1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file defines the interfaces that ARM uses to lower LLVM code into a 11 // selection DAG. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #define DEBUG_TYPE "arm-isel" 16 #include "ARMISelLowering.h" 17 #include "ARM.h" 18 #include "ARMCallingConv.h" 19 #include "ARMConstantPoolValue.h" 20 #include "ARMMachineFunctionInfo.h" 21 #include "ARMPerfectShuffle.h" 22 #include "ARMSubtarget.h" 23 #include "ARMTargetMachine.h" 24 #include "ARMTargetObjectFile.h" 25 #include "MCTargetDesc/ARMAddressingModes.h" 26 #include "llvm/CallingConv.h" 27 #include "llvm/Constants.h" 28 #include "llvm/Function.h" 29 #include "llvm/GlobalValue.h" 30 #include "llvm/Instruction.h" 31 #include "llvm/Instructions.h" 32 #include "llvm/Intrinsics.h" 33 #include "llvm/Type.h" 34 #include "llvm/CodeGen/CallingConvLower.h" 35 #include "llvm/CodeGen/IntrinsicLowering.h" 36 #include "llvm/CodeGen/MachineBasicBlock.h" 37 #include "llvm/CodeGen/MachineFrameInfo.h" 38 #include "llvm/CodeGen/MachineFunction.h" 39 #include "llvm/CodeGen/MachineInstrBuilder.h" 40 #include "llvm/CodeGen/MachineModuleInfo.h" 41 #include "llvm/CodeGen/MachineRegisterInfo.h" 42 #include "llvm/CodeGen/SelectionDAG.h" 43 #include "llvm/MC/MCSectionMachO.h" 44 #include "llvm/Target/TargetOptions.h" 45 #include "llvm/ADT/StringExtras.h" 46 #include "llvm/ADT/Statistic.h" 47 #include "llvm/Support/CommandLine.h" 48 #include "llvm/Support/ErrorHandling.h" 49 #include "llvm/Support/MathExtras.h" 50 #include "llvm/Support/raw_ostream.h" 51 using namespace llvm; 52 53 STATISTIC(NumTailCalls, "Number of tail calls"); 54 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt"); 55 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::SIGN_EXTEND_INREG, VT, Expand); 126 if (VT.isInteger()) { 127 setOperationAction(ISD::SHL, VT, Custom); 128 setOperationAction(ISD::SRA, VT, Custom); 129 setOperationAction(ISD::SRL, VT, Custom); 130 } 131 132 // Promote all bit-wise operations. 133 if (VT.isInteger() && VT != PromotedBitwiseVT) { 134 setOperationAction(ISD::AND, VT, Promote); 135 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT); 136 setOperationAction(ISD::OR, VT, Promote); 137 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT); 138 setOperationAction(ISD::XOR, VT, Promote); 139 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT); 140 } 141 142 // Neon does not support vector divide/remainder operations. 143 setOperationAction(ISD::SDIV, VT, Expand); 144 setOperationAction(ISD::UDIV, VT, Expand); 145 setOperationAction(ISD::FDIV, VT, Expand); 146 setOperationAction(ISD::SREM, VT, Expand); 147 setOperationAction(ISD::UREM, VT, Expand); 148 setOperationAction(ISD::FREM, VT, Expand); 149 } 150 151 void ARMTargetLowering::addDRTypeForNEON(MVT VT) { 152 addRegisterClass(VT, &ARM::DPRRegClass); 153 addTypeForNEON(VT, MVT::f64, MVT::v2i32); 154 } 155 156 void ARMTargetLowering::addQRTypeForNEON(MVT VT) { 157 addRegisterClass(VT, &ARM::QPRRegClass); 158 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32); 159 } 160 161 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) { 162 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin()) 163 return new TargetLoweringObjectFileMachO(); 164 165 return new ARMElfTargetObjectFile(); 166 } 167 168 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) 169 : TargetLowering(TM, createTLOF(TM)) { 170 Subtarget = &TM.getSubtarget<ARMSubtarget>(); 171 RegInfo = TM.getRegisterInfo(); 172 Itins = TM.getInstrItineraryData(); 173 174 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); 175 176 if (Subtarget->isTargetDarwin()) { 177 // Uses VFP for Thumb libfuncs if available. 178 if (Subtarget->isThumb() && Subtarget->hasVFP2()) { 179 // Single-precision floating-point arithmetic. 180 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp"); 181 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp"); 182 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp"); 183 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp"); 184 185 // Double-precision floating-point arithmetic. 186 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp"); 187 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp"); 188 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp"); 189 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp"); 190 191 // Single-precision comparisons. 192 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp"); 193 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp"); 194 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp"); 195 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp"); 196 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp"); 197 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp"); 198 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp"); 199 setLibcallName(RTLIB::O_F32, "__unordsf2vfp"); 200 201 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE); 202 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE); 203 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE); 204 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE); 205 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE); 206 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE); 207 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE); 208 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ); 209 210 // Double-precision comparisons. 211 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp"); 212 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp"); 213 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp"); 214 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp"); 215 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp"); 216 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp"); 217 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp"); 218 setLibcallName(RTLIB::O_F64, "__unorddf2vfp"); 219 220 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE); 221 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE); 222 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE); 223 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE); 224 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE); 225 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE); 226 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE); 227 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ); 228 229 // Floating-point to integer conversions. 230 // i64 conversions are done via library routines even when generating VFP 231 // instructions, so use the same ones. 232 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp"); 233 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp"); 234 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp"); 235 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp"); 236 237 // Conversions between floating types. 238 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp"); 239 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp"); 240 241 // Integer to floating-point conversions. 242 // i64 conversions are done via library routines even when generating VFP 243 // instructions, so use the same ones. 244 // FIXME: There appears to be some naming inconsistency in ARM libgcc: 245 // e.g., __floatunsidf vs. __floatunssidfvfp. 246 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp"); 247 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp"); 248 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp"); 249 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp"); 250 } 251 } 252 253 // These libcalls are not available in 32-bit. 254 setLibcallName(RTLIB::SHL_I128, 0); 255 setLibcallName(RTLIB::SRL_I128, 0); 256 setLibcallName(RTLIB::SRA_I128, 0); 257 258 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) { 259 // Double-precision floating-point arithmetic helper functions 260 // RTABI chapter 4.1.2, Table 2 261 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd"); 262 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv"); 263 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul"); 264 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub"); 265 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS); 266 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS); 267 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS); 268 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS); 269 270 // Double-precision floating-point comparison helper functions 271 // RTABI chapter 4.1.2, Table 3 272 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq"); 273 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE); 274 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq"); 275 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ); 276 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt"); 277 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE); 278 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple"); 279 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE); 280 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge"); 281 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE); 282 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt"); 283 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE); 284 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun"); 285 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE); 286 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun"); 287 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ); 288 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS); 289 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS); 290 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS); 291 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS); 292 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS); 293 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS); 294 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS); 295 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS); 296 297 // Single-precision floating-point arithmetic helper functions 298 // RTABI chapter 4.1.2, Table 4 299 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd"); 300 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv"); 301 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul"); 302 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub"); 303 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS); 304 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS); 305 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS); 306 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS); 307 308 // Single-precision floating-point comparison helper functions 309 // RTABI chapter 4.1.2, Table 5 310 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq"); 311 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE); 312 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq"); 313 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ); 314 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt"); 315 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE); 316 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple"); 317 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE); 318 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge"); 319 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE); 320 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt"); 321 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE); 322 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun"); 323 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE); 324 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun"); 325 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ); 326 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS); 327 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS); 328 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS); 329 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS); 330 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS); 331 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS); 332 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS); 333 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS); 334 335 // Floating-point to integer conversions. 336 // RTABI chapter 4.1.2, Table 6 337 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz"); 338 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz"); 339 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz"); 340 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz"); 341 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz"); 342 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz"); 343 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz"); 344 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz"); 345 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS); 346 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS); 347 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS); 348 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS); 349 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS); 350 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS); 351 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS); 352 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS); 353 354 // Conversions between floating types. 355 // RTABI chapter 4.1.2, Table 7 356 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f"); 357 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d"); 358 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS); 359 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS); 360 361 // Integer to floating-point conversions. 362 // RTABI chapter 4.1.2, Table 8 363 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d"); 364 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d"); 365 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d"); 366 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d"); 367 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f"); 368 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f"); 369 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f"); 370 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f"); 371 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS); 372 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS); 373 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS); 374 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS); 375 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS); 376 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS); 377 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS); 378 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS); 379 380 // Long long helper functions 381 // RTABI chapter 4.2, Table 9 382 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul"); 383 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl"); 384 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr"); 385 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr"); 386 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS); 387 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS); 388 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS); 389 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS); 390 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS); 391 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS); 392 393 // Integer division functions 394 // RTABI chapter 4.3.1 395 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv"); 396 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv"); 397 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv"); 398 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod"); 399 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv"); 400 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv"); 401 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv"); 402 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod"); 403 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS); 404 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS); 405 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS); 406 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS); 407 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS); 408 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS); 409 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS); 410 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS); 411 412 // Memory operations 413 // RTABI chapter 4.3.4 414 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy"); 415 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove"); 416 setLibcallName(RTLIB::MEMSET, "__aeabi_memset"); 417 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS); 418 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS); 419 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS); 420 } 421 422 // Use divmod compiler-rt calls for iOS 5.0 and later. 423 if (Subtarget->getTargetTriple().getOS() == Triple::IOS && 424 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) { 425 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4"); 426 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4"); 427 } 428 429 if (Subtarget->isThumb1Only()) 430 addRegisterClass(MVT::i32, &ARM::tGPRRegClass); 431 else 432 addRegisterClass(MVT::i32, &ARM::GPRRegClass); 433 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() && 434 !Subtarget->isThumb1Only()) { 435 addRegisterClass(MVT::f32, &ARM::SPRRegClass); 436 if (!Subtarget->isFPOnlySP()) 437 addRegisterClass(MVT::f64, &ARM::DPRRegClass); 438 439 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 440 } 441 442 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE; 443 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) { 444 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE; 445 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT) 446 setTruncStoreAction((MVT::SimpleValueType)VT, 447 (MVT::SimpleValueType)InnerVT, Expand); 448 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand); 449 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand); 450 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand); 451 } 452 453 setOperationAction(ISD::ConstantFP, MVT::f32, Custom); 454 455 if (Subtarget->hasNEON()) { 456 addDRTypeForNEON(MVT::v2f32); 457 addDRTypeForNEON(MVT::v8i8); 458 addDRTypeForNEON(MVT::v4i16); 459 addDRTypeForNEON(MVT::v2i32); 460 addDRTypeForNEON(MVT::v1i64); 461 462 addQRTypeForNEON(MVT::v4f32); 463 addQRTypeForNEON(MVT::v2f64); 464 addQRTypeForNEON(MVT::v16i8); 465 addQRTypeForNEON(MVT::v8i16); 466 addQRTypeForNEON(MVT::v4i32); 467 addQRTypeForNEON(MVT::v2i64); 468 469 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but 470 // neither Neon nor VFP support any arithmetic operations on it. 471 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively 472 // supported for v4f32. 473 setOperationAction(ISD::FADD, MVT::v2f64, Expand); 474 setOperationAction(ISD::FSUB, MVT::v2f64, Expand); 475 setOperationAction(ISD::FMUL, MVT::v2f64, Expand); 476 // FIXME: Code duplication: FDIV and FREM are expanded always, see 477 // ARMTargetLowering::addTypeForNEON method for details. 478 setOperationAction(ISD::FDIV, MVT::v2f64, Expand); 479 setOperationAction(ISD::FREM, MVT::v2f64, Expand); 480 // FIXME: Create unittest. 481 // In another words, find a way when "copysign" appears in DAG with vector 482 // operands. 483 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand); 484 // FIXME: Code duplication: SETCC has custom operation action, see 485 // ARMTargetLowering::addTypeForNEON method for details. 486 setOperationAction(ISD::SETCC, MVT::v2f64, Expand); 487 // FIXME: Create unittest for FNEG and for FABS. 488 setOperationAction(ISD::FNEG, MVT::v2f64, Expand); 489 setOperationAction(ISD::FABS, MVT::v2f64, Expand); 490 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand); 491 setOperationAction(ISD::FSIN, MVT::v2f64, Expand); 492 setOperationAction(ISD::FCOS, MVT::v2f64, Expand); 493 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand); 494 setOperationAction(ISD::FPOW, MVT::v2f64, Expand); 495 setOperationAction(ISD::FLOG, MVT::v2f64, Expand); 496 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand); 497 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand); 498 setOperationAction(ISD::FEXP, MVT::v2f64, Expand); 499 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand); 500 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR. 501 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand); 502 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand); 503 setOperationAction(ISD::FRINT, MVT::v2f64, Expand); 504 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand); 505 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand); 506 507 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand); 508 setOperationAction(ISD::FSIN, MVT::v4f32, Expand); 509 setOperationAction(ISD::FCOS, MVT::v4f32, Expand); 510 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand); 511 setOperationAction(ISD::FPOW, MVT::v4f32, Expand); 512 setOperationAction(ISD::FLOG, MVT::v4f32, Expand); 513 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand); 514 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand); 515 setOperationAction(ISD::FEXP, MVT::v4f32, Expand); 516 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand); 517 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand); 518 519 // Neon does not support some operations on v1i64 and v2i64 types. 520 setOperationAction(ISD::MUL, MVT::v1i64, Expand); 521 // Custom handling for some quad-vector types to detect VMULL. 522 setOperationAction(ISD::MUL, MVT::v8i16, Custom); 523 setOperationAction(ISD::MUL, MVT::v4i32, Custom); 524 setOperationAction(ISD::MUL, MVT::v2i64, Custom); 525 // Custom handling for some vector types to avoid expensive expansions 526 setOperationAction(ISD::SDIV, MVT::v4i16, Custom); 527 setOperationAction(ISD::SDIV, MVT::v8i8, Custom); 528 setOperationAction(ISD::UDIV, MVT::v4i16, Custom); 529 setOperationAction(ISD::UDIV, MVT::v8i8, Custom); 530 setOperationAction(ISD::SETCC, MVT::v1i64, Expand); 531 setOperationAction(ISD::SETCC, MVT::v2i64, Expand); 532 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with 533 // a destination type that is wider than the source, and nor does 534 // it have a FP_TO_[SU]INT instruction with a narrower destination than 535 // source. 536 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom); 537 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom); 538 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom); 539 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom); 540 541 setTargetDAGCombine(ISD::INTRINSIC_VOID); 542 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN); 543 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN); 544 setTargetDAGCombine(ISD::SHL); 545 setTargetDAGCombine(ISD::SRL); 546 setTargetDAGCombine(ISD::SRA); 547 setTargetDAGCombine(ISD::SIGN_EXTEND); 548 setTargetDAGCombine(ISD::ZERO_EXTEND); 549 setTargetDAGCombine(ISD::ANY_EXTEND); 550 setTargetDAGCombine(ISD::SELECT_CC); 551 setTargetDAGCombine(ISD::BUILD_VECTOR); 552 setTargetDAGCombine(ISD::VECTOR_SHUFFLE); 553 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT); 554 setTargetDAGCombine(ISD::STORE); 555 setTargetDAGCombine(ISD::FP_TO_SINT); 556 setTargetDAGCombine(ISD::FP_TO_UINT); 557 setTargetDAGCombine(ISD::FDIV); 558 559 // It is legal to extload from v4i8 to v4i16 or v4i32. 560 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8, 561 MVT::v4i16, MVT::v2i16, 562 MVT::v2i32}; 563 for (unsigned i = 0; i < 6; ++i) { 564 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal); 565 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal); 566 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal); 567 } 568 } 569 570 // ARM and Thumb2 support UMLAL/SMLAL. 571 if (!Subtarget->isThumb1Only()) 572 setTargetDAGCombine(ISD::ADDC); 573 574 575 computeRegisterProperties(); 576 577 // ARM does not have f32 extending load. 578 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand); 579 580 // ARM does not have i1 sign extending load. 581 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 582 583 // ARM supports all 4 flavors of integer indexed load / store. 584 if (!Subtarget->isThumb1Only()) { 585 for (unsigned im = (unsigned)ISD::PRE_INC; 586 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) { 587 setIndexedLoadAction(im, MVT::i1, Legal); 588 setIndexedLoadAction(im, MVT::i8, Legal); 589 setIndexedLoadAction(im, MVT::i16, Legal); 590 setIndexedLoadAction(im, MVT::i32, Legal); 591 setIndexedStoreAction(im, MVT::i1, Legal); 592 setIndexedStoreAction(im, MVT::i8, Legal); 593 setIndexedStoreAction(im, MVT::i16, Legal); 594 setIndexedStoreAction(im, MVT::i32, Legal); 595 } 596 } 597 598 // i64 operation support. 599 setOperationAction(ISD::MUL, MVT::i64, Expand); 600 setOperationAction(ISD::MULHU, MVT::i32, Expand); 601 if (Subtarget->isThumb1Only()) { 602 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand); 603 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand); 604 } 605 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops() 606 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP())) 607 setOperationAction(ISD::MULHS, MVT::i32, Expand); 608 609 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); 610 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); 611 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); 612 setOperationAction(ISD::SRL, MVT::i64, Custom); 613 setOperationAction(ISD::SRA, MVT::i64, Custom); 614 615 if (!Subtarget->isThumb1Only()) { 616 // FIXME: We should do this for Thumb1 as well. 617 setOperationAction(ISD::ADDC, MVT::i32, Custom); 618 setOperationAction(ISD::ADDE, MVT::i32, Custom); 619 setOperationAction(ISD::SUBC, MVT::i32, Custom); 620 setOperationAction(ISD::SUBE, MVT::i32, Custom); 621 } 622 623 // ARM does not have ROTL. 624 setOperationAction(ISD::ROTL, MVT::i32, Expand); 625 setOperationAction(ISD::CTTZ, MVT::i32, Custom); 626 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 627 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only()) 628 setOperationAction(ISD::CTLZ, MVT::i32, Expand); 629 630 // These just redirect to CTTZ and CTLZ on ARM. 631 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand); 632 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand); 633 634 // Only ARMv6 has BSWAP. 635 if (!Subtarget->hasV6Ops()) 636 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 637 638 // These are expanded into libcalls. 639 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) { 640 // v7M has a hardware divider 641 setOperationAction(ISD::SDIV, MVT::i32, Expand); 642 setOperationAction(ISD::UDIV, MVT::i32, Expand); 643 } 644 setOperationAction(ISD::SREM, MVT::i32, Expand); 645 setOperationAction(ISD::UREM, MVT::i32, Expand); 646 setOperationAction(ISD::SDIVREM, MVT::i32, Expand); 647 setOperationAction(ISD::UDIVREM, MVT::i32, Expand); 648 649 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 650 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 651 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom); 652 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 653 setOperationAction(ISD::BlockAddress, MVT::i32, Custom); 654 655 setOperationAction(ISD::TRAP, MVT::Other, Legal); 656 657 // Use the default implementation. 658 setOperationAction(ISD::VASTART, MVT::Other, Custom); 659 setOperationAction(ISD::VAARG, MVT::Other, Expand); 660 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 661 setOperationAction(ISD::VAEND, MVT::Other, Expand); 662 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 663 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 664 665 if (!Subtarget->isTargetDarwin()) { 666 // Non-Darwin platforms may return values in these registers via the 667 // personality function. 668 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand); 669 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand); 670 setExceptionPointerRegister(ARM::R0); 671 setExceptionSelectorRegister(ARM::R1); 672 } 673 674 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 675 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use 676 // the default expansion. 677 // FIXME: This should be checking for v6k, not just v6. 678 if (Subtarget->hasDataBarrier() || 679 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) { 680 // membarrier needs custom lowering; the rest are legal and handled 681 // normally. 682 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom); 683 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom); 684 // Custom lowering for 64-bit ops 685 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom); 686 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom); 687 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom); 688 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom); 689 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom); 690 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom); 691 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom); 692 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc. 693 setInsertFencesForAtomic(true); 694 } else { 695 // Set them all for expansion, which will force libcalls. 696 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand); 697 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand); 698 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand); 699 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand); 700 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand); 701 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand); 702 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand); 703 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand); 704 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand); 705 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand); 706 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand); 707 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand); 708 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand); 709 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand); 710 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the 711 // Unordered/Monotonic case. 712 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom); 713 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom); 714 // Since the libcalls include locking, fold in the fences 715 setShouldFoldAtomicFences(true); 716 } 717 718 setOperationAction(ISD::PREFETCH, MVT::Other, Custom); 719 720 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes. 721 if (!Subtarget->hasV6Ops()) { 722 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 723 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); 724 } 725 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 726 727 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() && 728 !Subtarget->isThumb1Only()) { 729 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR 730 // iff target supports vfp2. 731 setOperationAction(ISD::BITCAST, MVT::i64, Custom); 732 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom); 733 } 734 735 // We want to custom lower some of our intrinsics. 736 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); 737 if (Subtarget->isTargetDarwin()) { 738 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom); 739 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom); 740 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume"); 741 } 742 743 setOperationAction(ISD::SETCC, MVT::i32, Expand); 744 setOperationAction(ISD::SETCC, MVT::f32, Expand); 745 setOperationAction(ISD::SETCC, MVT::f64, Expand); 746 setOperationAction(ISD::SELECT, MVT::i32, Custom); 747 setOperationAction(ISD::SELECT, MVT::f32, Custom); 748 setOperationAction(ISD::SELECT, MVT::f64, Custom); 749 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom); 750 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 751 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 752 753 setOperationAction(ISD::BRCOND, MVT::Other, Expand); 754 setOperationAction(ISD::BR_CC, MVT::i32, Custom); 755 setOperationAction(ISD::BR_CC, MVT::f32, Custom); 756 setOperationAction(ISD::BR_CC, MVT::f64, Custom); 757 setOperationAction(ISD::BR_JT, MVT::Other, Custom); 758 759 // We don't support sin/cos/fmod/copysign/pow 760 setOperationAction(ISD::FSIN, MVT::f64, Expand); 761 setOperationAction(ISD::FSIN, MVT::f32, Expand); 762 setOperationAction(ISD::FCOS, MVT::f32, Expand); 763 setOperationAction(ISD::FCOS, MVT::f64, Expand); 764 setOperationAction(ISD::FREM, MVT::f64, Expand); 765 setOperationAction(ISD::FREM, MVT::f32, Expand); 766 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() && 767 !Subtarget->isThumb1Only()) { 768 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 769 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 770 } 771 setOperationAction(ISD::FPOW, MVT::f64, Expand); 772 setOperationAction(ISD::FPOW, MVT::f32, Expand); 773 774 if (!Subtarget->hasVFP4()) { 775 setOperationAction(ISD::FMA, MVT::f64, Expand); 776 setOperationAction(ISD::FMA, MVT::f32, Expand); 777 } 778 779 // Various VFP goodness 780 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) { 781 // int <-> fp are custom expanded into bit_convert + ARMISD ops. 782 if (Subtarget->hasVFP2()) { 783 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom); 784 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom); 785 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); 786 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 787 } 788 // Special handling for half-precision FP. 789 if (!Subtarget->hasFP16()) { 790 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand); 791 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand); 792 } 793 } 794 795 // We have target-specific dag combine patterns for the following nodes: 796 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine 797 setTargetDAGCombine(ISD::ADD); 798 setTargetDAGCombine(ISD::SUB); 799 setTargetDAGCombine(ISD::MUL); 800 setTargetDAGCombine(ISD::AND); 801 setTargetDAGCombine(ISD::OR); 802 setTargetDAGCombine(ISD::XOR); 803 804 if (Subtarget->hasV6Ops()) 805 setTargetDAGCombine(ISD::SRL); 806 807 setStackPointerRegisterToSaveRestore(ARM::SP); 808 809 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() || 810 !Subtarget->hasVFP2()) 811 setSchedulingPreference(Sched::RegPressure); 812 else 813 setSchedulingPreference(Sched::Hybrid); 814 815 //// temporary - rewrite interface to use type 816 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1; 817 maxStoresPerMemset = 16; 818 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4; 819 820 // On ARM arguments smaller than 4 bytes are extended, so all arguments 821 // are at least 4 bytes aligned. 822 setMinStackArgumentAlignment(4); 823 824 benefitFromCodePlacementOpt = true; 825 826 // Prefer likely predicted branches to selects on out-of-order cores. 827 predictableSelectIsExpensive = Subtarget->isCortexA9(); 828 829 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2); 830 } 831 832 // FIXME: It might make sense to define the representative register class as the 833 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is 834 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently, 835 // SPR's representative would be DPR_VFP2. This should work well if register 836 // pressure tracking were modified such that a register use would increment the 837 // pressure of the register class's representative and all of it's super 838 // classes' representatives transitively. We have not implemented this because 839 // of the difficulty prior to coalescing of modeling operand register classes 840 // due to the common occurrence of cross class copies and subregister insertions 841 // and extractions. 842 std::pair<const TargetRegisterClass*, uint8_t> 843 ARMTargetLowering::findRepresentativeClass(EVT VT) const{ 844 const TargetRegisterClass *RRC = 0; 845 uint8_t Cost = 1; 846 switch (VT.getSimpleVT().SimpleTy) { 847 default: 848 return TargetLowering::findRepresentativeClass(VT); 849 // Use DPR as representative register class for all floating point 850 // and vector types. Since there are 32 SPR registers and 32 DPR registers so 851 // the cost is 1 for both f32 and f64. 852 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16: 853 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32: 854 RRC = &ARM::DPRRegClass; 855 // When NEON is used for SP, only half of the register file is available 856 // because operations that define both SP and DP results will be constrained 857 // to the VFP2 class (D0-D15). We currently model this constraint prior to 858 // coalescing by double-counting the SP regs. See the FIXME above. 859 if (Subtarget->useNEONForSinglePrecisionFP()) 860 Cost = 2; 861 break; 862 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64: 863 case MVT::v4f32: case MVT::v2f64: 864 RRC = &ARM::DPRRegClass; 865 Cost = 2; 866 break; 867 case MVT::v4i64: 868 RRC = &ARM::DPRRegClass; 869 Cost = 4; 870 break; 871 case MVT::v8i64: 872 RRC = &ARM::DPRRegClass; 873 Cost = 8; 874 break; 875 } 876 return std::make_pair(RRC, Cost); 877 } 878 879 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const { 880 switch (Opcode) { 881 default: return 0; 882 case ARMISD::Wrapper: return "ARMISD::Wrapper"; 883 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN"; 884 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC"; 885 case ARMISD::WrapperJT: return "ARMISD::WrapperJT"; 886 case ARMISD::CALL: return "ARMISD::CALL"; 887 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED"; 888 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK"; 889 case ARMISD::tCALL: return "ARMISD::tCALL"; 890 case ARMISD::BRCOND: return "ARMISD::BRCOND"; 891 case ARMISD::BR_JT: return "ARMISD::BR_JT"; 892 case ARMISD::BR2_JT: return "ARMISD::BR2_JT"; 893 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG"; 894 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD"; 895 case ARMISD::CMP: return "ARMISD::CMP"; 896 case ARMISD::CMN: return "ARMISD::CMN"; 897 case ARMISD::CMPZ: return "ARMISD::CMPZ"; 898 case ARMISD::CMPFP: return "ARMISD::CMPFP"; 899 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0"; 900 case ARMISD::BCC_i64: return "ARMISD::BCC_i64"; 901 case ARMISD::FMSTAT: return "ARMISD::FMSTAT"; 902 903 case ARMISD::CMOV: return "ARMISD::CMOV"; 904 905 case ARMISD::RBIT: return "ARMISD::RBIT"; 906 907 case ARMISD::FTOSI: return "ARMISD::FTOSI"; 908 case ARMISD::FTOUI: return "ARMISD::FTOUI"; 909 case ARMISD::SITOF: return "ARMISD::SITOF"; 910 case ARMISD::UITOF: return "ARMISD::UITOF"; 911 912 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG"; 913 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG"; 914 case ARMISD::RRX: return "ARMISD::RRX"; 915 916 case ARMISD::ADDC: return "ARMISD::ADDC"; 917 case ARMISD::ADDE: return "ARMISD::ADDE"; 918 case ARMISD::SUBC: return "ARMISD::SUBC"; 919 case ARMISD::SUBE: return "ARMISD::SUBE"; 920 921 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD"; 922 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR"; 923 924 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP"; 925 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP"; 926 927 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN"; 928 929 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER"; 930 931 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC"; 932 933 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER"; 934 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR"; 935 936 case ARMISD::PRELOAD: return "ARMISD::PRELOAD"; 937 938 case ARMISD::VCEQ: return "ARMISD::VCEQ"; 939 case ARMISD::VCEQZ: return "ARMISD::VCEQZ"; 940 case ARMISD::VCGE: return "ARMISD::VCGE"; 941 case ARMISD::VCGEZ: return "ARMISD::VCGEZ"; 942 case ARMISD::VCLEZ: return "ARMISD::VCLEZ"; 943 case ARMISD::VCGEU: return "ARMISD::VCGEU"; 944 case ARMISD::VCGT: return "ARMISD::VCGT"; 945 case ARMISD::VCGTZ: return "ARMISD::VCGTZ"; 946 case ARMISD::VCLTZ: return "ARMISD::VCLTZ"; 947 case ARMISD::VCGTU: return "ARMISD::VCGTU"; 948 case ARMISD::VTST: return "ARMISD::VTST"; 949 950 case ARMISD::VSHL: return "ARMISD::VSHL"; 951 case ARMISD::VSHRs: return "ARMISD::VSHRs"; 952 case ARMISD::VSHRu: return "ARMISD::VSHRu"; 953 case ARMISD::VSHLLs: return "ARMISD::VSHLLs"; 954 case ARMISD::VSHLLu: return "ARMISD::VSHLLu"; 955 case ARMISD::VSHLLi: return "ARMISD::VSHLLi"; 956 case ARMISD::VSHRN: return "ARMISD::VSHRN"; 957 case ARMISD::VRSHRs: return "ARMISD::VRSHRs"; 958 case ARMISD::VRSHRu: return "ARMISD::VRSHRu"; 959 case ARMISD::VRSHRN: return "ARMISD::VRSHRN"; 960 case ARMISD::VQSHLs: return "ARMISD::VQSHLs"; 961 case ARMISD::VQSHLu: return "ARMISD::VQSHLu"; 962 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu"; 963 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs"; 964 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu"; 965 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu"; 966 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs"; 967 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu"; 968 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu"; 969 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu"; 970 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs"; 971 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM"; 972 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM"; 973 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM"; 974 case ARMISD::VDUP: return "ARMISD::VDUP"; 975 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE"; 976 case ARMISD::VEXT: return "ARMISD::VEXT"; 977 case ARMISD::VREV64: return "ARMISD::VREV64"; 978 case ARMISD::VREV32: return "ARMISD::VREV32"; 979 case ARMISD::VREV16: return "ARMISD::VREV16"; 980 case ARMISD::VZIP: return "ARMISD::VZIP"; 981 case ARMISD::VUZP: return "ARMISD::VUZP"; 982 case ARMISD::VTRN: return "ARMISD::VTRN"; 983 case ARMISD::VTBL1: return "ARMISD::VTBL1"; 984 case ARMISD::VTBL2: return "ARMISD::VTBL2"; 985 case ARMISD::VMULLs: return "ARMISD::VMULLs"; 986 case ARMISD::VMULLu: return "ARMISD::VMULLu"; 987 case ARMISD::UMLAL: return "ARMISD::UMLAL"; 988 case ARMISD::SMLAL: return "ARMISD::SMLAL"; 989 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR"; 990 case ARMISD::FMAX: return "ARMISD::FMAX"; 991 case ARMISD::FMIN: return "ARMISD::FMIN"; 992 case ARMISD::BFI: return "ARMISD::BFI"; 993 case ARMISD::VORRIMM: return "ARMISD::VORRIMM"; 994 case ARMISD::VBICIMM: return "ARMISD::VBICIMM"; 995 case ARMISD::VBSL: return "ARMISD::VBSL"; 996 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP"; 997 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP"; 998 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP"; 999 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD"; 1000 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD"; 1001 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD"; 1002 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD"; 1003 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD"; 1004 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD"; 1005 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD"; 1006 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD"; 1007 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD"; 1008 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD"; 1009 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD"; 1010 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD"; 1011 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD"; 1012 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD"; 1013 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD"; 1014 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD"; 1015 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD"; 1016 } 1017 } 1018 1019 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const { 1020 if (!VT.isVector()) return getPointerTy(); 1021 return VT.changeVectorElementTypeToInteger(); 1022 } 1023 1024 /// getRegClassFor - Return the register class that should be used for the 1025 /// specified value type. 1026 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const { 1027 // Map v4i64 to QQ registers but do not make the type legal. Similarly map 1028 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to 1029 // load / store 4 to 8 consecutive D registers. 1030 if (Subtarget->hasNEON()) { 1031 if (VT == MVT::v4i64) 1032 return &ARM::QQPRRegClass; 1033 if (VT == MVT::v8i64) 1034 return &ARM::QQQQPRRegClass; 1035 } 1036 return TargetLowering::getRegClassFor(VT); 1037 } 1038 1039 // Create a fast isel object. 1040 FastISel * 1041 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo, 1042 const TargetLibraryInfo *libInfo) const { 1043 return ARM::createFastISel(funcInfo, libInfo); 1044 } 1045 1046 /// getMaximalGlobalOffset - Returns the maximal possible offset which can 1047 /// be used for loads / stores from the global. 1048 unsigned ARMTargetLowering::getMaximalGlobalOffset() const { 1049 return (Subtarget->isThumb1Only() ? 127 : 4095); 1050 } 1051 1052 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const { 1053 unsigned NumVals = N->getNumValues(); 1054 if (!NumVals) 1055 return Sched::RegPressure; 1056 1057 for (unsigned i = 0; i != NumVals; ++i) { 1058 EVT VT = N->getValueType(i); 1059 if (VT == MVT::Glue || VT == MVT::Other) 1060 continue; 1061 if (VT.isFloatingPoint() || VT.isVector()) 1062 return Sched::ILP; 1063 } 1064 1065 if (!N->isMachineOpcode()) 1066 return Sched::RegPressure; 1067 1068 // Load are scheduled for latency even if there instruction itinerary 1069 // is not available. 1070 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1071 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode()); 1072 1073 if (MCID.getNumDefs() == 0) 1074 return Sched::RegPressure; 1075 if (!Itins->isEmpty() && 1076 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2) 1077 return Sched::ILP; 1078 1079 return Sched::RegPressure; 1080 } 1081 1082 //===----------------------------------------------------------------------===// 1083 // Lowering Code 1084 //===----------------------------------------------------------------------===// 1085 1086 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC 1087 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) { 1088 switch (CC) { 1089 default: llvm_unreachable("Unknown condition code!"); 1090 case ISD::SETNE: return ARMCC::NE; 1091 case ISD::SETEQ: return ARMCC::EQ; 1092 case ISD::SETGT: return ARMCC::GT; 1093 case ISD::SETGE: return ARMCC::GE; 1094 case ISD::SETLT: return ARMCC::LT; 1095 case ISD::SETLE: return ARMCC::LE; 1096 case ISD::SETUGT: return ARMCC::HI; 1097 case ISD::SETUGE: return ARMCC::HS; 1098 case ISD::SETULT: return ARMCC::LO; 1099 case ISD::SETULE: return ARMCC::LS; 1100 } 1101 } 1102 1103 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. 1104 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode, 1105 ARMCC::CondCodes &CondCode2) { 1106 CondCode2 = ARMCC::AL; 1107 switch (CC) { 1108 default: llvm_unreachable("Unknown FP condition!"); 1109 case ISD::SETEQ: 1110 case ISD::SETOEQ: CondCode = ARMCC::EQ; break; 1111 case ISD::SETGT: 1112 case ISD::SETOGT: CondCode = ARMCC::GT; break; 1113 case ISD::SETGE: 1114 case ISD::SETOGE: CondCode = ARMCC::GE; break; 1115 case ISD::SETOLT: CondCode = ARMCC::MI; break; 1116 case ISD::SETOLE: CondCode = ARMCC::LS; break; 1117 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break; 1118 case ISD::SETO: CondCode = ARMCC::VC; break; 1119 case ISD::SETUO: CondCode = ARMCC::VS; break; 1120 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break; 1121 case ISD::SETUGT: CondCode = ARMCC::HI; break; 1122 case ISD::SETUGE: CondCode = ARMCC::PL; break; 1123 case ISD::SETLT: 1124 case ISD::SETULT: CondCode = ARMCC::LT; break; 1125 case ISD::SETLE: 1126 case ISD::SETULE: CondCode = ARMCC::LE; break; 1127 case ISD::SETNE: 1128 case ISD::SETUNE: CondCode = ARMCC::NE; break; 1129 } 1130 } 1131 1132 //===----------------------------------------------------------------------===// 1133 // Calling Convention Implementation 1134 //===----------------------------------------------------------------------===// 1135 1136 #include "ARMGenCallingConv.inc" 1137 1138 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the 1139 /// given CallingConvention value. 1140 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC, 1141 bool Return, 1142 bool isVarArg) const { 1143 switch (CC) { 1144 default: 1145 llvm_unreachable("Unsupported calling convention"); 1146 case CallingConv::Fast: 1147 if (Subtarget->hasVFP2() && !isVarArg) { 1148 if (!Subtarget->isAAPCS_ABI()) 1149 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS); 1150 // For AAPCS ABI targets, just use VFP variant of the calling convention. 1151 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP); 1152 } 1153 // Fallthrough 1154 case CallingConv::C: { 1155 // Use target triple & subtarget features to do actual dispatch. 1156 if (!Subtarget->isAAPCS_ABI()) 1157 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS); 1158 else if (Subtarget->hasVFP2() && 1159 getTargetMachine().Options.FloatABIType == FloatABI::Hard && 1160 !isVarArg) 1161 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP); 1162 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS); 1163 } 1164 case CallingConv::ARM_AAPCS_VFP: 1165 if (!isVarArg) 1166 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP); 1167 // Fallthrough 1168 case CallingConv::ARM_AAPCS: 1169 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS); 1170 case CallingConv::ARM_APCS: 1171 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS); 1172 case CallingConv::GHC: 1173 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC); 1174 } 1175 } 1176 1177 /// LowerCallResult - Lower the result values of a call into the 1178 /// appropriate copies out of appropriate physical registers. 1179 SDValue 1180 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, 1181 CallingConv::ID CallConv, bool isVarArg, 1182 const SmallVectorImpl<ISD::InputArg> &Ins, 1183 DebugLoc dl, SelectionDAG &DAG, 1184 SmallVectorImpl<SDValue> &InVals) const { 1185 1186 // Assign locations to each value returned by this call. 1187 SmallVector<CCValAssign, 16> RVLocs; 1188 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1189 getTargetMachine(), RVLocs, *DAG.getContext(), Call); 1190 CCInfo.AnalyzeCallResult(Ins, 1191 CCAssignFnForNode(CallConv, /* Return*/ true, 1192 isVarArg)); 1193 1194 // Copy all of the result registers out of their specified physreg. 1195 for (unsigned i = 0; i != RVLocs.size(); ++i) { 1196 CCValAssign VA = RVLocs[i]; 1197 1198 SDValue Val; 1199 if (VA.needsCustom()) { 1200 // Handle f64 or half of a v2f64. 1201 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, 1202 InFlag); 1203 Chain = Lo.getValue(1); 1204 InFlag = Lo.getValue(2); 1205 VA = RVLocs[++i]; // skip ahead to next loc 1206 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, 1207 InFlag); 1208 Chain = Hi.getValue(1); 1209 InFlag = Hi.getValue(2); 1210 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi); 1211 1212 if (VA.getLocVT() == MVT::v2f64) { 1213 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64); 1214 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val, 1215 DAG.getConstant(0, MVT::i32)); 1216 1217 VA = RVLocs[++i]; // skip ahead to next loc 1218 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag); 1219 Chain = Lo.getValue(1); 1220 InFlag = Lo.getValue(2); 1221 VA = RVLocs[++i]; // skip ahead to next loc 1222 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag); 1223 Chain = Hi.getValue(1); 1224 InFlag = Hi.getValue(2); 1225 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi); 1226 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val, 1227 DAG.getConstant(1, MVT::i32)); 1228 } 1229 } else { 1230 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(), 1231 InFlag); 1232 Chain = Val.getValue(1); 1233 InFlag = Val.getValue(2); 1234 } 1235 1236 switch (VA.getLocInfo()) { 1237 default: llvm_unreachable("Unknown loc info!"); 1238 case CCValAssign::Full: break; 1239 case CCValAssign::BCvt: 1240 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val); 1241 break; 1242 } 1243 1244 InVals.push_back(Val); 1245 } 1246 1247 return Chain; 1248 } 1249 1250 /// LowerMemOpCallTo - Store the argument to the stack. 1251 SDValue 1252 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain, 1253 SDValue StackPtr, SDValue Arg, 1254 DebugLoc dl, SelectionDAG &DAG, 1255 const CCValAssign &VA, 1256 ISD::ArgFlagsTy Flags) const { 1257 unsigned LocMemOffset = VA.getLocMemOffset(); 1258 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset); 1259 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff); 1260 return DAG.getStore(Chain, dl, Arg, PtrOff, 1261 MachinePointerInfo::getStack(LocMemOffset), 1262 false, false, 0); 1263 } 1264 1265 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG, 1266 SDValue Chain, SDValue &Arg, 1267 RegsToPassVector &RegsToPass, 1268 CCValAssign &VA, CCValAssign &NextVA, 1269 SDValue &StackPtr, 1270 SmallVector<SDValue, 8> &MemOpChains, 1271 ISD::ArgFlagsTy Flags) const { 1272 1273 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl, 1274 DAG.getVTList(MVT::i32, MVT::i32), Arg); 1275 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd)); 1276 1277 if (NextVA.isRegLoc()) 1278 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1))); 1279 else { 1280 assert(NextVA.isMemLoc()); 1281 if (StackPtr.getNode() == 0) 1282 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 1283 1284 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1), 1285 dl, DAG, NextVA, 1286 Flags)); 1287 } 1288 } 1289 1290 /// LowerCall - Lowering a call into a callseq_start <- 1291 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter 1292 /// nodes. 1293 SDValue 1294 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, 1295 SmallVectorImpl<SDValue> &InVals) const { 1296 SelectionDAG &DAG = CLI.DAG; 1297 DebugLoc &dl = CLI.DL; 1298 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs; 1299 SmallVector<SDValue, 32> &OutVals = CLI.OutVals; 1300 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins; 1301 SDValue Chain = CLI.Chain; 1302 SDValue Callee = CLI.Callee; 1303 bool &isTailCall = CLI.IsTailCall; 1304 CallingConv::ID CallConv = CLI.CallConv; 1305 bool doesNotRet = CLI.DoesNotReturn; 1306 bool isVarArg = CLI.IsVarArg; 1307 1308 MachineFunction &MF = DAG.getMachineFunction(); 1309 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet(); 1310 bool IsSibCall = false; 1311 // Disable tail calls if they're not supported. 1312 if (!EnableARMTailCalls && !Subtarget->supportsTailCall()) 1313 isTailCall = false; 1314 if (isTailCall) { 1315 // Check if it's really possible to do a tail call. 1316 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, 1317 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(), 1318 Outs, OutVals, Ins, DAG); 1319 // We don't support GuaranteedTailCallOpt for ARM, only automatically 1320 // detected sibcalls. 1321 if (isTailCall) { 1322 ++NumTailCalls; 1323 IsSibCall = true; 1324 } 1325 } 1326 1327 // Analyze operands of the call, assigning locations to each operand. 1328 SmallVector<CCValAssign, 16> ArgLocs; 1329 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1330 getTargetMachine(), ArgLocs, *DAG.getContext(), Call); 1331 CCInfo.AnalyzeCallOperands(Outs, 1332 CCAssignFnForNode(CallConv, /* Return*/ false, 1333 isVarArg)); 1334 1335 // Get a count of how many bytes are to be pushed on the stack. 1336 unsigned NumBytes = CCInfo.getNextStackOffset(); 1337 1338 // For tail calls, memory operands are available in our caller's stack. 1339 if (IsSibCall) 1340 NumBytes = 0; 1341 1342 // Adjust the stack pointer for the new arguments... 1343 // These operations are automatically eliminated by the prolog/epilog pass 1344 if (!IsSibCall) 1345 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true)); 1346 1347 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 1348 1349 RegsToPassVector RegsToPass; 1350 SmallVector<SDValue, 8> MemOpChains; 1351 1352 // Walk the register/memloc assignments, inserting copies/loads. In the case 1353 // of tail call optimization, arguments are handled later. 1354 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); 1355 i != e; 1356 ++i, ++realArgIdx) { 1357 CCValAssign &VA = ArgLocs[i]; 1358 SDValue Arg = OutVals[realArgIdx]; 1359 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags; 1360 bool isByVal = Flags.isByVal(); 1361 1362 // Promote the value if needed. 1363 switch (VA.getLocInfo()) { 1364 default: llvm_unreachable("Unknown loc info!"); 1365 case CCValAssign::Full: break; 1366 case CCValAssign::SExt: 1367 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); 1368 break; 1369 case CCValAssign::ZExt: 1370 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); 1371 break; 1372 case CCValAssign::AExt: 1373 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg); 1374 break; 1375 case CCValAssign::BCvt: 1376 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg); 1377 break; 1378 } 1379 1380 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces 1381 if (VA.needsCustom()) { 1382 if (VA.getLocVT() == MVT::v2f64) { 1383 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1384 DAG.getConstant(0, MVT::i32)); 1385 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1386 DAG.getConstant(1, MVT::i32)); 1387 1388 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass, 1389 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags); 1390 1391 VA = ArgLocs[++i]; // skip ahead to next loc 1392 if (VA.isRegLoc()) { 1393 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass, 1394 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags); 1395 } else { 1396 assert(VA.isMemLoc()); 1397 1398 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1, 1399 dl, DAG, VA, Flags)); 1400 } 1401 } else { 1402 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i], 1403 StackPtr, MemOpChains, Flags); 1404 } 1405 } else if (VA.isRegLoc()) { 1406 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 1407 } else if (isByVal) { 1408 assert(VA.isMemLoc()); 1409 unsigned offset = 0; 1410 1411 // True if this byval aggregate will be split between registers 1412 // and memory. 1413 if (CCInfo.isFirstByValRegValid()) { 1414 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 1415 unsigned int i, j; 1416 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) { 1417 SDValue Const = DAG.getConstant(4*i, MVT::i32); 1418 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const); 1419 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg, 1420 MachinePointerInfo(), 1421 false, false, false, 0); 1422 MemOpChains.push_back(Load.getValue(1)); 1423 RegsToPass.push_back(std::make_pair(j, Load)); 1424 } 1425 offset = ARM::R4 - CCInfo.getFirstByValReg(); 1426 CCInfo.clearFirstByValReg(); 1427 } 1428 1429 if (Flags.getByValSize() - 4*offset > 0) { 1430 unsigned LocMemOffset = VA.getLocMemOffset(); 1431 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset); 1432 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, 1433 StkPtrOff); 1434 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset); 1435 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset); 1436 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, 1437 MVT::i32); 1438 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32); 1439 1440 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue); 1441 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode}; 1442 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs, 1443 Ops, array_lengthof(Ops))); 1444 } 1445 } else if (!IsSibCall) { 1446 assert(VA.isMemLoc()); 1447 1448 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg, 1449 dl, DAG, VA, Flags)); 1450 } 1451 } 1452 1453 if (!MemOpChains.empty()) 1454 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 1455 &MemOpChains[0], MemOpChains.size()); 1456 1457 // Build a sequence of copy-to-reg nodes chained together with token chain 1458 // and flag operands which copy the outgoing args into the appropriate regs. 1459 SDValue InFlag; 1460 // Tail call byval lowering might overwrite argument registers so in case of 1461 // tail call optimization the copies to registers are lowered later. 1462 if (!isTailCall) 1463 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 1464 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, 1465 RegsToPass[i].second, InFlag); 1466 InFlag = Chain.getValue(1); 1467 } 1468 1469 // For tail calls lower the arguments to the 'real' stack slot. 1470 if (isTailCall) { 1471 // Force all the incoming stack arguments to be loaded from the stack 1472 // before any new outgoing arguments are stored to the stack, because the 1473 // outgoing stack slots may alias the incoming argument stack slots, and 1474 // the alias isn't otherwise explicit. This is slightly more conservative 1475 // than necessary, because it means that each store effectively depends 1476 // on every argument instead of just those arguments it would clobber. 1477 1478 // Do not flag preceding copytoreg stuff together with the following stuff. 1479 InFlag = SDValue(); 1480 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 1481 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, 1482 RegsToPass[i].second, InFlag); 1483 InFlag = Chain.getValue(1); 1484 } 1485 InFlag =SDValue(); 1486 } 1487 1488 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 1489 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 1490 // node so that legalize doesn't hack it. 1491 bool isDirect = false; 1492 bool isARMFunc = false; 1493 bool isLocalARMFunc = false; 1494 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1495 1496 if (EnableARMLongCalls) { 1497 assert (getTargetMachine().getRelocationModel() == Reloc::Static 1498 && "long-calls with non-static relocation model!"); 1499 // Handle a global address or an external symbol. If it's not one of 1500 // those, the target's already in a register, so we don't need to do 1501 // anything extra. 1502 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 1503 const GlobalValue *GV = G->getGlobal(); 1504 // Create a constant pool entry for the callee address 1505 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1506 ARMConstantPoolValue *CPV = 1507 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0); 1508 1509 // Get the address of the callee into a register 1510 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1511 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1512 Callee = DAG.getLoad(getPointerTy(), dl, 1513 DAG.getEntryNode(), CPAddr, 1514 MachinePointerInfo::getConstantPool(), 1515 false, false, false, 0); 1516 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) { 1517 const char *Sym = S->getSymbol(); 1518 1519 // Create a constant pool entry for the callee address 1520 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1521 ARMConstantPoolValue *CPV = 1522 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym, 1523 ARMPCLabelIndex, 0); 1524 // Get the address of the callee into a register 1525 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1526 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1527 Callee = DAG.getLoad(getPointerTy(), dl, 1528 DAG.getEntryNode(), CPAddr, 1529 MachinePointerInfo::getConstantPool(), 1530 false, false, false, 0); 1531 } 1532 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 1533 const GlobalValue *GV = G->getGlobal(); 1534 isDirect = true; 1535 bool isExt = GV->isDeclaration() || GV->isWeakForLinker(); 1536 bool isStub = (isExt && Subtarget->isTargetDarwin()) && 1537 getTargetMachine().getRelocationModel() != Reloc::Static; 1538 isARMFunc = !Subtarget->isThumb() || isStub; 1539 // ARM call to a local ARM function is predicable. 1540 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking); 1541 // tBX takes a register source operand. 1542 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) { 1543 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1544 ARMConstantPoolValue *CPV = 1545 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4); 1546 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1547 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1548 Callee = DAG.getLoad(getPointerTy(), dl, 1549 DAG.getEntryNode(), CPAddr, 1550 MachinePointerInfo::getConstantPool(), 1551 false, false, false, 0); 1552 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 1553 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, 1554 getPointerTy(), Callee, PICLabel); 1555 } else { 1556 // On ELF targets for PIC code, direct calls should go through the PLT 1557 unsigned OpFlags = 0; 1558 if (Subtarget->isTargetELF() && 1559 getTargetMachine().getRelocationModel() == Reloc::PIC_) 1560 OpFlags = ARMII::MO_PLT; 1561 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags); 1562 } 1563 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 1564 isDirect = true; 1565 bool isStub = Subtarget->isTargetDarwin() && 1566 getTargetMachine().getRelocationModel() != Reloc::Static; 1567 isARMFunc = !Subtarget->isThumb() || isStub; 1568 // tBX takes a register source operand. 1569 const char *Sym = S->getSymbol(); 1570 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) { 1571 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1572 ARMConstantPoolValue *CPV = 1573 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym, 1574 ARMPCLabelIndex, 4); 1575 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1576 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1577 Callee = DAG.getLoad(getPointerTy(), dl, 1578 DAG.getEntryNode(), CPAddr, 1579 MachinePointerInfo::getConstantPool(), 1580 false, false, false, 0); 1581 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 1582 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, 1583 getPointerTy(), Callee, PICLabel); 1584 } else { 1585 unsigned OpFlags = 0; 1586 // On ELF targets for PIC code, direct calls should go through the PLT 1587 if (Subtarget->isTargetELF() && 1588 getTargetMachine().getRelocationModel() == Reloc::PIC_) 1589 OpFlags = ARMII::MO_PLT; 1590 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags); 1591 } 1592 } 1593 1594 // FIXME: handle tail calls differently. 1595 unsigned CallOpc; 1596 if (Subtarget->isThumb()) { 1597 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps()) 1598 CallOpc = ARMISD::CALL_NOLINK; 1599 else if (doesNotRet && isDirect && !isARMFunc && 1600 Subtarget->hasRAS() && !Subtarget->isThumb1Only()) 1601 // "mov lr, pc; b _foo" to avoid confusing the RSP 1602 CallOpc = ARMISD::CALL_NOLINK; 1603 else 1604 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL; 1605 } else { 1606 if (!isDirect && !Subtarget->hasV5TOps()) { 1607 CallOpc = ARMISD::CALL_NOLINK; 1608 } else if (doesNotRet && isDirect && Subtarget->hasRAS()) 1609 // "mov lr, pc; b _foo" to avoid confusing the RSP 1610 CallOpc = ARMISD::CALL_NOLINK; 1611 else 1612 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL; 1613 } 1614 1615 std::vector<SDValue> Ops; 1616 Ops.push_back(Chain); 1617 Ops.push_back(Callee); 1618 1619 // Add argument registers to the end of the list so that they are known live 1620 // into the call. 1621 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 1622 Ops.push_back(DAG.getRegister(RegsToPass[i].first, 1623 RegsToPass[i].second.getValueType())); 1624 1625 // Add a register mask operand representing the call-preserved registers. 1626 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo(); 1627 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv); 1628 assert(Mask && "Missing call preserved mask for calling convention"); 1629 Ops.push_back(DAG.getRegisterMask(Mask)); 1630 1631 if (InFlag.getNode()) 1632 Ops.push_back(InFlag); 1633 1634 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 1635 if (isTailCall) 1636 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size()); 1637 1638 // Returns a chain and a flag for retval copy to use. 1639 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size()); 1640 InFlag = Chain.getValue(1); 1641 1642 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true), 1643 DAG.getIntPtrConstant(0, true), InFlag); 1644 if (!Ins.empty()) 1645 InFlag = Chain.getValue(1); 1646 1647 // Handle result values, copying them out of physregs into vregs that we 1648 // return. 1649 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, 1650 dl, DAG, InVals); 1651 } 1652 1653 /// HandleByVal - Every parameter *after* a byval parameter is passed 1654 /// on the stack. Remember the next parameter register to allocate, 1655 /// and then confiscate the rest of the parameter registers to insure 1656 /// this. 1657 void 1658 ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const { 1659 unsigned reg = State->AllocateReg(GPRArgRegs, 4); 1660 assert((State->getCallOrPrologue() == Prologue || 1661 State->getCallOrPrologue() == Call) && 1662 "unhandled ParmContext"); 1663 if ((!State->isFirstByValRegValid()) && 1664 (ARM::R0 <= reg) && (reg <= ARM::R3)) { 1665 State->setFirstByValReg(reg); 1666 // At a call site, a byval parameter that is split between 1667 // registers and memory needs its size truncated here. In a 1668 // function prologue, such byval parameters are reassembled in 1669 // memory, and are not truncated. 1670 if (State->getCallOrPrologue() == Call) { 1671 unsigned excess = 4 * (ARM::R4 - reg); 1672 assert(size >= excess && "expected larger existing stack allocation"); 1673 size -= excess; 1674 } 1675 } 1676 // Confiscate any remaining parameter registers to preclude their 1677 // assignment to subsequent parameters. 1678 while (State->AllocateReg(GPRArgRegs, 4)) 1679 ; 1680 } 1681 1682 /// MatchingStackOffset - Return true if the given stack call argument is 1683 /// already available in the same position (relatively) of the caller's 1684 /// incoming argument stack. 1685 static 1686 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags, 1687 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI, 1688 const TargetInstrInfo *TII) { 1689 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8; 1690 int FI = INT_MAX; 1691 if (Arg.getOpcode() == ISD::CopyFromReg) { 1692 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg(); 1693 if (!TargetRegisterInfo::isVirtualRegister(VR)) 1694 return false; 1695 MachineInstr *Def = MRI->getVRegDef(VR); 1696 if (!Def) 1697 return false; 1698 if (!Flags.isByVal()) { 1699 if (!TII->isLoadFromStackSlot(Def, FI)) 1700 return false; 1701 } else { 1702 return false; 1703 } 1704 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) { 1705 if (Flags.isByVal()) 1706 // ByVal argument is passed in as a pointer but it's now being 1707 // dereferenced. e.g. 1708 // define @foo(%struct.X* %A) { 1709 // tail call @bar(%struct.X* byval %A) 1710 // } 1711 return false; 1712 SDValue Ptr = Ld->getBasePtr(); 1713 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr); 1714 if (!FINode) 1715 return false; 1716 FI = FINode->getIndex(); 1717 } else 1718 return false; 1719 1720 assert(FI != INT_MAX); 1721 if (!MFI->isFixedObjectIndex(FI)) 1722 return false; 1723 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI); 1724 } 1725 1726 /// IsEligibleForTailCallOptimization - Check whether the call is eligible 1727 /// for tail call optimization. Targets which want to do tail call 1728 /// optimization should implement this function. 1729 bool 1730 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee, 1731 CallingConv::ID CalleeCC, 1732 bool isVarArg, 1733 bool isCalleeStructRet, 1734 bool isCallerStructRet, 1735 const SmallVectorImpl<ISD::OutputArg> &Outs, 1736 const SmallVectorImpl<SDValue> &OutVals, 1737 const SmallVectorImpl<ISD::InputArg> &Ins, 1738 SelectionDAG& DAG) const { 1739 const Function *CallerF = DAG.getMachineFunction().getFunction(); 1740 CallingConv::ID CallerCC = CallerF->getCallingConv(); 1741 bool CCMatch = CallerCC == CalleeCC; 1742 1743 // Look for obvious safe cases to perform tail call optimization that do not 1744 // require ABI changes. This is what gcc calls sibcall. 1745 1746 // Do not sibcall optimize vararg calls unless the call site is not passing 1747 // any arguments. 1748 if (isVarArg && !Outs.empty()) 1749 return false; 1750 1751 // Also avoid sibcall optimization if either caller or callee uses struct 1752 // return semantics. 1753 if (isCalleeStructRet || isCallerStructRet) 1754 return false; 1755 1756 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo:: 1757 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as 1758 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation 1759 // support in the assembler and linker to be used. This would need to be 1760 // fixed to fully support tail calls in Thumb1. 1761 // 1762 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take 1763 // LR. This means if we need to reload LR, it takes an extra instructions, 1764 // which outweighs the value of the tail call; but here we don't know yet 1765 // whether LR is going to be used. Probably the right approach is to 1766 // generate the tail call here and turn it back into CALL/RET in 1767 // emitEpilogue if LR is used. 1768 1769 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls, 1770 // but we need to make sure there are enough registers; the only valid 1771 // registers are the 4 used for parameters. We don't currently do this 1772 // case. 1773 if (Subtarget->isThumb1Only()) 1774 return false; 1775 1776 // If the calling conventions do not match, then we'd better make sure the 1777 // results are returned in the same way as what the caller expects. 1778 if (!CCMatch) { 1779 SmallVector<CCValAssign, 16> RVLocs1; 1780 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), 1781 getTargetMachine(), RVLocs1, *DAG.getContext(), Call); 1782 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg)); 1783 1784 SmallVector<CCValAssign, 16> RVLocs2; 1785 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), 1786 getTargetMachine(), RVLocs2, *DAG.getContext(), Call); 1787 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg)); 1788 1789 if (RVLocs1.size() != RVLocs2.size()) 1790 return false; 1791 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) { 1792 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc()) 1793 return false; 1794 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo()) 1795 return false; 1796 if (RVLocs1[i].isRegLoc()) { 1797 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg()) 1798 return false; 1799 } else { 1800 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset()) 1801 return false; 1802 } 1803 } 1804 } 1805 1806 // If the callee takes no arguments then go on to check the results of the 1807 // call. 1808 if (!Outs.empty()) { 1809 // Check if stack adjustment is needed. For now, do not do this if any 1810 // argument is passed on the stack. 1811 SmallVector<CCValAssign, 16> ArgLocs; 1812 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), 1813 getTargetMachine(), ArgLocs, *DAG.getContext(), Call); 1814 CCInfo.AnalyzeCallOperands(Outs, 1815 CCAssignFnForNode(CalleeCC, false, isVarArg)); 1816 if (CCInfo.getNextStackOffset()) { 1817 MachineFunction &MF = DAG.getMachineFunction(); 1818 1819 // Check if the arguments are already laid out in the right way as 1820 // the caller's fixed stack objects. 1821 MachineFrameInfo *MFI = MF.getFrameInfo(); 1822 const MachineRegisterInfo *MRI = &MF.getRegInfo(); 1823 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1824 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); 1825 i != e; 1826 ++i, ++realArgIdx) { 1827 CCValAssign &VA = ArgLocs[i]; 1828 EVT RegVT = VA.getLocVT(); 1829 SDValue Arg = OutVals[realArgIdx]; 1830 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags; 1831 if (VA.getLocInfo() == CCValAssign::Indirect) 1832 return false; 1833 if (VA.needsCustom()) { 1834 // f64 and vector types are split into multiple registers or 1835 // register/stack-slot combinations. The types will not match 1836 // the registers; give up on memory f64 refs until we figure 1837 // out what to do about this. 1838 if (!VA.isRegLoc()) 1839 return false; 1840 if (!ArgLocs[++i].isRegLoc()) 1841 return false; 1842 if (RegVT == MVT::v2f64) { 1843 if (!ArgLocs[++i].isRegLoc()) 1844 return false; 1845 if (!ArgLocs[++i].isRegLoc()) 1846 return false; 1847 } 1848 } else if (!VA.isRegLoc()) { 1849 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags, 1850 MFI, MRI, TII)) 1851 return false; 1852 } 1853 } 1854 } 1855 } 1856 1857 return true; 1858 } 1859 1860 SDValue 1861 ARMTargetLowering::LowerReturn(SDValue Chain, 1862 CallingConv::ID CallConv, bool isVarArg, 1863 const SmallVectorImpl<ISD::OutputArg> &Outs, 1864 const SmallVectorImpl<SDValue> &OutVals, 1865 DebugLoc dl, SelectionDAG &DAG) const { 1866 1867 // CCValAssign - represent the assignment of the return value to a location. 1868 SmallVector<CCValAssign, 16> RVLocs; 1869 1870 // CCState - Info about the registers and stack slots. 1871 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1872 getTargetMachine(), RVLocs, *DAG.getContext(), Call); 1873 1874 // Analyze outgoing return values. 1875 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true, 1876 isVarArg)); 1877 1878 // If this is the first return lowered for this function, add 1879 // the regs to the liveout set for the function. 1880 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) { 1881 for (unsigned i = 0; i != RVLocs.size(); ++i) 1882 if (RVLocs[i].isRegLoc()) 1883 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg()); 1884 } 1885 1886 SDValue Flag; 1887 1888 // Copy the result values into the output registers. 1889 for (unsigned i = 0, realRVLocIdx = 0; 1890 i != RVLocs.size(); 1891 ++i, ++realRVLocIdx) { 1892 CCValAssign &VA = RVLocs[i]; 1893 assert(VA.isRegLoc() && "Can only return in registers!"); 1894 1895 SDValue Arg = OutVals[realRVLocIdx]; 1896 1897 switch (VA.getLocInfo()) { 1898 default: llvm_unreachable("Unknown loc info!"); 1899 case CCValAssign::Full: break; 1900 case CCValAssign::BCvt: 1901 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg); 1902 break; 1903 } 1904 1905 if (VA.needsCustom()) { 1906 if (VA.getLocVT() == MVT::v2f64) { 1907 // Extract the first half and return it in two registers. 1908 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1909 DAG.getConstant(0, MVT::i32)); 1910 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl, 1911 DAG.getVTList(MVT::i32, MVT::i32), Half); 1912 1913 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag); 1914 Flag = Chain.getValue(1); 1915 VA = RVLocs[++i]; // skip ahead to next loc 1916 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), 1917 HalfGPRs.getValue(1), Flag); 1918 Flag = Chain.getValue(1); 1919 VA = RVLocs[++i]; // skip ahead to next loc 1920 1921 // Extract the 2nd half and fall through to handle it as an f64 value. 1922 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1923 DAG.getConstant(1, MVT::i32)); 1924 } 1925 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is 1926 // available. 1927 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl, 1928 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1); 1929 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag); 1930 Flag = Chain.getValue(1); 1931 VA = RVLocs[++i]; // skip ahead to next loc 1932 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1), 1933 Flag); 1934 } else 1935 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag); 1936 1937 // Guarantee that all emitted copies are 1938 // stuck together, avoiding something bad. 1939 Flag = Chain.getValue(1); 1940 } 1941 1942 SDValue result; 1943 if (Flag.getNode()) 1944 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag); 1945 else // Return Void 1946 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain); 1947 1948 return result; 1949 } 1950 1951 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const { 1952 if (N->getNumValues() != 1) 1953 return false; 1954 if (!N->hasNUsesOfValue(1, 0)) 1955 return false; 1956 1957 SDValue TCChain = Chain; 1958 SDNode *Copy = *N->use_begin(); 1959 if (Copy->getOpcode() == ISD::CopyToReg) { 1960 // If the copy has a glue operand, we conservatively assume it isn't safe to 1961 // perform a tail call. 1962 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue) 1963 return false; 1964 TCChain = Copy->getOperand(0); 1965 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) { 1966 SDNode *VMov = Copy; 1967 // f64 returned in a pair of GPRs. 1968 SmallPtrSet<SDNode*, 2> Copies; 1969 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end(); 1970 UI != UE; ++UI) { 1971 if (UI->getOpcode() != ISD::CopyToReg) 1972 return false; 1973 Copies.insert(*UI); 1974 } 1975 if (Copies.size() > 2) 1976 return false; 1977 1978 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end(); 1979 UI != UE; ++UI) { 1980 SDValue UseChain = UI->getOperand(0); 1981 if (Copies.count(UseChain.getNode())) 1982 // Second CopyToReg 1983 Copy = *UI; 1984 else 1985 // First CopyToReg 1986 TCChain = UseChain; 1987 } 1988 } else if (Copy->getOpcode() == ISD::BITCAST) { 1989 // f32 returned in a single GPR. 1990 if (!Copy->hasOneUse()) 1991 return false; 1992 Copy = *Copy->use_begin(); 1993 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0)) 1994 return false; 1995 Chain = Copy->getOperand(0); 1996 } else { 1997 return false; 1998 } 1999 2000 bool HasRet = false; 2001 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end(); 2002 UI != UE; ++UI) { 2003 if (UI->getOpcode() != ARMISD::RET_FLAG) 2004 return false; 2005 HasRet = true; 2006 } 2007 2008 if (!HasRet) 2009 return false; 2010 2011 Chain = TCChain; 2012 return true; 2013 } 2014 2015 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const { 2016 if (!EnableARMTailCalls && !Subtarget->supportsTailCall()) 2017 return false; 2018 2019 if (!CI->isTailCall()) 2020 return false; 2021 2022 return !Subtarget->isThumb1Only(); 2023 } 2024 2025 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as 2026 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is 2027 // one of the above mentioned nodes. It has to be wrapped because otherwise 2028 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only 2029 // be used to form addressing mode. These wrapped nodes will be selected 2030 // into MOVi. 2031 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) { 2032 EVT PtrVT = Op.getValueType(); 2033 // FIXME there is no actual debug info here 2034 DebugLoc dl = Op.getDebugLoc(); 2035 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); 2036 SDValue Res; 2037 if (CP->isMachineConstantPoolEntry()) 2038 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT, 2039 CP->getAlignment()); 2040 else 2041 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, 2042 CP->getAlignment()); 2043 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res); 2044 } 2045 2046 unsigned ARMTargetLowering::getJumpTableEncoding() const { 2047 return MachineJumpTableInfo::EK_Inline; 2048 } 2049 2050 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, 2051 SelectionDAG &DAG) const { 2052 MachineFunction &MF = DAG.getMachineFunction(); 2053 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2054 unsigned ARMPCLabelIndex = 0; 2055 DebugLoc DL = Op.getDebugLoc(); 2056 EVT PtrVT = getPointerTy(); 2057 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress(); 2058 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2059 SDValue CPAddr; 2060 if (RelocM == Reloc::Static) { 2061 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4); 2062 } else { 2063 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8; 2064 ARMPCLabelIndex = AFI->createPICLabelUId(); 2065 ARMConstantPoolValue *CPV = 2066 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex, 2067 ARMCP::CPBlockAddress, PCAdj); 2068 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2069 } 2070 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr); 2071 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr, 2072 MachinePointerInfo::getConstantPool(), 2073 false, false, false, 0); 2074 if (RelocM == Reloc::Static) 2075 return Result; 2076 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2077 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel); 2078 } 2079 2080 // Lower ISD::GlobalTLSAddress using the "general dynamic" model 2081 SDValue 2082 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, 2083 SelectionDAG &DAG) const { 2084 DebugLoc dl = GA->getDebugLoc(); 2085 EVT PtrVT = getPointerTy(); 2086 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8; 2087 MachineFunction &MF = DAG.getMachineFunction(); 2088 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2089 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2090 ARMConstantPoolValue *CPV = 2091 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex, 2092 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true); 2093 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2094 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument); 2095 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, 2096 MachinePointerInfo::getConstantPool(), 2097 false, false, false, 0); 2098 SDValue Chain = Argument.getValue(1); 2099 2100 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2101 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel); 2102 2103 // call __tls_get_addr. 2104 ArgListTy Args; 2105 ArgListEntry Entry; 2106 Entry.Node = Argument; 2107 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext()); 2108 Args.push_back(Entry); 2109 // FIXME: is there useful debug info available here? 2110 TargetLowering::CallLoweringInfo CLI(Chain, 2111 (Type *) Type::getInt32Ty(*DAG.getContext()), 2112 false, false, false, false, 2113 0, CallingConv::C, /*isTailCall=*/false, 2114 /*doesNotRet=*/false, /*isReturnValueUsed=*/true, 2115 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl); 2116 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); 2117 return CallResult.first; 2118 } 2119 2120 // Lower ISD::GlobalTLSAddress using the "initial exec" or 2121 // "local exec" model. 2122 SDValue 2123 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA, 2124 SelectionDAG &DAG, 2125 TLSModel::Model model) const { 2126 const GlobalValue *GV = GA->getGlobal(); 2127 DebugLoc dl = GA->getDebugLoc(); 2128 SDValue Offset; 2129 SDValue Chain = DAG.getEntryNode(); 2130 EVT PtrVT = getPointerTy(); 2131 // Get the Thread Pointer 2132 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT); 2133 2134 if (model == TLSModel::InitialExec) { 2135 MachineFunction &MF = DAG.getMachineFunction(); 2136 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2137 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2138 // Initial exec model. 2139 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8; 2140 ARMConstantPoolValue *CPV = 2141 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex, 2142 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, 2143 true); 2144 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2145 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset); 2146 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, 2147 MachinePointerInfo::getConstantPool(), 2148 false, false, false, 0); 2149 Chain = Offset.getValue(1); 2150 2151 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2152 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel); 2153 2154 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, 2155 MachinePointerInfo::getConstantPool(), 2156 false, false, false, 0); 2157 } else { 2158 // local exec model 2159 assert(model == TLSModel::LocalExec); 2160 ARMConstantPoolValue *CPV = 2161 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF); 2162 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2163 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset); 2164 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, 2165 MachinePointerInfo::getConstantPool(), 2166 false, false, false, 0); 2167 } 2168 2169 // The address of the thread local variable is the add of the thread 2170 // pointer with the offset of the variable. 2171 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset); 2172 } 2173 2174 SDValue 2175 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const { 2176 // TODO: implement the "local dynamic" model 2177 assert(Subtarget->isTargetELF() && 2178 "TLS not implemented for non-ELF targets"); 2179 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 2180 2181 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal()); 2182 2183 switch (model) { 2184 case TLSModel::GeneralDynamic: 2185 case TLSModel::LocalDynamic: 2186 return LowerToTLSGeneralDynamicModel(GA, DAG); 2187 case TLSModel::InitialExec: 2188 case TLSModel::LocalExec: 2189 return LowerToTLSExecModels(GA, DAG, model); 2190 } 2191 llvm_unreachable("bogus TLS model"); 2192 } 2193 2194 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op, 2195 SelectionDAG &DAG) const { 2196 EVT PtrVT = getPointerTy(); 2197 DebugLoc dl = Op.getDebugLoc(); 2198 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 2199 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2200 if (RelocM == Reloc::PIC_) { 2201 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility(); 2202 ARMConstantPoolValue *CPV = 2203 ARMConstantPoolConstant::Create(GV, 2204 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT); 2205 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2206 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2207 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), 2208 CPAddr, 2209 MachinePointerInfo::getConstantPool(), 2210 false, false, false, 0); 2211 SDValue Chain = Result.getValue(1); 2212 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT); 2213 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT); 2214 if (!UseGOTOFF) 2215 Result = DAG.getLoad(PtrVT, dl, Chain, Result, 2216 MachinePointerInfo::getGOT(), 2217 false, false, false, 0); 2218 return Result; 2219 } 2220 2221 // If we have T2 ops, we can materialize the address directly via movt/movw 2222 // pair. This is always cheaper. 2223 if (Subtarget->useMovt()) { 2224 ++NumMovwMovt; 2225 // FIXME: Once remat is capable of dealing with instructions with register 2226 // operands, expand this into two nodes. 2227 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT, 2228 DAG.getTargetGlobalAddress(GV, dl, PtrVT)); 2229 } else { 2230 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4); 2231 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2232 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2233 MachinePointerInfo::getConstantPool(), 2234 false, false, false, 0); 2235 } 2236 } 2237 2238 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op, 2239 SelectionDAG &DAG) const { 2240 EVT PtrVT = getPointerTy(); 2241 DebugLoc dl = Op.getDebugLoc(); 2242 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 2243 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2244 MachineFunction &MF = DAG.getMachineFunction(); 2245 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2246 2247 // FIXME: Enable this for static codegen when tool issues are fixed. Also 2248 // update ARMFastISel::ARMMaterializeGV. 2249 if (Subtarget->useMovt() && RelocM != Reloc::Static) { 2250 ++NumMovwMovt; 2251 // FIXME: Once remat is capable of dealing with instructions with register 2252 // operands, expand this into two nodes. 2253 if (RelocM == Reloc::Static) 2254 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT, 2255 DAG.getTargetGlobalAddress(GV, dl, PtrVT)); 2256 2257 unsigned Wrapper = (RelocM == Reloc::PIC_) 2258 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN; 2259 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, 2260 DAG.getTargetGlobalAddress(GV, dl, PtrVT)); 2261 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) 2262 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result, 2263 MachinePointerInfo::getGOT(), 2264 false, false, false, 0); 2265 return Result; 2266 } 2267 2268 unsigned ARMPCLabelIndex = 0; 2269 SDValue CPAddr; 2270 if (RelocM == Reloc::Static) { 2271 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4); 2272 } else { 2273 ARMPCLabelIndex = AFI->createPICLabelUId(); 2274 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8); 2275 ARMConstantPoolValue *CPV = 2276 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 2277 PCAdj); 2278 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2279 } 2280 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2281 2282 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2283 MachinePointerInfo::getConstantPool(), 2284 false, false, false, 0); 2285 SDValue Chain = Result.getValue(1); 2286 2287 if (RelocM == Reloc::PIC_) { 2288 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2289 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 2290 } 2291 2292 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) 2293 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(), 2294 false, false, false, 0); 2295 2296 return Result; 2297 } 2298 2299 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op, 2300 SelectionDAG &DAG) const { 2301 assert(Subtarget->isTargetELF() && 2302 "GLOBAL OFFSET TABLE not implemented for non-ELF targets"); 2303 MachineFunction &MF = DAG.getMachineFunction(); 2304 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2305 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2306 EVT PtrVT = getPointerTy(); 2307 DebugLoc dl = Op.getDebugLoc(); 2308 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8; 2309 ARMConstantPoolValue *CPV = 2310 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_", 2311 ARMPCLabelIndex, PCAdj); 2312 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2313 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2314 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2315 MachinePointerInfo::getConstantPool(), 2316 false, false, false, 0); 2317 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2318 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 2319 } 2320 2321 SDValue 2322 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const { 2323 DebugLoc dl = Op.getDebugLoc(); 2324 SDValue Val = DAG.getConstant(0, MVT::i32); 2325 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, 2326 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0), 2327 Op.getOperand(1), Val); 2328 } 2329 2330 SDValue 2331 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const { 2332 DebugLoc dl = Op.getDebugLoc(); 2333 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0), 2334 Op.getOperand(1), DAG.getConstant(0, MVT::i32)); 2335 } 2336 2337 SDValue 2338 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG, 2339 const ARMSubtarget *Subtarget) const { 2340 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 2341 DebugLoc dl = Op.getDebugLoc(); 2342 switch (IntNo) { 2343 default: return SDValue(); // Don't custom lower most intrinsics. 2344 case Intrinsic::arm_thread_pointer: { 2345 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 2346 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT); 2347 } 2348 case Intrinsic::eh_sjlj_lsda: { 2349 MachineFunction &MF = DAG.getMachineFunction(); 2350 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2351 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2352 EVT PtrVT = getPointerTy(); 2353 DebugLoc dl = Op.getDebugLoc(); 2354 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2355 SDValue CPAddr; 2356 unsigned PCAdj = (RelocM != Reloc::PIC_) 2357 ? 0 : (Subtarget->isThumb() ? 4 : 8); 2358 ARMConstantPoolValue *CPV = 2359 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex, 2360 ARMCP::CPLSDA, PCAdj); 2361 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2362 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2363 SDValue Result = 2364 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2365 MachinePointerInfo::getConstantPool(), 2366 false, false, false, 0); 2367 2368 if (RelocM == Reloc::PIC_) { 2369 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2370 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 2371 } 2372 return Result; 2373 } 2374 case Intrinsic::arm_neon_vmulls: 2375 case Intrinsic::arm_neon_vmullu: { 2376 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls) 2377 ? ARMISD::VMULLs : ARMISD::VMULLu; 2378 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(), 2379 Op.getOperand(1), Op.getOperand(2)); 2380 } 2381 } 2382 } 2383 2384 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG, 2385 const ARMSubtarget *Subtarget) { 2386 DebugLoc dl = Op.getDebugLoc(); 2387 if (!Subtarget->hasDataBarrier()) { 2388 // Some ARMv6 cpus can support data barriers with an mcr instruction. 2389 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get 2390 // here. 2391 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() && 2392 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!"); 2393 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0), 2394 DAG.getConstant(0, MVT::i32)); 2395 } 2396 2397 SDValue Op5 = Op.getOperand(5); 2398 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0; 2399 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); 2400 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue(); 2401 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0); 2402 2403 ARM_MB::MemBOpt DMBOpt; 2404 if (isDeviceBarrier) 2405 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY; 2406 else 2407 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH; 2408 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0), 2409 DAG.getConstant(DMBOpt, MVT::i32)); 2410 } 2411 2412 2413 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG, 2414 const ARMSubtarget *Subtarget) { 2415 // FIXME: handle "fence singlethread" more efficiently. 2416 DebugLoc dl = Op.getDebugLoc(); 2417 if (!Subtarget->hasDataBarrier()) { 2418 // Some ARMv6 cpus can support data barriers with an mcr instruction. 2419 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get 2420 // here. 2421 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() && 2422 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!"); 2423 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0), 2424 DAG.getConstant(0, MVT::i32)); 2425 } 2426 2427 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0), 2428 DAG.getConstant(ARM_MB::ISH, MVT::i32)); 2429 } 2430 2431 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG, 2432 const ARMSubtarget *Subtarget) { 2433 // ARM pre v5TE and Thumb1 does not have preload instructions. 2434 if (!(Subtarget->isThumb2() || 2435 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps()))) 2436 // Just preserve the chain. 2437 return Op.getOperand(0); 2438 2439 DebugLoc dl = Op.getDebugLoc(); 2440 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1; 2441 if (!isRead && 2442 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension())) 2443 // ARMv7 with MP extension has PLDW. 2444 return Op.getOperand(0); 2445 2446 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue(); 2447 if (Subtarget->isThumb()) { 2448 // Invert the bits. 2449 isRead = ~isRead & 1; 2450 isData = ~isData & 1; 2451 } 2452 2453 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0), 2454 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32), 2455 DAG.getConstant(isData, MVT::i32)); 2456 } 2457 2458 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) { 2459 MachineFunction &MF = DAG.getMachineFunction(); 2460 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>(); 2461 2462 // vastart just stores the address of the VarArgsFrameIndex slot into the 2463 // memory location argument. 2464 DebugLoc dl = Op.getDebugLoc(); 2465 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 2466 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); 2467 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 2468 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), 2469 MachinePointerInfo(SV), false, false, 0); 2470 } 2471 2472 SDValue 2473 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA, 2474 SDValue &Root, SelectionDAG &DAG, 2475 DebugLoc dl) const { 2476 MachineFunction &MF = DAG.getMachineFunction(); 2477 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2478 2479 const TargetRegisterClass *RC; 2480 if (AFI->isThumb1OnlyFunction()) 2481 RC = &ARM::tGPRRegClass; 2482 else 2483 RC = &ARM::GPRRegClass; 2484 2485 // Transform the arguments stored in physical registers into virtual ones. 2486 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); 2487 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32); 2488 2489 SDValue ArgValue2; 2490 if (NextVA.isMemLoc()) { 2491 MachineFrameInfo *MFI = MF.getFrameInfo(); 2492 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true); 2493 2494 // Create load node to retrieve arguments from the stack. 2495 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2496 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, 2497 MachinePointerInfo::getFixedStack(FI), 2498 false, false, false, 0); 2499 } else { 2500 Reg = MF.addLiveIn(NextVA.getLocReg(), RC); 2501 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32); 2502 } 2503 2504 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2); 2505 } 2506 2507 void 2508 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF, 2509 unsigned &VARegSize, unsigned &VARegSaveSize) 2510 const { 2511 unsigned NumGPRs; 2512 if (CCInfo.isFirstByValRegValid()) 2513 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg(); 2514 else { 2515 unsigned int firstUnalloced; 2516 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs, 2517 sizeof(GPRArgRegs) / 2518 sizeof(GPRArgRegs[0])); 2519 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0; 2520 } 2521 2522 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment(); 2523 VARegSize = NumGPRs * 4; 2524 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1); 2525 } 2526 2527 // The remaining GPRs hold either the beginning of variable-argument 2528 // data, or the beginning of an aggregate passed by value (usuall 2529 // byval). Either way, we allocate stack slots adjacent to the data 2530 // provided by our caller, and store the unallocated registers there. 2531 // If this is a variadic function, the va_list pointer will begin with 2532 // these values; otherwise, this reassembles a (byval) structure that 2533 // was split between registers and memory. 2534 void 2535 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG, 2536 DebugLoc dl, SDValue &Chain, 2537 unsigned ArgOffset) const { 2538 MachineFunction &MF = DAG.getMachineFunction(); 2539 MachineFrameInfo *MFI = MF.getFrameInfo(); 2540 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2541 unsigned firstRegToSaveIndex; 2542 if (CCInfo.isFirstByValRegValid()) 2543 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0; 2544 else { 2545 firstRegToSaveIndex = CCInfo.getFirstUnallocated 2546 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0])); 2547 } 2548 2549 unsigned VARegSize, VARegSaveSize; 2550 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize); 2551 if (VARegSaveSize) { 2552 // If this function is vararg, store any remaining integer argument regs 2553 // to their spots on the stack so that they may be loaded by deferencing 2554 // the result of va_next. 2555 AFI->setVarArgsRegSaveSize(VARegSaveSize); 2556 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize, 2557 ArgOffset + VARegSaveSize 2558 - VARegSize, 2559 false)); 2560 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(), 2561 getPointerTy()); 2562 2563 SmallVector<SDValue, 4> MemOps; 2564 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) { 2565 const TargetRegisterClass *RC; 2566 if (AFI->isThumb1OnlyFunction()) 2567 RC = &ARM::tGPRRegClass; 2568 else 2569 RC = &ARM::GPRRegClass; 2570 2571 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC); 2572 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32); 2573 SDValue Store = 2574 DAG.getStore(Val.getValue(1), dl, Val, FIN, 2575 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()), 2576 false, false, 0); 2577 MemOps.push_back(Store); 2578 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN, 2579 DAG.getConstant(4, getPointerTy())); 2580 } 2581 if (!MemOps.empty()) 2582 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 2583 &MemOps[0], MemOps.size()); 2584 } else 2585 // This will point to the next argument passed via stack. 2586 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true)); 2587 } 2588 2589 SDValue 2590 ARMTargetLowering::LowerFormalArguments(SDValue Chain, 2591 CallingConv::ID CallConv, bool isVarArg, 2592 const SmallVectorImpl<ISD::InputArg> 2593 &Ins, 2594 DebugLoc dl, SelectionDAG &DAG, 2595 SmallVectorImpl<SDValue> &InVals) 2596 const { 2597 MachineFunction &MF = DAG.getMachineFunction(); 2598 MachineFrameInfo *MFI = MF.getFrameInfo(); 2599 2600 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2601 2602 // Assign locations to all of the incoming arguments. 2603 SmallVector<CCValAssign, 16> ArgLocs; 2604 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 2605 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue); 2606 CCInfo.AnalyzeFormalArguments(Ins, 2607 CCAssignFnForNode(CallConv, /* Return*/ false, 2608 isVarArg)); 2609 2610 SmallVector<SDValue, 16> ArgValues; 2611 int lastInsIndex = -1; 2612 2613 SDValue ArgValue; 2614 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 2615 CCValAssign &VA = ArgLocs[i]; 2616 2617 // Arguments stored in registers. 2618 if (VA.isRegLoc()) { 2619 EVT RegVT = VA.getLocVT(); 2620 2621 if (VA.needsCustom()) { 2622 // f64 and vector types are split up into multiple registers or 2623 // combinations of registers and stack slots. 2624 if (VA.getLocVT() == MVT::v2f64) { 2625 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i], 2626 Chain, DAG, dl); 2627 VA = ArgLocs[++i]; // skip ahead to next loc 2628 SDValue ArgValue2; 2629 if (VA.isMemLoc()) { 2630 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true); 2631 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2632 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN, 2633 MachinePointerInfo::getFixedStack(FI), 2634 false, false, false, 0); 2635 } else { 2636 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i], 2637 Chain, DAG, dl); 2638 } 2639 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64); 2640 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, 2641 ArgValue, ArgValue1, DAG.getIntPtrConstant(0)); 2642 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, 2643 ArgValue, ArgValue2, DAG.getIntPtrConstant(1)); 2644 } else 2645 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl); 2646 2647 } else { 2648 const TargetRegisterClass *RC; 2649 2650 if (RegVT == MVT::f32) 2651 RC = &ARM::SPRRegClass; 2652 else if (RegVT == MVT::f64) 2653 RC = &ARM::DPRRegClass; 2654 else if (RegVT == MVT::v2f64) 2655 RC = &ARM::QPRRegClass; 2656 else if (RegVT == MVT::i32) 2657 RC = AFI->isThumb1OnlyFunction() ? 2658 (const TargetRegisterClass*)&ARM::tGPRRegClass : 2659 (const TargetRegisterClass*)&ARM::GPRRegClass; 2660 else 2661 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering"); 2662 2663 // Transform the arguments in physical registers into virtual ones. 2664 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); 2665 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT); 2666 } 2667 2668 // If this is an 8 or 16-bit value, it is really passed promoted 2669 // to 32 bits. Insert an assert[sz]ext to capture this, then 2670 // truncate to the right size. 2671 switch (VA.getLocInfo()) { 2672 default: llvm_unreachable("Unknown loc info!"); 2673 case CCValAssign::Full: break; 2674 case CCValAssign::BCvt: 2675 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue); 2676 break; 2677 case CCValAssign::SExt: 2678 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue, 2679 DAG.getValueType(VA.getValVT())); 2680 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 2681 break; 2682 case CCValAssign::ZExt: 2683 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue, 2684 DAG.getValueType(VA.getValVT())); 2685 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 2686 break; 2687 } 2688 2689 InVals.push_back(ArgValue); 2690 2691 } else { // VA.isRegLoc() 2692 2693 // sanity check 2694 assert(VA.isMemLoc()); 2695 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered"); 2696 2697 int index = ArgLocs[i].getValNo(); 2698 2699 // Some Ins[] entries become multiple ArgLoc[] entries. 2700 // Process them only once. 2701 if (index != lastInsIndex) 2702 { 2703 ISD::ArgFlagsTy Flags = Ins[index].Flags; 2704 // FIXME: For now, all byval parameter objects are marked mutable. 2705 // This can be changed with more analysis. 2706 // In case of tail call optimization mark all arguments mutable. 2707 // Since they could be overwritten by lowering of arguments in case of 2708 // a tail call. 2709 if (Flags.isByVal()) { 2710 unsigned VARegSize, VARegSaveSize; 2711 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize); 2712 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0); 2713 unsigned Bytes = Flags.getByValSize() - VARegSize; 2714 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects. 2715 int FI = MFI->CreateFixedObject(Bytes, 2716 VA.getLocMemOffset(), false); 2717 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy())); 2718 } else { 2719 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8, 2720 VA.getLocMemOffset(), true); 2721 2722 // Create load nodes to retrieve arguments from the stack. 2723 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2724 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, 2725 MachinePointerInfo::getFixedStack(FI), 2726 false, false, false, 0)); 2727 } 2728 lastInsIndex = index; 2729 } 2730 } 2731 } 2732 2733 // varargs 2734 if (isVarArg) 2735 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset()); 2736 2737 return Chain; 2738 } 2739 2740 /// isFloatingPointZero - Return true if this is +0.0. 2741 static bool isFloatingPointZero(SDValue Op) { 2742 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op)) 2743 return CFP->getValueAPF().isPosZero(); 2744 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) { 2745 // Maybe this has already been legalized into the constant pool? 2746 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) { 2747 SDValue WrapperOp = Op.getOperand(1).getOperand(0); 2748 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp)) 2749 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal())) 2750 return CFP->getValueAPF().isPosZero(); 2751 } 2752 } 2753 return false; 2754 } 2755 2756 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for 2757 /// the given operands. 2758 SDValue 2759 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC, 2760 SDValue &ARMcc, SelectionDAG &DAG, 2761 DebugLoc dl) const { 2762 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) { 2763 unsigned C = RHSC->getZExtValue(); 2764 if (!isLegalICmpImmediate(C)) { 2765 // Constant does not fit, try adjusting it by one? 2766 switch (CC) { 2767 default: break; 2768 case ISD::SETLT: 2769 case ISD::SETGE: 2770 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) { 2771 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT; 2772 RHS = DAG.getConstant(C-1, MVT::i32); 2773 } 2774 break; 2775 case ISD::SETULT: 2776 case ISD::SETUGE: 2777 if (C != 0 && isLegalICmpImmediate(C-1)) { 2778 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT; 2779 RHS = DAG.getConstant(C-1, MVT::i32); 2780 } 2781 break; 2782 case ISD::SETLE: 2783 case ISD::SETGT: 2784 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) { 2785 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE; 2786 RHS = DAG.getConstant(C+1, MVT::i32); 2787 } 2788 break; 2789 case ISD::SETULE: 2790 case ISD::SETUGT: 2791 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) { 2792 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; 2793 RHS = DAG.getConstant(C+1, MVT::i32); 2794 } 2795 break; 2796 } 2797 } 2798 } 2799 2800 ARMCC::CondCodes CondCode = IntCCToARMCC(CC); 2801 ARMISD::NodeType CompareType; 2802 switch (CondCode) { 2803 default: 2804 CompareType = ARMISD::CMP; 2805 break; 2806 case ARMCC::EQ: 2807 case ARMCC::NE: 2808 // Uses only Z Flag 2809 CompareType = ARMISD::CMPZ; 2810 break; 2811 } 2812 ARMcc = DAG.getConstant(CondCode, MVT::i32); 2813 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS); 2814 } 2815 2816 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands. 2817 SDValue 2818 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG, 2819 DebugLoc dl) const { 2820 SDValue Cmp; 2821 if (!isFloatingPointZero(RHS)) 2822 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS); 2823 else 2824 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS); 2825 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp); 2826 } 2827 2828 /// duplicateCmp - Glue values can have only one use, so this function 2829 /// duplicates a comparison node. 2830 SDValue 2831 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const { 2832 unsigned Opc = Cmp.getOpcode(); 2833 DebugLoc DL = Cmp.getDebugLoc(); 2834 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ) 2835 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1)); 2836 2837 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation"); 2838 Cmp = Cmp.getOperand(0); 2839 Opc = Cmp.getOpcode(); 2840 if (Opc == ARMISD::CMPFP) 2841 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1)); 2842 else { 2843 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT"); 2844 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0)); 2845 } 2846 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp); 2847 } 2848 2849 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const { 2850 SDValue Cond = Op.getOperand(0); 2851 SDValue SelectTrue = Op.getOperand(1); 2852 SDValue SelectFalse = Op.getOperand(2); 2853 DebugLoc dl = Op.getDebugLoc(); 2854 2855 // Convert: 2856 // 2857 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond) 2858 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond) 2859 // 2860 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) { 2861 const ConstantSDNode *CMOVTrue = 2862 dyn_cast<ConstantSDNode>(Cond.getOperand(0)); 2863 const ConstantSDNode *CMOVFalse = 2864 dyn_cast<ConstantSDNode>(Cond.getOperand(1)); 2865 2866 if (CMOVTrue && CMOVFalse) { 2867 unsigned CMOVTrueVal = CMOVTrue->getZExtValue(); 2868 unsigned CMOVFalseVal = CMOVFalse->getZExtValue(); 2869 2870 SDValue True; 2871 SDValue False; 2872 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) { 2873 True = SelectTrue; 2874 False = SelectFalse; 2875 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) { 2876 True = SelectFalse; 2877 False = SelectTrue; 2878 } 2879 2880 if (True.getNode() && False.getNode()) { 2881 EVT VT = Op.getValueType(); 2882 SDValue ARMcc = Cond.getOperand(2); 2883 SDValue CCR = Cond.getOperand(3); 2884 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG); 2885 assert(True.getValueType() == VT); 2886 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp); 2887 } 2888 } 2889 } 2890 2891 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the 2892 // undefined bits before doing a full-word comparison with zero. 2893 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond, 2894 DAG.getConstant(1, Cond.getValueType())); 2895 2896 return DAG.getSelectCC(dl, Cond, 2897 DAG.getConstant(0, Cond.getValueType()), 2898 SelectTrue, SelectFalse, ISD::SETNE); 2899 } 2900 2901 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const { 2902 EVT VT = Op.getValueType(); 2903 SDValue LHS = Op.getOperand(0); 2904 SDValue RHS = Op.getOperand(1); 2905 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get(); 2906 SDValue TrueVal = Op.getOperand(2); 2907 SDValue FalseVal = Op.getOperand(3); 2908 DebugLoc dl = Op.getDebugLoc(); 2909 2910 if (LHS.getValueType() == MVT::i32) { 2911 SDValue ARMcc; 2912 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 2913 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl); 2914 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp); 2915 } 2916 2917 ARMCC::CondCodes CondCode, CondCode2; 2918 FPCCToARMCC(CC, CondCode, CondCode2); 2919 2920 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32); 2921 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl); 2922 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 2923 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, 2924 ARMcc, CCR, Cmp); 2925 if (CondCode2 != ARMCC::AL) { 2926 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32); 2927 // FIXME: Needs another CMP because flag can have but one use. 2928 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl); 2929 Result = DAG.getNode(ARMISD::CMOV, dl, VT, 2930 Result, TrueVal, ARMcc2, CCR, Cmp2); 2931 } 2932 return Result; 2933 } 2934 2935 /// canChangeToInt - Given the fp compare operand, return true if it is suitable 2936 /// to morph to an integer compare sequence. 2937 static bool canChangeToInt(SDValue Op, bool &SeenZero, 2938 const ARMSubtarget *Subtarget) { 2939 SDNode *N = Op.getNode(); 2940 if (!N->hasOneUse()) 2941 // Otherwise it requires moving the value from fp to integer registers. 2942 return false; 2943 if (!N->getNumValues()) 2944 return false; 2945 EVT VT = Op.getValueType(); 2946 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow()) 2947 // f32 case is generally profitable. f64 case only makes sense when vcmpe + 2948 // vmrs are very slow, e.g. cortex-a8. 2949 return false; 2950 2951 if (isFloatingPointZero(Op)) { 2952 SeenZero = true; 2953 return true; 2954 } 2955 return ISD::isNormalLoad(N); 2956 } 2957 2958 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) { 2959 if (isFloatingPointZero(Op)) 2960 return DAG.getConstant(0, MVT::i32); 2961 2962 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) 2963 return DAG.getLoad(MVT::i32, Op.getDebugLoc(), 2964 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(), 2965 Ld->isVolatile(), Ld->isNonTemporal(), 2966 Ld->isInvariant(), Ld->getAlignment()); 2967 2968 llvm_unreachable("Unknown VFP cmp argument!"); 2969 } 2970 2971 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG, 2972 SDValue &RetVal1, SDValue &RetVal2) { 2973 if (isFloatingPointZero(Op)) { 2974 RetVal1 = DAG.getConstant(0, MVT::i32); 2975 RetVal2 = DAG.getConstant(0, MVT::i32); 2976 return; 2977 } 2978 2979 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) { 2980 SDValue Ptr = Ld->getBasePtr(); 2981 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(), 2982 Ld->getChain(), Ptr, 2983 Ld->getPointerInfo(), 2984 Ld->isVolatile(), Ld->isNonTemporal(), 2985 Ld->isInvariant(), Ld->getAlignment()); 2986 2987 EVT PtrType = Ptr.getValueType(); 2988 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4); 2989 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(), 2990 PtrType, Ptr, DAG.getConstant(4, PtrType)); 2991 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(), 2992 Ld->getChain(), NewPtr, 2993 Ld->getPointerInfo().getWithOffset(4), 2994 Ld->isVolatile(), Ld->isNonTemporal(), 2995 Ld->isInvariant(), NewAlign); 2996 return; 2997 } 2998 2999 llvm_unreachable("Unknown VFP cmp argument!"); 3000 } 3001 3002 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some 3003 /// f32 and even f64 comparisons to integer ones. 3004 SDValue 3005 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const { 3006 SDValue Chain = Op.getOperand(0); 3007 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 3008 SDValue LHS = Op.getOperand(2); 3009 SDValue RHS = Op.getOperand(3); 3010 SDValue Dest = Op.getOperand(4); 3011 DebugLoc dl = Op.getDebugLoc(); 3012 3013 bool LHSSeenZero = false; 3014 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget); 3015 bool RHSSeenZero = false; 3016 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget); 3017 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) { 3018 // If unsafe fp math optimization is enabled and there are no other uses of 3019 // the CMP operands, and the condition code is EQ or NE, we can optimize it 3020 // to an integer comparison. 3021 if (CC == ISD::SETOEQ) 3022 CC = ISD::SETEQ; 3023 else if (CC == ISD::SETUNE) 3024 CC = ISD::SETNE; 3025 3026 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32); 3027 SDValue ARMcc; 3028 if (LHS.getValueType() == MVT::f32) { 3029 LHS = DAG.getNode(ISD::AND, dl, MVT::i32, 3030 bitcastf32Toi32(LHS, DAG), Mask); 3031 RHS = DAG.getNode(ISD::AND, dl, MVT::i32, 3032 bitcastf32Toi32(RHS, DAG), Mask); 3033 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl); 3034 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3035 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other, 3036 Chain, Dest, ARMcc, CCR, Cmp); 3037 } 3038 3039 SDValue LHS1, LHS2; 3040 SDValue RHS1, RHS2; 3041 expandf64Toi32(LHS, DAG, LHS1, LHS2); 3042 expandf64Toi32(RHS, DAG, RHS1, RHS2); 3043 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask); 3044 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask); 3045 ARMCC::CondCodes CondCode = IntCCToARMCC(CC); 3046 ARMcc = DAG.getConstant(CondCode, MVT::i32); 3047 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue); 3048 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest }; 3049 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7); 3050 } 3051 3052 return SDValue(); 3053 } 3054 3055 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const { 3056 SDValue Chain = Op.getOperand(0); 3057 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 3058 SDValue LHS = Op.getOperand(2); 3059 SDValue RHS = Op.getOperand(3); 3060 SDValue Dest = Op.getOperand(4); 3061 DebugLoc dl = Op.getDebugLoc(); 3062 3063 if (LHS.getValueType() == MVT::i32) { 3064 SDValue ARMcc; 3065 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl); 3066 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3067 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other, 3068 Chain, Dest, ARMcc, CCR, Cmp); 3069 } 3070 3071 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64); 3072 3073 if (getTargetMachine().Options.UnsafeFPMath && 3074 (CC == ISD::SETEQ || CC == ISD::SETOEQ || 3075 CC == ISD::SETNE || CC == ISD::SETUNE)) { 3076 SDValue Result = OptimizeVFPBrcond(Op, DAG); 3077 if (Result.getNode()) 3078 return Result; 3079 } 3080 3081 ARMCC::CondCodes CondCode, CondCode2; 3082 FPCCToARMCC(CC, CondCode, CondCode2); 3083 3084 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32); 3085 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl); 3086 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3087 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue); 3088 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp }; 3089 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5); 3090 if (CondCode2 != ARMCC::AL) { 3091 ARMcc = DAG.getConstant(CondCode2, MVT::i32); 3092 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) }; 3093 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5); 3094 } 3095 return Res; 3096 } 3097 3098 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const { 3099 SDValue Chain = Op.getOperand(0); 3100 SDValue Table = Op.getOperand(1); 3101 SDValue Index = Op.getOperand(2); 3102 DebugLoc dl = Op.getDebugLoc(); 3103 3104 EVT PTy = getPointerTy(); 3105 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table); 3106 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>(); 3107 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy); 3108 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy); 3109 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId); 3110 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy)); 3111 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table); 3112 if (Subtarget->isThumb2()) { 3113 // Thumb2 uses a two-level jump. That is, it jumps into the jump table 3114 // which does another jump to the destination. This also makes it easier 3115 // to translate it to TBB / TBH later. 3116 // FIXME: This might not work if the function is extremely large. 3117 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain, 3118 Addr, Op.getOperand(2), JTI, UId); 3119 } 3120 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) { 3121 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, 3122 MachinePointerInfo::getJumpTable(), 3123 false, false, false, 0); 3124 Chain = Addr.getValue(1); 3125 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table); 3126 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId); 3127 } else { 3128 Addr = DAG.getLoad(PTy, dl, Chain, Addr, 3129 MachinePointerInfo::getJumpTable(), 3130 false, false, false, 0); 3131 Chain = Addr.getValue(1); 3132 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId); 3133 } 3134 } 3135 3136 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) { 3137 EVT VT = Op.getValueType(); 3138 DebugLoc dl = Op.getDebugLoc(); 3139 3140 if (Op.getValueType().getVectorElementType() == MVT::i32) { 3141 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32) 3142 return Op; 3143 return DAG.UnrollVectorOp(Op.getNode()); 3144 } 3145 3146 assert(Op.getOperand(0).getValueType() == MVT::v4f32 && 3147 "Invalid type for custom lowering!"); 3148 if (VT != MVT::v4i16) 3149 return DAG.UnrollVectorOp(Op.getNode()); 3150 3151 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0)); 3152 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op); 3153 } 3154 3155 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) { 3156 EVT VT = Op.getValueType(); 3157 if (VT.isVector()) 3158 return LowerVectorFP_TO_INT(Op, DAG); 3159 3160 DebugLoc dl = Op.getDebugLoc(); 3161 unsigned Opc; 3162 3163 switch (Op.getOpcode()) { 3164 default: llvm_unreachable("Invalid opcode!"); 3165 case ISD::FP_TO_SINT: 3166 Opc = ARMISD::FTOSI; 3167 break; 3168 case ISD::FP_TO_UINT: 3169 Opc = ARMISD::FTOUI; 3170 break; 3171 } 3172 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0)); 3173 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); 3174 } 3175 3176 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) { 3177 EVT VT = Op.getValueType(); 3178 DebugLoc dl = Op.getDebugLoc(); 3179 3180 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) { 3181 if (VT.getVectorElementType() == MVT::f32) 3182 return Op; 3183 return DAG.UnrollVectorOp(Op.getNode()); 3184 } 3185 3186 assert(Op.getOperand(0).getValueType() == MVT::v4i16 && 3187 "Invalid type for custom lowering!"); 3188 if (VT != MVT::v4f32) 3189 return DAG.UnrollVectorOp(Op.getNode()); 3190 3191 unsigned CastOpc; 3192 unsigned Opc; 3193 switch (Op.getOpcode()) { 3194 default: llvm_unreachable("Invalid opcode!"); 3195 case ISD::SINT_TO_FP: 3196 CastOpc = ISD::SIGN_EXTEND; 3197 Opc = ISD::SINT_TO_FP; 3198 break; 3199 case ISD::UINT_TO_FP: 3200 CastOpc = ISD::ZERO_EXTEND; 3201 Opc = ISD::UINT_TO_FP; 3202 break; 3203 } 3204 3205 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0)); 3206 return DAG.getNode(Opc, dl, VT, Op); 3207 } 3208 3209 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) { 3210 EVT VT = Op.getValueType(); 3211 if (VT.isVector()) 3212 return LowerVectorINT_TO_FP(Op, DAG); 3213 3214 DebugLoc dl = Op.getDebugLoc(); 3215 unsigned Opc; 3216 3217 switch (Op.getOpcode()) { 3218 default: llvm_unreachable("Invalid opcode!"); 3219 case ISD::SINT_TO_FP: 3220 Opc = ARMISD::SITOF; 3221 break; 3222 case ISD::UINT_TO_FP: 3223 Opc = ARMISD::UITOF; 3224 break; 3225 } 3226 3227 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0)); 3228 return DAG.getNode(Opc, dl, VT, Op); 3229 } 3230 3231 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { 3232 // Implement fcopysign with a fabs and a conditional fneg. 3233 SDValue Tmp0 = Op.getOperand(0); 3234 SDValue Tmp1 = Op.getOperand(1); 3235 DebugLoc dl = Op.getDebugLoc(); 3236 EVT VT = Op.getValueType(); 3237 EVT SrcVT = Tmp1.getValueType(); 3238 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST || 3239 Tmp0.getOpcode() == ARMISD::VMOVDRR; 3240 bool UseNEON = !InGPR && Subtarget->hasNEON(); 3241 3242 if (UseNEON) { 3243 // Use VBSL to copy the sign bit. 3244 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80); 3245 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32, 3246 DAG.getTargetConstant(EncodedVal, MVT::i32)); 3247 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64; 3248 if (VT == MVT::f64) 3249 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT, 3250 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask), 3251 DAG.getConstant(32, MVT::i32)); 3252 else /*if (VT == MVT::f32)*/ 3253 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0); 3254 if (SrcVT == MVT::f32) { 3255 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1); 3256 if (VT == MVT::f64) 3257 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT, 3258 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1), 3259 DAG.getConstant(32, MVT::i32)); 3260 } else if (VT == MVT::f32) 3261 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64, 3262 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1), 3263 DAG.getConstant(32, MVT::i32)); 3264 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0); 3265 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1); 3266 3267 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff), 3268 MVT::i32); 3269 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes); 3270 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask, 3271 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes)); 3272 3273 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT, 3274 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask), 3275 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot)); 3276 if (VT == MVT::f32) { 3277 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res); 3278 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res, 3279 DAG.getConstant(0, MVT::i32)); 3280 } else { 3281 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res); 3282 } 3283 3284 return Res; 3285 } 3286 3287 // Bitcast operand 1 to i32. 3288 if (SrcVT == MVT::f64) 3289 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32), 3290 &Tmp1, 1).getValue(1); 3291 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1); 3292 3293 // Or in the signbit with integer operations. 3294 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32); 3295 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32); 3296 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1); 3297 if (VT == MVT::f32) { 3298 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32, 3299 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2); 3300 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, 3301 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1)); 3302 } 3303 3304 // f64: Or the high part with signbit and then combine two parts. 3305 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32), 3306 &Tmp0, 1); 3307 SDValue Lo = Tmp0.getValue(0); 3308 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2); 3309 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1); 3310 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi); 3311 } 3312 3313 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{ 3314 MachineFunction &MF = DAG.getMachineFunction(); 3315 MachineFrameInfo *MFI = MF.getFrameInfo(); 3316 MFI->setReturnAddressIsTaken(true); 3317 3318 EVT VT = Op.getValueType(); 3319 DebugLoc dl = Op.getDebugLoc(); 3320 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 3321 if (Depth) { 3322 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG); 3323 SDValue Offset = DAG.getConstant(4, MVT::i32); 3324 return DAG.getLoad(VT, dl, DAG.getEntryNode(), 3325 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset), 3326 MachinePointerInfo(), false, false, false, 0); 3327 } 3328 3329 // Return LR, which contains the return address. Mark it an implicit live-in. 3330 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32)); 3331 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT); 3332 } 3333 3334 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { 3335 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 3336 MFI->setFrameAddressIsTaken(true); 3337 3338 EVT VT = Op.getValueType(); 3339 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful 3340 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 3341 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin()) 3342 ? ARM::R7 : ARM::R11; 3343 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT); 3344 while (Depth--) 3345 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, 3346 MachinePointerInfo(), 3347 false, false, false, 0); 3348 return FrameAddr; 3349 } 3350 3351 /// ExpandBITCAST - If the target supports VFP, this function is called to 3352 /// expand a bit convert where either the source or destination type is i64 to 3353 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64 3354 /// operand type is illegal (e.g., v2f32 for a target that doesn't support 3355 /// vectors), since the legalizer won't know what to do with that. 3356 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) { 3357 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3358 DebugLoc dl = N->getDebugLoc(); 3359 SDValue Op = N->getOperand(0); 3360 3361 // This function is only supposed to be called for i64 types, either as the 3362 // source or destination of the bit convert. 3363 EVT SrcVT = Op.getValueType(); 3364 EVT DstVT = N->getValueType(0); 3365 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) && 3366 "ExpandBITCAST called for non-i64 type"); 3367 3368 // Turn i64->f64 into VMOVDRR. 3369 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) { 3370 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op, 3371 DAG.getConstant(0, MVT::i32)); 3372 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op, 3373 DAG.getConstant(1, MVT::i32)); 3374 return DAG.getNode(ISD::BITCAST, dl, DstVT, 3375 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi)); 3376 } 3377 3378 // Turn f64->i64 into VMOVRRD. 3379 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) { 3380 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl, 3381 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1); 3382 // Merge the pieces into a single i64 value. 3383 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1)); 3384 } 3385 3386 return SDValue(); 3387 } 3388 3389 /// getZeroVector - Returns a vector of specified type with all zero elements. 3390 /// Zero vectors are used to represent vector negation and in those cases 3391 /// will be implemented with the NEON VNEG instruction. However, VNEG does 3392 /// not support i64 elements, so sometimes the zero vectors will need to be 3393 /// explicitly constructed. Regardless, use a canonical VMOV to create the 3394 /// zero vector. 3395 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) { 3396 assert(VT.isVector() && "Expected a vector type"); 3397 // The canonical modified immediate encoding of a zero vector is....0! 3398 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32); 3399 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32; 3400 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal); 3401 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov); 3402 } 3403 3404 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two 3405 /// i32 values and take a 2 x i32 value to shift plus a shift amount. 3406 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op, 3407 SelectionDAG &DAG) const { 3408 assert(Op.getNumOperands() == 3 && "Not a double-shift!"); 3409 EVT VT = Op.getValueType(); 3410 unsigned VTBits = VT.getSizeInBits(); 3411 DebugLoc dl = Op.getDebugLoc(); 3412 SDValue ShOpLo = Op.getOperand(0); 3413 SDValue ShOpHi = Op.getOperand(1); 3414 SDValue ShAmt = Op.getOperand(2); 3415 SDValue ARMcc; 3416 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL; 3417 3418 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS); 3419 3420 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, 3421 DAG.getConstant(VTBits, MVT::i32), ShAmt); 3422 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt); 3423 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt, 3424 DAG.getConstant(VTBits, MVT::i32)); 3425 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt); 3426 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2); 3427 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt); 3428 3429 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3430 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE, 3431 ARMcc, DAG, dl); 3432 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt); 3433 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, 3434 CCR, Cmp); 3435 3436 SDValue Ops[2] = { Lo, Hi }; 3437 return DAG.getMergeValues(Ops, 2, dl); 3438 } 3439 3440 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two 3441 /// i32 values and take a 2 x i32 value to shift plus a shift amount. 3442 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op, 3443 SelectionDAG &DAG) const { 3444 assert(Op.getNumOperands() == 3 && "Not a double-shift!"); 3445 EVT VT = Op.getValueType(); 3446 unsigned VTBits = VT.getSizeInBits(); 3447 DebugLoc dl = Op.getDebugLoc(); 3448 SDValue ShOpLo = Op.getOperand(0); 3449 SDValue ShOpHi = Op.getOperand(1); 3450 SDValue ShAmt = Op.getOperand(2); 3451 SDValue ARMcc; 3452 3453 assert(Op.getOpcode() == ISD::SHL_PARTS); 3454 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, 3455 DAG.getConstant(VTBits, MVT::i32), ShAmt); 3456 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt); 3457 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt, 3458 DAG.getConstant(VTBits, MVT::i32)); 3459 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt); 3460 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt); 3461 3462 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2); 3463 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3464 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE, 3465 ARMcc, DAG, dl); 3466 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt); 3467 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc, 3468 CCR, Cmp); 3469 3470 SDValue Ops[2] = { Lo, Hi }; 3471 return DAG.getMergeValues(Ops, 2, dl); 3472 } 3473 3474 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op, 3475 SelectionDAG &DAG) const { 3476 // The rounding mode is in bits 23:22 of the FPSCR. 3477 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0 3478 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3) 3479 // so that the shift + and get folded into a bitfield extract. 3480 DebugLoc dl = Op.getDebugLoc(); 3481 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32, 3482 DAG.getConstant(Intrinsic::arm_get_fpscr, 3483 MVT::i32)); 3484 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR, 3485 DAG.getConstant(1U << 22, MVT::i32)); 3486 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds, 3487 DAG.getConstant(22, MVT::i32)); 3488 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE, 3489 DAG.getConstant(3, MVT::i32)); 3490 } 3491 3492 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG, 3493 const ARMSubtarget *ST) { 3494 EVT VT = N->getValueType(0); 3495 DebugLoc dl = N->getDebugLoc(); 3496 3497 if (!ST->hasV6T2Ops()) 3498 return SDValue(); 3499 3500 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0)); 3501 return DAG.getNode(ISD::CTLZ, dl, VT, rbit); 3502 } 3503 3504 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG, 3505 const ARMSubtarget *ST) { 3506 EVT VT = N->getValueType(0); 3507 DebugLoc dl = N->getDebugLoc(); 3508 3509 if (!VT.isVector()) 3510 return SDValue(); 3511 3512 // Lower vector shifts on NEON to use VSHL. 3513 assert(ST->hasNEON() && "unexpected vector shift"); 3514 3515 // Left shifts translate directly to the vshiftu intrinsic. 3516 if (N->getOpcode() == ISD::SHL) 3517 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, 3518 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32), 3519 N->getOperand(0), N->getOperand(1)); 3520 3521 assert((N->getOpcode() == ISD::SRA || 3522 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode"); 3523 3524 // NEON uses the same intrinsics for both left and right shifts. For 3525 // right shifts, the shift amounts are negative, so negate the vector of 3526 // shift amounts. 3527 EVT ShiftVT = N->getOperand(1).getValueType(); 3528 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT, 3529 getZeroVector(ShiftVT, DAG, dl), 3530 N->getOperand(1)); 3531 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ? 3532 Intrinsic::arm_neon_vshifts : 3533 Intrinsic::arm_neon_vshiftu); 3534 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, 3535 DAG.getConstant(vshiftInt, MVT::i32), 3536 N->getOperand(0), NegatedCount); 3537 } 3538 3539 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG, 3540 const ARMSubtarget *ST) { 3541 EVT VT = N->getValueType(0); 3542 DebugLoc dl = N->getDebugLoc(); 3543 3544 // We can get here for a node like i32 = ISD::SHL i32, i64 3545 if (VT != MVT::i64) 3546 return SDValue(); 3547 3548 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) && 3549 "Unknown shift to lower!"); 3550 3551 // We only lower SRA, SRL of 1 here, all others use generic lowering. 3552 if (!isa<ConstantSDNode>(N->getOperand(1)) || 3553 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1) 3554 return SDValue(); 3555 3556 // If we are in thumb mode, we don't have RRX. 3557 if (ST->isThumb1Only()) return SDValue(); 3558 3559 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr. 3560 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0), 3561 DAG.getConstant(0, MVT::i32)); 3562 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0), 3563 DAG.getConstant(1, MVT::i32)); 3564 3565 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and 3566 // captures the result into a carry flag. 3567 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG; 3568 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1); 3569 3570 // The low part is an ARMISD::RRX operand, which shifts the carry in. 3571 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1)); 3572 3573 // Merge the pieces into a single i64 value. 3574 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi); 3575 } 3576 3577 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) { 3578 SDValue TmpOp0, TmpOp1; 3579 bool Invert = false; 3580 bool Swap = false; 3581 unsigned Opc = 0; 3582 3583 SDValue Op0 = Op.getOperand(0); 3584 SDValue Op1 = Op.getOperand(1); 3585 SDValue CC = Op.getOperand(2); 3586 EVT VT = Op.getValueType(); 3587 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get(); 3588 DebugLoc dl = Op.getDebugLoc(); 3589 3590 if (Op.getOperand(1).getValueType().isFloatingPoint()) { 3591 switch (SetCCOpcode) { 3592 default: llvm_unreachable("Illegal FP comparison"); 3593 case ISD::SETUNE: 3594 case ISD::SETNE: Invert = true; // Fallthrough 3595 case ISD::SETOEQ: 3596 case ISD::SETEQ: Opc = ARMISD::VCEQ; break; 3597 case ISD::SETOLT: 3598 case ISD::SETLT: Swap = true; // Fallthrough 3599 case ISD::SETOGT: 3600 case ISD::SETGT: Opc = ARMISD::VCGT; break; 3601 case ISD::SETOLE: 3602 case ISD::SETLE: Swap = true; // Fallthrough 3603 case ISD::SETOGE: 3604 case ISD::SETGE: Opc = ARMISD::VCGE; break; 3605 case ISD::SETUGE: Swap = true; // Fallthrough 3606 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break; 3607 case ISD::SETUGT: Swap = true; // Fallthrough 3608 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break; 3609 case ISD::SETUEQ: Invert = true; // Fallthrough 3610 case ISD::SETONE: 3611 // Expand this to (OLT | OGT). 3612 TmpOp0 = Op0; 3613 TmpOp1 = Op1; 3614 Opc = ISD::OR; 3615 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0); 3616 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1); 3617 break; 3618 case ISD::SETUO: Invert = true; // Fallthrough 3619 case ISD::SETO: 3620 // Expand this to (OLT | OGE). 3621 TmpOp0 = Op0; 3622 TmpOp1 = Op1; 3623 Opc = ISD::OR; 3624 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0); 3625 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1); 3626 break; 3627 } 3628 } else { 3629 // Integer comparisons. 3630 switch (SetCCOpcode) { 3631 default: llvm_unreachable("Illegal integer comparison"); 3632 case ISD::SETNE: Invert = true; 3633 case ISD::SETEQ: Opc = ARMISD::VCEQ; break; 3634 case ISD::SETLT: Swap = true; 3635 case ISD::SETGT: Opc = ARMISD::VCGT; break; 3636 case ISD::SETLE: Swap = true; 3637 case ISD::SETGE: Opc = ARMISD::VCGE; break; 3638 case ISD::SETULT: Swap = true; 3639 case ISD::SETUGT: Opc = ARMISD::VCGTU; break; 3640 case ISD::SETULE: Swap = true; 3641 case ISD::SETUGE: Opc = ARMISD::VCGEU; break; 3642 } 3643 3644 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero). 3645 if (Opc == ARMISD::VCEQ) { 3646 3647 SDValue AndOp; 3648 if (ISD::isBuildVectorAllZeros(Op1.getNode())) 3649 AndOp = Op0; 3650 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) 3651 AndOp = Op1; 3652 3653 // Ignore bitconvert. 3654 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST) 3655 AndOp = AndOp.getOperand(0); 3656 3657 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) { 3658 Opc = ARMISD::VTST; 3659 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0)); 3660 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1)); 3661 Invert = !Invert; 3662 } 3663 } 3664 } 3665 3666 if (Swap) 3667 std::swap(Op0, Op1); 3668 3669 // If one of the operands is a constant vector zero, attempt to fold the 3670 // comparison to a specialized compare-against-zero form. 3671 SDValue SingleOp; 3672 if (ISD::isBuildVectorAllZeros(Op1.getNode())) 3673 SingleOp = Op0; 3674 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) { 3675 if (Opc == ARMISD::VCGE) 3676 Opc = ARMISD::VCLEZ; 3677 else if (Opc == ARMISD::VCGT) 3678 Opc = ARMISD::VCLTZ; 3679 SingleOp = Op1; 3680 } 3681 3682 SDValue Result; 3683 if (SingleOp.getNode()) { 3684 switch (Opc) { 3685 case ARMISD::VCEQ: 3686 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break; 3687 case ARMISD::VCGE: 3688 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break; 3689 case ARMISD::VCLEZ: 3690 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break; 3691 case ARMISD::VCGT: 3692 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break; 3693 case ARMISD::VCLTZ: 3694 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break; 3695 default: 3696 Result = DAG.getNode(Opc, dl, VT, Op0, Op1); 3697 } 3698 } else { 3699 Result = DAG.getNode(Opc, dl, VT, Op0, Op1); 3700 } 3701 3702 if (Invert) 3703 Result = DAG.getNOT(dl, Result, VT); 3704 3705 return Result; 3706 } 3707 3708 /// isNEONModifiedImm - Check if the specified splat value corresponds to a 3709 /// valid vector constant for a NEON instruction with a "modified immediate" 3710 /// operand (e.g., VMOV). If so, return the encoded value. 3711 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef, 3712 unsigned SplatBitSize, SelectionDAG &DAG, 3713 EVT &VT, bool is128Bits, NEONModImmType type) { 3714 unsigned OpCmode, Imm; 3715 3716 // SplatBitSize is set to the smallest size that splats the vector, so a 3717 // zero vector will always have SplatBitSize == 8. However, NEON modified 3718 // immediate instructions others than VMOV do not support the 8-bit encoding 3719 // of a zero vector, and the default encoding of zero is supposed to be the 3720 // 32-bit version. 3721 if (SplatBits == 0) 3722 SplatBitSize = 32; 3723 3724 switch (SplatBitSize) { 3725 case 8: 3726 if (type != VMOVModImm) 3727 return SDValue(); 3728 // Any 1-byte value is OK. Op=0, Cmode=1110. 3729 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big"); 3730 OpCmode = 0xe; 3731 Imm = SplatBits; 3732 VT = is128Bits ? MVT::v16i8 : MVT::v8i8; 3733 break; 3734 3735 case 16: 3736 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero. 3737 VT = is128Bits ? MVT::v8i16 : MVT::v4i16; 3738 if ((SplatBits & ~0xff) == 0) { 3739 // Value = 0x00nn: Op=x, Cmode=100x. 3740 OpCmode = 0x8; 3741 Imm = SplatBits; 3742 break; 3743 } 3744 if ((SplatBits & ~0xff00) == 0) { 3745 // Value = 0xnn00: Op=x, Cmode=101x. 3746 OpCmode = 0xa; 3747 Imm = SplatBits >> 8; 3748 break; 3749 } 3750 return SDValue(); 3751 3752 case 32: 3753 // NEON's 32-bit VMOV supports splat values where: 3754 // * only one byte is nonzero, or 3755 // * the least significant byte is 0xff and the second byte is nonzero, or 3756 // * the least significant 2 bytes are 0xff and the third is nonzero. 3757 VT = is128Bits ? MVT::v4i32 : MVT::v2i32; 3758 if ((SplatBits & ~0xff) == 0) { 3759 // Value = 0x000000nn: Op=x, Cmode=000x. 3760 OpCmode = 0; 3761 Imm = SplatBits; 3762 break; 3763 } 3764 if ((SplatBits & ~0xff00) == 0) { 3765 // Value = 0x0000nn00: Op=x, Cmode=001x. 3766 OpCmode = 0x2; 3767 Imm = SplatBits >> 8; 3768 break; 3769 } 3770 if ((SplatBits & ~0xff0000) == 0) { 3771 // Value = 0x00nn0000: Op=x, Cmode=010x. 3772 OpCmode = 0x4; 3773 Imm = SplatBits >> 16; 3774 break; 3775 } 3776 if ((SplatBits & ~0xff000000) == 0) { 3777 // Value = 0xnn000000: Op=x, Cmode=011x. 3778 OpCmode = 0x6; 3779 Imm = SplatBits >> 24; 3780 break; 3781 } 3782 3783 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC 3784 if (type == OtherModImm) return SDValue(); 3785 3786 if ((SplatBits & ~0xffff) == 0 && 3787 ((SplatBits | SplatUndef) & 0xff) == 0xff) { 3788 // Value = 0x0000nnff: Op=x, Cmode=1100. 3789 OpCmode = 0xc; 3790 Imm = SplatBits >> 8; 3791 SplatBits |= 0xff; 3792 break; 3793 } 3794 3795 if ((SplatBits & ~0xffffff) == 0 && 3796 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) { 3797 // Value = 0x00nnffff: Op=x, Cmode=1101. 3798 OpCmode = 0xd; 3799 Imm = SplatBits >> 16; 3800 SplatBits |= 0xffff; 3801 break; 3802 } 3803 3804 // Note: there are a few 32-bit splat values (specifically: 00ffff00, 3805 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not 3806 // VMOV.I32. A (very) minor optimization would be to replicate the value 3807 // and fall through here to test for a valid 64-bit splat. But, then the 3808 // caller would also need to check and handle the change in size. 3809 return SDValue(); 3810 3811 case 64: { 3812 if (type != VMOVModImm) 3813 return SDValue(); 3814 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff. 3815 uint64_t BitMask = 0xff; 3816 uint64_t Val = 0; 3817 unsigned ImmMask = 1; 3818 Imm = 0; 3819 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) { 3820 if (((SplatBits | SplatUndef) & BitMask) == BitMask) { 3821 Val |= BitMask; 3822 Imm |= ImmMask; 3823 } else if ((SplatBits & BitMask) != 0) { 3824 return SDValue(); 3825 } 3826 BitMask <<= 8; 3827 ImmMask <<= 1; 3828 } 3829 // Op=1, Cmode=1110. 3830 OpCmode = 0x1e; 3831 SplatBits = Val; 3832 VT = is128Bits ? MVT::v2i64 : MVT::v1i64; 3833 break; 3834 } 3835 3836 default: 3837 llvm_unreachable("unexpected size for isNEONModifiedImm"); 3838 } 3839 3840 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm); 3841 return DAG.getTargetConstant(EncodedVal, MVT::i32); 3842 } 3843 3844 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG, 3845 const ARMSubtarget *ST) const { 3846 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16()) 3847 return SDValue(); 3848 3849 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op); 3850 assert(Op.getValueType() == MVT::f32 && 3851 "ConstantFP custom lowering should only occur for f32."); 3852 3853 // Try splatting with a VMOV.f32... 3854 APFloat FPVal = CFP->getValueAPF(); 3855 int ImmVal = ARM_AM::getFP32Imm(FPVal); 3856 if (ImmVal != -1) { 3857 DebugLoc DL = Op.getDebugLoc(); 3858 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32); 3859 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32, 3860 NewVal); 3861 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant, 3862 DAG.getConstant(0, MVT::i32)); 3863 } 3864 3865 // If that fails, try a VMOV.i32 3866 EVT VMovVT; 3867 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue(); 3868 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false, 3869 VMOVModImm); 3870 if (NewVal != SDValue()) { 3871 DebugLoc DL = Op.getDebugLoc(); 3872 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT, 3873 NewVal); 3874 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32, 3875 VecConstant); 3876 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant, 3877 DAG.getConstant(0, MVT::i32)); 3878 } 3879 3880 // Finally, try a VMVN.i32 3881 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false, 3882 VMVNModImm); 3883 if (NewVal != SDValue()) { 3884 DebugLoc DL = Op.getDebugLoc(); 3885 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal); 3886 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32, 3887 VecConstant); 3888 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant, 3889 DAG.getConstant(0, MVT::i32)); 3890 } 3891 3892 return SDValue(); 3893 } 3894 3895 3896 static bool isVEXTMask(ArrayRef<int> M, EVT VT, 3897 bool &ReverseVEXT, unsigned &Imm) { 3898 unsigned NumElts = VT.getVectorNumElements(); 3899 ReverseVEXT = false; 3900 3901 // Assume that the first shuffle index is not UNDEF. Fail if it is. 3902 if (M[0] < 0) 3903 return false; 3904 3905 Imm = M[0]; 3906 3907 // If this is a VEXT shuffle, the immediate value is the index of the first 3908 // element. The other shuffle indices must be the successive elements after 3909 // the first one. 3910 unsigned ExpectedElt = Imm; 3911 for (unsigned i = 1; i < NumElts; ++i) { 3912 // Increment the expected index. If it wraps around, it may still be 3913 // a VEXT but the source vectors must be swapped. 3914 ExpectedElt += 1; 3915 if (ExpectedElt == NumElts * 2) { 3916 ExpectedElt = 0; 3917 ReverseVEXT = true; 3918 } 3919 3920 if (M[i] < 0) continue; // ignore UNDEF indices 3921 if (ExpectedElt != static_cast<unsigned>(M[i])) 3922 return false; 3923 } 3924 3925 // Adjust the index value if the source operands will be swapped. 3926 if (ReverseVEXT) 3927 Imm -= NumElts; 3928 3929 return true; 3930 } 3931 3932 /// isVREVMask - Check if a vector shuffle corresponds to a VREV 3933 /// instruction with the specified blocksize. (The order of the elements 3934 /// within each block of the vector is reversed.) 3935 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) { 3936 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) && 3937 "Only possible block sizes for VREV are: 16, 32, 64"); 3938 3939 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 3940 if (EltSz == 64) 3941 return false; 3942 3943 unsigned NumElts = VT.getVectorNumElements(); 3944 unsigned BlockElts = M[0] + 1; 3945 // If the first shuffle index is UNDEF, be optimistic. 3946 if (M[0] < 0) 3947 BlockElts = BlockSize / EltSz; 3948 3949 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz) 3950 return false; 3951 3952 for (unsigned i = 0; i < NumElts; ++i) { 3953 if (M[i] < 0) continue; // ignore UNDEF indices 3954 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts)) 3955 return false; 3956 } 3957 3958 return true; 3959 } 3960 3961 static bool isVTBLMask(ArrayRef<int> M, EVT VT) { 3962 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of 3963 // range, then 0 is placed into the resulting vector. So pretty much any mask 3964 // of 8 elements can work here. 3965 return VT == MVT::v8i8 && M.size() == 8; 3966 } 3967 3968 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 3969 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 3970 if (EltSz == 64) 3971 return false; 3972 3973 unsigned NumElts = VT.getVectorNumElements(); 3974 WhichResult = (M[0] == 0 ? 0 : 1); 3975 for (unsigned i = 0; i < NumElts; i += 2) { 3976 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) || 3977 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult)) 3978 return false; 3979 } 3980 return true; 3981 } 3982 3983 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of 3984 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 3985 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>. 3986 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){ 3987 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 3988 if (EltSz == 64) 3989 return false; 3990 3991 unsigned NumElts = VT.getVectorNumElements(); 3992 WhichResult = (M[0] == 0 ? 0 : 1); 3993 for (unsigned i = 0; i < NumElts; i += 2) { 3994 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) || 3995 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult)) 3996 return false; 3997 } 3998 return true; 3999 } 4000 4001 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 4002 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4003 if (EltSz == 64) 4004 return false; 4005 4006 unsigned NumElts = VT.getVectorNumElements(); 4007 WhichResult = (M[0] == 0 ? 0 : 1); 4008 for (unsigned i = 0; i != NumElts; ++i) { 4009 if (M[i] < 0) continue; // ignore UNDEF indices 4010 if ((unsigned) M[i] != 2 * i + WhichResult) 4011 return false; 4012 } 4013 4014 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4015 if (VT.is64BitVector() && EltSz == 32) 4016 return false; 4017 4018 return true; 4019 } 4020 4021 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of 4022 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 4023 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>, 4024 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){ 4025 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4026 if (EltSz == 64) 4027 return false; 4028 4029 unsigned Half = VT.getVectorNumElements() / 2; 4030 WhichResult = (M[0] == 0 ? 0 : 1); 4031 for (unsigned j = 0; j != 2; ++j) { 4032 unsigned Idx = WhichResult; 4033 for (unsigned i = 0; i != Half; ++i) { 4034 int MIdx = M[i + j * Half]; 4035 if (MIdx >= 0 && (unsigned) MIdx != Idx) 4036 return false; 4037 Idx += 2; 4038 } 4039 } 4040 4041 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4042 if (VT.is64BitVector() && EltSz == 32) 4043 return false; 4044 4045 return true; 4046 } 4047 4048 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 4049 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4050 if (EltSz == 64) 4051 return false; 4052 4053 unsigned NumElts = VT.getVectorNumElements(); 4054 WhichResult = (M[0] == 0 ? 0 : 1); 4055 unsigned Idx = WhichResult * NumElts / 2; 4056 for (unsigned i = 0; i != NumElts; i += 2) { 4057 if ((M[i] >= 0 && (unsigned) M[i] != Idx) || 4058 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts)) 4059 return false; 4060 Idx += 1; 4061 } 4062 4063 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4064 if (VT.is64BitVector() && EltSz == 32) 4065 return false; 4066 4067 return true; 4068 } 4069 4070 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of 4071 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 4072 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>. 4073 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){ 4074 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4075 if (EltSz == 64) 4076 return false; 4077 4078 unsigned NumElts = VT.getVectorNumElements(); 4079 WhichResult = (M[0] == 0 ? 0 : 1); 4080 unsigned Idx = WhichResult * NumElts / 2; 4081 for (unsigned i = 0; i != NumElts; i += 2) { 4082 if ((M[i] >= 0 && (unsigned) M[i] != Idx) || 4083 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx)) 4084 return false; 4085 Idx += 1; 4086 } 4087 4088 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4089 if (VT.is64BitVector() && EltSz == 32) 4090 return false; 4091 4092 return true; 4093 } 4094 4095 // If N is an integer constant that can be moved into a register in one 4096 // instruction, return an SDValue of such a constant (will become a MOV 4097 // instruction). Otherwise return null. 4098 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG, 4099 const ARMSubtarget *ST, DebugLoc dl) { 4100 uint64_t Val; 4101 if (!isa<ConstantSDNode>(N)) 4102 return SDValue(); 4103 Val = cast<ConstantSDNode>(N)->getZExtValue(); 4104 4105 if (ST->isThumb1Only()) { 4106 if (Val <= 255 || ~Val <= 255) 4107 return DAG.getConstant(Val, MVT::i32); 4108 } else { 4109 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1) 4110 return DAG.getConstant(Val, MVT::i32); 4111 } 4112 return SDValue(); 4113 } 4114 4115 // If this is a case we can't handle, return null and let the default 4116 // expansion code take care of it. 4117 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG, 4118 const ARMSubtarget *ST) const { 4119 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode()); 4120 DebugLoc dl = Op.getDebugLoc(); 4121 EVT VT = Op.getValueType(); 4122 4123 APInt SplatBits, SplatUndef; 4124 unsigned SplatBitSize; 4125 bool HasAnyUndefs; 4126 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) { 4127 if (SplatBitSize <= 64) { 4128 // Check if an immediate VMOV works. 4129 EVT VmovVT; 4130 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(), 4131 SplatUndef.getZExtValue(), SplatBitSize, 4132 DAG, VmovVT, VT.is128BitVector(), 4133 VMOVModImm); 4134 if (Val.getNode()) { 4135 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val); 4136 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov); 4137 } 4138 4139 // Try an immediate VMVN. 4140 uint64_t NegatedImm = (~SplatBits).getZExtValue(); 4141 Val = isNEONModifiedImm(NegatedImm, 4142 SplatUndef.getZExtValue(), SplatBitSize, 4143 DAG, VmovVT, VT.is128BitVector(), 4144 VMVNModImm); 4145 if (Val.getNode()) { 4146 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val); 4147 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov); 4148 } 4149 4150 // Use vmov.f32 to materialize other v2f32 and v4f32 splats. 4151 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) { 4152 int ImmVal = ARM_AM::getFP32Imm(SplatBits); 4153 if (ImmVal != -1) { 4154 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32); 4155 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val); 4156 } 4157 } 4158 } 4159 } 4160 4161 // Scan through the operands to see if only one value is used. 4162 // 4163 // As an optimisation, even if more than one value is used it may be more 4164 // profitable to splat with one value then change some lanes. 4165 // 4166 // Heuristically we decide to do this if the vector has a "dominant" value, 4167 // defined as splatted to more than half of the lanes. 4168 unsigned NumElts = VT.getVectorNumElements(); 4169 bool isOnlyLowElement = true; 4170 bool usesOnlyOneValue = true; 4171 bool hasDominantValue = false; 4172 bool isConstant = true; 4173 4174 // Map of the number of times a particular SDValue appears in the 4175 // element list. 4176 DenseMap<SDValue, unsigned> ValueCounts; 4177 SDValue Value; 4178 for (unsigned i = 0; i < NumElts; ++i) { 4179 SDValue V = Op.getOperand(i); 4180 if (V.getOpcode() == ISD::UNDEF) 4181 continue; 4182 if (i > 0) 4183 isOnlyLowElement = false; 4184 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V)) 4185 isConstant = false; 4186 4187 ValueCounts.insert(std::make_pair(V, 0)); 4188 unsigned &Count = ValueCounts[V]; 4189 4190 // Is this value dominant? (takes up more than half of the lanes) 4191 if (++Count > (NumElts / 2)) { 4192 hasDominantValue = true; 4193 Value = V; 4194 } 4195 } 4196 if (ValueCounts.size() != 1) 4197 usesOnlyOneValue = false; 4198 if (!Value.getNode() && ValueCounts.size() > 0) 4199 Value = ValueCounts.begin()->first; 4200 4201 if (ValueCounts.size() == 0) 4202 return DAG.getUNDEF(VT); 4203 4204 if (isOnlyLowElement) 4205 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value); 4206 4207 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 4208 4209 // Use VDUP for non-constant splats. For f32 constant splats, reduce to 4210 // i32 and try again. 4211 if (hasDominantValue && EltSize <= 32) { 4212 if (!isConstant) { 4213 SDValue N; 4214 4215 // If we are VDUPing a value that comes directly from a vector, that will 4216 // cause an unnecessary move to and from a GPR, where instead we could 4217 // just use VDUPLANE. 4218 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT) 4219 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT, 4220 Value->getOperand(0), Value->getOperand(1)); 4221 else 4222 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value); 4223 4224 if (!usesOnlyOneValue) { 4225 // The dominant value was splatted as 'N', but we now have to insert 4226 // all differing elements. 4227 for (unsigned I = 0; I < NumElts; ++I) { 4228 if (Op.getOperand(I) == Value) 4229 continue; 4230 SmallVector<SDValue, 3> Ops; 4231 Ops.push_back(N); 4232 Ops.push_back(Op.getOperand(I)); 4233 Ops.push_back(DAG.getConstant(I, MVT::i32)); 4234 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3); 4235 } 4236 } 4237 return N; 4238 } 4239 if (VT.getVectorElementType().isFloatingPoint()) { 4240 SmallVector<SDValue, 8> Ops; 4241 for (unsigned i = 0; i < NumElts; ++i) 4242 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32, 4243 Op.getOperand(i))); 4244 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts); 4245 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts); 4246 Val = LowerBUILD_VECTOR(Val, DAG, ST); 4247 if (Val.getNode()) 4248 return DAG.getNode(ISD::BITCAST, dl, VT, Val); 4249 } 4250 if (usesOnlyOneValue) { 4251 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl); 4252 if (isConstant && Val.getNode()) 4253 return DAG.getNode(ARMISD::VDUP, dl, VT, Val); 4254 } 4255 } 4256 4257 // If all elements are constants and the case above didn't get hit, fall back 4258 // to the default expansion, which will generate a load from the constant 4259 // pool. 4260 if (isConstant) 4261 return SDValue(); 4262 4263 // Empirical tests suggest this is rarely worth it for vectors of length <= 2. 4264 if (NumElts >= 4) { 4265 SDValue shuffle = ReconstructShuffle(Op, DAG); 4266 if (shuffle != SDValue()) 4267 return shuffle; 4268 } 4269 4270 // Vectors with 32- or 64-bit elements can be built by directly assigning 4271 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands 4272 // will be legalized. 4273 if (EltSize >= 32) { 4274 // Do the expansion with floating-point types, since that is what the VFP 4275 // registers are defined to use, and since i64 is not legal. 4276 EVT EltVT = EVT::getFloatingPointVT(EltSize); 4277 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts); 4278 SmallVector<SDValue, 8> Ops; 4279 for (unsigned i = 0; i < NumElts; ++i) 4280 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i))); 4281 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts); 4282 return DAG.getNode(ISD::BITCAST, dl, VT, Val); 4283 } 4284 4285 return SDValue(); 4286 } 4287 4288 // Gather data to see if the operation can be modelled as a 4289 // shuffle in combination with VEXTs. 4290 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op, 4291 SelectionDAG &DAG) const { 4292 DebugLoc dl = Op.getDebugLoc(); 4293 EVT VT = Op.getValueType(); 4294 unsigned NumElts = VT.getVectorNumElements(); 4295 4296 SmallVector<SDValue, 2> SourceVecs; 4297 SmallVector<unsigned, 2> MinElts; 4298 SmallVector<unsigned, 2> MaxElts; 4299 4300 for (unsigned i = 0; i < NumElts; ++i) { 4301 SDValue V = Op.getOperand(i); 4302 if (V.getOpcode() == ISD::UNDEF) 4303 continue; 4304 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) { 4305 // A shuffle can only come from building a vector from various 4306 // elements of other vectors. 4307 return SDValue(); 4308 } else if (V.getOperand(0).getValueType().getVectorElementType() != 4309 VT.getVectorElementType()) { 4310 // This code doesn't know how to handle shuffles where the vector 4311 // element types do not match (this happens because type legalization 4312 // promotes the return type of EXTRACT_VECTOR_ELT). 4313 // FIXME: It might be appropriate to extend this code to handle 4314 // mismatched types. 4315 return SDValue(); 4316 } 4317 4318 // Record this extraction against the appropriate vector if possible... 4319 SDValue SourceVec = V.getOperand(0); 4320 // If the element number isn't a constant, we can't effectively 4321 // analyze what's going on. 4322 if (!isa<ConstantSDNode>(V.getOperand(1))) 4323 return SDValue(); 4324 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue(); 4325 bool FoundSource = false; 4326 for (unsigned j = 0; j < SourceVecs.size(); ++j) { 4327 if (SourceVecs[j] == SourceVec) { 4328 if (MinElts[j] > EltNo) 4329 MinElts[j] = EltNo; 4330 if (MaxElts[j] < EltNo) 4331 MaxElts[j] = EltNo; 4332 FoundSource = true; 4333 break; 4334 } 4335 } 4336 4337 // Or record a new source if not... 4338 if (!FoundSource) { 4339 SourceVecs.push_back(SourceVec); 4340 MinElts.push_back(EltNo); 4341 MaxElts.push_back(EltNo); 4342 } 4343 } 4344 4345 // Currently only do something sane when at most two source vectors 4346 // involved. 4347 if (SourceVecs.size() > 2) 4348 return SDValue(); 4349 4350 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) }; 4351 int VEXTOffsets[2] = {0, 0}; 4352 4353 // This loop extracts the usage patterns of the source vectors 4354 // and prepares appropriate SDValues for a shuffle if possible. 4355 for (unsigned i = 0; i < SourceVecs.size(); ++i) { 4356 if (SourceVecs[i].getValueType() == VT) { 4357 // No VEXT necessary 4358 ShuffleSrcs[i] = SourceVecs[i]; 4359 VEXTOffsets[i] = 0; 4360 continue; 4361 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) { 4362 // It probably isn't worth padding out a smaller vector just to 4363 // break it down again in a shuffle. 4364 return SDValue(); 4365 } 4366 4367 // Since only 64-bit and 128-bit vectors are legal on ARM and 4368 // we've eliminated the other cases... 4369 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts && 4370 "unexpected vector sizes in ReconstructShuffle"); 4371 4372 if (MaxElts[i] - MinElts[i] >= NumElts) { 4373 // Span too large for a VEXT to cope 4374 return SDValue(); 4375 } 4376 4377 if (MinElts[i] >= NumElts) { 4378 // The extraction can just take the second half 4379 VEXTOffsets[i] = NumElts; 4380 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4381 SourceVecs[i], 4382 DAG.getIntPtrConstant(NumElts)); 4383 } else if (MaxElts[i] < NumElts) { 4384 // The extraction can just take the first half 4385 VEXTOffsets[i] = 0; 4386 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4387 SourceVecs[i], 4388 DAG.getIntPtrConstant(0)); 4389 } else { 4390 // An actual VEXT is needed 4391 VEXTOffsets[i] = MinElts[i]; 4392 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4393 SourceVecs[i], 4394 DAG.getIntPtrConstant(0)); 4395 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4396 SourceVecs[i], 4397 DAG.getIntPtrConstant(NumElts)); 4398 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2, 4399 DAG.getConstant(VEXTOffsets[i], MVT::i32)); 4400 } 4401 } 4402 4403 SmallVector<int, 8> Mask; 4404 4405 for (unsigned i = 0; i < NumElts; ++i) { 4406 SDValue Entry = Op.getOperand(i); 4407 if (Entry.getOpcode() == ISD::UNDEF) { 4408 Mask.push_back(-1); 4409 continue; 4410 } 4411 4412 SDValue ExtractVec = Entry.getOperand(0); 4413 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i) 4414 .getOperand(1))->getSExtValue(); 4415 if (ExtractVec == SourceVecs[0]) { 4416 Mask.push_back(ExtractElt - VEXTOffsets[0]); 4417 } else { 4418 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]); 4419 } 4420 } 4421 4422 // Final check before we try to produce nonsense... 4423 if (isShuffleMaskLegal(Mask, VT)) 4424 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1], 4425 &Mask[0]); 4426 4427 return SDValue(); 4428 } 4429 4430 /// isShuffleMaskLegal - Targets can use this to indicate that they only 4431 /// support *some* VECTOR_SHUFFLE operations, those with specific masks. 4432 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values 4433 /// are assumed to be legal. 4434 bool 4435 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M, 4436 EVT VT) const { 4437 if (VT.getVectorNumElements() == 4 && 4438 (VT.is128BitVector() || VT.is64BitVector())) { 4439 unsigned PFIndexes[4]; 4440 for (unsigned i = 0; i != 4; ++i) { 4441 if (M[i] < 0) 4442 PFIndexes[i] = 8; 4443 else 4444 PFIndexes[i] = M[i]; 4445 } 4446 4447 // Compute the index in the perfect shuffle table. 4448 unsigned PFTableIndex = 4449 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3]; 4450 unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; 4451 unsigned Cost = (PFEntry >> 30); 4452 4453 if (Cost <= 4) 4454 return true; 4455 } 4456 4457 bool ReverseVEXT; 4458 unsigned Imm, WhichResult; 4459 4460 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 4461 return (EltSize >= 32 || 4462 ShuffleVectorSDNode::isSplatMask(&M[0], VT) || 4463 isVREVMask(M, VT, 64) || 4464 isVREVMask(M, VT, 32) || 4465 isVREVMask(M, VT, 16) || 4466 isVEXTMask(M, VT, ReverseVEXT, Imm) || 4467 isVTBLMask(M, VT) || 4468 isVTRNMask(M, VT, WhichResult) || 4469 isVUZPMask(M, VT, WhichResult) || 4470 isVZIPMask(M, VT, WhichResult) || 4471 isVTRN_v_undef_Mask(M, VT, WhichResult) || 4472 isVUZP_v_undef_Mask(M, VT, WhichResult) || 4473 isVZIP_v_undef_Mask(M, VT, WhichResult)); 4474 } 4475 4476 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit 4477 /// the specified operations to build the shuffle. 4478 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS, 4479 SDValue RHS, SelectionDAG &DAG, 4480 DebugLoc dl) { 4481 unsigned OpNum = (PFEntry >> 26) & 0x0F; 4482 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1); 4483 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1); 4484 4485 enum { 4486 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3> 4487 OP_VREV, 4488 OP_VDUP0, 4489 OP_VDUP1, 4490 OP_VDUP2, 4491 OP_VDUP3, 4492 OP_VEXT1, 4493 OP_VEXT2, 4494 OP_VEXT3, 4495 OP_VUZPL, // VUZP, left result 4496 OP_VUZPR, // VUZP, right result 4497 OP_VZIPL, // VZIP, left result 4498 OP_VZIPR, // VZIP, right result 4499 OP_VTRNL, // VTRN, left result 4500 OP_VTRNR // VTRN, right result 4501 }; 4502 4503 if (OpNum == OP_COPY) { 4504 if (LHSID == (1*9+2)*9+3) return LHS; 4505 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!"); 4506 return RHS; 4507 } 4508 4509 SDValue OpLHS, OpRHS; 4510 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl); 4511 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl); 4512 EVT VT = OpLHS.getValueType(); 4513 4514 switch (OpNum) { 4515 default: llvm_unreachable("Unknown shuffle opcode!"); 4516 case OP_VREV: 4517 // VREV divides the vector in half and swaps within the half. 4518 if (VT.getVectorElementType() == MVT::i32 || 4519 VT.getVectorElementType() == MVT::f32) 4520 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS); 4521 // vrev <4 x i16> -> VREV32 4522 if (VT.getVectorElementType() == MVT::i16) 4523 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS); 4524 // vrev <4 x i8> -> VREV16 4525 assert(VT.getVectorElementType() == MVT::i8); 4526 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS); 4527 case OP_VDUP0: 4528 case OP_VDUP1: 4529 case OP_VDUP2: 4530 case OP_VDUP3: 4531 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, 4532 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32)); 4533 case OP_VEXT1: 4534 case OP_VEXT2: 4535 case OP_VEXT3: 4536 return DAG.getNode(ARMISD::VEXT, dl, VT, 4537 OpLHS, OpRHS, 4538 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32)); 4539 case OP_VUZPL: 4540 case OP_VUZPR: 4541 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 4542 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL); 4543 case OP_VZIPL: 4544 case OP_VZIPR: 4545 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 4546 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL); 4547 case OP_VTRNL: 4548 case OP_VTRNR: 4549 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 4550 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL); 4551 } 4552 } 4553 4554 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op, 4555 ArrayRef<int> ShuffleMask, 4556 SelectionDAG &DAG) { 4557 // Check to see if we can use the VTBL instruction. 4558 SDValue V1 = Op.getOperand(0); 4559 SDValue V2 = Op.getOperand(1); 4560 DebugLoc DL = Op.getDebugLoc(); 4561 4562 SmallVector<SDValue, 8> VTBLMask; 4563 for (ArrayRef<int>::iterator 4564 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I) 4565 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32)); 4566 4567 if (V2.getNode()->getOpcode() == ISD::UNDEF) 4568 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1, 4569 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, 4570 &VTBLMask[0], 8)); 4571 4572 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2, 4573 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, 4574 &VTBLMask[0], 8)); 4575 } 4576 4577 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) { 4578 SDValue V1 = Op.getOperand(0); 4579 SDValue V2 = Op.getOperand(1); 4580 DebugLoc dl = Op.getDebugLoc(); 4581 EVT VT = Op.getValueType(); 4582 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode()); 4583 4584 // Convert shuffles that are directly supported on NEON to target-specific 4585 // DAG nodes, instead of keeping them as shuffles and matching them again 4586 // during code selection. This is more efficient and avoids the possibility 4587 // of inconsistencies between legalization and selection. 4588 // FIXME: floating-point vectors should be canonicalized to integer vectors 4589 // of the same time so that they get CSEd properly. 4590 ArrayRef<int> ShuffleMask = SVN->getMask(); 4591 4592 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 4593 if (EltSize <= 32) { 4594 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) { 4595 int Lane = SVN->getSplatIndex(); 4596 // If this is undef splat, generate it via "just" vdup, if possible. 4597 if (Lane == -1) Lane = 0; 4598 4599 // Test if V1 is a SCALAR_TO_VECTOR. 4600 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) { 4601 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0)); 4602 } 4603 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR 4604 // (and probably will turn into a SCALAR_TO_VECTOR once legalization 4605 // reaches it). 4606 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR && 4607 !isa<ConstantSDNode>(V1.getOperand(0))) { 4608 bool IsScalarToVector = true; 4609 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i) 4610 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) { 4611 IsScalarToVector = false; 4612 break; 4613 } 4614 if (IsScalarToVector) 4615 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0)); 4616 } 4617 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1, 4618 DAG.getConstant(Lane, MVT::i32)); 4619 } 4620 4621 bool ReverseVEXT; 4622 unsigned Imm; 4623 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) { 4624 if (ReverseVEXT) 4625 std::swap(V1, V2); 4626 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2, 4627 DAG.getConstant(Imm, MVT::i32)); 4628 } 4629 4630 if (isVREVMask(ShuffleMask, VT, 64)) 4631 return DAG.getNode(ARMISD::VREV64, dl, VT, V1); 4632 if (isVREVMask(ShuffleMask, VT, 32)) 4633 return DAG.getNode(ARMISD::VREV32, dl, VT, V1); 4634 if (isVREVMask(ShuffleMask, VT, 16)) 4635 return DAG.getNode(ARMISD::VREV16, dl, VT, V1); 4636 4637 // Check for Neon shuffles that modify both input vectors in place. 4638 // If both results are used, i.e., if there are two shuffles with the same 4639 // source operands and with masks corresponding to both results of one of 4640 // these operations, DAG memoization will ensure that a single node is 4641 // used for both shuffles. 4642 unsigned WhichResult; 4643 if (isVTRNMask(ShuffleMask, VT, WhichResult)) 4644 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 4645 V1, V2).getValue(WhichResult); 4646 if (isVUZPMask(ShuffleMask, VT, WhichResult)) 4647 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 4648 V1, V2).getValue(WhichResult); 4649 if (isVZIPMask(ShuffleMask, VT, WhichResult)) 4650 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 4651 V1, V2).getValue(WhichResult); 4652 4653 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) 4654 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 4655 V1, V1).getValue(WhichResult); 4656 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) 4657 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 4658 V1, V1).getValue(WhichResult); 4659 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) 4660 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 4661 V1, V1).getValue(WhichResult); 4662 } 4663 4664 // If the shuffle is not directly supported and it has 4 elements, use 4665 // the PerfectShuffle-generated table to synthesize it from other shuffles. 4666 unsigned NumElts = VT.getVectorNumElements(); 4667 if (NumElts == 4) { 4668 unsigned PFIndexes[4]; 4669 for (unsigned i = 0; i != 4; ++i) { 4670 if (ShuffleMask[i] < 0) 4671 PFIndexes[i] = 8; 4672 else 4673 PFIndexes[i] = ShuffleMask[i]; 4674 } 4675 4676 // Compute the index in the perfect shuffle table. 4677 unsigned PFTableIndex = 4678 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3]; 4679 unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; 4680 unsigned Cost = (PFEntry >> 30); 4681 4682 if (Cost <= 4) 4683 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl); 4684 } 4685 4686 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs. 4687 if (EltSize >= 32) { 4688 // Do the expansion with floating-point types, since that is what the VFP 4689 // registers are defined to use, and since i64 is not legal. 4690 EVT EltVT = EVT::getFloatingPointVT(EltSize); 4691 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts); 4692 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1); 4693 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2); 4694 SmallVector<SDValue, 8> Ops; 4695 for (unsigned i = 0; i < NumElts; ++i) { 4696 if (ShuffleMask[i] < 0) 4697 Ops.push_back(DAG.getUNDEF(EltVT)); 4698 else 4699 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, 4700 ShuffleMask[i] < (int)NumElts ? V1 : V2, 4701 DAG.getConstant(ShuffleMask[i] & (NumElts-1), 4702 MVT::i32))); 4703 } 4704 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts); 4705 return DAG.getNode(ISD::BITCAST, dl, VT, Val); 4706 } 4707 4708 if (VT == MVT::v8i8) { 4709 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG); 4710 if (NewOp.getNode()) 4711 return NewOp; 4712 } 4713 4714 return SDValue(); 4715 } 4716 4717 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) { 4718 // INSERT_VECTOR_ELT is legal only for immediate indexes. 4719 SDValue Lane = Op.getOperand(2); 4720 if (!isa<ConstantSDNode>(Lane)) 4721 return SDValue(); 4722 4723 return Op; 4724 } 4725 4726 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) { 4727 // EXTRACT_VECTOR_ELT is legal only for immediate indexes. 4728 SDValue Lane = Op.getOperand(1); 4729 if (!isa<ConstantSDNode>(Lane)) 4730 return SDValue(); 4731 4732 SDValue Vec = Op.getOperand(0); 4733 if (Op.getValueType() == MVT::i32 && 4734 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) { 4735 DebugLoc dl = Op.getDebugLoc(); 4736 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane); 4737 } 4738 4739 return Op; 4740 } 4741 4742 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) { 4743 // The only time a CONCAT_VECTORS operation can have legal types is when 4744 // two 64-bit vectors are concatenated to a 128-bit vector. 4745 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 && 4746 "unexpected CONCAT_VECTORS"); 4747 DebugLoc dl = Op.getDebugLoc(); 4748 SDValue Val = DAG.getUNDEF(MVT::v2f64); 4749 SDValue Op0 = Op.getOperand(0); 4750 SDValue Op1 = Op.getOperand(1); 4751 if (Op0.getOpcode() != ISD::UNDEF) 4752 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val, 4753 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0), 4754 DAG.getIntPtrConstant(0)); 4755 if (Op1.getOpcode() != ISD::UNDEF) 4756 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val, 4757 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1), 4758 DAG.getIntPtrConstant(1)); 4759 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val); 4760 } 4761 4762 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each 4763 /// element has been zero/sign-extended, depending on the isSigned parameter, 4764 /// from an integer type half its size. 4765 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG, 4766 bool isSigned) { 4767 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32. 4768 EVT VT = N->getValueType(0); 4769 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) { 4770 SDNode *BVN = N->getOperand(0).getNode(); 4771 if (BVN->getValueType(0) != MVT::v4i32 || 4772 BVN->getOpcode() != ISD::BUILD_VECTOR) 4773 return false; 4774 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0; 4775 unsigned HiElt = 1 - LoElt; 4776 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt)); 4777 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt)); 4778 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2)); 4779 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2)); 4780 if (!Lo0 || !Hi0 || !Lo1 || !Hi1) 4781 return false; 4782 if (isSigned) { 4783 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 && 4784 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32) 4785 return true; 4786 } else { 4787 if (Hi0->isNullValue() && Hi1->isNullValue()) 4788 return true; 4789 } 4790 return false; 4791 } 4792 4793 if (N->getOpcode() != ISD::BUILD_VECTOR) 4794 return false; 4795 4796 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 4797 SDNode *Elt = N->getOperand(i).getNode(); 4798 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) { 4799 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 4800 unsigned HalfSize = EltSize / 2; 4801 if (isSigned) { 4802 if (!isIntN(HalfSize, C->getSExtValue())) 4803 return false; 4804 } else { 4805 if (!isUIntN(HalfSize, C->getZExtValue())) 4806 return false; 4807 } 4808 continue; 4809 } 4810 return false; 4811 } 4812 4813 return true; 4814 } 4815 4816 /// isSignExtended - Check if a node is a vector value that is sign-extended 4817 /// or a constant BUILD_VECTOR with sign-extended elements. 4818 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) { 4819 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N)) 4820 return true; 4821 if (isExtendedBUILD_VECTOR(N, DAG, true)) 4822 return true; 4823 return false; 4824 } 4825 4826 /// isZeroExtended - Check if a node is a vector value that is zero-extended 4827 /// or a constant BUILD_VECTOR with zero-extended elements. 4828 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) { 4829 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N)) 4830 return true; 4831 if (isExtendedBUILD_VECTOR(N, DAG, false)) 4832 return true; 4833 return false; 4834 } 4835 4836 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending 4837 /// load, or BUILD_VECTOR with extended elements, return the unextended value. 4838 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) { 4839 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND) 4840 return N->getOperand(0); 4841 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) 4842 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(), 4843 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(), 4844 LD->isNonTemporal(), LD->isInvariant(), 4845 LD->getAlignment()); 4846 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will 4847 // have been legalized as a BITCAST from v4i32. 4848 if (N->getOpcode() == ISD::BITCAST) { 4849 SDNode *BVN = N->getOperand(0).getNode(); 4850 assert(BVN->getOpcode() == ISD::BUILD_VECTOR && 4851 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR"); 4852 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0; 4853 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32, 4854 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2)); 4855 } 4856 // Construct a new BUILD_VECTOR with elements truncated to half the size. 4857 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR"); 4858 EVT VT = N->getValueType(0); 4859 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2; 4860 unsigned NumElts = VT.getVectorNumElements(); 4861 MVT TruncVT = MVT::getIntegerVT(EltSize); 4862 SmallVector<SDValue, 8> Ops; 4863 for (unsigned i = 0; i != NumElts; ++i) { 4864 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i)); 4865 const APInt &CInt = C->getAPIntValue(); 4866 // Element types smaller than 32 bits are not legal, so use i32 elements. 4867 // The values are implicitly truncated so sext vs. zext doesn't matter. 4868 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32)); 4869 } 4870 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), 4871 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts); 4872 } 4873 4874 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) { 4875 unsigned Opcode = N->getOpcode(); 4876 if (Opcode == ISD::ADD || Opcode == ISD::SUB) { 4877 SDNode *N0 = N->getOperand(0).getNode(); 4878 SDNode *N1 = N->getOperand(1).getNode(); 4879 return N0->hasOneUse() && N1->hasOneUse() && 4880 isSignExtended(N0, DAG) && isSignExtended(N1, DAG); 4881 } 4882 return false; 4883 } 4884 4885 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) { 4886 unsigned Opcode = N->getOpcode(); 4887 if (Opcode == ISD::ADD || Opcode == ISD::SUB) { 4888 SDNode *N0 = N->getOperand(0).getNode(); 4889 SDNode *N1 = N->getOperand(1).getNode(); 4890 return N0->hasOneUse() && N1->hasOneUse() && 4891 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG); 4892 } 4893 return false; 4894 } 4895 4896 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) { 4897 // Multiplications are only custom-lowered for 128-bit vectors so that 4898 // VMULL can be detected. Otherwise v2i64 multiplications are not legal. 4899 EVT VT = Op.getValueType(); 4900 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL"); 4901 SDNode *N0 = Op.getOperand(0).getNode(); 4902 SDNode *N1 = Op.getOperand(1).getNode(); 4903 unsigned NewOpc = 0; 4904 bool isMLA = false; 4905 bool isN0SExt = isSignExtended(N0, DAG); 4906 bool isN1SExt = isSignExtended(N1, DAG); 4907 if (isN0SExt && isN1SExt) 4908 NewOpc = ARMISD::VMULLs; 4909 else { 4910 bool isN0ZExt = isZeroExtended(N0, DAG); 4911 bool isN1ZExt = isZeroExtended(N1, DAG); 4912 if (isN0ZExt && isN1ZExt) 4913 NewOpc = ARMISD::VMULLu; 4914 else if (isN1SExt || isN1ZExt) { 4915 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these 4916 // into (s/zext A * s/zext C) + (s/zext B * s/zext C) 4917 if (isN1SExt && isAddSubSExt(N0, DAG)) { 4918 NewOpc = ARMISD::VMULLs; 4919 isMLA = true; 4920 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) { 4921 NewOpc = ARMISD::VMULLu; 4922 isMLA = true; 4923 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) { 4924 std::swap(N0, N1); 4925 NewOpc = ARMISD::VMULLu; 4926 isMLA = true; 4927 } 4928 } 4929 4930 if (!NewOpc) { 4931 if (VT == MVT::v2i64) 4932 // Fall through to expand this. It is not legal. 4933 return SDValue(); 4934 else 4935 // Other vector multiplications are legal. 4936 return Op; 4937 } 4938 } 4939 4940 // Legalize to a VMULL instruction. 4941 DebugLoc DL = Op.getDebugLoc(); 4942 SDValue Op0; 4943 SDValue Op1 = SkipExtension(N1, DAG); 4944 if (!isMLA) { 4945 Op0 = SkipExtension(N0, DAG); 4946 assert(Op0.getValueType().is64BitVector() && 4947 Op1.getValueType().is64BitVector() && 4948 "unexpected types for extended operands to VMULL"); 4949 return DAG.getNode(NewOpc, DL, VT, Op0, Op1); 4950 } 4951 4952 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during 4953 // isel lowering to take advantage of no-stall back to back vmul + vmla. 4954 // vmull q0, d4, d6 4955 // vmlal q0, d5, d6 4956 // is faster than 4957 // vaddl q0, d4, d5 4958 // vmovl q1, d6 4959 // vmul q0, q0, q1 4960 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG); 4961 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG); 4962 EVT Op1VT = Op1.getValueType(); 4963 return DAG.getNode(N0->getOpcode(), DL, VT, 4964 DAG.getNode(NewOpc, DL, VT, 4965 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1), 4966 DAG.getNode(NewOpc, DL, VT, 4967 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1)); 4968 } 4969 4970 static SDValue 4971 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) { 4972 // Convert to float 4973 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo)); 4974 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo)); 4975 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X); 4976 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y); 4977 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X); 4978 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y); 4979 // Get reciprocal estimate. 4980 // float4 recip = vrecpeq_f32(yf); 4981 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 4982 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y); 4983 // Because char has a smaller range than uchar, we can actually get away 4984 // without any newton steps. This requires that we use a weird bias 4985 // of 0xb000, however (again, this has been exhaustively tested). 4986 // float4 result = as_float4(as_int4(xf*recip) + 0xb000); 4987 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y); 4988 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X); 4989 Y = DAG.getConstant(0xb000, MVT::i32); 4990 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y); 4991 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y); 4992 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X); 4993 // Convert back to short. 4994 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X); 4995 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X); 4996 return X; 4997 } 4998 4999 static SDValue 5000 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) { 5001 SDValue N2; 5002 // Convert to float. 5003 // float4 yf = vcvt_f32_s32(vmovl_s16(y)); 5004 // float4 xf = vcvt_f32_s32(vmovl_s16(x)); 5005 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0); 5006 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1); 5007 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0); 5008 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1); 5009 5010 // Use reciprocal estimate and one refinement step. 5011 // float4 recip = vrecpeq_f32(yf); 5012 // recip *= vrecpsq_f32(yf, recip); 5013 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5014 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1); 5015 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5016 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), 5017 N1, N2); 5018 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); 5019 // Because short has a smaller range than ushort, we can actually get away 5020 // with only a single newton step. This requires that we use a weird bias 5021 // of 89, however (again, this has been exhaustively tested). 5022 // float4 result = as_float4(as_int4(xf*recip) + 0x89); 5023 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2); 5024 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0); 5025 N1 = DAG.getConstant(0x89, MVT::i32); 5026 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1); 5027 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1); 5028 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0); 5029 // Convert back to integer and return. 5030 // return vmovn_s32(vcvt_s32_f32(result)); 5031 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0); 5032 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0); 5033 return N0; 5034 } 5035 5036 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) { 5037 EVT VT = Op.getValueType(); 5038 assert((VT == MVT::v4i16 || VT == MVT::v8i8) && 5039 "unexpected type for custom-lowering ISD::SDIV"); 5040 5041 DebugLoc dl = Op.getDebugLoc(); 5042 SDValue N0 = Op.getOperand(0); 5043 SDValue N1 = Op.getOperand(1); 5044 SDValue N2, N3; 5045 5046 if (VT == MVT::v8i8) { 5047 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0); 5048 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1); 5049 5050 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5051 DAG.getIntPtrConstant(4)); 5052 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5053 DAG.getIntPtrConstant(4)); 5054 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5055 DAG.getIntPtrConstant(0)); 5056 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5057 DAG.getIntPtrConstant(0)); 5058 5059 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16 5060 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16 5061 5062 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2); 5063 N0 = LowerCONCAT_VECTORS(N0, DAG); 5064 5065 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0); 5066 return N0; 5067 } 5068 return LowerSDIV_v4i16(N0, N1, dl, DAG); 5069 } 5070 5071 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) { 5072 EVT VT = Op.getValueType(); 5073 assert((VT == MVT::v4i16 || VT == MVT::v8i8) && 5074 "unexpected type for custom-lowering ISD::UDIV"); 5075 5076 DebugLoc dl = Op.getDebugLoc(); 5077 SDValue N0 = Op.getOperand(0); 5078 SDValue N1 = Op.getOperand(1); 5079 SDValue N2, N3; 5080 5081 if (VT == MVT::v8i8) { 5082 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0); 5083 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1); 5084 5085 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5086 DAG.getIntPtrConstant(4)); 5087 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5088 DAG.getIntPtrConstant(4)); 5089 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5090 DAG.getIntPtrConstant(0)); 5091 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5092 DAG.getIntPtrConstant(0)); 5093 5094 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16 5095 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16 5096 5097 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2); 5098 N0 = LowerCONCAT_VECTORS(N0, DAG); 5099 5100 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8, 5101 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32), 5102 N0); 5103 return N0; 5104 } 5105 5106 // v4i16 sdiv ... Convert to float. 5107 // float4 yf = vcvt_f32_s32(vmovl_u16(y)); 5108 // float4 xf = vcvt_f32_s32(vmovl_u16(x)); 5109 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0); 5110 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1); 5111 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0); 5112 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1); 5113 5114 // Use reciprocal estimate and two refinement steps. 5115 // float4 recip = vrecpeq_f32(yf); 5116 // recip *= vrecpsq_f32(yf, recip); 5117 // recip *= vrecpsq_f32(yf, recip); 5118 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5119 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1); 5120 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5121 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), 5122 BN1, N2); 5123 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); 5124 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5125 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), 5126 BN1, N2); 5127 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); 5128 // Simply multiplying by the reciprocal estimate can leave us a few ulps 5129 // too low, so we add 2 ulps (exhaustive testing shows that this is enough, 5130 // and that it will never cause us to return an answer too large). 5131 // float4 result = as_float4(as_int4(xf*recip) + 2); 5132 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2); 5133 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0); 5134 N1 = DAG.getConstant(2, MVT::i32); 5135 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1); 5136 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1); 5137 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0); 5138 // Convert back to integer and return. 5139 // return vmovn_u32(vcvt_s32_f32(result)); 5140 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0); 5141 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0); 5142 return N0; 5143 } 5144 5145 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) { 5146 EVT VT = Op.getNode()->getValueType(0); 5147 SDVTList VTs = DAG.getVTList(VT, MVT::i32); 5148 5149 unsigned Opc; 5150 bool ExtraOp = false; 5151 switch (Op.getOpcode()) { 5152 default: llvm_unreachable("Invalid code"); 5153 case ISD::ADDC: Opc = ARMISD::ADDC; break; 5154 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break; 5155 case ISD::SUBC: Opc = ARMISD::SUBC; break; 5156 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break; 5157 } 5158 5159 if (!ExtraOp) 5160 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0), 5161 Op.getOperand(1)); 5162 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0), 5163 Op.getOperand(1), Op.getOperand(2)); 5164 } 5165 5166 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) { 5167 // Monotonic load/store is legal for all targets 5168 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic) 5169 return Op; 5170 5171 // Aquire/Release load/store is not legal for targets without a 5172 // dmb or equivalent available. 5173 return SDValue(); 5174 } 5175 5176 5177 static void 5178 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results, 5179 SelectionDAG &DAG, unsigned NewOp) { 5180 DebugLoc dl = Node->getDebugLoc(); 5181 assert (Node->getValueType(0) == MVT::i64 && 5182 "Only know how to expand i64 atomics"); 5183 5184 SmallVector<SDValue, 6> Ops; 5185 Ops.push_back(Node->getOperand(0)); // Chain 5186 Ops.push_back(Node->getOperand(1)); // Ptr 5187 // Low part of Val1 5188 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5189 Node->getOperand(2), DAG.getIntPtrConstant(0))); 5190 // High part of Val1 5191 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5192 Node->getOperand(2), DAG.getIntPtrConstant(1))); 5193 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) { 5194 // High part of Val1 5195 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5196 Node->getOperand(3), DAG.getIntPtrConstant(0))); 5197 // High part of Val2 5198 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5199 Node->getOperand(3), DAG.getIntPtrConstant(1))); 5200 } 5201 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other); 5202 SDValue Result = 5203 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64, 5204 cast<MemSDNode>(Node)->getMemOperand()); 5205 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) }; 5206 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2)); 5207 Results.push_back(Result.getValue(2)); 5208 } 5209 5210 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { 5211 switch (Op.getOpcode()) { 5212 default: llvm_unreachable("Don't know how to custom lower this!"); 5213 case ISD::ConstantPool: return LowerConstantPool(Op, DAG); 5214 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); 5215 case ISD::GlobalAddress: 5216 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) : 5217 LowerGlobalAddressELF(Op, DAG); 5218 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); 5219 case ISD::SELECT: return LowerSELECT(Op, DAG); 5220 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG); 5221 case ISD::BR_CC: return LowerBR_CC(Op, DAG); 5222 case ISD::BR_JT: return LowerBR_JT(Op, DAG); 5223 case ISD::VASTART: return LowerVASTART(Op, DAG); 5224 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget); 5225 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget); 5226 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget); 5227 case ISD::SINT_TO_FP: 5228 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG); 5229 case ISD::FP_TO_SINT: 5230 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG); 5231 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG); 5232 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG); 5233 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); 5234 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG); 5235 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG); 5236 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG); 5237 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG, 5238 Subtarget); 5239 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG); 5240 case ISD::SHL: 5241 case ISD::SRL: 5242 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget); 5243 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG); 5244 case ISD::SRL_PARTS: 5245 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG); 5246 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget); 5247 case ISD::SETCC: return LowerVSETCC(Op, DAG); 5248 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget); 5249 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget); 5250 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG); 5251 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG); 5252 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG); 5253 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG); 5254 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG); 5255 case ISD::MUL: return LowerMUL(Op, DAG); 5256 case ISD::SDIV: return LowerSDIV(Op, DAG); 5257 case ISD::UDIV: return LowerUDIV(Op, DAG); 5258 case ISD::ADDC: 5259 case ISD::ADDE: 5260 case ISD::SUBC: 5261 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG); 5262 case ISD::ATOMIC_LOAD: 5263 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG); 5264 } 5265 } 5266 5267 /// ReplaceNodeResults - Replace the results of node with an illegal result 5268 /// type with new values built out of custom code. 5269 void ARMTargetLowering::ReplaceNodeResults(SDNode *N, 5270 SmallVectorImpl<SDValue>&Results, 5271 SelectionDAG &DAG) const { 5272 SDValue Res; 5273 switch (N->getOpcode()) { 5274 default: 5275 llvm_unreachable("Don't know how to custom expand this!"); 5276 case ISD::BITCAST: 5277 Res = ExpandBITCAST(N, DAG); 5278 break; 5279 case ISD::SRL: 5280 case ISD::SRA: 5281 Res = Expand64BitShift(N, DAG, Subtarget); 5282 break; 5283 case ISD::ATOMIC_LOAD_ADD: 5284 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG); 5285 return; 5286 case ISD::ATOMIC_LOAD_AND: 5287 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG); 5288 return; 5289 case ISD::ATOMIC_LOAD_NAND: 5290 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG); 5291 return; 5292 case ISD::ATOMIC_LOAD_OR: 5293 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG); 5294 return; 5295 case ISD::ATOMIC_LOAD_SUB: 5296 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG); 5297 return; 5298 case ISD::ATOMIC_LOAD_XOR: 5299 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG); 5300 return; 5301 case ISD::ATOMIC_SWAP: 5302 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG); 5303 return; 5304 case ISD::ATOMIC_CMP_SWAP: 5305 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG); 5306 return; 5307 } 5308 if (Res.getNode()) 5309 Results.push_back(Res); 5310 } 5311 5312 //===----------------------------------------------------------------------===// 5313 // ARM Scheduler Hooks 5314 //===----------------------------------------------------------------------===// 5315 5316 MachineBasicBlock * 5317 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI, 5318 MachineBasicBlock *BB, 5319 unsigned Size) const { 5320 unsigned dest = MI->getOperand(0).getReg(); 5321 unsigned ptr = MI->getOperand(1).getReg(); 5322 unsigned oldval = MI->getOperand(2).getReg(); 5323 unsigned newval = MI->getOperand(3).getReg(); 5324 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5325 DebugLoc dl = MI->getDebugLoc(); 5326 bool isThumb2 = Subtarget->isThumb2(); 5327 5328 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 5329 unsigned scratch = MRI.createVirtualRegister(isThumb2 ? 5330 (const TargetRegisterClass*)&ARM::rGPRRegClass : 5331 (const TargetRegisterClass*)&ARM::GPRRegClass); 5332 5333 if (isThumb2) { 5334 MRI.constrainRegClass(dest, &ARM::rGPRRegClass); 5335 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass); 5336 MRI.constrainRegClass(newval, &ARM::rGPRRegClass); 5337 } 5338 5339 unsigned ldrOpc, strOpc; 5340 switch (Size) { 5341 default: llvm_unreachable("unsupported size for AtomicCmpSwap!"); 5342 case 1: 5343 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB; 5344 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB; 5345 break; 5346 case 2: 5347 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH; 5348 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH; 5349 break; 5350 case 4: 5351 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX; 5352 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX; 5353 break; 5354 } 5355 5356 MachineFunction *MF = BB->getParent(); 5357 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 5358 MachineFunction::iterator It = BB; 5359 ++It; // insert the new blocks after the current block 5360 5361 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 5362 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 5363 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5364 MF->insert(It, loop1MBB); 5365 MF->insert(It, loop2MBB); 5366 MF->insert(It, exitMBB); 5367 5368 // Transfer the remainder of BB and its successor edges to exitMBB. 5369 exitMBB->splice(exitMBB->begin(), BB, 5370 llvm::next(MachineBasicBlock::iterator(MI)), 5371 BB->end()); 5372 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 5373 5374 // thisMBB: 5375 // ... 5376 // fallthrough --> loop1MBB 5377 BB->addSuccessor(loop1MBB); 5378 5379 // loop1MBB: 5380 // ldrex dest, [ptr] 5381 // cmp dest, oldval 5382 // bne exitMBB 5383 BB = loop1MBB; 5384 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr); 5385 if (ldrOpc == ARM::t2LDREX) 5386 MIB.addImm(0); 5387 AddDefaultPred(MIB); 5388 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) 5389 .addReg(dest).addReg(oldval)); 5390 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5391 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5392 BB->addSuccessor(loop2MBB); 5393 BB->addSuccessor(exitMBB); 5394 5395 // loop2MBB: 5396 // strex scratch, newval, [ptr] 5397 // cmp scratch, #0 5398 // bne loop1MBB 5399 BB = loop2MBB; 5400 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr); 5401 if (strOpc == ARM::t2STREX) 5402 MIB.addImm(0); 5403 AddDefaultPred(MIB); 5404 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 5405 .addReg(scratch).addImm(0)); 5406 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5407 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5408 BB->addSuccessor(loop1MBB); 5409 BB->addSuccessor(exitMBB); 5410 5411 // exitMBB: 5412 // ... 5413 BB = exitMBB; 5414 5415 MI->eraseFromParent(); // The instruction is gone now. 5416 5417 return BB; 5418 } 5419 5420 MachineBasicBlock * 5421 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB, 5422 unsigned Size, unsigned BinOpcode) const { 5423 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0. 5424 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5425 5426 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 5427 MachineFunction *MF = BB->getParent(); 5428 MachineFunction::iterator It = BB; 5429 ++It; 5430 5431 unsigned dest = MI->getOperand(0).getReg(); 5432 unsigned ptr = MI->getOperand(1).getReg(); 5433 unsigned incr = MI->getOperand(2).getReg(); 5434 DebugLoc dl = MI->getDebugLoc(); 5435 bool isThumb2 = Subtarget->isThumb2(); 5436 5437 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 5438 if (isThumb2) { 5439 MRI.constrainRegClass(dest, &ARM::rGPRRegClass); 5440 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass); 5441 } 5442 5443 unsigned ldrOpc, strOpc; 5444 switch (Size) { 5445 default: llvm_unreachable("unsupported size for AtomicCmpSwap!"); 5446 case 1: 5447 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB; 5448 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB; 5449 break; 5450 case 2: 5451 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH; 5452 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH; 5453 break; 5454 case 4: 5455 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX; 5456 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX; 5457 break; 5458 } 5459 5460 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5461 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5462 MF->insert(It, loopMBB); 5463 MF->insert(It, exitMBB); 5464 5465 // Transfer the remainder of BB and its successor edges to exitMBB. 5466 exitMBB->splice(exitMBB->begin(), BB, 5467 llvm::next(MachineBasicBlock::iterator(MI)), 5468 BB->end()); 5469 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 5470 5471 const TargetRegisterClass *TRC = isThumb2 ? 5472 (const TargetRegisterClass*)&ARM::rGPRRegClass : 5473 (const TargetRegisterClass*)&ARM::GPRRegClass; 5474 unsigned scratch = MRI.createVirtualRegister(TRC); 5475 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC); 5476 5477 // thisMBB: 5478 // ... 5479 // fallthrough --> loopMBB 5480 BB->addSuccessor(loopMBB); 5481 5482 // loopMBB: 5483 // ldrex dest, ptr 5484 // <binop> scratch2, dest, incr 5485 // strex scratch, scratch2, ptr 5486 // cmp scratch, #0 5487 // bne- loopMBB 5488 // fallthrough --> exitMBB 5489 BB = loopMBB; 5490 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr); 5491 if (ldrOpc == ARM::t2LDREX) 5492 MIB.addImm(0); 5493 AddDefaultPred(MIB); 5494 if (BinOpcode) { 5495 // operand order needs to go the other way for NAND 5496 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr) 5497 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2). 5498 addReg(incr).addReg(dest)).addReg(0); 5499 else 5500 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2). 5501 addReg(dest).addReg(incr)).addReg(0); 5502 } 5503 5504 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr); 5505 if (strOpc == ARM::t2STREX) 5506 MIB.addImm(0); 5507 AddDefaultPred(MIB); 5508 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 5509 .addReg(scratch).addImm(0)); 5510 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5511 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5512 5513 BB->addSuccessor(loopMBB); 5514 BB->addSuccessor(exitMBB); 5515 5516 // exitMBB: 5517 // ... 5518 BB = exitMBB; 5519 5520 MI->eraseFromParent(); // The instruction is gone now. 5521 5522 return BB; 5523 } 5524 5525 MachineBasicBlock * 5526 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI, 5527 MachineBasicBlock *BB, 5528 unsigned Size, 5529 bool signExtend, 5530 ARMCC::CondCodes Cond) const { 5531 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5532 5533 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 5534 MachineFunction *MF = BB->getParent(); 5535 MachineFunction::iterator It = BB; 5536 ++It; 5537 5538 unsigned dest = MI->getOperand(0).getReg(); 5539 unsigned ptr = MI->getOperand(1).getReg(); 5540 unsigned incr = MI->getOperand(2).getReg(); 5541 unsigned oldval = dest; 5542 DebugLoc dl = MI->getDebugLoc(); 5543 bool isThumb2 = Subtarget->isThumb2(); 5544 5545 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 5546 if (isThumb2) { 5547 MRI.constrainRegClass(dest, &ARM::rGPRRegClass); 5548 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass); 5549 } 5550 5551 unsigned ldrOpc, strOpc, extendOpc; 5552 switch (Size) { 5553 default: llvm_unreachable("unsupported size for AtomicCmpSwap!"); 5554 case 1: 5555 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB; 5556 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB; 5557 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB; 5558 break; 5559 case 2: 5560 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH; 5561 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH; 5562 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH; 5563 break; 5564 case 4: 5565 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX; 5566 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX; 5567 extendOpc = 0; 5568 break; 5569 } 5570 5571 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5572 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5573 MF->insert(It, loopMBB); 5574 MF->insert(It, exitMBB); 5575 5576 // Transfer the remainder of BB and its successor edges to exitMBB. 5577 exitMBB->splice(exitMBB->begin(), BB, 5578 llvm::next(MachineBasicBlock::iterator(MI)), 5579 BB->end()); 5580 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 5581 5582 const TargetRegisterClass *TRC = isThumb2 ? 5583 (const TargetRegisterClass*)&ARM::rGPRRegClass : 5584 (const TargetRegisterClass*)&ARM::GPRRegClass; 5585 unsigned scratch = MRI.createVirtualRegister(TRC); 5586 unsigned scratch2 = MRI.createVirtualRegister(TRC); 5587 5588 // thisMBB: 5589 // ... 5590 // fallthrough --> loopMBB 5591 BB->addSuccessor(loopMBB); 5592 5593 // loopMBB: 5594 // ldrex dest, ptr 5595 // (sign extend dest, if required) 5596 // cmp dest, incr 5597 // cmov.cond scratch2, dest, incr 5598 // strex scratch, scratch2, ptr 5599 // cmp scratch, #0 5600 // bne- loopMBB 5601 // fallthrough --> exitMBB 5602 BB = loopMBB; 5603 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr); 5604 if (ldrOpc == ARM::t2LDREX) 5605 MIB.addImm(0); 5606 AddDefaultPred(MIB); 5607 5608 // Sign extend the value, if necessary. 5609 if (signExtend && extendOpc) { 5610 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass); 5611 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval) 5612 .addReg(dest) 5613 .addImm(0)); 5614 } 5615 5616 // Build compare and cmov instructions. 5617 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) 5618 .addReg(oldval).addReg(incr)); 5619 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2) 5620 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR); 5621 5622 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr); 5623 if (strOpc == ARM::t2STREX) 5624 MIB.addImm(0); 5625 AddDefaultPred(MIB); 5626 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 5627 .addReg(scratch).addImm(0)); 5628 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5629 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5630 5631 BB->addSuccessor(loopMBB); 5632 BB->addSuccessor(exitMBB); 5633 5634 // exitMBB: 5635 // ... 5636 BB = exitMBB; 5637 5638 MI->eraseFromParent(); // The instruction is gone now. 5639 5640 return BB; 5641 } 5642 5643 MachineBasicBlock * 5644 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB, 5645 unsigned Op1, unsigned Op2, 5646 bool NeedsCarry, bool IsCmpxchg) const { 5647 // This also handles ATOMIC_SWAP, indicated by Op1==0. 5648 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5649 5650 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 5651 MachineFunction *MF = BB->getParent(); 5652 MachineFunction::iterator It = BB; 5653 ++It; 5654 5655 unsigned destlo = MI->getOperand(0).getReg(); 5656 unsigned desthi = MI->getOperand(1).getReg(); 5657 unsigned ptr = MI->getOperand(2).getReg(); 5658 unsigned vallo = MI->getOperand(3).getReg(); 5659 unsigned valhi = MI->getOperand(4).getReg(); 5660 DebugLoc dl = MI->getDebugLoc(); 5661 bool isThumb2 = Subtarget->isThumb2(); 5662 5663 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 5664 if (isThumb2) { 5665 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass); 5666 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass); 5667 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass); 5668 } 5669 5670 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD; 5671 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD; 5672 5673 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5674 MachineBasicBlock *contBB = 0, *cont2BB = 0; 5675 if (IsCmpxchg) { 5676 contBB = MF->CreateMachineBasicBlock(LLVM_BB); 5677 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB); 5678 } 5679 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 5680 MF->insert(It, loopMBB); 5681 if (IsCmpxchg) { 5682 MF->insert(It, contBB); 5683 MF->insert(It, cont2BB); 5684 } 5685 MF->insert(It, exitMBB); 5686 5687 // Transfer the remainder of BB and its successor edges to exitMBB. 5688 exitMBB->splice(exitMBB->begin(), BB, 5689 llvm::next(MachineBasicBlock::iterator(MI)), 5690 BB->end()); 5691 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 5692 5693 const TargetRegisterClass *TRC = isThumb2 ? 5694 (const TargetRegisterClass*)&ARM::tGPRRegClass : 5695 (const TargetRegisterClass*)&ARM::GPRRegClass; 5696 unsigned storesuccess = MRI.createVirtualRegister(TRC); 5697 5698 // thisMBB: 5699 // ... 5700 // fallthrough --> loopMBB 5701 BB->addSuccessor(loopMBB); 5702 5703 // loopMBB: 5704 // ldrexd r2, r3, ptr 5705 // <binopa> r0, r2, incr 5706 // <binopb> r1, r3, incr 5707 // strexd storesuccess, r0, r1, ptr 5708 // cmp storesuccess, #0 5709 // bne- loopMBB 5710 // fallthrough --> exitMBB 5711 // 5712 // Note that the registers are explicitly specified because there is not any 5713 // way to force the register allocator to allocate a register pair. 5714 // 5715 // FIXME: The hardcoded registers are not necessary for Thumb2, but we 5716 // need to properly enforce the restriction that the two output registers 5717 // for ldrexd must be different. 5718 BB = loopMBB; 5719 // Load 5720 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc)) 5721 .addReg(ARM::R2, RegState::Define) 5722 .addReg(ARM::R3, RegState::Define).addReg(ptr)); 5723 // Copy r2/r3 into dest. (This copy will normally be coalesced.) 5724 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2); 5725 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3); 5726 5727 if (IsCmpxchg) { 5728 // Add early exit 5729 for (unsigned i = 0; i < 2; i++) { 5730 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : 5731 ARM::CMPrr)) 5732 .addReg(i == 0 ? destlo : desthi) 5733 .addReg(i == 0 ? vallo : valhi)); 5734 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5735 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5736 BB->addSuccessor(exitMBB); 5737 BB->addSuccessor(i == 0 ? contBB : cont2BB); 5738 BB = (i == 0 ? contBB : cont2BB); 5739 } 5740 5741 // Copy to physregs for strexd 5742 unsigned setlo = MI->getOperand(5).getReg(); 5743 unsigned sethi = MI->getOperand(6).getReg(); 5744 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo); 5745 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi); 5746 } else if (Op1) { 5747 // Perform binary operation 5748 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0) 5749 .addReg(destlo).addReg(vallo)) 5750 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry)); 5751 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1) 5752 .addReg(desthi).addReg(valhi)).addReg(0); 5753 } else { 5754 // Copy to physregs for strexd 5755 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo); 5756 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi); 5757 } 5758 5759 // Store 5760 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess) 5761 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr)); 5762 // Cmp+jump 5763 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 5764 .addReg(storesuccess).addImm(0)); 5765 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 5766 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 5767 5768 BB->addSuccessor(loopMBB); 5769 BB->addSuccessor(exitMBB); 5770 5771 // exitMBB: 5772 // ... 5773 BB = exitMBB; 5774 5775 MI->eraseFromParent(); // The instruction is gone now. 5776 5777 return BB; 5778 } 5779 5780 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and 5781 /// registers the function context. 5782 void ARMTargetLowering:: 5783 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB, 5784 MachineBasicBlock *DispatchBB, int FI) const { 5785 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5786 DebugLoc dl = MI->getDebugLoc(); 5787 MachineFunction *MF = MBB->getParent(); 5788 MachineRegisterInfo *MRI = &MF->getRegInfo(); 5789 MachineConstantPool *MCP = MF->getConstantPool(); 5790 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>(); 5791 const Function *F = MF->getFunction(); 5792 5793 bool isThumb = Subtarget->isThumb(); 5794 bool isThumb2 = Subtarget->isThumb2(); 5795 5796 unsigned PCLabelId = AFI->createPICLabelUId(); 5797 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8; 5798 ARMConstantPoolValue *CPV = 5799 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj); 5800 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4); 5801 5802 const TargetRegisterClass *TRC = isThumb ? 5803 (const TargetRegisterClass*)&ARM::tGPRRegClass : 5804 (const TargetRegisterClass*)&ARM::GPRRegClass; 5805 5806 // Grab constant pool and fixed stack memory operands. 5807 MachineMemOperand *CPMMO = 5808 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(), 5809 MachineMemOperand::MOLoad, 4, 4); 5810 5811 MachineMemOperand *FIMMOSt = 5812 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI), 5813 MachineMemOperand::MOStore, 4, 4); 5814 5815 // Load the address of the dispatch MBB into the jump buffer. 5816 if (isThumb2) { 5817 // Incoming value: jbuf 5818 // ldr.n r5, LCPI1_1 5819 // orr r5, r5, #1 5820 // add r5, pc 5821 // str r5, [$jbuf, #+4] ; &jbuf[1] 5822 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 5823 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1) 5824 .addConstantPoolIndex(CPI) 5825 .addMemOperand(CPMMO)); 5826 // Set the low bit because of thumb mode. 5827 unsigned NewVReg2 = MRI->createVirtualRegister(TRC); 5828 AddDefaultCC( 5829 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2) 5830 .addReg(NewVReg1, RegState::Kill) 5831 .addImm(0x01))); 5832 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 5833 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3) 5834 .addReg(NewVReg2, RegState::Kill) 5835 .addImm(PCLabelId); 5836 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12)) 5837 .addReg(NewVReg3, RegState::Kill) 5838 .addFrameIndex(FI) 5839 .addImm(36) // &jbuf[1] :: pc 5840 .addMemOperand(FIMMOSt)); 5841 } else if (isThumb) { 5842 // Incoming value: jbuf 5843 // ldr.n r1, LCPI1_4 5844 // add r1, pc 5845 // mov r2, #1 5846 // orrs r1, r2 5847 // add r2, $jbuf, #+4 ; &jbuf[1] 5848 // str r1, [r2] 5849 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 5850 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1) 5851 .addConstantPoolIndex(CPI) 5852 .addMemOperand(CPMMO)); 5853 unsigned NewVReg2 = MRI->createVirtualRegister(TRC); 5854 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2) 5855 .addReg(NewVReg1, RegState::Kill) 5856 .addImm(PCLabelId); 5857 // Set the low bit because of thumb mode. 5858 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 5859 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3) 5860 .addReg(ARM::CPSR, RegState::Define) 5861 .addImm(1)); 5862 unsigned NewVReg4 = MRI->createVirtualRegister(TRC); 5863 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4) 5864 .addReg(ARM::CPSR, RegState::Define) 5865 .addReg(NewVReg2, RegState::Kill) 5866 .addReg(NewVReg3, RegState::Kill)); 5867 unsigned NewVReg5 = MRI->createVirtualRegister(TRC); 5868 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5) 5869 .addFrameIndex(FI) 5870 .addImm(36)); // &jbuf[1] :: pc 5871 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi)) 5872 .addReg(NewVReg4, RegState::Kill) 5873 .addReg(NewVReg5, RegState::Kill) 5874 .addImm(0) 5875 .addMemOperand(FIMMOSt)); 5876 } else { 5877 // Incoming value: jbuf 5878 // ldr r1, LCPI1_1 5879 // add r1, pc, r1 5880 // str r1, [$jbuf, #+4] ; &jbuf[1] 5881 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 5882 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1) 5883 .addConstantPoolIndex(CPI) 5884 .addImm(0) 5885 .addMemOperand(CPMMO)); 5886 unsigned NewVReg2 = MRI->createVirtualRegister(TRC); 5887 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2) 5888 .addReg(NewVReg1, RegState::Kill) 5889 .addImm(PCLabelId)); 5890 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12)) 5891 .addReg(NewVReg2, RegState::Kill) 5892 .addFrameIndex(FI) 5893 .addImm(36) // &jbuf[1] :: pc 5894 .addMemOperand(FIMMOSt)); 5895 } 5896 } 5897 5898 MachineBasicBlock *ARMTargetLowering:: 5899 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const { 5900 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5901 DebugLoc dl = MI->getDebugLoc(); 5902 MachineFunction *MF = MBB->getParent(); 5903 MachineRegisterInfo *MRI = &MF->getRegInfo(); 5904 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>(); 5905 MachineFrameInfo *MFI = MF->getFrameInfo(); 5906 int FI = MFI->getFunctionContextIndex(); 5907 5908 const TargetRegisterClass *TRC = Subtarget->isThumb() ? 5909 (const TargetRegisterClass*)&ARM::tGPRRegClass : 5910 (const TargetRegisterClass*)&ARM::GPRnopcRegClass; 5911 5912 // Get a mapping of the call site numbers to all of the landing pads they're 5913 // associated with. 5914 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad; 5915 unsigned MaxCSNum = 0; 5916 MachineModuleInfo &MMI = MF->getMMI(); 5917 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; 5918 ++BB) { 5919 if (!BB->isLandingPad()) continue; 5920 5921 // FIXME: We should assert that the EH_LABEL is the first MI in the landing 5922 // pad. 5923 for (MachineBasicBlock::iterator 5924 II = BB->begin(), IE = BB->end(); II != IE; ++II) { 5925 if (!II->isEHLabel()) continue; 5926 5927 MCSymbol *Sym = II->getOperand(0).getMCSymbol(); 5928 if (!MMI.hasCallSiteLandingPad(Sym)) continue; 5929 5930 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym); 5931 for (SmallVectorImpl<unsigned>::iterator 5932 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end(); 5933 CSI != CSE; ++CSI) { 5934 CallSiteNumToLPad[*CSI].push_back(BB); 5935 MaxCSNum = std::max(MaxCSNum, *CSI); 5936 } 5937 break; 5938 } 5939 } 5940 5941 // Get an ordered list of the machine basic blocks for the jump table. 5942 std::vector<MachineBasicBlock*> LPadList; 5943 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs; 5944 LPadList.reserve(CallSiteNumToLPad.size()); 5945 for (unsigned I = 1; I <= MaxCSNum; ++I) { 5946 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I]; 5947 for (SmallVectorImpl<MachineBasicBlock*>::iterator 5948 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) { 5949 LPadList.push_back(*II); 5950 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end()); 5951 } 5952 } 5953 5954 assert(!LPadList.empty() && 5955 "No landing pad destinations for the dispatch jump table!"); 5956 5957 // Create the jump table and associated information. 5958 MachineJumpTableInfo *JTI = 5959 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline); 5960 unsigned MJTI = JTI->createJumpTableIndex(LPadList); 5961 unsigned UId = AFI->createJumpTableUId(); 5962 5963 // Create the MBBs for the dispatch code. 5964 5965 // Shove the dispatch's address into the return slot in the function context. 5966 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock(); 5967 DispatchBB->setIsLandingPad(); 5968 5969 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock(); 5970 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP)); 5971 DispatchBB->addSuccessor(TrapBB); 5972 5973 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock(); 5974 DispatchBB->addSuccessor(DispContBB); 5975 5976 // Insert and MBBs. 5977 MF->insert(MF->end(), DispatchBB); 5978 MF->insert(MF->end(), DispContBB); 5979 MF->insert(MF->end(), TrapBB); 5980 5981 // Insert code into the entry block that creates and registers the function 5982 // context. 5983 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI); 5984 5985 MachineMemOperand *FIMMOLd = 5986 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI), 5987 MachineMemOperand::MOLoad | 5988 MachineMemOperand::MOVolatile, 4, 4); 5989 5990 if (AFI->isThumb1OnlyFunction()) 5991 BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup)); 5992 else if (!Subtarget->hasVFP2()) 5993 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp)); 5994 else 5995 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup)); 5996 5997 unsigned NumLPads = LPadList.size(); 5998 if (Subtarget->isThumb2()) { 5999 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 6000 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1) 6001 .addFrameIndex(FI) 6002 .addImm(4) 6003 .addMemOperand(FIMMOLd)); 6004 6005 if (NumLPads < 256) { 6006 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri)) 6007 .addReg(NewVReg1) 6008 .addImm(LPadList.size())); 6009 } else { 6010 unsigned VReg1 = MRI->createVirtualRegister(TRC); 6011 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1) 6012 .addImm(NumLPads & 0xFFFF)); 6013 6014 unsigned VReg2 = VReg1; 6015 if ((NumLPads & 0xFFFF0000) != 0) { 6016 VReg2 = MRI->createVirtualRegister(TRC); 6017 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2) 6018 .addReg(VReg1) 6019 .addImm(NumLPads >> 16)); 6020 } 6021 6022 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr)) 6023 .addReg(NewVReg1) 6024 .addReg(VReg2)); 6025 } 6026 6027 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc)) 6028 .addMBB(TrapBB) 6029 .addImm(ARMCC::HI) 6030 .addReg(ARM::CPSR); 6031 6032 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 6033 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3) 6034 .addJumpTableIndex(MJTI) 6035 .addImm(UId)); 6036 6037 unsigned NewVReg4 = MRI->createVirtualRegister(TRC); 6038 AddDefaultCC( 6039 AddDefaultPred( 6040 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4) 6041 .addReg(NewVReg3, RegState::Kill) 6042 .addReg(NewVReg1) 6043 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2)))); 6044 6045 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT)) 6046 .addReg(NewVReg4, RegState::Kill) 6047 .addReg(NewVReg1) 6048 .addJumpTableIndex(MJTI) 6049 .addImm(UId); 6050 } else if (Subtarget->isThumb()) { 6051 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 6052 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1) 6053 .addFrameIndex(FI) 6054 .addImm(1) 6055 .addMemOperand(FIMMOLd)); 6056 6057 if (NumLPads < 256) { 6058 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8)) 6059 .addReg(NewVReg1) 6060 .addImm(NumLPads)); 6061 } else { 6062 MachineConstantPool *ConstantPool = MF->getConstantPool(); 6063 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext()); 6064 const Constant *C = ConstantInt::get(Int32Ty, NumLPads); 6065 6066 // MachineConstantPool wants an explicit alignment. 6067 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty); 6068 if (Align == 0) 6069 Align = getTargetData()->getTypeAllocSize(C->getType()); 6070 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align); 6071 6072 unsigned VReg1 = MRI->createVirtualRegister(TRC); 6073 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci)) 6074 .addReg(VReg1, RegState::Define) 6075 .addConstantPoolIndex(Idx)); 6076 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr)) 6077 .addReg(NewVReg1) 6078 .addReg(VReg1)); 6079 } 6080 6081 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc)) 6082 .addMBB(TrapBB) 6083 .addImm(ARMCC::HI) 6084 .addReg(ARM::CPSR); 6085 6086 unsigned NewVReg2 = MRI->createVirtualRegister(TRC); 6087 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2) 6088 .addReg(ARM::CPSR, RegState::Define) 6089 .addReg(NewVReg1) 6090 .addImm(2)); 6091 6092 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 6093 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3) 6094 .addJumpTableIndex(MJTI) 6095 .addImm(UId)); 6096 6097 unsigned NewVReg4 = MRI->createVirtualRegister(TRC); 6098 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4) 6099 .addReg(ARM::CPSR, RegState::Define) 6100 .addReg(NewVReg2, RegState::Kill) 6101 .addReg(NewVReg3)); 6102 6103 MachineMemOperand *JTMMOLd = 6104 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(), 6105 MachineMemOperand::MOLoad, 4, 4); 6106 6107 unsigned NewVReg5 = MRI->createVirtualRegister(TRC); 6108 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5) 6109 .addReg(NewVReg4, RegState::Kill) 6110 .addImm(0) 6111 .addMemOperand(JTMMOLd)); 6112 6113 unsigned NewVReg6 = MRI->createVirtualRegister(TRC); 6114 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6) 6115 .addReg(ARM::CPSR, RegState::Define) 6116 .addReg(NewVReg5, RegState::Kill) 6117 .addReg(NewVReg3)); 6118 6119 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr)) 6120 .addReg(NewVReg6, RegState::Kill) 6121 .addJumpTableIndex(MJTI) 6122 .addImm(UId); 6123 } else { 6124 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 6125 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1) 6126 .addFrameIndex(FI) 6127 .addImm(4) 6128 .addMemOperand(FIMMOLd)); 6129 6130 if (NumLPads < 256) { 6131 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri)) 6132 .addReg(NewVReg1) 6133 .addImm(NumLPads)); 6134 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) { 6135 unsigned VReg1 = MRI->createVirtualRegister(TRC); 6136 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1) 6137 .addImm(NumLPads & 0xFFFF)); 6138 6139 unsigned VReg2 = VReg1; 6140 if ((NumLPads & 0xFFFF0000) != 0) { 6141 VReg2 = MRI->createVirtualRegister(TRC); 6142 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2) 6143 .addReg(VReg1) 6144 .addImm(NumLPads >> 16)); 6145 } 6146 6147 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr)) 6148 .addReg(NewVReg1) 6149 .addReg(VReg2)); 6150 } else { 6151 MachineConstantPool *ConstantPool = MF->getConstantPool(); 6152 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext()); 6153 const Constant *C = ConstantInt::get(Int32Ty, NumLPads); 6154 6155 // MachineConstantPool wants an explicit alignment. 6156 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty); 6157 if (Align == 0) 6158 Align = getTargetData()->getTypeAllocSize(C->getType()); 6159 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align); 6160 6161 unsigned VReg1 = MRI->createVirtualRegister(TRC); 6162 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp)) 6163 .addReg(VReg1, RegState::Define) 6164 .addConstantPoolIndex(Idx) 6165 .addImm(0)); 6166 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr)) 6167 .addReg(NewVReg1) 6168 .addReg(VReg1, RegState::Kill)); 6169 } 6170 6171 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc)) 6172 .addMBB(TrapBB) 6173 .addImm(ARMCC::HI) 6174 .addReg(ARM::CPSR); 6175 6176 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 6177 AddDefaultCC( 6178 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3) 6179 .addReg(NewVReg1) 6180 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2)))); 6181 unsigned NewVReg4 = MRI->createVirtualRegister(TRC); 6182 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4) 6183 .addJumpTableIndex(MJTI) 6184 .addImm(UId)); 6185 6186 MachineMemOperand *JTMMOLd = 6187 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(), 6188 MachineMemOperand::MOLoad, 4, 4); 6189 unsigned NewVReg5 = MRI->createVirtualRegister(TRC); 6190 AddDefaultPred( 6191 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5) 6192 .addReg(NewVReg3, RegState::Kill) 6193 .addReg(NewVReg4) 6194 .addImm(0) 6195 .addMemOperand(JTMMOLd)); 6196 6197 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd)) 6198 .addReg(NewVReg5, RegState::Kill) 6199 .addReg(NewVReg4) 6200 .addJumpTableIndex(MJTI) 6201 .addImm(UId); 6202 } 6203 6204 // Add the jump table entries as successors to the MBB. 6205 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs; 6206 for (std::vector<MachineBasicBlock*>::iterator 6207 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) { 6208 MachineBasicBlock *CurMBB = *I; 6209 if (SeenMBBs.insert(CurMBB)) 6210 DispContBB->addSuccessor(CurMBB); 6211 } 6212 6213 // N.B. the order the invoke BBs are processed in doesn't matter here. 6214 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII); 6215 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo(); 6216 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF); 6217 SmallVector<MachineBasicBlock*, 64> MBBLPads; 6218 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator 6219 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) { 6220 MachineBasicBlock *BB = *I; 6221 6222 // Remove the landing pad successor from the invoke block and replace it 6223 // with the new dispatch block. 6224 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(), 6225 BB->succ_end()); 6226 while (!Successors.empty()) { 6227 MachineBasicBlock *SMBB = Successors.pop_back_val(); 6228 if (SMBB->isLandingPad()) { 6229 BB->removeSuccessor(SMBB); 6230 MBBLPads.push_back(SMBB); 6231 } 6232 } 6233 6234 BB->addSuccessor(DispatchBB); 6235 6236 // Find the invoke call and mark all of the callee-saved registers as 6237 // 'implicit defined' so that they're spilled. This prevents code from 6238 // moving instructions to before the EH block, where they will never be 6239 // executed. 6240 for (MachineBasicBlock::reverse_iterator 6241 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) { 6242 if (!II->isCall()) continue; 6243 6244 DenseMap<unsigned, bool> DefRegs; 6245 for (MachineInstr::mop_iterator 6246 OI = II->operands_begin(), OE = II->operands_end(); 6247 OI != OE; ++OI) { 6248 if (!OI->isReg()) continue; 6249 DefRegs[OI->getReg()] = true; 6250 } 6251 6252 MachineInstrBuilder MIB(&*II); 6253 6254 for (unsigned i = 0; SavedRegs[i] != 0; ++i) { 6255 unsigned Reg = SavedRegs[i]; 6256 if (Subtarget->isThumb2() && 6257 !ARM::tGPRRegClass.contains(Reg) && 6258 !ARM::hGPRRegClass.contains(Reg)) 6259 continue; 6260 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg)) 6261 continue; 6262 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg)) 6263 continue; 6264 if (!DefRegs[Reg]) 6265 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead); 6266 } 6267 6268 break; 6269 } 6270 } 6271 6272 // Mark all former landing pads as non-landing pads. The dispatch is the only 6273 // landing pad now. 6274 for (SmallVectorImpl<MachineBasicBlock*>::iterator 6275 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I) 6276 (*I)->setIsLandingPad(false); 6277 6278 // The instruction is gone now. 6279 MI->eraseFromParent(); 6280 6281 return MBB; 6282 } 6283 6284 static 6285 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) { 6286 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(), 6287 E = MBB->succ_end(); I != E; ++I) 6288 if (*I != Succ) 6289 return *I; 6290 llvm_unreachable("Expecting a BB with two successors!"); 6291 } 6292 6293 MachineBasicBlock *ARMTargetLowering:: 6294 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const { 6295 // This pseudo instruction has 3 operands: dst, src, size 6296 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold(). 6297 // Otherwise, we will generate unrolled scalar copies. 6298 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 6299 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 6300 MachineFunction::iterator It = BB; 6301 ++It; 6302 6303 unsigned dest = MI->getOperand(0).getReg(); 6304 unsigned src = MI->getOperand(1).getReg(); 6305 unsigned SizeVal = MI->getOperand(2).getImm(); 6306 unsigned Align = MI->getOperand(3).getImm(); 6307 DebugLoc dl = MI->getDebugLoc(); 6308 6309 bool isThumb2 = Subtarget->isThumb2(); 6310 MachineFunction *MF = BB->getParent(); 6311 MachineRegisterInfo &MRI = MF->getRegInfo(); 6312 unsigned ldrOpc, strOpc, UnitSize = 0; 6313 6314 const TargetRegisterClass *TRC = isThumb2 ? 6315 (const TargetRegisterClass*)&ARM::tGPRRegClass : 6316 (const TargetRegisterClass*)&ARM::GPRRegClass; 6317 const TargetRegisterClass *TRC_Vec = 0; 6318 6319 if (Align & 1) { 6320 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM; 6321 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM; 6322 UnitSize = 1; 6323 } else if (Align & 2) { 6324 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST; 6325 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST; 6326 UnitSize = 2; 6327 } else { 6328 // Check whether we can use NEON instructions. 6329 if (!MF->getFunction()->hasFnAttr(Attribute::NoImplicitFloat) && 6330 Subtarget->hasNEON()) { 6331 if ((Align % 16 == 0) && SizeVal >= 16) { 6332 ldrOpc = ARM::VLD1q32wb_fixed; 6333 strOpc = ARM::VST1q32wb_fixed; 6334 UnitSize = 16; 6335 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass; 6336 } 6337 else if ((Align % 8 == 0) && SizeVal >= 8) { 6338 ldrOpc = ARM::VLD1d32wb_fixed; 6339 strOpc = ARM::VST1d32wb_fixed; 6340 UnitSize = 8; 6341 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass; 6342 } 6343 } 6344 // Can't use NEON instructions. 6345 if (UnitSize == 0) { 6346 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM; 6347 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM; 6348 UnitSize = 4; 6349 } 6350 } 6351 6352 unsigned BytesLeft = SizeVal % UnitSize; 6353 unsigned LoopSize = SizeVal - BytesLeft; 6354 6355 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) { 6356 // Use LDR and STR to copy. 6357 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize) 6358 // [destOut] = STR_POST(scratch, destIn, UnitSize) 6359 unsigned srcIn = src; 6360 unsigned destIn = dest; 6361 for (unsigned i = 0; i < LoopSize; i+=UnitSize) { 6362 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC); 6363 unsigned srcOut = MRI.createVirtualRegister(TRC); 6364 unsigned destOut = MRI.createVirtualRegister(TRC); 6365 if (UnitSize >= 8) { 6366 AddDefaultPred(BuildMI(*BB, MI, dl, 6367 TII->get(ldrOpc), scratch) 6368 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0)); 6369 6370 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut) 6371 .addReg(destIn).addImm(0).addReg(scratch)); 6372 } else if (isThumb2) { 6373 AddDefaultPred(BuildMI(*BB, MI, dl, 6374 TII->get(ldrOpc), scratch) 6375 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize)); 6376 6377 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut) 6378 .addReg(scratch).addReg(destIn) 6379 .addImm(UnitSize)); 6380 } else { 6381 AddDefaultPred(BuildMI(*BB, MI, dl, 6382 TII->get(ldrOpc), scratch) 6383 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0) 6384 .addImm(UnitSize)); 6385 6386 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut) 6387 .addReg(scratch).addReg(destIn) 6388 .addReg(0).addImm(UnitSize)); 6389 } 6390 srcIn = srcOut; 6391 destIn = destOut; 6392 } 6393 6394 // Handle the leftover bytes with LDRB and STRB. 6395 // [scratch, srcOut] = LDRB_POST(srcIn, 1) 6396 // [destOut] = STRB_POST(scratch, destIn, 1) 6397 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM; 6398 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM; 6399 for (unsigned i = 0; i < BytesLeft; i++) { 6400 unsigned scratch = MRI.createVirtualRegister(TRC); 6401 unsigned srcOut = MRI.createVirtualRegister(TRC); 6402 unsigned destOut = MRI.createVirtualRegister(TRC); 6403 if (isThumb2) { 6404 AddDefaultPred(BuildMI(*BB, MI, dl, 6405 TII->get(ldrOpc),scratch) 6406 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1)); 6407 6408 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut) 6409 .addReg(scratch).addReg(destIn) 6410 .addReg(0).addImm(1)); 6411 } else { 6412 AddDefaultPred(BuildMI(*BB, MI, dl, 6413 TII->get(ldrOpc),scratch) 6414 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1)); 6415 6416 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut) 6417 .addReg(scratch).addReg(destIn) 6418 .addReg(0).addImm(1)); 6419 } 6420 srcIn = srcOut; 6421 destIn = destOut; 6422 } 6423 MI->eraseFromParent(); // The instruction is gone now. 6424 return BB; 6425 } 6426 6427 // Expand the pseudo op to a loop. 6428 // thisMBB: 6429 // ... 6430 // movw varEnd, # --> with thumb2 6431 // movt varEnd, # 6432 // ldrcp varEnd, idx --> without thumb2 6433 // fallthrough --> loopMBB 6434 // loopMBB: 6435 // PHI varPhi, varEnd, varLoop 6436 // PHI srcPhi, src, srcLoop 6437 // PHI destPhi, dst, destLoop 6438 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize) 6439 // [destLoop] = STR_POST(scratch, destPhi, UnitSize) 6440 // subs varLoop, varPhi, #UnitSize 6441 // bne loopMBB 6442 // fallthrough --> exitMBB 6443 // exitMBB: 6444 // epilogue to handle left-over bytes 6445 // [scratch, srcOut] = LDRB_POST(srcLoop, 1) 6446 // [destOut] = STRB_POST(scratch, destLoop, 1) 6447 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 6448 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 6449 MF->insert(It, loopMBB); 6450 MF->insert(It, exitMBB); 6451 6452 // Transfer the remainder of BB and its successor edges to exitMBB. 6453 exitMBB->splice(exitMBB->begin(), BB, 6454 llvm::next(MachineBasicBlock::iterator(MI)), 6455 BB->end()); 6456 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 6457 6458 // Load an immediate to varEnd. 6459 unsigned varEnd = MRI.createVirtualRegister(TRC); 6460 if (isThumb2) { 6461 unsigned VReg1 = varEnd; 6462 if ((LoopSize & 0xFFFF0000) != 0) 6463 VReg1 = MRI.createVirtualRegister(TRC); 6464 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1) 6465 .addImm(LoopSize & 0xFFFF)); 6466 6467 if ((LoopSize & 0xFFFF0000) != 0) 6468 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd) 6469 .addReg(VReg1) 6470 .addImm(LoopSize >> 16)); 6471 } else { 6472 MachineConstantPool *ConstantPool = MF->getConstantPool(); 6473 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext()); 6474 const Constant *C = ConstantInt::get(Int32Ty, LoopSize); 6475 6476 // MachineConstantPool wants an explicit alignment. 6477 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty); 6478 if (Align == 0) 6479 Align = getTargetData()->getTypeAllocSize(C->getType()); 6480 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align); 6481 6482 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp)) 6483 .addReg(varEnd, RegState::Define) 6484 .addConstantPoolIndex(Idx) 6485 .addImm(0)); 6486 } 6487 BB->addSuccessor(loopMBB); 6488 6489 // Generate the loop body: 6490 // varPhi = PHI(varLoop, varEnd) 6491 // srcPhi = PHI(srcLoop, src) 6492 // destPhi = PHI(destLoop, dst) 6493 MachineBasicBlock *entryBB = BB; 6494 BB = loopMBB; 6495 unsigned varLoop = MRI.createVirtualRegister(TRC); 6496 unsigned varPhi = MRI.createVirtualRegister(TRC); 6497 unsigned srcLoop = MRI.createVirtualRegister(TRC); 6498 unsigned srcPhi = MRI.createVirtualRegister(TRC); 6499 unsigned destLoop = MRI.createVirtualRegister(TRC); 6500 unsigned destPhi = MRI.createVirtualRegister(TRC); 6501 6502 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi) 6503 .addReg(varLoop).addMBB(loopMBB) 6504 .addReg(varEnd).addMBB(entryBB); 6505 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi) 6506 .addReg(srcLoop).addMBB(loopMBB) 6507 .addReg(src).addMBB(entryBB); 6508 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi) 6509 .addReg(destLoop).addMBB(loopMBB) 6510 .addReg(dest).addMBB(entryBB); 6511 6512 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize) 6513 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz) 6514 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC); 6515 if (UnitSize >= 8) { 6516 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch) 6517 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0)); 6518 6519 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop) 6520 .addReg(destPhi).addImm(0).addReg(scratch)); 6521 } else if (isThumb2) { 6522 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch) 6523 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize)); 6524 6525 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop) 6526 .addReg(scratch).addReg(destPhi) 6527 .addImm(UnitSize)); 6528 } else { 6529 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch) 6530 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0) 6531 .addImm(UnitSize)); 6532 6533 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop) 6534 .addReg(scratch).addReg(destPhi) 6535 .addReg(0).addImm(UnitSize)); 6536 } 6537 6538 // Decrement loop variable by UnitSize. 6539 MachineInstrBuilder MIB = BuildMI(BB, dl, 6540 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop); 6541 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize))); 6542 MIB->getOperand(5).setReg(ARM::CPSR); 6543 MIB->getOperand(5).setIsDef(true); 6544 6545 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 6546 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 6547 6548 // loopMBB can loop back to loopMBB or fall through to exitMBB. 6549 BB->addSuccessor(loopMBB); 6550 BB->addSuccessor(exitMBB); 6551 6552 // Add epilogue to handle BytesLeft. 6553 BB = exitMBB; 6554 MachineInstr *StartOfExit = exitMBB->begin(); 6555 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM; 6556 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM; 6557 6558 // [scratch, srcOut] = LDRB_POST(srcLoop, 1) 6559 // [destOut] = STRB_POST(scratch, destLoop, 1) 6560 unsigned srcIn = srcLoop; 6561 unsigned destIn = destLoop; 6562 for (unsigned i = 0; i < BytesLeft; i++) { 6563 unsigned scratch = MRI.createVirtualRegister(TRC); 6564 unsigned srcOut = MRI.createVirtualRegister(TRC); 6565 unsigned destOut = MRI.createVirtualRegister(TRC); 6566 if (isThumb2) { 6567 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, 6568 TII->get(ldrOpc),scratch) 6569 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1)); 6570 6571 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut) 6572 .addReg(scratch).addReg(destIn) 6573 .addImm(1)); 6574 } else { 6575 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, 6576 TII->get(ldrOpc),scratch) 6577 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1)); 6578 6579 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut) 6580 .addReg(scratch).addReg(destIn) 6581 .addReg(0).addImm(1)); 6582 } 6583 srcIn = srcOut; 6584 destIn = destOut; 6585 } 6586 6587 MI->eraseFromParent(); // The instruction is gone now. 6588 return BB; 6589 } 6590 6591 MachineBasicBlock * 6592 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 6593 MachineBasicBlock *BB) const { 6594 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 6595 DebugLoc dl = MI->getDebugLoc(); 6596 bool isThumb2 = Subtarget->isThumb2(); 6597 switch (MI->getOpcode()) { 6598 default: { 6599 MI->dump(); 6600 llvm_unreachable("Unexpected instr type to insert"); 6601 } 6602 // The Thumb2 pre-indexed stores have the same MI operands, they just 6603 // define them differently in the .td files from the isel patterns, so 6604 // they need pseudos. 6605 case ARM::t2STR_preidx: 6606 MI->setDesc(TII->get(ARM::t2STR_PRE)); 6607 return BB; 6608 case ARM::t2STRB_preidx: 6609 MI->setDesc(TII->get(ARM::t2STRB_PRE)); 6610 return BB; 6611 case ARM::t2STRH_preidx: 6612 MI->setDesc(TII->get(ARM::t2STRH_PRE)); 6613 return BB; 6614 6615 case ARM::STRi_preidx: 6616 case ARM::STRBi_preidx: { 6617 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ? 6618 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM; 6619 // Decode the offset. 6620 unsigned Offset = MI->getOperand(4).getImm(); 6621 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub; 6622 Offset = ARM_AM::getAM2Offset(Offset); 6623 if (isSub) 6624 Offset = -Offset; 6625 6626 MachineMemOperand *MMO = *MI->memoperands_begin(); 6627 BuildMI(*BB, MI, dl, TII->get(NewOpc)) 6628 .addOperand(MI->getOperand(0)) // Rn_wb 6629 .addOperand(MI->getOperand(1)) // Rt 6630 .addOperand(MI->getOperand(2)) // Rn 6631 .addImm(Offset) // offset (skip GPR==zero_reg) 6632 .addOperand(MI->getOperand(5)) // pred 6633 .addOperand(MI->getOperand(6)) 6634 .addMemOperand(MMO); 6635 MI->eraseFromParent(); 6636 return BB; 6637 } 6638 case ARM::STRr_preidx: 6639 case ARM::STRBr_preidx: 6640 case ARM::STRH_preidx: { 6641 unsigned NewOpc; 6642 switch (MI->getOpcode()) { 6643 default: llvm_unreachable("unexpected opcode!"); 6644 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break; 6645 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break; 6646 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break; 6647 } 6648 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc)); 6649 for (unsigned i = 0; i < MI->getNumOperands(); ++i) 6650 MIB.addOperand(MI->getOperand(i)); 6651 MI->eraseFromParent(); 6652 return BB; 6653 } 6654 case ARM::ATOMIC_LOAD_ADD_I8: 6655 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr); 6656 case ARM::ATOMIC_LOAD_ADD_I16: 6657 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr); 6658 case ARM::ATOMIC_LOAD_ADD_I32: 6659 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr); 6660 6661 case ARM::ATOMIC_LOAD_AND_I8: 6662 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr); 6663 case ARM::ATOMIC_LOAD_AND_I16: 6664 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr); 6665 case ARM::ATOMIC_LOAD_AND_I32: 6666 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr); 6667 6668 case ARM::ATOMIC_LOAD_OR_I8: 6669 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr); 6670 case ARM::ATOMIC_LOAD_OR_I16: 6671 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr); 6672 case ARM::ATOMIC_LOAD_OR_I32: 6673 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr); 6674 6675 case ARM::ATOMIC_LOAD_XOR_I8: 6676 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr); 6677 case ARM::ATOMIC_LOAD_XOR_I16: 6678 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr); 6679 case ARM::ATOMIC_LOAD_XOR_I32: 6680 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr); 6681 6682 case ARM::ATOMIC_LOAD_NAND_I8: 6683 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr); 6684 case ARM::ATOMIC_LOAD_NAND_I16: 6685 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr); 6686 case ARM::ATOMIC_LOAD_NAND_I32: 6687 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr); 6688 6689 case ARM::ATOMIC_LOAD_SUB_I8: 6690 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr); 6691 case ARM::ATOMIC_LOAD_SUB_I16: 6692 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr); 6693 case ARM::ATOMIC_LOAD_SUB_I32: 6694 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr); 6695 6696 case ARM::ATOMIC_LOAD_MIN_I8: 6697 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT); 6698 case ARM::ATOMIC_LOAD_MIN_I16: 6699 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT); 6700 case ARM::ATOMIC_LOAD_MIN_I32: 6701 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT); 6702 6703 case ARM::ATOMIC_LOAD_MAX_I8: 6704 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT); 6705 case ARM::ATOMIC_LOAD_MAX_I16: 6706 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT); 6707 case ARM::ATOMIC_LOAD_MAX_I32: 6708 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT); 6709 6710 case ARM::ATOMIC_LOAD_UMIN_I8: 6711 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO); 6712 case ARM::ATOMIC_LOAD_UMIN_I16: 6713 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO); 6714 case ARM::ATOMIC_LOAD_UMIN_I32: 6715 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO); 6716 6717 case ARM::ATOMIC_LOAD_UMAX_I8: 6718 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI); 6719 case ARM::ATOMIC_LOAD_UMAX_I16: 6720 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI); 6721 case ARM::ATOMIC_LOAD_UMAX_I32: 6722 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI); 6723 6724 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0); 6725 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0); 6726 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0); 6727 6728 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1); 6729 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2); 6730 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4); 6731 6732 6733 case ARM::ATOMADD6432: 6734 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr, 6735 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr, 6736 /*NeedsCarry*/ true); 6737 case ARM::ATOMSUB6432: 6738 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr, 6739 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr, 6740 /*NeedsCarry*/ true); 6741 case ARM::ATOMOR6432: 6742 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr, 6743 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr); 6744 case ARM::ATOMXOR6432: 6745 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr, 6746 isThumb2 ? ARM::t2EORrr : ARM::EORrr); 6747 case ARM::ATOMAND6432: 6748 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr, 6749 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr); 6750 case ARM::ATOMSWAP6432: 6751 return EmitAtomicBinary64(MI, BB, 0, 0, false); 6752 case ARM::ATOMCMPXCHG6432: 6753 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr, 6754 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr, 6755 /*NeedsCarry*/ false, /*IsCmpxchg*/true); 6756 6757 case ARM::tMOVCCr_pseudo: { 6758 // To "insert" a SELECT_CC instruction, we actually have to insert the 6759 // diamond control-flow pattern. The incoming instruction knows the 6760 // destination vreg to set, the condition code register to branch on, the 6761 // true/false values to select between, and a branch opcode to use. 6762 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 6763 MachineFunction::iterator It = BB; 6764 ++It; 6765 6766 // thisMBB: 6767 // ... 6768 // TrueVal = ... 6769 // cmpTY ccX, r1, r2 6770 // bCC copy1MBB 6771 // fallthrough --> copy0MBB 6772 MachineBasicBlock *thisMBB = BB; 6773 MachineFunction *F = BB->getParent(); 6774 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 6775 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); 6776 F->insert(It, copy0MBB); 6777 F->insert(It, sinkMBB); 6778 6779 // Transfer the remainder of BB and its successor edges to sinkMBB. 6780 sinkMBB->splice(sinkMBB->begin(), BB, 6781 llvm::next(MachineBasicBlock::iterator(MI)), 6782 BB->end()); 6783 sinkMBB->transferSuccessorsAndUpdatePHIs(BB); 6784 6785 BB->addSuccessor(copy0MBB); 6786 BB->addSuccessor(sinkMBB); 6787 6788 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB) 6789 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg()); 6790 6791 // copy0MBB: 6792 // %FalseValue = ... 6793 // # fallthrough to sinkMBB 6794 BB = copy0MBB; 6795 6796 // Update machine-CFG edges 6797 BB->addSuccessor(sinkMBB); 6798 6799 // sinkMBB: 6800 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] 6801 // ... 6802 BB = sinkMBB; 6803 BuildMI(*BB, BB->begin(), dl, 6804 TII->get(ARM::PHI), MI->getOperand(0).getReg()) 6805 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB) 6806 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB); 6807 6808 MI->eraseFromParent(); // The pseudo instruction is gone now. 6809 return BB; 6810 } 6811 6812 case ARM::BCCi64: 6813 case ARM::BCCZi64: { 6814 // If there is an unconditional branch to the other successor, remove it. 6815 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end()); 6816 6817 // Compare both parts that make up the double comparison separately for 6818 // equality. 6819 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64; 6820 6821 unsigned LHS1 = MI->getOperand(1).getReg(); 6822 unsigned LHS2 = MI->getOperand(2).getReg(); 6823 if (RHSisZero) { 6824 AddDefaultPred(BuildMI(BB, dl, 6825 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 6826 .addReg(LHS1).addImm(0)); 6827 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 6828 .addReg(LHS2).addImm(0) 6829 .addImm(ARMCC::EQ).addReg(ARM::CPSR); 6830 } else { 6831 unsigned RHS1 = MI->getOperand(3).getReg(); 6832 unsigned RHS2 = MI->getOperand(4).getReg(); 6833 AddDefaultPred(BuildMI(BB, dl, 6834 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) 6835 .addReg(LHS1).addReg(RHS1)); 6836 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) 6837 .addReg(LHS2).addReg(RHS2) 6838 .addImm(ARMCC::EQ).addReg(ARM::CPSR); 6839 } 6840 6841 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB(); 6842 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB); 6843 if (MI->getOperand(0).getImm() == ARMCC::NE) 6844 std::swap(destMBB, exitMBB); 6845 6846 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 6847 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR); 6848 if (isThumb2) 6849 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB)); 6850 else 6851 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB); 6852 6853 MI->eraseFromParent(); // The pseudo instruction is gone now. 6854 return BB; 6855 } 6856 6857 case ARM::Int_eh_sjlj_setjmp: 6858 case ARM::Int_eh_sjlj_setjmp_nofp: 6859 case ARM::tInt_eh_sjlj_setjmp: 6860 case ARM::t2Int_eh_sjlj_setjmp: 6861 case ARM::t2Int_eh_sjlj_setjmp_nofp: 6862 EmitSjLjDispatchBlock(MI, BB); 6863 return BB; 6864 6865 case ARM::ABS: 6866 case ARM::t2ABS: { 6867 // To insert an ABS instruction, we have to insert the 6868 // diamond control-flow pattern. The incoming instruction knows the 6869 // source vreg to test against 0, the destination vreg to set, 6870 // the condition code register to branch on, the 6871 // true/false values to select between, and a branch opcode to use. 6872 // It transforms 6873 // V1 = ABS V0 6874 // into 6875 // V2 = MOVS V0 6876 // BCC (branch to SinkBB if V0 >= 0) 6877 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0) 6878 // SinkBB: V1 = PHI(V2, V3) 6879 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 6880 MachineFunction::iterator BBI = BB; 6881 ++BBI; 6882 MachineFunction *Fn = BB->getParent(); 6883 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB); 6884 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB); 6885 Fn->insert(BBI, RSBBB); 6886 Fn->insert(BBI, SinkBB); 6887 6888 unsigned int ABSSrcReg = MI->getOperand(1).getReg(); 6889 unsigned int ABSDstReg = MI->getOperand(0).getReg(); 6890 bool isThumb2 = Subtarget->isThumb2(); 6891 MachineRegisterInfo &MRI = Fn->getRegInfo(); 6892 // In Thumb mode S must not be specified if source register is the SP or 6893 // PC and if destination register is the SP, so restrict register class 6894 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ? 6895 (const TargetRegisterClass*)&ARM::rGPRRegClass : 6896 (const TargetRegisterClass*)&ARM::GPRRegClass); 6897 6898 // Transfer the remainder of BB and its successor edges to sinkMBB. 6899 SinkBB->splice(SinkBB->begin(), BB, 6900 llvm::next(MachineBasicBlock::iterator(MI)), 6901 BB->end()); 6902 SinkBB->transferSuccessorsAndUpdatePHIs(BB); 6903 6904 BB->addSuccessor(RSBBB); 6905 BB->addSuccessor(SinkBB); 6906 6907 // fall through to SinkMBB 6908 RSBBB->addSuccessor(SinkBB); 6909 6910 // insert a cmp at the end of BB 6911 AddDefaultPred(BuildMI(BB, dl, 6912 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 6913 .addReg(ABSSrcReg).addImm(0)); 6914 6915 // insert a bcc with opposite CC to ARMCC::MI at the end of BB 6916 BuildMI(BB, dl, 6917 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB) 6918 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR); 6919 6920 // insert rsbri in RSBBB 6921 // Note: BCC and rsbri will be converted into predicated rsbmi 6922 // by if-conversion pass 6923 BuildMI(*RSBBB, RSBBB->begin(), dl, 6924 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg) 6925 .addReg(ABSSrcReg, RegState::Kill) 6926 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0); 6927 6928 // insert PHI in SinkBB, 6929 // reuse ABSDstReg to not change uses of ABS instruction 6930 BuildMI(*SinkBB, SinkBB->begin(), dl, 6931 TII->get(ARM::PHI), ABSDstReg) 6932 .addReg(NewRsbDstReg).addMBB(RSBBB) 6933 .addReg(ABSSrcReg).addMBB(BB); 6934 6935 // remove ABS instruction 6936 MI->eraseFromParent(); 6937 6938 // return last added BB 6939 return SinkBB; 6940 } 6941 case ARM::COPY_STRUCT_BYVAL_I32: 6942 ++NumLoopByVals; 6943 return EmitStructByval(MI, BB); 6944 } 6945 } 6946 6947 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI, 6948 SDNode *Node) const { 6949 if (!MI->hasPostISelHook()) { 6950 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) && 6951 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'"); 6952 return; 6953 } 6954 6955 const MCInstrDesc *MCID = &MI->getDesc(); 6956 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB, 6957 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional 6958 // operand is still set to noreg. If needed, set the optional operand's 6959 // register to CPSR, and remove the redundant implicit def. 6960 // 6961 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>). 6962 6963 // Rename pseudo opcodes. 6964 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode()); 6965 if (NewOpc) { 6966 const ARMBaseInstrInfo *TII = 6967 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo()); 6968 MCID = &TII->get(NewOpc); 6969 6970 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 && 6971 "converted opcode should be the same except for cc_out"); 6972 6973 MI->setDesc(*MCID); 6974 6975 // Add the optional cc_out operand 6976 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true)); 6977 } 6978 unsigned ccOutIdx = MCID->getNumOperands() - 1; 6979 6980 // Any ARM instruction that sets the 's' bit should specify an optional 6981 // "cc_out" operand in the last operand position. 6982 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) { 6983 assert(!NewOpc && "Optional cc_out operand required"); 6984 return; 6985 } 6986 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it 6987 // since we already have an optional CPSR def. 6988 bool definesCPSR = false; 6989 bool deadCPSR = false; 6990 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands(); 6991 i != e; ++i) { 6992 const MachineOperand &MO = MI->getOperand(i); 6993 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) { 6994 definesCPSR = true; 6995 if (MO.isDead()) 6996 deadCPSR = true; 6997 MI->RemoveOperand(i); 6998 break; 6999 } 7000 } 7001 if (!definesCPSR) { 7002 assert(!NewOpc && "Optional cc_out operand required"); 7003 return; 7004 } 7005 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag"); 7006 if (deadCPSR) { 7007 assert(!MI->getOperand(ccOutIdx).getReg() && 7008 "expect uninitialized optional cc_out operand"); 7009 return; 7010 } 7011 7012 // If this instruction was defined with an optional CPSR def and its dag node 7013 // had a live implicit CPSR def, then activate the optional CPSR def. 7014 MachineOperand &MO = MI->getOperand(ccOutIdx); 7015 MO.setReg(ARM::CPSR); 7016 MO.setIsDef(true); 7017 } 7018 7019 //===----------------------------------------------------------------------===// 7020 // ARM Optimization Hooks 7021 //===----------------------------------------------------------------------===// 7022 7023 // Helper function that checks if N is a null or all ones constant. 7024 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) { 7025 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N); 7026 if (!C) 7027 return false; 7028 return AllOnes ? C->isAllOnesValue() : C->isNullValue(); 7029 } 7030 7031 // Return true if N is conditionally 0 or all ones. 7032 // Detects these expressions where cc is an i1 value: 7033 // 7034 // (select cc 0, y) [AllOnes=0] 7035 // (select cc y, 0) [AllOnes=0] 7036 // (zext cc) [AllOnes=0] 7037 // (sext cc) [AllOnes=0/1] 7038 // (select cc -1, y) [AllOnes=1] 7039 // (select cc y, -1) [AllOnes=1] 7040 // 7041 // Invert is set when N is the null/all ones constant when CC is false. 7042 // OtherOp is set to the alternative value of N. 7043 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes, 7044 SDValue &CC, bool &Invert, 7045 SDValue &OtherOp, 7046 SelectionDAG &DAG) { 7047 switch (N->getOpcode()) { 7048 default: return false; 7049 case ISD::SELECT: { 7050 CC = N->getOperand(0); 7051 SDValue N1 = N->getOperand(1); 7052 SDValue N2 = N->getOperand(2); 7053 if (isZeroOrAllOnes(N1, AllOnes)) { 7054 Invert = false; 7055 OtherOp = N2; 7056 return true; 7057 } 7058 if (isZeroOrAllOnes(N2, AllOnes)) { 7059 Invert = true; 7060 OtherOp = N1; 7061 return true; 7062 } 7063 return false; 7064 } 7065 case ISD::ZERO_EXTEND: 7066 // (zext cc) can never be the all ones value. 7067 if (AllOnes) 7068 return false; 7069 // Fall through. 7070 case ISD::SIGN_EXTEND: { 7071 EVT VT = N->getValueType(0); 7072 CC = N->getOperand(0); 7073 if (CC.getValueType() != MVT::i1) 7074 return false; 7075 Invert = !AllOnes; 7076 if (AllOnes) 7077 // When looking for an AllOnes constant, N is an sext, and the 'other' 7078 // value is 0. 7079 OtherOp = DAG.getConstant(0, VT); 7080 else if (N->getOpcode() == ISD::ZERO_EXTEND) 7081 // When looking for a 0 constant, N can be zext or sext. 7082 OtherOp = DAG.getConstant(1, VT); 7083 else 7084 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT); 7085 return true; 7086 } 7087 } 7088 } 7089 7090 // Combine a constant select operand into its use: 7091 // 7092 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c)) 7093 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c)) 7094 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1] 7095 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c)) 7096 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c)) 7097 // 7098 // The transform is rejected if the select doesn't have a constant operand that 7099 // is null, or all ones when AllOnes is set. 7100 // 7101 // Also recognize sext/zext from i1: 7102 // 7103 // (add (zext cc), x) -> (select cc (add x, 1), x) 7104 // (add (sext cc), x) -> (select cc (add x, -1), x) 7105 // 7106 // These transformations eventually create predicated instructions. 7107 // 7108 // @param N The node to transform. 7109 // @param Slct The N operand that is a select. 7110 // @param OtherOp The other N operand (x above). 7111 // @param DCI Context. 7112 // @param AllOnes Require the select constant to be all ones instead of null. 7113 // @returns The new node, or SDValue() on failure. 7114 static 7115 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp, 7116 TargetLowering::DAGCombinerInfo &DCI, 7117 bool AllOnes = false) { 7118 SelectionDAG &DAG = DCI.DAG; 7119 EVT VT = N->getValueType(0); 7120 SDValue NonConstantVal; 7121 SDValue CCOp; 7122 bool SwapSelectOps; 7123 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps, 7124 NonConstantVal, DAG)) 7125 return SDValue(); 7126 7127 // Slct is now know to be the desired identity constant when CC is true. 7128 SDValue TrueVal = OtherOp; 7129 SDValue FalseVal = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT, 7130 OtherOp, NonConstantVal); 7131 // Unless SwapSelectOps says CC should be false. 7132 if (SwapSelectOps) 7133 std::swap(TrueVal, FalseVal); 7134 7135 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT, 7136 CCOp, TrueVal, FalseVal); 7137 } 7138 7139 // Attempt combineSelectAndUse on each operand of a commutative operator N. 7140 static 7141 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes, 7142 TargetLowering::DAGCombinerInfo &DCI) { 7143 SDValue N0 = N->getOperand(0); 7144 SDValue N1 = N->getOperand(1); 7145 if (N0.getNode()->hasOneUse()) { 7146 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes); 7147 if (Result.getNode()) 7148 return Result; 7149 } 7150 if (N1.getNode()->hasOneUse()) { 7151 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes); 7152 if (Result.getNode()) 7153 return Result; 7154 } 7155 return SDValue(); 7156 } 7157 7158 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction 7159 // (only after legalization). 7160 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1, 7161 TargetLowering::DAGCombinerInfo &DCI, 7162 const ARMSubtarget *Subtarget) { 7163 7164 // Only perform optimization if after legalize, and if NEON is available. We 7165 // also expected both operands to be BUILD_VECTORs. 7166 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON() 7167 || N0.getOpcode() != ISD::BUILD_VECTOR 7168 || N1.getOpcode() != ISD::BUILD_VECTOR) 7169 return SDValue(); 7170 7171 // Check output type since VPADDL operand elements can only be 8, 16, or 32. 7172 EVT VT = N->getValueType(0); 7173 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64) 7174 return SDValue(); 7175 7176 // Check that the vector operands are of the right form. 7177 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR 7178 // operands, where N is the size of the formed vector. 7179 // Each EXTRACT_VECTOR should have the same input vector and odd or even 7180 // index such that we have a pair wise add pattern. 7181 7182 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing. 7183 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT) 7184 return SDValue(); 7185 SDValue Vec = N0->getOperand(0)->getOperand(0); 7186 SDNode *V = Vec.getNode(); 7187 unsigned nextIndex = 0; 7188 7189 // For each operands to the ADD which are BUILD_VECTORs, 7190 // check to see if each of their operands are an EXTRACT_VECTOR with 7191 // the same vector and appropriate index. 7192 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) { 7193 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT 7194 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) { 7195 7196 SDValue ExtVec0 = N0->getOperand(i); 7197 SDValue ExtVec1 = N1->getOperand(i); 7198 7199 // First operand is the vector, verify its the same. 7200 if (V != ExtVec0->getOperand(0).getNode() || 7201 V != ExtVec1->getOperand(0).getNode()) 7202 return SDValue(); 7203 7204 // Second is the constant, verify its correct. 7205 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1)); 7206 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1)); 7207 7208 // For the constant, we want to see all the even or all the odd. 7209 if (!C0 || !C1 || C0->getZExtValue() != nextIndex 7210 || C1->getZExtValue() != nextIndex+1) 7211 return SDValue(); 7212 7213 // Increment index. 7214 nextIndex+=2; 7215 } else 7216 return SDValue(); 7217 } 7218 7219 // Create VPADDL node. 7220 SelectionDAG &DAG = DCI.DAG; 7221 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 7222 7223 // Build operand list. 7224 SmallVector<SDValue, 8> Ops; 7225 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, 7226 TLI.getPointerTy())); 7227 7228 // Input is the vector. 7229 Ops.push_back(Vec); 7230 7231 // Get widened type and narrowed type. 7232 MVT widenType; 7233 unsigned numElem = VT.getVectorNumElements(); 7234 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) { 7235 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break; 7236 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break; 7237 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break; 7238 default: 7239 llvm_unreachable("Invalid vector element type for padd optimization."); 7240 } 7241 7242 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(), 7243 widenType, &Ops[0], Ops.size()); 7244 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp); 7245 } 7246 7247 static SDValue findMUL_LOHI(SDValue V) { 7248 if (V->getOpcode() == ISD::UMUL_LOHI || 7249 V->getOpcode() == ISD::SMUL_LOHI) 7250 return V; 7251 return SDValue(); 7252 } 7253 7254 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode, 7255 TargetLowering::DAGCombinerInfo &DCI, 7256 const ARMSubtarget *Subtarget) { 7257 7258 if (Subtarget->isThumb1Only()) return SDValue(); 7259 7260 // Only perform the checks after legalize when the pattern is available. 7261 if (DCI.isBeforeLegalize()) return SDValue(); 7262 7263 // Look for multiply add opportunities. 7264 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where 7265 // each add nodes consumes a value from ISD::UMUL_LOHI and there is 7266 // a glue link from the first add to the second add. 7267 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by 7268 // a S/UMLAL instruction. 7269 // loAdd UMUL_LOHI 7270 // \ / :lo \ :hi 7271 // \ / \ [no multiline comment] 7272 // ADDC | hiAdd 7273 // \ :glue / / 7274 // \ / / 7275 // ADDE 7276 // 7277 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC"); 7278 SDValue AddcOp0 = AddcNode->getOperand(0); 7279 SDValue AddcOp1 = AddcNode->getOperand(1); 7280 7281 // Check if the two operands are from the same mul_lohi node. 7282 if (AddcOp0.getNode() == AddcOp1.getNode()) 7283 return SDValue(); 7284 7285 assert(AddcNode->getNumValues() == 2 && 7286 AddcNode->getValueType(0) == MVT::i32 && 7287 AddcNode->getValueType(1) == MVT::Glue && 7288 "Expect ADDC with two result values: i32, glue"); 7289 7290 // Check that the ADDC adds the low result of the S/UMUL_LOHI. 7291 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI && 7292 AddcOp0->getOpcode() != ISD::SMUL_LOHI && 7293 AddcOp1->getOpcode() != ISD::UMUL_LOHI && 7294 AddcOp1->getOpcode() != ISD::SMUL_LOHI) 7295 return SDValue(); 7296 7297 // Look for the glued ADDE. 7298 SDNode* AddeNode = AddcNode->getGluedUser(); 7299 if (AddeNode == NULL) 7300 return SDValue(); 7301 7302 // Make sure it is really an ADDE. 7303 if (AddeNode->getOpcode() != ISD::ADDE) 7304 return SDValue(); 7305 7306 assert(AddeNode->getNumOperands() == 3 && 7307 AddeNode->getOperand(2).getValueType() == MVT::Glue && 7308 "ADDE node has the wrong inputs"); 7309 7310 // Check for the triangle shape. 7311 SDValue AddeOp0 = AddeNode->getOperand(0); 7312 SDValue AddeOp1 = AddeNode->getOperand(1); 7313 7314 // Make sure that the ADDE operands are not coming from the same node. 7315 if (AddeOp0.getNode() == AddeOp1.getNode()) 7316 return SDValue(); 7317 7318 // Find the MUL_LOHI node walking up ADDE's operands. 7319 bool IsLeftOperandMUL = false; 7320 SDValue MULOp = findMUL_LOHI(AddeOp0); 7321 if (MULOp == SDValue()) 7322 MULOp = findMUL_LOHI(AddeOp1); 7323 else 7324 IsLeftOperandMUL = true; 7325 if (MULOp == SDValue()) 7326 return SDValue(); 7327 7328 // Figure out the right opcode. 7329 unsigned Opc = MULOp->getOpcode(); 7330 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL; 7331 7332 // Figure out the high and low input values to the MLAL node. 7333 SDValue* HiMul = &MULOp; 7334 SDValue* HiAdd = NULL; 7335 SDValue* LoMul = NULL; 7336 SDValue* LowAdd = NULL; 7337 7338 if (IsLeftOperandMUL) 7339 HiAdd = &AddeOp1; 7340 else 7341 HiAdd = &AddeOp0; 7342 7343 7344 if (AddcOp0->getOpcode() == Opc) { 7345 LoMul = &AddcOp0; 7346 LowAdd = &AddcOp1; 7347 } 7348 if (AddcOp1->getOpcode() == Opc) { 7349 LoMul = &AddcOp1; 7350 LowAdd = &AddcOp0; 7351 } 7352 7353 if (LoMul == NULL) 7354 return SDValue(); 7355 7356 if (LoMul->getNode() != HiMul->getNode()) 7357 return SDValue(); 7358 7359 // Create the merged node. 7360 SelectionDAG &DAG = DCI.DAG; 7361 7362 // Build operand list. 7363 SmallVector<SDValue, 8> Ops; 7364 Ops.push_back(LoMul->getOperand(0)); 7365 Ops.push_back(LoMul->getOperand(1)); 7366 Ops.push_back(*LowAdd); 7367 Ops.push_back(*HiAdd); 7368 7369 SDValue MLALNode = DAG.getNode(FinalOpc, AddcNode->getDebugLoc(), 7370 DAG.getVTList(MVT::i32, MVT::i32), 7371 &Ops[0], Ops.size()); 7372 7373 // Replace the ADDs' nodes uses by the MLA node's values. 7374 SDValue HiMLALResult(MLALNode.getNode(), 1); 7375 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult); 7376 7377 SDValue LoMLALResult(MLALNode.getNode(), 0); 7378 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult); 7379 7380 // Return original node to notify the driver to stop replacing. 7381 SDValue resNode(AddcNode, 0); 7382 return resNode; 7383 } 7384 7385 /// PerformADDCCombine - Target-specific dag combine transform from 7386 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL. 7387 static SDValue PerformADDCCombine(SDNode *N, 7388 TargetLowering::DAGCombinerInfo &DCI, 7389 const ARMSubtarget *Subtarget) { 7390 7391 return AddCombineTo64bitMLAL(N, DCI, Subtarget); 7392 7393 } 7394 7395 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with 7396 /// operands N0 and N1. This is a helper for PerformADDCombine that is 7397 /// called with the default operands, and if that fails, with commuted 7398 /// operands. 7399 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1, 7400 TargetLowering::DAGCombinerInfo &DCI, 7401 const ARMSubtarget *Subtarget){ 7402 7403 // Attempt to create vpaddl for this add. 7404 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget); 7405 if (Result.getNode()) 7406 return Result; 7407 7408 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c)) 7409 if (N0.getNode()->hasOneUse()) { 7410 SDValue Result = combineSelectAndUse(N, N0, N1, DCI); 7411 if (Result.getNode()) return Result; 7412 } 7413 return SDValue(); 7414 } 7415 7416 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD. 7417 /// 7418 static SDValue PerformADDCombine(SDNode *N, 7419 TargetLowering::DAGCombinerInfo &DCI, 7420 const ARMSubtarget *Subtarget) { 7421 SDValue N0 = N->getOperand(0); 7422 SDValue N1 = N->getOperand(1); 7423 7424 // First try with the default operand order. 7425 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget); 7426 if (Result.getNode()) 7427 return Result; 7428 7429 // If that didn't work, try again with the operands commuted. 7430 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget); 7431 } 7432 7433 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB. 7434 /// 7435 static SDValue PerformSUBCombine(SDNode *N, 7436 TargetLowering::DAGCombinerInfo &DCI) { 7437 SDValue N0 = N->getOperand(0); 7438 SDValue N1 = N->getOperand(1); 7439 7440 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c)) 7441 if (N1.getNode()->hasOneUse()) { 7442 SDValue Result = combineSelectAndUse(N, N1, N0, DCI); 7443 if (Result.getNode()) return Result; 7444 } 7445 7446 return SDValue(); 7447 } 7448 7449 /// PerformVMULCombine 7450 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the 7451 /// special multiplier accumulator forwarding. 7452 /// vmul d3, d0, d2 7453 /// vmla d3, d1, d2 7454 /// is faster than 7455 /// vadd d3, d0, d1 7456 /// vmul d3, d3, d2 7457 static SDValue PerformVMULCombine(SDNode *N, 7458 TargetLowering::DAGCombinerInfo &DCI, 7459 const ARMSubtarget *Subtarget) { 7460 if (!Subtarget->hasVMLxForwarding()) 7461 return SDValue(); 7462 7463 SelectionDAG &DAG = DCI.DAG; 7464 SDValue N0 = N->getOperand(0); 7465 SDValue N1 = N->getOperand(1); 7466 unsigned Opcode = N0.getOpcode(); 7467 if (Opcode != ISD::ADD && Opcode != ISD::SUB && 7468 Opcode != ISD::FADD && Opcode != ISD::FSUB) { 7469 Opcode = N1.getOpcode(); 7470 if (Opcode != ISD::ADD && Opcode != ISD::SUB && 7471 Opcode != ISD::FADD && Opcode != ISD::FSUB) 7472 return SDValue(); 7473 std::swap(N0, N1); 7474 } 7475 7476 EVT VT = N->getValueType(0); 7477 DebugLoc DL = N->getDebugLoc(); 7478 SDValue N00 = N0->getOperand(0); 7479 SDValue N01 = N0->getOperand(1); 7480 return DAG.getNode(Opcode, DL, VT, 7481 DAG.getNode(ISD::MUL, DL, VT, N00, N1), 7482 DAG.getNode(ISD::MUL, DL, VT, N01, N1)); 7483 } 7484 7485 static SDValue PerformMULCombine(SDNode *N, 7486 TargetLowering::DAGCombinerInfo &DCI, 7487 const ARMSubtarget *Subtarget) { 7488 SelectionDAG &DAG = DCI.DAG; 7489 7490 if (Subtarget->isThumb1Only()) 7491 return SDValue(); 7492 7493 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer()) 7494 return SDValue(); 7495 7496 EVT VT = N->getValueType(0); 7497 if (VT.is64BitVector() || VT.is128BitVector()) 7498 return PerformVMULCombine(N, DCI, Subtarget); 7499 if (VT != MVT::i32) 7500 return SDValue(); 7501 7502 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)); 7503 if (!C) 7504 return SDValue(); 7505 7506 int64_t MulAmt = C->getSExtValue(); 7507 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt); 7508 7509 ShiftAmt = ShiftAmt & (32 - 1); 7510 SDValue V = N->getOperand(0); 7511 DebugLoc DL = N->getDebugLoc(); 7512 7513 SDValue Res; 7514 MulAmt >>= ShiftAmt; 7515 7516 if (MulAmt >= 0) { 7517 if (isPowerOf2_32(MulAmt - 1)) { 7518 // (mul x, 2^N + 1) => (add (shl x, N), x) 7519 Res = DAG.getNode(ISD::ADD, DL, VT, 7520 V, 7521 DAG.getNode(ISD::SHL, DL, VT, 7522 V, 7523 DAG.getConstant(Log2_32(MulAmt - 1), 7524 MVT::i32))); 7525 } else if (isPowerOf2_32(MulAmt + 1)) { 7526 // (mul x, 2^N - 1) => (sub (shl x, N), x) 7527 Res = DAG.getNode(ISD::SUB, DL, VT, 7528 DAG.getNode(ISD::SHL, DL, VT, 7529 V, 7530 DAG.getConstant(Log2_32(MulAmt + 1), 7531 MVT::i32)), 7532 V); 7533 } else 7534 return SDValue(); 7535 } else { 7536 uint64_t MulAmtAbs = -MulAmt; 7537 if (isPowerOf2_32(MulAmtAbs + 1)) { 7538 // (mul x, -(2^N - 1)) => (sub x, (shl x, N)) 7539 Res = DAG.getNode(ISD::SUB, DL, VT, 7540 V, 7541 DAG.getNode(ISD::SHL, DL, VT, 7542 V, 7543 DAG.getConstant(Log2_32(MulAmtAbs + 1), 7544 MVT::i32))); 7545 } else if (isPowerOf2_32(MulAmtAbs - 1)) { 7546 // (mul x, -(2^N + 1)) => - (add (shl x, N), x) 7547 Res = DAG.getNode(ISD::ADD, DL, VT, 7548 V, 7549 DAG.getNode(ISD::SHL, DL, VT, 7550 V, 7551 DAG.getConstant(Log2_32(MulAmtAbs-1), 7552 MVT::i32))); 7553 Res = DAG.getNode(ISD::SUB, DL, VT, 7554 DAG.getConstant(0, MVT::i32),Res); 7555 7556 } else 7557 return SDValue(); 7558 } 7559 7560 if (ShiftAmt != 0) 7561 Res = DAG.getNode(ISD::SHL, DL, VT, 7562 Res, DAG.getConstant(ShiftAmt, MVT::i32)); 7563 7564 // Do not add new nodes to DAG combiner worklist. 7565 DCI.CombineTo(N, Res, false); 7566 return SDValue(); 7567 } 7568 7569 static SDValue PerformANDCombine(SDNode *N, 7570 TargetLowering::DAGCombinerInfo &DCI, 7571 const ARMSubtarget *Subtarget) { 7572 7573 // Attempt to use immediate-form VBIC 7574 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1)); 7575 DebugLoc dl = N->getDebugLoc(); 7576 EVT VT = N->getValueType(0); 7577 SelectionDAG &DAG = DCI.DAG; 7578 7579 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT)) 7580 return SDValue(); 7581 7582 APInt SplatBits, SplatUndef; 7583 unsigned SplatBitSize; 7584 bool HasAnyUndefs; 7585 if (BVN && 7586 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) { 7587 if (SplatBitSize <= 64) { 7588 EVT VbicVT; 7589 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(), 7590 SplatUndef.getZExtValue(), SplatBitSize, 7591 DAG, VbicVT, VT.is128BitVector(), 7592 OtherModImm); 7593 if (Val.getNode()) { 7594 SDValue Input = 7595 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0)); 7596 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val); 7597 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic); 7598 } 7599 } 7600 } 7601 7602 if (!Subtarget->isThumb1Only()) { 7603 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) 7604 SDValue Result = combineSelectAndUseCommutative(N, true, DCI); 7605 if (Result.getNode()) 7606 return Result; 7607 } 7608 7609 return SDValue(); 7610 } 7611 7612 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR 7613 static SDValue PerformORCombine(SDNode *N, 7614 TargetLowering::DAGCombinerInfo &DCI, 7615 const ARMSubtarget *Subtarget) { 7616 // Attempt to use immediate-form VORR 7617 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1)); 7618 DebugLoc dl = N->getDebugLoc(); 7619 EVT VT = N->getValueType(0); 7620 SelectionDAG &DAG = DCI.DAG; 7621 7622 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT)) 7623 return SDValue(); 7624 7625 APInt SplatBits, SplatUndef; 7626 unsigned SplatBitSize; 7627 bool HasAnyUndefs; 7628 if (BVN && Subtarget->hasNEON() && 7629 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) { 7630 if (SplatBitSize <= 64) { 7631 EVT VorrVT; 7632 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(), 7633 SplatUndef.getZExtValue(), SplatBitSize, 7634 DAG, VorrVT, VT.is128BitVector(), 7635 OtherModImm); 7636 if (Val.getNode()) { 7637 SDValue Input = 7638 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0)); 7639 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val); 7640 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr); 7641 } 7642 } 7643 } 7644 7645 if (!Subtarget->isThumb1Only()) { 7646 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c)) 7647 SDValue Result = combineSelectAndUseCommutative(N, false, DCI); 7648 if (Result.getNode()) 7649 return Result; 7650 } 7651 7652 // The code below optimizes (or (and X, Y), Z). 7653 // The AND operand needs to have a single user to make these optimizations 7654 // profitable. 7655 SDValue N0 = N->getOperand(0); 7656 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse()) 7657 return SDValue(); 7658 SDValue N1 = N->getOperand(1); 7659 7660 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant. 7661 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() && 7662 DAG.getTargetLoweringInfo().isTypeLegal(VT)) { 7663 APInt SplatUndef; 7664 unsigned SplatBitSize; 7665 bool HasAnyUndefs; 7666 7667 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1)); 7668 APInt SplatBits0; 7669 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize, 7670 HasAnyUndefs) && !HasAnyUndefs) { 7671 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1)); 7672 APInt SplatBits1; 7673 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize, 7674 HasAnyUndefs) && !HasAnyUndefs && 7675 SplatBits0 == ~SplatBits1) { 7676 // Canonicalize the vector type to make instruction selection simpler. 7677 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32; 7678 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT, 7679 N0->getOperand(1), N0->getOperand(0), 7680 N1->getOperand(0)); 7681 return DAG.getNode(ISD::BITCAST, dl, VT, Result); 7682 } 7683 } 7684 } 7685 7686 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when 7687 // reasonable. 7688 7689 // BFI is only available on V6T2+ 7690 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops()) 7691 return SDValue(); 7692 7693 DebugLoc DL = N->getDebugLoc(); 7694 // 1) or (and A, mask), val => ARMbfi A, val, mask 7695 // iff (val & mask) == val 7696 // 7697 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask 7698 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2) 7699 // && mask == ~mask2 7700 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2) 7701 // && ~mask == mask2 7702 // (i.e., copy a bitfield value into another bitfield of the same width) 7703 7704 if (VT != MVT::i32) 7705 return SDValue(); 7706 7707 SDValue N00 = N0.getOperand(0); 7708 7709 // The value and the mask need to be constants so we can verify this is 7710 // actually a bitfield set. If the mask is 0xffff, we can do better 7711 // via a movt instruction, so don't use BFI in that case. 7712 SDValue MaskOp = N0.getOperand(1); 7713 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp); 7714 if (!MaskC) 7715 return SDValue(); 7716 unsigned Mask = MaskC->getZExtValue(); 7717 if (Mask == 0xffff) 7718 return SDValue(); 7719 SDValue Res; 7720 // Case (1): or (and A, mask), val => ARMbfi A, val, mask 7721 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 7722 if (N1C) { 7723 unsigned Val = N1C->getZExtValue(); 7724 if ((Val & ~Mask) != Val) 7725 return SDValue(); 7726 7727 if (ARM::isBitFieldInvertedMask(Mask)) { 7728 Val >>= CountTrailingZeros_32(~Mask); 7729 7730 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, 7731 DAG.getConstant(Val, MVT::i32), 7732 DAG.getConstant(Mask, MVT::i32)); 7733 7734 // Do not add new nodes to DAG combiner worklist. 7735 DCI.CombineTo(N, Res, false); 7736 return SDValue(); 7737 } 7738 } else if (N1.getOpcode() == ISD::AND) { 7739 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask 7740 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); 7741 if (!N11C) 7742 return SDValue(); 7743 unsigned Mask2 = N11C->getZExtValue(); 7744 7745 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern 7746 // as is to match. 7747 if (ARM::isBitFieldInvertedMask(Mask) && 7748 (Mask == ~Mask2)) { 7749 // The pack halfword instruction works better for masks that fit it, 7750 // so use that when it's available. 7751 if (Subtarget->hasT2ExtractPack() && 7752 (Mask == 0xffff || Mask == 0xffff0000)) 7753 return SDValue(); 7754 // 2a 7755 unsigned amt = CountTrailingZeros_32(Mask2); 7756 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0), 7757 DAG.getConstant(amt, MVT::i32)); 7758 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res, 7759 DAG.getConstant(Mask, MVT::i32)); 7760 // Do not add new nodes to DAG combiner worklist. 7761 DCI.CombineTo(N, Res, false); 7762 return SDValue(); 7763 } else if (ARM::isBitFieldInvertedMask(~Mask) && 7764 (~Mask == Mask2)) { 7765 // The pack halfword instruction works better for masks that fit it, 7766 // so use that when it's available. 7767 if (Subtarget->hasT2ExtractPack() && 7768 (Mask2 == 0xffff || Mask2 == 0xffff0000)) 7769 return SDValue(); 7770 // 2b 7771 unsigned lsb = CountTrailingZeros_32(Mask); 7772 Res = DAG.getNode(ISD::SRL, DL, VT, N00, 7773 DAG.getConstant(lsb, MVT::i32)); 7774 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res, 7775 DAG.getConstant(Mask2, MVT::i32)); 7776 // Do not add new nodes to DAG combiner worklist. 7777 DCI.CombineTo(N, Res, false); 7778 return SDValue(); 7779 } 7780 } 7781 7782 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) && 7783 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) && 7784 ARM::isBitFieldInvertedMask(~Mask)) { 7785 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask 7786 // where lsb(mask) == #shamt and masked bits of B are known zero. 7787 SDValue ShAmt = N00.getOperand(1); 7788 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue(); 7789 unsigned LSB = CountTrailingZeros_32(Mask); 7790 if (ShAmtC != LSB) 7791 return SDValue(); 7792 7793 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0), 7794 DAG.getConstant(~Mask, MVT::i32)); 7795 7796 // Do not add new nodes to DAG combiner worklist. 7797 DCI.CombineTo(N, Res, false); 7798 } 7799 7800 return SDValue(); 7801 } 7802 7803 static SDValue PerformXORCombine(SDNode *N, 7804 TargetLowering::DAGCombinerInfo &DCI, 7805 const ARMSubtarget *Subtarget) { 7806 EVT VT = N->getValueType(0); 7807 SelectionDAG &DAG = DCI.DAG; 7808 7809 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT)) 7810 return SDValue(); 7811 7812 if (!Subtarget->isThumb1Only()) { 7813 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c)) 7814 SDValue Result = combineSelectAndUseCommutative(N, false, DCI); 7815 if (Result.getNode()) 7816 return Result; 7817 } 7818 7819 return SDValue(); 7820 } 7821 7822 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff 7823 /// the bits being cleared by the AND are not demanded by the BFI. 7824 static SDValue PerformBFICombine(SDNode *N, 7825 TargetLowering::DAGCombinerInfo &DCI) { 7826 SDValue N1 = N->getOperand(1); 7827 if (N1.getOpcode() == ISD::AND) { 7828 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); 7829 if (!N11C) 7830 return SDValue(); 7831 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue(); 7832 unsigned LSB = CountTrailingZeros_32(~InvMask); 7833 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB; 7834 unsigned Mask = (1 << Width)-1; 7835 unsigned Mask2 = N11C->getZExtValue(); 7836 if ((Mask & (~Mask2)) == 0) 7837 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0), 7838 N->getOperand(0), N1.getOperand(0), 7839 N->getOperand(2)); 7840 } 7841 return SDValue(); 7842 } 7843 7844 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for 7845 /// ARMISD::VMOVRRD. 7846 static SDValue PerformVMOVRRDCombine(SDNode *N, 7847 TargetLowering::DAGCombinerInfo &DCI) { 7848 // vmovrrd(vmovdrr x, y) -> x,y 7849 SDValue InDouble = N->getOperand(0); 7850 if (InDouble.getOpcode() == ARMISD::VMOVDRR) 7851 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1)); 7852 7853 // vmovrrd(load f64) -> (load i32), (load i32) 7854 SDNode *InNode = InDouble.getNode(); 7855 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() && 7856 InNode->getValueType(0) == MVT::f64 && 7857 InNode->getOperand(1).getOpcode() == ISD::FrameIndex && 7858 !cast<LoadSDNode>(InNode)->isVolatile()) { 7859 // TODO: Should this be done for non-FrameIndex operands? 7860 LoadSDNode *LD = cast<LoadSDNode>(InNode); 7861 7862 SelectionDAG &DAG = DCI.DAG; 7863 DebugLoc DL = LD->getDebugLoc(); 7864 SDValue BasePtr = LD->getBasePtr(); 7865 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr, 7866 LD->getPointerInfo(), LD->isVolatile(), 7867 LD->isNonTemporal(), LD->isInvariant(), 7868 LD->getAlignment()); 7869 7870 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr, 7871 DAG.getConstant(4, MVT::i32)); 7872 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr, 7873 LD->getPointerInfo(), LD->isVolatile(), 7874 LD->isNonTemporal(), LD->isInvariant(), 7875 std::min(4U, LD->getAlignment() / 2)); 7876 7877 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1)); 7878 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2); 7879 DCI.RemoveFromWorklist(LD); 7880 DAG.DeleteNode(LD); 7881 return Result; 7882 } 7883 7884 return SDValue(); 7885 } 7886 7887 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for 7888 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands. 7889 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) { 7890 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X) 7891 SDValue Op0 = N->getOperand(0); 7892 SDValue Op1 = N->getOperand(1); 7893 if (Op0.getOpcode() == ISD::BITCAST) 7894 Op0 = Op0.getOperand(0); 7895 if (Op1.getOpcode() == ISD::BITCAST) 7896 Op1 = Op1.getOperand(0); 7897 if (Op0.getOpcode() == ARMISD::VMOVRRD && 7898 Op0.getNode() == Op1.getNode() && 7899 Op0.getResNo() == 0 && Op1.getResNo() == 1) 7900 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), 7901 N->getValueType(0), Op0.getOperand(0)); 7902 return SDValue(); 7903 } 7904 7905 /// PerformSTORECombine - Target-specific dag combine xforms for 7906 /// ISD::STORE. 7907 static SDValue PerformSTORECombine(SDNode *N, 7908 TargetLowering::DAGCombinerInfo &DCI) { 7909 StoreSDNode *St = cast<StoreSDNode>(N); 7910 if (St->isVolatile()) 7911 return SDValue(); 7912 7913 // Optimize trunc store (of multiple scalars) to shuffle and store. First, 7914 // pack all of the elements in one place. Next, store to memory in fewer 7915 // chunks. 7916 SDValue StVal = St->getValue(); 7917 EVT VT = StVal.getValueType(); 7918 if (St->isTruncatingStore() && VT.isVector()) { 7919 SelectionDAG &DAG = DCI.DAG; 7920 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 7921 EVT StVT = St->getMemoryVT(); 7922 unsigned NumElems = VT.getVectorNumElements(); 7923 assert(StVT != VT && "Cannot truncate to the same type"); 7924 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits(); 7925 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits(); 7926 7927 // From, To sizes and ElemCount must be pow of two 7928 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue(); 7929 7930 // We are going to use the original vector elt for storing. 7931 // Accumulated smaller vector elements must be a multiple of the store size. 7932 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue(); 7933 7934 unsigned SizeRatio = FromEltSz / ToEltSz; 7935 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits()); 7936 7937 // Create a type on which we perform the shuffle. 7938 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(), 7939 NumElems*SizeRatio); 7940 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits()); 7941 7942 DebugLoc DL = St->getDebugLoc(); 7943 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal); 7944 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1); 7945 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio; 7946 7947 // Can't shuffle using an illegal type. 7948 if (!TLI.isTypeLegal(WideVecVT)) return SDValue(); 7949 7950 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec, 7951 DAG.getUNDEF(WideVec.getValueType()), 7952 ShuffleVec.data()); 7953 // At this point all of the data is stored at the bottom of the 7954 // register. We now need to save it to mem. 7955 7956 // Find the largest store unit 7957 MVT StoreType = MVT::i8; 7958 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE; 7959 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) { 7960 MVT Tp = (MVT::SimpleValueType)tp; 7961 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz) 7962 StoreType = Tp; 7963 } 7964 // Didn't find a legal store type. 7965 if (!TLI.isTypeLegal(StoreType)) 7966 return SDValue(); 7967 7968 // Bitcast the original vector into a vector of store-size units 7969 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(), 7970 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits()); 7971 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits()); 7972 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff); 7973 SmallVector<SDValue, 8> Chains; 7974 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8, 7975 TLI.getPointerTy()); 7976 SDValue BasePtr = St->getBasePtr(); 7977 7978 // Perform one or more big stores into memory. 7979 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits(); 7980 for (unsigned I = 0; I < E; I++) { 7981 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, 7982 StoreType, ShuffWide, 7983 DAG.getIntPtrConstant(I)); 7984 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr, 7985 St->getPointerInfo(), St->isVolatile(), 7986 St->isNonTemporal(), St->getAlignment()); 7987 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr, 7988 Increment); 7989 Chains.push_back(Ch); 7990 } 7991 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0], 7992 Chains.size()); 7993 } 7994 7995 if (!ISD::isNormalStore(St)) 7996 return SDValue(); 7997 7998 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and 7999 // ARM stores of arguments in the same cache line. 8000 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR && 8001 StVal.getNode()->hasOneUse()) { 8002 SelectionDAG &DAG = DCI.DAG; 8003 DebugLoc DL = St->getDebugLoc(); 8004 SDValue BasePtr = St->getBasePtr(); 8005 SDValue NewST1 = DAG.getStore(St->getChain(), DL, 8006 StVal.getNode()->getOperand(0), BasePtr, 8007 St->getPointerInfo(), St->isVolatile(), 8008 St->isNonTemporal(), St->getAlignment()); 8009 8010 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr, 8011 DAG.getConstant(4, MVT::i32)); 8012 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1), 8013 OffsetPtr, St->getPointerInfo(), St->isVolatile(), 8014 St->isNonTemporal(), 8015 std::min(4U, St->getAlignment() / 2)); 8016 } 8017 8018 if (StVal.getValueType() != MVT::i64 || 8019 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT) 8020 return SDValue(); 8021 8022 // Bitcast an i64 store extracted from a vector to f64. 8023 // Otherwise, the i64 value will be legalized to a pair of i32 values. 8024 SelectionDAG &DAG = DCI.DAG; 8025 DebugLoc dl = StVal.getDebugLoc(); 8026 SDValue IntVec = StVal.getOperand(0); 8027 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, 8028 IntVec.getValueType().getVectorNumElements()); 8029 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec); 8030 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, 8031 Vec, StVal.getOperand(1)); 8032 dl = N->getDebugLoc(); 8033 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt); 8034 // Make the DAGCombiner fold the bitcasts. 8035 DCI.AddToWorklist(Vec.getNode()); 8036 DCI.AddToWorklist(ExtElt.getNode()); 8037 DCI.AddToWorklist(V.getNode()); 8038 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(), 8039 St->getPointerInfo(), St->isVolatile(), 8040 St->isNonTemporal(), St->getAlignment(), 8041 St->getTBAAInfo()); 8042 } 8043 8044 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node 8045 /// are normal, non-volatile loads. If so, it is profitable to bitcast an 8046 /// i64 vector to have f64 elements, since the value can then be loaded 8047 /// directly into a VFP register. 8048 static bool hasNormalLoadOperand(SDNode *N) { 8049 unsigned NumElts = N->getValueType(0).getVectorNumElements(); 8050 for (unsigned i = 0; i < NumElts; ++i) { 8051 SDNode *Elt = N->getOperand(i).getNode(); 8052 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile()) 8053 return true; 8054 } 8055 return false; 8056 } 8057 8058 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for 8059 /// ISD::BUILD_VECTOR. 8060 static SDValue PerformBUILD_VECTORCombine(SDNode *N, 8061 TargetLowering::DAGCombinerInfo &DCI){ 8062 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X): 8063 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value 8064 // into a pair of GPRs, which is fine when the value is used as a scalar, 8065 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD. 8066 SelectionDAG &DAG = DCI.DAG; 8067 if (N->getNumOperands() == 2) { 8068 SDValue RV = PerformVMOVDRRCombine(N, DAG); 8069 if (RV.getNode()) 8070 return RV; 8071 } 8072 8073 // Load i64 elements as f64 values so that type legalization does not split 8074 // them up into i32 values. 8075 EVT VT = N->getValueType(0); 8076 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N)) 8077 return SDValue(); 8078 DebugLoc dl = N->getDebugLoc(); 8079 SmallVector<SDValue, 8> Ops; 8080 unsigned NumElts = VT.getVectorNumElements(); 8081 for (unsigned i = 0; i < NumElts; ++i) { 8082 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i)); 8083 Ops.push_back(V); 8084 // Make the DAGCombiner fold the bitcast. 8085 DCI.AddToWorklist(V.getNode()); 8086 } 8087 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts); 8088 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts); 8089 return DAG.getNode(ISD::BITCAST, dl, VT, BV); 8090 } 8091 8092 /// PerformInsertEltCombine - Target-specific dag combine xforms for 8093 /// ISD::INSERT_VECTOR_ELT. 8094 static SDValue PerformInsertEltCombine(SDNode *N, 8095 TargetLowering::DAGCombinerInfo &DCI) { 8096 // Bitcast an i64 load inserted into a vector to f64. 8097 // Otherwise, the i64 value will be legalized to a pair of i32 values. 8098 EVT VT = N->getValueType(0); 8099 SDNode *Elt = N->getOperand(1).getNode(); 8100 if (VT.getVectorElementType() != MVT::i64 || 8101 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile()) 8102 return SDValue(); 8103 8104 SelectionDAG &DAG = DCI.DAG; 8105 DebugLoc dl = N->getDebugLoc(); 8106 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, 8107 VT.getVectorNumElements()); 8108 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0)); 8109 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1)); 8110 // Make the DAGCombiner fold the bitcasts. 8111 DCI.AddToWorklist(Vec.getNode()); 8112 DCI.AddToWorklist(V.getNode()); 8113 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT, 8114 Vec, V, N->getOperand(2)); 8115 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt); 8116 } 8117 8118 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for 8119 /// ISD::VECTOR_SHUFFLE. 8120 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) { 8121 // The LLVM shufflevector instruction does not require the shuffle mask 8122 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does 8123 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the 8124 // operands do not match the mask length, they are extended by concatenating 8125 // them with undef vectors. That is probably the right thing for other 8126 // targets, but for NEON it is better to concatenate two double-register 8127 // size vector operands into a single quad-register size vector. Do that 8128 // transformation here: 8129 // shuffle(concat(v1, undef), concat(v2, undef)) -> 8130 // shuffle(concat(v1, v2), undef) 8131 SDValue Op0 = N->getOperand(0); 8132 SDValue Op1 = N->getOperand(1); 8133 if (Op0.getOpcode() != ISD::CONCAT_VECTORS || 8134 Op1.getOpcode() != ISD::CONCAT_VECTORS || 8135 Op0.getNumOperands() != 2 || 8136 Op1.getNumOperands() != 2) 8137 return SDValue(); 8138 SDValue Concat0Op1 = Op0.getOperand(1); 8139 SDValue Concat1Op1 = Op1.getOperand(1); 8140 if (Concat0Op1.getOpcode() != ISD::UNDEF || 8141 Concat1Op1.getOpcode() != ISD::UNDEF) 8142 return SDValue(); 8143 // Skip the transformation if any of the types are illegal. 8144 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 8145 EVT VT = N->getValueType(0); 8146 if (!TLI.isTypeLegal(VT) || 8147 !TLI.isTypeLegal(Concat0Op1.getValueType()) || 8148 !TLI.isTypeLegal(Concat1Op1.getValueType())) 8149 return SDValue(); 8150 8151 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT, 8152 Op0.getOperand(0), Op1.getOperand(0)); 8153 // Translate the shuffle mask. 8154 SmallVector<int, 16> NewMask; 8155 unsigned NumElts = VT.getVectorNumElements(); 8156 unsigned HalfElts = NumElts/2; 8157 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 8158 for (unsigned n = 0; n < NumElts; ++n) { 8159 int MaskElt = SVN->getMaskElt(n); 8160 int NewElt = -1; 8161 if (MaskElt < (int)HalfElts) 8162 NewElt = MaskElt; 8163 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts)) 8164 NewElt = HalfElts + MaskElt - NumElts; 8165 NewMask.push_back(NewElt); 8166 } 8167 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat, 8168 DAG.getUNDEF(VT), NewMask.data()); 8169 } 8170 8171 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and 8172 /// NEON load/store intrinsics to merge base address updates. 8173 static SDValue CombineBaseUpdate(SDNode *N, 8174 TargetLowering::DAGCombinerInfo &DCI) { 8175 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer()) 8176 return SDValue(); 8177 8178 SelectionDAG &DAG = DCI.DAG; 8179 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID || 8180 N->getOpcode() == ISD::INTRINSIC_W_CHAIN); 8181 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1); 8182 SDValue Addr = N->getOperand(AddrOpIdx); 8183 8184 // Search for a use of the address operand that is an increment. 8185 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), 8186 UE = Addr.getNode()->use_end(); UI != UE; ++UI) { 8187 SDNode *User = *UI; 8188 if (User->getOpcode() != ISD::ADD || 8189 UI.getUse().getResNo() != Addr.getResNo()) 8190 continue; 8191 8192 // Check that the add is independent of the load/store. Otherwise, folding 8193 // it would create a cycle. 8194 if (User->isPredecessorOf(N) || N->isPredecessorOf(User)) 8195 continue; 8196 8197 // Find the new opcode for the updating load/store. 8198 bool isLoad = true; 8199 bool isLaneOp = false; 8200 unsigned NewOpc = 0; 8201 unsigned NumVecs = 0; 8202 if (isIntrinsic) { 8203 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); 8204 switch (IntNo) { 8205 default: llvm_unreachable("unexpected intrinsic for Neon base update"); 8206 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD; 8207 NumVecs = 1; break; 8208 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD; 8209 NumVecs = 2; break; 8210 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD; 8211 NumVecs = 3; break; 8212 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD; 8213 NumVecs = 4; break; 8214 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD; 8215 NumVecs = 2; isLaneOp = true; break; 8216 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD; 8217 NumVecs = 3; isLaneOp = true; break; 8218 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD; 8219 NumVecs = 4; isLaneOp = true; break; 8220 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD; 8221 NumVecs = 1; isLoad = false; break; 8222 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD; 8223 NumVecs = 2; isLoad = false; break; 8224 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD; 8225 NumVecs = 3; isLoad = false; break; 8226 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD; 8227 NumVecs = 4; isLoad = false; break; 8228 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD; 8229 NumVecs = 2; isLoad = false; isLaneOp = true; break; 8230 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD; 8231 NumVecs = 3; isLoad = false; isLaneOp = true; break; 8232 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD; 8233 NumVecs = 4; isLoad = false; isLaneOp = true; break; 8234 } 8235 } else { 8236 isLaneOp = true; 8237 switch (N->getOpcode()) { 8238 default: llvm_unreachable("unexpected opcode for Neon base update"); 8239 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break; 8240 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break; 8241 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break; 8242 } 8243 } 8244 8245 // Find the size of memory referenced by the load/store. 8246 EVT VecTy; 8247 if (isLoad) 8248 VecTy = N->getValueType(0); 8249 else 8250 VecTy = N->getOperand(AddrOpIdx+1).getValueType(); 8251 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8; 8252 if (isLaneOp) 8253 NumBytes /= VecTy.getVectorNumElements(); 8254 8255 // If the increment is a constant, it must match the memory ref size. 8256 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0); 8257 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) { 8258 uint64_t IncVal = CInc->getZExtValue(); 8259 if (IncVal != NumBytes) 8260 continue; 8261 } else if (NumBytes >= 3 * 16) { 8262 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two 8263 // separate instructions that make it harder to use a non-constant update. 8264 continue; 8265 } 8266 8267 // Create the new updating load/store node. 8268 EVT Tys[6]; 8269 unsigned NumResultVecs = (isLoad ? NumVecs : 0); 8270 unsigned n; 8271 for (n = 0; n < NumResultVecs; ++n) 8272 Tys[n] = VecTy; 8273 Tys[n++] = MVT::i32; 8274 Tys[n] = MVT::Other; 8275 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2); 8276 SmallVector<SDValue, 8> Ops; 8277 Ops.push_back(N->getOperand(0)); // incoming chain 8278 Ops.push_back(N->getOperand(AddrOpIdx)); 8279 Ops.push_back(Inc); 8280 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) { 8281 Ops.push_back(N->getOperand(i)); 8282 } 8283 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N); 8284 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys, 8285 Ops.data(), Ops.size(), 8286 MemInt->getMemoryVT(), 8287 MemInt->getMemOperand()); 8288 8289 // Update the uses. 8290 std::vector<SDValue> NewResults; 8291 for (unsigned i = 0; i < NumResultVecs; ++i) { 8292 NewResults.push_back(SDValue(UpdN.getNode(), i)); 8293 } 8294 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain 8295 DCI.CombineTo(N, NewResults); 8296 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs)); 8297 8298 break; 8299 } 8300 return SDValue(); 8301 } 8302 8303 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a 8304 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic 8305 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and 8306 /// return true. 8307 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { 8308 SelectionDAG &DAG = DCI.DAG; 8309 EVT VT = N->getValueType(0); 8310 // vldN-dup instructions only support 64-bit vectors for N > 1. 8311 if (!VT.is64BitVector()) 8312 return false; 8313 8314 // Check if the VDUPLANE operand is a vldN-dup intrinsic. 8315 SDNode *VLD = N->getOperand(0).getNode(); 8316 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN) 8317 return false; 8318 unsigned NumVecs = 0; 8319 unsigned NewOpc = 0; 8320 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue(); 8321 if (IntNo == Intrinsic::arm_neon_vld2lane) { 8322 NumVecs = 2; 8323 NewOpc = ARMISD::VLD2DUP; 8324 } else if (IntNo == Intrinsic::arm_neon_vld3lane) { 8325 NumVecs = 3; 8326 NewOpc = ARMISD::VLD3DUP; 8327 } else if (IntNo == Intrinsic::arm_neon_vld4lane) { 8328 NumVecs = 4; 8329 NewOpc = ARMISD::VLD4DUP; 8330 } else { 8331 return false; 8332 } 8333 8334 // First check that all the vldN-lane uses are VDUPLANEs and that the lane 8335 // numbers match the load. 8336 unsigned VLDLaneNo = 8337 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue(); 8338 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end(); 8339 UI != UE; ++UI) { 8340 // Ignore uses of the chain result. 8341 if (UI.getUse().getResNo() == NumVecs) 8342 continue; 8343 SDNode *User = *UI; 8344 if (User->getOpcode() != ARMISD::VDUPLANE || 8345 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue()) 8346 return false; 8347 } 8348 8349 // Create the vldN-dup node. 8350 EVT Tys[5]; 8351 unsigned n; 8352 for (n = 0; n < NumVecs; ++n) 8353 Tys[n] = VT; 8354 Tys[n] = MVT::Other; 8355 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1); 8356 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) }; 8357 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD); 8358 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys, 8359 Ops, 2, VLDMemInt->getMemoryVT(), 8360 VLDMemInt->getMemOperand()); 8361 8362 // Update the uses. 8363 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end(); 8364 UI != UE; ++UI) { 8365 unsigned ResNo = UI.getUse().getResNo(); 8366 // Ignore uses of the chain result. 8367 if (ResNo == NumVecs) 8368 continue; 8369 SDNode *User = *UI; 8370 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo)); 8371 } 8372 8373 // Now the vldN-lane intrinsic is dead except for its chain result. 8374 // Update uses of the chain. 8375 std::vector<SDValue> VLDDupResults; 8376 for (unsigned n = 0; n < NumVecs; ++n) 8377 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n)); 8378 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs)); 8379 DCI.CombineTo(VLD, VLDDupResults); 8380 8381 return true; 8382 } 8383 8384 /// PerformVDUPLANECombine - Target-specific dag combine xforms for 8385 /// ARMISD::VDUPLANE. 8386 static SDValue PerformVDUPLANECombine(SDNode *N, 8387 TargetLowering::DAGCombinerInfo &DCI) { 8388 SDValue Op = N->getOperand(0); 8389 8390 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses 8391 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation. 8392 if (CombineVLDDUP(N, DCI)) 8393 return SDValue(N, 0); 8394 8395 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is 8396 // redundant. Ignore bit_converts for now; element sizes are checked below. 8397 while (Op.getOpcode() == ISD::BITCAST) 8398 Op = Op.getOperand(0); 8399 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM) 8400 return SDValue(); 8401 8402 // Make sure the VMOV element size is not bigger than the VDUPLANE elements. 8403 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits(); 8404 // The canonical VMOV for a zero vector uses a 32-bit element size. 8405 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 8406 unsigned EltBits; 8407 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0) 8408 EltSize = 8; 8409 EVT VT = N->getValueType(0); 8410 if (EltSize > VT.getVectorElementType().getSizeInBits()) 8411 return SDValue(); 8412 8413 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op); 8414 } 8415 8416 // isConstVecPow2 - Return true if each vector element is a power of 2, all 8417 // elements are the same constant, C, and Log2(C) ranges from 1 to 32. 8418 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C) 8419 { 8420 integerPart cN; 8421 integerPart c0 = 0; 8422 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements(); 8423 I != E; I++) { 8424 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I)); 8425 if (!C) 8426 return false; 8427 8428 bool isExact; 8429 APFloat APF = C->getValueAPF(); 8430 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact) 8431 != APFloat::opOK || !isExact) 8432 return false; 8433 8434 c0 = (I == 0) ? cN : c0; 8435 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32) 8436 return false; 8437 } 8438 C = c0; 8439 return true; 8440 } 8441 8442 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD) 8443 /// can replace combinations of VMUL and VCVT (floating-point to integer) 8444 /// when the VMUL has a constant operand that is a power of 2. 8445 /// 8446 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>): 8447 /// vmul.f32 d16, d17, d16 8448 /// vcvt.s32.f32 d16, d16 8449 /// becomes: 8450 /// vcvt.s32.f32 d16, d16, #3 8451 static SDValue PerformVCVTCombine(SDNode *N, 8452 TargetLowering::DAGCombinerInfo &DCI, 8453 const ARMSubtarget *Subtarget) { 8454 SelectionDAG &DAG = DCI.DAG; 8455 SDValue Op = N->getOperand(0); 8456 8457 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() || 8458 Op.getOpcode() != ISD::FMUL) 8459 return SDValue(); 8460 8461 uint64_t C; 8462 SDValue N0 = Op->getOperand(0); 8463 SDValue ConstVec = Op->getOperand(1); 8464 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT; 8465 8466 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR || 8467 !isConstVecPow2(ConstVec, isSigned, C)) 8468 return SDValue(); 8469 8470 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs : 8471 Intrinsic::arm_neon_vcvtfp2fxu; 8472 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(), 8473 N->getValueType(0), 8474 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0, 8475 DAG.getConstant(Log2_64(C), MVT::i32)); 8476 } 8477 8478 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD) 8479 /// can replace combinations of VCVT (integer to floating-point) and VDIV 8480 /// when the VDIV has a constant operand that is a power of 2. 8481 /// 8482 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>): 8483 /// vcvt.f32.s32 d16, d16 8484 /// vdiv.f32 d16, d17, d16 8485 /// becomes: 8486 /// vcvt.f32.s32 d16, d16, #3 8487 static SDValue PerformVDIVCombine(SDNode *N, 8488 TargetLowering::DAGCombinerInfo &DCI, 8489 const ARMSubtarget *Subtarget) { 8490 SelectionDAG &DAG = DCI.DAG; 8491 SDValue Op = N->getOperand(0); 8492 unsigned OpOpcode = Op.getNode()->getOpcode(); 8493 8494 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() || 8495 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP)) 8496 return SDValue(); 8497 8498 uint64_t C; 8499 SDValue ConstVec = N->getOperand(1); 8500 bool isSigned = OpOpcode == ISD::SINT_TO_FP; 8501 8502 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR || 8503 !isConstVecPow2(ConstVec, isSigned, C)) 8504 return SDValue(); 8505 8506 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp : 8507 Intrinsic::arm_neon_vcvtfxu2fp; 8508 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(), 8509 Op.getValueType(), 8510 DAG.getConstant(IntrinsicOpcode, MVT::i32), 8511 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32)); 8512 } 8513 8514 /// Getvshiftimm - Check if this is a valid build_vector for the immediate 8515 /// operand of a vector shift operation, where all the elements of the 8516 /// build_vector must have the same constant integer value. 8517 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) { 8518 // Ignore bit_converts. 8519 while (Op.getOpcode() == ISD::BITCAST) 8520 Op = Op.getOperand(0); 8521 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode()); 8522 APInt SplatBits, SplatUndef; 8523 unsigned SplatBitSize; 8524 bool HasAnyUndefs; 8525 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, 8526 HasAnyUndefs, ElementBits) || 8527 SplatBitSize > ElementBits) 8528 return false; 8529 Cnt = SplatBits.getSExtValue(); 8530 return true; 8531 } 8532 8533 /// isVShiftLImm - Check if this is a valid build_vector for the immediate 8534 /// operand of a vector shift left operation. That value must be in the range: 8535 /// 0 <= Value < ElementBits for a left shift; or 8536 /// 0 <= Value <= ElementBits for a long left shift. 8537 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) { 8538 assert(VT.isVector() && "vector shift count is not a vector type"); 8539 unsigned ElementBits = VT.getVectorElementType().getSizeInBits(); 8540 if (! getVShiftImm(Op, ElementBits, Cnt)) 8541 return false; 8542 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits); 8543 } 8544 8545 /// isVShiftRImm - Check if this is a valid build_vector for the immediate 8546 /// operand of a vector shift right operation. For a shift opcode, the value 8547 /// is positive, but for an intrinsic the value count must be negative. The 8548 /// absolute value must be in the range: 8549 /// 1 <= |Value| <= ElementBits for a right shift; or 8550 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift. 8551 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic, 8552 int64_t &Cnt) { 8553 assert(VT.isVector() && "vector shift count is not a vector type"); 8554 unsigned ElementBits = VT.getVectorElementType().getSizeInBits(); 8555 if (! getVShiftImm(Op, ElementBits, Cnt)) 8556 return false; 8557 if (isIntrinsic) 8558 Cnt = -Cnt; 8559 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits)); 8560 } 8561 8562 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics. 8563 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) { 8564 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); 8565 switch (IntNo) { 8566 default: 8567 // Don't do anything for most intrinsics. 8568 break; 8569 8570 // Vector shifts: check for immediate versions and lower them. 8571 // Note: This is done during DAG combining instead of DAG legalizing because 8572 // the build_vectors for 64-bit vector element shift counts are generally 8573 // not legal, and it is hard to see their values after they get legalized to 8574 // loads from a constant pool. 8575 case Intrinsic::arm_neon_vshifts: 8576 case Intrinsic::arm_neon_vshiftu: 8577 case Intrinsic::arm_neon_vshiftls: 8578 case Intrinsic::arm_neon_vshiftlu: 8579 case Intrinsic::arm_neon_vshiftn: 8580 case Intrinsic::arm_neon_vrshifts: 8581 case Intrinsic::arm_neon_vrshiftu: 8582 case Intrinsic::arm_neon_vrshiftn: 8583 case Intrinsic::arm_neon_vqshifts: 8584 case Intrinsic::arm_neon_vqshiftu: 8585 case Intrinsic::arm_neon_vqshiftsu: 8586 case Intrinsic::arm_neon_vqshiftns: 8587 case Intrinsic::arm_neon_vqshiftnu: 8588 case Intrinsic::arm_neon_vqshiftnsu: 8589 case Intrinsic::arm_neon_vqrshiftns: 8590 case Intrinsic::arm_neon_vqrshiftnu: 8591 case Intrinsic::arm_neon_vqrshiftnsu: { 8592 EVT VT = N->getOperand(1).getValueType(); 8593 int64_t Cnt; 8594 unsigned VShiftOpc = 0; 8595 8596 switch (IntNo) { 8597 case Intrinsic::arm_neon_vshifts: 8598 case Intrinsic::arm_neon_vshiftu: 8599 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) { 8600 VShiftOpc = ARMISD::VSHL; 8601 break; 8602 } 8603 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) { 8604 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ? 8605 ARMISD::VSHRs : ARMISD::VSHRu); 8606 break; 8607 } 8608 return SDValue(); 8609 8610 case Intrinsic::arm_neon_vshiftls: 8611 case Intrinsic::arm_neon_vshiftlu: 8612 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt)) 8613 break; 8614 llvm_unreachable("invalid shift count for vshll intrinsic"); 8615 8616 case Intrinsic::arm_neon_vrshifts: 8617 case Intrinsic::arm_neon_vrshiftu: 8618 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) 8619 break; 8620 return SDValue(); 8621 8622 case Intrinsic::arm_neon_vqshifts: 8623 case Intrinsic::arm_neon_vqshiftu: 8624 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) 8625 break; 8626 return SDValue(); 8627 8628 case Intrinsic::arm_neon_vqshiftsu: 8629 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) 8630 break; 8631 llvm_unreachable("invalid shift count for vqshlu intrinsic"); 8632 8633 case Intrinsic::arm_neon_vshiftn: 8634 case Intrinsic::arm_neon_vrshiftn: 8635 case Intrinsic::arm_neon_vqshiftns: 8636 case Intrinsic::arm_neon_vqshiftnu: 8637 case Intrinsic::arm_neon_vqshiftnsu: 8638 case Intrinsic::arm_neon_vqrshiftns: 8639 case Intrinsic::arm_neon_vqrshiftnu: 8640 case Intrinsic::arm_neon_vqrshiftnsu: 8641 // Narrowing shifts require an immediate right shift. 8642 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt)) 8643 break; 8644 llvm_unreachable("invalid shift count for narrowing vector shift " 8645 "intrinsic"); 8646 8647 default: 8648 llvm_unreachable("unhandled vector shift"); 8649 } 8650 8651 switch (IntNo) { 8652 case Intrinsic::arm_neon_vshifts: 8653 case Intrinsic::arm_neon_vshiftu: 8654 // Opcode already set above. 8655 break; 8656 case Intrinsic::arm_neon_vshiftls: 8657 case Intrinsic::arm_neon_vshiftlu: 8658 if (Cnt == VT.getVectorElementType().getSizeInBits()) 8659 VShiftOpc = ARMISD::VSHLLi; 8660 else 8661 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ? 8662 ARMISD::VSHLLs : ARMISD::VSHLLu); 8663 break; 8664 case Intrinsic::arm_neon_vshiftn: 8665 VShiftOpc = ARMISD::VSHRN; break; 8666 case Intrinsic::arm_neon_vrshifts: 8667 VShiftOpc = ARMISD::VRSHRs; break; 8668 case Intrinsic::arm_neon_vrshiftu: 8669 VShiftOpc = ARMISD::VRSHRu; break; 8670 case Intrinsic::arm_neon_vrshiftn: 8671 VShiftOpc = ARMISD::VRSHRN; break; 8672 case Intrinsic::arm_neon_vqshifts: 8673 VShiftOpc = ARMISD::VQSHLs; break; 8674 case Intrinsic::arm_neon_vqshiftu: 8675 VShiftOpc = ARMISD::VQSHLu; break; 8676 case Intrinsic::arm_neon_vqshiftsu: 8677 VShiftOpc = ARMISD::VQSHLsu; break; 8678 case Intrinsic::arm_neon_vqshiftns: 8679 VShiftOpc = ARMISD::VQSHRNs; break; 8680 case Intrinsic::arm_neon_vqshiftnu: 8681 VShiftOpc = ARMISD::VQSHRNu; break; 8682 case Intrinsic::arm_neon_vqshiftnsu: 8683 VShiftOpc = ARMISD::VQSHRNsu; break; 8684 case Intrinsic::arm_neon_vqrshiftns: 8685 VShiftOpc = ARMISD::VQRSHRNs; break; 8686 case Intrinsic::arm_neon_vqrshiftnu: 8687 VShiftOpc = ARMISD::VQRSHRNu; break; 8688 case Intrinsic::arm_neon_vqrshiftnsu: 8689 VShiftOpc = ARMISD::VQRSHRNsu; break; 8690 } 8691 8692 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0), 8693 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32)); 8694 } 8695 8696 case Intrinsic::arm_neon_vshiftins: { 8697 EVT VT = N->getOperand(1).getValueType(); 8698 int64_t Cnt; 8699 unsigned VShiftOpc = 0; 8700 8701 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt)) 8702 VShiftOpc = ARMISD::VSLI; 8703 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt)) 8704 VShiftOpc = ARMISD::VSRI; 8705 else { 8706 llvm_unreachable("invalid shift count for vsli/vsri intrinsic"); 8707 } 8708 8709 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0), 8710 N->getOperand(1), N->getOperand(2), 8711 DAG.getConstant(Cnt, MVT::i32)); 8712 } 8713 8714 case Intrinsic::arm_neon_vqrshifts: 8715 case Intrinsic::arm_neon_vqrshiftu: 8716 // No immediate versions of these to check for. 8717 break; 8718 } 8719 8720 return SDValue(); 8721 } 8722 8723 /// PerformShiftCombine - Checks for immediate versions of vector shifts and 8724 /// lowers them. As with the vector shift intrinsics, this is done during DAG 8725 /// combining instead of DAG legalizing because the build_vectors for 64-bit 8726 /// vector element shift counts are generally not legal, and it is hard to see 8727 /// their values after they get legalized to loads from a constant pool. 8728 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG, 8729 const ARMSubtarget *ST) { 8730 EVT VT = N->getValueType(0); 8731 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) { 8732 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high 8733 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16. 8734 SDValue N1 = N->getOperand(1); 8735 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) { 8736 SDValue N0 = N->getOperand(0); 8737 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP && 8738 DAG.MaskedValueIsZero(N0.getOperand(0), 8739 APInt::getHighBitsSet(32, 16))) 8740 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1); 8741 } 8742 } 8743 8744 // Nothing to be done for scalar shifts. 8745 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 8746 if (!VT.isVector() || !TLI.isTypeLegal(VT)) 8747 return SDValue(); 8748 8749 assert(ST->hasNEON() && "unexpected vector shift"); 8750 int64_t Cnt; 8751 8752 switch (N->getOpcode()) { 8753 default: llvm_unreachable("unexpected shift opcode"); 8754 8755 case ISD::SHL: 8756 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) 8757 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0), 8758 DAG.getConstant(Cnt, MVT::i32)); 8759 break; 8760 8761 case ISD::SRA: 8762 case ISD::SRL: 8763 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) { 8764 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ? 8765 ARMISD::VSHRs : ARMISD::VSHRu); 8766 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0), 8767 DAG.getConstant(Cnt, MVT::i32)); 8768 } 8769 } 8770 return SDValue(); 8771 } 8772 8773 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND, 8774 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND. 8775 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG, 8776 const ARMSubtarget *ST) { 8777 SDValue N0 = N->getOperand(0); 8778 8779 // Check for sign- and zero-extensions of vector extract operations of 8- 8780 // and 16-bit vector elements. NEON supports these directly. They are 8781 // handled during DAG combining because type legalization will promote them 8782 // to 32-bit types and it is messy to recognize the operations after that. 8783 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) { 8784 SDValue Vec = N0.getOperand(0); 8785 SDValue Lane = N0.getOperand(1); 8786 EVT VT = N->getValueType(0); 8787 EVT EltVT = N0.getValueType(); 8788 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 8789 8790 if (VT == MVT::i32 && 8791 (EltVT == MVT::i8 || EltVT == MVT::i16) && 8792 TLI.isTypeLegal(Vec.getValueType()) && 8793 isa<ConstantSDNode>(Lane)) { 8794 8795 unsigned Opc = 0; 8796 switch (N->getOpcode()) { 8797 default: llvm_unreachable("unexpected opcode"); 8798 case ISD::SIGN_EXTEND: 8799 Opc = ARMISD::VGETLANEs; 8800 break; 8801 case ISD::ZERO_EXTEND: 8802 case ISD::ANY_EXTEND: 8803 Opc = ARMISD::VGETLANEu; 8804 break; 8805 } 8806 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane); 8807 } 8808 } 8809 8810 return SDValue(); 8811 } 8812 8813 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC 8814 /// to match f32 max/min patterns to use NEON vmax/vmin instructions. 8815 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG, 8816 const ARMSubtarget *ST) { 8817 // If the target supports NEON, try to use vmax/vmin instructions for f32 8818 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set, 8819 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is 8820 // a NaN; only do the transformation when it matches that behavior. 8821 8822 // For now only do this when using NEON for FP operations; if using VFP, it 8823 // is not obvious that the benefit outweighs the cost of switching to the 8824 // NEON pipeline. 8825 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() || 8826 N->getValueType(0) != MVT::f32) 8827 return SDValue(); 8828 8829 SDValue CondLHS = N->getOperand(0); 8830 SDValue CondRHS = N->getOperand(1); 8831 SDValue LHS = N->getOperand(2); 8832 SDValue RHS = N->getOperand(3); 8833 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get(); 8834 8835 unsigned Opcode = 0; 8836 bool IsReversed; 8837 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) { 8838 IsReversed = false; // x CC y ? x : y 8839 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) { 8840 IsReversed = true ; // x CC y ? y : x 8841 } else { 8842 return SDValue(); 8843 } 8844 8845 bool IsUnordered; 8846 switch (CC) { 8847 default: break; 8848 case ISD::SETOLT: 8849 case ISD::SETOLE: 8850 case ISD::SETLT: 8851 case ISD::SETLE: 8852 case ISD::SETULT: 8853 case ISD::SETULE: 8854 // If LHS is NaN, an ordered comparison will be false and the result will 8855 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS 8856 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN. 8857 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE); 8858 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS)) 8859 break; 8860 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin 8861 // will return -0, so vmin can only be used for unsafe math or if one of 8862 // the operands is known to be nonzero. 8863 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) && 8864 !DAG.getTarget().Options.UnsafeFPMath && 8865 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) 8866 break; 8867 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN; 8868 break; 8869 8870 case ISD::SETOGT: 8871 case ISD::SETOGE: 8872 case ISD::SETGT: 8873 case ISD::SETGE: 8874 case ISD::SETUGT: 8875 case ISD::SETUGE: 8876 // If LHS is NaN, an ordered comparison will be false and the result will 8877 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS 8878 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN. 8879 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE); 8880 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS)) 8881 break; 8882 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax 8883 // will return +0, so vmax can only be used for unsafe math or if one of 8884 // the operands is known to be nonzero. 8885 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) && 8886 !DAG.getTarget().Options.UnsafeFPMath && 8887 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) 8888 break; 8889 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX; 8890 break; 8891 } 8892 8893 if (!Opcode) 8894 return SDValue(); 8895 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS); 8896 } 8897 8898 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV. 8899 SDValue 8900 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const { 8901 SDValue Cmp = N->getOperand(4); 8902 if (Cmp.getOpcode() != ARMISD::CMPZ) 8903 // Only looking at EQ and NE cases. 8904 return SDValue(); 8905 8906 EVT VT = N->getValueType(0); 8907 DebugLoc dl = N->getDebugLoc(); 8908 SDValue LHS = Cmp.getOperand(0); 8909 SDValue RHS = Cmp.getOperand(1); 8910 SDValue FalseVal = N->getOperand(0); 8911 SDValue TrueVal = N->getOperand(1); 8912 SDValue ARMcc = N->getOperand(2); 8913 ARMCC::CondCodes CC = 8914 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue(); 8915 8916 // Simplify 8917 // mov r1, r0 8918 // cmp r1, x 8919 // mov r0, y 8920 // moveq r0, x 8921 // to 8922 // cmp r0, x 8923 // movne r0, y 8924 // 8925 // mov r1, r0 8926 // cmp r1, x 8927 // mov r0, x 8928 // movne r0, y 8929 // to 8930 // cmp r0, x 8931 // movne r0, y 8932 /// FIXME: Turn this into a target neutral optimization? 8933 SDValue Res; 8934 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) { 8935 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc, 8936 N->getOperand(3), Cmp); 8937 } else if (CC == ARMCC::EQ && TrueVal == RHS) { 8938 SDValue ARMcc; 8939 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl); 8940 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc, 8941 N->getOperand(3), NewCmp); 8942 } 8943 8944 if (Res.getNode()) { 8945 APInt KnownZero, KnownOne; 8946 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne); 8947 // Capture demanded bits information that would be otherwise lost. 8948 if (KnownZero == 0xfffffffe) 8949 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res, 8950 DAG.getValueType(MVT::i1)); 8951 else if (KnownZero == 0xffffff00) 8952 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res, 8953 DAG.getValueType(MVT::i8)); 8954 else if (KnownZero == 0xffff0000) 8955 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res, 8956 DAG.getValueType(MVT::i16)); 8957 } 8958 8959 return Res; 8960 } 8961 8962 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N, 8963 DAGCombinerInfo &DCI) const { 8964 switch (N->getOpcode()) { 8965 default: break; 8966 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget); 8967 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget); 8968 case ISD::SUB: return PerformSUBCombine(N, DCI); 8969 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget); 8970 case ISD::OR: return PerformORCombine(N, DCI, Subtarget); 8971 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget); 8972 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget); 8973 case ARMISD::BFI: return PerformBFICombine(N, DCI); 8974 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI); 8975 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG); 8976 case ISD::STORE: return PerformSTORECombine(N, DCI); 8977 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI); 8978 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI); 8979 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG); 8980 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI); 8981 case ISD::FP_TO_SINT: 8982 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget); 8983 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget); 8984 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG); 8985 case ISD::SHL: 8986 case ISD::SRA: 8987 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget); 8988 case ISD::SIGN_EXTEND: 8989 case ISD::ZERO_EXTEND: 8990 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget); 8991 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget); 8992 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG); 8993 case ARMISD::VLD2DUP: 8994 case ARMISD::VLD3DUP: 8995 case ARMISD::VLD4DUP: 8996 return CombineBaseUpdate(N, DCI); 8997 case ISD::INTRINSIC_VOID: 8998 case ISD::INTRINSIC_W_CHAIN: 8999 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) { 9000 case Intrinsic::arm_neon_vld1: 9001 case Intrinsic::arm_neon_vld2: 9002 case Intrinsic::arm_neon_vld3: 9003 case Intrinsic::arm_neon_vld4: 9004 case Intrinsic::arm_neon_vld2lane: 9005 case Intrinsic::arm_neon_vld3lane: 9006 case Intrinsic::arm_neon_vld4lane: 9007 case Intrinsic::arm_neon_vst1: 9008 case Intrinsic::arm_neon_vst2: 9009 case Intrinsic::arm_neon_vst3: 9010 case Intrinsic::arm_neon_vst4: 9011 case Intrinsic::arm_neon_vst2lane: 9012 case Intrinsic::arm_neon_vst3lane: 9013 case Intrinsic::arm_neon_vst4lane: 9014 return CombineBaseUpdate(N, DCI); 9015 default: break; 9016 } 9017 break; 9018 } 9019 return SDValue(); 9020 } 9021 9022 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc, 9023 EVT VT) const { 9024 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE); 9025 } 9026 9027 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const { 9028 if (!Subtarget->allowsUnalignedMem()) 9029 return false; 9030 9031 switch (VT.getSimpleVT().SimpleTy) { 9032 default: 9033 return false; 9034 case MVT::i8: 9035 case MVT::i16: 9036 case MVT::i32: 9037 return true; 9038 case MVT::f64: 9039 return Subtarget->hasNEON(); 9040 // FIXME: VLD1 etc with standard alignment is legal. 9041 } 9042 } 9043 9044 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign, 9045 unsigned AlignCheck) { 9046 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) && 9047 (DstAlign == 0 || DstAlign % AlignCheck == 0)); 9048 } 9049 9050 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size, 9051 unsigned DstAlign, unsigned SrcAlign, 9052 bool IsZeroVal, 9053 bool MemcpyStrSrc, 9054 MachineFunction &MF) const { 9055 const Function *F = MF.getFunction(); 9056 9057 // See if we can use NEON instructions for this... 9058 if (IsZeroVal && 9059 !F->hasFnAttr(Attribute::NoImplicitFloat) && 9060 Subtarget->hasNEON()) { 9061 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) { 9062 return MVT::v4i32; 9063 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) { 9064 return MVT::v2i32; 9065 } 9066 } 9067 9068 // Lowering to i32/i16 if the size permits. 9069 if (Size >= 4) { 9070 return MVT::i32; 9071 } else if (Size >= 2) { 9072 return MVT::i16; 9073 } 9074 9075 // Let the target-independent logic figure it out. 9076 return MVT::Other; 9077 } 9078 9079 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) { 9080 if (V < 0) 9081 return false; 9082 9083 unsigned Scale = 1; 9084 switch (VT.getSimpleVT().SimpleTy) { 9085 default: return false; 9086 case MVT::i1: 9087 case MVT::i8: 9088 // Scale == 1; 9089 break; 9090 case MVT::i16: 9091 // Scale == 2; 9092 Scale = 2; 9093 break; 9094 case MVT::i32: 9095 // Scale == 4; 9096 Scale = 4; 9097 break; 9098 } 9099 9100 if ((V & (Scale - 1)) != 0) 9101 return false; 9102 V /= Scale; 9103 return V == (V & ((1LL << 5) - 1)); 9104 } 9105 9106 static bool isLegalT2AddressImmediate(int64_t V, EVT VT, 9107 const ARMSubtarget *Subtarget) { 9108 bool isNeg = false; 9109 if (V < 0) { 9110 isNeg = true; 9111 V = - V; 9112 } 9113 9114 switch (VT.getSimpleVT().SimpleTy) { 9115 default: return false; 9116 case MVT::i1: 9117 case MVT::i8: 9118 case MVT::i16: 9119 case MVT::i32: 9120 // + imm12 or - imm8 9121 if (isNeg) 9122 return V == (V & ((1LL << 8) - 1)); 9123 return V == (V & ((1LL << 12) - 1)); 9124 case MVT::f32: 9125 case MVT::f64: 9126 // Same as ARM mode. FIXME: NEON? 9127 if (!Subtarget->hasVFP2()) 9128 return false; 9129 if ((V & 3) != 0) 9130 return false; 9131 V >>= 2; 9132 return V == (V & ((1LL << 8) - 1)); 9133 } 9134 } 9135 9136 /// isLegalAddressImmediate - Return true if the integer value can be used 9137 /// as the offset of the target addressing mode for load / store of the 9138 /// given type. 9139 static bool isLegalAddressImmediate(int64_t V, EVT VT, 9140 const ARMSubtarget *Subtarget) { 9141 if (V == 0) 9142 return true; 9143 9144 if (!VT.isSimple()) 9145 return false; 9146 9147 if (Subtarget->isThumb1Only()) 9148 return isLegalT1AddressImmediate(V, VT); 9149 else if (Subtarget->isThumb2()) 9150 return isLegalT2AddressImmediate(V, VT, Subtarget); 9151 9152 // ARM mode. 9153 if (V < 0) 9154 V = - V; 9155 switch (VT.getSimpleVT().SimpleTy) { 9156 default: return false; 9157 case MVT::i1: 9158 case MVT::i8: 9159 case MVT::i32: 9160 // +- imm12 9161 return V == (V & ((1LL << 12) - 1)); 9162 case MVT::i16: 9163 // +- imm8 9164 return V == (V & ((1LL << 8) - 1)); 9165 case MVT::f32: 9166 case MVT::f64: 9167 if (!Subtarget->hasVFP2()) // FIXME: NEON? 9168 return false; 9169 if ((V & 3) != 0) 9170 return false; 9171 V >>= 2; 9172 return V == (V & ((1LL << 8) - 1)); 9173 } 9174 } 9175 9176 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM, 9177 EVT VT) const { 9178 int Scale = AM.Scale; 9179 if (Scale < 0) 9180 return false; 9181 9182 switch (VT.getSimpleVT().SimpleTy) { 9183 default: return false; 9184 case MVT::i1: 9185 case MVT::i8: 9186 case MVT::i16: 9187 case MVT::i32: 9188 if (Scale == 1) 9189 return true; 9190 // r + r << imm 9191 Scale = Scale & ~1; 9192 return Scale == 2 || Scale == 4 || Scale == 8; 9193 case MVT::i64: 9194 // r + r 9195 if (((unsigned)AM.HasBaseReg + Scale) <= 2) 9196 return true; 9197 return false; 9198 case MVT::isVoid: 9199 // Note, we allow "void" uses (basically, uses that aren't loads or 9200 // stores), because arm allows folding a scale into many arithmetic 9201 // operations. This should be made more precise and revisited later. 9202 9203 // Allow r << imm, but the imm has to be a multiple of two. 9204 if (Scale & 1) return false; 9205 return isPowerOf2_32(Scale); 9206 } 9207 } 9208 9209 /// isLegalAddressingMode - Return true if the addressing mode represented 9210 /// by AM is legal for this target, for a load/store of the specified type. 9211 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM, 9212 Type *Ty) const { 9213 EVT VT = getValueType(Ty, true); 9214 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget)) 9215 return false; 9216 9217 // Can never fold addr of global into load/store. 9218 if (AM.BaseGV) 9219 return false; 9220 9221 switch (AM.Scale) { 9222 case 0: // no scale reg, must be "r+i" or "r", or "i". 9223 break; 9224 case 1: 9225 if (Subtarget->isThumb1Only()) 9226 return false; 9227 // FALL THROUGH. 9228 default: 9229 // ARM doesn't support any R+R*scale+imm addr modes. 9230 if (AM.BaseOffs) 9231 return false; 9232 9233 if (!VT.isSimple()) 9234 return false; 9235 9236 if (Subtarget->isThumb2()) 9237 return isLegalT2ScaledAddressingMode(AM, VT); 9238 9239 int Scale = AM.Scale; 9240 switch (VT.getSimpleVT().SimpleTy) { 9241 default: return false; 9242 case MVT::i1: 9243 case MVT::i8: 9244 case MVT::i32: 9245 if (Scale < 0) Scale = -Scale; 9246 if (Scale == 1) 9247 return true; 9248 // r + r << imm 9249 return isPowerOf2_32(Scale & ~1); 9250 case MVT::i16: 9251 case MVT::i64: 9252 // r + r 9253 if (((unsigned)AM.HasBaseReg + Scale) <= 2) 9254 return true; 9255 return false; 9256 9257 case MVT::isVoid: 9258 // Note, we allow "void" uses (basically, uses that aren't loads or 9259 // stores), because arm allows folding a scale into many arithmetic 9260 // operations. This should be made more precise and revisited later. 9261 9262 // Allow r << imm, but the imm has to be a multiple of two. 9263 if (Scale & 1) return false; 9264 return isPowerOf2_32(Scale); 9265 } 9266 } 9267 return true; 9268 } 9269 9270 /// isLegalICmpImmediate - Return true if the specified immediate is legal 9271 /// icmp immediate, that is the target has icmp instructions which can compare 9272 /// a register against the immediate without having to materialize the 9273 /// immediate into a register. 9274 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const { 9275 // Thumb2 and ARM modes can use cmn for negative immediates. 9276 if (!Subtarget->isThumb()) 9277 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1; 9278 if (Subtarget->isThumb2()) 9279 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1; 9280 // Thumb1 doesn't have cmn, and only 8-bit immediates. 9281 return Imm >= 0 && Imm <= 255; 9282 } 9283 9284 /// isLegalAddImmediate - Return true if the specified immediate is a legal add 9285 /// *or sub* immediate, that is the target has add or sub instructions which can 9286 /// add a register with the immediate without having to materialize the 9287 /// immediate into a register. 9288 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const { 9289 // Same encoding for add/sub, just flip the sign. 9290 int64_t AbsImm = llvm::abs64(Imm); 9291 if (!Subtarget->isThumb()) 9292 return ARM_AM::getSOImmVal(AbsImm) != -1; 9293 if (Subtarget->isThumb2()) 9294 return ARM_AM::getT2SOImmVal(AbsImm) != -1; 9295 // Thumb1 only has 8-bit unsigned immediate. 9296 return AbsImm >= 0 && AbsImm <= 255; 9297 } 9298 9299 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT, 9300 bool isSEXTLoad, SDValue &Base, 9301 SDValue &Offset, bool &isInc, 9302 SelectionDAG &DAG) { 9303 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB) 9304 return false; 9305 9306 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) { 9307 // AddressingMode 3 9308 Base = Ptr->getOperand(0); 9309 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 9310 int RHSC = (int)RHS->getZExtValue(); 9311 if (RHSC < 0 && RHSC > -256) { 9312 assert(Ptr->getOpcode() == ISD::ADD); 9313 isInc = false; 9314 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 9315 return true; 9316 } 9317 } 9318 isInc = (Ptr->getOpcode() == ISD::ADD); 9319 Offset = Ptr->getOperand(1); 9320 return true; 9321 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) { 9322 // AddressingMode 2 9323 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 9324 int RHSC = (int)RHS->getZExtValue(); 9325 if (RHSC < 0 && RHSC > -0x1000) { 9326 assert(Ptr->getOpcode() == ISD::ADD); 9327 isInc = false; 9328 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 9329 Base = Ptr->getOperand(0); 9330 return true; 9331 } 9332 } 9333 9334 if (Ptr->getOpcode() == ISD::ADD) { 9335 isInc = true; 9336 ARM_AM::ShiftOpc ShOpcVal= 9337 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode()); 9338 if (ShOpcVal != ARM_AM::no_shift) { 9339 Base = Ptr->getOperand(1); 9340 Offset = Ptr->getOperand(0); 9341 } else { 9342 Base = Ptr->getOperand(0); 9343 Offset = Ptr->getOperand(1); 9344 } 9345 return true; 9346 } 9347 9348 isInc = (Ptr->getOpcode() == ISD::ADD); 9349 Base = Ptr->getOperand(0); 9350 Offset = Ptr->getOperand(1); 9351 return true; 9352 } 9353 9354 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store. 9355 return false; 9356 } 9357 9358 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT, 9359 bool isSEXTLoad, SDValue &Base, 9360 SDValue &Offset, bool &isInc, 9361 SelectionDAG &DAG) { 9362 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB) 9363 return false; 9364 9365 Base = Ptr->getOperand(0); 9366 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 9367 int RHSC = (int)RHS->getZExtValue(); 9368 if (RHSC < 0 && RHSC > -0x100) { // 8 bits. 9369 assert(Ptr->getOpcode() == ISD::ADD); 9370 isInc = false; 9371 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 9372 return true; 9373 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero. 9374 isInc = Ptr->getOpcode() == ISD::ADD; 9375 Offset = DAG.getConstant(RHSC, RHS->getValueType(0)); 9376 return true; 9377 } 9378 } 9379 9380 return false; 9381 } 9382 9383 /// getPreIndexedAddressParts - returns true by value, base pointer and 9384 /// offset pointer and addressing mode by reference if the node's address 9385 /// can be legally represented as pre-indexed load / store address. 9386 bool 9387 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base, 9388 SDValue &Offset, 9389 ISD::MemIndexedMode &AM, 9390 SelectionDAG &DAG) const { 9391 if (Subtarget->isThumb1Only()) 9392 return false; 9393 9394 EVT VT; 9395 SDValue Ptr; 9396 bool isSEXTLoad = false; 9397 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 9398 Ptr = LD->getBasePtr(); 9399 VT = LD->getMemoryVT(); 9400 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD; 9401 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 9402 Ptr = ST->getBasePtr(); 9403 VT = ST->getMemoryVT(); 9404 } else 9405 return false; 9406 9407 bool isInc; 9408 bool isLegal = false; 9409 if (Subtarget->isThumb2()) 9410 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base, 9411 Offset, isInc, DAG); 9412 else 9413 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base, 9414 Offset, isInc, DAG); 9415 if (!isLegal) 9416 return false; 9417 9418 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC; 9419 return true; 9420 } 9421 9422 /// getPostIndexedAddressParts - returns true by value, base pointer and 9423 /// offset pointer and addressing mode by reference if this node can be 9424 /// combined with a load / store to form a post-indexed load / store. 9425 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op, 9426 SDValue &Base, 9427 SDValue &Offset, 9428 ISD::MemIndexedMode &AM, 9429 SelectionDAG &DAG) const { 9430 if (Subtarget->isThumb1Only()) 9431 return false; 9432 9433 EVT VT; 9434 SDValue Ptr; 9435 bool isSEXTLoad = false; 9436 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 9437 VT = LD->getMemoryVT(); 9438 Ptr = LD->getBasePtr(); 9439 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD; 9440 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 9441 VT = ST->getMemoryVT(); 9442 Ptr = ST->getBasePtr(); 9443 } else 9444 return false; 9445 9446 bool isInc; 9447 bool isLegal = false; 9448 if (Subtarget->isThumb2()) 9449 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset, 9450 isInc, DAG); 9451 else 9452 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset, 9453 isInc, DAG); 9454 if (!isLegal) 9455 return false; 9456 9457 if (Ptr != Base) { 9458 // Swap base ptr and offset to catch more post-index load / store when 9459 // it's legal. In Thumb2 mode, offset must be an immediate. 9460 if (Ptr == Offset && Op->getOpcode() == ISD::ADD && 9461 !Subtarget->isThumb2()) 9462 std::swap(Base, Offset); 9463 9464 // Post-indexed load / store update the base pointer. 9465 if (Ptr != Base) 9466 return false; 9467 } 9468 9469 AM = isInc ? ISD::POST_INC : ISD::POST_DEC; 9470 return true; 9471 } 9472 9473 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 9474 APInt &KnownZero, 9475 APInt &KnownOne, 9476 const SelectionDAG &DAG, 9477 unsigned Depth) const { 9478 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); 9479 switch (Op.getOpcode()) { 9480 default: break; 9481 case ARMISD::CMOV: { 9482 // Bits are known zero/one if known on the LHS and RHS. 9483 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1); 9484 if (KnownZero == 0 && KnownOne == 0) return; 9485 9486 APInt KnownZeroRHS, KnownOneRHS; 9487 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1); 9488 KnownZero &= KnownZeroRHS; 9489 KnownOne &= KnownOneRHS; 9490 return; 9491 } 9492 } 9493 } 9494 9495 //===----------------------------------------------------------------------===// 9496 // ARM Inline Assembly Support 9497 //===----------------------------------------------------------------------===// 9498 9499 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const { 9500 // Looking for "rev" which is V6+. 9501 if (!Subtarget->hasV6Ops()) 9502 return false; 9503 9504 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue()); 9505 std::string AsmStr = IA->getAsmString(); 9506 SmallVector<StringRef, 4> AsmPieces; 9507 SplitString(AsmStr, AsmPieces, ";\n"); 9508 9509 switch (AsmPieces.size()) { 9510 default: return false; 9511 case 1: 9512 AsmStr = AsmPieces[0]; 9513 AsmPieces.clear(); 9514 SplitString(AsmStr, AsmPieces, " \t,"); 9515 9516 // rev $0, $1 9517 if (AsmPieces.size() == 3 && 9518 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" && 9519 IA->getConstraintString().compare(0, 4, "=l,l") == 0) { 9520 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType()); 9521 if (Ty && Ty->getBitWidth() == 32) 9522 return IntrinsicLowering::LowerToByteSwap(CI); 9523 } 9524 break; 9525 } 9526 9527 return false; 9528 } 9529 9530 /// getConstraintType - Given a constraint letter, return the type of 9531 /// constraint it is for this target. 9532 ARMTargetLowering::ConstraintType 9533 ARMTargetLowering::getConstraintType(const std::string &Constraint) const { 9534 if (Constraint.size() == 1) { 9535 switch (Constraint[0]) { 9536 default: break; 9537 case 'l': return C_RegisterClass; 9538 case 'w': return C_RegisterClass; 9539 case 'h': return C_RegisterClass; 9540 case 'x': return C_RegisterClass; 9541 case 't': return C_RegisterClass; 9542 case 'j': return C_Other; // Constant for movw. 9543 // An address with a single base register. Due to the way we 9544 // currently handle addresses it is the same as an 'r' memory constraint. 9545 case 'Q': return C_Memory; 9546 } 9547 } else if (Constraint.size() == 2) { 9548 switch (Constraint[0]) { 9549 default: break; 9550 // All 'U+' constraints are addresses. 9551 case 'U': return C_Memory; 9552 } 9553 } 9554 return TargetLowering::getConstraintType(Constraint); 9555 } 9556 9557 /// Examine constraint type and operand type and determine a weight value. 9558 /// This object must already have been set up with the operand type 9559 /// and the current alternative constraint selected. 9560 TargetLowering::ConstraintWeight 9561 ARMTargetLowering::getSingleConstraintMatchWeight( 9562 AsmOperandInfo &info, const char *constraint) const { 9563 ConstraintWeight weight = CW_Invalid; 9564 Value *CallOperandVal = info.CallOperandVal; 9565 // If we don't have a value, we can't do a match, 9566 // but allow it at the lowest weight. 9567 if (CallOperandVal == NULL) 9568 return CW_Default; 9569 Type *type = CallOperandVal->getType(); 9570 // Look at the constraint type. 9571 switch (*constraint) { 9572 default: 9573 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); 9574 break; 9575 case 'l': 9576 if (type->isIntegerTy()) { 9577 if (Subtarget->isThumb()) 9578 weight = CW_SpecificReg; 9579 else 9580 weight = CW_Register; 9581 } 9582 break; 9583 case 'w': 9584 if (type->isFloatingPointTy()) 9585 weight = CW_Register; 9586 break; 9587 } 9588 return weight; 9589 } 9590 9591 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair; 9592 RCPair 9593 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint, 9594 EVT VT) const { 9595 if (Constraint.size() == 1) { 9596 // GCC ARM Constraint Letters 9597 switch (Constraint[0]) { 9598 case 'l': // Low regs or general regs. 9599 if (Subtarget->isThumb()) 9600 return RCPair(0U, &ARM::tGPRRegClass); 9601 return RCPair(0U, &ARM::GPRRegClass); 9602 case 'h': // High regs or no regs. 9603 if (Subtarget->isThumb()) 9604 return RCPair(0U, &ARM::hGPRRegClass); 9605 break; 9606 case 'r': 9607 return RCPair(0U, &ARM::GPRRegClass); 9608 case 'w': 9609 if (VT == MVT::f32) 9610 return RCPair(0U, &ARM::SPRRegClass); 9611 if (VT.getSizeInBits() == 64) 9612 return RCPair(0U, &ARM::DPRRegClass); 9613 if (VT.getSizeInBits() == 128) 9614 return RCPair(0U, &ARM::QPRRegClass); 9615 break; 9616 case 'x': 9617 if (VT == MVT::f32) 9618 return RCPair(0U, &ARM::SPR_8RegClass); 9619 if (VT.getSizeInBits() == 64) 9620 return RCPair(0U, &ARM::DPR_8RegClass); 9621 if (VT.getSizeInBits() == 128) 9622 return RCPair(0U, &ARM::QPR_8RegClass); 9623 break; 9624 case 't': 9625 if (VT == MVT::f32) 9626 return RCPair(0U, &ARM::SPRRegClass); 9627 break; 9628 } 9629 } 9630 if (StringRef("{cc}").equals_lower(Constraint)) 9631 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass); 9632 9633 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 9634 } 9635 9636 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 9637 /// vector. If it is invalid, don't add anything to Ops. 9638 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op, 9639 std::string &Constraint, 9640 std::vector<SDValue>&Ops, 9641 SelectionDAG &DAG) const { 9642 SDValue Result(0, 0); 9643 9644 // Currently only support length 1 constraints. 9645 if (Constraint.length() != 1) return; 9646 9647 char ConstraintLetter = Constraint[0]; 9648 switch (ConstraintLetter) { 9649 default: break; 9650 case 'j': 9651 case 'I': case 'J': case 'K': case 'L': 9652 case 'M': case 'N': case 'O': 9653 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 9654 if (!C) 9655 return; 9656 9657 int64_t CVal64 = C->getSExtValue(); 9658 int CVal = (int) CVal64; 9659 // None of these constraints allow values larger than 32 bits. Check 9660 // that the value fits in an int. 9661 if (CVal != CVal64) 9662 return; 9663 9664 switch (ConstraintLetter) { 9665 case 'j': 9666 // Constant suitable for movw, must be between 0 and 9667 // 65535. 9668 if (Subtarget->hasV6T2Ops()) 9669 if (CVal >= 0 && CVal <= 65535) 9670 break; 9671 return; 9672 case 'I': 9673 if (Subtarget->isThumb1Only()) { 9674 // This must be a constant between 0 and 255, for ADD 9675 // immediates. 9676 if (CVal >= 0 && CVal <= 255) 9677 break; 9678 } else if (Subtarget->isThumb2()) { 9679 // A constant that can be used as an immediate value in a 9680 // data-processing instruction. 9681 if (ARM_AM::getT2SOImmVal(CVal) != -1) 9682 break; 9683 } else { 9684 // A constant that can be used as an immediate value in a 9685 // data-processing instruction. 9686 if (ARM_AM::getSOImmVal(CVal) != -1) 9687 break; 9688 } 9689 return; 9690 9691 case 'J': 9692 if (Subtarget->isThumb()) { // FIXME thumb2 9693 // This must be a constant between -255 and -1, for negated ADD 9694 // immediates. This can be used in GCC with an "n" modifier that 9695 // prints the negated value, for use with SUB instructions. It is 9696 // not useful otherwise but is implemented for compatibility. 9697 if (CVal >= -255 && CVal <= -1) 9698 break; 9699 } else { 9700 // This must be a constant between -4095 and 4095. It is not clear 9701 // what this constraint is intended for. Implemented for 9702 // compatibility with GCC. 9703 if (CVal >= -4095 && CVal <= 4095) 9704 break; 9705 } 9706 return; 9707 9708 case 'K': 9709 if (Subtarget->isThumb1Only()) { 9710 // A 32-bit value where only one byte has a nonzero value. Exclude 9711 // zero to match GCC. This constraint is used by GCC internally for 9712 // constants that can be loaded with a move/shift combination. 9713 // It is not useful otherwise but is implemented for compatibility. 9714 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal)) 9715 break; 9716 } else if (Subtarget->isThumb2()) { 9717 // A constant whose bitwise inverse can be used as an immediate 9718 // value in a data-processing instruction. This can be used in GCC 9719 // with a "B" modifier that prints the inverted value, for use with 9720 // BIC and MVN instructions. It is not useful otherwise but is 9721 // implemented for compatibility. 9722 if (ARM_AM::getT2SOImmVal(~CVal) != -1) 9723 break; 9724 } else { 9725 // A constant whose bitwise inverse can be used as an immediate 9726 // value in a data-processing instruction. This can be used in GCC 9727 // with a "B" modifier that prints the inverted value, for use with 9728 // BIC and MVN instructions. It is not useful otherwise but is 9729 // implemented for compatibility. 9730 if (ARM_AM::getSOImmVal(~CVal) != -1) 9731 break; 9732 } 9733 return; 9734 9735 case 'L': 9736 if (Subtarget->isThumb1Only()) { 9737 // This must be a constant between -7 and 7, 9738 // for 3-operand ADD/SUB immediate instructions. 9739 if (CVal >= -7 && CVal < 7) 9740 break; 9741 } else if (Subtarget->isThumb2()) { 9742 // A constant whose negation can be used as an immediate value in a 9743 // data-processing instruction. This can be used in GCC with an "n" 9744 // modifier that prints the negated value, for use with SUB 9745 // instructions. It is not useful otherwise but is implemented for 9746 // compatibility. 9747 if (ARM_AM::getT2SOImmVal(-CVal) != -1) 9748 break; 9749 } else { 9750 // A constant whose negation can be used as an immediate value in a 9751 // data-processing instruction. This can be used in GCC with an "n" 9752 // modifier that prints the negated value, for use with SUB 9753 // instructions. It is not useful otherwise but is implemented for 9754 // compatibility. 9755 if (ARM_AM::getSOImmVal(-CVal) != -1) 9756 break; 9757 } 9758 return; 9759 9760 case 'M': 9761 if (Subtarget->isThumb()) { // FIXME thumb2 9762 // This must be a multiple of 4 between 0 and 1020, for 9763 // ADD sp + immediate. 9764 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0)) 9765 break; 9766 } else { 9767 // A power of two or a constant between 0 and 32. This is used in 9768 // GCC for the shift amount on shifted register operands, but it is 9769 // useful in general for any shift amounts. 9770 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0)) 9771 break; 9772 } 9773 return; 9774 9775 case 'N': 9776 if (Subtarget->isThumb()) { // FIXME thumb2 9777 // This must be a constant between 0 and 31, for shift amounts. 9778 if (CVal >= 0 && CVal <= 31) 9779 break; 9780 } 9781 return; 9782 9783 case 'O': 9784 if (Subtarget->isThumb()) { // FIXME thumb2 9785 // This must be a multiple of 4 between -508 and 508, for 9786 // ADD/SUB sp = sp + immediate. 9787 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0)) 9788 break; 9789 } 9790 return; 9791 } 9792 Result = DAG.getTargetConstant(CVal, Op.getValueType()); 9793 break; 9794 } 9795 9796 if (Result.getNode()) { 9797 Ops.push_back(Result); 9798 return; 9799 } 9800 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); 9801 } 9802 9803 bool 9804 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 9805 // The ARM target isn't yet aware of offsets. 9806 return false; 9807 } 9808 9809 bool ARM::isBitFieldInvertedMask(unsigned v) { 9810 if (v == 0xffffffff) 9811 return 0; 9812 // there can be 1's on either or both "outsides", all the "inside" 9813 // bits must be 0's 9814 unsigned int lsb = 0, msb = 31; 9815 while (v & (1 << msb)) --msb; 9816 while (v & (1 << lsb)) ++lsb; 9817 for (unsigned int i = lsb; i <= msb; ++i) { 9818 if (v & (1 << i)) 9819 return 0; 9820 } 9821 return 1; 9822 } 9823 9824 /// isFPImmLegal - Returns true if the target can instruction select the 9825 /// specified FP immediate natively. If false, the legalizer will 9826 /// materialize the FP immediate as a load from a constant pool. 9827 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const { 9828 if (!Subtarget->hasVFP3()) 9829 return false; 9830 if (VT == MVT::f32) 9831 return ARM_AM::getFP32Imm(Imm) != -1; 9832 if (VT == MVT::f64) 9833 return ARM_AM::getFP64Imm(Imm) != -1; 9834 return false; 9835 } 9836 9837 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as 9838 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment 9839 /// specified in the intrinsic calls. 9840 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, 9841 const CallInst &I, 9842 unsigned Intrinsic) const { 9843 switch (Intrinsic) { 9844 case Intrinsic::arm_neon_vld1: 9845 case Intrinsic::arm_neon_vld2: 9846 case Intrinsic::arm_neon_vld3: 9847 case Intrinsic::arm_neon_vld4: 9848 case Intrinsic::arm_neon_vld2lane: 9849 case Intrinsic::arm_neon_vld3lane: 9850 case Intrinsic::arm_neon_vld4lane: { 9851 Info.opc = ISD::INTRINSIC_W_CHAIN; 9852 // Conservatively set memVT to the entire set of vectors loaded. 9853 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8; 9854 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts); 9855 Info.ptrVal = I.getArgOperand(0); 9856 Info.offset = 0; 9857 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1); 9858 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue(); 9859 Info.vol = false; // volatile loads with NEON intrinsics not supported 9860 Info.readMem = true; 9861 Info.writeMem = false; 9862 return true; 9863 } 9864 case Intrinsic::arm_neon_vst1: 9865 case Intrinsic::arm_neon_vst2: 9866 case Intrinsic::arm_neon_vst3: 9867 case Intrinsic::arm_neon_vst4: 9868 case Intrinsic::arm_neon_vst2lane: 9869 case Intrinsic::arm_neon_vst3lane: 9870 case Intrinsic::arm_neon_vst4lane: { 9871 Info.opc = ISD::INTRINSIC_VOID; 9872 // Conservatively set memVT to the entire set of vectors stored. 9873 unsigned NumElts = 0; 9874 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) { 9875 Type *ArgTy = I.getArgOperand(ArgI)->getType(); 9876 if (!ArgTy->isVectorTy()) 9877 break; 9878 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8; 9879 } 9880 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts); 9881 Info.ptrVal = I.getArgOperand(0); 9882 Info.offset = 0; 9883 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1); 9884 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue(); 9885 Info.vol = false; // volatile stores with NEON intrinsics not supported 9886 Info.readMem = false; 9887 Info.writeMem = true; 9888 return true; 9889 } 9890 case Intrinsic::arm_strexd: { 9891 Info.opc = ISD::INTRINSIC_W_CHAIN; 9892 Info.memVT = MVT::i64; 9893 Info.ptrVal = I.getArgOperand(2); 9894 Info.offset = 0; 9895 Info.align = 8; 9896 Info.vol = true; 9897 Info.readMem = false; 9898 Info.writeMem = true; 9899 return true; 9900 } 9901 case Intrinsic::arm_ldrexd: { 9902 Info.opc = ISD::INTRINSIC_W_CHAIN; 9903 Info.memVT = MVT::i64; 9904 Info.ptrVal = I.getArgOperand(0); 9905 Info.offset = 0; 9906 Info.align = 8; 9907 Info.vol = true; 9908 Info.readMem = true; 9909 Info.writeMem = false; 9910 return true; 9911 } 9912 default: 9913 break; 9914 } 9915 9916 return false; 9917 } 9918