If it's not one of 1549 // those, the target's already in a register, so we don't need to do 1550 // anything extra. 1551 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 1552 const GlobalValue *GV = G->getGlobal(); 1553 // Create a constant pool entry for the callee address 1554 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1555 ARMConstantPoolValue *CPV = 1556 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0); 1557 1558 // Get the address of the callee into a register 1559 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1560 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1561 Callee = DAG.getLoad(getPointerTy(), dl, 1562 DAG.getEntryNode(), CPAddr, 1563 MachinePointerInfo::getConstantPool(), 1564 false, false, false, 0); 1565 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) { 1566 const char *Sym = S->getSymbol(); 1567 1568 // Create a constant pool entry for the callee address 1569 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1570 ARMConstantPoolValue *CPV = 1571 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym, 1572 ARMPCLabelIndex, 0); 1573 // Get the address of the callee into a register 1574 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1575 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1576 Callee = DAG.getLoad(getPointerTy(), dl, 1577 DAG.getEntryNode(), CPAddr, 1578 MachinePointerInfo::getConstantPool(), 1579 false, false, false, 0); 1580 } 1581 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 1582 const GlobalValue *GV = G->getGlobal(); 1583 isDirect = true; 1584 bool isExt = GV->isDeclaration() || GV->isWeakForLinker(); 1585 bool isStub = (isExt && Subtarget->isTargetDarwin()) && 1586 getTargetMachine().getRelocationModel() != Reloc::Static; 1587 isARMFunc = !Subtarget->isThumb() || isStub; 1588 // ARM call to a local ARM function is predicable. 1589 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking); 1590 // tBX takes a register source operand. 1591 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) { 1592 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1593 ARMConstantPoolValue *CPV = 1594 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4); 1595 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1596 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1597 Callee = DAG.getLoad(getPointerTy(), dl, 1598 DAG.getEntryNode(), CPAddr, 1599 MachinePointerInfo::getConstantPool(), 1600 false, false, false, 0); 1601 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 1602 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, 1603 getPointerTy(), Callee, PICLabel); 1604 } else { 1605 // On ELF targets for PIC code, direct calls should go through the PLT 1606 unsigned OpFlags = 0; 1607 if (Subtarget->isTargetELF() && 1608 getTargetMachine().getRelocationModel() == Reloc::PIC_) 1609 OpFlags = ARMII::MO_PLT; 1610 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags); 1611 } 1612 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 1613 isDirect = true; 1614 bool isStub = Subtarget->isTargetDarwin() && 1615 getTargetMachine().getRelocationModel() != Reloc::Static; 1616 isARMFunc = !Subtarget->isThumb() || isStub; 1617 // tBX takes a register source operand. 1618 const char *Sym = S->getSymbol(); 1619 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) { 1620 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1621 ARMConstantPoolValue *CPV = 1622 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym, 1623 ARMPCLabelIndex, 4); 1624 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1625 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1626 Callee = DAG.getLoad(getPointerTy(), dl, 1627 DAG.getEntryNode(), CPAddr, 1628 MachinePointerInfo::getConstantPool(), 1629 false, false, false, 0); 1630 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 1631 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, 1632 getPointerTy(), Callee, PICLabel); 1633 } else { 1634 unsigned OpFlags = 0; 1635 // On ELF targets for PIC code, direct calls should go through the PLT 1636 if (Subtarget->isTargetELF() && 1637 getTargetMachine().getRelocationModel() == Reloc::PIC_) 1638 OpFlags = ARMII::MO_PLT; 1639 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags); 1640 } 1641 } 1642 1643 // FIXME: handle tail calls differently. 1644 unsigned CallOpc; 1645 bool HasMinSizeAttr = MF.getFunction()->getAttributes(). 1646 hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize); 1647 if (Subtarget->isThumb()) { 1648 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps()) 1649 CallOpc = ARMISD::CALL_NOLINK; 1650 else 1651 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL; 1652 } else { 1653 if (!isDirect && !Subtarget->hasV5TOps()) 1654 CallOpc = ARMISD::CALL_NOLINK; 1655 else if (doesNotRet && isDirect && Subtarget->hasRAS() && 1656 // Emit regular call when code size is the priority 1657 !HasMinSizeAttr) 1658 // "mov lr, pc; b _foo" to avoid confusing the RSP 1659 CallOpc = ARMISD::CALL_NOLINK; 1660 else 1661 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL; 1662 } 1663 1664 std::vector<SDValue> Ops; 1665 Ops.push_back(Chain); 1666 Ops.push_back(Callee); 1667 1668 // Add argument registers to the end of the list so that they are known live 1669 // into the call. 1670 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 1671 Ops.push_back(DAG.getRegister(RegsToPass[i].first, 1672 RegsToPass[i].second.getValueType())); 1673 1674 // Add a register mask operand representing the call-preserved registers. 1675 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo(); 1676 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv); 1677 assert(Mask && "Missing call preserved mask for calling convention"); 1678 Ops.push_back(DAG.getRegisterMask(Mask)); 1679 1680 if (InFlag.getNode()) 1681 Ops.push_back(InFlag); 1682 1683 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 1684 if (isTailCall) 1685 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size()); 1686 1687 // Returns a chain and a flag for retval copy to use. 1688 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size()); 1689 InFlag = Chain.getValue(1); 1690 1691 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true), 1692 DAG.getIntPtrConstant(0, true), InFlag); 1693 if (!Ins.empty()) 1694 InFlag = Chain.getValue(1); 1695 1696 // Handle result values, copying them out of physregs into vregs that we 1697 // return. 1698 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, 1699 dl, DAG, InVals); 1700 } 1701 1702 /// HandleByVal - Every parameter *after* a byval parameter is passed 1703 /// on the stack. Remember the next parameter register to allocate, 1704 /// and then confiscate the rest of the parameter registers to insure 1705 /// this. 1706 void 1707 ARMTargetLowering::HandleByVal( 1708 CCState *State, unsigned &size, unsigned Align) const { 1709 unsigned reg = State->AllocateReg(GPRArgRegs, 4); 1710 assert((State->getCallOrPrologue() == Prologue || 1711 State->getCallOrPrologue() == Call) && 1712 "unhandled ParmContext"); 1713 if ((!State->isFirstByValRegValid()) && 1714 (ARM::R0 <= reg) && (reg <= ARM::R3)) { 1715 if (Subtarget->isAAPCS_ABI() && Align > 4) { 1716 unsigned AlignInRegs = Align / 4; 1717 unsigned Waste = (ARM::R4 - reg) % AlignInRegs; 1718 for (unsigned i = 0; i < Waste; ++i) 1719 reg = State->AllocateReg(GPRArgRegs, 4); 1720 } 1721 if (reg != 0) { 1722 State->setFirstByValReg(reg); 1723 // At a call site, a byval parameter that is split between 1724 // registers and memory needs its size truncated here. In a 1725 // function prologue, such byval parameters are reassembled in 1726 // memory, and are not truncated. 1727 if (State->getCallOrPrologue() == Call) { 1728 unsigned excess = 4 * (ARM::R4 - reg); 1729 assert(size >= excess && "expected larger existing stack allocation"); 1730 size -= excess; 1731 } 1732 } 1733 } 1734 // Confiscate any remaining parameter registers to preclude their 1735 // assignment to subsequent parameters. 1736 while (State->AllocateReg(GPRArgRegs, 4)) 1737 ; 1738 } 1739 1740 /// MatchingStackOffset - Return true if the given stack call argument is 1741 /// already available in the same position (relatively) of the caller's 1742 /// incoming argument stack. 1743 static 1744 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags, 1745 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI, 1746 const TargetInstrInfo *TII) { 1747 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8; 1748 int FI = INT_MAX; 1749 if (Arg.getOpcode() == ISD::CopyFromReg) { 1750 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg(); 1751 if (!TargetRegisterInfo::isVirtualRegister(VR)) 1752 return false; 1753 MachineInstr *Def = MRI->getVRegDef(VR); 1754 if (!Def) 1755 return false; 1756 if (!Flags.isByVal()) { 1757 if (!TII->isLoadFromStackSlot(Def, FI)) 1758 return false; 1759 } else { 1760 return false; 1761 } 1762 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) { 1763 if (Flags.isByVal()) 1764 // ByVal argument is passed in as a pointer but it's now being 1765 // dereferenced. e.g. 1766 // define @foo(%struct.X* %A) { 1767 // tail call @bar(%struct.X* byval %A) 1768 // } 1769 return false; 1770 SDValue Ptr = Ld->getBasePtr(); 1771 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr); 1772 if (!FINode) 1773 return false; 1774 FI = FINode->getIndex(); 1775 } else 1776 return false; 1777 1778 assert(FI != INT_MAX); 1779 if (!MFI->isFixedObjectIndex(FI)) 1780 return false; 1781 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI); 1782 } 1783 1784 /// IsEligibleForTailCallOptimization - Check whether the call is eligible 1785 /// for tail call optimization. Targets which want to do tail call 1786 /// optimization should implement this function. 1787 bool 1788 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee, 1789 CallingConv::ID CalleeCC, 1790 bool isVarArg, 1791 bool isCalleeStructRet, 1792 bool isCallerStructRet, 1793 const SmallVectorImpl<ISD::OutputArg> &Outs, 1794 const SmallVectorImpl<SDValue> &OutVals, 1795 const SmallVectorImpl<ISD::InputArg> &Ins, 1796 SelectionDAG& DAG) const { 1797 const Function *CallerF = DAG.getMachineFunction().getFunction(); 1798 CallingConv::ID CallerCC = CallerF->getCallingConv(); 1799 bool CCMatch = CallerCC == CalleeCC; 1800 1801 // Look for obvious safe cases to perform tail call optimization that do not 1802 // require ABI changes. This is what gcc calls sibcall. 1803 1804 // Do not sibcall optimize vararg calls unless the call site is not passing 1805 // any arguments. 1806 if (isVarArg && !Outs.empty()) 1807 return false; 1808 1809 // Also avoid sibcall optimization if either caller or callee uses struct 1810 // return semantics. 1811 if (isCalleeStructRet || isCallerStructRet) 1812 return false; 1813 1814 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo:: 1815 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as 1816 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation 1817 // support in the assembler and linker to be used. This would need to be 1818 // fixed to fully support tail calls in Thumb1. 1819 // 1820 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take 1821 // LR. This means if we need to reload LR, it takes an extra instructions, 1822 // which outweighs the value of the tail call; but here we don't know yet 1823 // whether LR is going to be used. Probably the right approach is to 1824 // generate the tail call here and turn it back into CALL/RET in 1825 // emitEpilogue if LR is used. 1826 1827 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls, 1828 // but we need to make sure there are enough registers; the only valid 1829 // registers are the 4 used for parameters. We don't currently do this 1830 // case. 1831 if (Subtarget->isThumb1Only()) 1832 return false; 1833 1834 // If the calling conventions do not match, then we'd better make sure the 1835 // results are returned in the same way as what the caller expects. 1836 if (!CCMatch) { 1837 SmallVector<CCValAssign, 16> RVLocs1; 1838 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), 1839 getTargetMachine(), RVLocs1, *DAG.getContext(), Call); 1840 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg)); 1841 1842 SmallVector<CCValAssign, 16> RVLocs2; 1843 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), 1844 getTargetMachine(), RVLocs2, *DAG.getContext(), Call); 1845 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg)); 1846 1847 if (RVLocs1.size() != RVLocs2.size()) 1848 return false; 1849 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) { 1850 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc()) 1851 return false; 1852 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo()) 1853 return false; 1854 if (RVLocs1[i].isRegLoc()) { 1855 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg()) 1856 return false; 1857 } else { 1858 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset()) 1859 return false; 1860 } 1861 } 1862 } 1863 1864 // If Caller's vararg or byval argument has been split between registers and 1865 // stack, do not perform tail call, since part of the argument is in caller's 1866 // local frame. 1867 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction(). 1868 getInfo<ARMFunctionInfo>(); 1869 if (AFI_Caller->getVarArgsRegSaveSize()) 1870 return false; 1871 1872 // If the callee takes no arguments then go on to check the results of the 1873 // call. 1874 if (!Outs.empty()) { 1875 // Check if stack adjustment is needed. For now, do not do this if any 1876 // argument is passed on the stack. 1877 SmallVector<CCValAssign, 16> ArgLocs; 1878 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), 1879 getTargetMachine(), ArgLocs, *DAG.getContext(), Call); 1880 CCInfo.AnalyzeCallOperands(Outs, 1881 CCAssignFnForNode(CalleeCC, false, isVarArg)); 1882 if (CCInfo.getNextStackOffset()) { 1883 MachineFunction &MF = DAG.getMachineFunction(); 1884 1885 // Check if the arguments are already laid out in the right way as 1886 // the caller's fixed stack objects. 1887 MachineFrameInfo *MFI = MF.getFrameInfo(); 1888 const MachineRegisterInfo *MRI = &MF.getRegInfo(); 1889 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1890 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); 1891 i != e; 1892 ++i, ++realArgIdx) { 1893 CCValAssign &VA = ArgLocs[i]; 1894 EVT RegVT = VA.getLocVT(); 1895 SDValue Arg = OutVals[realArgIdx]; 1896 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags; 1897 if (VA.getLocInfo() == CCValAssign::Indirect) 1898 return false; 1899 if (VA.needsCustom()) { 1900 // f64 and vector types are split into multiple registers or 1901 // register/stack-slot combinations. The types will not match 1902 // the registers; give up on memory f64 refs until we figure 1903 // out what to do about this. 1904 if (!VA.isRegLoc()) 1905 return false; 1906 if (!ArgLocs[++i].isRegLoc()) 1907 return false; 1908 if (RegVT == MVT::v2f64) { 1909 if (!ArgLocs[++i].isRegLoc()) 1910 return false; 1911 if (!ArgLocs[++i].isRegLoc()) 1912 return false; 1913 } 1914 } else if (!VA.isRegLoc()) { 1915 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags, 1916 MFI, MRI, TII)) 1917 return false; 1918 } 1919 } 1920 } 1921 } 1922 1923 return true; 1924 } 1925 1926 bool 1927 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv, 1928 MachineFunction &MF, bool isVarArg, 1929 const SmallVectorImpl<ISD::OutputArg> &Outs, 1930 LLVMContext &Context) const { 1931 SmallVector<CCValAssign, 16> RVLocs; 1932 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context); 1933 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true, 1934 isVarArg)); 1935 } 1936 1937 SDValue 1938 ARMTargetLowering::LowerReturn(SDValue Chain, 1939 CallingConv::ID CallConv, bool isVarArg, 1940 const SmallVectorImpl<ISD::OutputArg> &Outs, 1941 const SmallVectorImpl<SDValue> &OutVals, 1942 DebugLoc dl, SelectionDAG &DAG) const { 1943 1944 // CCValAssign - represent the assignment of the return value to a location. 1945 SmallVector<CCValAssign, 16> RVLocs; 1946 1947 // CCState - Info about the registers and stack slots. 1948 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1949 getTargetMachine(), RVLocs, *DAG.getContext(), Call); 1950 1951 // Analyze outgoing return values. 1952 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true, 1953 isVarArg)); 1954 1955 SDValue Flag; 1956 SmallVector<SDValue, 4> RetOps; 1957 RetOps.push_back(Chain); // Operand #0 = Chain (updated below) 1958 1959 // Copy the result values into the output registers. 1960 for (unsigned i = 0, realRVLocIdx = 0; 1961 i != RVLocs.size(); 1962 ++i, ++realRVLocIdx) { 1963 CCValAssign &VA = RVLocs[i]; 1964 assert(VA.isRegLoc() && "Can only return in registers!"); 1965 1966 SDValue Arg = OutVals[realRVLocIdx]; 1967 1968 switch (VA.getLocInfo()) { 1969 default: llvm_unreachable("Unknown loc info!"); 1970 case CCValAssign::Full: break; 1971 case CCValAssign::BCvt: 1972 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg); 1973 break; 1974 } 1975 1976 if (VA.needsCustom()) { 1977 if (VA.getLocVT() == MVT::v2f64) { 1978 // Extract the first half and return it in two registers. 1979 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1980 DAG.getConstant(0, MVT::i32)); 1981 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl, 1982 DAG.getVTList(MVT::i32, MVT::i32), Half); 1983 1984 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag); 1985 Flag = Chain.getValue(1); 1986 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 1987 VA = RVLocs[++i]; // skip ahead to next loc 1988 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), 1989 HalfGPRs.getValue(1), Flag); 1990 Flag = Chain.getValue(1); 1991 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 1992 VA = RVLocs[++i]; // skip ahead to next loc 1993 1994 // Extract the 2nd half and fall through to handle it as an f64 value. 1995 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1996 DAG.getConstant(1, MVT::i32)); 1997 } 1998 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is 1999 // available. 2000 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl, 2001 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1); 2002 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag); 2003 Flag = Chain.getValue(1); 2004 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 2005 VA = RVLocs[++i]; // skip ahead to next loc 2006 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1), 2007 Flag); 2008 } else 2009 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag); 2010 2011 // Guarantee that all emitted copies are 2012 // stuck together, avoiding something bad. 2013 Flag = Chain.getValue(1); 2014 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 2015 } 2016 2017 // Update chain and glue. 2018 RetOps[0] = Chain; 2019 if (Flag.getNode()) 2020 RetOps.push_back(Flag); 2021 2022 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, 2023 RetOps.data(), RetOps.size()); 2024 } 2025 2026 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const { 2027 if (N->getNumValues() != 1) 2028 return false; 2029 if (!N->hasNUsesOfValue(1, 0)) 2030 return false; 2031 2032 SDValue TCChain = Chain; 2033 SDNode *Copy = *N->use_begin(); 2034 if (Copy->getOpcode() == ISD::CopyToReg) { 2035 // If the copy has a glue operand, we conservatively assume it isn't safe to 2036 // perform a tail call. 2037 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue) 2038 return false; 2039 TCChain = Copy->getOperand(0); 2040 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) { 2041 SDNode *VMov = Copy; 2042 // f64 returned in a pair of GPRs. 2043 SmallPtrSet<SDNode*, 2> Copies; 2044 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end(); 2045 UI != UE; ++UI) { 2046 if (UI->getOpcode() != ISD::CopyToReg) 2047 return false; 2048 Copies.insert(*UI); 2049 } 2050 if (Copies.size() > 2) 2051 return false; 2052 2053 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end(); 2054 UI != UE; ++UI) { 2055 SDValue UseChain = UI->getOperand(0); 2056 if (Copies.count(UseChain.getNode())) 2057 // Second CopyToReg 2058 Copy = *UI; 2059 else 2060 // First CopyToReg 2061 TCChain = UseChain; 2062 } 2063 } else if (Copy->getOpcode() == ISD::BITCAST) { 2064 // f32 returned in a single GPR. 2065 if (!Copy->hasOneUse()) 2066 return false; 2067 Copy = *Copy->use_begin(); 2068 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0)) 2069 return false; 2070 Chain = Copy->getOperand(0); 2071 } else { 2072 return false; 2073 } 2074 2075 bool HasRet = false; 2076 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end(); 2077 UI != UE; ++UI) { 2078 if (UI->getOpcode() != ARMISD::RET_FLAG) 2079 return false; 2080 HasRet = true; 2081 } 2082 2083 if (!HasRet) 2084 return false; 2085 2086 Chain = TCChain; 2087 return true; 2088 } 2089 2090 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const { 2091 if (!EnableARMTailCalls && !Subtarget->supportsTailCall()) 2092 return false; 2093 2094 if (!CI->isTailCall()) 2095 return false; 2096 2097 return !Subtarget->isThumb1Only(); 2098 } 2099 2100 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as 2101 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is 2102 // one of the above mentioned nodes. It has to be wrapped because otherwise 2103 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only 2104 // be used to form addressing mode. These wrapped nodes will be selected 2105 // into MOVi. 2106 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) { 2107 EVT PtrVT = Op.getValueType(); 2108 // FIXME there is no actual debug info here 2109 DebugLoc dl = Op.getDebugLoc(); 2110 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); 2111 SDValue Res; 2112 if (CP->isMachineConstantPoolEntry()) 2113 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT, 2114 CP->getAlignment()); 2115 else 2116 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, 2117 CP->getAlignment()); 2118 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res); 2119 } 2120 2121 unsigned ARMTargetLowering::getJumpTableEncoding() const { 2122 return MachineJumpTableInfo::EK_Inline; 2123 } 2124 2125 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, 2126 SelectionDAG &DAG) const { 2127 MachineFunction &MF = DAG.getMachineFunction(); 2128 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2129 unsigned ARMPCLabelIndex = 0; 2130 DebugLoc DL = Op.getDebugLoc(); 2131 EVT PtrVT = getPointerTy(); 2132 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress(); 2133 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2134 SDValue CPAddr; 2135 if (RelocM == Reloc::Static) { 2136 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4); 2137 } else { 2138 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8; 2139 ARMPCLabelIndex = AFI->createPICLabelUId(); 2140 ARMConstantPoolValue *CPV = 2141 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex, 2142 ARMCP::CPBlockAddress, PCAdj); 2143 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2144 } 2145 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr); 2146 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr, 2147 MachinePointerInfo::getConstantPool(), 2148 false, false, false, 0); 2149 if (RelocM == Reloc::Static) 2150 return Result; 2151 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2152 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel); 2153 } 2154 2155 // Lower ISD::GlobalTLSAddress using the "general dynamic" model 2156 SDValue 2157 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, 2158 SelectionDAG &DAG) const { 2159 DebugLoc dl = GA->getDebugLoc(); 2160 EVT PtrVT = getPointerTy(); 2161 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8; 2162 MachineFunction &MF = DAG.getMachineFunction(); 2163 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2164 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2165 ARMConstantPoolValue *CPV = 2166 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex, 2167 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true); 2168 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2169 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument); 2170 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, 2171 MachinePointerInfo::getConstantPool(), 2172 false, false, false, 0); 2173 SDValue Chain = Argument.getValue(1); 2174 2175 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2176 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel); 2177 2178 // call __tls_get_addr. 2179 ArgListTy Args; 2180 ArgListEntry Entry; 2181 Entry.Node = Argument; 2182 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext()); 2183 Args.push_back(Entry); 2184 // FIXME: is there useful debug info available here? 2185 TargetLowering::CallLoweringInfo CLI(Chain, 2186 (Type *) Type::getInt32Ty(*DAG.getContext()), 2187 false, false, false, false, 2188 0, CallingConv::C, /*isTailCall=*/false, 2189 /*doesNotRet=*/false, /*isReturnValueUsed=*/true, 2190 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl); 2191 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); 2192 return CallResult.first; 2193 } 2194 2195 // Lower ISD::GlobalTLSAddress using the "initial exec" or 2196 // "local exec" model. 2197 SDValue 2198 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA, 2199 SelectionDAG &DAG, 2200 TLSModel::Model model) const { 2201 const GlobalValue *GV = GA->getGlobal(); 2202 DebugLoc dl = GA->getDebugLoc(); 2203 SDValue Offset; 2204 SDValue Chain = DAG.getEntryNode(); 2205 EVT PtrVT = getPointerTy(); 2206 // Get the Thread Pointer 2207 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT); 2208 2209 if (model == TLSModel::InitialExec) { 2210 MachineFunction &MF = DAG.getMachineFunction(); 2211 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2212 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2213 // Initial exec model. 2214 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8; 2215 ARMConstantPoolValue *CPV = 2216 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex, 2217 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, 2218 true); 2219 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2220 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset); 2221 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, 2222 MachinePointerInfo::getConstantPool(), 2223 false, false, false, 0); 2224 Chain = Offset.getValue(1); 2225 2226 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2227 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel); 2228 2229 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, 2230 MachinePointerInfo::getConstantPool(), 2231 false, false, false, 0); 2232 } else { 2233 // local exec model 2234 assert(model == TLSModel::LocalExec); 2235 ARMConstantPoolValue *CPV = 2236 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF); 2237 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2238 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset); 2239 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, 2240 MachinePointerInfo::getConstantPool(), 2241 false, false, false, 0); 2242 } 2243 2244 // The address of the thread local variable is the add of the thread 2245 // pointer with the offset of the variable. 2246 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset); 2247 } 2248 2249 SDValue 2250 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const { 2251 // TODO: implement the "local dynamic" model 2252 assert(Subtarget->isTargetELF() && 2253 "TLS not implemented for non-ELF targets"); 2254 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 2255 2256 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal()); 2257 2258 switch (model) { 2259 case TLSModel::GeneralDynamic: 2260 case TLSModel::LocalDynamic: 2261 return LowerToTLSGeneralDynamicModel(GA, DAG); 2262 case TLSModel::InitialExec: 2263 case TLSModel::LocalExec: 2264 return LowerToTLSExecModels(GA, DAG, model); 2265 } 2266 llvm_unreachable("bogus TLS model"); 2267 } 2268 2269 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op, 2270 SelectionDAG &DAG) const { 2271 EVT PtrVT = getPointerTy(); 2272 DebugLoc dl = Op.getDebugLoc(); 2273 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 2274 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) { 2275 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility(); 2276 ARMConstantPoolValue *CPV = 2277 ARMConstantPoolConstant::Create(GV, 2278 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT); 2279 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2280 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2281 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), 2282 CPAddr, 2283 MachinePointerInfo::getConstantPool(), 2284 false, false, false, 0); 2285 SDValue Chain = Result.getValue(1); 2286 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT); 2287 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT); 2288 if (!UseGOTOFF) 2289 Result = DAG.getLoad(PtrVT, dl, Chain, Result, 2290 MachinePointerInfo::getGOT(), 2291 false, false, false, 0); 2292 return Result; 2293 } 2294 2295 // If we have T2 ops, we can materialize the address directly via movt/movw 2296 // pair. This is always cheaper. 2297 if (Subtarget->useMovt()) { 2298 ++NumMovwMovt; 2299 // FIXME: Once remat is capable of dealing with instructions with register 2300 // operands, expand this into two nodes. 2301 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT, 2302 DAG.getTargetGlobalAddress(GV, dl, PtrVT)); 2303 } else { 2304 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4); 2305 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2306 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2307 MachinePointerInfo::getConstantPool(), 2308 false, false, false, 0); 2309 } 2310 } 2311 2312 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op, 2313 SelectionDAG &DAG) const { 2314 EVT PtrVT = getPointerTy(); 2315 DebugLoc dl = Op.getDebugLoc(); 2316 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 2317 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2318 2319 // FIXME: Enable this for static codegen when tool issues are fixed. Also 2320 // update ARMFastISel::ARMMaterializeGV. 2321 if (Subtarget->useMovt() && RelocM != Reloc::Static) { 2322 ++NumMovwMovt; 2323 // FIXME: Once remat is capable of dealing with instructions with register 2324 // operands, expand this into two nodes. 2325 if (RelocM == Reloc::Static) 2326 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT, 2327 DAG.getTargetGlobalAddress(GV, dl, PtrVT)); 2328 2329 unsigned Wrapper = (RelocM == Reloc::PIC_) 2330 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN; 2331 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, 2332 DAG.getTargetGlobalAddress(GV, dl, PtrVT)); 2333 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) 2334 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result, 2335 MachinePointerInfo::getGOT(), 2336 false, false, false, 0); 2337 return Result; 2338 } 2339 2340 unsigned ARMPCLabelIndex = 0; 2341 SDValue CPAddr; 2342 if (RelocM == Reloc::Static) { 2343 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4); 2344 } else { 2345 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>(); 2346 ARMPCLabelIndex = AFI->createPICLabelUId(); 2347 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8); 2348 ARMConstantPoolValue *CPV = 2349 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 2350 PCAdj); 2351 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2352 } 2353 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2354 2355 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2356 MachinePointerInfo::getConstantPool(), 2357 false, false, false, 0); 2358 SDValue Chain = Result.getValue(1); 2359 2360 if (RelocM == Reloc::PIC_) { 2361 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2362 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 2363 } 2364 2365 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) 2366 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(), 2367 false, false, false, 0); 2368 2369 return Result; 2370 } 2371 2372 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op, 2373 SelectionDAG &DAG) const { 2374 assert(Subtarget->isTargetELF() && 2375 "GLOBAL OFFSET TABLE not implemented for non-ELF targets"); 2376 MachineFunction &MF = DAG.getMachineFunction(); 2377 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2378 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2379 EVT PtrVT = getPointerTy(); 2380 DebugLoc dl = Op.getDebugLoc(); 2381 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8; 2382 ARMConstantPoolValue *CPV = 2383 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_", 2384 ARMPCLabelIndex, PCAdj); 2385 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2386 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2387 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2388 MachinePointerInfo::getConstantPool(), 2389 false, false, false, 0); 2390 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2391 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 2392 } 2393 2394 SDValue 2395 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const { 2396 DebugLoc dl = Op.getDebugLoc(); 2397 SDValue Val = DAG.getConstant(0, MVT::i32); 2398 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, 2399 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0), 2400 Op.getOperand(1), Val); 2401 } 2402 2403 SDValue 2404 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const { 2405 DebugLoc dl = Op.getDebugLoc(); 2406 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0), 2407 Op.getOperand(1), DAG.getConstant(0, MVT::i32)); 2408 } 2409 2410 SDValue 2411 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG, 2412 const ARMSubtarget *Subtarget) const { 2413 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 2414 DebugLoc dl = Op.getDebugLoc(); 2415 switch (IntNo) { 2416 default: return SDValue(); // Don't custom lower most intrinsics. 2417 case Intrinsic::arm_thread_pointer: { 2418 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 2419 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT); 2420 } 2421 case Intrinsic::eh_sjlj_lsda: { 2422 MachineFunction &MF = DAG.getMachineFunction(); 2423 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2424 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2425 EVT PtrVT = getPointerTy(); 2426 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2427 SDValue CPAddr; 2428 unsigned PCAdj = (RelocM != Reloc::PIC_) 2429 ? 0 : (Subtarget->isThumb() ? 4 : 8); 2430 ARMConstantPoolValue *CPV = 2431 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex, 2432 ARMCP::CPLSDA, PCAdj); 2433 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2434 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2435 SDValue Result = 2436 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2437 MachinePointerInfo::getConstantPool(), 2438 false, false, false, 0); 2439 2440 if (RelocM == Reloc::PIC_) { 2441 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2442 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 2443 } 2444 return Result; 2445 } 2446 case Intrinsic::arm_neon_vmulls: 2447 case Intrinsic::arm_neon_vmullu: { 2448 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls) 2449 ? ARMISD::VMULLs : ARMISD::VMULLu; 2450 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(), 2451 Op.getOperand(1), Op.getOperand(2)); 2452 } 2453 } 2454 } 2455 2456 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG, 2457 const ARMSubtarget *Subtarget) { 2458 DebugLoc dl = Op.getDebugLoc(); 2459 if (!Subtarget->hasDataBarrier()) { 2460 // Some ARMv6 cpus can support data barriers with an mcr instruction. 2461 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get 2462 // here. 2463 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() && 2464 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!"); 2465 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0), 2466 DAG.getConstant(0, MVT::i32)); 2467 } 2468 2469 SDValue Op5 = Op.getOperand(5); 2470 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0; 2471 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); 2472 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue(); 2473 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0); 2474 2475 ARM_MB::MemBOpt DMBOpt; 2476 if (isDeviceBarrier) 2477 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY; 2478 else 2479 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH; 2480 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0), 2481 DAG.getConstant(DMBOpt, MVT::i32)); 2482 } 2483 2484 2485 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG, 2486 const ARMSubtarget *Subtarget) { 2487 // FIXME: handle "fence singlethread" more efficiently. 2488 DebugLoc dl = Op.getDebugLoc(); 2489 if (!Subtarget->hasDataBarrier()) { 2490 // Some ARMv6 cpus can support data barriers with an mcr instruction. 2491 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get 2492 // here. 2493 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() && 2494 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!"); 2495 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0), 2496 DAG.getConstant(0, MVT::i32)); 2497 } 2498 2499 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0), 2500 DAG.getConstant(ARM_MB::ISH, MVT::i32)); 2501 } 2502 2503 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG, 2504 const ARMSubtarget *Subtarget) { 2505 // ARM pre v5TE and Thumb1 does not have preload instructions. 2506 if (!(Subtarget->isThumb2() || 2507 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps()))) 2508 // Just preserve the chain. 2509 return Op.getOperand(0); 2510 2511 DebugLoc dl = Op.getDebugLoc(); 2512 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1; 2513 if (!isRead && 2514 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension())) 2515 // ARMv7 with MP extension has PLDW. 2516 return Op.getOperand(0); 2517 2518 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue(); 2519 if (Subtarget->isThumb()) { 2520 // Invert the bits. 2521 isRead = ~isRead & 1; 2522 isData = ~isData & 1; 2523 } 2524 2525 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0), 2526 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32), 2527 DAG.getConstant(isData, MVT::i32)); 2528 } 2529 2530 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) { 2531 MachineFunction &MF = DAG.getMachineFunction(); 2532 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>(); 2533 2534 // vastart just stores the address of the VarArgsFrameIndex slot into the 2535 // memory location argument. 2536 DebugLoc dl = Op.getDebugLoc(); 2537 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 2538 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); 2539 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 2540 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), 2541 MachinePointerInfo(SV), false, false, 0); 2542 } 2543 2544 SDValue 2545 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA, 2546 SDValue &Root, SelectionDAG &DAG, 2547 DebugLoc dl) const { 2548 MachineFunction &MF = DAG.getMachineFunction(); 2549 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2550 2551 const TargetRegisterClass *RC; 2552 if (AFI->isThumb1OnlyFunction()) 2553 RC = &ARM::tGPRRegClass; 2554 else 2555 RC = &ARM::GPRRegClass; 2556 2557 // Transform the arguments stored in physical registers into virtual ones. 2558 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); 2559 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32); 2560 2561 SDValue ArgValue2; 2562 if (NextVA.isMemLoc()) { 2563 MachineFrameInfo *MFI = MF.getFrameInfo(); 2564 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true); 2565 2566 // Create load node to retrieve arguments from the stack. 2567 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2568 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, 2569 MachinePointerInfo::getFixedStack(FI), 2570 false, false, false, 0); 2571 } else { 2572 Reg = MF.addLiveIn(NextVA.getLocReg(), RC); 2573 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32); 2574 } 2575 2576 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2); 2577 } 2578 2579 void 2580 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF, 2581 unsigned &VARegSize, unsigned &VARegSaveSize) 2582 const { 2583 unsigned NumGPRs; 2584 if (CCInfo.isFirstByValRegValid()) 2585 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg(); 2586 else { 2587 unsigned int firstUnalloced; 2588 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs, 2589 sizeof(GPRArgRegs) / 2590 sizeof(GPRArgRegs[0])); 2591 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0; 2592 } 2593 2594 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment(); 2595 VARegSize = NumGPRs * 4; 2596 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1); 2597 } 2598 2599 // The remaining GPRs hold either the beginning of variable-argument 2600 // data, or the beginning of an aggregate passed by value (usually 2601 // byval). Either way, we allocate stack slots adjacent to the data 2602 // provided by our caller, and store the unallocated registers there. 2603 // If this is a variadic function, the va_list pointer will begin with 2604 // these values; otherwise, this reassembles a (byval) structure that 2605 // was split between registers and memory. 2606 void 2607 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG, 2608 DebugLoc dl, SDValue &Chain, 2609 const Value *OrigArg, 2610 unsigned OffsetFromOrigArg, 2611 unsigned ArgOffset, 2612 bool ForceMutable) const { 2613 MachineFunction &MF = DAG.getMachineFunction(); 2614 MachineFrameInfo *MFI = MF.getFrameInfo(); 2615 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2616 unsigned firstRegToSaveIndex; 2617 if (CCInfo.isFirstByValRegValid()) 2618 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0; 2619 else { 2620 firstRegToSaveIndex = CCInfo.getFirstUnallocated 2621 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0])); 2622 } 2623 2624 unsigned VARegSize, VARegSaveSize; 2625 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize); 2626 if (VARegSaveSize) { 2627 // If this function is vararg, store any remaining integer argument regs 2628 // to their spots on the stack so that they may be loaded by deferencing 2629 // the result of va_next. 2630 AFI->setVarArgsRegSaveSize(VARegSaveSize); 2631 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize, 2632 ArgOffset + VARegSaveSize 2633 - VARegSize, 2634 false)); 2635 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(), 2636 getPointerTy()); 2637 2638 SmallVector<SDValue, 4> MemOps; 2639 for (unsigned i = 0; firstRegToSaveIndex < 4; ++firstRegToSaveIndex, ++i) { 2640 const TargetRegisterClass *RC; 2641 if (AFI->isThumb1OnlyFunction()) 2642 RC = &ARM::tGPRRegClass; 2643 else 2644 RC = &ARM::GPRRegClass; 2645 2646 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC); 2647 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32); 2648 SDValue Store = 2649 DAG.getStore(Val.getValue(1), dl, Val, FIN, 2650 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i), 2651 false, false, 0); 2652 MemOps.push_back(Store); 2653 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN, 2654 DAG.getConstant(4, getPointerTy())); 2655 } 2656 if (!MemOps.empty()) 2657 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 2658 &MemOps[0], MemOps.size()); 2659 } else 2660 // This will point to the next argument passed via stack. 2661 AFI->setVarArgsFrameIndex( 2662 MFI->CreateFixedObject(4, ArgOffset, !ForceMutable)); 2663 } 2664 2665 SDValue 2666 ARMTargetLowering::LowerFormalArguments(SDValue Chain, 2667 CallingConv::ID CallConv, bool isVarArg, 2668 const SmallVectorImpl<ISD::InputArg> 2669 &Ins, 2670 DebugLoc dl, SelectionDAG &DAG, 2671 SmallVectorImpl<SDValue> &InVals) 2672 const { 2673 MachineFunction &MF = DAG.getMachineFunction(); 2674 MachineFrameInfo *MFI = MF.getFrameInfo(); 2675 2676 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2677 2678 // Assign locations to all of the incoming arguments. 2679 SmallVector<CCValAssign, 16> ArgLocs; 2680 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 2681 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue); 2682 CCInfo.AnalyzeFormalArguments(Ins, 2683 CCAssignFnForNode(CallConv, /* Return*/ false, 2684 isVarArg)); 2685 2686 SmallVector<SDValue, 16> ArgValues; 2687 int lastInsIndex = -1; 2688 SDValue ArgValue; 2689 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin(); 2690 unsigned CurArgIdx = 0; 2691 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 2692 CCValAssign &VA = ArgLocs[i]; 2693 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx); 2694 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex; 2695 // Arguments stored in registers. 2696 if (VA.isRegLoc()) { 2697 EVT RegVT = VA.getLocVT(); 2698 2699 if (VA.needsCustom()) { 2700 // f64 and vector types are split up into multiple registers or 2701 // combinations of registers and stack slots. 2702 if (VA.getLocVT() == MVT::v2f64) { 2703 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i], 2704 Chain, DAG, dl); 2705 VA = ArgLocs[++i]; // skip ahead to next loc 2706 SDValue ArgValue2; 2707 if (VA.isMemLoc()) { 2708 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true); 2709 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2710 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN, 2711 MachinePointerInfo::getFixedStack(FI), 2712 false, false, false, 0); 2713 } else { 2714 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i], 2715 Chain, DAG, dl); 2716 } 2717 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64); 2718 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, 2719 ArgValue, ArgValue1, DAG.getIntPtrConstant(0)); 2720 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, 2721 ArgValue, ArgValue2, DAG.getIntPtrConstant(1)); 2722 } else 2723 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl); 2724 2725 } else { 2726 const TargetRegisterClass *RC; 2727 2728 if (RegVT == MVT::f32) 2729 RC = &ARM::SPRRegClass; 2730 else if (RegVT == MVT::f64) 2731 RC = &ARM::DPRRegClass; 2732 else if (RegVT == MVT::v2f64) 2733 RC = &ARM::QPRRegClass; 2734 else if (RegVT == MVT::i32) 2735 RC = AFI->isThumb1OnlyFunction() ? 2736 (const TargetRegisterClass*)&ARM::tGPRRegClass : 2737 (const TargetRegisterClass*)&ARM::GPRRegClass; 2738 else 2739 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering"); 2740 2741 // Transform the arguments in physical registers into virtual ones. 2742 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); 2743 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT); 2744 } 2745 2746 // If this is an 8 or 16-bit value, it is really passed promoted 2747 // to 32 bits. Insert an assert[sz]ext to capture this, then 2748 // truncate to the right size. 2749 switch (VA.getLocInfo()) { 2750 default: llvm_unreachable("Unknown loc info!"); 2751 case CCValAssign::Full: break; 2752 case CCValAssign::BCvt: 2753 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue); 2754 break; 2755 case CCValAssign::SExt: 2756 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue, 2757 DAG.getValueType(VA.getValVT())); 2758 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 2759 break; 2760 case CCValAssign::ZExt: 2761 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue, 2762 DAG.getValueType(VA.getValVT())); 2763 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 2764 break; 2765 } 2766 2767 InVals.push_back(ArgValue); 2768 2769 } else { // VA.isRegLoc() 2770 2771 // sanity check 2772 assert(VA.isMemLoc()); 2773 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered"); 2774 2775 int index = ArgLocs[i].getValNo(); 2776 2777 // Some Ins[] entries become multiple ArgLoc[] entries. 2778 // Process them only once. 2779 if (index != lastInsIndex) 2780 { 2781 ISD::ArgFlagsTy Flags = Ins[index].Flags; 2782 // FIXME: For now, all byval parameter objects are marked mutable. 2783 // This can be changed with more analysis. 2784 // In case of tail call optimization mark all arguments mutable. 2785 // Since they could be overwritten by lowering of arguments in case of 2786 // a tail call. 2787 if (Flags.isByVal()) { 2788 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2789 if (!AFI->getVarArgsFrameIndex()) { 2790 VarArgStyleRegisters(CCInfo, DAG, 2791 dl, Chain, CurOrigArg, 2792 Ins[VA.getValNo()].PartOffset, 2793 VA.getLocMemOffset(), 2794 true /*force mutable frames*/); 2795 int VAFrameIndex = AFI->getVarArgsFrameIndex(); 2796 InVals.push_back(DAG.getFrameIndex(VAFrameIndex, getPointerTy())); 2797 } else { 2798 int FI = MFI->CreateFixedObject(Flags.getByValSize(), 2799 VA.getLocMemOffset(), false); 2800 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy())); 2801 } 2802 } else { 2803 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8, 2804 VA.getLocMemOffset(), true); 2805 2806 // Create load nodes to retrieve arguments from the stack. 2807 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2808 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, 2809 MachinePointerInfo::getFixedStack(FI), 2810 false, false, false, 0)); 2811 } 2812 lastInsIndex = index; 2813 } 2814 } 2815 } 2816 2817 // varargs 2818 if (isVarArg) 2819 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0, 0, 2820 CCInfo.getNextStackOffset()); 2821 2822 return Chain; 2823 } 2824 2825 /// isFloatingPointZero - Return true if this is +0.0. 2826 static bool isFloatingPointZero(SDValue Op) { 2827 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op)) 2828 return CFP->getValueAPF().isPosZero(); 2829 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) { 2830 // Maybe this has already been legalized into the constant pool? 2831 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) { 2832 SDValue WrapperOp = Op.getOperand(1).getOperand(0); 2833 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp)) 2834 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal())) 2835 return CFP->getValueAPF().isPosZero(); 2836 } 2837 } 2838 return false; 2839 } 2840 2841 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for 2842 /// the given operands. 2843 SDValue 2844 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC, 2845 SDValue &ARMcc, SelectionDAG &DAG, 2846 DebugLoc dl) const { 2847 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) { 2848 unsigned C = RHSC->getZExtValue(); 2849 if (!isLegalICmpImmediate(C)) { 2850 // Constant does not fit, try adjusting it by one? 2851 switch (CC) { 2852 default: break; 2853 case ISD::SETLT: 2854 case ISD::SETGE: 2855 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) { 2856 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT; 2857 RHS = DAG.getConstant(C-1, MVT::i32); 2858 } 2859 break; 2860 case ISD::SETULT: 2861 case ISD::SETUGE: 2862 if (C != 0 && isLegalICmpImmediate(C-1)) { 2863 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT; 2864 RHS = DAG.getConstant(C-1, MVT::i32); 2865 } 2866 break; 2867 case ISD::SETLE: 2868 case ISD::SETGT: 2869 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) { 2870 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE; 2871 RHS = DAG.getConstant(C+1, MVT::i32); 2872 } 2873 break; 2874 case ISD::SETULE: 2875 case ISD::SETUGT: 2876 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) { 2877 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; 2878 RHS = DAG.getConstant(C+1, MVT::i32); 2879 } 2880 break; 2881 } 2882 } 2883 } 2884 2885 ARMCC::CondCodes CondCode = IntCCToARMCC(CC); 2886 ARMISD::NodeType CompareType; 2887 switch (CondCode) { 2888 default: 2889 CompareType = ARMISD::CMP; 2890 break; 2891 case ARMCC::EQ: 2892 case ARMCC::NE: 2893 // Uses only Z Flag 2894 CompareType = ARMISD::CMPZ; 2895 break; 2896 } 2897 ARMcc = DAG.getConstant(CondCode, MVT::i32); 2898 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS); 2899 } 2900 2901 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands. 2902 SDValue 2903 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG, 2904 DebugLoc dl) const { 2905 SDValue Cmp; 2906 if (!isFloatingPointZero(RHS)) 2907 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS); 2908 else 2909 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS); 2910 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp); 2911 } 2912 2913 /// duplicateCmp - Glue values can have only one use, so this function 2914 /// duplicates a comparison node. 2915 SDValue 2916 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const { 2917 unsigned Opc = Cmp.getOpcode(); 2918 DebugLoc DL = Cmp.getDebugLoc(); 2919 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ) 2920 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1)); 2921 2922 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation"); 2923 Cmp = Cmp.getOperand(0); 2924 Opc = Cmp.getOpcode(); 2925 if (Opc == ARMISD::CMPFP) 2926 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1)); 2927 else { 2928 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT"); 2929 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0)); 2930 } 2931 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp); 2932 } 2933 2934 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const { 2935 SDValue Cond = Op.getOperand(0); 2936 SDValue SelectTrue = Op.getOperand(1); 2937 SDValue SelectFalse = Op.getOperand(2); 2938 DebugLoc dl = Op.getDebugLoc(); 2939 2940 // Convert: 2941 // 2942 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond) 2943 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond) 2944 // 2945 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) { 2946 const ConstantSDNode *CMOVTrue = 2947 dyn_cast<ConstantSDNode>(Cond.getOperand(0)); 2948 const ConstantSDNode *CMOVFalse = 2949 dyn_cast<ConstantSDNode>(Cond.getOperand(1)); 2950 2951 if (CMOVTrue && CMOVFalse) { 2952 unsigned CMOVTrueVal = CMOVTrue->getZExtValue(); 2953 unsigned CMOVFalseVal = CMOVFalse->getZExtValue(); 2954 2955 SDValue True; 2956 SDValue False; 2957 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) { 2958 True = SelectTrue; 2959 False = SelectFalse; 2960 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) { 2961 True = SelectFalse; 2962 False = SelectTrue; 2963 } 2964 2965 if (True.getNode() && False.getNode()) { 2966 EVT VT = Op.getValueType(); 2967 SDValue ARMcc = Cond.getOperand(2); 2968 SDValue CCR = Cond.getOperand(3); 2969 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG); 2970 assert(True.getValueType() == VT); 2971 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp); 2972 } 2973 } 2974 } 2975 2976 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the 2977 // undefined bits before doing a full-word comparison with zero. 2978 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond, 2979 DAG.getConstant(1, Cond.getValueType())); 2980 2981 return DAG.getSelectCC(dl, Cond, 2982 DAG.getConstant(0, Cond.getValueType()), 2983 SelectTrue, SelectFalse, ISD::SETNE); 2984 } 2985 2986 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const { 2987 EVT VT = Op.getValueType(); 2988 SDValue LHS = Op.getOperand(0); 2989 SDValue RHS = Op.getOperand(1); 2990 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get(); 2991 SDValue TrueVal = Op.getOperand(2); 2992 SDValue FalseVal = Op.getOperand(3); 2993 DebugLoc dl = Op.getDebugLoc(); 2994 2995 if (LHS.getValueType() == MVT::i32) { 2996 SDValue ARMcc; 2997 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 2998 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl); 2999 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp); 3000 } 3001 3002 ARMCC::CondCodes CondCode, CondCode2; 3003 FPCCToARMCC(CC, CondCode, CondCode2); 3004 3005 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32); 3006 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl); 3007 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3008 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, 3009 ARMcc, CCR, Cmp); 3010 if (CondCode2 != ARMCC::AL) { 3011 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32); 3012 // FIXME: Needs another CMP because flag can have but one use. 3013 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl); 3014 Result = DAG.getNode(ARMISD::CMOV, dl, VT, 3015 Result, TrueVal, ARMcc2, CCR, Cmp2); 3016 } 3017 return Result; 3018 } 3019 3020 /// canChangeToInt - Given the fp compare operand, return true if it is suitable 3021 /// to morph to an integer compare sequence. 3022 static bool canChangeToInt(SDValue Op, bool &SeenZero, 3023 const ARMSubtarget *Subtarget) { 3024 SDNode *N = Op.getNode(); 3025 if (!N->hasOneUse()) 3026 // Otherwise it requires moving the value from fp to integer registers. 3027 return false; 3028 if (!N->getNumValues()) 3029 return false; 3030 EVT VT = Op.getValueType(); 3031 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow()) 3032 // f32 case is generally profitable. f64 case only makes sense when vcmpe + 3033 // vmrs are very slow, e.g. cortex-a8. 3034 return false; 3035 3036 if (isFloatingPointZero(Op)) { 3037 SeenZero = true; 3038 return true; 3039 } 3040 return ISD::isNormalLoad(N); 3041 } 3042 3043 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) { 3044 if (isFloatingPointZero(Op)) 3045 return DAG.getConstant(0, MVT::i32); 3046 3047 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) 3048 return DAG.getLoad(MVT::i32, Op.getDebugLoc(), 3049 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(), 3050 Ld->isVolatile(), Ld->isNonTemporal(), 3051 Ld->isInvariant(), Ld->getAlignment()); 3052 3053 llvm_unreachable("Unknown VFP cmp argument!"); 3054 } 3055 3056 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG, 3057 SDValue &RetVal1, SDValue &RetVal2) { 3058 if (isFloatingPointZero(Op)) { 3059 RetVal1 = DAG.getConstant(0, MVT::i32); 3060 RetVal2 = DAG.getConstant(0, MVT::i32); 3061 return; 3062 } 3063 3064 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) { 3065 SDValue Ptr = Ld->getBasePtr(); 3066 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(), 3067 Ld->getChain(), Ptr, 3068 Ld->getPointerInfo(), 3069 Ld->isVolatile(), Ld->isNonTemporal(), 3070 Ld->isInvariant(), Ld->getAlignment()); 3071 3072 EVT PtrType = Ptr.getValueType(); 3073 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4); 3074 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(), 3075 PtrType, Ptr, DAG.getConstant(4, PtrType)); 3076 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(), 3077 Ld->getChain(), NewPtr, 3078 Ld->getPointerInfo().getWithOffset(4), 3079 Ld->isVolatile(), Ld->isNonTemporal(), 3080 Ld->isInvariant(), NewAlign); 3081 return; 3082 } 3083 3084 llvm_unreachable("Unknown VFP cmp argument!"); 3085 } 3086 3087 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some 3088 /// f32 and even f64 comparisons to integer ones. 3089 SDValue 3090 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const { 3091 SDValue Chain = Op.getOperand(0); 3092 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 3093 SDValue LHS = Op.getOperand(2); 3094 SDValue RHS = Op.getOperand(3); 3095 SDValue Dest = Op.getOperand(4); 3096 DebugLoc dl = Op.getDebugLoc(); 3097 3098 bool LHSSeenZero = false; 3099 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget); 3100 bool RHSSeenZero = false; 3101 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget); 3102 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) { 3103 // If unsafe fp math optimization is enabled and there are no other uses of 3104 // the CMP operands, and the condition code is EQ or NE, we can optimize it 3105 // to an integer comparison. 3106 if (CC == ISD::SETOEQ) 3107 CC = ISD::SETEQ; 3108 else if (CC == ISD::SETUNE) 3109 CC = ISD::SETNE; 3110 3111 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32); 3112 SDValue ARMcc; 3113 if (LHS.getValueType() == MVT::f32) { 3114 LHS = DAG.getNode(ISD::AND, dl, MVT::i32, 3115 bitcastf32Toi32(LHS, DAG), Mask); 3116 RHS = DAG.getNode(ISD::AND, dl, MVT::i32, 3117 bitcastf32Toi32(RHS, DAG), Mask); 3118 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl); 3119 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3120 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other, 3121 Chain, Dest, ARMcc, CCR, Cmp); 3122 } 3123 3124 SDValue LHS1, LHS2; 3125 SDValue RHS1, RHS2; 3126 expandf64Toi32(LHS, DAG, LHS1, LHS2); 3127 expandf64Toi32(RHS, DAG, RHS1, RHS2); 3128 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask); 3129 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask); 3130 ARMCC::CondCodes CondCode = IntCCToARMCC(CC); 3131 ARMcc = DAG.getConstant(CondCode, MVT::i32); 3132 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue); 3133 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest }; 3134 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7); 3135 } 3136 3137 return SDValue(); 3138 } 3139 3140 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const { 3141 SDValue Chain = Op.getOperand(0); 3142 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 3143 SDValue LHS = Op.getOperand(2); 3144 SDValue RHS = Op.getOperand(3); 3145 SDValue Dest = Op.getOperand(4); 3146 DebugLoc dl = Op.getDebugLoc(); 3147 3148 if (LHS.getValueType() == MVT::i32) { 3149 SDValue ARMcc; 3150 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl); 3151 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3152 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other, 3153 Chain, Dest, ARMcc, CCR, Cmp); 3154 } 3155 3156 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64); 3157 3158 if (getTargetMachine().Options.UnsafeFPMath && 3159 (CC == ISD::SETEQ || CC == ISD::SETOEQ || 3160 CC == ISD::SETNE || CC == ISD::SETUNE)) { 3161 SDValue Result = OptimizeVFPBrcond(Op, DAG); 3162 if (Result.getNode()) 3163 return Result; 3164 } 3165 3166 ARMCC::CondCodes CondCode, CondCode2; 3167 FPCCToARMCC(CC, CondCode, CondCode2); 3168 3169 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32); 3170 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl); 3171 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3172 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue); 3173 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp }; 3174 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5); 3175 if (CondCode2 != ARMCC::AL) { 3176 ARMcc = DAG.getConstant(CondCode2, MVT::i32); 3177 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) }; 3178 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5); 3179 } 3180 return Res; 3181 } 3182 3183 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const { 3184 SDValue Chain = Op.getOperand(0); 3185 SDValue Table = Op.getOperand(1); 3186 SDValue Index = Op.getOperand(2); 3187 DebugLoc dl = Op.getDebugLoc(); 3188 3189 EVT PTy = getPointerTy(); 3190 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table); 3191 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>(); 3192 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy); 3193 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy); 3194 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId); 3195 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy)); 3196 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table); 3197 if (Subtarget->isThumb2()) { 3198 // Thumb2 uses a two-level jump. That is, it jumps into the jump table 3199 // which does another jump to the destination. This also makes it easier 3200 // to translate it to TBB / TBH later. 3201 // FIXME: This might not work if the function is extremely large. 3202 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain, 3203 Addr, Op.getOperand(2), JTI, UId); 3204 } 3205 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) { 3206 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, 3207 MachinePointerInfo::getJumpTable(), 3208 false, false, false, 0); 3209 Chain = Addr.getValue(1); 3210 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table); 3211 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId); 3212 } else { 3213 Addr = DAG.getLoad(PTy, dl, Chain, Addr, 3214 MachinePointerInfo::getJumpTable(), 3215 false, false, false, 0); 3216 Chain = Addr.getValue(1); 3217 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId); 3218 } 3219 } 3220 3221 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) { 3222 EVT VT = Op.getValueType(); 3223 DebugLoc dl = Op.getDebugLoc(); 3224 3225 if (Op.getValueType().getVectorElementType() == MVT::i32) { 3226 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32) 3227 return Op; 3228 return DAG.UnrollVectorOp(Op.getNode()); 3229 } 3230 3231 assert(Op.getOperand(0).getValueType() == MVT::v4f32 && 3232 "Invalid type for custom lowering!"); 3233 if (VT != MVT::v4i16) 3234 return DAG.UnrollVectorOp(Op.getNode()); 3235 3236 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0)); 3237 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op); 3238 } 3239 3240 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) { 3241 EVT VT = Op.getValueType(); 3242 if (VT.isVector()) 3243 return LowerVectorFP_TO_INT(Op, DAG); 3244 3245 DebugLoc dl = Op.getDebugLoc(); 3246 unsigned Opc; 3247 3248 switch (Op.getOpcode()) { 3249 default: llvm_unreachable("Invalid opcode!"); 3250 case ISD::FP_TO_SINT: 3251 Opc = ARMISD::FTOSI; 3252 break; 3253 case ISD::FP_TO_UINT: 3254 Opc = ARMISD::FTOUI; 3255 break; 3256 } 3257 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0)); 3258 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); 3259 } 3260 3261 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) { 3262 EVT VT = Op.getValueType(); 3263 DebugLoc dl = Op.getDebugLoc(); 3264 3265 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) { 3266 if (VT.getVectorElementType() == MVT::f32) 3267 return Op; 3268 return DAG.UnrollVectorOp(Op.getNode()); 3269 } 3270 3271 assert(Op.getOperand(0).getValueType() == MVT::v4i16 && 3272 "Invalid type for custom lowering!"); 3273 if (VT != MVT::v4f32) 3274 return DAG.UnrollVectorOp(Op.getNode()); 3275 3276 unsigned CastOpc; 3277 unsigned Opc; 3278 switch (Op.getOpcode()) { 3279 default: llvm_unreachable("Invalid opcode!"); 3280 case ISD::SINT_TO_FP: 3281 CastOpc = ISD::SIGN_EXTEND; 3282 Opc = ISD::SINT_TO_FP; 3283 break; 3284 case ISD::UINT_TO_FP: 3285 CastOpc = ISD::ZERO_EXTEND; 3286 Opc = ISD::UINT_TO_FP; 3287 break; 3288 } 3289 3290 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0)); 3291 return DAG.getNode(Opc, dl, VT, Op); 3292 } 3293 3294 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) { 3295 EVT VT = Op.getValueType(); 3296 if (VT.isVector()) 3297 return LowerVectorINT_TO_FP(Op, DAG); 3298 3299 DebugLoc dl = Op.getDebugLoc(); 3300 unsigned Opc; 3301 3302 switch (Op.getOpcode()) { 3303 default: llvm_unreachable("Invalid opcode!"); 3304 case ISD::SINT_TO_FP: 3305 Opc = ARMISD::SITOF; 3306 break; 3307 case ISD::UINT_TO_FP: 3308 Opc = ARMISD::UITOF; 3309 break; 3310 } 3311 3312 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0)); 3313 return DAG.getNode(Opc, dl, VT, Op); 3314 } 3315 3316 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { 3317 // Implement fcopysign with a fabs and a conditional fneg. 3318 SDValue Tmp0 = Op.getOperand(0); 3319 SDValue Tmp1 = Op.getOperand(1); 3320 DebugLoc dl = Op.getDebugLoc(); 3321 EVT VT = Op.getValueType(); 3322 EVT SrcVT = Tmp1.getValueType(); 3323 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST || 3324 Tmp0.getOpcode() == ARMISD::VMOVDRR; 3325 bool UseNEON = !InGPR && Subtarget->hasNEON(); 3326 3327 if (UseNEON) { 3328 // Use VBSL to copy the sign bit. 3329 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80); 3330 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32, 3331 DAG.getTargetConstant(EncodedVal, MVT::i32)); 3332 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64; 3333 if (VT == MVT::f64) 3334 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT, 3335 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask), 3336 DAG.getConstant(32, MVT::i32)); 3337 else /*if (VT == MVT::f32)*/ 3338 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0); 3339 if (SrcVT == MVT::f32) { 3340 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1); 3341 if (VT == MVT::f64) 3342 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT, 3343 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1), 3344 DAG.getConstant(32, MVT::i32)); 3345 } else if (VT == MVT::f32) 3346 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64, 3347 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1), 3348 DAG.getConstant(32, MVT::i32)); 3349 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0); 3350 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1); 3351 3352 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff), 3353 MVT::i32); 3354 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes); 3355 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask, 3356 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes)); 3357 3358 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT, 3359 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask), 3360 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot)); 3361 if (VT == MVT::f32) { 3362 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res); 3363 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res, 3364 DAG.getConstant(0, MVT::i32)); 3365 } else { 3366 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res); 3367 } 3368 3369 return Res; 3370 } 3371 3372 // Bitcast operand 1 to i32. 3373 if (SrcVT == MVT::f64) 3374 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32), 3375 &Tmp1, 1).getValue(1); 3376 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1); 3377 3378 // Or in the signbit with integer operations. 3379 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32); 3380 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32); 3381 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1); 3382 if (VT == MVT::f32) { 3383 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32, 3384 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2); 3385 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, 3386 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1)); 3387 } 3388 3389 // f64: Or the high part with signbit and then combine two parts. 3390 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32), 3391 &Tmp0, 1); 3392 SDValue Lo = Tmp0.getValue(0); 3393 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2); 3394 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1); 3395 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi); 3396 } 3397 3398 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{ 3399 MachineFunction &MF = DAG.getMachineFunction(); 3400 MachineFrameInfo *MFI = MF.getFrameInfo(); 3401 MFI->setReturnAddressIsTaken(true); 3402 3403 EVT VT = Op.getValueType(); 3404 DebugLoc dl = Op.getDebugLoc(); 3405 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 3406 if (Depth) { 3407 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG); 3408 SDValue Offset = DAG.getConstant(4, MVT::i32); 3409 return DAG.getLoad(VT, dl, DAG.getEntryNode(), 3410 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset), 3411 MachinePointerInfo(), false, false, false, 0); 3412 } 3413 3414 // Return LR, which contains the return address. Mark it an implicit live-in. 3415 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32)); 3416 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT); 3417 } 3418 3419 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { 3420 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 3421 MFI->setFrameAddressIsTaken(true); 3422 3423 EVT VT = Op.getValueType(); 3424 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful 3425 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 3426 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin()) 3427 ? ARM::R7 : ARM::R11; 3428 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT); 3429 while (Depth--) 3430 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, 3431 MachinePointerInfo(), 3432 false, false, false, 0); 3433 return FrameAddr; 3434 } 3435 3436 /// ExpandBITCAST - If the target supports VFP, this function is called to 3437 /// expand a bit convert where either the source or destination type is i64 to 3438 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64 3439 /// operand type is illegal (e.g., v2f32 for a target that doesn't support 3440 /// vectors), since the legalizer won't know what to do with that. 3441 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) { 3442 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3443 DebugLoc dl = N->getDebugLoc(); 3444 SDValue Op = N->getOperand(0); 3445 3446 // This function is only supposed to be called for i64 types, either as the 3447 // source or destination of the bit convert. 3448 EVT SrcVT = Op.getValueType(); 3449 EVT DstVT = N->getValueType(0); 3450 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) && 3451 "ExpandBITCAST called for non-i64 type"); 3452 3453 // Turn i64->f64 into VMOVDRR. 3454 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) { 3455 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op, 3456 DAG.getConstant(0, MVT::i32)); 3457 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op, 3458 DAG.getConstant(1, MVT::i32)); 3459 return DAG.getNode(ISD::BITCAST, dl, DstVT, 3460 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi)); 3461 } 3462 3463 // Turn f64->i64 into VMOVRRD. 3464 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) { 3465 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl, 3466 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1); 3467 // Merge the pieces into a single i64 value. 3468 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1)); 3469 } 3470 3471 return SDValue(); 3472 } 3473 3474 /// getZeroVector - Returns a vector of specified type with all zero elements. 3475 /// Zero vectors are used to represent vector negation and in those cases 3476 /// will be implemented with the NEON VNEG instruction. However, VNEG does 3477 /// not support i64 elements, so sometimes the zero vectors will need to be 3478 /// explicitly constructed. Regardless, use a canonical VMOV to create the 3479 /// zero vector. 3480 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) { 3481 assert(VT.isVector() && "Expected a vector type"); 3482 // The canonical modified immediate encoding of a zero vector is....0! 3483 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32); 3484 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32; 3485 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal); 3486 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov); 3487 } 3488 3489 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two 3490 /// i32 values and take a 2 x i32 value to shift plus a shift amount. 3491 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op, 3492 SelectionDAG &DAG) const { 3493 assert(Op.getNumOperands() == 3 && "Not a double-shift!"); 3494 EVT VT = Op.getValueType(); 3495 unsigned VTBits = VT.getSizeInBits(); 3496 DebugLoc dl = Op.getDebugLoc(); 3497 SDValue ShOpLo = Op.getOperand(0); 3498 SDValue ShOpHi = Op.getOperand(1); 3499 SDValue ShAmt = Op.getOperand(2); 3500 SDValue ARMcc; 3501 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL; 3502 3503 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS); 3504 3505 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, 3506 DAG.getConstant(VTBits, MVT::i32), ShAmt); 3507 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt); 3508 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt, 3509 DAG.getConstant(VTBits, MVT::i32)); 3510 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt); 3511 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2); 3512 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt); 3513 3514 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3515 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE, 3516 ARMcc, DAG, dl); 3517 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt); 3518 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, 3519 CCR, Cmp); 3520 3521 SDValue Ops[2] = { Lo, Hi }; 3522 return DAG.getMergeValues(Ops, 2, dl); 3523 } 3524 3525 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two 3526 /// i32 values and take a 2 x i32 value to shift plus a shift amount. 3527 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op, 3528 SelectionDAG &DAG) const { 3529 assert(Op.getNumOperands() == 3 && "Not a double-shift!"); 3530 EVT VT = Op.getValueType(); 3531 unsigned VTBits = VT.getSizeInBits(); 3532 DebugLoc dl = Op.getDebugLoc(); 3533 SDValue ShOpLo = Op.getOperand(0); 3534 SDValue ShOpHi = Op.getOperand(1); 3535 SDValue ShAmt = Op.getOperand(2); 3536 SDValue ARMcc; 3537 3538 assert(Op.getOpcode() == ISD::SHL_PARTS); 3539 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, 3540 DAG.getConstant(VTBits, MVT::i32), ShAmt); 3541 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt); 3542 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt, 3543 DAG.getConstant(VTBits, MVT::i32)); 3544 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt); 3545 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt); 3546 3547 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2); 3548 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3549 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE, 3550 ARMcc, DAG, dl); 3551 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt); 3552 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc, 3553 CCR, Cmp); 3554 3555 SDValue Ops[2] = { Lo, Hi }; 3556 return DAG.getMergeValues(Ops, 2, dl); 3557 } 3558 3559 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op, 3560 SelectionDAG &DAG) const { 3561 // The rounding mode is in bits 23:22 of the FPSCR. 3562 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0 3563 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3) 3564 // so that the shift + and get folded into a bitfield extract. 3565 DebugLoc dl = Op.getDebugLoc(); 3566 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32, 3567 DAG.getConstant(Intrinsic::arm_get_fpscr, 3568 MVT::i32)); 3569 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR, 3570 DAG.getConstant(1U << 22, MVT::i32)); 3571 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds, 3572 DAG.getConstant(22, MVT::i32)); 3573 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE, 3574 DAG.getConstant(3, MVT::i32)); 3575 } 3576 3577 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG, 3578 const ARMSubtarget *ST) { 3579 EVT VT = N->getValueType(0); 3580 DebugLoc dl = N->getDebugLoc(); 3581 3582 if (!ST->hasV6T2Ops()) 3583 return SDValue(); 3584 3585 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0)); 3586 return DAG.getNode(ISD::CTLZ, dl, VT, rbit); 3587 } 3588 3589 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count 3590 /// for each 16-bit element from operand, repeated. The basic idea is to 3591 /// leverage vcnt to get the 8-bit counts, gather and add the results. 3592 /// 3593 /// Trace for v4i16: 3594 /// input = [v0 v1 v2 v3 ] (vi 16-bit element) 3595 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element) 3596 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi) 3597 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6] 3598 /// [b0 b1 b2 b3 b4 b5 b6 b7] 3599 /// +[b1 b0 b3 b2 b5 b4 b7 b6] 3600 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0, 3601 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits) 3602 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) { 3603 EVT VT = N->getValueType(0); 3604 DebugLoc DL = N->getDebugLoc(); 3605 3606 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8; 3607 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0)); 3608 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0); 3609 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1); 3610 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2); 3611 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3); 3612 } 3613 3614 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the 3615 /// bit-count for each 16-bit element from the operand. We need slightly 3616 /// different sequencing for v4i16 and v8i16 to stay within NEON's available 3617 /// 64/128-bit registers. 3618 /// 3619 /// Trace for v4i16: 3620 /// input = [v0 v1 v2 v3 ] (vi 16-bit element) 3621 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi) 3622 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ] 3623 /// v4i16:Extracted = [k0 k1 k2 k3 ] 3624 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) { 3625 EVT VT = N->getValueType(0); 3626 DebugLoc DL = N->getDebugLoc(); 3627 3628 SDValue BitCounts = getCTPOP16BitCounts(N, DAG); 3629 if (VT.is64BitVector()) { 3630 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts); 3631 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended, 3632 DAG.getIntPtrConstant(0)); 3633 } else { 3634 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8, 3635 BitCounts, DAG.getIntPtrConstant(0)); 3636 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted); 3637 } 3638 } 3639 3640 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the 3641 /// bit-count for each 32-bit element from the operand. The idea here is 3642 /// to split the vector into 16-bit elements, leverage the 16-bit count 3643 /// routine, and then combine the results. 3644 /// 3645 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged): 3646 /// input = [v0 v1 ] (vi: 32-bit elements) 3647 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1]) 3648 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi) 3649 /// vrev: N0 = [k1 k0 k3 k2 ] 3650 /// [k0 k1 k2 k3 ] 3651 /// N1 =+[k1 k0 k3 k2 ] 3652 /// [k0 k2 k1 k3 ] 3653 /// N2 =+[k1 k3 k0 k2 ] 3654 /// [k0 k2 k1 k3 ] 3655 /// Extended =+[k1 k3 k0 k2 ] 3656 /// [k0 k2 ] 3657 /// Extracted=+[k1 k3 ] 3658 /// 3659 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) { 3660 EVT VT = N->getValueType(0); 3661 DebugLoc DL = N->getDebugLoc(); 3662 3663 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16; 3664 3665 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0)); 3666 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG); 3667 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16); 3668 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0); 3669 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1); 3670 3671 if (VT.is64BitVector()) { 3672 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2); 3673 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended, 3674 DAG.getIntPtrConstant(0)); 3675 } else { 3676 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2, 3677 DAG.getIntPtrConstant(0)); 3678 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted); 3679 } 3680 } 3681 3682 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG, 3683 const ARMSubtarget *ST) { 3684 EVT VT = N->getValueType(0); 3685 3686 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON."); 3687 assert((VT == MVT::v2i32 || VT == MVT::v4i32 || 3688 VT == MVT::v4i16 || VT == MVT::v8i16) && 3689 "Unexpected type for custom ctpop lowering"); 3690 3691 if (VT.getVectorElementType() == MVT::i32) 3692 return lowerCTPOP32BitElements(N, DAG); 3693 else 3694 return lowerCTPOP16BitElements(N, DAG); 3695 } 3696 3697 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG, 3698 const ARMSubtarget *ST) { 3699 EVT VT = N->getValueType(0); 3700 DebugLoc dl = N->getDebugLoc(); 3701 3702 if (!VT.isVector()) 3703 return SDValue(); 3704 3705 // Lower vector shifts on NEON to use VSHL. 3706 assert(ST->hasNEON() && "unexpected vector shift"); 3707 3708 // Left shifts translate directly to the vshiftu intrinsic. 3709 if (N->getOpcode() == ISD::SHL) 3710 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, 3711 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32), 3712 N->getOperand(0), N->getOperand(1)); 3713 3714 assert((N->getOpcode() == ISD::SRA || 3715 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode"); 3716 3717 // NEON uses the same intrinsics for both left and right shifts. For 3718 // right shifts, the shift amounts are negative, so negate the vector of 3719 // shift amounts. 3720 EVT ShiftVT = N->getOperand(1).getValueType(); 3721 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT, 3722 getZeroVector(ShiftVT, DAG, dl), 3723 N->getOperand(1)); 3724 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ? 3725 Intrinsic::arm_neon_vshifts : 3726 Intrinsic::arm_neon_vshiftu); 3727 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, 3728 DAG.getConstant(vshiftInt, MVT::i32), 3729 N->getOperand(0), NegatedCount); 3730 } 3731 3732 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG, 3733 const ARMSubtarget *ST) { 3734 EVT VT = N->getValueType(0); 3735 DebugLoc dl = N->getDebugLoc(); 3736 3737 // We can get here for a node like i32 = ISD::SHL i32, i64 3738 if (VT != MVT::i64) 3739 return SDValue(); 3740 3741 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) && 3742 "Unknown shift to lower!"); 3743 3744 // We only lower SRA, SRL of 1 here, all others use generic lowering. 3745 if (!isa<ConstantSDNode>(N->getOperand(1)) || 3746 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1) 3747 return SDValue(); 3748 3749 // If we are in thumb mode, we don't have RRX. 3750 if (ST->isThumb1Only()) return SDValue(); 3751 3752 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr. 3753 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0), 3754 DAG.getConstant(0, MVT::i32)); 3755 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0), 3756 DAG.getConstant(1, MVT::i32)); 3757 3758 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and 3759 // captures the result into a carry flag. 3760 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG; 3761 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1); 3762 3763 // The low part is an ARMISD::RRX operand, which shifts the carry in. 3764 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1)); 3765 3766 // Merge the pieces into a single i64 value. 3767 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi); 3768 } 3769 3770 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) { 3771 SDValue TmpOp0, TmpOp1; 3772 bool Invert = false; 3773 bool Swap = false; 3774 unsigned Opc = 0; 3775 3776 SDValue Op0 = Op.getOperand(0); 3777 SDValue Op1 = Op.getOperand(1); 3778 SDValue CC = Op.getOperand(2); 3779 EVT VT = Op.getValueType(); 3780 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get(); 3781 DebugLoc dl = Op.getDebugLoc(); 3782 3783 if (Op.getOperand(1).getValueType().isFloatingPoint()) { 3784 switch (SetCCOpcode) { 3785 default: llvm_unreachable("Illegal FP comparison"); 3786 case ISD::SETUNE: 3787 case ISD::SETNE: Invert = true; // Fallthrough 3788 case ISD::SETOEQ: 3789 case ISD::SETEQ: Opc = ARMISD::VCEQ; break; 3790 case ISD::SETOLT: 3791 case ISD::SETLT: Swap = true; // Fallthrough 3792 case ISD::SETOGT: 3793 case ISD::SETGT: Opc = ARMISD::VCGT; break; 3794 case ISD::SETOLE: 3795 case ISD::SETLE: Swap = true; // Fallthrough 3796 case ISD::SETOGE: 3797 case ISD::SETGE: Opc = ARMISD::VCGE; break; 3798 case ISD::SETUGE: Swap = true; // Fallthrough 3799 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break; 3800 case ISD::SETUGT: Swap = true; // Fallthrough 3801 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break; 3802 case ISD::SETUEQ: Invert = true; // Fallthrough 3803 case ISD::SETONE: 3804 // Expand this to (OLT | OGT). 3805 TmpOp0 = Op0; 3806 TmpOp1 = Op1; 3807 Opc = ISD::OR; 3808 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0); 3809 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1); 3810 break; 3811 case ISD::SETUO: Invert = true; // Fallthrough 3812 case ISD::SETO: 3813 // Expand this to (OLT | OGE). 3814 TmpOp0 = Op0; 3815 TmpOp1 = Op1; 3816 Opc = ISD::OR; 3817 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0); 3818 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1); 3819 break; 3820 } 3821 } else { 3822 // Integer comparisons. 3823 switch (SetCCOpcode) { 3824 default: llvm_unreachable("Illegal integer comparison"); 3825 case ISD::SETNE: Invert = true; 3826 case ISD::SETEQ: Opc = ARMISD::VCEQ; break; 3827 case ISD::SETLT: Swap = true; 3828 case ISD::SETGT: Opc = ARMISD::VCGT; break; 3829 case ISD::SETLE: Swap = true; 3830 case ISD::SETGE: Opc = ARMISD::VCGE; break; 3831 case ISD::SETULT: Swap = true; 3832 case ISD::SETUGT: Opc = ARMISD::VCGTU; break; 3833 case ISD::SETULE: Swap = true; 3834 case ISD::SETUGE: Opc = ARMISD::VCGEU; break; 3835 } 3836 3837 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero). 3838 if (Opc == ARMISD::VCEQ) { 3839 3840 SDValue AndOp; 3841 if (ISD::isBuildVectorAllZeros(Op1.getNode())) 3842 AndOp = Op0; 3843 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) 3844 AndOp = Op1; 3845 3846 // Ignore bitconvert. 3847 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST) 3848 AndOp = AndOp.getOperand(0); 3849 3850 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) { 3851 Opc = ARMISD::VTST; 3852 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0)); 3853 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1)); 3854 Invert = !Invert; 3855 } 3856 } 3857 } 3858 3859 if (Swap) 3860 std::swap(Op0, Op1); 3861 3862 // If one of the operands is a constant vector zero, attempt to fold the 3863 // comparison to a specialized compare-against-zero form. 3864 SDValue SingleOp; 3865 if (ISD::isBuildVectorAllZeros(Op1.getNode())) 3866 SingleOp = Op0; 3867 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) { 3868 if (Opc == ARMISD::VCGE) 3869 Opc = ARMISD::VCLEZ; 3870 else if (Opc == ARMISD::VCGT) 3871 Opc = ARMISD::VCLTZ; 3872 SingleOp = Op1; 3873 } 3874 3875 SDValue Result; 3876 if (SingleOp.getNode()) { 3877 switch (Opc) { 3878 case ARMISD::VCEQ: 3879 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break; 3880 case ARMISD::VCGE: 3881 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break; 3882 case ARMISD::VCLEZ: 3883 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break; 3884 case ARMISD::VCGT: 3885 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break; 3886 case ARMISD::VCLTZ: 3887 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break; 3888 default: 3889 Result = DAG.getNode(Opc, dl, VT, Op0, Op1); 3890 } 3891 } else { 3892 Result = DAG.getNode(Opc, dl, VT, Op0, Op1); 3893 } 3894 3895 if (Invert) 3896 Result = DAG.getNOT(dl, Result, VT); 3897 3898 return Result; 3899 } 3900 3901 /// isNEONModifiedImm - Check if the specified splat value corresponds to a 3902 /// valid vector constant for a NEON instruction with a "modified immediate" 3903 /// operand (e.g., VMOV). If so, return the encoded value. 3904 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef, 3905 unsigned SplatBitSize, SelectionDAG &DAG, 3906 EVT &VT, bool is128Bits, NEONModImmType type) { 3907 unsigned OpCmode, Imm; 3908 3909 // SplatBitSize is set to the smallest size that splats the vector, so a 3910 // zero vector will always have SplatBitSize == 8. However, NEON modified 3911 // immediate instructions others than VMOV do not support the 8-bit encoding 3912 // of a zero vector, and the default encoding of zero is supposed to be the 3913 // 32-bit version. 3914 if (SplatBits == 0) 3915 SplatBitSize = 32; 3916 3917 switch (SplatBitSize) { 3918 case 8: 3919 if (type != VMOVModImm) 3920 return SDValue(); 3921 // Any 1-byte value is OK. Op=0, Cmode=1110. 3922 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big"); 3923 OpCmode = 0xe; 3924 Imm = SplatBits; 3925 VT = is128Bits ? MVT::v16i8 : MVT::v8i8; 3926 break; 3927 3928 case 16: 3929 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero. 3930 VT = is128Bits ? MVT::v8i16 : MVT::v4i16; 3931 if ((SplatBits & ~0xff) == 0) { 3932 // Value = 0x00nn: Op=x, Cmode=100x. 3933 OpCmode = 0x8; 3934 Imm = SplatBits; 3935 break; 3936 } 3937 if ((SplatBits & ~0xff00) == 0) { 3938 // Value = 0xnn00: Op=x, Cmode=101x. 3939 OpCmode = 0xa; 3940 Imm = SplatBits >> 8; 3941 break; 3942 } 3943 return SDValue(); 3944 3945 case 32: 3946 // NEON's 32-bit VMOV supports splat values where: 3947 // * only one byte is nonzero, or 3948 // * the least significant byte is 0xff and the second byte is nonzero, or 3949 // * the least significant 2 bytes are 0xff and the third is nonzero. 3950 VT = is128Bits ? MVT::v4i32 : MVT::v2i32; 3951 if ((SplatBits & ~0xff) == 0) { 3952 // Value = 0x000000nn: Op=x, Cmode=000x. 3953 OpCmode = 0; 3954 Imm = SplatBits; 3955 break; 3956 } 3957 if ((SplatBits & ~0xff00) == 0) { 3958 // Value = 0x0000nn00: Op=x, Cmode=001x. 3959 OpCmode = 0x2; 3960 Imm = SplatBits >> 8; 3961 break; 3962 } 3963 if ((SplatBits & ~0xff0000) == 0) { 3964 // Value = 0x00nn0000: Op=x, Cmode=010x. 3965 OpCmode = 0x4; 3966 Imm = SplatBits >> 16; 3967 break; 3968 } 3969 if ((SplatBits & ~0xff000000) == 0) { 3970 // Value = 0xnn000000: Op=x, Cmode=011x. 3971 OpCmode = 0x6; 3972 Imm = SplatBits >> 24; 3973 break; 3974 } 3975 3976 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC 3977 if (type == OtherModImm) return SDValue(); 3978 3979 if ((SplatBits & ~0xffff) == 0 && 3980 ((SplatBits | SplatUndef) & 0xff) == 0xff) { 3981 // Value = 0x0000nnff: Op=x, Cmode=1100. 3982 OpCmode = 0xc; 3983 Imm = SplatBits >> 8; 3984 SplatBits |= 0xff; 3985 break; 3986 } 3987 3988 if ((SplatBits & ~0xffffff) == 0 && 3989 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) { 3990 // Value = 0x00nnffff: Op=x, Cmode=1101. 3991 OpCmode = 0xd; 3992 Imm = SplatBits >> 16; 3993 SplatBits |= 0xffff; 3994 break; 3995 } 3996 3997 // Note: there are a few 32-bit splat values (specifically: 00ffff00, 3998 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not 3999 // VMOV.I32. A (very) minor optimization would be to replicate the value 4000 // and fall through here to test for a valid 64-bit splat. But, then the 4001 // caller would also need to check and handle the change in size. 4002 return SDValue(); 4003 4004 case 64: { 4005 if (type != VMOVModImm) 4006 return SDValue(); 4007 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff. 4008 uint64_t BitMask = 0xff; 4009 uint64_t Val = 0; 4010 unsigned ImmMask = 1; 4011 Imm = 0; 4012 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) { 4013 if (((SplatBits | SplatUndef) & BitMask) == BitMask) { 4014 Val |= BitMask; 4015 Imm |= ImmMask; 4016 } else if ((SplatBits & BitMask) != 0) { 4017 return SDValue(); 4018 } 4019 BitMask <<= 8; 4020 ImmMask <<= 1; 4021 } 4022 // Op=1, Cmode=1110. 4023 OpCmode = 0x1e; 4024 SplatBits = Val; 4025 VT = is128Bits ? MVT::v2i64 : MVT::v1i64; 4026 break; 4027 } 4028 4029 default: 4030 llvm_unreachable("unexpected size for isNEONModifiedImm"); 4031 } 4032 4033 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm); 4034 return DAG.getTargetConstant(EncodedVal, MVT::i32); 4035 } 4036 4037 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG, 4038 const ARMSubtarget *ST) const { 4039 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16()) 4040 return SDValue(); 4041 4042 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op); 4043 assert(Op.getValueType() == MVT::f32 && 4044 "ConstantFP custom lowering should only occur for f32."); 4045 4046 // Try splatting with a VMOV.f32... 4047 APFloat FPVal = CFP->getValueAPF(); 4048 int ImmVal = ARM_AM::getFP32Imm(FPVal); 4049 if (ImmVal != -1) { 4050 DebugLoc DL = Op.getDebugLoc(); 4051 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32); 4052 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32, 4053 NewVal); 4054 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant, 4055 DAG.getConstant(0, MVT::i32)); 4056 } 4057 4058 // If that fails, try a VMOV.i32 4059 EVT VMovVT; 4060 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue(); 4061 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false, 4062 VMOVModImm); 4063 if (NewVal != SDValue()) { 4064 DebugLoc DL = Op.getDebugLoc(); 4065 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT, 4066 NewVal); 4067 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32, 4068 VecConstant); 4069 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant, 4070 DAG.getConstant(0, MVT::i32)); 4071 } 4072 4073 // Finally, try a VMVN.i32 4074 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false, 4075 VMVNModImm); 4076 if (NewVal != SDValue()) { 4077 DebugLoc DL = Op.getDebugLoc(); 4078 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal); 4079 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32, 4080 VecConstant); 4081 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant, 4082 DAG.getConstant(0, MVT::i32)); 4083 } 4084 4085 return SDValue(); 4086 } 4087 4088 // check if an VEXT instruction can handle the shuffle mask when the 4089 // vector sources of the shuffle are the same. 4090 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) { 4091 unsigned NumElts = VT.getVectorNumElements(); 4092 4093 // Assume that the first shuffle index is not UNDEF. Fail if it is. 4094 if (M[0] < 0) 4095 return false; 4096 4097 Imm = M[0]; 4098 4099 // If this is a VEXT shuffle, the immediate value is the index of the first 4100 // element. The other shuffle indices must be the successive elements after 4101 // the first one. 4102 unsigned ExpectedElt = Imm; 4103 for (unsigned i = 1; i < NumElts; ++i) { 4104 // Increment the expected index. If it wraps around, just follow it 4105 // back to index zero and keep going. 4106 ++ExpectedElt; 4107 if (ExpectedElt == NumElts) 4108 ExpectedElt = 0; 4109 4110 if (M[i] < 0) continue; // ignore UNDEF indices 4111 if (ExpectedElt != static_cast<unsigned>(M[i])) 4112 return false; 4113 } 4114 4115 return true; 4116 } 4117 4118 4119 static bool isVEXTMask(ArrayRef<int> M, EVT VT, 4120 bool &ReverseVEXT, unsigned &Imm) { 4121 unsigned NumElts = VT.getVectorNumElements(); 4122 ReverseVEXT = false; 4123 4124 // Assume that the first shuffle index is not UNDEF. Fail if it is. 4125 if (M[0] < 0) 4126 return false; 4127 4128 Imm = M[0]; 4129 4130 // If this is a VEXT shuffle, the immediate value is the index of the first 4131 // element. The other shuffle indices must be the successive elements after 4132 // the first one. 4133 unsigned ExpectedElt = Imm; 4134 for (unsigned i = 1; i < NumElts; ++i) { 4135 // Increment the expected index. If it wraps around, it may still be 4136 // a VEXT but the source vectors must be swapped. 4137 ExpectedElt += 1; 4138 if (ExpectedElt == NumElts * 2) { 4139 ExpectedElt = 0; 4140 ReverseVEXT = true; 4141 } 4142 4143 if (M[i] < 0) continue; // ignore UNDEF indices 4144 if (ExpectedElt != static_cast<unsigned>(M[i])) 4145 return false; 4146 } 4147 4148 // Adjust the index value if the source operands will be swapped. 4149 if (ReverseVEXT) 4150 Imm -= NumElts; 4151 4152 return true; 4153 } 4154 4155 /// isVREVMask - Check if a vector shuffle corresponds to a VREV 4156 /// instruction with the specified blocksize. (The order of the elements 4157 /// within each block of the vector is reversed.) 4158 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) { 4159 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) && 4160 "Only possible block sizes for VREV are: 16, 32, 64"); 4161 4162 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4163 if (EltSz == 64) 4164 return false; 4165 4166 unsigned NumElts = VT.getVectorNumElements(); 4167 unsigned BlockElts = M[0] + 1; 4168 // If the first shuffle index is UNDEF, be optimistic. 4169 if (M[0] < 0) 4170 BlockElts = BlockSize / EltSz; 4171 4172 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz) 4173 return false; 4174 4175 for (unsigned i = 0; i < NumElts; ++i) { 4176 if (M[i] < 0) continue; // ignore UNDEF indices 4177 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts)) 4178 return false; 4179 } 4180 4181 return true; 4182 } 4183 4184 static bool isVTBLMask(ArrayRef<int> M, EVT VT) { 4185 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of 4186 // range, then 0 is placed into the resulting vector. So pretty much any mask 4187 // of 8 elements can work here. 4188 return VT == MVT::v8i8 && M.size() == 8; 4189 } 4190 4191 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 4192 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4193 if (EltSz == 64) 4194 return false; 4195 4196 unsigned NumElts = VT.getVectorNumElements(); 4197 WhichResult = (M[0] == 0 ? 0 : 1); 4198 for (unsigned i = 0; i < NumElts; i += 2) { 4199 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) || 4200 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult)) 4201 return false; 4202 } 4203 return true; 4204 } 4205 4206 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of 4207 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 4208 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>. 4209 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){ 4210 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4211 if (EltSz == 64) 4212 return false; 4213 4214 unsigned NumElts = VT.getVectorNumElements(); 4215 WhichResult = (M[0] == 0 ? 0 : 1); 4216 for (unsigned i = 0; i < NumElts; i += 2) { 4217 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) || 4218 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult)) 4219 return false; 4220 } 4221 return true; 4222 } 4223 4224 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 4225 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4226 if (EltSz == 64) 4227 return false; 4228 4229 unsigned NumElts = VT.getVectorNumElements(); 4230 WhichResult = (M[0] == 0 ? 0 : 1); 4231 for (unsigned i = 0; i != NumElts; ++i) { 4232 if (M[i] < 0) continue; // ignore UNDEF indices 4233 if ((unsigned) M[i] != 2 * i + WhichResult) 4234 return false; 4235 } 4236 4237 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4238 if (VT.is64BitVector() && EltSz == 32) 4239 return false; 4240 4241 return true; 4242 } 4243 4244 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of 4245 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 4246 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>, 4247 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){ 4248 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4249 if (EltSz == 64) 4250 return false; 4251 4252 unsigned Half = VT.getVectorNumElements() / 2; 4253 WhichResult = (M[0] == 0 ? 0 : 1); 4254 for (unsigned j = 0; j != 2; ++j) { 4255 unsigned Idx = WhichResult; 4256 for (unsigned i = 0; i != Half; ++i) { 4257 int MIdx = M[i + j * Half]; 4258 if (MIdx >= 0 && (unsigned) MIdx != Idx) 4259 return false; 4260 Idx += 2; 4261 } 4262 } 4263 4264 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4265 if (VT.is64BitVector() && EltSz == 32) 4266 return false; 4267 4268 return true; 4269 } 4270 4271 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 4272 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4273 if (EltSz == 64) 4274 return false; 4275 4276 unsigned NumElts = VT.getVectorNumElements(); 4277 WhichResult = (M[0] == 0 ? 0 : 1); 4278 unsigned Idx = WhichResult * NumElts / 2; 4279 for (unsigned i = 0; i != NumElts; i += 2) { 4280 if ((M[i] >= 0 && (unsigned) M[i] != Idx) || 4281 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts)) 4282 return false; 4283 Idx += 1; 4284 } 4285 4286 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4287 if (VT.is64BitVector() && EltSz == 32) 4288 return false; 4289 4290 return true; 4291 } 4292 4293 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of 4294 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 4295 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>. 4296 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){ 4297 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4298 if (EltSz == 64) 4299 return false; 4300 4301 unsigned NumElts = VT.getVectorNumElements(); 4302 WhichResult = (M[0] == 0 ? 0 : 1); 4303 unsigned Idx = WhichResult * NumElts / 2; 4304 for (unsigned i = 0; i != NumElts; i += 2) { 4305 if ((M[i] >= 0 && (unsigned) M[i] != Idx) || 4306 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx)) 4307 return false; 4308 Idx += 1; 4309 } 4310 4311 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4312 if (VT.is64BitVector() && EltSz == 32) 4313 return false; 4314 4315 return true; 4316 } 4317 4318 /// \return true if this is a reverse operation on an vector. 4319 static bool isReverseMask(ArrayRef<int> M, EVT VT) { 4320 unsigned NumElts = VT.getVectorNumElements(); 4321 // Make sure the mask has the right size. 4322 if (NumElts != M.size()) 4323 return false; 4324 4325 // Look for <15, ..., 3, -1, 1, 0>. 4326 for (unsigned i = 0; i != NumElts; ++i) 4327 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i)) 4328 return false; 4329 4330 return true; 4331 } 4332 4333 // If N is an integer constant that can be moved into a register in one 4334 // instruction, return an SDValue of such a constant (will become a MOV 4335 // instruction). Otherwise return null. 4336 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG, 4337 const ARMSubtarget *ST, DebugLoc dl) { 4338 uint64_t Val; 4339 if (!isa<ConstantSDNode>(N)) 4340 return SDValue(); 4341 Val = cast<ConstantSDNode>(N)->getZExtValue(); 4342 4343 if (ST->isThumb1Only()) { 4344 if (Val <= 255 || ~Val <= 255) 4345 return DAG.getConstant(Val, MVT::i32); 4346 } else { 4347 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1) 4348 return DAG.getConstant(Val, MVT::i32); 4349 } 4350 return SDValue(); 4351 } 4352 4353 // If this is a case we can't handle, return null and let the default 4354 // expansion code take care of it. 4355 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG, 4356 const ARMSubtarget *ST) const { 4357 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode()); 4358 DebugLoc dl = Op.getDebugLoc(); 4359 EVT VT = Op.getValueType(); 4360 4361 APInt SplatBits, SplatUndef; 4362 unsigned SplatBitSize; 4363 bool HasAnyUndefs; 4364 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) { 4365 if (SplatBitSize <= 64) { 4366 // Check if an immediate VMOV works. 4367 EVT VmovVT; 4368 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(), 4369 SplatUndef.getZExtValue(), SplatBitSize, 4370 DAG, VmovVT, VT.is128BitVector(), 4371 VMOVModImm); 4372 if (Val.getNode()) { 4373 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val); 4374 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov); 4375 } 4376 4377 // Try an immediate VMVN. 4378 uint64_t NegatedImm = (~SplatBits).getZExtValue(); 4379 Val = isNEONModifiedImm(NegatedImm, 4380 SplatUndef.getZExtValue(), SplatBitSize, 4381 DAG, VmovVT, VT.is128BitVector(), 4382 VMVNModImm); 4383 if (Val.getNode()) { 4384 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val); 4385 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov); 4386 } 4387 4388 // Use vmov.f32 to materialize other v2f32 and v4f32 splats. 4389 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) { 4390 int ImmVal = ARM_AM::getFP32Imm(SplatBits); 4391 if (ImmVal != -1) { 4392 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32); 4393 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val); 4394 } 4395 } 4396 } 4397 } 4398 4399 // Scan through the operands to see if only one value is used. 4400 // 4401 // As an optimisation, even if more than one value is used it may be more 4402 // profitable to splat with one value then change some lanes. 4403 // 4404 // Heuristically we decide to do this if the vector has a "dominant" value, 4405 // defined as splatted to more than half of the lanes. 4406 unsigned NumElts = VT.getVectorNumElements(); 4407 bool isOnlyLowElement = true; 4408 bool usesOnlyOneValue = true; 4409 bool hasDominantValue = false; 4410 bool isConstant = true; 4411 4412 // Map of the number of times a particular SDValue appears in the 4413 // element list. 4414 DenseMap<SDValue, unsigned> ValueCounts; 4415 SDValue Value; 4416 for (unsigned i = 0; i < NumElts; ++i) { 4417 SDValue V = Op.getOperand(i); 4418 if (V.getOpcode() == ISD::UNDEF) 4419 continue; 4420 if (i > 0) 4421 isOnlyLowElement = false; 4422 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V)) 4423 isConstant = false; 4424 4425 ValueCounts.insert(std::make_pair(V, 0)); 4426 unsigned &Count = ValueCounts[V]; 4427 4428 // Is this value dominant? (takes up more than half of the lanes) 4429 if (++Count > (NumElts / 2)) { 4430 hasDominantValue = true; 4431 Value = V; 4432 } 4433 } 4434 if (ValueCounts.size() != 1) 4435 usesOnlyOneValue = false; 4436 if (!Value.getNode() && ValueCounts.size() > 0) 4437 Value = ValueCounts.begin()->first; 4438 4439 if (ValueCounts.size() == 0) 4440 return DAG.getUNDEF(VT); 4441 4442 if (isOnlyLowElement) 4443 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value); 4444 4445 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 4446 4447 // Use VDUP for non-constant splats. For f32 constant splats, reduce to 4448 // i32 and try again. 4449 if (hasDominantValue && EltSize <= 32) { 4450 if (!isConstant) { 4451 SDValue N; 4452 4453 // If we are VDUPing a value that comes directly from a vector, that will 4454 // cause an unnecessary move to and from a GPR, where instead we could 4455 // just use VDUPLANE. We can only do this if the lane being extracted 4456 // is at a constant index, as the VDUP from lane instructions only have 4457 // constant-index forms. 4458 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT && 4459 isa<ConstantSDNode>(Value->getOperand(1))) { 4460 // We need to create a new undef vector to use for the VDUPLANE if the 4461 // size of the vector from which we get the value is different than the 4462 // size of the vector that we need to create. We will insert the element 4463 // such that the register coalescer will remove unnecessary copies. 4464 if (VT != Value->getOperand(0).getValueType()) { 4465 ConstantSDNode *constIndex; 4466 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1)); 4467 assert(constIndex && "The index is not a constant!"); 4468 unsigned index = constIndex->getAPIntValue().getLimitedValue() % 4469 VT.getVectorNumElements(); 4470 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT, 4471 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT), 4472 Value, DAG.getConstant(index, MVT::i32)), 4473 DAG.getConstant(index, MVT::i32)); 4474 } else 4475 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT, 4476 Value->getOperand(0), Value->getOperand(1)); 4477 } else 4478 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value); 4479 4480 if (!usesOnlyOneValue) { 4481 // The dominant value was splatted as 'N', but we now have to insert 4482 // all differing elements. 4483 for (unsigned I = 0; I < NumElts; ++I) { 4484 if (Op.getOperand(I) == Value) 4485 continue; 4486 SmallVector<SDValue, 3> Ops; 4487 Ops.push_back(N); 4488 Ops.push_back(Op.getOperand(I)); 4489 Ops.push_back(DAG.getConstant(I, MVT::i32)); 4490 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3); 4491 } 4492 } 4493 return N; 4494 } 4495 if (VT.getVectorElementType().isFloatingPoint()) { 4496 SmallVector<SDValue, 8> Ops; 4497 for (unsigned i = 0; i < NumElts; ++i) 4498 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32, 4499 Op.getOperand(i))); 4500 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts); 4501 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts); 4502 Val = LowerBUILD_VECTOR(Val, DAG, ST); 4503 if (Val.getNode()) 4504 return DAG.getNode(ISD::BITCAST, dl, VT, Val); 4505 } 4506 if (usesOnlyOneValue) { 4507 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl); 4508 if (isConstant && Val.getNode()) 4509 return DAG.getNode(ARMISD::VDUP, dl, VT, Val); 4510 } 4511 } 4512 4513 // If all elements are constants and the case above didn't get hit, fall back 4514 // to the default expansion, which will generate a load from the constant 4515 // pool. 4516 if (isConstant) 4517 return SDValue(); 4518 4519 // Empirical tests suggest this is rarely worth it for vectors of length <= 2. 4520 if (NumElts >= 4) { 4521 SDValue shuffle = ReconstructShuffle(Op, DAG); 4522 if (shuffle != SDValue()) 4523 return shuffle; 4524 } 4525 4526 // Vectors with 32- or 64-bit elements can be built by directly assigning 4527 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands 4528 // will be legalized. 4529 if (EltSize >= 32) { 4530 // Do the expansion with floating-point types, since that is what the VFP 4531 // registers are defined to use, and since i64 is not legal. 4532 EVT EltVT = EVT::getFloatingPointVT(EltSize); 4533 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts); 4534 SmallVector<SDValue, 8> Ops; 4535 for (unsigned i = 0; i < NumElts; ++i) 4536 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i))); 4537 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts); 4538 return DAG.getNode(ISD::BITCAST, dl, VT, Val); 4539 } 4540 4541 return SDValue(); 4542 } 4543 4544 // Gather data to see if the operation can be modelled as a 4545 // shuffle in combination with VEXTs. 4546 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op, 4547 SelectionDAG &DAG) const { 4548 DebugLoc dl = Op.getDebugLoc(); 4549 EVT VT = Op.getValueType(); 4550 unsigned NumElts = VT.getVectorNumElements(); 4551 4552 SmallVector<SDValue, 2> SourceVecs; 4553 SmallVector<unsigned, 2> MinElts; 4554 SmallVector<unsigned, 2> MaxElts; 4555 4556 for (unsigned i = 0; i < NumElts; ++i) { 4557 SDValue V = Op.getOperand(i); 4558 if (V.getOpcode() == ISD::UNDEF) 4559 continue; 4560 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) { 4561 // A shuffle can only come from building a vector from various 4562 // elements of other vectors. 4563 return SDValue(); 4564 } else if (V.getOperand(0).getValueType().getVectorElementType() != 4565 VT.getVectorElementType()) { 4566 // This code doesn't know how to handle shuffles where the vector 4567 // element types do not match (this happens because type legalization 4568 // promotes the return type of EXTRACT_VECTOR_ELT). 4569 // FIXME: It might be appropriate to extend this code to handle 4570 // mismatched types. 4571 return SDValue(); 4572 } 4573 4574 // Record this extraction against the appropriate vector if possible... 4575 SDValue SourceVec = V.getOperand(0); 4576 // If the element number isn't a constant, we can't effectively 4577 // analyze what's going on. 4578 if (!isa<ConstantSDNode>(V.getOperand(1))) 4579 return SDValue(); 4580 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue(); 4581 bool FoundSource = false; 4582 for (unsigned j = 0; j < SourceVecs.size(); ++j) { 4583 if (SourceVecs[j] == SourceVec) { 4584 if (MinElts[j] > EltNo) 4585 MinElts[j] = EltNo; 4586 if (MaxElts[j] < EltNo) 4587 MaxElts[j] = EltNo; 4588 FoundSource = true; 4589 break; 4590 } 4591 } 4592 4593 // Or record a new source if not... 4594 if (!FoundSource) { 4595 SourceVecs.push_back(SourceVec); 4596 MinElts.push_back(EltNo); 4597 MaxElts.push_back(EltNo); 4598 } 4599 } 4600 4601 // Currently only do something sane when at most two source vectors 4602 // involved. 4603 if (SourceVecs.size() > 2) 4604 return SDValue(); 4605 4606 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) }; 4607 int VEXTOffsets[2] = {0, 0}; 4608 4609 // This loop extracts the usage patterns of the source vectors 4610 // and prepares appropriate SDValues for a shuffle if possible. 4611 for (unsigned i = 0; i < SourceVecs.size(); ++i) { 4612 if (SourceVecs[i].getValueType() == VT) { 4613 // No VEXT necessary 4614 ShuffleSrcs[i] = SourceVecs[i]; 4615 VEXTOffsets[i] = 0; 4616 continue; 4617 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) { 4618 // It probably isn't worth padding out a smaller vector just to 4619 // break it down again in a shuffle. 4620 return SDValue(); 4621 } 4622 4623 // Since only 64-bit and 128-bit vectors are legal on ARM and 4624 // we've eliminated the other cases... 4625 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts && 4626 "unexpected vector sizes in ReconstructShuffle"); 4627 4628 if (MaxElts[i] - MinElts[i] >= NumElts) { 4629 // Span too large for a VEXT to cope 4630 return SDValue(); 4631 } 4632 4633 if (MinElts[i] >= NumElts) { 4634 // The extraction can just take the second half 4635 VEXTOffsets[i] = NumElts; 4636 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4637 SourceVecs[i], 4638 DAG.getIntPtrConstant(NumElts)); 4639 } else if (MaxElts[i] < NumElts) { 4640 // The extraction can just take the first half 4641 VEXTOffsets[i] = 0; 4642 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4643 SourceVecs[i], 4644 DAG.getIntPtrConstant(0)); 4645 } else { 4646 // An actual VEXT is needed 4647 VEXTOffsets[i] = MinElts[i]; 4648 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4649 SourceVecs[i], 4650 DAG.getIntPtrConstant(0)); 4651 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4652 SourceVecs[i], 4653 DAG.getIntPtrConstant(NumElts)); 4654 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2, 4655 DAG.getConstant(VEXTOffsets[i], MVT::i32)); 4656 } 4657 } 4658 4659 SmallVector<int, 8> Mask; 4660 4661 for (unsigned i = 0; i < NumElts; ++i) { 4662 SDValue Entry = Op.getOperand(i); 4663 if (Entry.getOpcode() == ISD::UNDEF) { 4664 Mask.push_back(-1); 4665 continue; 4666 } 4667 4668 SDValue ExtractVec = Entry.getOperand(0); 4669 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i) 4670 .getOperand(1))->getSExtValue(); 4671 if (ExtractVec == SourceVecs[0]) { 4672 Mask.push_back(ExtractElt - VEXTOffsets[0]); 4673 } else { 4674 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]); 4675 } 4676 } 4677 4678 // Final check before we try to produce nonsense... 4679 if (isShuffleMaskLegal(Mask, VT)) 4680 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1], 4681 &Mask[0]); 4682 4683 return SDValue(); 4684 } 4685 4686 /// isShuffleMaskLegal - Targets can use this to indicate that they only 4687 /// support *some* VECTOR_SHUFFLE operations, those with specific masks. 4688 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values 4689 /// are assumed to be legal. 4690 bool 4691 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M, 4692 EVT VT) const { 4693 if (VT.getVectorNumElements() == 4 && 4694 (VT.is128BitVector() || VT.is64BitVector())) { 4695 unsigned PFIndexes[4]; 4696 for (unsigned i = 0; i != 4; ++i) { 4697 if (M[i] < 0) 4698 PFIndexes[i] = 8; 4699 else 4700 PFIndexes[i] = M[i]; 4701 } 4702 4703 // Compute the index in the perfect shuffle table. 4704 unsigned PFTableIndex = 4705 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3]; 4706 unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; 4707 unsigned Cost = (PFEntry >> 30); 4708 4709 if (Cost <= 4) 4710 return true; 4711 } 4712 4713 bool ReverseVEXT; 4714 unsigned Imm, WhichResult; 4715 4716 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 4717 return (EltSize >= 32 || 4718 ShuffleVectorSDNode::isSplatMask(&M[0], VT) || 4719 isVREVMask(M, VT, 64) || 4720 isVREVMask(M, VT, 32) || 4721 isVREVMask(M, VT, 16) || 4722 isVEXTMask(M, VT, ReverseVEXT, Imm) || 4723 isVTBLMask(M, VT) || 4724 isVTRNMask(M, VT, WhichResult) || 4725 isVUZPMask(M, VT, WhichResult) || 4726 isVZIPMask(M, VT, WhichResult) || 4727 isVTRN_v_undef_Mask(M, VT, WhichResult) || 4728 isVUZP_v_undef_Mask(M, VT, WhichResult) || 4729 isVZIP_v_undef_Mask(M, VT, WhichResult) || 4730 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT))); 4731 } 4732 4733 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit 4734 /// the specified operations to build the shuffle. 4735 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS, 4736 SDValue RHS, SelectionDAG &DAG, 4737 DebugLoc dl) { 4738 unsigned OpNum = (PFEntry >> 26) & 0x0F; 4739 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1); 4740 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1); 4741 4742 enum { 4743 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3> 4744 OP_VREV, 4745 OP_VDUP0, 4746 OP_VDUP1, 4747 OP_VDUP2, 4748 OP_VDUP3, 4749 OP_VEXT1, 4750 OP_VEXT2, 4751 OP_VEXT3, 4752 OP_VUZPL, // VUZP, left result 4753 OP_VUZPR, // VUZP, right result 4754 OP_VZIPL, // VZIP, left result 4755 OP_VZIPR, // VZIP, right result 4756 OP_VTRNL, // VTRN, left result 4757 OP_VTRNR // VTRN, right result 4758 }; 4759 4760 if (OpNum == OP_COPY) { 4761 if (LHSID == (1*9+2)*9+3) return LHS; 4762 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!"); 4763 return RHS; 4764 } 4765 4766 SDValue OpLHS, OpRHS; 4767 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl); 4768 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl); 4769 EVT VT = OpLHS.getValueType(); 4770 4771 switch (OpNum) { 4772 default: llvm_unreachable("Unknown shuffle opcode!"); 4773 case OP_VREV: 4774 // VREV divides the vector in half and swaps within the half. 4775 if (VT.getVectorElementType() == MVT::i32 || 4776 VT.getVectorElementType() == MVT::f32) 4777 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS); 4778 // vrev <4 x i16> -> VREV32 4779 if (VT.getVectorElementType() == MVT::i16) 4780 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS); 4781 // vrev <4 x i8> -> VREV16 4782 assert(VT.getVectorElementType() == MVT::i8); 4783 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS); 4784 case OP_VDUP0: 4785 case OP_VDUP1: 4786 case OP_VDUP2: 4787 case OP_VDUP3: 4788 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, 4789 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32)); 4790 case OP_VEXT1: 4791 case OP_VEXT2: 4792 case OP_VEXT3: 4793 return DAG.getNode(ARMISD::VEXT, dl, VT, 4794 OpLHS, OpRHS, 4795 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32)); 4796 case OP_VUZPL: 4797 case OP_VUZPR: 4798 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 4799 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL); 4800 case OP_VZIPL: 4801 case OP_VZIPR: 4802 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 4803 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL); 4804 case OP_VTRNL: 4805 case OP_VTRNR: 4806 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 4807 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL); 4808 } 4809 } 4810 4811 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op, 4812 ArrayRef<int> ShuffleMask, 4813 SelectionDAG &DAG) { 4814 // Check to see if we can use the VTBL instruction. 4815 SDValue V1 = Op.getOperand(0); 4816 SDValue V2 = Op.getOperand(1); 4817 DebugLoc DL = Op.getDebugLoc(); 4818 4819 SmallVector<SDValue, 8> VTBLMask; 4820 for (ArrayRef<int>::iterator 4821 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I) 4822 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32)); 4823 4824 if (V2.getNode()->getOpcode() == ISD::UNDEF) 4825 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1, 4826 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, 4827 &VTBLMask[0], 8)); 4828 4829 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2, 4830 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, 4831 &VTBLMask[0], 8)); 4832 } 4833 4834 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op, 4835 SelectionDAG &DAG) { 4836 DebugLoc DL = Op.getDebugLoc(); 4837 SDValue OpLHS = Op.getOperand(0); 4838 EVT VT = OpLHS.getValueType(); 4839 4840 assert((VT == MVT::v8i16 || VT == MVT::v16i8) && 4841 "Expect an v8i16/v16i8 type"); 4842 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS); 4843 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now, 4844 // extract the first 8 bytes into the top double word and the last 8 bytes 4845 // into the bottom double word. The v8i16 case is similar. 4846 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4; 4847 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS, 4848 DAG.getConstant(ExtractNum, MVT::i32)); 4849 } 4850 4851 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) { 4852 SDValue V1 = Op.getOperand(0); 4853 SDValue V2 = Op.getOperand(1); 4854 DebugLoc dl = Op.getDebugLoc(); 4855 EVT VT = Op.getValueType(); 4856 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode()); 4857 4858 // Convert shuffles that are directly supported on NEON to target-specific 4859 // DAG nodes, instead of keeping them as shuffles and matching them again 4860 // during code selection. This is more efficient and avoids the possibility 4861 // of inconsistencies between legalization and selection. 4862 // FIXME: floating-point vectors should be canonicalized to integer vectors 4863 // of the same time so that they get CSEd properly. 4864 ArrayRef<int> ShuffleMask = SVN->getMask(); 4865 4866 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 4867 if (EltSize <= 32) { 4868 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) { 4869 int Lane = SVN->getSplatIndex(); 4870 // If this is undef splat, generate it via "just" vdup, if possible. 4871 if (Lane == -1) Lane = 0; 4872 4873 // Test if V1 is a SCALAR_TO_VECTOR. 4874 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) { 4875 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0)); 4876 } 4877 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR 4878 // (and probably will turn into a SCALAR_TO_VECTOR once legalization 4879 // reaches it). 4880 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR && 4881 !isa<ConstantSDNode>(V1.getOperand(0))) { 4882 bool IsScalarToVector = true; 4883 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i) 4884 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) { 4885 IsScalarToVector = false; 4886 break; 4887 } 4888 if (IsScalarToVector) 4889 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0)); 4890 } 4891 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1, 4892 DAG.getConstant(Lane, MVT::i32)); 4893 } 4894 4895 bool ReverseVEXT; 4896 unsigned Imm; 4897 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) { 4898 if (ReverseVEXT) 4899 std::swap(V1, V2); 4900 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2, 4901 DAG.getConstant(Imm, MVT::i32)); 4902 } 4903 4904 if (isVREVMask(ShuffleMask, VT, 64)) 4905 return DAG.getNode(ARMISD::VREV64, dl, VT, V1); 4906 if (isVREVMask(ShuffleMask, VT, 32)) 4907 return DAG.getNode(ARMISD::VREV32, dl, VT, V1); 4908 if (isVREVMask(ShuffleMask, VT, 16)) 4909 return DAG.getNode(ARMISD::VREV16, dl, VT, V1); 4910 4911 if (V2->getOpcode() == ISD::UNDEF && 4912 isSingletonVEXTMask(ShuffleMask, VT, Imm)) { 4913 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1, 4914 DAG.getConstant(Imm, MVT::i32)); 4915 } 4916 4917 // Check for Neon shuffles that modify both input vectors in place. 4918 // If both results are used, i.e., if there are two shuffles with the same 4919 // source operands and with masks corresponding to both results of one of 4920 // these operations, DAG memoization will ensure that a single node is 4921 // used for both shuffles. 4922 unsigned WhichResult; 4923 if (isVTRNMask(ShuffleMask, VT, WhichResult)) 4924 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 4925 V1, V2).getValue(WhichResult); 4926 if (isVUZPMask(ShuffleMask, VT, WhichResult)) 4927 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 4928 V1, V2).getValue(WhichResult); 4929 if (isVZIPMask(ShuffleMask, VT, WhichResult)) 4930 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 4931 V1, V2).getValue(WhichResult); 4932 4933 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) 4934 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 4935 V1, V1).getValue(WhichResult); 4936 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) 4937 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 4938 V1, V1).getValue(WhichResult); 4939 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) 4940 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 4941 V1, V1).getValue(WhichResult); 4942 } 4943 4944 // If the shuffle is not directly supported and it has 4 elements, use 4945 // the PerfectShuffle-generated table to synthesize it from other shuffles. 4946 unsigned NumElts = VT.getVectorNumElements(); 4947 if (NumElts == 4) { 4948 unsigned PFIndexes[4]; 4949 for (unsigned i = 0; i != 4; ++i) { 4950 if (ShuffleMask[i] < 0) 4951 PFIndexes[i] = 8; 4952 else 4953 PFIndexes[i] = ShuffleMask[i]; 4954 } 4955 4956 // Compute the index in the perfect shuffle table. 4957 unsigned PFTableIndex = 4958 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3]; 4959 unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; 4960 unsigned Cost = (PFEntry >> 30); 4961 4962 if (Cost <= 4) 4963 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl); 4964 } 4965 4966 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs. 4967 if (EltSize >= 32) { 4968 // Do the expansion with floating-point types, since that is what the VFP 4969 // registers are defined to use, and since i64 is not legal. 4970 EVT EltVT = EVT::getFloatingPointVT(EltSize); 4971 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts); 4972 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1); 4973 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2); 4974 SmallVector<SDValue, 8> Ops; 4975 for (unsigned i = 0; i < NumElts; ++i) { 4976 if (ShuffleMask[i] < 0) 4977 Ops.push_back(DAG.getUNDEF(EltVT)); 4978 else 4979 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, 4980 ShuffleMask[i] < (int)NumElts ? V1 : V2, 4981 DAG.getConstant(ShuffleMask[i] & (NumElts-1), 4982 MVT::i32))); 4983 } 4984 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts); 4985 return DAG.getNode(ISD::BITCAST, dl, VT, Val); 4986 } 4987 4988 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT)) 4989 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG); 4990 4991 if (VT == MVT::v8i8) { 4992 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG); 4993 if (NewOp.getNode()) 4994 return NewOp; 4995 } 4996 4997 return SDValue(); 4998 } 4999 5000 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) { 5001 // INSERT_VECTOR_ELT is legal only for immediate indexes. 5002 SDValue Lane = Op.getOperand(2); 5003 if (!isa<ConstantSDNode>(Lane)) 5004 return SDValue(); 5005 5006 return Op; 5007 } 5008 5009 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) { 5010 // EXTRACT_VECTOR_ELT is legal only for immediate indexes. 5011 SDValue Lane = Op.getOperand(1); 5012 if (!isa<ConstantSDNode>(Lane)) 5013 return SDValue(); 5014 5015 SDValue Vec = Op.getOperand(0); 5016 if (Op.getValueType() == MVT::i32 && 5017 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) { 5018 DebugLoc dl = Op.getDebugLoc(); 5019 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane); 5020 } 5021 5022 return Op; 5023 } 5024 5025 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) { 5026 // The only time a CONCAT_VECTORS operation can have legal types is when 5027 // two 64-bit vectors are concatenated to a 128-bit vector. 5028 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 && 5029 "unexpected CONCAT_VECTORS"); 5030 DebugLoc dl = Op.getDebugLoc(); 5031 SDValue Val = DAG.getUNDEF(MVT::v2f64); 5032 SDValue Op0 = Op.getOperand(0); 5033 SDValue Op1 = Op.getOperand(1); 5034 if (Op0.getOpcode() != ISD::UNDEF) 5035 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val, 5036 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0), 5037 DAG.getIntPtrConstant(0)); 5038 if (Op1.getOpcode() != ISD::UNDEF) 5039 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val, 5040 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1), 5041 DAG.getIntPtrConstant(1)); 5042 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val); 5043 } 5044 5045 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each 5046 /// element has been zero/sign-extended, depending on the isSigned parameter, 5047 /// from an integer type half its size. 5048 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG, 5049 bool isSigned) { 5050 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32. 5051 EVT VT = N->getValueType(0); 5052 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) { 5053 SDNode *BVN = N->getOperand(0).getNode(); 5054 if (BVN->getValueType(0) != MVT::v4i32 || 5055 BVN->getOpcode() != ISD::BUILD_VECTOR) 5056 return false; 5057 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0; 5058 unsigned HiElt = 1 - LoElt; 5059 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt)); 5060 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt)); 5061 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2)); 5062 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2)); 5063 if (!Lo0 || !Hi0 || !Lo1 || !Hi1) 5064 return false; 5065 if (isSigned) { 5066 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 && 5067 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32) 5068 return true; 5069 } else { 5070 if (Hi0->isNullValue() && Hi1->isNullValue()) 5071 return true; 5072 } 5073 return false; 5074 } 5075 5076 if (N->getOpcode() != ISD::BUILD_VECTOR) 5077 return false; 5078 5079 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 5080 SDNode *Elt = N->getOperand(i).getNode(); 5081 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) { 5082 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 5083 unsigned HalfSize = EltSize / 2; 5084 if (isSigned) { 5085 if (!isIntN(HalfSize, C->getSExtValue())) 5086 return false; 5087 } else { 5088 if (!isUIntN(HalfSize, C->getZExtValue())) 5089 return false; 5090 } 5091 continue; 5092 } 5093 return false; 5094 } 5095 5096 return true; 5097 } 5098 5099 /// isSignExtended - Check if a node is a vector value that is sign-extended 5100 /// or a constant BUILD_VECTOR with sign-extended elements. 5101 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) { 5102 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N)) 5103 return true; 5104 if (isExtendedBUILD_VECTOR(N, DAG, true)) 5105 return true; 5106 return false; 5107 } 5108 5109 /// isZeroExtended - Check if a node is a vector value that is zero-extended 5110 /// or a constant BUILD_VECTOR with zero-extended elements. 5111 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) { 5112 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N)) 5113 return true; 5114 if (isExtendedBUILD_VECTOR(N, DAG, false)) 5115 return true; 5116 return false; 5117 } 5118 5119 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total 5120 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL. 5121 /// We insert the required extension here to get the vector to fill a D register. 5122 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG, 5123 const EVT &OrigTy, 5124 const EVT &ExtTy, 5125 unsigned ExtOpcode) { 5126 // The vector originally had a size of OrigTy. It was then extended to ExtTy. 5127 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than 5128 // 64-bits we need to insert a new extension so that it will be 64-bits. 5129 assert(ExtTy.is128BitVector() && "Unexpected extension size"); 5130 if (OrigTy.getSizeInBits() >= 64) 5131 return N; 5132 5133 // Must extend size to at least 64 bits to be used as an operand for VMULL. 5134 MVT::SimpleValueType OrigSimpleTy = OrigTy.getSimpleVT().SimpleTy; 5135 EVT NewVT; 5136 switch (OrigSimpleTy) { 5137 default: llvm_unreachable("Unexpected Orig Vector Type"); 5138 case MVT::v2i8: 5139 case MVT::v2i16: 5140 NewVT = MVT::v2i32; 5141 break; 5142 case MVT::v4i8: 5143 NewVT = MVT::v4i16; 5144 break; 5145 } 5146 return DAG.getNode(ExtOpcode, N->getDebugLoc(), NewVT, N); 5147 } 5148 5149 /// SkipLoadExtensionForVMULL - return a load of the original vector size that 5150 /// does not do any sign/zero extension. If the original vector is less 5151 /// than 64 bits, an appropriate extension will be added after the load to 5152 /// reach a total size of 64 bits. We have to add the extension separately 5153 /// because ARM does not have a sign/zero extending load for vectors. 5154 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) { 5155 SDValue NonExtendingLoad = 5156 DAG.getLoad(LD->getMemoryVT(), LD->getDebugLoc(), LD->getChain(), 5157 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(), 5158 LD->isNonTemporal(), LD->isInvariant(), 5159 LD->getAlignment()); 5160 unsigned ExtOp = 0; 5161 switch (LD->getExtensionType()) { 5162 default: llvm_unreachable("Unexpected LoadExtType"); 5163 case ISD::EXTLOAD: 5164 case ISD::SEXTLOAD: ExtOp = ISD::SIGN_EXTEND; break; 5165 case ISD::ZEXTLOAD: ExtOp = ISD::ZERO_EXTEND; break; 5166 } 5167 MVT::SimpleValueType MemType = LD->getMemoryVT().getSimpleVT().SimpleTy; 5168 MVT::SimpleValueType ExtType = LD->getValueType(0).getSimpleVT().SimpleTy; 5169 return AddRequiredExtensionForVMULL(NonExtendingLoad, DAG, 5170 MemType, ExtType, ExtOp); 5171 } 5172 5173 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND, 5174 /// extending load, or BUILD_VECTOR with extended elements, return the 5175 /// unextended value. The unextended vector should be 64 bits so that it can 5176 /// be used as an operand to a VMULL instruction. If the original vector size 5177 /// before extension is less than 64 bits we add a an extension to resize 5178 /// the vector to 64 bits. 5179 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) { 5180 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND) 5181 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG, 5182 N->getOperand(0)->getValueType(0), 5183 N->getValueType(0), 5184 N->getOpcode()); 5185 5186 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) 5187 return SkipLoadExtensionForVMULL(LD, DAG); 5188 5189 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will 5190 // have been legalized as a BITCAST from v4i32. 5191 if (N->getOpcode() == ISD::BITCAST) { 5192 SDNode *BVN = N->getOperand(0).getNode(); 5193 assert(BVN->getOpcode() == ISD::BUILD_VECTOR && 5194 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR"); 5195 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0; 5196 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32, 5197 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2)); 5198 } 5199 // Construct a new BUILD_VECTOR with elements truncated to half the size. 5200 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR"); 5201 EVT VT = N->getValueType(0); 5202 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2; 5203 unsigned NumElts = VT.getVectorNumElements(); 5204 MVT TruncVT = MVT::getIntegerVT(EltSize); 5205 SmallVector<SDValue, 8> Ops; 5206 for (unsigned i = 0; i != NumElts; ++i) { 5207 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i)); 5208 const APInt &CInt = C->getAPIntValue(); 5209 // Element types smaller than 32 bits are not legal, so use i32 elements. 5210 // The values are implicitly truncated so sext vs. zext doesn't matter. 5211 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32)); 5212 } 5213 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), 5214 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts); 5215 } 5216 5217 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) { 5218 unsigned Opcode = N->getOpcode(); 5219 if (Opcode == ISD::ADD || Opcode == ISD::SUB) { 5220 SDNode *N0 = N->getOperand(0).getNode(); 5221 SDNode *N1 = N->getOperand(1).getNode(); 5222 return N0->hasOneUse() && N1->hasOneUse() && 5223 isSignExtended(N0, DAG) && isSignExtended(N1, DAG); 5224 } 5225 return false; 5226 } 5227 5228 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) { 5229 unsigned Opcode = N->getOpcode(); 5230 if (Opcode == ISD::ADD || Opcode == ISD::SUB) { 5231 SDNode *N0 = N->getOperand(0).getNode(); 5232 SDNode *N1 = N->getOperand(1).getNode(); 5233 return N0->hasOneUse() && N1->hasOneUse() && 5234 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG); 5235 } 5236 return false; 5237 } 5238 5239 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) { 5240 // Multiplications are only custom-lowered for 128-bit vectors so that 5241 // VMULL can be detected. Otherwise v2i64 multiplications are not legal. 5242 EVT VT = Op.getValueType(); 5243 assert(VT.is128BitVector() && VT.isInteger() && 5244 "unexpected type for custom-lowering ISD::MUL"); 5245 SDNode *N0 = Op.getOperand(0).getNode(); 5246 SDNode *N1 = Op.getOperand(1).getNode(); 5247 unsigned NewOpc = 0; 5248 bool isMLA = false; 5249 bool isN0SExt = isSignExtended(N0, DAG); 5250 bool isN1SExt = isSignExtended(N1, DAG); 5251 if (isN0SExt && isN1SExt) 5252 NewOpc = ARMISD::VMULLs; 5253 else { 5254 bool isN0ZExt = isZeroExtended(N0, DAG); 5255 bool isN1ZExt = isZeroExtended(N1, DAG); 5256 if (isN0ZExt && isN1ZExt) 5257 NewOpc = ARMISD::VMULLu; 5258 else if (isN1SExt || isN1ZExt) { 5259 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these 5260 // into (s/zext A * s/zext C) + (s/zext B * s/zext C) 5261 if (isN1SExt && isAddSubSExt(N0, DAG)) { 5262 NewOpc = ARMISD::VMULLs; 5263 isMLA = true; 5264 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) { 5265 NewOpc = ARMISD::VMULLu; 5266 isMLA = true; 5267 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) { 5268 std::swap(N0, N1); 5269 NewOpc = ARMISD::VMULLu; 5270 isMLA = true; 5271 } 5272 } 5273 5274 if (!NewOpc) { 5275 if (VT == MVT::v2i64) 5276 // Fall through to expand this. It is not legal. 5277 return SDValue(); 5278 else 5279 // Other vector multiplications are legal. 5280 return Op; 5281 } 5282 } 5283 5284 // Legalize to a VMULL instruction. 5285 DebugLoc DL = Op.getDebugLoc(); 5286 SDValue Op0; 5287 SDValue Op1 = SkipExtensionForVMULL(N1, DAG); 5288 if (!isMLA) { 5289 Op0 = SkipExtensionForVMULL(N0, DAG); 5290 assert(Op0.getValueType().is64BitVector() && 5291 Op1.getValueType().is64BitVector() && 5292 "unexpected types for extended operands to VMULL"); 5293 return DAG.getNode(NewOpc, DL, VT, Op0, Op1); 5294 } 5295 5296 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during 5297 // isel lowering to take advantage of no-stall back to back vmul + vmla. 5298 // vmull q0, d4, d6 5299 // vmlal q0, d5, d6 5300 // is faster than 5301 // vaddl q0, d4, d5 5302 // vmovl q1, d6 5303 // vmul q0, q0, q1 5304 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG); 5305 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG); 5306 EVT Op1VT = Op1.getValueType(); 5307 return DAG.getNode(N0->getOpcode(), DL, VT, 5308 DAG.getNode(NewOpc, DL, VT, 5309 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1), 5310 DAG.getNode(NewOpc, DL, VT, 5311 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1)); 5312 } 5313 5314 static SDValue 5315 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) { 5316 // Convert to float 5317 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo)); 5318 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo)); 5319 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X); 5320 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y); 5321 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X); 5322 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y); 5323 // Get reciprocal estimate. 5324 // float4 recip = vrecpeq_f32(yf); 5325 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5326 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y); 5327 // Because char has a smaller range than uchar, we can actually get away 5328 // without any newton steps. This requires that we use a weird bias 5329 // of 0xb000, however (again, this has been exhaustively tested). 5330 // float4 result = as_float4(as_int4(xf*recip) + 0xb000); 5331 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y); 5332 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X); 5333 Y = DAG.getConstant(0xb000, MVT::i32); 5334 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y); 5335 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y); 5336 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X); 5337 // Convert back to short. 5338 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X); 5339 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X); 5340 return X; 5341 } 5342 5343 static SDValue 5344 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) { 5345 SDValue N2; 5346 // Convert to float. 5347 // float4 yf = vcvt_f32_s32(vmovl_s16(y)); 5348 // float4 xf = vcvt_f32_s32(vmovl_s16(x)); 5349 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0); 5350 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1); 5351 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0); 5352 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1); 5353 5354 // Use reciprocal estimate and one refinement step. 5355 // float4 recip = vrecpeq_f32(yf); 5356 // recip *= vrecpsq_f32(yf, recip); 5357 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5358 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1); 5359 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5360 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), 5361 N1, N2); 5362 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); 5363 // Because short has a smaller range than ushort, we can actually get away 5364 // with only a single newton step. This requires that we use a weird bias 5365 // of 89, however (again, this has been exhaustively tested). 5366 // float4 result = as_float4(as_int4(xf*recip) + 0x89); 5367 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2); 5368 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0); 5369 N1 = DAG.getConstant(0x89, MVT::i32); 5370 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1); 5371 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1); 5372 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0); 5373 // Convert back to integer and return. 5374 // return vmovn_s32(vcvt_s32_f32(result)); 5375 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0); 5376 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0); 5377 return N0; 5378 } 5379 5380 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) { 5381 EVT VT = Op.getValueType(); 5382 assert((VT == MVT::v4i16 || VT == MVT::v8i8) && 5383 "unexpected type for custom-lowering ISD::SDIV"); 5384 5385 DebugLoc dl = Op.getDebugLoc(); 5386 SDValue N0 = Op.getOperand(0); 5387 SDValue N1 = Op.getOperand(1); 5388 SDValue N2, N3; 5389 5390 if (VT == MVT::v8i8) { 5391 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0); 5392 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1); 5393 5394 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5395 DAG.getIntPtrConstant(4)); 5396 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5397 DAG.getIntPtrConstant(4)); 5398 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5399 DAG.getIntPtrConstant(0)); 5400 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5401 DAG.getIntPtrConstant(0)); 5402 5403 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16 5404 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16 5405 5406 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2); 5407 N0 = LowerCONCAT_VECTORS(N0, DAG); 5408 5409 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0); 5410 return N0; 5411 } 5412 return LowerSDIV_v4i16(N0, N1, dl, DAG); 5413 } 5414 5415 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) { 5416 EVT VT = Op.getValueType(); 5417 assert((VT == MVT::v4i16 || VT == MVT::v8i8) && 5418 "unexpected type for custom-lowering ISD::UDIV"); 5419 5420 DebugLoc dl = Op.getDebugLoc(); 5421 SDValue N0 = Op.getOperand(0); 5422 SDValue N1 = Op.getOperand(1); 5423 SDValue N2, N3; 5424 5425 if (VT == MVT::v8i8) { 5426 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0); 5427 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1); 5428 5429 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5430 DAG.getIntPtrConstant(4)); 5431 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5432 DAG.getIntPtrConstant(4)); 5433 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5434 DAG.getIntPtrConstant(0)); 5435 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5436 DAG.getIntPtrConstant(0)); 5437 5438 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16 5439 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16 5440 5441 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2); 5442 N0 = LowerCONCAT_VECTORS(N0, DAG); 5443 5444 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8, 5445 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32), 5446 N0); 5447 return N0; 5448 } 5449 5450 // v4i16 sdiv ... Convert to float. 5451 // float4 yf = vcvt_f32_s32(vmovl_u16(y)); 5452 // float4 xf = vcvt_f32_s32(vmovl_u16(x)); 5453 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0); 5454 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1); 5455 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0); 5456 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1); 5457 5458 // Use reciprocal estimate and two refinement steps. 5459 // float4 recip = vrecpeq_f32(yf); 5460 // recip *= vrecpsq_f32(yf, recip); 5461 // recip *= vrecpsq_f32(yf, recip); 5462 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5463 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1); 5464 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5465 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), 5466 BN1, N2); 5467 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); 5468 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5469 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), 5470 BN1, N2); 5471 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); 5472 // Simply multiplying by the reciprocal estimate can leave us a few ulps 5473 // too low, so we add 2 ulps (exhaustive testing shows that this is enough, 5474 // and that it will never cause us to return an answer too large). 5475 // float4 result = as_float4(as_int4(xf*recip) + 2); 5476 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2); 5477 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0); 5478 N1 = DAG.getConstant(2, MVT::i32); 5479 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1); 5480 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1); 5481 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0); 5482 // Convert back to integer and return. 5483 // return vmovn_u32(vcvt_s32_f32(result)); 5484 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0); 5485 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0); 5486 return N0; 5487 } 5488 5489 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) { 5490 EVT VT = Op.getNode()->getValueType(0); 5491 SDVTList VTs = DAG.getVTList(VT, MVT::i32); 5492 5493 unsigned Opc; 5494 bool ExtraOp = false; 5495 switch (Op.getOpcode()) { 5496 default: llvm_unreachable("Invalid code"); 5497 case ISD::ADDC: Opc = ARMISD::ADDC; break; 5498 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break; 5499 case ISD::SUBC: Opc = ARMISD::SUBC; break; 5500 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break; 5501 } 5502 5503 if (!ExtraOp) 5504 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0), 5505 Op.getOperand(1)); 5506 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0), 5507 Op.getOperand(1), Op.getOperand(2)); 5508 } 5509 5510 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) { 5511 // Monotonic load/store is legal for all targets 5512 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic) 5513 return Op; 5514 5515 // Aquire/Release load/store is not legal for targets without a 5516 // dmb or equivalent available. 5517 return SDValue(); 5518 } 5519 5520 5521 static void 5522 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results, 5523 SelectionDAG &DAG, unsigned NewOp) { 5524 DebugLoc dl = Node->getDebugLoc(); 5525 assert (Node->getValueType(0) == MVT::i64 && 5526 "Only know how to expand i64 atomics"); 5527 5528 SmallVector<SDValue, 6> Ops; 5529 Ops.push_back(Node->getOperand(0)); // Chain 5530 Ops.push_back(Node->getOperand(1)); // Ptr 5531 // Low part of Val1 5532 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5533 Node->getOperand(2), DAG.getIntPtrConstant(0))); 5534 // High part of Val1 5535 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5536 Node->getOperand(2), DAG.getIntPtrConstant(1))); 5537 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) { 5538 // High part of Val1 5539 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5540 Node->getOperand(3), DAG.getIntPtrConstant(0))); 5541 // High part of Val2 5542 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5543 Node->getOperand(3), DAG.getIntPtrConstant(1))); 5544 } 5545 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other); 5546 SDValue Result = 5547 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64, 5548 cast<MemSDNode>(Node)->getMemOperand()); 5549 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) }; 5550 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2)); 5551 Results.push_back(Result.getValue(2)); 5552 } 5553 5554 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { 5555 switch (Op.getOpcode()) { 5556 default: llvm_unreachable("Don't know how to custom lower this!"); 5557 case ISD::ConstantPool: return LowerConstantPool(Op, DAG); 5558 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); 5559 case ISD::GlobalAddress: 5560 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) : 5561 LowerGlobalAddressELF(Op, DAG); 5562 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); 5563 case ISD::SELECT: return LowerSELECT(Op, DAG); 5564 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG); 5565 case ISD::BR_CC: return LowerBR_CC(Op, DAG); 5566 case ISD::BR_JT: return LowerBR_JT(Op, DAG); 5567 case ISD::VASTART: return LowerVASTART(Op, DAG); 5568 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget); 5569 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget); 5570 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget); 5571 case ISD::SINT_TO_FP: 5572 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG); 5573 case ISD::FP_TO_SINT: 5574 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG); 5575 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG); 5576 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG); 5577 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); 5578 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG); 5579 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG); 5580 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG); 5581 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG, 5582 Subtarget); 5583 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG); 5584 case ISD::SHL: 5585 case ISD::SRL: 5586 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget); 5587 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG); 5588 case ISD::SRL_PARTS: 5589 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG); 5590 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget); 5591 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget); 5592 case ISD::SETCC: return LowerVSETCC(Op, DAG); 5593 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget); 5594 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget); 5595 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG); 5596 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG); 5597 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG); 5598 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG); 5599 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG); 5600 case ISD::MUL: return LowerMUL(Op, DAG); 5601 case ISD::SDIV: return LowerSDIV(Op, DAG); 5602 case ISD::UDIV: return LowerUDIV(Op, DAG); 5603 case ISD::ADDC: 5604 case ISD::ADDE: 5605 case ISD::SUBC: 5606 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG); 5607 case ISD::ATOMIC_LOAD: 5608 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG); 5609 } 5610 } 5611 5612 /// ReplaceNodeResults - Replace the results of node with an illegal result 5613 /// type with new values built out of custom code. 5614 void ARMTargetLowering::ReplaceNodeResults(SDNode *N, 5615 SmallVectorImpl<SDValue>&Results, 5616 SelectionDAG &DAG) const { 5617 SDValue Res; 5618 switch (N->getOpcode()) { 5619 default: 5620 llvm_unreachable("Don't know how to custom expand this!"); 5621 case ISD::BITCAST: 5622 Res = ExpandBITCAST(N, DAG); 5623 break; 5624 case ISD::SRL: 5625 case ISD::SRA: 5626 Res = Expand64BitShift(N, DAG, Subtarget); 5627 break; 5628 case ISD::ATOMIC_LOAD_ADD: 5629 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG); 5630 return; 5631 case ISD::ATOMIC_LOAD_AND: 5632 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG); 5633 return; 5634 case ISD::ATOMIC_LOAD_NAND: 5635 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG); 5636 return; 5637 case ISD::ATOMIC_LOAD_OR: 5638 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG); 5639 return; 5640 case ISD::ATOMIC_LOAD_SUB: 5641 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG); 5642 return; 5643 case ISD::ATOMIC_LOAD_XOR: 5644 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG); 5645 return; 5646 case ISD::ATOMIC_SWAP: 5647 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG); 5648 return; 5649 case ISD::ATOMIC_CMP_SWAP: 5650 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG); 5651 return; 5652 case ISD::ATOMIC_LOAD_MIN: 5653 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMIN64_DAG); 5654 return; 5655 case ISD::ATOMIC_LOAD_UMIN: 5656 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMIN64_DAG); 5657 return; 5658 case ISD::ATOMIC_LOAD_MAX: 5659 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMAX64_DAG); 5660 return; 5661 case ISD::ATOMIC_LOAD_UMAX: 5662 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMAX64_DAG); 5663 return; 5664 } 5665 if (Res.getNode()) 5666 Results.push_back(Res); 5667 } 5668 5669 //===----------------------------------------------------------------------===// 5670 // ARM Scheduler Hooks 5671 //===----------------------------------------------------------------------===// 5672 5673 MachineBasicBlock * 5674 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI, 5675 MachineBasicBlock *BB, 5676 unsigned Size) const { 5677 unsigned dest = MI->getOperand(0).getReg(); 5678 unsigned ptr = MI->getOperand(1).getReg(); 5679 unsigned oldval = MI->getOperand(2).getReg(); 5680 unsigned newval = MI->getOperand(3).getReg(); 5681 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5682 DebugLoc dl = MI->getDebugLoc(); 5683 bool isThumb2 = Subtarget->isThumb2(); 5684 5685 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 5686 unsigned scratch = MRI.createVirtualRegister(isThumb2 ? 5687 (const TargetRegisterClass*)&ARM::rGPRRegClass : 5688 (const TargetRegisterClass*)&ARM::GPRRegClass); 5689 5690 if (isThumb2) { 5691 MRI.constrainRegClass(dest, &ARM::rGPRRegClass); 5692 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass); 5693 MRI.constrainRegClass(newval, &ARM::rGPRRegClass); 5694 } 5695 5696 unsigned ldrOpc, strOpc; 5697 switch (Size) { 5698 default: llvm_unreachable("unsupported size for AtomicCmpSwap!"); 5699 case 1: 5700 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB; 5701 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB; 5702 break; 5703 case 2: 5704 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH; 5705 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH; 5706 break; 5707 case 4: 5708 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX; 5709 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX; 5710 break; 5711 } 5712 5713 MachineFunction *MF = BB->getParent(); 5714 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 5715 MachineFunction::iterator It = BB; 5716 ++It; // insert the new blocks after the current block 5717 5718 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 5719 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 5720 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5721 MF->insert(It, loop1MBB); 5722 MF->insert(It, loop2MBB); 5723 MF->insert(It, exitMBB); 5724 5725 // Transfer the remainder of BB and its successor edges to exitMBB. 5726 exitMBB->splice(exitMBB->begin(), BB, 5727 llvm::next(MachineBasicBlock::iterator(MI)), 5728 BB->end()); 5729 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 5730 5731 // thisMBB: 5732 // ... 5733 // fallthrough --> loop1MBB 5734 BB->addSuccessor(loop1MBB); 5735 5736 // loop1MBB: 5737 // ldrex dest, [ptr] 5738 // cmp dest, oldval 5739 // bne exitMBB 5740 BB = loop1MBB; 5741 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr); 5742 if (ldrOpc == ARM::t2LDREX) 5743 MIB.addImm(0); 5744 AddDefaultPred(MIB); 5745 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) 5746 .addReg(dest).addReg(oldval)); 5747 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5748 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5749 BB->addSuccessor(loop2MBB); 5750 BB->addSuccessor(exitMBB); 5751 5752 // loop2MBB: 5753 // strex scratch, newval, [ptr] 5754 // cmp scratch, #0 5755 // bne loop1MBB 5756 BB = loop2MBB; 5757 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr); 5758 if (strOpc == ARM::t2STREX) 5759 MIB.addImm(0); 5760 AddDefaultPred(MIB); 5761 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 5762 .addReg(scratch).addImm(0)); 5763 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5764 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5765 BB->addSuccessor(loop1MBB); 5766 BB->addSuccessor(exitMBB); 5767 5768 // exitMBB: 5769 // ... 5770 BB = exitMBB; 5771 5772 MI->eraseFromParent(); // The instruction is gone now. 5773 5774 return BB; 5775 } 5776 5777 MachineBasicBlock * 5778 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB, 5779 unsigned Size, unsigned BinOpcode) const { 5780 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0. 5781 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5782 5783 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 5784 MachineFunction *MF = BB->getParent(); 5785 MachineFunction::iterator It = BB; 5786 ++It; 5787 5788 unsigned dest = MI->getOperand(0).getReg(); 5789 unsigned ptr = MI->getOperand(1).getReg(); 5790 unsigned incr = MI->getOperand(2).getReg(); 5791 DebugLoc dl = MI->getDebugLoc(); 5792 bool isThumb2 = Subtarget->isThumb2(); 5793 5794 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 5795 if (isThumb2) { 5796 MRI.constrainRegClass(dest, &ARM::rGPRRegClass); 5797 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass); 5798 } 5799 5800 unsigned ldrOpc, strOpc; 5801 switch (Size) { 5802 default: llvm_unreachable("unsupported size for AtomicCmpSwap!"); 5803 case 1: 5804 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB; 5805 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB; 5806 break; 5807 case 2: 5808 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH; 5809 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH; 5810 break; 5811 case 4: 5812 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX; 5813 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX; 5814 break; 5815 } 5816 5817 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5818 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5819 MF->insert(It, loopMBB); 5820 MF->insert(It, exitMBB); 5821 5822 // Transfer the remainder of BB and its successor edges to exitMBB. 5823 exitMBB->splice(exitMBB->begin(), BB, 5824 llvm::next(MachineBasicBlock::iterator(MI)), 5825 BB->end()); 5826 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 5827 5828 const TargetRegisterClass *TRC = isThumb2 ? 5829 (const TargetRegisterClass*)&ARM::rGPRRegClass : 5830 (const TargetRegisterClass*)&ARM::GPRRegClass; 5831 unsigned scratch = MRI.createVirtualRegister(TRC); 5832 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC); 5833 5834 // thisMBB: 5835 // ... 5836 // fallthrough --> loopMBB 5837 BB->addSuccessor(loopMBB); 5838 5839 // loopMBB: 5840 // ldrex dest, ptr 5841 // <binop> scratch2, dest, incr 5842 // strex scratch, scratch2, ptr 5843 // cmp scratch, #0 5844 // bne- loopMBB 5845 // fallthrough --> exitMBB 5846 BB = loopMBB; 5847 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr); 5848 if (ldrOpc == ARM::t2LDREX) 5849 MIB.addImm(0); 5850 AddDefaultPred(MIB); 5851 if (BinOpcode) { 5852 // operand order needs to go the other way for NAND 5853 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr) 5854 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2). 5855 addReg(incr).addReg(dest)).addReg(0); 5856 else 5857 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2). 5858 addReg(dest).addReg(incr)).addReg(0); 5859 } 5860 5861 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr); 5862 if (strOpc == ARM::t2STREX) 5863 MIB.addImm(0); 5864 AddDefaultPred(MIB); 5865 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 5866 .addReg(scratch).addImm(0)); 5867 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5868 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5869 5870 BB->addSuccessor(loopMBB); 5871 BB->addSuccessor(exitMBB); 5872 5873 // exitMBB: 5874 // ... 5875 BB = exitMBB; 5876 5877 MI->eraseFromParent(); // The instruction is gone now. 5878 5879 return BB; 5880 } 5881 5882 MachineBasicBlock * 5883 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI, 5884 MachineBasicBlock *BB, 5885 unsigned Size, 5886 bool signExtend, 5887 ARMCC::CondCodes Cond) const { 5888 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5889 5890 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 5891 MachineFunction *MF = BB->getParent(); 5892 MachineFunction::iterator It = BB; 5893 ++It; 5894 5895 unsigned dest = MI->getOperand(0).getReg(); 5896 unsigned ptr = MI->getOperand(1).getReg(); 5897 unsigned incr = MI->getOperand(2).getReg(); 5898 unsigned oldval = dest; 5899 DebugLoc dl = MI->getDebugLoc(); 5900 bool isThumb2 = Subtarget->isThumb2(); 5901 5902 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 5903 if (isThumb2) { 5904 MRI.constrainRegClass(dest, &ARM::rGPRRegClass); 5905 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass); 5906 } 5907 5908 unsigned ldrOpc, strOpc, extendOpc; 5909 switch (Size) { 5910 default: llvm_unreachable("unsupported size for AtomicCmpSwap!"); 5911 case 1: 5912 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB; 5913 strOpc = isThumb2 ? ARM::t2STREXB :