1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file defines the interfaces that ARM uses to lower LLVM code into a 11 // selection DAG. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #define DEBUG_TYPE "arm-isel" 16 #include "ARMISelLowering.h" 17 #include "ARM.h" 18 #include "ARMCallingConv.h" 19 #include "ARMConstantPoolValue.h" 20 #include "ARMMachineFunctionInfo.h" 21 #include "ARMPerfectShuffle.h" 22 #include "ARMSubtarget.h" 23 #include "ARMTargetMachine.h" 24 #include "ARMTargetObjectFile.h" 25 #include "MCTargetDesc/ARMAddressingModes.h" 26 #include "llvm/ADT/Statistic.h" 27 #include "llvm/ADT/StringExtras.h" 28 #include "llvm/CodeGen/CallingConvLower.h" 29 #include "llvm/CodeGen/IntrinsicLowering.h" 30 #include "llvm/CodeGen/MachineBasicBlock.h" 31 #include "llvm/CodeGen/MachineFrameInfo.h" 32 #include "llvm/CodeGen/MachineFunction.h" 33 #include "llvm/CodeGen/MachineInstrBuilder.h" 34 #include "llvm/CodeGen/MachineModuleInfo.h" 35 #include "llvm/CodeGen/MachineRegisterInfo.h" 36 #include "llvm/CodeGen/SelectionDAG.h" 37 #include "llvm/IR/CallingConv.h" 38 #include "llvm/IR/Constants.h" 39 #include "llvm/IR/Function.h" 40 #include "llvm/IR/GlobalValue.h" 41 #include "llvm/IR/Instruction.h" 42 #include "llvm/IR/Instructions.h" 43 #include "llvm/IR/Intrinsics.h" 44 #include "llvm/IR/Type.h" 45 #include "llvm/MC/MCSectionMachO.h" 46 #include "llvm/Support/CommandLine.h" 47 #include "llvm/Support/ErrorHandling.h" 48 #include "llvm/Support/MathExtras.h" 49 #include "llvm/Support/raw_ostream.h" 50 #include "llvm/Target/TargetOptions.h" 51 using namespace llvm; 52 53 STATISTIC(NumTailCalls, "Number of tail calls"); 54 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt"); 55 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments"); 56 57 // This option should go away when tail calls fully work. 58 static cl::opt<bool> 59 EnableARMTailCalls("arm-tail-calls", cl::Hidden, 60 cl::desc("Generate tail calls (TEMPORARY OPTION)."), 61 cl::init(false)); 62 63 cl::opt<bool> 64 EnableARMLongCalls("arm-long-calls", cl::Hidden, 65 cl::desc("Generate calls via indirect call instructions"), 66 cl::init(false)); 67 68 static cl::opt<bool> 69 ARMInterworking("arm-interworking", cl::Hidden, 70 cl::desc("Enable / disable ARM interworking (for debugging only)"), 71 cl::init(true)); 72 73 namespace { 74 class ARMCCState : public CCState { 75 public: 76 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF, 77 const TargetMachine &TM, SmallVectorImpl<CCValAssign> &locs, 78 LLVMContext &C, ParmContext PC) 79 : CCState(CC, isVarArg, MF, TM, locs, C) { 80 assert(((PC == Call) || (PC == Prologue)) && 81 "ARMCCState users must specify whether their context is call" 82 "or prologue generation."); 83 CallOrPrologue = PC; 84 } 85 }; 86 } 87 88 // The APCS parameter registers. 89 static const uint16_t GPRArgRegs[] = { 90 ARM::R0, ARM::R1, ARM::R2, ARM::R3 91 }; 92 93 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT, 94 MVT PromotedBitwiseVT) { 95 if (VT != PromotedLdStVT) { 96 setOperationAction(ISD::LOAD, VT, Promote); 97 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT); 98 99 setOperationAction(ISD::STORE, VT, Promote); 100 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT); 101 } 102 103 MVT ElemTy = VT.getVectorElementType(); 104 if (ElemTy != MVT::i64 && ElemTy != MVT::f64) 105 setOperationAction(ISD::SETCC, VT, Custom); 106 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); 107 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); 108 if (ElemTy == MVT::i32) { 109 setOperationAction(ISD::SINT_TO_FP, VT, Custom); 110 setOperationAction(ISD::UINT_TO_FP, VT, Custom); 111 setOperationAction(ISD::FP_TO_SINT, VT, Custom); 112 setOperationAction(ISD::FP_TO_UINT, VT, Custom); 113 } else { 114 setOperationAction(ISD::SINT_TO_FP, VT, Expand); 115 setOperationAction(ISD::UINT_TO_FP, VT, Expand); 116 setOperationAction(ISD::FP_TO_SINT, VT, Expand); 117 setOperationAction(ISD::FP_TO_UINT, VT, Expand); 118 } 119 setOperationAction(ISD::BUILD_VECTOR, VT, Custom); 120 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); 121 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal); 122 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal); 123 setOperationAction(ISD::SELECT, VT, Expand); 124 setOperationAction(ISD::SELECT_CC, VT, Expand); 125 setOperationAction(ISD::VSELECT, VT, Expand); 126 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand); 127 if (VT.isInteger()) { 128 setOperationAction(ISD::SHL, VT, Custom); 129 setOperationAction(ISD::SRA, VT, Custom); 130 setOperationAction(ISD::SRL, VT, Custom); 131 } 132 133 // Promote all bit-wise operations. 134 if (VT.isInteger() && VT != PromotedBitwiseVT) { 135 setOperationAction(ISD::AND, VT, Promote); 136 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT); 137 setOperationAction(ISD::OR, VT, Promote); 138 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT); 139 setOperationAction(ISD::XOR, VT, Promote); 140 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT); 141 } 142 143 // Neon does not support vector divide/remainder operations. 144 setOperationAction(ISD::SDIV, VT, Expand); 145 setOperationAction(ISD::UDIV, VT, Expand); 146 setOperationAction(ISD::FDIV, VT, Expand); 147 setOperationAction(ISD::SREM, VT, Expand); 148 setOperationAction(ISD::UREM, VT, Expand); 149 setOperationAction(ISD::FREM, VT, Expand); 150 } 151 152 void ARMTargetLowering::addDRTypeForNEON(MVT VT) { 153 addRegisterClass(VT, &ARM::DPRRegClass); 154 addTypeForNEON(VT, MVT::f64, MVT::v2i32); 155 } 156 157 void ARMTargetLowering::addQRTypeForNEON(MVT VT) { 158 addRegisterClass(VT, &ARM::QPRRegClass); 159 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32); 160 } 161 162 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) { 163 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin()) 164 return new TargetLoweringObjectFileMachO(); 165 166 return new ARMElfTargetObjectFile(); 167 } 168 169 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) 170 : TargetLowering(TM, createTLOF(TM)) { 171 Subtarget = &TM.getSubtarget<ARMSubtarget>(); 172 RegInfo = TM.getRegisterInfo(); 173 Itins = TM.getInstrItineraryData(); 174 175 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); 176 177 if (Subtarget->isTargetDarwin()) { 178 // Uses VFP for Thumb libfuncs if available. 179 if (Subtarget->isThumb() && Subtarget->hasVFP2()) { 180 // Single-precision floating-point arithmetic. 181 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp"); 182 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp"); 183 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp"); 184 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp"); 185 186 // Double-precision floating-point arithmetic. 187 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp"); 188 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp"); 189 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp"); 190 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp"); 191 192 // Single-precision comparisons. 193 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp"); 194 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp"); 195 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp"); 196 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp"); 197 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp"); 198 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp"); 199 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp"); 200 setLibcallName(RTLIB::O_F32, "__unordsf2vfp"); 201 202 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE); 203 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE); 204 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE); 205 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE); 206 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE); 207 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE); 208 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE); 209 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ); 210 211 // Double-precision comparisons. 212 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp"); 213 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp"); 214 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp"); 215 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp"); 216 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp"); 217 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp"); 218 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp"); 219 setLibcallName(RTLIB::O_F64, "__unorddf2vfp"); 220 221 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE); 222 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE); 223 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE); 224 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE); 225 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE); 226 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE); 227 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE); 228 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ); 229 230 // Floating-point to integer conversions. 231 // i64 conversions are done via library routines even when generating VFP 232 // instructions, so use the same ones. 233 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp"); 234 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp"); 235 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp"); 236 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp"); 237 238 // Conversions between floating types. 239 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp"); 240 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp"); 241 242 // Integer to floating-point conversions. 243 // i64 conversions are done via library routines even when generating VFP 244 // instructions, so use the same ones. 245 // FIXME: There appears to be some naming inconsistency in ARM libgcc: 246 // e.g., __floatunsidf vs. __floatunssidfvfp. 247 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp"); 248 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp"); 249 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp"); 250 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp"); 251 } 252 } 253 254 // These libcalls are not available in 32-bit. 255 setLibcallName(RTLIB::SHL_I128, 0); 256 setLibcallName(RTLIB::SRL_I128, 0); 257 setLibcallName(RTLIB::SRA_I128, 0); 258 259 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) { 260 // Double-precision floating-point arithmetic helper functions 261 // RTABI chapter 4.1.2, Table 2 262 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd"); 263 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv"); 264 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul"); 265 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub"); 266 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS); 267 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS); 268 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS); 269 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS); 270 271 // Double-precision floating-point comparison helper functions 272 // RTABI chapter 4.1.2, Table 3 273 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq"); 274 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE); 275 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq"); 276 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ); 277 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt"); 278 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE); 279 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple"); 280 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE); 281 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge"); 282 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE); 283 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt"); 284 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE); 285 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun"); 286 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE); 287 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun"); 288 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ); 289 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS); 290 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS); 291 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS); 292 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS); 293 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS); 294 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS); 295 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS); 296 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS); 297 298 // Single-precision floating-point arithmetic helper functions 299 // RTABI chapter 4.1.2, Table 4 300 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd"); 301 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv"); 302 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul"); 303 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub"); 304 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS); 305 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS); 306 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS); 307 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS); 308 309 // Single-precision floating-point comparison helper functions 310 // RTABI chapter 4.1.2, Table 5 311 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq"); 312 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE); 313 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq"); 314 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ); 315 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt"); 316 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE); 317 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple"); 318 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE); 319 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge"); 320 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE); 321 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt"); 322 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE); 323 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun"); 324 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE); 325 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun"); 326 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ); 327 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS); 328 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS); 329 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS); 330 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS); 331 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS); 332 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS); 333 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS); 334 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS); 335 336 // Floating-point to integer conversions. 337 // RTABI chapter 4.1.2, Table 6 338 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz"); 339 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz"); 340 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz"); 341 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz"); 342 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz"); 343 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz"); 344 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz"); 345 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz"); 346 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS); 347 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS); 348 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS); 349 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS); 350 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS); 351 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS); 352 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS); 353 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS); 354 355 // Conversions between floating types. 356 // RTABI chapter 4.1.2, Table 7 357 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f"); 358 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d"); 359 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS); 360 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS); 361 362 // Integer to floating-point conversions. 363 // RTABI chapter 4.1.2, Table 8 364 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d"); 365 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d"); 366 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d"); 367 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d"); 368 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f"); 369 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f"); 370 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f"); 371 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f"); 372 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS); 373 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS); 374 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS); 375 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS); 376 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS); 377 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS); 378 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS); 379 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS); 380 381 // Long long helper functions 382 // RTABI chapter 4.2, Table 9 383 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul"); 384 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl"); 385 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr"); 386 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr"); 387 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS); 388 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS); 389 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS); 390 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS); 391 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS); 392 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS); 393 394 // Integer division functions 395 // RTABI chapter 4.3.1 396 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv"); 397 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv"); 398 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv"); 399 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod"); 400 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv"); 401 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv"); 402 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv"); 403 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod"); 404 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS); 405 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS); 406 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS); 407 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS); 408 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS); 409 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS); 410 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS); 411 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS); 412 413 // Memory operations 414 // RTABI chapter 4.3.4 415 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy"); 416 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove"); 417 setLibcallName(RTLIB::MEMSET, "__aeabi_memset"); 418 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS); 419 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS); 420 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS); 421 } 422 423 // Use divmod compiler-rt calls for iOS 5.0 and later. 424 if (Subtarget->getTargetTriple().getOS() == Triple::IOS && 425 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) { 426 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4"); 427 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4"); 428 } 429 430 if (Subtarget->isThumb1Only()) 431 addRegisterClass(MVT::i32, &ARM::tGPRRegClass); 432 else 433 addRegisterClass(MVT::i32, &ARM::GPRRegClass); 434 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() && 435 !Subtarget->isThumb1Only()) { 436 addRegisterClass(MVT::f32, &ARM::SPRRegClass); 437 if (!Subtarget->isFPOnlySP()) 438 addRegisterClass(MVT::f64, &ARM::DPRRegClass); 439 440 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 441 } 442 443 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE; 444 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) { 445 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE; 446 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT) 447 setTruncStoreAction((MVT::SimpleValueType)VT, 448 (MVT::SimpleValueType)InnerVT, Expand); 449 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand); 450 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand); 451 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand); 452 } 453 454 setOperationAction(ISD::ConstantFP, MVT::f32, Custom); 455 456 if (Subtarget->hasNEON()) { 457 addDRTypeForNEON(MVT::v2f32); 458 addDRTypeForNEON(MVT::v8i8); 459 addDRTypeForNEON(MVT::v4i16); 460 addDRTypeForNEON(MVT::v2i32); 461 addDRTypeForNEON(MVT::v1i64); 462 463 addQRTypeForNEON(MVT::v4f32); 464 addQRTypeForNEON(MVT::v2f64); 465 addQRTypeForNEON(MVT::v16i8); 466 addQRTypeForNEON(MVT::v8i16); 467 addQRTypeForNEON(MVT::v4i32); 468 addQRTypeForNEON(MVT::v2i64); 469 470 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but 471 // neither Neon nor VFP support any arithmetic operations on it. 472 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively 473 // supported for v4f32. 474 setOperationAction(ISD::FADD, MVT::v2f64, Expand); 475 setOperationAction(ISD::FSUB, MVT::v2f64, Expand); 476 setOperationAction(ISD::FMUL, MVT::v2f64, Expand); 477 // FIXME: Code duplication: FDIV and FREM are expanded always, see 478 // ARMTargetLowering::addTypeForNEON method for details. 479 setOperationAction(ISD::FDIV, MVT::v2f64, Expand); 480 setOperationAction(ISD::FREM, MVT::v2f64, Expand); 481 // FIXME: Create unittest. 482 // In another words, find a way when "copysign" appears in DAG with vector 483 // operands. 484 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand); 485 // FIXME: Code duplication: SETCC has custom operation action, see 486 // ARMTargetLowering::addTypeForNEON method for details. 487 setOperationAction(ISD::SETCC, MVT::v2f64, Expand); 488 // FIXME: Create unittest for FNEG and for FABS. 489 setOperationAction(ISD::FNEG, MVT::v2f64, Expand); 490 setOperationAction(ISD::FABS, MVT::v2f64, Expand); 491 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand); 492 setOperationAction(ISD::FSIN, MVT::v2f64, Expand); 493 setOperationAction(ISD::FCOS, MVT::v2f64, Expand); 494 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand); 495 setOperationAction(ISD::FPOW, MVT::v2f64, Expand); 496 setOperationAction(ISD::FLOG, MVT::v2f64, Expand); 497 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand); 498 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand); 499 setOperationAction(ISD::FEXP, MVT::v2f64, Expand); 500 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand); 501 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR. 502 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand); 503 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand); 504 setOperationAction(ISD::FRINT, MVT::v2f64, Expand); 505 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand); 506 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand); 507 setOperationAction(ISD::FMA, MVT::v2f64, Expand); 508 509 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand); 510 setOperationAction(ISD::FSIN, MVT::v4f32, Expand); 511 setOperationAction(ISD::FCOS, MVT::v4f32, Expand); 512 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand); 513 setOperationAction(ISD::FPOW, MVT::v4f32, Expand); 514 setOperationAction(ISD::FLOG, MVT::v4f32, Expand); 515 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand); 516 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand); 517 setOperationAction(ISD::FEXP, MVT::v4f32, Expand); 518 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand); 519 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand); 520 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand); 521 setOperationAction(ISD::FRINT, MVT::v4f32, Expand); 522 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand); 523 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand); 524 525 // Mark v2f32 intrinsics. 526 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand); 527 setOperationAction(ISD::FSIN, MVT::v2f32, Expand); 528 setOperationAction(ISD::FCOS, MVT::v2f32, Expand); 529 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand); 530 setOperationAction(ISD::FPOW, MVT::v2f32, Expand); 531 setOperationAction(ISD::FLOG, MVT::v2f32, Expand); 532 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand); 533 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand); 534 setOperationAction(ISD::FEXP, MVT::v2f32, Expand); 535 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand); 536 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand); 537 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand); 538 setOperationAction(ISD::FRINT, MVT::v2f32, Expand); 539 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand); 540 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand); 541 542 // Neon does not support some operations on v1i64 and v2i64 types. 543 setOperationAction(ISD::MUL, MVT::v1i64, Expand); 544 // Custom handling for some quad-vector types to detect VMULL. 545 setOperationAction(ISD::MUL, MVT::v8i16, Custom); 546 setOperationAction(ISD::MUL, MVT::v4i32, Custom); 547 setOperationAction(ISD::MUL, MVT::v2i64, Custom); 548 // Custom handling for some vector types to avoid expensive expansions 549 setOperationAction(ISD::SDIV, MVT::v4i16, Custom); 550 setOperationAction(ISD::SDIV, MVT::v8i8, Custom); 551 setOperationAction(ISD::UDIV, MVT::v4i16, Custom); 552 setOperationAction(ISD::UDIV, MVT::v8i8, Custom); 553 setOperationAction(ISD::SETCC, MVT::v1i64, Expand); 554 setOperationAction(ISD::SETCC, MVT::v2i64, Expand); 555 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with 556 // a destination type that is wider than the source, and nor does 557 // it have a FP_TO_[SU]INT instruction with a narrower destination than 558 // source. 559 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom); 560 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom); 561 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom); 562 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom); 563 564 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand); 565 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand); 566 567 // Custom expand long extensions to vectors. 568 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom); 569 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom); 570 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom); 571 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom); 572 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom); 573 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom); 574 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom); 575 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom); 576 577 // NEON does not have single instruction CTPOP for vectors with element 578 // types wider than 8-bits. However, custom lowering can leverage the 579 // v8i8/v16i8 vcnt instruction. 580 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom); 581 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom); 582 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom); 583 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom); 584 585 // NEON only has FMA instructions as of VFP4. 586 if (!Subtarget->hasVFP4()) { 587 setOperationAction(ISD::FMA, MVT::v2f32, Expand); 588 setOperationAction(ISD::FMA, MVT::v4f32, Expand); 589 } 590 591 setTargetDAGCombine(ISD::INTRINSIC_VOID); 592 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN); 593 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN); 594 setTargetDAGCombine(ISD::SHL); 595 setTargetDAGCombine(ISD::SRL); 596 setTargetDAGCombine(ISD::SRA); 597 setTargetDAGCombine(ISD::SIGN_EXTEND); 598 setTargetDAGCombine(ISD::ZERO_EXTEND); 599 setTargetDAGCombine(ISD::ANY_EXTEND); 600 setTargetDAGCombine(ISD::SELECT_CC); 601 setTargetDAGCombine(ISD::BUILD_VECTOR); 602 setTargetDAGCombine(ISD::VECTOR_SHUFFLE); 603 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT); 604 setTargetDAGCombine(ISD::STORE); 605 setTargetDAGCombine(ISD::FP_TO_SINT); 606 setTargetDAGCombine(ISD::FP_TO_UINT); 607 setTargetDAGCombine(ISD::FDIV); 608 609 // It is legal to extload from v4i8 to v4i16 or v4i32. 610 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8, 611 MVT::v4i16, MVT::v2i16, 612 MVT::v2i32}; 613 for (unsigned i = 0; i < 6; ++i) { 614 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal); 615 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal); 616 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal); 617 } 618 } 619 620 // ARM and Thumb2 support UMLAL/SMLAL. 621 if (!Subtarget->isThumb1Only()) 622 setTargetDAGCombine(ISD::ADDC); 623 624 625 computeRegisterProperties(); 626 627 // ARM does not have f32 extending load. 628 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand); 629 630 // ARM does not have i1 sign extending load. 631 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 632 633 // ARM supports all 4 flavors of integer indexed load / store. 634 if (!Subtarget->isThumb1Only()) { 635 for (unsigned im = (unsigned)ISD::PRE_INC; 636 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) { 637 setIndexedLoadAction(im, MVT::i1, Legal); 638 setIndexedLoadAction(im, MVT::i8, Legal); 639 setIndexedLoadAction(im, MVT::i16, Legal); 640 setIndexedLoadAction(im, MVT::i32, Legal); 641 setIndexedStoreAction(im, MVT::i1, Legal); 642 setIndexedStoreAction(im, MVT::i8, Legal); 643 setIndexedStoreAction(im, MVT::i16, Legal); 644 setIndexedStoreAction(im, MVT::i32, Legal); 645 } 646 } 647 648 // i64 operation support. 649 setOperationAction(ISD::MUL, MVT::i64, Expand); 650 setOperationAction(ISD::MULHU, MVT::i32, Expand); 651 if (Subtarget->isThumb1Only()) { 652 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand); 653 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand); 654 } 655 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops() 656 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP())) 657 setOperationAction(ISD::MULHS, MVT::i32, Expand); 658 659 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); 660 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); 661 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); 662 setOperationAction(ISD::SRL, MVT::i64, Custom); 663 setOperationAction(ISD::SRA, MVT::i64, Custom); 664 665 if (!Subtarget->isThumb1Only()) { 666 // FIXME: We should do this for Thumb1 as well. 667 setOperationAction(ISD::ADDC, MVT::i32, Custom); 668 setOperationAction(ISD::ADDE, MVT::i32, Custom); 669 setOperationAction(ISD::SUBC, MVT::i32, Custom); 670 setOperationAction(ISD::SUBE, MVT::i32, Custom); 671 } 672 673 // ARM does not have ROTL. 674 setOperationAction(ISD::ROTL, MVT::i32, Expand); 675 setOperationAction(ISD::CTTZ, MVT::i32, Custom); 676 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 677 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only()) 678 setOperationAction(ISD::CTLZ, MVT::i32, Expand); 679 680 // These just redirect to CTTZ and CTLZ on ARM. 681 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand); 682 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand); 683 684 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom); 685 686 // Only ARMv6 has BSWAP. 687 if (!Subtarget->hasV6Ops()) 688 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 689 690 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) && 691 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) { 692 // These are expanded into libcalls if the cpu doesn't have HW divider. 693 setOperationAction(ISD::SDIV, MVT::i32, Expand); 694 setOperationAction(ISD::UDIV, MVT::i32, Expand); 695 } 696 697 // FIXME: Also set divmod for SREM on EABI 698 setOperationAction(ISD::SREM, MVT::i32, Expand); 699 setOperationAction(ISD::UREM, MVT::i32, Expand); 700 // Register based DivRem for AEABI (RTABI 4.2) 701 if (Subtarget->isTargetAEABI()) { 702 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod"); 703 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod"); 704 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod"); 705 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod"); 706 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod"); 707 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod"); 708 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod"); 709 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod"); 710 711 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS); 712 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS); 713 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS); 714 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS); 715 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS); 716 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS); 717 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS); 718 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS); 719 720 setOperationAction(ISD::SDIVREM, MVT::i32, Custom); 721 setOperationAction(ISD::UDIVREM, MVT::i32, Custom); 722 } else { 723 setOperationAction(ISD::SDIVREM, MVT::i32, Expand); 724 setOperationAction(ISD::UDIVREM, MVT::i32, Expand); 725 } 726 727 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 728 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 729 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom); 730 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 731 setOperationAction(ISD::BlockAddress, MVT::i32, Custom); 732 733 setOperationAction(ISD::TRAP, MVT::Other, Legal); 734 735 // Use the default implementation. 736 setOperationAction(ISD::VASTART, MVT::Other, Custom); 737 setOperationAction(ISD::VAARG, MVT::Other, Expand); 738 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 739 setOperationAction(ISD::VAEND, MVT::Other, Expand); 740 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 741 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 742 743 if (!Subtarget->isTargetDarwin()) { 744 // Non-Darwin platforms may return values in these registers via the 745 // personality function. 746 setExceptionPointerRegister(ARM::R0); 747 setExceptionSelectorRegister(ARM::R1); 748 } 749 750 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 751 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use 752 // the default expansion. 753 // FIXME: This should be checking for v6k, not just v6. 754 if (Subtarget->hasDataBarrier() || 755 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) { 756 // membarrier needs custom lowering; the rest are legal and handled 757 // normally. 758 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom); 759 // Custom lowering for 64-bit ops 760 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom); 761 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom); 762 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom); 763 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom); 764 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom); 765 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom); 766 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i64, Custom); 767 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i64, Custom); 768 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom); 769 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom); 770 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom); 771 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc. 772 setInsertFencesForAtomic(true); 773 } else { 774 // Set them all for expansion, which will force libcalls. 775 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand); 776 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand); 777 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand); 778 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand); 779 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand); 780 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand); 781 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand); 782 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand); 783 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand); 784 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand); 785 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand); 786 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand); 787 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand); 788 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the 789 // Unordered/Monotonic case. 790 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom); 791 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom); 792 } 793 794 setOperationAction(ISD::PREFETCH, MVT::Other, Custom); 795 796 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes. 797 if (!Subtarget->hasV6Ops()) { 798 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 799 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); 800 } 801 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 802 803 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() && 804 !Subtarget->isThumb1Only()) { 805 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR 806 // iff target supports vfp2. 807 setOperationAction(ISD::BITCAST, MVT::i64, Custom); 808 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom); 809 } 810 811 // We want to custom lower some of our intrinsics. 812 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); 813 if (Subtarget->isTargetDarwin()) { 814 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom); 815 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom); 816 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume"); 817 } 818 819 setOperationAction(ISD::SETCC, MVT::i32, Expand); 820 setOperationAction(ISD::SETCC, MVT::f32, Expand); 821 setOperationAction(ISD::SETCC, MVT::f64, Expand); 822 setOperationAction(ISD::SELECT, MVT::i32, Custom); 823 setOperationAction(ISD::SELECT, MVT::f32, Custom); 824 setOperationAction(ISD::SELECT, MVT::f64, Custom); 825 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom); 826 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 827 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 828 829 setOperationAction(ISD::BRCOND, MVT::Other, Expand); 830 setOperationAction(ISD::BR_CC, MVT::i32, Custom); 831 setOperationAction(ISD::BR_CC, MVT::f32, Custom); 832 setOperationAction(ISD::BR_CC, MVT::f64, Custom); 833 setOperationAction(ISD::BR_JT, MVT::Other, Custom); 834 835 // We don't support sin/cos/fmod/copysign/pow 836 setOperationAction(ISD::FSIN, MVT::f64, Expand); 837 setOperationAction(ISD::FSIN, MVT::f32, Expand); 838 setOperationAction(ISD::FCOS, MVT::f32, Expand); 839 setOperationAction(ISD::FCOS, MVT::f64, Expand); 840 setOperationAction(ISD::FSINCOS, MVT::f64, Expand); 841 setOperationAction(ISD::FSINCOS, MVT::f32, Expand); 842 setOperationAction(ISD::FREM, MVT::f64, Expand); 843 setOperationAction(ISD::FREM, MVT::f32, Expand); 844 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() && 845 !Subtarget->isThumb1Only()) { 846 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 847 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 848 } 849 setOperationAction(ISD::FPOW, MVT::f64, Expand); 850 setOperationAction(ISD::FPOW, MVT::f32, Expand); 851 852 if (!Subtarget->hasVFP4()) { 853 setOperationAction(ISD::FMA, MVT::f64, Expand); 854 setOperationAction(ISD::FMA, MVT::f32, Expand); 855 } 856 857 // Various VFP goodness 858 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) { 859 // int <-> fp are custom expanded into bit_convert + ARMISD ops. 860 if (Subtarget->hasVFP2()) { 861 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom); 862 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom); 863 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); 864 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 865 } 866 // Special handling for half-precision FP. 867 if (!Subtarget->hasFP16()) { 868 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand); 869 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand); 870 } 871 } 872 873 // We have target-specific dag combine patterns for the following nodes: 874 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine 875 setTargetDAGCombine(ISD::ADD); 876 setTargetDAGCombine(ISD::SUB); 877 setTargetDAGCombine(ISD::MUL); 878 setTargetDAGCombine(ISD::AND); 879 setTargetDAGCombine(ISD::OR); 880 setTargetDAGCombine(ISD::XOR); 881 882 if (Subtarget->hasV6Ops()) 883 setTargetDAGCombine(ISD::SRL); 884 885 setStackPointerRegisterToSaveRestore(ARM::SP); 886 887 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() || 888 !Subtarget->hasVFP2()) 889 setSchedulingPreference(Sched::RegPressure); 890 else 891 setSchedulingPreference(Sched::Hybrid); 892 893 //// temporary - rewrite interface to use type 894 MaxStoresPerMemset = 8; 895 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4; 896 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores 897 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2; 898 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores 899 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2; 900 901 // On ARM arguments smaller than 4 bytes are extended, so all arguments 902 // are at least 4 bytes aligned. 903 setMinStackArgumentAlignment(4); 904 905 // Prefer likely predicted branches to selects on out-of-order cores. 906 PredictableSelectIsExpensive = Subtarget->isLikeA9(); 907 908 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2); 909 } 910 911 // FIXME: It might make sense to define the representative register class as the 912 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is 913 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently, 914 // SPR's representative would be DPR_VFP2. This should work well if register 915 // pressure tracking were modified such that a register use would increment the 916 // pressure of the register class's representative and all of it's super 917 // classes' representatives transitively. We have not implemented this because 918 // of the difficulty prior to coalescing of modeling operand register classes 919 // due to the common occurrence of cross class copies and subregister insertions 920 // and extractions. 921 std::pair<const TargetRegisterClass*, uint8_t> 922 ARMTargetLowering::findRepresentativeClass(MVT VT) const{ 923 const TargetRegisterClass *RRC = 0; 924 uint8_t Cost = 1; 925 switch (VT.SimpleTy) { 926 default: 927 return TargetLowering::findRepresentativeClass(VT); 928 // Use DPR as representative register class for all floating point 929 // and vector types. Since there are 32 SPR registers and 32 DPR registers so 930 // the cost is 1 for both f32 and f64. 931 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16: 932 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32: 933 RRC = &ARM::DPRRegClass; 934 // When NEON is used for SP, only half of the register file is available 935 // because operations that define both SP and DP results will be constrained 936 // to the VFP2 class (D0-D15). We currently model this constraint prior to 937 // coalescing by double-counting the SP regs. See the FIXME above. 938 if (Subtarget->useNEONForSinglePrecisionFP()) 939 Cost = 2; 940 break; 941 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64: 942 case MVT::v4f32: case MVT::v2f64: 943 RRC = &ARM::DPRRegClass; 944 Cost = 2; 945 break; 946 case MVT::v4i64: 947 RRC = &ARM::DPRRegClass; 948 Cost = 4; 949 break; 950 case MVT::v8i64: 951 RRC = &ARM::DPRRegClass; 952 Cost = 8; 953 break; 954 } 955 return std::make_pair(RRC, Cost); 956 } 957 958 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const { 959 switch (Opcode) { 960 default: return 0; 961 case ARMISD::Wrapper: return "ARMISD::Wrapper"; 962 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN"; 963 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC"; 964 case ARMISD::WrapperJT: return "ARMISD::WrapperJT"; 965 case ARMISD::CALL: return "ARMISD::CALL"; 966 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED"; 967 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK"; 968 case ARMISD::tCALL: return "ARMISD::tCALL"; 969 case ARMISD::BRCOND: return "ARMISD::BRCOND"; 970 case ARMISD::BR_JT: return "ARMISD::BR_JT"; 971 case ARMISD::BR2_JT: return "ARMISD::BR2_JT"; 972 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG"; 973 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD"; 974 case ARMISD::CMP: return "ARMISD::CMP"; 975 case ARMISD::CMN: return "ARMISD::CMN"; 976 case ARMISD::CMPZ: return "ARMISD::CMPZ"; 977 case ARMISD::CMPFP: return "ARMISD::CMPFP"; 978 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0"; 979 case ARMISD::BCC_i64: return "ARMISD::BCC_i64"; 980 case ARMISD::FMSTAT: return "ARMISD::FMSTAT"; 981 982 case ARMISD::CMOV: return "ARMISD::CMOV"; 983 984 case ARMISD::RBIT: return "ARMISD::RBIT"; 985 986 case ARMISD::FTOSI: return "ARMISD::FTOSI"; 987 case ARMISD::FTOUI: return "ARMISD::FTOUI"; 988 case ARMISD::SITOF: return "ARMISD::SITOF"; 989 case ARMISD::UITOF: return "ARMISD::UITOF"; 990 991 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG"; 992 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG"; 993 case ARMISD::RRX: return "ARMISD::RRX"; 994 995 case ARMISD::ADDC: return "ARMISD::ADDC"; 996 case ARMISD::ADDE: return "ARMISD::ADDE"; 997 case ARMISD::SUBC: return "ARMISD::SUBC"; 998 case ARMISD::SUBE: return "ARMISD::SUBE"; 999 1000 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD"; 1001 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR"; 1002 1003 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP"; 1004 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP"; 1005 1006 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN"; 1007 1008 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER"; 1009 1010 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC"; 1011 1012 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER"; 1013 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR"; 1014 1015 case ARMISD::PRELOAD: return "ARMISD::PRELOAD"; 1016 1017 case ARMISD::VCEQ: return "ARMISD::VCEQ"; 1018 case ARMISD::VCEQZ: return "ARMISD::VCEQZ"; 1019 case ARMISD::VCGE: return "ARMISD::VCGE"; 1020 case ARMISD::VCGEZ: return "ARMISD::VCGEZ"; 1021 case ARMISD::VCLEZ: return "ARMISD::VCLEZ"; 1022 case ARMISD::VCGEU: return "ARMISD::VCGEU"; 1023 case ARMISD::VCGT: return "ARMISD::VCGT"; 1024 case ARMISD::VCGTZ: return "ARMISD::VCGTZ"; 1025 case ARMISD::VCLTZ: return "ARMISD::VCLTZ"; 1026 case ARMISD::VCGTU: return "ARMISD::VCGTU"; 1027 case ARMISD::VTST: return "ARMISD::VTST"; 1028 1029 case ARMISD::VSHL: return "ARMISD::VSHL"; 1030 case ARMISD::VSHRs: return "ARMISD::VSHRs"; 1031 case ARMISD::VSHRu: return "ARMISD::VSHRu"; 1032 case ARMISD::VSHLLs: return "ARMISD::VSHLLs"; 1033 case ARMISD::VSHLLu: return "ARMISD::VSHLLu"; 1034 case ARMISD::VSHLLi: return "ARMISD::VSHLLi"; 1035 case ARMISD::VSHRN: return "ARMISD::VSHRN"; 1036 case ARMISD::VRSHRs: return "ARMISD::VRSHRs"; 1037 case ARMISD::VRSHRu: return "ARMISD::VRSHRu"; 1038 case ARMISD::VRSHRN: return "ARMISD::VRSHRN"; 1039 case ARMISD::VQSHLs: return "ARMISD::VQSHLs"; 1040 case ARMISD::VQSHLu: return "ARMISD::VQSHLu"; 1041 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu"; 1042 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs"; 1043 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu"; 1044 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu"; 1045 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs"; 1046 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu"; 1047 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu"; 1048 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu"; 1049 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs"; 1050 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM"; 1051 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM"; 1052 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM"; 1053 case ARMISD::VDUP: return "ARMISD::VDUP"; 1054 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE"; 1055 case ARMISD::VEXT: return "ARMISD::VEXT"; 1056 case ARMISD::VREV64: return "ARMISD::VREV64"; 1057 case ARMISD::VREV32: return "ARMISD::VREV32"; 1058 case ARMISD::VREV16: return "ARMISD::VREV16"; 1059 case ARMISD::VZIP: return "ARMISD::VZIP"; 1060 case ARMISD::VUZP: return "ARMISD::VUZP"; 1061 case ARMISD::VTRN: return "ARMISD::VTRN"; 1062 case ARMISD::VTBL1: return "ARMISD::VTBL1"; 1063 case ARMISD::VTBL2: return "ARMISD::VTBL2"; 1064 case ARMISD::VMULLs: return "ARMISD::VMULLs"; 1065 case ARMISD::VMULLu: return "ARMISD::VMULLu"; 1066 case ARMISD::UMLAL: return "ARMISD::UMLAL"; 1067 case ARMISD::SMLAL: return "ARMISD::SMLAL"; 1068 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR"; 1069 case ARMISD::FMAX: return "ARMISD::FMAX"; 1070 case ARMISD::FMIN: return "ARMISD::FMIN"; 1071 case ARMISD::BFI: return "ARMISD::BFI"; 1072 case ARMISD::VORRIMM: return "ARMISD::VORRIMM"; 1073 case ARMISD::VBICIMM: return "ARMISD::VBICIMM"; 1074 case ARMISD::VBSL: return "ARMISD::VBSL"; 1075 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP"; 1076 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP"; 1077 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP"; 1078 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD"; 1079 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD"; 1080 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD"; 1081 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD"; 1082 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD"; 1083 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD"; 1084 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD"; 1085 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD"; 1086 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD"; 1087 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD"; 1088 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD"; 1089 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD"; 1090 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD"; 1091 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD"; 1092 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD"; 1093 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD"; 1094 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD"; 1095 1096 case ARMISD::ATOMADD64_DAG: return "ATOMADD64_DAG"; 1097 case ARMISD::ATOMSUB64_DAG: return "ATOMSUB64_DAG"; 1098 case ARMISD::ATOMOR64_DAG: return "ATOMOR64_DAG"; 1099 case ARMISD::ATOMXOR64_DAG: return "ATOMXOR64_DAG"; 1100 case ARMISD::ATOMAND64_DAG: return "ATOMAND64_DAG"; 1101 case ARMISD::ATOMNAND64_DAG: return "ATOMNAND64_DAG"; 1102 case ARMISD::ATOMSWAP64_DAG: return "ATOMSWAP64_DAG"; 1103 case ARMISD::ATOMCMPXCHG64_DAG: return "ATOMCMPXCHG64_DAG"; 1104 case ARMISD::ATOMMIN64_DAG: return "ATOMMIN64_DAG"; 1105 case ARMISD::ATOMUMIN64_DAG: return "ATOMUMIN64_DAG"; 1106 case ARMISD::ATOMMAX64_DAG: return "ATOMMAX64_DAG"; 1107 case ARMISD::ATOMUMAX64_DAG: return "ATOMUMAX64_DAG"; 1108 } 1109 } 1110 1111 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const { 1112 if (!VT.isVector()) return getPointerTy(); 1113 return VT.changeVectorElementTypeToInteger(); 1114 } 1115 1116 /// getRegClassFor - Return the register class that should be used for the 1117 /// specified value type. 1118 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const { 1119 // Map v4i64 to QQ registers but do not make the type legal. Similarly map 1120 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to 1121 // load / store 4 to 8 consecutive D registers. 1122 if (Subtarget->hasNEON()) { 1123 if (VT == MVT::v4i64) 1124 return &ARM::QQPRRegClass; 1125 if (VT == MVT::v8i64) 1126 return &ARM::QQQQPRRegClass; 1127 } 1128 return TargetLowering::getRegClassFor(VT); 1129 } 1130 1131 // Create a fast isel object. 1132 FastISel * 1133 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo, 1134 const TargetLibraryInfo *libInfo) const { 1135 return ARM::createFastISel(funcInfo, libInfo); 1136 } 1137 1138 /// getMaximalGlobalOffset - Returns the maximal possible offset which can 1139 /// be used for loads / stores from the global. 1140 unsigned ARMTargetLowering::getMaximalGlobalOffset() const { 1141 return (Subtarget->isThumb1Only() ? 127 : 4095); 1142 } 1143 1144 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const { 1145 unsigned NumVals = N->getNumValues(); 1146 if (!NumVals) 1147 return Sched::RegPressure; 1148 1149 for (unsigned i = 0; i != NumVals; ++i) { 1150 EVT VT = N->getValueType(i); 1151 if (VT == MVT::Glue || VT == MVT::Other) 1152 continue; 1153 if (VT.isFloatingPoint() || VT.isVector()) 1154 return Sched::ILP; 1155 } 1156 1157 if (!N->isMachineOpcode()) 1158 return Sched::RegPressure; 1159 1160 // Load are scheduled for latency even if there instruction itinerary 1161 // is not available. 1162 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1163 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode()); 1164 1165 if (MCID.getNumDefs() == 0) 1166 return Sched::RegPressure; 1167 if (!Itins->isEmpty() && 1168 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2) 1169 return Sched::ILP; 1170 1171 return Sched::RegPressure; 1172 } 1173 1174 //===----------------------------------------------------------------------===// 1175 // Lowering Code 1176 //===----------------------------------------------------------------------===// 1177 1178 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC 1179 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) { 1180 switch (CC) { 1181 default: llvm_unreachable("Unknown condition code!"); 1182 case ISD::SETNE: return ARMCC::NE; 1183 case ISD::SETEQ: return ARMCC::EQ; 1184 case ISD::SETGT: return ARMCC::GT; 1185 case ISD::SETGE: return ARMCC::GE; 1186 case ISD::SETLT: return ARMCC::LT; 1187 case ISD::SETLE: return ARMCC::LE; 1188 case ISD::SETUGT: return ARMCC::HI; 1189 case ISD::SETUGE: return ARMCC::HS; 1190 case ISD::SETULT: return ARMCC::LO; 1191 case ISD::SETULE: return ARMCC::LS; 1192 } 1193 } 1194 1195 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. 1196 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode, 1197 ARMCC::CondCodes &CondCode2) { 1198 CondCode2 = ARMCC::AL; 1199 switch (CC) { 1200 default: llvm_unreachable("Unknown FP condition!"); 1201 case ISD::SETEQ: 1202 case ISD::SETOEQ: CondCode = ARMCC::EQ; break; 1203 case ISD::SETGT: 1204 case ISD::SETOGT: CondCode = ARMCC::GT; break; 1205 case ISD::SETGE: 1206 case ISD::SETOGE: CondCode = ARMCC::GE; break; 1207 case ISD::SETOLT: CondCode = ARMCC::MI; break; 1208 case ISD::SETOLE: CondCode = ARMCC::LS; break; 1209 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break; 1210 case ISD::SETO: CondCode = ARMCC::VC; break; 1211 case ISD::SETUO: CondCode = ARMCC::VS; break; 1212 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break; 1213 case ISD::SETUGT: CondCode = ARMCC::HI; break; 1214 case ISD::SETUGE: CondCode = ARMCC::PL; break; 1215 case ISD::SETLT: 1216 case ISD::SETULT: CondCode = ARMCC::LT; break; 1217 case ISD::SETLE: 1218 case ISD::SETULE: CondCode = ARMCC::LE; break; 1219 case ISD::SETNE: 1220 case ISD::SETUNE: CondCode = ARMCC::NE; break; 1221 } 1222 } 1223 1224 //===----------------------------------------------------------------------===// 1225 // Calling Convention Implementation 1226 //===----------------------------------------------------------------------===// 1227 1228 #include "ARMGenCallingConv.inc" 1229 1230 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the 1231 /// given CallingConvention value. 1232 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC, 1233 bool Return, 1234 bool isVarArg) const { 1235 switch (CC) { 1236 default: 1237 llvm_unreachable("Unsupported calling convention"); 1238 case CallingConv::Fast: 1239 if (Subtarget->hasVFP2() && !isVarArg) { 1240 if (!Subtarget->isAAPCS_ABI()) 1241 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS); 1242 // For AAPCS ABI targets, just use VFP variant of the calling convention. 1243 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP); 1244 } 1245 // Fallthrough 1246 case CallingConv::C: { 1247 // Use target triple & subtarget features to do actual dispatch. 1248 if (!Subtarget->isAAPCS_ABI()) 1249 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS); 1250 else if (Subtarget->hasVFP2() && 1251 getTargetMachine().Options.FloatABIType == FloatABI::Hard && 1252 !isVarArg) 1253 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP); 1254 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS); 1255 } 1256 case CallingConv::ARM_AAPCS_VFP: 1257 if (!isVarArg) 1258 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP); 1259 // Fallthrough 1260 case CallingConv::ARM_AAPCS: 1261 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS); 1262 case CallingConv::ARM_APCS: 1263 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS); 1264 case CallingConv::GHC: 1265 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC); 1266 } 1267 } 1268 1269 /// LowerCallResult - Lower the result values of a call into the 1270 /// appropriate copies out of appropriate physical registers. 1271 SDValue 1272 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, 1273 CallingConv::ID CallConv, bool isVarArg, 1274 const SmallVectorImpl<ISD::InputArg> &Ins, 1275 SDLoc dl, SelectionDAG &DAG, 1276 SmallVectorImpl<SDValue> &InVals, 1277 bool isThisReturn, SDValue ThisVal) const { 1278 1279 // Assign locations to each value returned by this call. 1280 SmallVector<CCValAssign, 16> RVLocs; 1281 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1282 getTargetMachine(), RVLocs, *DAG.getContext(), Call); 1283 CCInfo.AnalyzeCallResult(Ins, 1284 CCAssignFnForNode(CallConv, /* Return*/ true, 1285 isVarArg)); 1286 1287 // Copy all of the result registers out of their specified physreg. 1288 for (unsigned i = 0; i != RVLocs.size(); ++i) { 1289 CCValAssign VA = RVLocs[i]; 1290 1291 // Pass 'this' value directly from the argument to return value, to avoid 1292 // reg unit interference 1293 if (i == 0 && isThisReturn) { 1294 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 && 1295 "unexpected return calling convention register assignment"); 1296 InVals.push_back(ThisVal); 1297 continue; 1298 } 1299 1300 SDValue Val; 1301 if (VA.needsCustom()) { 1302 // Handle f64 or half of a v2f64. 1303 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, 1304 InFlag); 1305 Chain = Lo.getValue(1); 1306 InFlag = Lo.getValue(2); 1307 VA = RVLocs[++i]; // skip ahead to next loc 1308 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, 1309 InFlag); 1310 Chain = Hi.getValue(1); 1311 InFlag = Hi.getValue(2); 1312 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi); 1313 1314 if (VA.getLocVT() == MVT::v2f64) { 1315 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64); 1316 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val, 1317 DAG.getConstant(0, MVT::i32)); 1318 1319 VA = RVLocs[++i]; // skip ahead to next loc 1320 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag); 1321 Chain = Lo.getValue(1); 1322 InFlag = Lo.getValue(2); 1323 VA = RVLocs[++i]; // skip ahead to next loc 1324 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag); 1325 Chain = Hi.getValue(1); 1326 InFlag = Hi.getValue(2); 1327 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi); 1328 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val, 1329 DAG.getConstant(1, MVT::i32)); 1330 } 1331 } else { 1332 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(), 1333 InFlag); 1334 Chain = Val.getValue(1); 1335 InFlag = Val.getValue(2); 1336 } 1337 1338 switch (VA.getLocInfo()) { 1339 default: llvm_unreachable("Unknown loc info!"); 1340 case CCValAssign::Full: break; 1341 case CCValAssign::BCvt: 1342 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val); 1343 break; 1344 } 1345 1346 InVals.push_back(Val); 1347 } 1348 1349 return Chain; 1350 } 1351 1352 /// LowerMemOpCallTo - Store the argument to the stack. 1353 SDValue 1354 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain, 1355 SDValue StackPtr, SDValue Arg, 1356 SDLoc dl, SelectionDAG &DAG, 1357 const CCValAssign &VA, 1358 ISD::ArgFlagsTy Flags) const { 1359 unsigned LocMemOffset = VA.getLocMemOffset(); 1360 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset); 1361 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff); 1362 return DAG.getStore(Chain, dl, Arg, PtrOff, 1363 MachinePointerInfo::getStack(LocMemOffset), 1364 false, false, 0); 1365 } 1366 1367 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG, 1368 SDValue Chain, SDValue &Arg, 1369 RegsToPassVector &RegsToPass, 1370 CCValAssign &VA, CCValAssign &NextVA, 1371 SDValue &StackPtr, 1372 SmallVectorImpl<SDValue> &MemOpChains, 1373 ISD::ArgFlagsTy Flags) const { 1374 1375 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl, 1376 DAG.getVTList(MVT::i32, MVT::i32), Arg); 1377 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd)); 1378 1379 if (NextVA.isRegLoc()) 1380 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1))); 1381 else { 1382 assert(NextVA.isMemLoc()); 1383 if (StackPtr.getNode() == 0) 1384 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 1385 1386 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1), 1387 dl, DAG, NextVA, 1388 Flags)); 1389 } 1390 } 1391 1392 /// LowerCall - Lowering a call into a callseq_start <- 1393 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter 1394 /// nodes. 1395 SDValue 1396 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, 1397 SmallVectorImpl<SDValue> &InVals) const { 1398 SelectionDAG &DAG = CLI.DAG; 1399 SDLoc &dl = CLI.DL; 1400 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; 1401 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; 1402 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; 1403 SDValue Chain = CLI.Chain; 1404 SDValue Callee = CLI.Callee; 1405 bool &isTailCall = CLI.IsTailCall; 1406 CallingConv::ID CallConv = CLI.CallConv; 1407 bool doesNotRet = CLI.DoesNotReturn; 1408 bool isVarArg = CLI.IsVarArg; 1409 1410 MachineFunction &MF = DAG.getMachineFunction(); 1411 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet(); 1412 bool isThisReturn = false; 1413 bool isSibCall = false; 1414 // Disable tail calls if they're not supported. 1415 if (!EnableARMTailCalls && !Subtarget->supportsTailCall()) 1416 isTailCall = false; 1417 if (isTailCall) { 1418 // Check if it's really possible to do a tail call. 1419 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, 1420 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(), 1421 Outs, OutVals, Ins, DAG); 1422 // We don't support GuaranteedTailCallOpt for ARM, only automatically 1423 // detected sibcalls. 1424 if (isTailCall) { 1425 ++NumTailCalls; 1426 isSibCall = true; 1427 } 1428 } 1429 1430 // Analyze operands of the call, assigning locations to each operand. 1431 SmallVector<CCValAssign, 16> ArgLocs; 1432 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 1433 getTargetMachine(), ArgLocs, *DAG.getContext(), Call); 1434 CCInfo.AnalyzeCallOperands(Outs, 1435 CCAssignFnForNode(CallConv, /* Return*/ false, 1436 isVarArg)); 1437 1438 // Get a count of how many bytes are to be pushed on the stack. 1439 unsigned NumBytes = CCInfo.getNextStackOffset(); 1440 1441 // For tail calls, memory operands are available in our caller's stack. 1442 if (isSibCall) 1443 NumBytes = 0; 1444 1445 // Adjust the stack pointer for the new arguments... 1446 // These operations are automatically eliminated by the prolog/epilog pass 1447 if (!isSibCall) 1448 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), 1449 dl); 1450 1451 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 1452 1453 RegsToPassVector RegsToPass; 1454 SmallVector<SDValue, 8> MemOpChains; 1455 1456 // Walk the register/memloc assignments, inserting copies/loads. In the case 1457 // of tail call optimization, arguments are handled later. 1458 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); 1459 i != e; 1460 ++i, ++realArgIdx) { 1461 CCValAssign &VA = ArgLocs[i]; 1462 SDValue Arg = OutVals[realArgIdx]; 1463 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags; 1464 bool isByVal = Flags.isByVal(); 1465 1466 // Promote the value if needed. 1467 switch (VA.getLocInfo()) { 1468 default: llvm_unreachable("Unknown loc info!"); 1469 case CCValAssign::Full: break; 1470 case CCValAssign::SExt: 1471 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); 1472 break; 1473 case CCValAssign::ZExt: 1474 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); 1475 break; 1476 case CCValAssign::AExt: 1477 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg); 1478 break; 1479 case CCValAssign::BCvt: 1480 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg); 1481 break; 1482 } 1483 1484 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces 1485 if (VA.needsCustom()) { 1486 if (VA.getLocVT() == MVT::v2f64) { 1487 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1488 DAG.getConstant(0, MVT::i32)); 1489 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1490 DAG.getConstant(1, MVT::i32)); 1491 1492 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass, 1493 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags); 1494 1495 VA = ArgLocs[++i]; // skip ahead to next loc 1496 if (VA.isRegLoc()) { 1497 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass, 1498 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags); 1499 } else { 1500 assert(VA.isMemLoc()); 1501 1502 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1, 1503 dl, DAG, VA, Flags)); 1504 } 1505 } else { 1506 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i], 1507 StackPtr, MemOpChains, Flags); 1508 } 1509 } else if (VA.isRegLoc()) { 1510 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) { 1511 assert(VA.getLocVT() == MVT::i32 && 1512 "unexpected calling convention register assignment"); 1513 assert(!Ins.empty() && Ins[0].VT == MVT::i32 && 1514 "unexpected use of 'returned'"); 1515 isThisReturn = true; 1516 } 1517 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 1518 } else if (isByVal) { 1519 assert(VA.isMemLoc()); 1520 unsigned offset = 0; 1521 1522 // True if this byval aggregate will be split between registers 1523 // and memory. 1524 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount(); 1525 unsigned CurByValIdx = CCInfo.getInRegsParamsProceed(); 1526 1527 if (CurByValIdx < ByValArgsCount) { 1528 1529 unsigned RegBegin, RegEnd; 1530 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd); 1531 1532 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 1533 unsigned int i, j; 1534 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) { 1535 SDValue Const = DAG.getConstant(4*i, MVT::i32); 1536 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const); 1537 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg, 1538 MachinePointerInfo(), 1539 false, false, false, 0); 1540 MemOpChains.push_back(Load.getValue(1)); 1541 RegsToPass.push_back(std::make_pair(j, Load)); 1542 } 1543 1544 // If parameter size outsides register area, "offset" value 1545 // helps us to calculate stack slot for remained part properly. 1546 offset = RegEnd - RegBegin; 1547 1548 CCInfo.nextInRegsParam(); 1549 } 1550 1551 if (Flags.getByValSize() > 4*offset) { 1552 unsigned LocMemOffset = VA.getLocMemOffset(); 1553 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset); 1554 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, 1555 StkPtrOff); 1556 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset); 1557 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset); 1558 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, 1559 MVT::i32); 1560 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32); 1561 1562 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue); 1563 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode}; 1564 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs, 1565 Ops, array_lengthof(Ops))); 1566 } 1567 } else if (!isSibCall) { 1568 assert(VA.isMemLoc()); 1569 1570 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg, 1571 dl, DAG, VA, Flags)); 1572 } 1573 } 1574 1575 if (!MemOpChains.empty()) 1576 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 1577 &MemOpChains[0], MemOpChains.size()); 1578 1579 // Build a sequence of copy-to-reg nodes chained together with token chain 1580 // and flag operands which copy the outgoing args into the appropriate regs. 1581 SDValue InFlag; 1582 // Tail call byval lowering might overwrite argument registers so in case of 1583 // tail call optimization the copies to registers are lowered later. 1584 if (!isTailCall) 1585 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 1586 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, 1587 RegsToPass[i].second, InFlag); 1588 InFlag = Chain.getValue(1); 1589 } 1590 1591 // For tail calls lower the arguments to the 'real' stack slot. 1592 if (isTailCall) { 1593 // Force all the incoming stack arguments to be loaded from the stack 1594 // before any new outgoing arguments are stored to the stack, because the 1595 // outgoing stack slots may alias the incoming argument stack slots, and 1596 // the alias isn't otherwise explicit. This is slightly more conservative 1597 // than necessary, because it means that each store effectively depends 1598 // on every argument instead of just those arguments it would clobber. 1599 1600 // Do not flag preceding copytoreg stuff together with the following stuff. 1601 InFlag = SDValue(); 1602 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 1603 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, 1604 RegsToPass[i].second, InFlag); 1605 InFlag = Chain.getValue(1); 1606 } 1607 InFlag = SDValue(); 1608 } 1609 1610 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 1611 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 1612 // node so that legalize doesn't hack it. 1613 bool isDirect = false; 1614 bool isARMFunc = false; 1615 bool isLocalARMFunc = false; 1616 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1617 1618 if (EnableARMLongCalls) { 1619 assert (getTargetMachine().getRelocationModel() == Reloc::Static 1620 && "long-calls with non-static relocation model!"); 1621 // Handle a global address or an external symbol. If it's not one of 1622 // those, the target's already in a register, so we don't need to do 1623 // anything extra. 1624 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 1625 const GlobalValue *GV = G->getGlobal(); 1626 // Create a constant pool entry for the callee address 1627 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1628 ARMConstantPoolValue *CPV = 1629 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0); 1630 1631 // Get the address of the callee into a register 1632 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1633 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1634 Callee = DAG.getLoad(getPointerTy(), dl, 1635 DAG.getEntryNode(), CPAddr, 1636 MachinePointerInfo::getConstantPool(), 1637 false, false, false, 0); 1638 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) { 1639 const char *Sym = S->getSymbol(); 1640 1641 // Create a constant pool entry for the callee address 1642 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1643 ARMConstantPoolValue *CPV = 1644 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym, 1645 ARMPCLabelIndex, 0); 1646 // Get the address of the callee into a register 1647 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1648 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1649 Callee = DAG.getLoad(getPointerTy(), dl, 1650 DAG.getEntryNode(), CPAddr, 1651 MachinePointerInfo::getConstantPool(), 1652 false, false, false, 0); 1653 } 1654 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 1655 const GlobalValue *GV = G->getGlobal(); 1656 isDirect = true; 1657 bool isExt = GV->isDeclaration() || GV->isWeakForLinker(); 1658 bool isStub = (isExt && Subtarget->isTargetDarwin()) && 1659 getTargetMachine().getRelocationModel() != Reloc::Static; 1660 isARMFunc = !Subtarget->isThumb() || isStub; 1661 // ARM call to a local ARM function is predicable. 1662 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking); 1663 // tBX takes a register source operand. 1664 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) { 1665 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1666 ARMConstantPoolValue *CPV = 1667 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4); 1668 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1669 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1670 Callee = DAG.getLoad(getPointerTy(), dl, 1671 DAG.getEntryNode(), CPAddr, 1672 MachinePointerInfo::getConstantPool(), 1673 false, false, false, 0); 1674 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 1675 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, 1676 getPointerTy(), Callee, PICLabel); 1677 } else { 1678 // On ELF targets for PIC code, direct calls should go through the PLT 1679 unsigned OpFlags = 0; 1680 if (Subtarget->isTargetELF() && 1681 getTargetMachine().getRelocationModel() == Reloc::PIC_) 1682 OpFlags = ARMII::MO_PLT; 1683 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags); 1684 } 1685 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 1686 isDirect = true; 1687 bool isStub = Subtarget->isTargetDarwin() && 1688 getTargetMachine().getRelocationModel() != Reloc::Static; 1689 isARMFunc = !Subtarget->isThumb() || isStub; 1690 // tBX takes a register source operand. 1691 const char *Sym = S->getSymbol(); 1692 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) { 1693 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 1694 ARMConstantPoolValue *CPV = 1695 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym, 1696 ARMPCLabelIndex, 4); 1697 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 1698 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1699 Callee = DAG.getLoad(getPointerTy(), dl, 1700 DAG.getEntryNode(), CPAddr, 1701 MachinePointerInfo::getConstantPool(), 1702 false, false, false, 0); 1703 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 1704 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, 1705 getPointerTy(), Callee, PICLabel); 1706 } else { 1707 unsigned OpFlags = 0; 1708 // On ELF targets for PIC code, direct calls should go through the PLT 1709 if (Subtarget->isTargetELF() && 1710 getTargetMachine().getRelocationModel() == Reloc::PIC_) 1711 OpFlags = ARMII::MO_PLT; 1712 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags); 1713 } 1714 } 1715 1716 // FIXME: handle tail calls differently. 1717 unsigned CallOpc; 1718 bool HasMinSizeAttr = MF.getFunction()->getAttributes(). 1719 hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize); 1720 if (Subtarget->isThumb()) { 1721 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps()) 1722 CallOpc = ARMISD::CALL_NOLINK; 1723 else 1724 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL; 1725 } else { 1726 if (!isDirect && !Subtarget->hasV5TOps()) 1727 CallOpc = ARMISD::CALL_NOLINK; 1728 else if (doesNotRet && isDirect && Subtarget->hasRAS() && 1729 // Emit regular call when code size is the priority 1730 !HasMinSizeAttr) 1731 // "mov lr, pc; b _foo" to avoid confusing the RSP 1732 CallOpc = ARMISD::CALL_NOLINK; 1733 else 1734 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL; 1735 } 1736 1737 std::vector<SDValue> Ops; 1738 Ops.push_back(Chain); 1739 Ops.push_back(Callee); 1740 1741 // Add argument registers to the end of the list so that they are known live 1742 // into the call. 1743 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 1744 Ops.push_back(DAG.getRegister(RegsToPass[i].first, 1745 RegsToPass[i].second.getValueType())); 1746 1747 // Add a register mask operand representing the call-preserved registers. 1748 const uint32_t *Mask; 1749 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo(); 1750 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI); 1751 if (isThisReturn) { 1752 // For 'this' returns, use the R0-preserving mask if applicable 1753 Mask = ARI->getThisReturnPreservedMask(CallConv); 1754 if (!Mask) { 1755 // Set isThisReturn to false if the calling convention is not one that 1756 // allows 'returned' to be modeled in this way, so LowerCallResult does 1757 // not try to pass 'this' straight through 1758 isThisReturn = false; 1759 Mask = ARI->getCallPreservedMask(CallConv); 1760 } 1761 } else 1762 Mask = ARI->getCallPreservedMask(CallConv); 1763 1764 assert(Mask && "Missing call preserved mask for calling convention"); 1765 Ops.push_back(DAG.getRegisterMask(Mask)); 1766 1767 if (InFlag.getNode()) 1768 Ops.push_back(InFlag); 1769 1770 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 1771 if (isTailCall) 1772 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size()); 1773 1774 // Returns a chain and a flag for retval copy to use. 1775 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size()); 1776 InFlag = Chain.getValue(1); 1777 1778 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true), 1779 DAG.getIntPtrConstant(0, true), InFlag, dl); 1780 if (!Ins.empty()) 1781 InFlag = Chain.getValue(1); 1782 1783 // Handle result values, copying them out of physregs into vregs that we 1784 // return. 1785 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG, 1786 InVals, isThisReturn, 1787 isThisReturn ? OutVals[0] : SDValue()); 1788 } 1789 1790 /// HandleByVal - Every parameter *after* a byval parameter is passed 1791 /// on the stack. Remember the next parameter register to allocate, 1792 /// and then confiscate the rest of the parameter registers to insure 1793 /// this. 1794 void 1795 ARMTargetLowering::HandleByVal( 1796 CCState *State, unsigned &size, unsigned Align) const { 1797 unsigned reg = State->AllocateReg(GPRArgRegs, 4); 1798 assert((State->getCallOrPrologue() == Prologue || 1799 State->getCallOrPrologue() == Call) && 1800 "unhandled ParmContext"); 1801 1802 // For in-prologue parameters handling, we also introduce stack offset 1803 // for byval registers: see CallingConvLower.cpp, CCState::HandleByVal. 1804 // This behaviour outsides AAPCS rules (5.5 Parameters Passing) of how 1805 // NSAA should be evaluted (NSAA means "next stacked argument address"). 1806 // So: NextStackOffset = NSAAOffset + SizeOfByValParamsStoredInRegs. 1807 // Then: NSAAOffset = NextStackOffset - SizeOfByValParamsStoredInRegs. 1808 unsigned NSAAOffset = State->getNextStackOffset(); 1809 if (State->getCallOrPrologue() != Call) { 1810 for (unsigned i = 0, e = State->getInRegsParamsCount(); i != e; ++i) { 1811 unsigned RB, RE; 1812 State->getInRegsParamInfo(i, RB, RE); 1813 assert(NSAAOffset >= (RE-RB)*4 && 1814 "Stack offset for byval regs doesn't introduced anymore?"); 1815 NSAAOffset -= (RE-RB)*4; 1816 } 1817 } 1818 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) { 1819 if (Subtarget->isAAPCS_ABI() && Align > 4) { 1820 unsigned AlignInRegs = Align / 4; 1821 unsigned Waste = (ARM::R4 - reg) % AlignInRegs; 1822 for (unsigned i = 0; i < Waste; ++i) 1823 reg = State->AllocateReg(GPRArgRegs, 4); 1824 } 1825 if (reg != 0) { 1826 unsigned excess = 4 * (ARM::R4 - reg); 1827 1828 // Special case when NSAA != SP and parameter size greater than size of 1829 // all remained GPR regs. In that case we can't split parameter, we must 1830 // send it to stack. We also must set NCRN to R4, so waste all 1831 // remained registers. 1832 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) { 1833 while (State->AllocateReg(GPRArgRegs, 4)) 1834 ; 1835 return; 1836 } 1837 1838 // First register for byval parameter is the first register that wasn't 1839 // allocated before this method call, so it would be "reg". 1840 // If parameter is small enough to be saved in range [reg, r4), then 1841 // the end (first after last) register would be reg + param-size-in-regs, 1842 // else parameter would be splitted between registers and stack, 1843 // end register would be r4 in this case. 1844 unsigned ByValRegBegin = reg; 1845 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4; 1846 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd); 1847 // Note, first register is allocated in the beginning of function already, 1848 // allocate remained amount of registers we need. 1849 for (unsigned i = reg+1; i != ByValRegEnd; ++i) 1850 State->AllocateReg(GPRArgRegs, 4); 1851 // At a call site, a byval parameter that is split between 1852 // registers and memory needs its size truncated here. In a 1853 // function prologue, such byval parameters are reassembled in 1854 // memory, and are not truncated. 1855 if (State->getCallOrPrologue() == Call) { 1856 // Make remained size equal to 0 in case, when 1857 // the whole structure may be stored into registers. 1858 if (size < excess) 1859 size = 0; 1860 else 1861 size -= excess; 1862 } 1863 } 1864 } 1865 } 1866 1867 /// MatchingStackOffset - Return true if the given stack call argument is 1868 /// already available in the same position (relatively) of the caller's 1869 /// incoming argument stack. 1870 static 1871 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags, 1872 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI, 1873 const TargetInstrInfo *TII) { 1874 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8; 1875 int FI = INT_MAX; 1876 if (Arg.getOpcode() == ISD::CopyFromReg) { 1877 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg(); 1878 if (!TargetRegisterInfo::isVirtualRegister(VR)) 1879 return false; 1880 MachineInstr *Def = MRI->getVRegDef(VR); 1881 if (!Def) 1882 return false; 1883 if (!Flags.isByVal()) { 1884 if (!TII->isLoadFromStackSlot(Def, FI)) 1885 return false; 1886 } else { 1887 return false; 1888 } 1889 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) { 1890 if (Flags.isByVal()) 1891 // ByVal argument is passed in as a pointer but it's now being 1892 // dereferenced. e.g. 1893 // define @foo(%struct.X* %A) { 1894 // tail call @bar(%struct.X* byval %A) 1895 // } 1896 return false; 1897 SDValue Ptr = Ld->getBasePtr(); 1898 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr); 1899 if (!FINode) 1900 return false; 1901 FI = FINode->getIndex(); 1902 } else 1903 return false; 1904 1905 assert(FI != INT_MAX); 1906 if (!MFI->isFixedObjectIndex(FI)) 1907 return false; 1908 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI); 1909 } 1910 1911 /// IsEligibleForTailCallOptimization - Check whether the call is eligible 1912 /// for tail call optimization. Targets which want to do tail call 1913 /// optimization should implement this function. 1914 bool 1915 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee, 1916 CallingConv::ID CalleeCC, 1917 bool isVarArg, 1918 bool isCalleeStructRet, 1919 bool isCallerStructRet, 1920 const SmallVectorImpl<ISD::OutputArg> &Outs, 1921 const SmallVectorImpl<SDValue> &OutVals, 1922 const SmallVectorImpl<ISD::InputArg> &Ins, 1923 SelectionDAG& DAG) const { 1924 const Function *CallerF = DAG.getMachineFunction().getFunction(); 1925 CallingConv::ID CallerCC = CallerF->getCallingConv(); 1926 bool CCMatch = CallerCC == CalleeCC; 1927 1928 // Look for obvious safe cases to perform tail call optimization that do not 1929 // require ABI changes. This is what gcc calls sibcall. 1930 1931 // Do not sibcall optimize vararg calls unless the call site is not passing 1932 // any arguments. 1933 if (isVarArg && !Outs.empty()) 1934 return false; 1935 1936 // Also avoid sibcall optimization if either caller or callee uses struct 1937 // return semantics. 1938 if (isCalleeStructRet || isCallerStructRet) 1939 return false; 1940 1941 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo:: 1942 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as 1943 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation 1944 // support in the assembler and linker to be used. This would need to be 1945 // fixed to fully support tail calls in Thumb1. 1946 // 1947 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take 1948 // LR. This means if we need to reload LR, it takes an extra instructions, 1949 // which outweighs the value of the tail call; but here we don't know yet 1950 // whether LR is going to be used. Probably the right approach is to 1951 // generate the tail call here and turn it back into CALL/RET in 1952 // emitEpilogue if LR is used. 1953 1954 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls, 1955 // but we need to make sure there are enough registers; the only valid 1956 // registers are the 4 used for parameters. We don't currently do this 1957 // case. 1958 if (Subtarget->isThumb1Only()) 1959 return false; 1960 1961 // If the calling conventions do not match, then we'd better make sure the 1962 // results are returned in the same way as what the caller expects. 1963 if (!CCMatch) { 1964 SmallVector<CCValAssign, 16> RVLocs1; 1965 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), 1966 getTargetMachine(), RVLocs1, *DAG.getContext(), Call); 1967 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg)); 1968 1969 SmallVector<CCValAssign, 16> RVLocs2; 1970 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), 1971 getTargetMachine(), RVLocs2, *DAG.getContext(), Call); 1972 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg)); 1973 1974 if (RVLocs1.size() != RVLocs2.size()) 1975 return false; 1976 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) { 1977 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc()) 1978 return false; 1979 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo()) 1980 return false; 1981 if (RVLocs1[i].isRegLoc()) { 1982 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg()) 1983 return false; 1984 } else { 1985 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset()) 1986 return false; 1987 } 1988 } 1989 } 1990 1991 // If Caller's vararg or byval argument has been split between registers and 1992 // stack, do not perform tail call, since part of the argument is in caller's 1993 // local frame. 1994 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction(). 1995 getInfo<ARMFunctionInfo>(); 1996 if (AFI_Caller->getArgRegsSaveSize()) 1997 return false; 1998 1999 // If the callee takes no arguments then go on to check the results of the 2000 // call. 2001 if (!Outs.empty()) { 2002 // Check if stack adjustment is needed. For now, do not do this if any 2003 // argument is passed on the stack. 2004 SmallVector<CCValAssign, 16> ArgLocs; 2005 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), 2006 getTargetMachine(), ArgLocs, *DAG.getContext(), Call); 2007 CCInfo.AnalyzeCallOperands(Outs, 2008 CCAssignFnForNode(CalleeCC, false, isVarArg)); 2009 if (CCInfo.getNextStackOffset()) { 2010 MachineFunction &MF = DAG.getMachineFunction(); 2011 2012 // Check if the arguments are already laid out in the right way as 2013 // the caller's fixed stack objects. 2014 MachineFrameInfo *MFI = MF.getFrameInfo(); 2015 const MachineRegisterInfo *MRI = &MF.getRegInfo(); 2016 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 2017 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); 2018 i != e; 2019 ++i, ++realArgIdx) { 2020 CCValAssign &VA = ArgLocs[i]; 2021 EVT RegVT = VA.getLocVT(); 2022 SDValue Arg = OutVals[realArgIdx]; 2023 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags; 2024 if (VA.getLocInfo() == CCValAssign::Indirect) 2025 return false; 2026 if (VA.needsCustom()) { 2027 // f64 and vector types are split into multiple registers or 2028 // register/stack-slot combinations. The types will not match 2029 // the registers; give up on memory f64 refs until we figure 2030 // out what to do about this. 2031 if (!VA.isRegLoc()) 2032 return false; 2033 if (!ArgLocs[++i].isRegLoc()) 2034 return false; 2035 if (RegVT == MVT::v2f64) { 2036 if (!ArgLocs[++i].isRegLoc()) 2037 return false; 2038 if (!ArgLocs[++i].isRegLoc()) 2039 return false; 2040 } 2041 } else if (!VA.isRegLoc()) { 2042 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags, 2043 MFI, MRI, TII)) 2044 return false; 2045 } 2046 } 2047 } 2048 } 2049 2050 return true; 2051 } 2052 2053 bool 2054 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv, 2055 MachineFunction &MF, bool isVarArg, 2056 const SmallVectorImpl<ISD::OutputArg> &Outs, 2057 LLVMContext &Context) const { 2058 SmallVector<CCValAssign, 16> RVLocs; 2059 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context); 2060 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true, 2061 isVarArg)); 2062 } 2063 2064 SDValue 2065 ARMTargetLowering::LowerReturn(SDValue Chain, 2066 CallingConv::ID CallConv, bool isVarArg, 2067 const SmallVectorImpl<ISD::OutputArg> &Outs, 2068 const SmallVectorImpl<SDValue> &OutVals, 2069 SDLoc dl, SelectionDAG &DAG) const { 2070 2071 // CCValAssign - represent the assignment of the return value to a location. 2072 SmallVector<CCValAssign, 16> RVLocs; 2073 2074 // CCState - Info about the registers and stack slots. 2075 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 2076 getTargetMachine(), RVLocs, *DAG.getContext(), Call); 2077 2078 // Analyze outgoing return values. 2079 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true, 2080 isVarArg)); 2081 2082 SDValue Flag; 2083 SmallVector<SDValue, 4> RetOps; 2084 RetOps.push_back(Chain); // Operand #0 = Chain (updated below) 2085 2086 // Copy the result values into the output registers. 2087 for (unsigned i = 0, realRVLocIdx = 0; 2088 i != RVLocs.size(); 2089 ++i, ++realRVLocIdx) { 2090 CCValAssign &VA = RVLocs[i]; 2091 assert(VA.isRegLoc() && "Can only return in registers!"); 2092 2093 SDValue Arg = OutVals[realRVLocIdx]; 2094 2095 switch (VA.getLocInfo()) { 2096 default: llvm_unreachable("Unknown loc info!"); 2097 case CCValAssign::Full: break; 2098 case CCValAssign::BCvt: 2099 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg); 2100 break; 2101 } 2102 2103 if (VA.needsCustom()) { 2104 if (VA.getLocVT() == MVT::v2f64) { 2105 // Extract the first half and return it in two registers. 2106 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 2107 DAG.getConstant(0, MVT::i32)); 2108 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl, 2109 DAG.getVTList(MVT::i32, MVT::i32), Half); 2110 2111 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag); 2112 Flag = Chain.getValue(1); 2113 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 2114 VA = RVLocs[++i]; // skip ahead to next loc 2115 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), 2116 HalfGPRs.getValue(1), Flag); 2117 Flag = Chain.getValue(1); 2118 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 2119 VA = RVLocs[++i]; // skip ahead to next loc 2120 2121 // Extract the 2nd half and fall through to handle it as an f64 value. 2122 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 2123 DAG.getConstant(1, MVT::i32)); 2124 } 2125 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is 2126 // available. 2127 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl, 2128 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1); 2129 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag); 2130 Flag = Chain.getValue(1); 2131 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 2132 VA = RVLocs[++i]; // skip ahead to next loc 2133 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1), 2134 Flag); 2135 } else 2136 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag); 2137 2138 // Guarantee that all emitted copies are 2139 // stuck together, avoiding something bad. 2140 Flag = Chain.getValue(1); 2141 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 2142 } 2143 2144 // Update chain and glue. 2145 RetOps[0] = Chain; 2146 if (Flag.getNode()) 2147 RetOps.push_back(Flag); 2148 2149 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, 2150 RetOps.data(), RetOps.size()); 2151 } 2152 2153 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const { 2154 if (N->getNumValues() != 1) 2155 return false; 2156 if (!N->hasNUsesOfValue(1, 0)) 2157 return false; 2158 2159 SDValue TCChain = Chain; 2160 SDNode *Copy = *N->use_begin(); 2161 if (Copy->getOpcode() == ISD::CopyToReg) { 2162 // If the copy has a glue operand, we conservatively assume it isn't safe to 2163 // perform a tail call. 2164 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue) 2165 return false; 2166 TCChain = Copy->getOperand(0); 2167 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) { 2168 SDNode *VMov = Copy; 2169 // f64 returned in a pair of GPRs. 2170 SmallPtrSet<SDNode*, 2> Copies; 2171 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end(); 2172 UI != UE; ++UI) { 2173 if (UI->getOpcode() != ISD::CopyToReg) 2174 return false; 2175 Copies.insert(*UI); 2176 } 2177 if (Copies.size() > 2) 2178 return false; 2179 2180 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end(); 2181 UI != UE; ++UI) { 2182 SDValue UseChain = UI->getOperand(0); 2183 if (Copies.count(UseChain.getNode())) 2184 // Second CopyToReg 2185 Copy = *UI; 2186 else 2187 // First CopyToReg 2188 TCChain = UseChain; 2189 } 2190 } else if (Copy->getOpcode() == ISD::BITCAST) { 2191 // f32 returned in a single GPR. 2192 if (!Copy->hasOneUse()) 2193 return false; 2194 Copy = *Copy->use_begin(); 2195 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0)) 2196 return false; 2197 TCChain = Copy->getOperand(0); 2198 } else { 2199 return false; 2200 } 2201 2202 bool HasRet = false; 2203 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end(); 2204 UI != UE; ++UI) { 2205 if (UI->getOpcode() != ARMISD::RET_FLAG) 2206 return false; 2207 HasRet = true; 2208 } 2209 2210 if (!HasRet) 2211 return false; 2212 2213 Chain = TCChain; 2214 return true; 2215 } 2216 2217 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const { 2218 if (!EnableARMTailCalls && !Subtarget->supportsTailCall()) 2219 return false; 2220 2221 if (!CI->isTailCall()) 2222 return false; 2223 2224 return !Subtarget->isThumb1Only(); 2225 } 2226 2227 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as 2228 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is 2229 // one of the above mentioned nodes. It has to be wrapped because otherwise 2230 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only 2231 // be used to form addressing mode. These wrapped nodes will be selected 2232 // into MOVi. 2233 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) { 2234 EVT PtrVT = Op.getValueType(); 2235 // FIXME there is no actual debug info here 2236 SDLoc dl(Op); 2237 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); 2238 SDValue Res; 2239 if (CP->isMachineConstantPoolEntry()) 2240 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT, 2241 CP->getAlignment()); 2242 else 2243 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, 2244 CP->getAlignment()); 2245 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res); 2246 } 2247 2248 unsigned ARMTargetLowering::getJumpTableEncoding() const { 2249 return MachineJumpTableInfo::EK_Inline; 2250 } 2251 2252 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, 2253 SelectionDAG &DAG) const { 2254 MachineFunction &MF = DAG.getMachineFunction(); 2255 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2256 unsigned ARMPCLabelIndex = 0; 2257 SDLoc DL(Op); 2258 EVT PtrVT = getPointerTy(); 2259 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress(); 2260 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2261 SDValue CPAddr; 2262 if (RelocM == Reloc::Static) { 2263 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4); 2264 } else { 2265 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8; 2266 ARMPCLabelIndex = AFI->createPICLabelUId(); 2267 ARMConstantPoolValue *CPV = 2268 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex, 2269 ARMCP::CPBlockAddress, PCAdj); 2270 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2271 } 2272 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr); 2273 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr, 2274 MachinePointerInfo::getConstantPool(), 2275 false, false, false, 0); 2276 if (RelocM == Reloc::Static) 2277 return Result; 2278 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2279 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel); 2280 } 2281 2282 // Lower ISD::GlobalTLSAddress using the "general dynamic" model 2283 SDValue 2284 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, 2285 SelectionDAG &DAG) const { 2286 SDLoc dl(GA); 2287 EVT PtrVT = getPointerTy(); 2288 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8; 2289 MachineFunction &MF = DAG.getMachineFunction(); 2290 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2291 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2292 ARMConstantPoolValue *CPV = 2293 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex, 2294 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true); 2295 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2296 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument); 2297 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, 2298 MachinePointerInfo::getConstantPool(), 2299 false, false, false, 0); 2300 SDValue Chain = Argument.getValue(1); 2301 2302 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2303 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel); 2304 2305 // call __tls_get_addr. 2306 ArgListTy Args; 2307 ArgListEntry Entry; 2308 Entry.Node = Argument; 2309 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext()); 2310 Args.push_back(Entry); 2311 // FIXME: is there useful debug info available here? 2312 TargetLowering::CallLoweringInfo CLI(Chain, 2313 (Type *) Type::getInt32Ty(*DAG.getContext()), 2314 false, false, false, false, 2315 0, CallingConv::C, /*isTailCall=*/false, 2316 /*doesNotRet=*/false, /*isReturnValueUsed=*/true, 2317 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl); 2318 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); 2319 return CallResult.first; 2320 } 2321 2322 // Lower ISD::GlobalTLSAddress using the "initial exec" or 2323 // "local exec" model. 2324 SDValue 2325 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA, 2326 SelectionDAG &DAG, 2327 TLSModel::Model model) const { 2328 const GlobalValue *GV = GA->getGlobal(); 2329 SDLoc dl(GA); 2330 SDValue Offset; 2331 SDValue Chain = DAG.getEntryNode(); 2332 EVT PtrVT = getPointerTy(); 2333 // Get the Thread Pointer 2334 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT); 2335 2336 if (model == TLSModel::InitialExec) { 2337 MachineFunction &MF = DAG.getMachineFunction(); 2338 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2339 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2340 // Initial exec model. 2341 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8; 2342 ARMConstantPoolValue *CPV = 2343 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex, 2344 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF, 2345 true); 2346 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2347 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset); 2348 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, 2349 MachinePointerInfo::getConstantPool(), 2350 false, false, false, 0); 2351 Chain = Offset.getValue(1); 2352 2353 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2354 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel); 2355 2356 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, 2357 MachinePointerInfo::getConstantPool(), 2358 false, false, false, 0); 2359 } else { 2360 // local exec model 2361 assert(model == TLSModel::LocalExec); 2362 ARMConstantPoolValue *CPV = 2363 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF); 2364 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2365 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset); 2366 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, 2367 MachinePointerInfo::getConstantPool(), 2368 false, false, false, 0); 2369 } 2370 2371 // The address of the thread local variable is the add of the thread 2372 // pointer with the offset of the variable. 2373 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset); 2374 } 2375 2376 SDValue 2377 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const { 2378 // TODO: implement the "local dynamic" model 2379 assert(Subtarget->isTargetELF() && 2380 "TLS not implemented for non-ELF targets"); 2381 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 2382 2383 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal()); 2384 2385 switch (model) { 2386 case TLSModel::GeneralDynamic: 2387 case TLSModel::LocalDynamic: 2388 return LowerToTLSGeneralDynamicModel(GA, DAG); 2389 case TLSModel::InitialExec: 2390 case TLSModel::LocalExec: 2391 return LowerToTLSExecModels(GA, DAG, model); 2392 } 2393 llvm_unreachable("bogus TLS model"); 2394 } 2395 2396 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op, 2397 SelectionDAG &DAG) const { 2398 EVT PtrVT = getPointerTy(); 2399 SDLoc dl(Op); 2400 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 2401 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) { 2402 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility(); 2403 ARMConstantPoolValue *CPV = 2404 ARMConstantPoolConstant::Create(GV, 2405 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT); 2406 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2407 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2408 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), 2409 CPAddr, 2410 MachinePointerInfo::getConstantPool(), 2411 false, false, false, 0); 2412 SDValue Chain = Result.getValue(1); 2413 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT); 2414 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT); 2415 if (!UseGOTOFF) 2416 Result = DAG.getLoad(PtrVT, dl, Chain, Result, 2417 MachinePointerInfo::getGOT(), 2418 false, false, false, 0); 2419 return Result; 2420 } 2421 2422 // If we have T2 ops, we can materialize the address directly via movt/movw 2423 // pair. This is always cheaper. 2424 if (Subtarget->useMovt()) { 2425 ++NumMovwMovt; 2426 // FIXME: Once remat is capable of dealing with instructions with register 2427 // operands, expand this into two nodes. 2428 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT, 2429 DAG.getTargetGlobalAddress(GV, dl, PtrVT)); 2430 } else { 2431 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4); 2432 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2433 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2434 MachinePointerInfo::getConstantPool(), 2435 false, false, false, 0); 2436 } 2437 } 2438 2439 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op, 2440 SelectionDAG &DAG) const { 2441 EVT PtrVT = getPointerTy(); 2442 SDLoc dl(Op); 2443 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 2444 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2445 2446 // FIXME: Enable this for static codegen when tool issues are fixed. Also 2447 // update ARMFastISel::ARMMaterializeGV. 2448 if (Subtarget->useMovt() && RelocM != Reloc::Static) { 2449 ++NumMovwMovt; 2450 // FIXME: Once remat is capable of dealing with instructions with register 2451 // operands, expand this into two nodes. 2452 if (RelocM == Reloc::Static) 2453 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT, 2454 DAG.getTargetGlobalAddress(GV, dl, PtrVT)); 2455 2456 unsigned Wrapper = (RelocM == Reloc::PIC_) 2457 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN; 2458 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, 2459 DAG.getTargetGlobalAddress(GV, dl, PtrVT)); 2460 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) 2461 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result, 2462 MachinePointerInfo::getGOT(), 2463 false, false, false, 0); 2464 return Result; 2465 } 2466 2467 unsigned ARMPCLabelIndex = 0; 2468 SDValue CPAddr; 2469 if (RelocM == Reloc::Static) { 2470 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4); 2471 } else { 2472 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>(); 2473 ARMPCLabelIndex = AFI->createPICLabelUId(); 2474 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8); 2475 ARMConstantPoolValue *CPV = 2476 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 2477 PCAdj); 2478 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2479 } 2480 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2481 2482 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2483 MachinePointerInfo::getConstantPool(), 2484 false, false, false, 0); 2485 SDValue Chain = Result.getValue(1); 2486 2487 if (RelocM == Reloc::PIC_) { 2488 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2489 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 2490 } 2491 2492 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) 2493 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(), 2494 false, false, false, 0); 2495 2496 return Result; 2497 } 2498 2499 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op, 2500 SelectionDAG &DAG) const { 2501 assert(Subtarget->isTargetELF() && 2502 "GLOBAL OFFSET TABLE not implemented for non-ELF targets"); 2503 MachineFunction &MF = DAG.getMachineFunction(); 2504 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2505 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2506 EVT PtrVT = getPointerTy(); 2507 SDLoc dl(Op); 2508 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8; 2509 ARMConstantPoolValue *CPV = 2510 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_", 2511 ARMPCLabelIndex, PCAdj); 2512 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2513 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2514 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2515 MachinePointerInfo::getConstantPool(), 2516 false, false, false, 0); 2517 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2518 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 2519 } 2520 2521 SDValue 2522 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const { 2523 SDLoc dl(Op); 2524 SDValue Val = DAG.getConstant(0, MVT::i32); 2525 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, 2526 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0), 2527 Op.getOperand(1), Val); 2528 } 2529 2530 SDValue 2531 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const { 2532 SDLoc dl(Op); 2533 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0), 2534 Op.getOperand(1), DAG.getConstant(0, MVT::i32)); 2535 } 2536 2537 SDValue 2538 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG, 2539 const ARMSubtarget *Subtarget) const { 2540 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 2541 SDLoc dl(Op); 2542 switch (IntNo) { 2543 default: return SDValue(); // Don't custom lower most intrinsics. 2544 case Intrinsic::arm_thread_pointer: { 2545 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 2546 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT); 2547 } 2548 case Intrinsic::eh_sjlj_lsda: { 2549 MachineFunction &MF = DAG.getMachineFunction(); 2550 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2551 unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); 2552 EVT PtrVT = getPointerTy(); 2553 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 2554 SDValue CPAddr; 2555 unsigned PCAdj = (RelocM != Reloc::PIC_) 2556 ? 0 : (Subtarget->isThumb() ? 4 : 8); 2557 ARMConstantPoolValue *CPV = 2558 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex, 2559 ARMCP::CPLSDA, PCAdj); 2560 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 2561 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 2562 SDValue Result = 2563 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, 2564 MachinePointerInfo::getConstantPool(), 2565 false, false, false, 0); 2566 2567 if (RelocM == Reloc::PIC_) { 2568 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32); 2569 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 2570 } 2571 return Result; 2572 } 2573 case Intrinsic::arm_neon_vmulls: 2574 case Intrinsic::arm_neon_vmullu: { 2575 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls) 2576 ? ARMISD::VMULLs : ARMISD::VMULLu; 2577 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(), 2578 Op.getOperand(1), Op.getOperand(2)); 2579 } 2580 } 2581 } 2582 2583 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG, 2584 const ARMSubtarget *Subtarget) { 2585 // FIXME: handle "fence singlethread" more efficiently. 2586 SDLoc dl(Op); 2587 if (!Subtarget->hasDataBarrier()) { 2588 // Some ARMv6 cpus can support data barriers with an mcr instruction. 2589 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get 2590 // here. 2591 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() && 2592 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!"); 2593 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0), 2594 DAG.getConstant(0, MVT::i32)); 2595 } 2596 2597 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1)); 2598 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue()); 2599 unsigned Domain = ARM_MB::ISH; 2600 if (Subtarget->isSwift() && Ord == Release) { 2601 // Swift happens to implement ISHST barriers in a way that's compatible with 2602 // Release semantics but weaker than ISH so we'd be fools not to use 2603 // it. Beware: other processors probably don't! 2604 Domain = ARM_MB::ISHST; 2605 } 2606 2607 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0), 2608 DAG.getConstant(Domain, MVT::i32)); 2609 } 2610 2611 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG, 2612 const ARMSubtarget *Subtarget) { 2613 // ARM pre v5TE and Thumb1 does not have preload instructions. 2614 if (!(Subtarget->isThumb2() || 2615 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps()))) 2616 // Just preserve the chain. 2617 return Op.getOperand(0); 2618 2619 SDLoc dl(Op); 2620 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1; 2621 if (!isRead && 2622 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension())) 2623 // ARMv7 with MP extension has PLDW. 2624 return Op.getOperand(0); 2625 2626 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue(); 2627 if (Subtarget->isThumb()) { 2628 // Invert the bits. 2629 isRead = ~isRead & 1; 2630 isData = ~isData & 1; 2631 } 2632 2633 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0), 2634 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32), 2635 DAG.getConstant(isData, MVT::i32)); 2636 } 2637 2638 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) { 2639 MachineFunction &MF = DAG.getMachineFunction(); 2640 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>(); 2641 2642 // vastart just stores the address of the VarArgsFrameIndex slot into the 2643 // memory location argument. 2644 SDLoc dl(Op); 2645 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 2646 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); 2647 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 2648 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), 2649 MachinePointerInfo(SV), false, false, 0); 2650 } 2651 2652 SDValue 2653 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA, 2654 SDValue &Root, SelectionDAG &DAG, 2655 SDLoc dl) const { 2656 MachineFunction &MF = DAG.getMachineFunction(); 2657 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2658 2659 const TargetRegisterClass *RC; 2660 if (AFI->isThumb1OnlyFunction()) 2661 RC = &ARM::tGPRRegClass; 2662 else 2663 RC = &ARM::GPRRegClass; 2664 2665 // Transform the arguments stored in physical registers into virtual ones. 2666 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); 2667 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32); 2668 2669 SDValue ArgValue2; 2670 if (NextVA.isMemLoc()) { 2671 MachineFrameInfo *MFI = MF.getFrameInfo(); 2672 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true); 2673 2674 // Create load node to retrieve arguments from the stack. 2675 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2676 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, 2677 MachinePointerInfo::getFixedStack(FI), 2678 false, false, false, 0); 2679 } else { 2680 Reg = MF.addLiveIn(NextVA.getLocReg(), RC); 2681 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32); 2682 } 2683 2684 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2); 2685 } 2686 2687 void 2688 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF, 2689 unsigned InRegsParamRecordIdx, 2690 unsigned ArgSize, 2691 unsigned &ArgRegsSize, 2692 unsigned &ArgRegsSaveSize) 2693 const { 2694 unsigned NumGPRs; 2695 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) { 2696 unsigned RBegin, REnd; 2697 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd); 2698 NumGPRs = REnd - RBegin; 2699 } else { 2700 unsigned int firstUnalloced; 2701 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs, 2702 sizeof(GPRArgRegs) / 2703 sizeof(GPRArgRegs[0])); 2704 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0; 2705 } 2706 2707 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment(); 2708 ArgRegsSize = NumGPRs * 4; 2709 2710 // If parameter is split between stack and GPRs... 2711 if (NumGPRs && Align == 8 && 2712 (ArgRegsSize < ArgSize || 2713 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) { 2714 // Add padding for part of param recovered from GPRs, so 2715 // its last byte must be at address K*8 - 1. 2716 // We need to do it, since remained (stack) part of parameter has 2717 // stack alignment, and we need to "attach" "GPRs head" without gaps 2718 // to it: 2719 // Stack: 2720 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes... 2721 // [ [padding] [GPRs head] ] [ Tail passed via stack .... 2722 // 2723 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2724 unsigned Padding = 2725 ((ArgRegsSize + AFI->getArgRegsSaveSize() + Align - 1) & ~(Align-1)) - 2726 (ArgRegsSize + AFI->getArgRegsSaveSize()); 2727 ArgRegsSaveSize = ArgRegsSize + Padding; 2728 } else 2729 // We don't need to extend regs save size for byval parameters if they 2730 // are passed via GPRs only. 2731 ArgRegsSaveSize = ArgRegsSize; 2732 } 2733 2734 // The remaining GPRs hold either the beginning of variable-argument 2735 // data, or the beginning of an aggregate passed by value (usually 2736 // byval). Either way, we allocate stack slots adjacent to the data 2737 // provided by our caller, and store the unallocated registers there. 2738 // If this is a variadic function, the va_list pointer will begin with 2739 // these values; otherwise, this reassembles a (byval) structure that 2740 // was split between registers and memory. 2741 // Return: The frame index registers were stored into. 2742 int 2743 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG, 2744 SDLoc dl, SDValue &Chain, 2745 const Value *OrigArg, 2746 unsigned InRegsParamRecordIdx, 2747 unsigned OffsetFromOrigArg, 2748 unsigned ArgOffset, 2749 unsigned ArgSize, 2750 bool ForceMutable) const { 2751 2752 // Currently, two use-cases possible: 2753 // Case #1. Non var-args function, and we meet first byval parameter. 2754 // Setup first unallocated register as first byval register; 2755 // eat all remained registers 2756 // (these two actions are performed by HandleByVal method). 2757 // Then, here, we initialize stack frame with 2758 // "store-reg" instructions. 2759 // Case #2. Var-args function, that doesn't contain byval parameters. 2760 // The same: eat all remained unallocated registers, 2761 // initialize stack frame. 2762 2763 MachineFunction &MF = DAG.getMachineFunction(); 2764 MachineFrameInfo *MFI = MF.getFrameInfo(); 2765 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2766 unsigned firstRegToSaveIndex, lastRegToSaveIndex; 2767 unsigned RBegin, REnd; 2768 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) { 2769 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd); 2770 firstRegToSaveIndex = RBegin - ARM::R0; 2771 lastRegToSaveIndex = REnd - ARM::R0; 2772 } else { 2773 firstRegToSaveIndex = CCInfo.getFirstUnallocated 2774 (GPRArgRegs, array_lengthof(GPRArgRegs)); 2775 lastRegToSaveIndex = 4; 2776 } 2777 2778 unsigned ArgRegsSize, ArgRegsSaveSize; 2779 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize, 2780 ArgRegsSize, ArgRegsSaveSize); 2781 2782 // Store any by-val regs to their spots on the stack so that they may be 2783 // loaded by deferencing the result of formal parameter pointer or va_next. 2784 // Note: once stack area for byval/varargs registers 2785 // was initialized, it can't be initialized again. 2786 if (ArgRegsSaveSize) { 2787 2788 unsigned Padding = ArgRegsSaveSize - ArgRegsSize; 2789 2790 if (Padding) { 2791 assert(AFI->getStoredByValParamsPadding() == 0 && 2792 "The only parameter may be padded."); 2793 AFI->setStoredByValParamsPadding(Padding); 2794 } 2795 2796 int FrameIndex = MFI->CreateFixedObject( 2797 ArgRegsSaveSize, 2798 Padding + ArgOffset, 2799 false); 2800 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy()); 2801 2802 SmallVector<SDValue, 4> MemOps; 2803 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex; 2804 ++firstRegToSaveIndex, ++i) { 2805 const TargetRegisterClass *RC; 2806 if (AFI->isThumb1OnlyFunction()) 2807 RC = &ARM::tGPRRegClass; 2808 else 2809 RC = &ARM::GPRRegClass; 2810 2811 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC); 2812 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32); 2813 SDValue Store = 2814 DAG.getStore(Val.getValue(1), dl, Val, FIN, 2815 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i), 2816 false, false, 0); 2817 MemOps.push_back(Store); 2818 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN, 2819 DAG.getConstant(4, getPointerTy())); 2820 } 2821 2822 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize()); 2823 2824 if (!MemOps.empty()) 2825 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 2826 &MemOps[0], MemOps.size()); 2827 return FrameIndex; 2828 } else 2829 // This will point to the next argument passed via stack. 2830 return MFI->CreateFixedObject( 2831 4, AFI->getStoredByValParamsPadding() + ArgOffset, !ForceMutable); 2832 } 2833 2834 // Setup stack frame, the va_list pointer will start from. 2835 void 2836 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG, 2837 SDLoc dl, SDValue &Chain, 2838 unsigned ArgOffset, 2839 bool ForceMutable) const { 2840 MachineFunction &MF = DAG.getMachineFunction(); 2841 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2842 2843 // Try to store any remaining integer argument regs 2844 // to their spots on the stack so that they may be loaded by deferencing 2845 // the result of va_next. 2846 // If there is no regs to be stored, just point address after last 2847 // argument passed via stack. 2848 int FrameIndex = 2849 StoreByValRegs(CCInfo, DAG, dl, Chain, 0, CCInfo.getInRegsParamsCount(), 2850 0, ArgOffset, 0, ForceMutable); 2851 2852 AFI->setVarArgsFrameIndex(FrameIndex); 2853 } 2854 2855 SDValue 2856 ARMTargetLowering::LowerFormalArguments(SDValue Chain, 2857 CallingConv::ID CallConv, bool isVarArg, 2858 const SmallVectorImpl<ISD::InputArg> 2859 &Ins, 2860 SDLoc dl, SelectionDAG &DAG, 2861 SmallVectorImpl<SDValue> &InVals) 2862 const { 2863 MachineFunction &MF = DAG.getMachineFunction(); 2864 MachineFrameInfo *MFI = MF.getFrameInfo(); 2865 2866 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 2867 2868 // Assign locations to all of the incoming arguments. 2869 SmallVector<CCValAssign, 16> ArgLocs; 2870 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), 2871 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue); 2872 CCInfo.AnalyzeFormalArguments(Ins, 2873 CCAssignFnForNode(CallConv, /* Return*/ false, 2874 isVarArg)); 2875 2876 SmallVector<SDValue, 16> ArgValues; 2877 int lastInsIndex = -1; 2878 SDValue ArgValue; 2879 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin(); 2880 unsigned CurArgIdx = 0; 2881 2882 // Initially ArgRegsSaveSize is zero. 2883 // Then we increase this value each time we meet byval parameter. 2884 // We also increase this value in case of varargs function. 2885 AFI->setArgRegsSaveSize(0); 2886 2887 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 2888 CCValAssign &VA = ArgLocs[i]; 2889 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx); 2890 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex; 2891 // Arguments stored in registers. 2892 if (VA.isRegLoc()) { 2893 EVT RegVT = VA.getLocVT(); 2894 2895 if (VA.needsCustom()) { 2896 // f64 and vector types are split up into multiple registers or 2897 // combinations of registers and stack slots. 2898 if (VA.getLocVT() == MVT::v2f64) { 2899 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i], 2900 Chain, DAG, dl); 2901 VA = ArgLocs[++i]; // skip ahead to next loc 2902 SDValue ArgValue2; 2903 if (VA.isMemLoc()) { 2904 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true); 2905 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2906 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN, 2907 MachinePointerInfo::getFixedStack(FI), 2908 false, false, false, 0); 2909 } else { 2910 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i], 2911 Chain, DAG, dl); 2912 } 2913 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64); 2914 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, 2915 ArgValue, ArgValue1, DAG.getIntPtrConstant(0)); 2916 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, 2917 ArgValue, ArgValue2, DAG.getIntPtrConstant(1)); 2918 } else 2919 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl); 2920 2921 } else { 2922 const TargetRegisterClass *RC; 2923 2924 if (RegVT == MVT::f32) 2925 RC = &ARM::SPRRegClass; 2926 else if (RegVT == MVT::f64) 2927 RC = &ARM::DPRRegClass; 2928 else if (RegVT == MVT::v2f64) 2929 RC = &ARM::QPRRegClass; 2930 else if (RegVT == MVT::i32) 2931 RC = AFI->isThumb1OnlyFunction() ? 2932 (const TargetRegisterClass*)&ARM::tGPRRegClass : 2933 (const TargetRegisterClass*)&ARM::GPRRegClass; 2934 else 2935 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering"); 2936 2937 // Transform the arguments in physical registers into virtual ones. 2938 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); 2939 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT); 2940 } 2941 2942 // If this is an 8 or 16-bit value, it is really passed promoted 2943 // to 32 bits. Insert an assert[sz]ext to capture this, then 2944 // truncate to the right size. 2945 switch (VA.getLocInfo()) { 2946 default: llvm_unreachable("Unknown loc info!"); 2947 case CCValAssign::Full: break; 2948 case CCValAssign::BCvt: 2949 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue); 2950 break; 2951 case CCValAssign::SExt: 2952 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue, 2953 DAG.getValueType(VA.getValVT())); 2954 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 2955 break; 2956 case CCValAssign::ZExt: 2957 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue, 2958 DAG.getValueType(VA.getValVT())); 2959 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 2960 break; 2961 } 2962 2963 InVals.push_back(ArgValue); 2964 2965 } else { // VA.isRegLoc() 2966 2967 // sanity check 2968 assert(VA.isMemLoc()); 2969 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered"); 2970 2971 int index = ArgLocs[i].getValNo(); 2972 2973 // Some Ins[] entries become multiple ArgLoc[] entries. 2974 // Process them only once. 2975 if (index != lastInsIndex) 2976 { 2977 ISD::ArgFlagsTy Flags = Ins[index].Flags; 2978 // FIXME: For now, all byval parameter objects are marked mutable. 2979 // This can be changed with more analysis. 2980 // In case of tail call optimization mark all arguments mutable. 2981 // Since they could be overwritten by lowering of arguments in case of 2982 // a tail call. 2983 if (Flags.isByVal()) { 2984 unsigned CurByValIndex = CCInfo.getInRegsParamsProceed(); 2985 int FrameIndex = StoreByValRegs( 2986 CCInfo, DAG, dl, Chain, CurOrigArg, 2987 CurByValIndex, 2988 Ins[VA.getValNo()].PartOffset, 2989 VA.getLocMemOffset(), 2990 Flags.getByValSize(), 2991 true /*force mutable frames*/); 2992 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy())); 2993 CCInfo.nextInRegsParam(); 2994 } else { 2995 unsigned FIOffset = VA.getLocMemOffset() + 2996 AFI->getStoredByValParamsPadding(); 2997 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8, 2998 FIOffset, true); 2999 3000 // Create load nodes to retrieve arguments from the stack. 3001 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 3002 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, 3003 MachinePointerInfo::getFixedStack(FI), 3004 false, false, false, 0)); 3005 } 3006 lastInsIndex = index; 3007 } 3008 } 3009 } 3010 3011 // varargs 3012 if (isVarArg) 3013 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 3014 CCInfo.getNextStackOffset()); 3015 3016 return Chain; 3017 } 3018 3019 /// isFloatingPointZero - Return true if this is +0.0. 3020 static bool isFloatingPointZero(SDValue Op) { 3021 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op)) 3022 return CFP->getValueAPF().isPosZero(); 3023 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) { 3024 // Maybe this has already been legalized into the constant pool? 3025 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) { 3026 SDValue WrapperOp = Op.getOperand(1).getOperand(0); 3027 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp)) 3028 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal())) 3029 return CFP->getValueAPF().isPosZero(); 3030 } 3031 } 3032 return false; 3033 } 3034 3035 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for 3036 /// the given operands. 3037 SDValue 3038 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC, 3039 SDValue &ARMcc, SelectionDAG &DAG, 3040 SDLoc dl) const { 3041 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) { 3042 unsigned C = RHSC->getZExtValue(); 3043 if (!isLegalICmpImmediate(C)) { 3044 // Constant does not fit, try adjusting it by one? 3045 switch (CC) { 3046 default: break; 3047 case ISD::SETLT: 3048 case ISD::SETGE: 3049 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) { 3050 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT; 3051 RHS = DAG.getConstant(C-1, MVT::i32); 3052 } 3053 break; 3054 case ISD::SETULT: 3055 case ISD::SETUGE: 3056 if (C != 0 && isLegalICmpImmediate(C-1)) { 3057 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT; 3058 RHS = DAG.getConstant(C-1, MVT::i32); 3059 } 3060 break; 3061 case ISD::SETLE: 3062 case ISD::SETGT: 3063 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) { 3064 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE; 3065 RHS = DAG.getConstant(C+1, MVT::i32); 3066 } 3067 break; 3068 case ISD::SETULE: 3069 case ISD::SETUGT: 3070 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) { 3071 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; 3072 RHS = DAG.getConstant(C+1, MVT::i32); 3073 } 3074 break; 3075 } 3076 } 3077 } 3078 3079 ARMCC::CondCodes CondCode = IntCCToARMCC(CC); 3080 ARMISD::NodeType CompareType; 3081 switch (CondCode) { 3082 default: 3083 CompareType = ARMISD::CMP; 3084 break; 3085 case ARMCC::EQ: 3086 case ARMCC::NE: 3087 // Uses only Z Flag 3088 CompareType = ARMISD::CMPZ; 3089 break; 3090 } 3091 ARMcc = DAG.getConstant(CondCode, MVT::i32); 3092 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS); 3093 } 3094 3095 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands. 3096 SDValue 3097 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG, 3098 SDLoc dl) const { 3099 SDValue Cmp; 3100 if (!isFloatingPointZero(RHS)) 3101 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS); 3102 else 3103 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS); 3104 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp); 3105 } 3106 3107 /// duplicateCmp - Glue values can have only one use, so this function 3108 /// duplicates a comparison node. 3109 SDValue 3110 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const { 3111 unsigned Opc = Cmp.getOpcode(); 3112 SDLoc DL(Cmp); 3113 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ) 3114 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1)); 3115 3116 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation"); 3117 Cmp = Cmp.getOperand(0); 3118 Opc = Cmp.getOpcode(); 3119 if (Opc == ARMISD::CMPFP) 3120 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1)); 3121 else { 3122 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT"); 3123 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0)); 3124 } 3125 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp); 3126 } 3127 3128 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const { 3129 SDValue Cond = Op.getOperand(0); 3130 SDValue SelectTrue = Op.getOperand(1); 3131 SDValue SelectFalse = Op.getOperand(2); 3132 SDLoc dl(Op); 3133 3134 // Convert: 3135 // 3136 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond) 3137 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond) 3138 // 3139 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) { 3140 const ConstantSDNode *CMOVTrue = 3141 dyn_cast<ConstantSDNode>(Cond.getOperand(0)); 3142 const ConstantSDNode *CMOVFalse = 3143 dyn_cast<ConstantSDNode>(Cond.getOperand(1)); 3144 3145 if (CMOVTrue && CMOVFalse) { 3146 unsigned CMOVTrueVal = CMOVTrue->getZExtValue(); 3147 unsigned CMOVFalseVal = CMOVFalse->getZExtValue(); 3148 3149 SDValue True; 3150 SDValue False; 3151 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) { 3152 True = SelectTrue; 3153 False = SelectFalse; 3154 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) { 3155 True = SelectFalse; 3156 False = SelectTrue; 3157 } 3158 3159 if (True.getNode() && False.getNode()) { 3160 EVT VT = Op.getValueType(); 3161 SDValue ARMcc = Cond.getOperand(2); 3162 SDValue CCR = Cond.getOperand(3); 3163 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG); 3164 assert(True.getValueType() == VT); 3165 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp); 3166 } 3167 } 3168 } 3169 3170 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the 3171 // undefined bits before doing a full-word comparison with zero. 3172 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond, 3173 DAG.getConstant(1, Cond.getValueType())); 3174 3175 return DAG.getSelectCC(dl, Cond, 3176 DAG.getConstant(0, Cond.getValueType()), 3177 SelectTrue, SelectFalse, ISD::SETNE); 3178 } 3179 3180 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const { 3181 EVT VT = Op.getValueType(); 3182 SDValue LHS = Op.getOperand(0); 3183 SDValue RHS = Op.getOperand(1); 3184 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get(); 3185 SDValue TrueVal = Op.getOperand(2); 3186 SDValue FalseVal = Op.getOperand(3); 3187 SDLoc dl(Op); 3188 3189 if (LHS.getValueType() == MVT::i32) { 3190 SDValue ARMcc; 3191 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3192 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl); 3193 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp); 3194 } 3195 3196 ARMCC::CondCodes CondCode, CondCode2; 3197 FPCCToARMCC(CC, CondCode, CondCode2); 3198 3199 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32); 3200 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl); 3201 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3202 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, 3203 ARMcc, CCR, Cmp); 3204 if (CondCode2 != ARMCC::AL) { 3205 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32); 3206 // FIXME: Needs another CMP because flag can have but one use. 3207 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl); 3208 Result = DAG.getNode(ARMISD::CMOV, dl, VT, 3209 Result, TrueVal, ARMcc2, CCR, Cmp2); 3210 } 3211 return Result; 3212 } 3213 3214 /// canChangeToInt - Given the fp compare operand, return true if it is suitable 3215 /// to morph to an integer compare sequence. 3216 static bool canChangeToInt(SDValue Op, bool &SeenZero, 3217 const ARMSubtarget *Subtarget) { 3218 SDNode *N = Op.getNode(); 3219 if (!N->hasOneUse()) 3220 // Otherwise it requires moving the value from fp to integer registers. 3221 return false; 3222 if (!N->getNumValues()) 3223 return false; 3224 EVT VT = Op.getValueType(); 3225 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow()) 3226 // f32 case is generally profitable. f64 case only makes sense when vcmpe + 3227 // vmrs are very slow, e.g. cortex-a8. 3228 return false; 3229 3230 if (isFloatingPointZero(Op)) { 3231 SeenZero = true; 3232 return true; 3233 } 3234 return ISD::isNormalLoad(N); 3235 } 3236 3237 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) { 3238 if (isFloatingPointZero(Op)) 3239 return DAG.getConstant(0, MVT::i32); 3240 3241 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) 3242 return DAG.getLoad(MVT::i32, SDLoc(Op), 3243 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(), 3244 Ld->isVolatile(), Ld->isNonTemporal(), 3245 Ld->isInvariant(), Ld->getAlignment()); 3246 3247 llvm_unreachable("Unknown VFP cmp argument!"); 3248 } 3249 3250 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG, 3251 SDValue &RetVal1, SDValue &RetVal2) { 3252 if (isFloatingPointZero(Op)) { 3253 RetVal1 = DAG.getConstant(0, MVT::i32); 3254 RetVal2 = DAG.getConstant(0, MVT::i32); 3255 return; 3256 } 3257 3258 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) { 3259 SDValue Ptr = Ld->getBasePtr(); 3260 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op), 3261 Ld->getChain(), Ptr, 3262 Ld->getPointerInfo(), 3263 Ld->isVolatile(), Ld->isNonTemporal(), 3264 Ld->isInvariant(), Ld->getAlignment()); 3265 3266 EVT PtrType = Ptr.getValueType(); 3267 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4); 3268 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op), 3269 PtrType, Ptr, DAG.getConstant(4, PtrType)); 3270 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op), 3271 Ld->getChain(), NewPtr, 3272 Ld->getPointerInfo().getWithOffset(4), 3273 Ld->isVolatile(), Ld->isNonTemporal(), 3274 Ld->isInvariant(), NewAlign); 3275 return; 3276 } 3277 3278 llvm_unreachable("Unknown VFP cmp argument!"); 3279 } 3280 3281 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some 3282 /// f32 and even f64 comparisons to integer ones. 3283 SDValue 3284 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const { 3285 SDValue Chain = Op.getOperand(0); 3286 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 3287 SDValue LHS = Op.getOperand(2); 3288 SDValue RHS = Op.getOperand(3); 3289 SDValue Dest = Op.getOperand(4); 3290 SDLoc dl(Op); 3291 3292 bool LHSSeenZero = false; 3293 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget); 3294 bool RHSSeenZero = false; 3295 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget); 3296 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) { 3297 // If unsafe fp math optimization is enabled and there are no other uses of 3298 // the CMP operands, and the condition code is EQ or NE, we can optimize it 3299 // to an integer comparison. 3300 if (CC == ISD::SETOEQ) 3301 CC = ISD::SETEQ; 3302 else if (CC == ISD::SETUNE) 3303 CC = ISD::SETNE; 3304 3305 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32); 3306 SDValue ARMcc; 3307 if (LHS.getValueType() == MVT::f32) { 3308 LHS = DAG.getNode(ISD::AND, dl, MVT::i32, 3309 bitcastf32Toi32(LHS, DAG), Mask); 3310 RHS = DAG.getNode(ISD::AND, dl, MVT::i32, 3311 bitcastf32Toi32(RHS, DAG), Mask); 3312 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl); 3313 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3314 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other, 3315 Chain, Dest, ARMcc, CCR, Cmp); 3316 } 3317 3318 SDValue LHS1, LHS2; 3319 SDValue RHS1, RHS2; 3320 expandf64Toi32(LHS, DAG, LHS1, LHS2); 3321 expandf64Toi32(RHS, DAG, RHS1, RHS2); 3322 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask); 3323 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask); 3324 ARMCC::CondCodes CondCode = IntCCToARMCC(CC); 3325 ARMcc = DAG.getConstant(CondCode, MVT::i32); 3326 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue); 3327 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest }; 3328 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7); 3329 } 3330 3331 return SDValue(); 3332 } 3333 3334 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const { 3335 SDValue Chain = Op.getOperand(0); 3336 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 3337 SDValue LHS = Op.getOperand(2); 3338 SDValue RHS = Op.getOperand(3); 3339 SDValue Dest = Op.getOperand(4); 3340 SDLoc dl(Op); 3341 3342 if (LHS.getValueType() == MVT::i32) { 3343 SDValue ARMcc; 3344 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl); 3345 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3346 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other, 3347 Chain, Dest, ARMcc, CCR, Cmp); 3348 } 3349 3350 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64); 3351 3352 if (getTargetMachine().Options.UnsafeFPMath && 3353 (CC == ISD::SETEQ || CC == ISD::SETOEQ || 3354 CC == ISD::SETNE || CC == ISD::SETUNE)) { 3355 SDValue Result = OptimizeVFPBrcond(Op, DAG); 3356 if (Result.getNode()) 3357 return Result; 3358 } 3359 3360 ARMCC::CondCodes CondCode, CondCode2; 3361 FPCCToARMCC(CC, CondCode, CondCode2); 3362 3363 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32); 3364 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl); 3365 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3366 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue); 3367 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp }; 3368 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5); 3369 if (CondCode2 != ARMCC::AL) { 3370 ARMcc = DAG.getConstant(CondCode2, MVT::i32); 3371 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) }; 3372 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5); 3373 } 3374 return Res; 3375 } 3376 3377 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const { 3378 SDValue Chain = Op.getOperand(0); 3379 SDValue Table = Op.getOperand(1); 3380 SDValue Index = Op.getOperand(2); 3381 SDLoc dl(Op); 3382 3383 EVT PTy = getPointerTy(); 3384 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table); 3385 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>(); 3386 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy); 3387 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy); 3388 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId); 3389 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy)); 3390 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table); 3391 if (Subtarget->isThumb2()) { 3392 // Thumb2 uses a two-level jump. That is, it jumps into the jump table 3393 // which does another jump to the destination. This also makes it easier 3394 // to translate it to TBB / TBH later. 3395 // FIXME: This might not work if the function is extremely large. 3396 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain, 3397 Addr, Op.getOperand(2), JTI, UId); 3398 } 3399 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) { 3400 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, 3401 MachinePointerInfo::getJumpTable(), 3402 false, false, false, 0); 3403 Chain = Addr.getValue(1); 3404 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table); 3405 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId); 3406 } else { 3407 Addr = DAG.getLoad(PTy, dl, Chain, Addr, 3408 MachinePointerInfo::getJumpTable(), 3409 false, false, false, 0); 3410 Chain = Addr.getValue(1); 3411 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId); 3412 } 3413 } 3414 3415 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) { 3416 EVT VT = Op.getValueType(); 3417 SDLoc dl(Op); 3418 3419 if (Op.getValueType().getVectorElementType() == MVT::i32) { 3420 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32) 3421 return Op; 3422 return DAG.UnrollVectorOp(Op.getNode()); 3423 } 3424 3425 assert(Op.getOperand(0).getValueType() == MVT::v4f32 && 3426 "Invalid type for custom lowering!"); 3427 if (VT != MVT::v4i16) 3428 return DAG.UnrollVectorOp(Op.getNode()); 3429 3430 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0)); 3431 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op); 3432 } 3433 3434 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) { 3435 EVT VT = Op.getValueType(); 3436 if (VT.isVector()) 3437 return LowerVectorFP_TO_INT(Op, DAG); 3438 3439 SDLoc dl(Op); 3440 unsigned Opc; 3441 3442 switch (Op.getOpcode()) { 3443 default: llvm_unreachable("Invalid opcode!"); 3444 case ISD::FP_TO_SINT: 3445 Opc = ARMISD::FTOSI; 3446 break; 3447 case ISD::FP_TO_UINT: 3448 Opc = ARMISD::FTOUI; 3449 break; 3450 } 3451 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0)); 3452 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op); 3453 } 3454 3455 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) { 3456 EVT VT = Op.getValueType(); 3457 SDLoc dl(Op); 3458 3459 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) { 3460 if (VT.getVectorElementType() == MVT::f32) 3461 return Op; 3462 return DAG.UnrollVectorOp(Op.getNode()); 3463 } 3464 3465 assert(Op.getOperand(0).getValueType() == MVT::v4i16 && 3466 "Invalid type for custom lowering!"); 3467 if (VT != MVT::v4f32) 3468 return DAG.UnrollVectorOp(Op.getNode()); 3469 3470 unsigned CastOpc; 3471 unsigned Opc; 3472 switch (Op.getOpcode()) { 3473 default: llvm_unreachable("Invalid opcode!"); 3474 case ISD::SINT_TO_FP: 3475 CastOpc = ISD::SIGN_EXTEND; 3476 Opc = ISD::SINT_TO_FP; 3477 break; 3478 case ISD::UINT_TO_FP: 3479 CastOpc = ISD::ZERO_EXTEND; 3480 Opc = ISD::UINT_TO_FP; 3481 break; 3482 } 3483 3484 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0)); 3485 return DAG.getNode(Opc, dl, VT, Op); 3486 } 3487 3488 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) { 3489 EVT VT = Op.getValueType(); 3490 if (VT.isVector()) 3491 return LowerVectorINT_TO_FP(Op, DAG); 3492 3493 SDLoc dl(Op); 3494 unsigned Opc; 3495 3496 switch (Op.getOpcode()) { 3497 default: llvm_unreachable("Invalid opcode!"); 3498 case ISD::SINT_TO_FP: 3499 Opc = ARMISD::SITOF; 3500 break; 3501 case ISD::UINT_TO_FP: 3502 Opc = ARMISD::UITOF; 3503 break; 3504 } 3505 3506 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0)); 3507 return DAG.getNode(Opc, dl, VT, Op); 3508 } 3509 3510 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { 3511 // Implement fcopysign with a fabs and a conditional fneg. 3512 SDValue Tmp0 = Op.getOperand(0); 3513 SDValue Tmp1 = Op.getOperand(1); 3514 SDLoc dl(Op); 3515 EVT VT = Op.getValueType(); 3516 EVT SrcVT = Tmp1.getValueType(); 3517 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST || 3518 Tmp0.getOpcode() == ARMISD::VMOVDRR; 3519 bool UseNEON = !InGPR && Subtarget->hasNEON(); 3520 3521 if (UseNEON) { 3522 // Use VBSL to copy the sign bit. 3523 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80); 3524 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32, 3525 DAG.getTargetConstant(EncodedVal, MVT::i32)); 3526 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64; 3527 if (VT == MVT::f64) 3528 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT, 3529 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask), 3530 DAG.getConstant(32, MVT::i32)); 3531 else /*if (VT == MVT::f32)*/ 3532 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0); 3533 if (SrcVT == MVT::f32) { 3534 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1); 3535 if (VT == MVT::f64) 3536 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT, 3537 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1), 3538 DAG.getConstant(32, MVT::i32)); 3539 } else if (VT == MVT::f32) 3540 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64, 3541 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1), 3542 DAG.getConstant(32, MVT::i32)); 3543 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0); 3544 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1); 3545 3546 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff), 3547 MVT::i32); 3548 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes); 3549 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask, 3550 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes)); 3551 3552 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT, 3553 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask), 3554 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot)); 3555 if (VT == MVT::f32) { 3556 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res); 3557 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res, 3558 DAG.getConstant(0, MVT::i32)); 3559 } else { 3560 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res); 3561 } 3562 3563 return Res; 3564 } 3565 3566 // Bitcast operand 1 to i32. 3567 if (SrcVT == MVT::f64) 3568 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32), 3569 &Tmp1, 1).getValue(1); 3570 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1); 3571 3572 // Or in the signbit with integer operations. 3573 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32); 3574 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32); 3575 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1); 3576 if (VT == MVT::f32) { 3577 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32, 3578 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2); 3579 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, 3580 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1)); 3581 } 3582 3583 // f64: Or the high part with signbit and then combine two parts. 3584 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32), 3585 &Tmp0, 1); 3586 SDValue Lo = Tmp0.getValue(0); 3587 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2); 3588 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1); 3589 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi); 3590 } 3591 3592 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{ 3593 MachineFunction &MF = DAG.getMachineFunction(); 3594 MachineFrameInfo *MFI = MF.getFrameInfo(); 3595 MFI->setReturnAddressIsTaken(true); 3596 3597 EVT VT = Op.getValueType(); 3598 SDLoc dl(Op); 3599 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 3600 if (Depth) { 3601 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG); 3602 SDValue Offset = DAG.getConstant(4, MVT::i32); 3603 return DAG.getLoad(VT, dl, DAG.getEntryNode(), 3604 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset), 3605 MachinePointerInfo(), false, false, false, 0); 3606 } 3607 3608 // Return LR, which contains the return address. Mark it an implicit live-in. 3609 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32)); 3610 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT); 3611 } 3612 3613 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { 3614 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 3615 MFI->setFrameAddressIsTaken(true); 3616 3617 EVT VT = Op.getValueType(); 3618 SDLoc dl(Op); // FIXME probably not meaningful 3619 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 3620 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin()) 3621 ? ARM::R7 : ARM::R11; 3622 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT); 3623 while (Depth--) 3624 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, 3625 MachinePointerInfo(), 3626 false, false, false, 0); 3627 return FrameAddr; 3628 } 3629 3630 /// Custom Expand long vector extensions, where size(DestVec) > 2*size(SrcVec), 3631 /// and size(DestVec) > 128-bits. 3632 /// This is achieved by doing the one extension from the SrcVec, splitting the 3633 /// result, extending these parts, and then concatenating these into the 3634 /// destination. 3635 static SDValue ExpandVectorExtension(SDNode *N, SelectionDAG &DAG) { 3636 SDValue Op = N->getOperand(0); 3637 EVT SrcVT = Op.getValueType(); 3638 EVT DestVT = N->getValueType(0); 3639 3640 assert(DestVT.getSizeInBits() > 128 && 3641 "Custom sext/zext expansion needs >128-bit vector."); 3642 // If this is a normal length extension, use the default expansion. 3643 if (SrcVT.getSizeInBits()*4 != DestVT.getSizeInBits() && 3644 SrcVT.getSizeInBits()*8 != DestVT.getSizeInBits()) 3645 return SDValue(); 3646 3647 SDLoc dl(N); 3648 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits(); 3649 unsigned DestEltSize = DestVT.getVectorElementType().getSizeInBits(); 3650 unsigned NumElts = SrcVT.getVectorNumElements(); 3651 LLVMContext &Ctx = *DAG.getContext(); 3652 SDValue Mid, SplitLo, SplitHi, ExtLo, ExtHi; 3653 3654 EVT MidVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2), 3655 NumElts); 3656 EVT SplitVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2), 3657 NumElts/2); 3658 EVT ExtVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, DestEltSize), 3659 NumElts/2); 3660 3661 Mid = DAG.getNode(N->getOpcode(), dl, MidVT, Op); 3662 SplitLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid, 3663 DAG.getIntPtrConstant(0)); 3664 SplitHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid, 3665 DAG.getIntPtrConstant(NumElts/2)); 3666 ExtLo = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitLo); 3667 ExtHi = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitHi); 3668 return DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, ExtLo, ExtHi); 3669 } 3670 3671 /// ExpandBITCAST - If the target supports VFP, this function is called to 3672 /// expand a bit convert where either the source or destination type is i64 to 3673 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64 3674 /// operand type is illegal (e.g., v2f32 for a target that doesn't support 3675 /// vectors), since the legalizer won't know what to do with that. 3676 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) { 3677 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3678 SDLoc dl(N); 3679 SDValue Op = N->getOperand(0); 3680 3681 // This function is only supposed to be called for i64 types, either as the 3682 // source or destination of the bit convert. 3683 EVT SrcVT = Op.getValueType(); 3684 EVT DstVT = N->getValueType(0); 3685 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) && 3686 "ExpandBITCAST called for non-i64 type"); 3687 3688 // Turn i64->f64 into VMOVDRR. 3689 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) { 3690 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op, 3691 DAG.getConstant(0, MVT::i32)); 3692 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op, 3693 DAG.getConstant(1, MVT::i32)); 3694 return DAG.getNode(ISD::BITCAST, dl, DstVT, 3695 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi)); 3696 } 3697 3698 // Turn f64->i64 into VMOVRRD. 3699 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) { 3700 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl, 3701 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1); 3702 // Merge the pieces into a single i64 value. 3703 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1)); 3704 } 3705 3706 return SDValue(); 3707 } 3708 3709 /// getZeroVector - Returns a vector of specified type with all zero elements. 3710 /// Zero vectors are used to represent vector negation and in those cases 3711 /// will be implemented with the NEON VNEG instruction. However, VNEG does 3712 /// not support i64 elements, so sometimes the zero vectors will need to be 3713 /// explicitly constructed. Regardless, use a canonical VMOV to create the 3714 /// zero vector. 3715 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) { 3716 assert(VT.isVector() && "Expected a vector type"); 3717 // The canonical modified immediate encoding of a zero vector is....0! 3718 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32); 3719 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32; 3720 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal); 3721 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov); 3722 } 3723 3724 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two 3725 /// i32 values and take a 2 x i32 value to shift plus a shift amount. 3726 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op, 3727 SelectionDAG &DAG) const { 3728 assert(Op.getNumOperands() == 3 && "Not a double-shift!"); 3729 EVT VT = Op.getValueType(); 3730 unsigned VTBits = VT.getSizeInBits(); 3731 SDLoc dl(Op); 3732 SDValue ShOpLo = Op.getOperand(0); 3733 SDValue ShOpHi = Op.getOperand(1); 3734 SDValue ShAmt = Op.getOperand(2); 3735 SDValue ARMcc; 3736 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL; 3737 3738 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS); 3739 3740 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, 3741 DAG.getConstant(VTBits, MVT::i32), ShAmt); 3742 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt); 3743 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt, 3744 DAG.getConstant(VTBits, MVT::i32)); 3745 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt); 3746 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2); 3747 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt); 3748 3749 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3750 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE, 3751 ARMcc, DAG, dl); 3752 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt); 3753 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, 3754 CCR, Cmp); 3755 3756 SDValue Ops[2] = { Lo, Hi }; 3757 return DAG.getMergeValues(Ops, 2, dl); 3758 } 3759 3760 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two 3761 /// i32 values and take a 2 x i32 value to shift plus a shift amount. 3762 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op, 3763 SelectionDAG &DAG) const { 3764 assert(Op.getNumOperands() == 3 && "Not a double-shift!"); 3765 EVT VT = Op.getValueType(); 3766 unsigned VTBits = VT.getSizeInBits(); 3767 SDLoc dl(Op); 3768 SDValue ShOpLo = Op.getOperand(0); 3769 SDValue ShOpHi = Op.getOperand(1); 3770 SDValue ShAmt = Op.getOperand(2); 3771 SDValue ARMcc; 3772 3773 assert(Op.getOpcode() == ISD::SHL_PARTS); 3774 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, 3775 DAG.getConstant(VTBits, MVT::i32), ShAmt); 3776 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt); 3777 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt, 3778 DAG.getConstant(VTBits, MVT::i32)); 3779 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt); 3780 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt); 3781 3782 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2); 3783 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 3784 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE, 3785 ARMcc, DAG, dl); 3786 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt); 3787 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc, 3788 CCR, Cmp); 3789 3790 SDValue Ops[2] = { Lo, Hi }; 3791 return DAG.getMergeValues(Ops, 2, dl); 3792 } 3793 3794 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op, 3795 SelectionDAG &DAG) const { 3796 // The rounding mode is in bits 23:22 of the FPSCR. 3797 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0 3798 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3) 3799 // so that the shift + and get folded into a bitfield extract. 3800 SDLoc dl(Op); 3801 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32, 3802 DAG.getConstant(Intrinsic::arm_get_fpscr, 3803 MVT::i32)); 3804 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR, 3805 DAG.getConstant(1U << 22, MVT::i32)); 3806 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds, 3807 DAG.getConstant(22, MVT::i32)); 3808 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE, 3809 DAG.getConstant(3, MVT::i32)); 3810 } 3811 3812 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG, 3813 const ARMSubtarget *ST) { 3814 EVT VT = N->getValueType(0); 3815 SDLoc dl(N); 3816 3817 if (!ST->hasV6T2Ops()) 3818 return SDValue(); 3819 3820 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0)); 3821 return DAG.getNode(ISD::CTLZ, dl, VT, rbit); 3822 } 3823 3824 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count 3825 /// for each 16-bit element from operand, repeated. The basic idea is to 3826 /// leverage vcnt to get the 8-bit counts, gather and add the results. 3827 /// 3828 /// Trace for v4i16: 3829 /// input = [v0 v1 v2 v3 ] (vi 16-bit element) 3830 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element) 3831 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi) 3832 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6] 3833 /// [b0 b1 b2 b3 b4 b5 b6 b7] 3834 /// +[b1 b0 b3 b2 b5 b4 b7 b6] 3835 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0, 3836 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits) 3837 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) { 3838 EVT VT = N->getValueType(0); 3839 SDLoc DL(N); 3840 3841 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8; 3842 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0)); 3843 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0); 3844 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1); 3845 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2); 3846 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3); 3847 } 3848 3849 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the 3850 /// bit-count for each 16-bit element from the operand. We need slightly 3851 /// different sequencing for v4i16 and v8i16 to stay within NEON's available 3852 /// 64/128-bit registers. 3853 /// 3854 /// Trace for v4i16: 3855 /// input = [v0 v1 v2 v3 ] (vi 16-bit element) 3856 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi) 3857 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ] 3858 /// v4i16:Extracted = [k0 k1 k2 k3 ] 3859 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) { 3860 EVT VT = N->getValueType(0); 3861 SDLoc DL(N); 3862 3863 SDValue BitCounts = getCTPOP16BitCounts(N, DAG); 3864 if (VT.is64BitVector()) { 3865 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts); 3866 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended, 3867 DAG.getIntPtrConstant(0)); 3868 } else { 3869 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8, 3870 BitCounts, DAG.getIntPtrConstant(0)); 3871 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted); 3872 } 3873 } 3874 3875 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the 3876 /// bit-count for each 32-bit element from the operand. The idea here is 3877 /// to split the vector into 16-bit elements, leverage the 16-bit count 3878 /// routine, and then combine the results. 3879 /// 3880 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged): 3881 /// input = [v0 v1 ] (vi: 32-bit elements) 3882 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1]) 3883 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi) 3884 /// vrev: N0 = [k1 k0 k3 k2 ] 3885 /// [k0 k1 k2 k3 ] 3886 /// N1 =+[k1 k0 k3 k2 ] 3887 /// [k0 k2 k1 k3 ] 3888 /// N2 =+[k1 k3 k0 k2 ] 3889 /// [k0 k2 k1 k3 ] 3890 /// Extended =+[k1 k3 k0 k2 ] 3891 /// [k0 k2 ] 3892 /// Extracted=+[k1 k3 ] 3893 /// 3894 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) { 3895 EVT VT = N->getValueType(0); 3896 SDLoc DL(N); 3897 3898 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16; 3899 3900 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0)); 3901 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG); 3902 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16); 3903 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0); 3904 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1); 3905 3906 if (VT.is64BitVector()) { 3907 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2); 3908 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended, 3909 DAG.getIntPtrConstant(0)); 3910 } else { 3911 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2, 3912 DAG.getIntPtrConstant(0)); 3913 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted); 3914 } 3915 } 3916 3917 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG, 3918 const ARMSubtarget *ST) { 3919 EVT VT = N->getValueType(0); 3920 3921 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON."); 3922 assert((VT == MVT::v2i32 || VT == MVT::v4i32 || 3923 VT == MVT::v4i16 || VT == MVT::v8i16) && 3924 "Unexpected type for custom ctpop lowering"); 3925 3926 if (VT.getVectorElementType() == MVT::i32) 3927 return lowerCTPOP32BitElements(N, DAG); 3928 else 3929 return lowerCTPOP16BitElements(N, DAG); 3930 } 3931 3932 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG, 3933 const ARMSubtarget *ST) { 3934 EVT VT = N->getValueType(0); 3935 SDLoc dl(N); 3936 3937 if (!VT.isVector()) 3938 return SDValue(); 3939 3940 // Lower vector shifts on NEON to use VSHL. 3941 assert(ST->hasNEON() && "unexpected vector shift"); 3942 3943 // Left shifts translate directly to the vshiftu intrinsic. 3944 if (N->getOpcode() == ISD::SHL) 3945 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, 3946 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32), 3947 N->getOperand(0), N->getOperand(1)); 3948 3949 assert((N->getOpcode() == ISD::SRA || 3950 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode"); 3951 3952 // NEON uses the same intrinsics for both left and right shifts. For 3953 // right shifts, the shift amounts are negative, so negate the vector of 3954 // shift amounts. 3955 EVT ShiftVT = N->getOperand(1).getValueType(); 3956 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT, 3957 getZeroVector(ShiftVT, DAG, dl), 3958 N->getOperand(1)); 3959 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ? 3960 Intrinsic::arm_neon_vshifts : 3961 Intrinsic::arm_neon_vshiftu); 3962 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, 3963 DAG.getConstant(vshiftInt, MVT::i32), 3964 N->getOperand(0), NegatedCount); 3965 } 3966 3967 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG, 3968 const ARMSubtarget *ST) { 3969 EVT VT = N->getValueType(0); 3970 SDLoc dl(N); 3971 3972 // We can get here for a node like i32 = ISD::SHL i32, i64 3973 if (VT != MVT::i64) 3974 return SDValue(); 3975 3976 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) && 3977 "Unknown shift to lower!"); 3978 3979 // We only lower SRA, SRL of 1 here, all others use generic lowering. 3980 if (!isa<ConstantSDNode>(N->getOperand(1)) || 3981 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1) 3982 return SDValue(); 3983 3984 // If we are in thumb mode, we don't have RRX. 3985 if (ST->isThumb1Only()) return SDValue(); 3986 3987 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr. 3988 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0), 3989 DAG.getConstant(0, MVT::i32)); 3990 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0), 3991 DAG.getConstant(1, MVT::i32)); 3992 3993 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and 3994 // captures the result into a carry flag. 3995 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG; 3996 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1); 3997 3998 // The low part is an ARMISD::RRX operand, which shifts the carry in. 3999 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1)); 4000 4001 // Merge the pieces into a single i64 value. 4002 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi); 4003 } 4004 4005 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) { 4006 SDValue TmpOp0, TmpOp1; 4007 bool Invert = false; 4008 bool Swap = false; 4009 unsigned Opc = 0; 4010 4011 SDValue Op0 = Op.getOperand(0); 4012 SDValue Op1 = Op.getOperand(1); 4013 SDValue CC = Op.getOperand(2); 4014 EVT VT = Op.getValueType(); 4015 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get(); 4016 SDLoc dl(Op); 4017 4018 if (Op.getOperand(1).getValueType().isFloatingPoint()) { 4019 switch (SetCCOpcode) { 4020 default: llvm_unreachable("Illegal FP comparison"); 4021 case ISD::SETUNE: 4022 case ISD::SETNE: Invert = true; // Fallthrough 4023 case ISD::SETOEQ: 4024 case ISD::SETEQ: Opc = ARMISD::VCEQ; break; 4025 case ISD::SETOLT: 4026 case ISD::SETLT: Swap = true; // Fallthrough 4027 case ISD::SETOGT: 4028 case ISD::SETGT: Opc = ARMISD::VCGT; break; 4029 case ISD::SETOLE: 4030 case ISD::SETLE: Swap = true; // Fallthrough 4031 case ISD::SETOGE: 4032 case ISD::SETGE: Opc = ARMISD::VCGE; break; 4033 case ISD::SETUGE: Swap = true; // Fallthrough 4034 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break; 4035 case ISD::SETUGT: Swap = true; // Fallthrough 4036 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break; 4037 case ISD::SETUEQ: Invert = true; // Fallthrough 4038 case ISD::SETONE: 4039 // Expand this to (OLT | OGT). 4040 TmpOp0 = Op0; 4041 TmpOp1 = Op1; 4042 Opc = ISD::OR; 4043 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0); 4044 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1); 4045 break; 4046 case ISD::SETUO: Invert = true; // Fallthrough 4047 case ISD::SETO: 4048 // Expand this to (OLT | OGE). 4049 TmpOp0 = Op0; 4050 TmpOp1 = Op1; 4051 Opc = ISD::OR; 4052 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0); 4053 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1); 4054 break; 4055 } 4056 } else { 4057 // Integer comparisons. 4058 switch (SetCCOpcode) { 4059 default: llvm_unreachable("Illegal integer comparison"); 4060 case ISD::SETNE: Invert = true; 4061 case ISD::SETEQ: Opc = ARMISD::VCEQ; break; 4062 case ISD::SETLT: Swap = true; 4063 case ISD::SETGT: Opc = ARMISD::VCGT; break; 4064 case ISD::SETLE: Swap = true; 4065 case ISD::SETGE: Opc = ARMISD::VCGE; break; 4066 case ISD::SETULT: Swap = true; 4067 case ISD::SETUGT: Opc = ARMISD::VCGTU; break; 4068 case ISD::SETULE: Swap = true; 4069 case ISD::SETUGE: Opc = ARMISD::VCGEU; break; 4070 } 4071 4072 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero). 4073 if (Opc == ARMISD::VCEQ) { 4074 4075 SDValue AndOp; 4076 if (ISD::isBuildVectorAllZeros(Op1.getNode())) 4077 AndOp = Op0; 4078 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) 4079 AndOp = Op1; 4080 4081 // Ignore bitconvert. 4082 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST) 4083 AndOp = AndOp.getOperand(0); 4084 4085 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) { 4086 Opc = ARMISD::VTST; 4087 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0)); 4088 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1)); 4089 Invert = !Invert; 4090 } 4091 } 4092 } 4093 4094 if (Swap) 4095 std::swap(Op0, Op1); 4096 4097 // If one of the operands is a constant vector zero, attempt to fold the 4098 // comparison to a specialized compare-against-zero form. 4099 SDValue SingleOp; 4100 if (ISD::isBuildVectorAllZeros(Op1.getNode())) 4101 SingleOp = Op0; 4102 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) { 4103 if (Opc == ARMISD::VCGE) 4104 Opc = ARMISD::VCLEZ; 4105 else if (Opc == ARMISD::VCGT) 4106 Opc = ARMISD::VCLTZ; 4107 SingleOp = Op1; 4108 } 4109 4110 SDValue Result; 4111 if (SingleOp.getNode()) { 4112 switch (Opc) { 4113 case ARMISD::VCEQ: 4114 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break; 4115 case ARMISD::VCGE: 4116 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break; 4117 case ARMISD::VCLEZ: 4118 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break; 4119 case ARMISD::VCGT: 4120 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break; 4121 case ARMISD::VCLTZ: 4122 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break; 4123 default: 4124 Result = DAG.getNode(Opc, dl, VT, Op0, Op1); 4125 } 4126 } else { 4127 Result = DAG.getNode(Opc, dl, VT, Op0, Op1); 4128 } 4129 4130 if (Invert) 4131 Result = DAG.getNOT(dl, Result, VT); 4132 4133 return Result; 4134 } 4135 4136 /// isNEONModifiedImm - Check if the specified splat value corresponds to a 4137 /// valid vector constant for a NEON instruction with a "modified immediate" 4138 /// operand (e.g., VMOV). If so, return the encoded value. 4139 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef, 4140 unsigned SplatBitSize, SelectionDAG &DAG, 4141 EVT &VT, bool is128Bits, NEONModImmType type) { 4142 unsigned OpCmode, Imm; 4143 4144 // SplatBitSize is set to the smallest size that splats the vector, so a 4145 // zero vector will always have SplatBitSize == 8. However, NEON modified 4146 // immediate instructions others than VMOV do not support the 8-bit encoding 4147 // of a zero vector, and the default encoding of zero is supposed to be the 4148 // 32-bit version. 4149 if (SplatBits == 0) 4150 SplatBitSize = 32; 4151 4152 switch (SplatBitSize) { 4153 case 8: 4154 if (type != VMOVModImm) 4155 return SDValue(); 4156 // Any 1-byte value is OK. Op=0, Cmode=1110. 4157 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big"); 4158 OpCmode = 0xe; 4159 Imm = SplatBits; 4160 VT = is128Bits ? MVT::v16i8 : MVT::v8i8; 4161 break; 4162 4163 case 16: 4164 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero. 4165 VT = is128Bits ? MVT::v8i16 : MVT::v4i16; 4166 if ((SplatBits & ~0xff) == 0) { 4167 // Value = 0x00nn: Op=x, Cmode=100x. 4168 OpCmode = 0x8; 4169 Imm = SplatBits; 4170 break; 4171 } 4172 if ((SplatBits & ~0xff00) == 0) { 4173 // Value = 0xnn00: Op=x, Cmode=101x. 4174 OpCmode = 0xa; 4175 Imm = SplatBits >> 8; 4176 break; 4177 } 4178 return SDValue(); 4179 4180 case 32: 4181 // NEON's 32-bit VMOV supports splat values where: 4182 // * only one byte is nonzero, or 4183 // * the least significant byte is 0xff and the second byte is nonzero, or 4184 // * the least significant 2 bytes are 0xff and the third is nonzero. 4185 VT = is128Bits ? MVT::v4i32 : MVT::v2i32; 4186 if ((SplatBits & ~0xff) == 0) { 4187 // Value = 0x000000nn: Op=x, Cmode=000x. 4188 OpCmode = 0; 4189 Imm = SplatBits; 4190 break; 4191 } 4192 if ((SplatBits & ~0xff00) == 0) { 4193 // Value = 0x0000nn00: Op=x, Cmode=001x. 4194 OpCmode = 0x2; 4195 Imm = SplatBits >> 8; 4196 break; 4197 } 4198 if ((SplatBits & ~0xff0000) == 0) { 4199 // Value = 0x00nn0000: Op=x, Cmode=010x. 4200 OpCmode = 0x4; 4201 Imm = SplatBits >> 16; 4202 break; 4203 } 4204 if ((SplatBits & ~0xff000000) == 0) { 4205 // Value = 0xnn000000: Op=x, Cmode=011x. 4206 OpCmode = 0x6; 4207 Imm = SplatBits >> 24; 4208 break; 4209 } 4210 4211 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC 4212 if (type == OtherModImm) return SDValue(); 4213 4214 if ((SplatBits & ~0xffff) == 0 && 4215 ((SplatBits | SplatUndef) & 0xff) == 0xff) { 4216 // Value = 0x0000nnff: Op=x, Cmode=1100. 4217 OpCmode = 0xc; 4218 Imm = SplatBits >> 8; 4219 SplatBits |= 0xff; 4220 break; 4221 } 4222 4223 if ((SplatBits & ~0xffffff) == 0 && 4224 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) { 4225 // Value = 0x00nnffff: Op=x, Cmode=1101. 4226 OpCmode = 0xd; 4227 Imm = SplatBits >> 16; 4228 SplatBits |= 0xffff; 4229 break; 4230 } 4231 4232 // Note: there are a few 32-bit splat values (specifically: 00ffff00, 4233 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not 4234 // VMOV.I32. A (very) minor optimization would be to replicate the value 4235 // and fall through here to test for a valid 64-bit splat. But, then the 4236 // caller would also need to check and handle the change in size. 4237 return SDValue(); 4238 4239 case 64: { 4240 if (type != VMOVModImm) 4241 return SDValue(); 4242 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff. 4243 uint64_t BitMask = 0xff; 4244 uint64_t Val = 0; 4245 unsigned ImmMask = 1; 4246 Imm = 0; 4247 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) { 4248 if (((SplatBits | SplatUndef) & BitMask) == BitMask) { 4249 Val |= BitMask; 4250 Imm |= ImmMask; 4251 } else if ((SplatBits & BitMask) != 0) { 4252 return SDValue(); 4253 } 4254 BitMask <<= 8; 4255 ImmMask <<= 1; 4256 } 4257 // Op=1, Cmode=1110. 4258 OpCmode = 0x1e; 4259 SplatBits = Val; 4260 VT = is128Bits ? MVT::v2i64 : MVT::v1i64; 4261 break; 4262 } 4263 4264 default: 4265 llvm_unreachable("unexpected size for isNEONModifiedImm"); 4266 } 4267 4268 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm); 4269 return DAG.getTargetConstant(EncodedVal, MVT::i32); 4270 } 4271 4272 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG, 4273 const ARMSubtarget *ST) const { 4274 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16()) 4275 return SDValue(); 4276 4277 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op); 4278 assert(Op.getValueType() == MVT::f32 && 4279 "ConstantFP custom lowering should only occur for f32."); 4280 4281 // Try splatting with a VMOV.f32... 4282 APFloat FPVal = CFP->getValueAPF(); 4283 int ImmVal = ARM_AM::getFP32Imm(FPVal); 4284 if (ImmVal != -1) { 4285 SDLoc DL(Op); 4286 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32); 4287 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32, 4288 NewVal); 4289 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant, 4290 DAG.getConstant(0, MVT::i32)); 4291 } 4292 4293 // If that fails, try a VMOV.i32 4294 EVT VMovVT; 4295 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue(); 4296 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false, 4297 VMOVModImm); 4298 if (NewVal != SDValue()) { 4299 SDLoc DL(Op); 4300 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT, 4301 NewVal); 4302 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32, 4303 VecConstant); 4304 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant, 4305 DAG.getConstant(0, MVT::i32)); 4306 } 4307 4308 // Finally, try a VMVN.i32 4309 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false, 4310 VMVNModImm); 4311 if (NewVal != SDValue()) { 4312 SDLoc DL(Op); 4313 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal); 4314 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32, 4315 VecConstant); 4316 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant, 4317 DAG.getConstant(0, MVT::i32)); 4318 } 4319 4320 return SDValue(); 4321 } 4322 4323 // check if an VEXT instruction can handle the shuffle mask when the 4324 // vector sources of the shuffle are the same. 4325 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) { 4326 unsigned NumElts = VT.getVectorNumElements(); 4327 4328 // Assume that the first shuffle index is not UNDEF. Fail if it is. 4329 if (M[0] < 0) 4330 return false; 4331 4332 Imm = M[0]; 4333 4334 // If this is a VEXT shuffle, the immediate value is the index of the first 4335 // element. The other shuffle indices must be the successive elements after 4336 // the first one. 4337 unsigned ExpectedElt = Imm; 4338 for (unsigned i = 1; i < NumElts; ++i) { 4339 // Increment the expected index. If it wraps around, just follow it 4340 // back to index zero and keep going. 4341 ++ExpectedElt; 4342 if (ExpectedElt == NumElts) 4343 ExpectedElt = 0; 4344 4345 if (M[i] < 0) continue; // ignore UNDEF indices 4346 if (ExpectedElt != static_cast<unsigned>(M[i])) 4347 return false; 4348 } 4349 4350 return true; 4351 } 4352 4353 4354 static bool isVEXTMask(ArrayRef<int> M, EVT VT, 4355 bool &ReverseVEXT, unsigned &Imm) { 4356 unsigned NumElts = VT.getVectorNumElements(); 4357 ReverseVEXT = false; 4358 4359 // Assume that the first shuffle index is not UNDEF. Fail if it is. 4360 if (M[0] < 0) 4361 return false; 4362 4363 Imm = M[0]; 4364 4365 // If this is a VEXT shuffle, the immediate value is the index of the first 4366 // element. The other shuffle indices must be the successive elements after 4367 // the first one. 4368 unsigned ExpectedElt = Imm; 4369 for (unsigned i = 1; i < NumElts; ++i) { 4370 // Increment the expected index. If it wraps around, it may still be 4371 // a VEXT but the source vectors must be swapped. 4372 ExpectedElt += 1; 4373 if (ExpectedElt == NumElts * 2) { 4374 ExpectedElt = 0; 4375 ReverseVEXT = true; 4376 } 4377 4378 if (M[i] < 0) continue; // ignore UNDEF indices 4379 if (ExpectedElt != static_cast<unsigned>(M[i])) 4380 return false; 4381 } 4382 4383 // Adjust the index value if the source operands will be swapped. 4384 if (ReverseVEXT) 4385 Imm -= NumElts; 4386 4387 return true; 4388 } 4389 4390 /// isVREVMask - Check if a vector shuffle corresponds to a VREV 4391 /// instruction with the specified blocksize. (The order of the elements 4392 /// within each block of the vector is reversed.) 4393 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) { 4394 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) && 4395 "Only possible block sizes for VREV are: 16, 32, 64"); 4396 4397 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4398 if (EltSz == 64) 4399 return false; 4400 4401 unsigned NumElts = VT.getVectorNumElements(); 4402 unsigned BlockElts = M[0] + 1; 4403 // If the first shuffle index is UNDEF, be optimistic. 4404 if (M[0] < 0) 4405 BlockElts = BlockSize / EltSz; 4406 4407 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz) 4408 return false; 4409 4410 for (unsigned i = 0; i < NumElts; ++i) { 4411 if (M[i] < 0) continue; // ignore UNDEF indices 4412 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts)) 4413 return false; 4414 } 4415 4416 return true; 4417 } 4418 4419 static bool isVTBLMask(ArrayRef<int> M, EVT VT) { 4420 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of 4421 // range, then 0 is placed into the resulting vector. So pretty much any mask 4422 // of 8 elements can work here. 4423 return VT == MVT::v8i8 && M.size() == 8; 4424 } 4425 4426 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 4427 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4428 if (EltSz == 64) 4429 return false; 4430 4431 unsigned NumElts = VT.getVectorNumElements(); 4432 WhichResult = (M[0] == 0 ? 0 : 1); 4433 for (unsigned i = 0; i < NumElts; i += 2) { 4434 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) || 4435 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult)) 4436 return false; 4437 } 4438 return true; 4439 } 4440 4441 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of 4442 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 4443 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>. 4444 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){ 4445 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4446 if (EltSz == 64) 4447 return false; 4448 4449 unsigned NumElts = VT.getVectorNumElements(); 4450 WhichResult = (M[0] == 0 ? 0 : 1); 4451 for (unsigned i = 0; i < NumElts; i += 2) { 4452 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) || 4453 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult)) 4454 return false; 4455 } 4456 return true; 4457 } 4458 4459 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 4460 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4461 if (EltSz == 64) 4462 return false; 4463 4464 unsigned NumElts = VT.getVectorNumElements(); 4465 WhichResult = (M[0] == 0 ? 0 : 1); 4466 for (unsigned i = 0; i != NumElts; ++i) { 4467 if (M[i] < 0) continue; // ignore UNDEF indices 4468 if ((unsigned) M[i] != 2 * i + WhichResult) 4469 return false; 4470 } 4471 4472 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4473 if (VT.is64BitVector() && EltSz == 32) 4474 return false; 4475 4476 return true; 4477 } 4478 4479 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of 4480 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 4481 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>, 4482 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){ 4483 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4484 if (EltSz == 64) 4485 return false; 4486 4487 unsigned Half = VT.getVectorNumElements() / 2; 4488 WhichResult = (M[0] == 0 ? 0 : 1); 4489 for (unsigned j = 0; j != 2; ++j) { 4490 unsigned Idx = WhichResult; 4491 for (unsigned i = 0; i != Half; ++i) { 4492 int MIdx = M[i + j * Half]; 4493 if (MIdx >= 0 && (unsigned) MIdx != Idx) 4494 return false; 4495 Idx += 2; 4496 } 4497 } 4498 4499 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4500 if (VT.is64BitVector() && EltSz == 32) 4501 return false; 4502 4503 return true; 4504 } 4505 4506 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 4507 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4508 if (EltSz == 64) 4509 return false; 4510 4511 unsigned NumElts = VT.getVectorNumElements(); 4512 WhichResult = (M[0] == 0 ? 0 : 1); 4513 unsigned Idx = WhichResult * NumElts / 2; 4514 for (unsigned i = 0; i != NumElts; i += 2) { 4515 if ((M[i] >= 0 && (unsigned) M[i] != Idx) || 4516 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts)) 4517 return false; 4518 Idx += 1; 4519 } 4520 4521 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4522 if (VT.is64BitVector() && EltSz == 32) 4523 return false; 4524 4525 return true; 4526 } 4527 4528 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of 4529 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 4530 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>. 4531 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){ 4532 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 4533 if (EltSz == 64) 4534 return false; 4535 4536 unsigned NumElts = VT.getVectorNumElements(); 4537 WhichResult = (M[0] == 0 ? 0 : 1); 4538 unsigned Idx = WhichResult * NumElts / 2; 4539 for (unsigned i = 0; i != NumElts; i += 2) { 4540 if ((M[i] >= 0 && (unsigned) M[i] != Idx) || 4541 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx)) 4542 return false; 4543 Idx += 1; 4544 } 4545 4546 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 4547 if (VT.is64BitVector() && EltSz == 32) 4548 return false; 4549 4550 return true; 4551 } 4552 4553 /// \return true if this is a reverse operation on an vector. 4554 static bool isReverseMask(ArrayRef<int> M, EVT VT) { 4555 unsigned NumElts = VT.getVectorNumElements(); 4556 // Make sure the mask has the right size. 4557 if (NumElts != M.size()) 4558 return false; 4559 4560 // Look for <15, ..., 3, -1, 1, 0>. 4561 for (unsigned i = 0; i != NumElts; ++i) 4562 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i)) 4563 return false; 4564 4565 return true; 4566 } 4567 4568 // If N is an integer constant that can be moved into a register in one 4569 // instruction, return an SDValue of such a constant (will become a MOV 4570 // instruction). Otherwise return null. 4571 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG, 4572 const ARMSubtarget *ST, SDLoc dl) { 4573 uint64_t Val; 4574 if (!isa<ConstantSDNode>(N)) 4575 return SDValue(); 4576 Val = cast<ConstantSDNode>(N)->getZExtValue(); 4577 4578 if (ST->isThumb1Only()) { 4579 if (Val <= 255 || ~Val <= 255) 4580 return DAG.getConstant(Val, MVT::i32); 4581 } else { 4582 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1) 4583 return DAG.getConstant(Val, MVT::i32); 4584 } 4585 return SDValue(); 4586 } 4587 4588 // If this is a case we can't handle, return null and let the default 4589 // expansion code take care of it. 4590 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG, 4591 const ARMSubtarget *ST) const { 4592 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode()); 4593 SDLoc dl(Op); 4594 EVT VT = Op.getValueType(); 4595 4596 APInt SplatBits, SplatUndef; 4597 unsigned SplatBitSize; 4598 bool HasAnyUndefs; 4599 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) { 4600 if (SplatBitSize <= 64) { 4601 // Check if an immediate VMOV works. 4602 EVT VmovVT; 4603 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(), 4604 SplatUndef.getZExtValue(), SplatBitSize, 4605 DAG, VmovVT, VT.is128BitVector(), 4606 VMOVModImm); 4607 if (Val.getNode()) { 4608 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val); 4609 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov); 4610 } 4611 4612 // Try an immediate VMVN. 4613 uint64_t NegatedImm = (~SplatBits).getZExtValue(); 4614 Val = isNEONModifiedImm(NegatedImm, 4615 SplatUndef.getZExtValue(), SplatBitSize, 4616 DAG, VmovVT, VT.is128BitVector(), 4617 VMVNModImm); 4618 if (Val.getNode()) { 4619 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val); 4620 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov); 4621 } 4622 4623 // Use vmov.f32 to materialize other v2f32 and v4f32 splats. 4624 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) { 4625 int ImmVal = ARM_AM::getFP32Imm(SplatBits); 4626 if (ImmVal != -1) { 4627 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32); 4628 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val); 4629 } 4630 } 4631 } 4632 } 4633 4634 // Scan through the operands to see if only one value is used. 4635 // 4636 // As an optimisation, even if more than one value is used it may be more 4637 // profitable to splat with one value then change some lanes. 4638 // 4639 // Heuristically we decide to do this if the vector has a "dominant" value, 4640 // defined as splatted to more than half of the lanes. 4641 unsigned NumElts = VT.getVectorNumElements(); 4642 bool isOnlyLowElement = true; 4643 bool usesOnlyOneValue = true; 4644 bool hasDominantValue = false; 4645 bool isConstant = true; 4646 4647 // Map of the number of times a particular SDValue appears in the 4648 // element list. 4649 DenseMap<SDValue, unsigned> ValueCounts; 4650 SDValue Value; 4651 for (unsigned i = 0; i < NumElts; ++i) { 4652 SDValue V = Op.getOperand(i); 4653 if (V.getOpcode() == ISD::UNDEF) 4654 continue; 4655 if (i > 0) 4656 isOnlyLowElement = false; 4657 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V)) 4658 isConstant = false; 4659 4660 ValueCounts.insert(std::make_pair(V, 0)); 4661 unsigned &Count = ValueCounts[V]; 4662 4663 // Is this value dominant? (takes up more than half of the lanes) 4664 if (++Count > (NumElts / 2)) { 4665 hasDominantValue = true; 4666 Value = V; 4667 } 4668 } 4669 if (ValueCounts.size() != 1) 4670 usesOnlyOneValue = false; 4671 if (!Value.getNode() && ValueCounts.size() > 0) 4672 Value = ValueCounts.begin()->first; 4673 4674 if (ValueCounts.size() == 0) 4675 return DAG.getUNDEF(VT); 4676 4677 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR. 4678 // Keep going if we are hitting this case. 4679 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode())) 4680 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value); 4681 4682 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 4683 4684 // Use VDUP for non-constant splats. For f32 constant splats, reduce to 4685 // i32 and try again. 4686 if (hasDominantValue && EltSize <= 32) { 4687 if (!isConstant) { 4688 SDValue N; 4689 4690 // If we are VDUPing a value that comes directly from a vector, that will 4691 // cause an unnecessary move to and from a GPR, where instead we could 4692 // just use VDUPLANE. We can only do this if the lane being extracted 4693 // is at a constant index, as the VDUP from lane instructions only have 4694 // constant-index forms. 4695 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT && 4696 isa<ConstantSDNode>(Value->getOperand(1))) { 4697 // We need to create a new undef vector to use for the VDUPLANE if the 4698 // size of the vector from which we get the value is different than the 4699 // size of the vector that we need to create. We will insert the element 4700 // such that the register coalescer will remove unnecessary copies. 4701 if (VT != Value->getOperand(0).getValueType()) { 4702 ConstantSDNode *constIndex; 4703 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1)); 4704 assert(constIndex && "The index is not a constant!"); 4705 unsigned index = constIndex->getAPIntValue().getLimitedValue() % 4706 VT.getVectorNumElements(); 4707 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT, 4708 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT), 4709 Value, DAG.getConstant(index, MVT::i32)), 4710 DAG.getConstant(index, MVT::i32)); 4711 } else 4712 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT, 4713 Value->getOperand(0), Value->getOperand(1)); 4714 } else 4715 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value); 4716 4717 if (!usesOnlyOneValue) { 4718 // The dominant value was splatted as 'N', but we now have to insert 4719 // all differing elements. 4720 for (unsigned I = 0; I < NumElts; ++I) { 4721 if (Op.getOperand(I) == Value) 4722 continue; 4723 SmallVector<SDValue, 3> Ops; 4724 Ops.push_back(N); 4725 Ops.push_back(Op.getOperand(I)); 4726 Ops.push_back(DAG.getConstant(I, MVT::i32)); 4727 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3); 4728 } 4729 } 4730 return N; 4731 } 4732 if (VT.getVectorElementType().isFloatingPoint()) { 4733 SmallVector<SDValue, 8> Ops; 4734 for (unsigned i = 0; i < NumElts; ++i) 4735 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32, 4736 Op.getOperand(i))); 4737 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts); 4738 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts); 4739 Val = LowerBUILD_VECTOR(Val, DAG, ST); 4740 if (Val.getNode()) 4741 return DAG.getNode(ISD::BITCAST, dl, VT, Val); 4742 } 4743 if (usesOnlyOneValue) { 4744 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl); 4745 if (isConstant && Val.getNode()) 4746 return DAG.getNode(ARMISD::VDUP, dl, VT, Val); 4747 } 4748 } 4749 4750 // If all elements are constants and the case above didn't get hit, fall back 4751 // to the default expansion, which will generate a load from the constant 4752 // pool. 4753 if (isConstant) 4754 return SDValue(); 4755 4756 // Empirical tests suggest this is rarely worth it for vectors of length <= 2. 4757 if (NumElts >= 4) { 4758 SDValue shuffle = ReconstructShuffle(Op, DAG); 4759 if (shuffle != SDValue()) 4760 return shuffle; 4761 } 4762 4763 // Vectors with 32- or 64-bit elements can be built by directly assigning 4764 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands 4765 // will be legalized. 4766 if (EltSize >= 32) { 4767 // Do the expansion with floating-point types, since that is what the VFP 4768 // registers are defined to use, and since i64 is not legal. 4769 EVT EltVT = EVT::getFloatingPointVT(EltSize); 4770 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts); 4771 SmallVector<SDValue, 8> Ops; 4772 for (unsigned i = 0; i < NumElts; ++i) 4773 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i))); 4774 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts); 4775 return DAG.getNode(ISD::BITCAST, dl, VT, Val); 4776 } 4777 4778 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we 4779 // know the default expansion would otherwise fall back on something even 4780 // worse. For a vector with one or two non-undef values, that's 4781 // scalar_to_vector for the elements followed by a shuffle (provided the 4782 // shuffle is valid for the target) and materialization element by element 4783 // on the stack followed by a load for everything else. 4784 if (!isConstant && !usesOnlyOneValue) { 4785 SDValue Vec = DAG.getUNDEF(VT); 4786 for (unsigned i = 0 ; i < NumElts; ++i) { 4787 SDValue V = Op.getOperand(i); 4788 if (V.getOpcode() == ISD::UNDEF) 4789 continue; 4790 SDValue LaneIdx = DAG.getConstant(i, MVT::i32); 4791 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx); 4792 } 4793 return Vec; 4794 } 4795 4796 return SDValue(); 4797 } 4798 4799 // Gather data to see if the operation can be modelled as a 4800 // shuffle in combination with VEXTs. 4801 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op, 4802 SelectionDAG &DAG) const { 4803 SDLoc dl(Op); 4804 EVT VT = Op.getValueType(); 4805 unsigned NumElts = VT.getVectorNumElements(); 4806 4807 SmallVector<SDValue, 2> SourceVecs; 4808 SmallVector<unsigned, 2> MinElts; 4809 SmallVector<unsigned, 2> MaxElts; 4810 4811 for (unsigned i = 0; i < NumElts; ++i) { 4812 SDValue V = Op.getOperand(i); 4813 if (V.getOpcode() == ISD::UNDEF) 4814 continue; 4815 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) { 4816 // A shuffle can only come from building a vector from various 4817 // elements of other vectors. 4818 return SDValue(); 4819 } else if (V.getOperand(0).getValueType().getVectorElementType() != 4820 VT.getVectorElementType()) { 4821 // This code doesn't know how to handle shuffles where the vector 4822 // element types do not match (this happens because type legalization 4823 // promotes the return type of EXTRACT_VECTOR_ELT). 4824 // FIXME: It might be appropriate to extend this code to handle 4825 // mismatched types. 4826 return SDValue(); 4827 } 4828 4829 // Record this extraction against the appropriate vector if possible... 4830 SDValue SourceVec = V.getOperand(0); 4831 // If the element number isn't a constant, we can't effectively 4832 // analyze what's going on. 4833 if (!isa<ConstantSDNode>(V.getOperand(1))) 4834 return SDValue(); 4835 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue(); 4836 bool FoundSource = false; 4837 for (unsigned j = 0; j < SourceVecs.size(); ++j) { 4838 if (SourceVecs[j] == SourceVec) { 4839 if (MinElts[j] > EltNo) 4840 MinElts[j] = EltNo; 4841 if (MaxElts[j] < EltNo) 4842 MaxElts[j] = EltNo; 4843 FoundSource = true; 4844 break; 4845 } 4846 } 4847 4848 // Or record a new source if not... 4849 if (!FoundSource) { 4850 SourceVecs.push_back(SourceVec); 4851 MinElts.push_back(EltNo); 4852 MaxElts.push_back(EltNo); 4853 } 4854 } 4855 4856 // Currently only do something sane when at most two source vectors 4857 // involved. 4858 if (SourceVecs.size() > 2) 4859 return SDValue(); 4860 4861 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) }; 4862 int VEXTOffsets[2] = {0, 0}; 4863 4864 // This loop extracts the usage patterns of the source vectors 4865 // and prepares appropriate SDValues for a shuffle if possible. 4866 for (unsigned i = 0; i < SourceVecs.size(); ++i) { 4867 if (SourceVecs[i].getValueType() == VT) { 4868 // No VEXT necessary 4869 ShuffleSrcs[i] = SourceVecs[i]; 4870 VEXTOffsets[i] = 0; 4871 continue; 4872 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) { 4873 // It probably isn't worth padding out a smaller vector just to 4874 // break it down again in a shuffle. 4875 return SDValue(); 4876 } 4877 4878 // Since only 64-bit and 128-bit vectors are legal on ARM and 4879 // we've eliminated the other cases... 4880 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts && 4881 "unexpected vector sizes in ReconstructShuffle"); 4882 4883 if (MaxElts[i] - MinElts[i] >= NumElts) { 4884 // Span too large for a VEXT to cope 4885 return SDValue(); 4886 } 4887 4888 if (MinElts[i] >= NumElts) { 4889 // The extraction can just take the second half 4890 VEXTOffsets[i] = NumElts; 4891 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4892 SourceVecs[i], 4893 DAG.getIntPtrConstant(NumElts)); 4894 } else if (MaxElts[i] < NumElts) { 4895 // The extraction can just take the first half 4896 VEXTOffsets[i] = 0; 4897 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4898 SourceVecs[i], 4899 DAG.getIntPtrConstant(0)); 4900 } else { 4901 // An actual VEXT is needed 4902 VEXTOffsets[i] = MinElts[i]; 4903 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4904 SourceVecs[i], 4905 DAG.getIntPtrConstant(0)); 4906 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, 4907 SourceVecs[i], 4908 DAG.getIntPtrConstant(NumElts)); 4909 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2, 4910 DAG.getConstant(VEXTOffsets[i], MVT::i32)); 4911 } 4912 } 4913 4914 SmallVector<int, 8> Mask; 4915 4916 for (unsigned i = 0; i < NumElts; ++i) { 4917 SDValue Entry = Op.getOperand(i); 4918 if (Entry.getOpcode() == ISD::UNDEF) { 4919 Mask.push_back(-1); 4920 continue; 4921 } 4922 4923 SDValue ExtractVec = Entry.getOperand(0); 4924 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i) 4925 .getOperand(1))->getSExtValue(); 4926 if (ExtractVec == SourceVecs[0]) { 4927 Mask.push_back(ExtractElt - VEXTOffsets[0]); 4928 } else { 4929 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]); 4930 } 4931 } 4932 4933 // Final check before we try to produce nonsense... 4934 if (isShuffleMaskLegal(Mask, VT)) 4935 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1], 4936 &Mask[0]); 4937 4938 return SDValue(); 4939 } 4940 4941 /// isShuffleMaskLegal - Targets can use this to indicate that they only 4942 /// support *some* VECTOR_SHUFFLE operations, those with specific masks. 4943 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values 4944 /// are assumed to be legal. 4945 bool 4946 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M, 4947 EVT VT) const { 4948 if (VT.getVectorNumElements() == 4 && 4949 (VT.is128BitVector() || VT.is64BitVector())) { 4950 unsigned PFIndexes[4]; 4951 for (unsigned i = 0; i != 4; ++i) { 4952 if (M[i] < 0) 4953 PFIndexes[i] = 8; 4954 else 4955 PFIndexes[i] = M[i]; 4956 } 4957 4958 // Compute the index in the perfect shuffle table. 4959 unsigned PFTableIndex = 4960 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3]; 4961 unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; 4962 unsigned Cost = (PFEntry >> 30); 4963 4964 if (Cost <= 4) 4965 return true; 4966 } 4967 4968 bool ReverseVEXT; 4969 unsigned Imm, WhichResult; 4970 4971 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 4972 return (EltSize >= 32 || 4973 ShuffleVectorSDNode::isSplatMask(&M[0], VT) || 4974 isVREVMask(M, VT, 64) || 4975 isVREVMask(M, VT, 32) || 4976 isVREVMask(M, VT, 16) || 4977 isVEXTMask(M, VT, ReverseVEXT, Imm) || 4978 isVTBLMask(M, VT) || 4979 isVTRNMask(M, VT, WhichResult) || 4980 isVUZPMask(M, VT, WhichResult) || 4981 isVZIPMask(M, VT, WhichResult) || 4982 isVTRN_v_undef_Mask(M, VT, WhichResult) || 4983 isVUZP_v_undef_Mask(M, VT, WhichResult) || 4984 isVZIP_v_undef_Mask(M, VT, WhichResult) || 4985 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT))); 4986 } 4987 4988 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit 4989 /// the specified operations to build the shuffle. 4990 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS, 4991 SDValue RHS, SelectionDAG &DAG, 4992 SDLoc dl) { 4993 unsigned OpNum = (PFEntry >> 26) & 0x0F; 4994 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1); 4995 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1); 4996 4997 enum { 4998 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3> 4999 OP_VREV, 5000 OP_VDUP0, 5001 OP_VDUP1, 5002 OP_VDUP2, 5003 OP_VDUP3, 5004 OP_VEXT1, 5005 OP_VEXT2, 5006 OP_VEXT3, 5007 OP_VUZPL, // VUZP, left result 5008 OP_VUZPR, // VUZP, right result 5009 OP_VZIPL, // VZIP, left result 5010 OP_VZIPR, // VZIP, right result 5011 OP_VTRNL, // VTRN, left result 5012 OP_VTRNR // VTRN, right result 5013 }; 5014 5015 if (OpNum == OP_COPY) { 5016 if (LHSID == (1*9+2)*9+3) return LHS; 5017 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!"); 5018 return RHS; 5019 } 5020 5021 SDValue OpLHS, OpRHS; 5022 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl); 5023 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl); 5024 EVT VT = OpLHS.getValueType(); 5025 5026 switch (OpNum) { 5027 default: llvm_unreachable("Unknown shuffle opcode!"); 5028 case OP_VREV: 5029 // VREV divides the vector in half and swaps within the half. 5030 if (VT.getVectorElementType() == MVT::i32 || 5031 VT.getVectorElementType() == MVT::f32) 5032 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS); 5033 // vrev <4 x i16> -> VREV32 5034 if (VT.getVectorElementType() == MVT::i16) 5035 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS); 5036 // vrev <4 x i8> -> VREV16 5037 assert(VT.getVectorElementType() == MVT::i8); 5038 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS); 5039 case OP_VDUP0: 5040 case OP_VDUP1: 5041 case OP_VDUP2: 5042 case OP_VDUP3: 5043 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, 5044 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32)); 5045 case OP_VEXT1: 5046 case OP_VEXT2: 5047 case OP_VEXT3: 5048 return DAG.getNode(ARMISD::VEXT, dl, VT, 5049 OpLHS, OpRHS, 5050 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32)); 5051 case OP_VUZPL: 5052 case OP_VUZPR: 5053 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 5054 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL); 5055 case OP_VZIPL: 5056 case OP_VZIPR: 5057 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 5058 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL); 5059 case OP_VTRNL: 5060 case OP_VTRNR: 5061 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 5062 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL); 5063 } 5064 } 5065 5066 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op, 5067 ArrayRef<int> ShuffleMask, 5068 SelectionDAG &DAG) { 5069 // Check to see if we can use the VTBL instruction. 5070 SDValue V1 = Op.getOperand(0); 5071 SDValue V2 = Op.getOperand(1); 5072 SDLoc DL(Op); 5073 5074 SmallVector<SDValue, 8> VTBLMask; 5075 for (ArrayRef<int>::iterator 5076 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I) 5077 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32)); 5078 5079 if (V2.getNode()->getOpcode() == ISD::UNDEF) 5080 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1, 5081 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, 5082 &VTBLMask[0], 8)); 5083 5084 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2, 5085 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, 5086 &VTBLMask[0], 8)); 5087 } 5088 5089 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op, 5090 SelectionDAG &DAG) { 5091 SDLoc DL(Op); 5092 SDValue OpLHS = Op.getOperand(0); 5093 EVT VT = OpLHS.getValueType(); 5094 5095 assert((VT == MVT::v8i16 || VT == MVT::v16i8) && 5096 "Expect an v8i16/v16i8 type"); 5097 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS); 5098 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now, 5099 // extract the first 8 bytes into the top double word and the last 8 bytes 5100 // into the bottom double word. The v8i16 case is similar. 5101 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4; 5102 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS, 5103 DAG.getConstant(ExtractNum, MVT::i32)); 5104 } 5105 5106 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) { 5107 SDValue V1 = Op.getOperand(0); 5108 SDValue V2 = Op.getOperand(1); 5109 SDLoc dl(Op); 5110 EVT VT = Op.getValueType(); 5111 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode()); 5112 5113 // Convert shuffles that are directly supported on NEON to target-specific 5114 // DAG nodes, instead of keeping them as shuffles and matching them again 5115 // during code selection. This is more efficient and avoids the possibility 5116 // of inconsistencies between legalization and selection. 5117 // FIXME: floating-point vectors should be canonicalized to integer vectors 5118 // of the same time so that they get CSEd properly. 5119 ArrayRef<int> ShuffleMask = SVN->getMask(); 5120 5121 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 5122 if (EltSize <= 32) { 5123 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) { 5124 int Lane = SVN->getSplatIndex(); 5125 // If this is undef splat, generate it via "just" vdup, if possible. 5126 if (Lane == -1) Lane = 0; 5127 5128 // Test if V1 is a SCALAR_TO_VECTOR. 5129 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) { 5130 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0)); 5131 } 5132 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR 5133 // (and probably will turn into a SCALAR_TO_VECTOR once legalization 5134 // reaches it). 5135 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR && 5136 !isa<ConstantSDNode>(V1.getOperand(0))) { 5137 bool IsScalarToVector = true; 5138 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i) 5139 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) { 5140 IsScalarToVector = false; 5141 break; 5142 } 5143 if (IsScalarToVector) 5144 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0)); 5145 } 5146 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1, 5147 DAG.getConstant(Lane, MVT::i32)); 5148 } 5149 5150 bool ReverseVEXT; 5151 unsigned Imm; 5152 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) { 5153 if (ReverseVEXT) 5154 std::swap(V1, V2); 5155 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2, 5156 DAG.getConstant(Imm, MVT::i32)); 5157 } 5158 5159 if (isVREVMask(ShuffleMask, VT, 64)) 5160 return DAG.getNode(ARMISD::VREV64, dl, VT, V1); 5161 if (isVREVMask(ShuffleMask, VT, 32)) 5162 return DAG.getNode(ARMISD::VREV32, dl, VT, V1); 5163 if (isVREVMask(ShuffleMask, VT, 16)) 5164 return DAG.getNode(ARMISD::VREV16, dl, VT, V1); 5165 5166 if (V2->getOpcode() == ISD::UNDEF && 5167 isSingletonVEXTMask(ShuffleMask, VT, Imm)) { 5168 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1, 5169 DAG.getConstant(Imm, MVT::i32)); 5170 } 5171 5172 // Check for Neon shuffles that modify both input vectors in place. 5173 // If both results are used, i.e., if there are two shuffles with the same 5174 // source operands and with masks corresponding to both results of one of 5175 // these operations, DAG memoization will ensure that a single node is 5176 // used for both shuffles. 5177 unsigned WhichResult; 5178 if (isVTRNMask(ShuffleMask, VT, WhichResult)) 5179 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 5180 V1, V2).getValue(WhichResult); 5181 if (isVUZPMask(ShuffleMask, VT, WhichResult)) 5182 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 5183 V1, V2).getValue(WhichResult); 5184 if (isVZIPMask(ShuffleMask, VT, WhichResult)) 5185 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 5186 V1, V2).getValue(WhichResult); 5187 5188 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) 5189 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 5190 V1, V1).getValue(WhichResult); 5191 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) 5192 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 5193 V1, V1).getValue(WhichResult); 5194 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) 5195 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 5196 V1, V1).getValue(WhichResult); 5197 } 5198 5199 // If the shuffle is not directly supported and it has 4 elements, use 5200 // the PerfectShuffle-generated table to synthesize it from other shuffles. 5201 unsigned NumElts = VT.getVectorNumElements(); 5202 if (NumElts == 4) { 5203 unsigned PFIndexes[4]; 5204 for (unsigned i = 0; i != 4; ++i) { 5205 if (ShuffleMask[i] < 0) 5206 PFIndexes[i] = 8; 5207 else 5208 PFIndexes[i] = ShuffleMask[i]; 5209 } 5210 5211 // Compute the index in the perfect shuffle table. 5212 unsigned PFTableIndex = 5213 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3]; 5214 unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; 5215 unsigned Cost = (PFEntry >> 30); 5216 5217 if (Cost <= 4) 5218 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl); 5219 } 5220 5221 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs. 5222 if (EltSize >= 32) { 5223 // Do the expansion with floating-point types, since that is what the VFP 5224 // registers are defined to use, and since i64 is not legal. 5225 EVT EltVT = EVT::getFloatingPointVT(EltSize); 5226 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts); 5227 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1); 5228 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2); 5229 SmallVector<SDValue, 8> Ops; 5230 for (unsigned i = 0; i < NumElts; ++i) { 5231 if (ShuffleMask[i] < 0) 5232 Ops.push_back(DAG.getUNDEF(EltVT)); 5233 else 5234 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, 5235 ShuffleMask[i] < (int)NumElts ? V1 : V2, 5236 DAG.getConstant(ShuffleMask[i] & (NumElts-1), 5237 MVT::i32))); 5238 } 5239 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts); 5240 return DAG.getNode(ISD::BITCAST, dl, VT, Val); 5241 } 5242 5243 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT)) 5244 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG); 5245 5246 if (VT == MVT::v8i8) { 5247 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG); 5248 if (NewOp.getNode()) 5249 return NewOp; 5250 } 5251 5252 return SDValue(); 5253 } 5254 5255 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) { 5256 // INSERT_VECTOR_ELT is legal only for immediate indexes. 5257 SDValue Lane = Op.getOperand(2); 5258 if (!isa<ConstantSDNode>(Lane)) 5259 return SDValue(); 5260 5261 return Op; 5262 } 5263 5264 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) { 5265 // EXTRACT_VECTOR_ELT is legal only for immediate indexes. 5266 SDValue Lane = Op.getOperand(1); 5267 if (!isa<ConstantSDNode>(Lane)) 5268 return SDValue(); 5269 5270 SDValue Vec = Op.getOperand(0); 5271 if (Op.getValueType() == MVT::i32 && 5272 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) { 5273 SDLoc dl(Op); 5274 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane); 5275 } 5276 5277 return Op; 5278 } 5279 5280 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) { 5281 // The only time a CONCAT_VECTORS operation can have legal types is when 5282 // two 64-bit vectors are concatenated to a 128-bit vector. 5283 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 && 5284 "unexpected CONCAT_VECTORS"); 5285 SDLoc dl(Op); 5286 SDValue Val = DAG.getUNDEF(MVT::v2f64); 5287 SDValue Op0 = Op.getOperand(0); 5288 SDValue Op1 = Op.getOperand(1); 5289 if (Op0.getOpcode() != ISD::UNDEF) 5290 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val, 5291 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0), 5292 DAG.getIntPtrConstant(0)); 5293 if (Op1.getOpcode() != ISD::UNDEF) 5294 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val, 5295 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1), 5296 DAG.getIntPtrConstant(1)); 5297 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val); 5298 } 5299 5300 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each 5301 /// element has been zero/sign-extended, depending on the isSigned parameter, 5302 /// from an integer type half its size. 5303 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG, 5304 bool isSigned) { 5305 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32. 5306 EVT VT = N->getValueType(0); 5307 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) { 5308 SDNode *BVN = N->getOperand(0).getNode(); 5309 if (BVN->getValueType(0) != MVT::v4i32 || 5310 BVN->getOpcode() != ISD::BUILD_VECTOR) 5311 return false; 5312 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0; 5313 unsigned HiElt = 1 - LoElt; 5314 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt)); 5315 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt)); 5316 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2)); 5317 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2)); 5318 if (!Lo0 || !Hi0 || !Lo1 || !Hi1) 5319 return false; 5320 if (isSigned) { 5321 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 && 5322 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32) 5323 return true; 5324 } else { 5325 if (Hi0->isNullValue() && Hi1->isNullValue()) 5326 return true; 5327 } 5328 return false; 5329 } 5330 5331 if (N->getOpcode() != ISD::BUILD_VECTOR) 5332 return false; 5333 5334 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 5335 SDNode *Elt = N->getOperand(i).getNode(); 5336 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) { 5337 unsigned EltSize = VT.getVectorElementType().getSizeInBits(); 5338 unsigned HalfSize = EltSize / 2; 5339 if (isSigned) { 5340 if (!isIntN(HalfSize, C->getSExtValue())) 5341 return false; 5342 } else { 5343 if (!isUIntN(HalfSize, C->getZExtValue())) 5344 return false; 5345 } 5346 continue; 5347 } 5348 return false; 5349 } 5350 5351 return true; 5352 } 5353 5354 /// isSignExtended - Check if a node is a vector value that is sign-extended 5355 /// or a constant BUILD_VECTOR with sign-extended elements. 5356 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) { 5357 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N)) 5358 return true; 5359 if (isExtendedBUILD_VECTOR(N, DAG, true)) 5360 return true; 5361 return false; 5362 } 5363 5364 /// isZeroExtended - Check if a node is a vector value that is zero-extended 5365 /// or a constant BUILD_VECTOR with zero-extended elements. 5366 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) { 5367 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N)) 5368 return true; 5369 if (isExtendedBUILD_VECTOR(N, DAG, false)) 5370 return true; 5371 return false; 5372 } 5373 5374 static EVT getExtensionTo64Bits(const EVT &OrigVT) { 5375 if (OrigVT.getSizeInBits() >= 64) 5376 return OrigVT; 5377 5378 assert(OrigVT.isSimple() && "Expecting a simple value type"); 5379 5380 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy; 5381 switch (OrigSimpleTy) { 5382 default: llvm_unreachable("Unexpected Vector Type"); 5383 case MVT::v2i8: 5384 case MVT::v2i16: 5385 return MVT::v2i32; 5386 case MVT::v4i8: 5387 return MVT::v4i16; 5388 } 5389 } 5390 5391 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total 5392 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL. 5393 /// We insert the required extension here to get the vector to fill a D register. 5394 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG, 5395 const EVT &OrigTy, 5396 const EVT &ExtTy, 5397 unsigned ExtOpcode) { 5398 // The vector originally had a size of OrigTy. It was then extended to ExtTy. 5399 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than 5400 // 64-bits we need to insert a new extension so that it will be 64-bits. 5401 assert(ExtTy.is128BitVector() && "Unexpected extension size"); 5402 if (OrigTy.getSizeInBits() >= 64) 5403 return N; 5404 5405 // Must extend size to at least 64 bits to be used as an operand for VMULL. 5406 EVT NewVT = getExtensionTo64Bits(OrigTy); 5407 5408 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N); 5409 } 5410 5411 /// SkipLoadExtensionForVMULL - return a load of the original vector size that 5412 /// does not do any sign/zero extension. If the original vector is less 5413 /// than 64 bits, an appropriate extension will be added after the load to 5414 /// reach a total size of 64 bits. We have to add the extension separately 5415 /// because ARM does not have a sign/zero extending load for vectors. 5416 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) { 5417 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT()); 5418 5419 // The load already has the right type. 5420 if (ExtendedTy == LD->getMemoryVT()) 5421 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(), 5422 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(), 5423 LD->isNonTemporal(), LD->isInvariant(), 5424 LD->getAlignment()); 5425 5426 // We need to create a zextload/sextload. We cannot just create a load 5427 // followed by a zext/zext node because LowerMUL is also run during normal 5428 // operation legalization where we can't create illegal types. 5429 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy, 5430 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(), 5431 LD->getMemoryVT(), LD->isVolatile(), 5432 LD->isNonTemporal(), LD->getAlignment()); 5433 } 5434 5435 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND, 5436 /// extending load, or BUILD_VECTOR with extended elements, return the 5437 /// unextended value. The unextended vector should be 64 bits so that it can 5438 /// be used as an operand to a VMULL instruction. If the original vector size 5439 /// before extension is less than 64 bits we add a an extension to resize 5440 /// the vector to 64 bits. 5441 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) { 5442 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND) 5443 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG, 5444 N->getOperand(0)->getValueType(0), 5445 N->getValueType(0), 5446 N->getOpcode()); 5447 5448 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) 5449 return SkipLoadExtensionForVMULL(LD, DAG); 5450 5451 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will 5452 // have been legalized as a BITCAST from v4i32. 5453 if (N->getOpcode() == ISD::BITCAST) { 5454 SDNode *BVN = N->getOperand(0).getNode(); 5455 assert(BVN->getOpcode() == ISD::BUILD_VECTOR && 5456 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR"); 5457 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0; 5458 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32, 5459 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2)); 5460 } 5461 // Construct a new BUILD_VECTOR with elements truncated to half the size. 5462 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR"); 5463 EVT VT = N->getValueType(0); 5464 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2; 5465 unsigned NumElts = VT.getVectorNumElements(); 5466 MVT TruncVT = MVT::getIntegerVT(EltSize); 5467 SmallVector<SDValue, 8> Ops; 5468 for (unsigned i = 0; i != NumElts; ++i) { 5469 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i)); 5470 const APInt &CInt = C->getAPIntValue(); 5471 // Element types smaller than 32 bits are not legal, so use i32 elements. 5472 // The values are implicitly truncated so sext vs. zext doesn't matter. 5473 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32)); 5474 } 5475 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), 5476 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts); 5477 } 5478 5479 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) { 5480 unsigned Opcode = N->getOpcode(); 5481 if (Opcode == ISD::ADD || Opcode == ISD::SUB) { 5482 SDNode *N0 = N->getOperand(0).getNode(); 5483 SDNode *N1 = N->getOperand(1).getNode(); 5484 return N0->hasOneUse() && N1->hasOneUse() && 5485 isSignExtended(N0, DAG) && isSignExtended(N1, DAG); 5486 } 5487 return false; 5488 } 5489 5490 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) { 5491 unsigned Opcode = N->getOpcode(); 5492 if (Opcode == ISD::ADD || Opcode == ISD::SUB) { 5493 SDNode *N0 = N->getOperand(0).getNode(); 5494 SDNode *N1 = N->getOperand(1).getNode(); 5495 return N0->hasOneUse() && N1->hasOneUse() && 5496 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG); 5497 } 5498 return false; 5499 } 5500 5501 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) { 5502 // Multiplications are only custom-lowered for 128-bit vectors so that 5503 // VMULL can be detected. Otherwise v2i64 multiplications are not legal. 5504 EVT VT = Op.getValueType(); 5505 assert(VT.is128BitVector() && VT.isInteger() && 5506 "unexpected type for custom-lowering ISD::MUL"); 5507 SDNode *N0 = Op.getOperand(0).getNode(); 5508 SDNode *N1 = Op.getOperand(1).getNode(); 5509 unsigned NewOpc = 0; 5510 bool isMLA = false; 5511 bool isN0SExt = isSignExtended(N0, DAG); 5512 bool isN1SExt = isSignExtended(N1, DAG); 5513 if (isN0SExt && isN1SExt) 5514 NewOpc = ARMISD::VMULLs; 5515 else { 5516 bool isN0ZExt = isZeroExtended(N0, DAG); 5517 bool isN1ZExt = isZeroExtended(N1, DAG); 5518 if (isN0ZExt && isN1ZExt) 5519 NewOpc = ARMISD::VMULLu; 5520 else if (isN1SExt || isN1ZExt) { 5521 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these 5522 // into (s/zext A * s/zext C) + (s/zext B * s/zext C) 5523 if (isN1SExt && isAddSubSExt(N0, DAG)) { 5524 NewOpc = ARMISD::VMULLs; 5525 isMLA = true; 5526 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) { 5527 NewOpc = ARMISD::VMULLu; 5528 isMLA = true; 5529 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) { 5530 std::swap(N0, N1); 5531 NewOpc = ARMISD::VMULLu; 5532 isMLA = true; 5533 } 5534 } 5535 5536 if (!NewOpc) { 5537 if (VT == MVT::v2i64) 5538 // Fall through to expand this. It is not legal. 5539 return SDValue(); 5540 else 5541 // Other vector multiplications are legal. 5542 return Op; 5543 } 5544 } 5545 5546 // Legalize to a VMULL instruction. 5547 SDLoc DL(Op); 5548 SDValue Op0; 5549 SDValue Op1 = SkipExtensionForVMULL(N1, DAG); 5550 if (!isMLA) { 5551 Op0 = SkipExtensionForVMULL(N0, DAG); 5552 assert(Op0.getValueType().is64BitVector() && 5553 Op1.getValueType().is64BitVector() && 5554 "unexpected types for extended operands to VMULL"); 5555 return DAG.getNode(NewOpc, DL, VT, Op0, Op1); 5556 } 5557 5558 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during 5559 // isel lowering to take advantage of no-stall back to back vmul + vmla. 5560 // vmull q0, d4, d6 5561 // vmlal q0, d5, d6 5562 // is faster than 5563 // vaddl q0, d4, d5 5564 // vmovl q1, d6 5565 // vmul q0, q0, q1 5566 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG); 5567 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG); 5568 EVT Op1VT = Op1.getValueType(); 5569 return DAG.getNode(N0->getOpcode(), DL, VT, 5570 DAG.getNode(NewOpc, DL, VT, 5571 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1), 5572 DAG.getNode(NewOpc, DL, VT, 5573 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1)); 5574 } 5575 5576 static SDValue 5577 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) { 5578 // Convert to float 5579 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo)); 5580 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo)); 5581 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X); 5582 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y); 5583 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X); 5584 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y); 5585 // Get reciprocal estimate. 5586 // float4 recip = vrecpeq_f32(yf); 5587 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5588 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y); 5589 // Because char has a smaller range than uchar, we can actually get away 5590 // without any newton steps. This requires that we use a weird bias 5591 // of 0xb000, however (again, this has been exhaustively tested). 5592 // float4 result = as_float4(as_int4(xf*recip) + 0xb000); 5593 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y); 5594 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X); 5595 Y = DAG.getConstant(0xb000, MVT::i32); 5596 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y); 5597 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y); 5598 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X); 5599 // Convert back to short. 5600 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X); 5601 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X); 5602 return X; 5603 } 5604 5605 static SDValue 5606 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) { 5607 SDValue N2; 5608 // Convert to float. 5609 // float4 yf = vcvt_f32_s32(vmovl_s16(y)); 5610 // float4 xf = vcvt_f32_s32(vmovl_s16(x)); 5611 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0); 5612 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1); 5613 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0); 5614 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1); 5615 5616 // Use reciprocal estimate and one refinement step. 5617 // float4 recip = vrecpeq_f32(yf); 5618 // recip *= vrecpsq_f32(yf, recip); 5619 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5620 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1); 5621 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5622 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), 5623 N1, N2); 5624 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); 5625 // Because short has a smaller range than ushort, we can actually get away 5626 // with only a single newton step. This requires that we use a weird bias 5627 // of 89, however (again, this has been exhaustively tested). 5628 // float4 result = as_float4(as_int4(xf*recip) + 0x89); 5629 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2); 5630 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0); 5631 N1 = DAG.getConstant(0x89, MVT::i32); 5632 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1); 5633 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1); 5634 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0); 5635 // Convert back to integer and return. 5636 // return vmovn_s32(vcvt_s32_f32(result)); 5637 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0); 5638 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0); 5639 return N0; 5640 } 5641 5642 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) { 5643 EVT VT = Op.getValueType(); 5644 assert((VT == MVT::v4i16 || VT == MVT::v8i8) && 5645 "unexpected type for custom-lowering ISD::SDIV"); 5646 5647 SDLoc dl(Op); 5648 SDValue N0 = Op.getOperand(0); 5649 SDValue N1 = Op.getOperand(1); 5650 SDValue N2, N3; 5651 5652 if (VT == MVT::v8i8) { 5653 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0); 5654 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1); 5655 5656 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5657 DAG.getIntPtrConstant(4)); 5658 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5659 DAG.getIntPtrConstant(4)); 5660 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5661 DAG.getIntPtrConstant(0)); 5662 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5663 DAG.getIntPtrConstant(0)); 5664 5665 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16 5666 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16 5667 5668 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2); 5669 N0 = LowerCONCAT_VECTORS(N0, DAG); 5670 5671 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0); 5672 return N0; 5673 } 5674 return LowerSDIV_v4i16(N0, N1, dl, DAG); 5675 } 5676 5677 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) { 5678 EVT VT = Op.getValueType(); 5679 assert((VT == MVT::v4i16 || VT == MVT::v8i8) && 5680 "unexpected type for custom-lowering ISD::UDIV"); 5681 5682 SDLoc dl(Op); 5683 SDValue N0 = Op.getOperand(0); 5684 SDValue N1 = Op.getOperand(1); 5685 SDValue N2, N3; 5686 5687 if (VT == MVT::v8i8) { 5688 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0); 5689 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1); 5690 5691 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5692 DAG.getIntPtrConstant(4)); 5693 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5694 DAG.getIntPtrConstant(4)); 5695 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0, 5696 DAG.getIntPtrConstant(0)); 5697 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1, 5698 DAG.getIntPtrConstant(0)); 5699 5700 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16 5701 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16 5702 5703 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2); 5704 N0 = LowerCONCAT_VECTORS(N0, DAG); 5705 5706 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8, 5707 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32), 5708 N0); 5709 return N0; 5710 } 5711 5712 // v4i16 sdiv ... Convert to float. 5713 // float4 yf = vcvt_f32_s32(vmovl_u16(y)); 5714 // float4 xf = vcvt_f32_s32(vmovl_u16(x)); 5715 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0); 5716 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1); 5717 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0); 5718 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1); 5719 5720 // Use reciprocal estimate and two refinement steps. 5721 // float4 recip = vrecpeq_f32(yf); 5722 // recip *= vrecpsq_f32(yf, recip); 5723 // recip *= vrecpsq_f32(yf, recip); 5724 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5725 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1); 5726 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5727 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), 5728 BN1, N2); 5729 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); 5730 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32, 5731 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32), 5732 BN1, N2); 5733 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2); 5734 // Simply multiplying by the reciprocal estimate can leave us a few ulps 5735 // too low, so we add 2 ulps (exhaustive testing shows that this is enough, 5736 // and that it will never cause us to return an answer too large). 5737 // float4 result = as_float4(as_int4(xf*recip) + 2); 5738 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2); 5739 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0); 5740 N1 = DAG.getConstant(2, MVT::i32); 5741 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1); 5742 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1); 5743 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0); 5744 // Convert back to integer and return. 5745 // return vmovn_u32(vcvt_s32_f32(result)); 5746 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0); 5747 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0); 5748 return N0; 5749 } 5750 5751 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) { 5752 EVT VT = Op.getNode()->getValueType(0); 5753 SDVTList VTs = DAG.getVTList(VT, MVT::i32); 5754 5755 unsigned Opc; 5756 bool ExtraOp = false; 5757 switch (Op.getOpcode()) { 5758 default: llvm_unreachable("Invalid code"); 5759 case ISD::ADDC: Opc = ARMISD::ADDC; break; 5760 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break; 5761 case ISD::SUBC: Opc = ARMISD::SUBC; break; 5762 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break; 5763 } 5764 5765 if (!ExtraOp) 5766 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), 5767 Op.getOperand(1)); 5768 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), 5769 Op.getOperand(1), Op.getOperand(2)); 5770 } 5771 5772 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) { 5773 // Monotonic load/store is legal for all targets 5774 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic) 5775 return Op; 5776 5777 // Aquire/Release load/store is not legal for targets without a 5778 // dmb or equivalent available. 5779 return SDValue(); 5780 } 5781 5782 static void 5783 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results, 5784 SelectionDAG &DAG, unsigned NewOp) { 5785 SDLoc dl(Node); 5786 assert (Node->getValueType(0) == MVT::i64 && 5787 "Only know how to expand i64 atomics"); 5788 5789 SmallVector<SDValue, 6> Ops; 5790 Ops.push_back(Node->getOperand(0)); // Chain 5791 Ops.push_back(Node->getOperand(1)); // Ptr 5792 // Low part of Val1 5793 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5794 Node->getOperand(2), DAG.getIntPtrConstant(0))); 5795 // High part of Val1 5796 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5797 Node->getOperand(2), DAG.getIntPtrConstant(1))); 5798 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) { 5799 // High part of Val1 5800 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5801 Node->getOperand(3), DAG.getIntPtrConstant(0))); 5802 // High part of Val2 5803 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, 5804 Node->getOperand(3), DAG.getIntPtrConstant(1))); 5805 } 5806 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other); 5807 SDValue Result = 5808 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64, 5809 cast<MemSDNode>(Node)->getMemOperand()); 5810 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) }; 5811 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2)); 5812 Results.push_back(Result.getValue(2)); 5813 } 5814 5815 static void ReplaceREADCYCLECOUNTER(SDNode *N, 5816 SmallVectorImpl<SDValue> &Results, 5817 SelectionDAG &DAG, 5818 const ARMSubtarget *Subtarget) { 5819 SDLoc DL(N); 5820 SDValue Cycles32, OutChain; 5821 5822 if (Subtarget->hasPerfMon()) { 5823 // Under Power Management extensions, the cycle-count is: 5824 // mrc p15, #0, <Rt>, c9, c13, #0 5825 SDValue Ops[] = { N->getOperand(0), // Chain 5826 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32), 5827 DAG.getConstant(15, MVT::i32), 5828 DAG.getConstant(0, MVT::i32), 5829 DAG.getConstant(9, MVT::i32), 5830 DAG.getConstant(13, MVT::i32), 5831 DAG.getConstant(0, MVT::i32) 5832 }; 5833 5834 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL, 5835 DAG.getVTList(MVT::i32, MVT::Other), &Ops[0], 5836 array_lengthof(Ops)); 5837 OutChain = Cycles32.getValue(1); 5838 } else { 5839 // Intrinsic is defined to return 0 on unsupported platforms. Technically 5840 // there are older ARM CPUs that have implementation-specific ways of 5841 // obtaining this information (FIXME!). 5842 Cycles32 = DAG.getConstant(0, MVT::i32); 5843 OutChain = DAG.getEntryNode(); 5844 } 5845 5846 5847 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, 5848 Cycles32, DAG.getConstant(0, MVT::i32)); 5849 Results.push_back(Cycles64); 5850 Results.push_back(OutChain); 5851 } 5852 5853 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { 5854 switch (Op.getOpcode()) { 5855 default: llvm_unreachable("Don't know how to custom lower this!"); 5856 case ISD::ConstantPool: return LowerConstantPool(Op, DAG); 5857 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); 5858 case ISD::GlobalAddress: 5859 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) : 5860 LowerGlobalAddressELF(Op, DAG); 5861 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); 5862 case ISD::SELECT: return LowerSELECT(Op, DAG); 5863 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG); 5864 case ISD::BR_CC: return LowerBR_CC(Op, DAG); 5865 case ISD::BR_JT: return LowerBR_JT(Op, DAG); 5866 case ISD::VASTART: return LowerVASTART(Op, DAG); 5867 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget); 5868 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget); 5869 case ISD::SINT_TO_FP: 5870 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG); 5871 case ISD::FP_TO_SINT: 5872 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG); 5873 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG); 5874 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG); 5875 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); 5876 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG); 5877 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG); 5878 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG); 5879 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG, 5880 Subtarget); 5881 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG); 5882 case ISD::SHL: 5883 case ISD::SRL: 5884 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget); 5885 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG); 5886 case ISD::SRL_PARTS: 5887 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG); 5888 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget); 5889 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget); 5890 case ISD::SETCC: return LowerVSETCC(Op, DAG); 5891 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget); 5892 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget); 5893 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG); 5894 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG); 5895 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG); 5896 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG); 5897 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG); 5898 case ISD::MUL: return LowerMUL(Op, DAG); 5899 case ISD::SDIV: return LowerSDIV(Op, DAG); 5900 case ISD::UDIV: return LowerUDIV(Op, DAG); 5901 case ISD::ADDC: 5902 case ISD::ADDE: 5903 case ISD::SUBC: 5904 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG); 5905 case ISD::ATOMIC_LOAD: 5906 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG); 5907 case ISD::SDIVREM: 5908 case ISD::UDIVREM: return LowerDivRem(Op, DAG); 5909 } 5910 } 5911 5912 /// ReplaceNodeResults - Replace the results of node with an illegal result 5913 /// type with new values built out of custom code. 5914 void ARMTargetLowering::ReplaceNodeResults(SDNode *N, 5915 SmallVectorImpl<SDValue>&Results, 5916 SelectionDAG &DAG) const { 5917 SDValue Res; 5918 switch (N->getOpcode()) { 5919 default: 5920 llvm_unreachable("Don't know how to custom expand this!"); 5921 case ISD::BITCAST: 5922 Res = ExpandBITCAST(N, DAG); 5923 break; 5924 case ISD::SIGN_EXTEND: 5925 case ISD::ZERO_EXTEND: 5926 Res = ExpandVectorExtension(N, DAG); 5927 break; 5928 case ISD::SRL: 5929 case ISD::SRA: 5930 Res = Expand64BitShift(N, DAG, Subtarget); 5931 break; 5932 case ISD::READCYCLECOUNTER: 5933 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget); 5934 return; 5935 case ISD::ATOMIC_LOAD_ADD: 5936 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG); 5937 return; 5938 case ISD::ATOMIC_LOAD_AND: 5939 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG); 5940 return; 5941 case ISD::ATOMIC_LOAD_NAND: 5942 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG); 5943 return; 5944 case ISD::ATOMIC_LOAD_OR: 5945 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG); 5946 return; 5947 case ISD::ATOMIC_LOAD_SUB: 5948 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG); 5949 return; 5950 case ISD::ATOMIC_LOAD_XOR: 5951 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG); 5952 return; 5953 case ISD::ATOMIC_SWAP: 5954 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG); 5955 return; 5956 case ISD::ATOMIC_CMP_SWAP: 5957 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG); 5958 return; 5959 case ISD::ATOMIC_LOAD_MIN: 5960 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMIN64_DAG); 5961 return; 5962 case ISD::ATOMIC_LOAD_UMIN: 5963 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMIN64_DAG); 5964 return; 5965 case ISD::ATOMIC_LOAD_MAX: 5966 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMMAX64_DAG); 5967 return; 5968 case ISD::ATOMIC_LOAD_UMAX: 5969 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMUMAX64_DAG); 5970 return; 5971 } 5972 if (Res.getNode()) 5973 Results.push_back(Res); 5974 } 5975 5976 //===----------------------------------------------------------------------===// 5977 // ARM Scheduler Hooks 5978 //===----------------------------------------------------------------------===// 5979 5980 MachineBasicBlock * 5981 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI, 5982 MachineBasicBlock *BB, 5983 unsigned Size) const { 5984 unsigned dest = MI->getOperand(0).getReg(); 5985 unsigned ptr = MI->getOperand(1).getReg(); 5986 unsigned oldval = MI->getOperand(2).getReg(); 5987 unsigned newval = MI->getOperand(3).getReg(); 5988 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 5989 DebugLoc dl = MI->getDebugLoc(); 5990 bool isThumb2 = Subtarget->isThumb2(); 5991 5992 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 5993 unsigned scratch = MRI.createVirtualRegister(isThumb2 ? 5994 (const TargetRegisterClass*)&ARM::rGPRRegClass : 5995 (const TargetRegisterClass*)&ARM::GPRRegClass); 5996 5997 if (isThumb2) { 5998 MRI.constrainRegClass(dest, &ARM::rGPRRegClass); 5999 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass); 6000 MRI.constrainRegClass(newval, &ARM::rGPRRegClass); 6001 } 6002 6003 unsigned ldrOpc, strOpc; 6004 switch (Size) { 6005 default: llvm_unreachable("unsupported size for AtomicCmpSwap!"); 6006 case 1: 6007 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB; 6008 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB; 6009 break; 6010 case 2: 6011 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH; 6012 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH; 6013 break; 6014 case 4: 6015 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX; 6016 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX; 6017 break; 6018 } 6019 6020 MachineFunction *MF = BB->getParent(); 6021 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 6022 MachineFunction::iterator It = BB; 6023 ++It; // insert the new blocks after the current block 6024 6025 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 6026 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 6027 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 6028 MF->insert(It, loop1MBB); 6029 MF->insert(It, loop2MBB); 6030 MF->insert(It, exitMBB); 6031 6032 // Transfer the remainder of BB and its successor edges to exitMBB. 6033 exitMBB->splice(exitMBB->begin(), BB, 6034 llvm::next(MachineBasicBlock::iterator(MI)), 6035 BB->end()); 6036 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 6037 6038 // thisMBB: 6039 // ... 6040 // fallthrough --> loop1MBB 6041 BB->addSuccessor(loop1MBB); 6042 6043 // loop1MBB: 6044 // ldrex dest, [ptr] 6045 // cmp dest, oldval 6046 // bne exitMBB 6047 BB = loop1MBB; 6048 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr); 6049 if (ldrOpc == ARM::t2LDREX) 6050 MIB.addImm(0); 6051 AddDefaultPred(MIB); 6052 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) 6053 .addReg(dest).addReg(oldval)); 6054 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 6055 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 6056 BB->addSuccessor(loop2MBB); 6057 BB->addSuccessor(exitMBB); 6058 6059 // loop2MBB: 6060 // strex scratch, newval, [ptr] 6061 // cmp scratch, #0 6062 // bne loop1MBB 6063 BB = loop2MBB; 6064 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr); 6065 if (strOpc == ARM::t2STREX) 6066 MIB.addImm(0); 6067 AddDefaultPred(MIB); 6068 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 6069 .addReg(scratch).addImm(0)); 6070 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 6071 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 6072 BB->addSuccessor(loop1MBB); 6073 BB->addSuccessor(exitMBB); 6074 6075 // exitMBB: 6076 // ... 6077 BB = exitMBB; 6078 6079 MI->eraseFromParent(); // The instruction is gone now. 6080 6081 return BB; 6082 } 6083 6084 MachineBasicBlock * 6085 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB, 6086 unsigned Size, unsigned BinOpcode) const { 6087 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0. 6088 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 6089 6090 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 6091 MachineFunction *MF = BB->getParent(); 6092 MachineFunction::iterator It = BB; 6093 ++It; 6094 6095 unsigned dest = MI->getOperand(0).getReg(); 6096 unsigned ptr = MI->getOperand(1).getReg(); 6097 unsigned incr = MI->getOperand(2).getReg(); 6098 DebugLoc dl = MI->getDebugLoc(); 6099 bool isThumb2 = Subtarget->isThumb2(); 6100 6101 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 6102 if (isThumb2) { 6103 MRI.constrainRegClass(dest, &ARM::rGPRRegClass); 6104 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass); 6105 } 6106 6107 unsigned ldrOpc, strOpc; 6108 switch (Size) { 6109 default: llvm_unreachable("unsupported size for AtomicCmpSwap!"); 6110 case 1: 6111 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB; 6112 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB; 6113 break; 6114 case 2: 6115 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH; 6116 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH; 6117 break; 6118 case 4: 6119 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX; 6120 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX; 6121 break; 6122 } 6123 6124 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 6125 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 6126 MF->insert(It, loopMBB); 6127 MF->insert(It, exitMBB); 6128 6129 // Transfer the remainder of BB and its successor edges to exitMBB. 6130 exitMBB->splice(exitMBB->begin(), BB, 6131 llvm::next(MachineBasicBlock::iterator(MI)), 6132 BB->end()); 6133 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 6134 6135 const TargetRegisterClass *TRC = isThumb2 ? 6136 (const TargetRegisterClass*)&ARM::rGPRRegClass : 6137 (const TargetRegisterClass*)&ARM::GPRRegClass; 6138 unsigned scratch = MRI.createVirtualRegister(TRC); 6139 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC); 6140 6141 // thisMBB: 6142 // ... 6143 // fallthrough --> loopMBB 6144 BB->addSuccessor(loopMBB); 6145 6146 // loopMBB: 6147 // ldrex dest, ptr 6148 // <binop> scratch2, dest, incr 6149 // strex scratch, scratch2, ptr 6150 // cmp scratch, #0 6151 // bne- loopMBB 6152 // fallthrough --> exitMBB 6153 BB = loopMBB; 6154 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr); 6155 if (ldrOpc == ARM::t2LDREX) 6156 MIB.addImm(0); 6157 AddDefaultPred(MIB); 6158 if (BinOpcode) { 6159 // operand order needs to go the other way for NAND 6160 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr) 6161 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2). 6162 addReg(incr).addReg(dest)).addReg(0); 6163 else 6164 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2). 6165 addReg(dest).addReg(incr)).addReg(0); 6166 } 6167 6168 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr); 6169 if (strOpc == ARM::t2STREX) 6170 MIB.addImm(0); 6171 AddDefaultPred(MIB); 6172 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 6173 .addReg(scratch).addImm(0)); 6174 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 6175 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 6176 6177 BB->addSuccessor(loopMBB); 6178 BB->addSuccessor(exitMBB); 6179 6180 // exitMBB: 6181 // ... 6182 BB = exitMBB; 6183 6184 MI->eraseFromParent(); // The instruction is gone now. 6185 6186 return BB; 6187 } 6188 6189 MachineBasicBlock * 6190 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI, 6191 MachineBasicBlock *BB, 6192 unsigned Size, 6193 bool signExtend, 6194 ARMCC::CondCodes Cond) const { 6195 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 6196 6197 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 6198 MachineFunction *MF = BB->getParent(); 6199 MachineFunction::iterator It = BB; 6200 ++It; 6201 6202 unsigned dest = MI->getOperand(0).getReg(); 6203 unsigned ptr = MI->getOperand(1).getReg(); 6204 unsigned incr = MI->getOperand(2).getReg(); 6205 unsigned oldval = dest; 6206 DebugLoc dl = MI->getDebugLoc(); 6207 bool isThumb2 = Subtarget->isThumb2(); 6208 6209 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 6210 if (isThumb2) { 6211 MRI.constrainRegClass(dest, &ARM::rGPRRegClass); 6212 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass); 6213 } 6214 6215 unsigned ldrOpc, strOpc, extendOpc; 6216 switch (Size) { 6217 default: llvm_unreachable("unsupported size for AtomicCmpSwap!"); 6218 case 1: 6219 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB; 6220 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB; 6221 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB; 6222 break; 6223 case 2: 6224 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH; 6225 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH; 6226 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH; 6227 break; 6228 case 4: 6229 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX; 6230 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX; 6231 extendOpc = 0; 6232 break; 6233 } 6234 6235 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 6236 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 6237 MF->insert(It, loopMBB); 6238 MF->insert(It, exitMBB); 6239 6240 // Transfer the remainder of BB and its successor edges to exitMBB. 6241 exitMBB->splice(exitMBB->begin(), BB, 6242 llvm::next(MachineBasicBlock::iterator(MI)), 6243 BB->end()); 6244 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 6245 6246 const TargetRegisterClass *TRC = isThumb2 ? 6247 (const TargetRegisterClass*)&ARM::rGPRRegClass : 6248 (const TargetRegisterClass*)&ARM::GPRRegClass; 6249 unsigned scratch = MRI.createVirtualRegister(TRC); 6250 unsigned scratch2 = MRI.createVirtualRegister(TRC); 6251 6252 // thisMBB: 6253 // ... 6254 // fallthrough --> loopMBB 6255 BB->addSuccessor(loopMBB); 6256 6257 // loopMBB: 6258 // ldrex dest, ptr 6259 // (sign extend dest, if required) 6260 // cmp dest, incr 6261 // cmov.cond scratch2, incr, dest 6262 // strex scratch, scratch2, ptr 6263 // cmp scratch, #0 6264 // bne- loopMBB 6265 // fallthrough --> exitMBB 6266 BB = loopMBB; 6267 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr); 6268 if (ldrOpc == ARM::t2LDREX) 6269 MIB.addImm(0); 6270 AddDefaultPred(MIB); 6271 6272 // Sign extend the value, if necessary. 6273 if (signExtend && extendOpc) { 6274 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass); 6275 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval) 6276 .addReg(dest) 6277 .addImm(0)); 6278 } 6279 6280 // Build compare and cmov instructions. 6281 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) 6282 .addReg(oldval).addReg(incr)); 6283 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2) 6284 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR); 6285 6286 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr); 6287 if (strOpc == ARM::t2STREX) 6288 MIB.addImm(0); 6289 AddDefaultPred(MIB); 6290 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 6291 .addReg(scratch).addImm(0)); 6292 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 6293 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 6294 6295 BB->addSuccessor(loopMBB); 6296 BB->addSuccessor(exitMBB); 6297 6298 // exitMBB: 6299 // ... 6300 BB = exitMBB; 6301 6302 MI->eraseFromParent(); // The instruction is gone now. 6303 6304 return BB; 6305 } 6306 6307 MachineBasicBlock * 6308 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB, 6309 unsigned Op1, unsigned Op2, 6310 bool NeedsCarry, bool IsCmpxchg, 6311 bool IsMinMax, ARMCC::CondCodes CC) const { 6312 // This also handles ATOMIC_SWAP, indicated by Op1==0. 6313 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 6314 6315 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 6316 MachineFunction *MF = BB->getParent(); 6317 MachineFunction::iterator It = BB; 6318 ++It; 6319 6320 unsigned destlo = MI->getOperand(0).getReg(); 6321 unsigned desthi = MI->getOperand(1).getReg(); 6322 unsigned ptr = MI->getOperand(2).getReg(); 6323 unsigned vallo = MI->getOperand(3).getReg(); 6324 unsigned valhi = MI->getOperand(4).getReg(); 6325 DebugLoc dl = MI->getDebugLoc(); 6326 bool isThumb2 = Subtarget->isThumb2(); 6327 6328 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); 6329 if (isThumb2) { 6330 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass); 6331 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass); 6332 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass); 6333 } 6334 6335 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 6336 MachineBasicBlock *contBB = 0, *cont2BB = 0; 6337 if (IsCmpxchg || IsMinMax) 6338 contBB = MF->CreateMachineBasicBlock(LLVM_BB); 6339 if (IsCmpxchg) 6340 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB); 6341 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 6342 6343 MF->insert(It, loopMBB); 6344 if (IsCmpxchg || IsMinMax) MF->insert(It, contBB); 6345 if (IsCmpxchg) MF->insert(It, cont2BB); 6346 MF->insert(It, exitMBB); 6347 6348 // Transfer the remainder of BB and its successor edges to exitMBB. 6349 exitMBB->splice(exitMBB->begin(), BB, 6350 llvm::next(MachineBasicBlock::iterator(MI)), 6351 BB->end()); 6352 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 6353 6354 const TargetRegisterClass *TRC = isThumb2 ? 6355 (const TargetRegisterClass*)&ARM::tGPRRegClass : 6356 (const TargetRegisterClass*)&ARM::GPRRegClass; 6357 unsigned storesuccess = MRI.createVirtualRegister(TRC); 6358 6359 // thisMBB: 6360 // ... 6361 // fallthrough --> loopMBB 6362 BB->addSuccessor(loopMBB); 6363 6364 // loopMBB: 6365 // ldrexd r2, r3, ptr 6366 // <binopa> r0, r2, incr 6367 // <binopb> r1, r3, incr 6368 // strexd storesuccess, r0, r1, ptr 6369 // cmp storesuccess, #0 6370 // bne- loopMBB 6371 // fallthrough --> exitMBB 6372 BB = loopMBB; 6373 6374 // Load 6375 if (isThumb2) { 6376 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2LDREXD)) 6377 .addReg(destlo, RegState::Define) 6378 .addReg(desthi, RegState::Define) 6379 .addReg(ptr)); 6380 } else { 6381 unsigned GPRPair0 = MRI.createVirtualRegister(&ARM::GPRPairRegClass); 6382 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDREXD)) 6383 .addReg(GPRPair0, RegState::Define).addReg(ptr)); 6384 // Copy r2/r3 into dest. (This copy will normally be coalesced.) 6385 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo) 6386 .addReg(GPRPair0, 0, ARM::gsub_0); 6387 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi) 6388 .addReg(GPRPair0, 0, ARM::gsub_1); 6389 } 6390 6391 unsigned StoreLo, StoreHi; 6392 if (IsCmpxchg) { 6393 // Add early exit 6394 for (unsigned i = 0; i < 2; i++) { 6395 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : 6396 ARM::CMPrr)) 6397 .addReg(i == 0 ? destlo : desthi) 6398 .addReg(i == 0 ? vallo : valhi)); 6399 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 6400 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 6401 BB->addSuccessor(exitMBB); 6402 BB->addSuccessor(i == 0 ? contBB : cont2BB); 6403 BB = (i == 0 ? contBB : cont2BB); 6404 } 6405 6406 // Copy to physregs for strexd 6407 StoreLo = MI->getOperand(5).getReg(); 6408 StoreHi = MI->getOperand(6).getReg(); 6409 } else if (Op1) { 6410 // Perform binary operation 6411 unsigned tmpRegLo = MRI.createVirtualRegister(TRC); 6412 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), tmpRegLo) 6413 .addReg(destlo).addReg(vallo)) 6414 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry)); 6415 unsigned tmpRegHi = MRI.createVirtualRegister(TRC); 6416 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), tmpRegHi) 6417 .addReg(desthi).addReg(valhi)) 6418 .addReg(IsMinMax ? ARM::CPSR : 0, getDefRegState(IsMinMax)); 6419 6420 StoreLo = tmpRegLo; 6421 StoreHi = tmpRegHi; 6422 } else { 6423 // Copy to physregs for strexd 6424 StoreLo = vallo; 6425 StoreHi = valhi; 6426 } 6427 if (IsMinMax) { 6428 // Compare and branch to exit block. 6429 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 6430 .addMBB(exitMBB).addImm(CC).addReg(ARM::CPSR); 6431 BB->addSuccessor(exitMBB); 6432 BB->addSuccessor(contBB); 6433 BB = contBB; 6434 StoreLo = vallo; 6435 StoreHi = valhi; 6436 } 6437 6438 // Store 6439 if (isThumb2) { 6440 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2STREXD), storesuccess) 6441 .addReg(StoreLo).addReg(StoreHi).addReg(ptr)); 6442 } else { 6443 // Marshal a pair... 6444 unsigned StorePair = MRI.createVirtualRegister(&ARM::GPRPairRegClass); 6445 unsigned UndefPair = MRI.createVirtualRegister(&ARM::GPRPairRegClass); 6446 unsigned r1 = MRI.createVirtualRegister(&ARM::GPRPairRegClass); 6447 BuildMI(BB, dl, TII->get(TargetOpcode::IMPLICIT_DEF), UndefPair); 6448 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), r1) 6449 .addReg(UndefPair) 6450 .addReg(StoreLo) 6451 .addImm(ARM::gsub_0); 6452 BuildMI(BB, dl, TII->get(TargetOpcode::INSERT_SUBREG), StorePair) 6453 .addReg(r1) 6454 .addReg(StoreHi) 6455 .addImm(ARM::gsub_1); 6456 6457 // ...and store it 6458 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::STREXD), storesuccess) 6459 .addReg(StorePair).addReg(ptr)); 6460 } 6461 // Cmp+jump 6462 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 6463 .addReg(storesuccess).addImm(0)); 6464 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 6465 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 6466 6467 BB->addSuccessor(loopMBB); 6468 BB->addSuccessor(exitMBB); 6469 6470 // exitMBB: 6471 // ... 6472 BB = exitMBB; 6473 6474 MI->eraseFromParent(); // The instruction is gone now. 6475 6476 return BB; 6477 } 6478 6479 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and 6480 /// registers the function context. 6481 void ARMTargetLowering:: 6482 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB, 6483 MachineBasicBlock *DispatchBB, int FI) const { 6484 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 6485 DebugLoc dl = MI->getDebugLoc(); 6486 MachineFunction *MF = MBB->getParent(); 6487 MachineRegisterInfo *MRI = &MF->getRegInfo(); 6488 MachineConstantPool *MCP = MF->getConstantPool(); 6489 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>(); 6490 const Function *F = MF->getFunction(); 6491 6492 bool isThumb = Subtarget->isThumb(); 6493 bool isThumb2 = Subtarget->isThumb2(); 6494 6495 unsigned PCLabelId = AFI->createPICLabelUId(); 6496 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8; 6497 ARMConstantPoolValue *CPV = 6498 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj); 6499 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4); 6500 6501 const TargetRegisterClass *TRC = isThumb ? 6502 (const TargetRegisterClass*)&ARM::tGPRRegClass : 6503 (const TargetRegisterClass*)&ARM::GPRRegClass; 6504 6505 // Grab constant pool and fixed stack memory operands. 6506 MachineMemOperand *CPMMO = 6507 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(), 6508 MachineMemOperand::MOLoad, 4, 4); 6509 6510 MachineMemOperand *FIMMOSt = 6511 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI), 6512 MachineMemOperand::MOStore, 4, 4); 6513 6514 // Load the address of the dispatch MBB into the jump buffer. 6515 if (isThumb2) { 6516 // Incoming value: jbuf 6517 // ldr.n r5, LCPI1_1 6518 // orr r5, r5, #1 6519 // add r5, pc 6520 // str r5, [$jbuf, #+4] ; &jbuf[1] 6521 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 6522 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1) 6523 .addConstantPoolIndex(CPI) 6524 .addMemOperand(CPMMO)); 6525 // Set the low bit because of thumb mode. 6526 unsigned NewVReg2 = MRI->createVirtualRegister(TRC); 6527 AddDefaultCC( 6528 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2) 6529 .addReg(NewVReg1, RegState::Kill) 6530 .addImm(0x01))); 6531 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 6532 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3) 6533 .addReg(NewVReg2, RegState::Kill) 6534 .addImm(PCLabelId); 6535 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12)) 6536 .addReg(NewVReg3, RegState::Kill) 6537 .addFrameIndex(FI) 6538 .addImm(36) // &jbuf[1] :: pc 6539 .addMemOperand(FIMMOSt)); 6540 } else if (isThumb) { 6541 // Incoming value: jbuf 6542 // ldr.n r1, LCPI1_4 6543 // add r1, pc 6544 // mov r2, #1 6545 // orrs r1, r2 6546 // add r2, $jbuf, #+4 ; &jbuf[1] 6547 // str r1, [r2] 6548 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 6549 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1) 6550 .addConstantPoolIndex(CPI) 6551 .addMemOperand(CPMMO)); 6552 unsigned NewVReg2 = MRI->createVirtualRegister(TRC); 6553 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2) 6554 .addReg(NewVReg1, RegState::Kill) 6555 .addImm(PCLabelId); 6556 // Set the low bit because of thumb mode. 6557 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 6558 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3) 6559 .addReg(ARM::CPSR, RegState::Define) 6560 .addImm(1)); 6561 unsigned NewVReg4 = MRI->createVirtualRegister(TRC); 6562 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4) 6563 .addReg(ARM::CPSR, RegState::Define) 6564 .addReg(NewVReg2, RegState::Kill) 6565 .addReg(NewVReg3, RegState::Kill)); 6566 unsigned NewVReg5 = MRI->createVirtualRegister(TRC); 6567 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5) 6568 .addFrameIndex(FI) 6569 .addImm(36)); // &jbuf[1] :: pc 6570 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi)) 6571 .addReg(NewVReg4, RegState::Kill) 6572 .addReg(NewVReg5, RegState::Kill) 6573 .addImm(0) 6574 .addMemOperand(FIMMOSt)); 6575 } else { 6576 // Incoming value: jbuf 6577 // ldr r1, LCPI1_1 6578 // add r1, pc, r1 6579 // str r1, [$jbuf, #+4] ; &jbuf[1] 6580 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 6581 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1) 6582 .addConstantPoolIndex(CPI) 6583 .addImm(0) 6584 .addMemOperand(CPMMO)); 6585 unsigned NewVReg2 = MRI->createVirtualRegister(TRC); 6586 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2) 6587 .addReg(NewVReg1, RegState::Kill) 6588 .addImm(PCLabelId)); 6589 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12)) 6590 .addReg(NewVReg2, RegState::Kill) 6591 .addFrameIndex(FI) 6592 .addImm(36) // &jbuf[1] :: pc 6593 .addMemOperand(FIMMOSt)); 6594 } 6595 } 6596 6597 MachineBasicBlock *ARMTargetLowering:: 6598 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const { 6599 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 6600 DebugLoc dl = MI->getDebugLoc(); 6601 MachineFunction *MF = MBB->getParent(); 6602 MachineRegisterInfo *MRI = &MF->getRegInfo(); 6603 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>(); 6604 MachineFrameInfo *MFI = MF->getFrameInfo(); 6605 int FI = MFI->getFunctionContextIndex(); 6606 6607 const TargetRegisterClass *TRC = Subtarget->isThumb() ? 6608 (const TargetRegisterClass*)&ARM::tGPRRegClass : 6609 (const TargetRegisterClass*)&ARM::GPRnopcRegClass; 6610 6611 // Get a mapping of the call site numbers to all of the landing pads they're 6612 // associated with. 6613 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad; 6614 unsigned MaxCSNum = 0; 6615 MachineModuleInfo &MMI = MF->getMMI(); 6616 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; 6617 ++BB) { 6618 if (!BB->isLandingPad()) continue; 6619 6620 // FIXME: We should assert that the EH_LABEL is the first MI in the landing 6621 // pad. 6622 for (MachineBasicBlock::iterator 6623 II = BB->begin(), IE = BB->end(); II != IE; ++II) { 6624 if (!II->isEHLabel()) continue; 6625 6626 MCSymbol *Sym = II->getOperand(0).getMCSymbol(); 6627 if (!MMI.hasCallSiteLandingPad(Sym)) continue; 6628 6629 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym); 6630 for (SmallVectorImpl<unsigned>::iterator 6631 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end(); 6632 CSI != CSE; ++CSI) { 6633 CallSiteNumToLPad[*CSI].push_back(BB); 6634 MaxCSNum = std::max(MaxCSNum, *CSI); 6635 } 6636 break; 6637 } 6638 } 6639 6640 // Get an ordered list of the machine basic blocks for the jump table. 6641 std::vector<MachineBasicBlock*> LPadList; 6642 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs; 6643 LPadList.reserve(CallSiteNumToLPad.size()); 6644 for (unsigned I = 1; I <= MaxCSNum; ++I) { 6645 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I]; 6646 for (SmallVectorImpl<MachineBasicBlock*>::iterator 6647 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) { 6648 LPadList.push_back(*II); 6649 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end()); 6650 } 6651 } 6652 6653 assert(!LPadList.empty() && 6654 "No landing pad destinations for the dispatch jump table!"); 6655 6656 // Create the jump table and associated information. 6657 MachineJumpTableInfo *JTI = 6658 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline); 6659 unsigned MJTI = JTI->createJumpTableIndex(LPadList); 6660 unsigned UId = AFI->createJumpTableUId(); 6661 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 6662 6663 // Create the MBBs for the dispatch code. 6664 6665 // Shove the dispatch's address into the return slot in the function context. 6666 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock(); 6667 DispatchBB->setIsLandingPad(); 6668 6669 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock(); 6670 unsigned trap_opcode; 6671 if (Subtarget->isThumb()) 6672 trap_opcode = ARM::tTRAP; 6673 else 6674 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP; 6675 6676 BuildMI(TrapBB, dl, TII->get(trap_opcode)); 6677 DispatchBB->addSuccessor(TrapBB); 6678 6679 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock(); 6680 DispatchBB->addSuccessor(DispContBB); 6681 6682 // Insert and MBBs. 6683 MF->insert(MF->end(), DispatchBB); 6684 MF->insert(MF->end(), DispContBB); 6685 MF->insert(MF->end(), TrapBB); 6686 6687 // Insert code into the entry block that creates and registers the function 6688 // context. 6689 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI); 6690 6691 MachineMemOperand *FIMMOLd = 6692 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI), 6693 MachineMemOperand::MOLoad | 6694 MachineMemOperand::MOVolatile, 4, 4); 6695 6696 MachineInstrBuilder MIB; 6697 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup)); 6698 6699 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII); 6700 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo(); 6701 6702 // Add a register mask with no preserved registers. This results in all 6703 // registers being marked as clobbered. 6704 MIB.addRegMask(RI.getNoPreservedMask()); 6705 6706 unsigned NumLPads = LPadList.size(); 6707 if (Subtarget->isThumb2()) { 6708 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 6709 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1) 6710 .addFrameIndex(FI) 6711 .addImm(4) 6712 .addMemOperand(FIMMOLd)); 6713 6714 if (NumLPads < 256) { 6715 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri)) 6716 .addReg(NewVReg1) 6717 .addImm(LPadList.size())); 6718 } else { 6719 unsigned VReg1 = MRI->createVirtualRegister(TRC); 6720 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1) 6721 .addImm(NumLPads & 0xFFFF)); 6722 6723 unsigned VReg2 = VReg1; 6724 if ((NumLPads & 0xFFFF0000) != 0) { 6725 VReg2 = MRI->createVirtualRegister(TRC); 6726 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2) 6727 .addReg(VReg1) 6728 .addImm(NumLPads >> 16)); 6729 } 6730 6731 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr)) 6732 .addReg(NewVReg1) 6733 .addReg(VReg2)); 6734 } 6735 6736 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc)) 6737 .addMBB(TrapBB) 6738 .addImm(ARMCC::HI) 6739 .addReg(ARM::CPSR); 6740 6741 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 6742 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3) 6743 .addJumpTableIndex(MJTI) 6744 .addImm(UId)); 6745 6746 unsigned NewVReg4 = MRI->createVirtualRegister(TRC); 6747 AddDefaultCC( 6748 AddDefaultPred( 6749 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4) 6750 .addReg(NewVReg3, RegState::Kill) 6751 .addReg(NewVReg1) 6752 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2)))); 6753 6754 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT)) 6755 .addReg(NewVReg4, RegState::Kill) 6756 .addReg(NewVReg1) 6757 .addJumpTableIndex(MJTI) 6758 .addImm(UId); 6759 } else if (Subtarget->isThumb()) { 6760 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 6761 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1) 6762 .addFrameIndex(FI) 6763 .addImm(1) 6764 .addMemOperand(FIMMOLd)); 6765 6766 if (NumLPads < 256) { 6767 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8)) 6768 .addReg(NewVReg1) 6769 .addImm(NumLPads)); 6770 } else { 6771 MachineConstantPool *ConstantPool = MF->getConstantPool(); 6772 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext()); 6773 const Constant *C = ConstantInt::get(Int32Ty, NumLPads); 6774 6775 // MachineConstantPool wants an explicit alignment. 6776 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty); 6777 if (Align == 0) 6778 Align = getDataLayout()->getTypeAllocSize(C->getType()); 6779 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align); 6780 6781 unsigned VReg1 = MRI->createVirtualRegister(TRC); 6782 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci)) 6783 .addReg(VReg1, RegState::Define) 6784 .addConstantPoolIndex(Idx)); 6785 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr)) 6786 .addReg(NewVReg1) 6787 .addReg(VReg1)); 6788 } 6789 6790 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc)) 6791 .addMBB(TrapBB) 6792 .addImm(ARMCC::HI) 6793 .addReg(ARM::CPSR); 6794 6795 unsigned NewVReg2 = MRI->createVirtualRegister(TRC); 6796 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2) 6797 .addReg(ARM::CPSR, RegState::Define) 6798 .addReg(NewVReg1) 6799 .addImm(2)); 6800 6801 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 6802 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3) 6803 .addJumpTableIndex(MJTI) 6804 .addImm(UId)); 6805 6806 unsigned NewVReg4 = MRI->createVirtualRegister(TRC); 6807 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4) 6808 .addReg(ARM::CPSR, RegState::Define) 6809 .addReg(NewVReg2, RegState::Kill) 6810 .addReg(NewVReg3)); 6811 6812 MachineMemOperand *JTMMOLd = 6813 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(), 6814 MachineMemOperand::MOLoad, 4, 4); 6815 6816 unsigned NewVReg5 = MRI->createVirtualRegister(TRC); 6817 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5) 6818 .addReg(NewVReg4, RegState::Kill) 6819 .addImm(0) 6820 .addMemOperand(JTMMOLd)); 6821 6822 unsigned NewVReg6 = NewVReg5; 6823 if (RelocM == Reloc::PIC_) { 6824 NewVReg6 = MRI->createVirtualRegister(TRC); 6825 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6) 6826 .addReg(ARM::CPSR, RegState::Define) 6827 .addReg(NewVReg5, RegState::Kill) 6828 .addReg(NewVReg3)); 6829 } 6830 6831 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr)) 6832 .addReg(NewVReg6, RegState::Kill) 6833 .addJumpTableIndex(MJTI) 6834 .addImm(UId); 6835 } else { 6836 unsigned NewVReg1 = MRI->createVirtualRegister(TRC); 6837 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1) 6838 .addFrameIndex(FI) 6839 .addImm(4) 6840 .addMemOperand(FIMMOLd)); 6841 6842 if (NumLPads < 256) { 6843 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri)) 6844 .addReg(NewVReg1) 6845 .addImm(NumLPads)); 6846 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) { 6847 unsigned VReg1 = MRI->createVirtualRegister(TRC); 6848 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1) 6849 .addImm(NumLPads & 0xFFFF)); 6850 6851 unsigned VReg2 = VReg1; 6852 if ((NumLPads & 0xFFFF0000) != 0) { 6853 VReg2 = MRI->createVirtualRegister(TRC); 6854 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2) 6855 .addReg(VReg1) 6856 .addImm(NumLPads >> 16)); 6857 } 6858 6859 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr)) 6860 .addReg(NewVReg1) 6861 .addReg(VReg2)); 6862 } else { 6863 MachineConstantPool *ConstantPool = MF->getConstantPool(); 6864 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext()); 6865 const Constant *C = ConstantInt::get(Int32Ty, NumLPads); 6866 6867 // MachineConstantPool wants an explicit alignment. 6868 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty); 6869 if (Align == 0) 6870 Align = getDataLayout()->getTypeAllocSize(C->getType()); 6871 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align); 6872 6873 unsigned VReg1 = MRI->createVirtualRegister(TRC); 6874 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp)) 6875 .addReg(VReg1, RegState::Define) 6876 .addConstantPoolIndex(Idx) 6877 .addImm(0)); 6878 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr)) 6879 .addReg(NewVReg1) 6880 .addReg(VReg1, RegState::Kill)); 6881 } 6882 6883 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc)) 6884 .addMBB(TrapBB) 6885 .addImm(ARMCC::HI) 6886 .addReg(ARM::CPSR); 6887 6888 unsigned NewVReg3 = MRI->createVirtualRegister(TRC); 6889 AddDefaultCC( 6890 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3) 6891 .addReg(NewVReg1) 6892 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2)))); 6893 unsigned NewVReg4 = MRI->createVirtualRegister(TRC); 6894 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4) 6895 .addJumpTableIndex(MJTI) 6896 .addImm(UId)); 6897 6898 MachineMemOperand *JTMMOLd = 6899 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(), 6900 MachineMemOperand::MOLoad, 4, 4); 6901 unsigned NewVReg5 = MRI->createVirtualRegister(TRC); 6902 AddDefaultPred( 6903 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5) 6904 .addReg(NewVReg3, RegState::Kill) 6905 .addReg(NewVReg4) 6906 .addImm(0) 6907 .addMemOperand(JTMMOLd)); 6908 6909 if (RelocM == Reloc::PIC_) { 6910 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd)) 6911 .addReg(NewVReg5, RegState::Kill) 6912 .addReg(NewVReg4) 6913 .addJumpTableIndex(MJTI) 6914 .addImm(UId); 6915 } else { 6916 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr)) 6917 .addReg(NewVReg5, RegState::Kill) 6918 .addJumpTableIndex(MJTI) 6919 .addImm(UId); 6920 } 6921 } 6922 6923 // Add the jump table entries as successors to the MBB. 6924 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs; 6925 for (std::vector<MachineBasicBlock*>::iterator 6926 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) { 6927 MachineBasicBlock *CurMBB = *I; 6928 if (SeenMBBs.insert(CurMBB)) 6929 DispContBB->addSuccessor(CurMBB); 6930 } 6931 6932 // N.B. the order the invoke BBs are processed in doesn't matter here. 6933 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF); 6934 SmallVector<MachineBasicBlock*, 64> MBBLPads; 6935 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator 6936 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) { 6937 MachineBasicBlock *BB = *I; 6938 6939 // Remove the landing pad successor from the invoke block and replace it 6940 // with the new dispatch block. 6941 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(), 6942 BB->succ_end()); 6943 while (!Successors.empty()) { 6944 MachineBasicBlock *SMBB = Successors.pop_back_val(); 6945 if (SMBB->isLandingPad()) { 6946 BB->removeSuccessor(SMBB); 6947 MBBLPads.push_back(SMBB); 6948 } 6949 } 6950 6951 BB->addSuccessor(DispatchBB); 6952 6953 // Find the invoke call and mark all of the callee-saved registers as 6954 // 'implicit defined' so that they're spilled. This prevents code from 6955 // moving instructions to before the EH block, where they will never be 6956 // executed. 6957 for (MachineBasicBlock::reverse_iterator 6958 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) { 6959 if (!II->isCall()) continue; 6960 6961 DenseMap<unsigned, bool> DefRegs; 6962 for (MachineInstr::mop_iterator 6963 OI = II->operands_begin(), OE = II->operands_end(); 6964 OI != OE; ++OI) { 6965 if (!OI->isReg()) continue; 6966 DefRegs[OI->getReg()] = true; 6967 } 6968 6969 MachineInstrBuilder MIB(*MF, &*II); 6970 6971 for (unsigned i = 0; SavedRegs[i] != 0; ++i) { 6972 unsigned Reg = SavedRegs[i]; 6973 if (Subtarget->isThumb2() && 6974 !ARM::tGPRRegClass.contains(Reg) && 6975 !ARM::hGPRRegClass.contains(Reg)) 6976 continue; 6977 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg)) 6978 continue; 6979 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg)) 6980 continue; 6981 if (!DefRegs[Reg]) 6982 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead); 6983 } 6984 6985 break; 6986 } 6987 } 6988 6989 // Mark all former landing pads as non-landing pads. The dispatch is the only 6990 // landing pad now. 6991 for (SmallVectorImpl<MachineBasicBlock*>::iterator 6992 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I) 6993 (*I)->setIsLandingPad(false); 6994 6995 // The instruction is gone now. 6996 MI->eraseFromParent(); 6997 6998 return MBB; 6999 } 7000 7001 static 7002 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) { 7003 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(), 7004 E = MBB->succ_end(); I != E; ++I) 7005 if (*I != Succ) 7006 return *I; 7007 llvm_unreachable("Expecting a BB with two successors!"); 7008 } 7009 7010 MachineBasicBlock *ARMTargetLowering:: 7011 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const { 7012 // This pseudo instruction has 3 operands: dst, src, size 7013 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold(). 7014 // Otherwise, we will generate unrolled scalar copies. 7015 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 7016 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 7017 MachineFunction::iterator It = BB; 7018 ++It; 7019 7020 unsigned dest = MI->getOperand(0).getReg(); 7021 unsigned src = MI->getOperand(1).getReg(); 7022 unsigned SizeVal = MI->getOperand(2).getImm(); 7023 unsigned Align = MI->getOperand(3).getImm(); 7024 DebugLoc dl = MI->getDebugLoc(); 7025 7026 bool isThumb2 = Subtarget->isThumb2(); 7027 MachineFunction *MF = BB->getParent(); 7028 MachineRegisterInfo &MRI = MF->getRegInfo(); 7029 unsigned ldrOpc, strOpc, UnitSize = 0; 7030 7031 const TargetRegisterClass *TRC = isThumb2 ? 7032 (const TargetRegisterClass*)&ARM::tGPRRegClass : 7033 (const TargetRegisterClass*)&ARM::GPRRegClass; 7034 const TargetRegisterClass *TRC_Vec = 0; 7035 7036 if (Align & 1) { 7037 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM; 7038 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM; 7039 UnitSize = 1; 7040 } else if (Align & 2) { 7041 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST; 7042 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST; 7043 UnitSize = 2; 7044 } else { 7045 // Check whether we can use NEON instructions. 7046 if (!MF->getFunction()->getAttributes(). 7047 hasAttribute(AttributeSet::FunctionIndex, 7048 Attribute::NoImplicitFloat) && 7049 Subtarget->hasNEON()) { 7050 if ((Align % 16 == 0) && SizeVal >= 16) { 7051 ldrOpc = ARM::VLD1q32wb_fixed; 7052 strOpc = ARM::VST1q32wb_fixed; 7053 UnitSize = 16; 7054 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass; 7055 } 7056 else if ((Align % 8 == 0) && SizeVal >= 8) { 7057 ldrOpc = ARM::VLD1d32wb_fixed; 7058 strOpc = ARM::VST1d32wb_fixed; 7059 UnitSize = 8; 7060 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass; 7061 } 7062 } 7063 // Can't use NEON instructions. 7064 if (UnitSize == 0) { 7065 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM; 7066 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM; 7067 UnitSize = 4; 7068 } 7069 } 7070 7071 unsigned BytesLeft = SizeVal % UnitSize; 7072 unsigned LoopSize = SizeVal - BytesLeft; 7073 7074 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) { 7075 // Use LDR and STR to copy. 7076 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize) 7077 // [destOut] = STR_POST(scratch, destIn, UnitSize) 7078 unsigned srcIn = src; 7079 unsigned destIn = dest; 7080 for (unsigned i = 0; i < LoopSize; i+=UnitSize) { 7081 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC); 7082 unsigned srcOut = MRI.createVirtualRegister(TRC); 7083 unsigned destOut = MRI.createVirtualRegister(TRC); 7084 if (UnitSize >= 8) { 7085 AddDefaultPred(BuildMI(*BB, MI, dl, 7086 TII->get(ldrOpc), scratch) 7087 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0)); 7088 7089 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut) 7090 .addReg(destIn).addImm(0).addReg(scratch)); 7091 } else if (isThumb2) { 7092 AddDefaultPred(BuildMI(*BB, MI, dl, 7093 TII->get(ldrOpc), scratch) 7094 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize)); 7095 7096 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut) 7097 .addReg(scratch).addReg(destIn) 7098 .addImm(UnitSize)); 7099 } else { 7100 AddDefaultPred(BuildMI(*BB, MI, dl, 7101 TII->get(ldrOpc), scratch) 7102 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0) 7103 .addImm(UnitSize)); 7104 7105 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut) 7106 .addReg(scratch).addReg(destIn) 7107 .addReg(0).addImm(UnitSize)); 7108 } 7109 srcIn = srcOut; 7110 destIn = destOut; 7111 } 7112 7113 // Handle the leftover bytes with LDRB and STRB. 7114 // [scratch, srcOut] = LDRB_POST(srcIn, 1) 7115 // [destOut] = STRB_POST(scratch, destIn, 1) 7116 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM; 7117 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM; 7118 for (unsigned i = 0; i < BytesLeft; i++) { 7119 unsigned scratch = MRI.createVirtualRegister(TRC); 7120 unsigned srcOut = MRI.createVirtualRegister(TRC); 7121 unsigned destOut = MRI.createVirtualRegister(TRC); 7122 if (isThumb2) { 7123 AddDefaultPred(BuildMI(*BB, MI, dl, 7124 TII->get(ldrOpc),scratch) 7125 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1)); 7126 7127 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut) 7128 .addReg(scratch).addReg(destIn) 7129 .addReg(0).addImm(1)); 7130 } else { 7131 AddDefaultPred(BuildMI(*BB, MI, dl, 7132 TII->get(ldrOpc),scratch) 7133 .addReg(srcOut, RegState::Define).addReg(srcIn) 7134 .addReg(0).addImm(1)); 7135 7136 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut) 7137 .addReg(scratch).addReg(destIn) 7138 .addReg(0).addImm(1)); 7139 } 7140 srcIn = srcOut; 7141 destIn = destOut; 7142 } 7143 MI->eraseFromParent(); // The instruction is gone now. 7144 return BB; 7145 } 7146 7147 // Expand the pseudo op to a loop. 7148 // thisMBB: 7149 // ... 7150 // movw varEnd, # --> with thumb2 7151 // movt varEnd, # 7152 // ldrcp varEnd, idx --> without thumb2 7153 // fallthrough --> loopMBB 7154 // loopMBB: 7155 // PHI varPhi, varEnd, varLoop 7156 // PHI srcPhi, src, srcLoop 7157 // PHI destPhi, dst, destLoop 7158 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize) 7159 // [destLoop] = STR_POST(scratch, destPhi, UnitSize) 7160 // subs varLoop, varPhi, #UnitSize 7161 // bne loopMBB 7162 // fallthrough --> exitMBB 7163 // exitMBB: 7164 // epilogue to handle left-over bytes 7165 // [scratch, srcOut] = LDRB_POST(srcLoop, 1) 7166 // [destOut] = STRB_POST(scratch, destLoop, 1) 7167 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 7168 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 7169 MF->insert(It, loopMBB); 7170 MF->insert(It, exitMBB); 7171 7172 // Transfer the remainder of BB and its successor edges to exitMBB. 7173 exitMBB->splice(exitMBB->begin(), BB, 7174 llvm::next(MachineBasicBlock::iterator(MI)), 7175 BB->end()); 7176 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 7177 7178 // Load an immediate to varEnd. 7179 unsigned varEnd = MRI.createVirtualRegister(TRC); 7180 if (isThumb2) { 7181 unsigned VReg1 = varEnd; 7182 if ((LoopSize & 0xFFFF0000) != 0) 7183 VReg1 = MRI.createVirtualRegister(TRC); 7184 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1) 7185 .addImm(LoopSize & 0xFFFF)); 7186 7187 if ((LoopSize & 0xFFFF0000) != 0) 7188 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd) 7189 .addReg(VReg1) 7190 .addImm(LoopSize >> 16)); 7191 } else { 7192 MachineConstantPool *ConstantPool = MF->getConstantPool(); 7193 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext()); 7194 const Constant *C = ConstantInt::get(Int32Ty, LoopSize); 7195 7196 // MachineConstantPool wants an explicit alignment. 7197 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty); 7198 if (Align == 0) 7199 Align = getDataLayout()->getTypeAllocSize(C->getType()); 7200 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align); 7201 7202 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp)) 7203 .addReg(varEnd, RegState::Define) 7204 .addConstantPoolIndex(Idx) 7205 .addImm(0)); 7206 } 7207 BB->addSuccessor(loopMBB); 7208 7209 // Generate the loop body: 7210 // varPhi = PHI(varLoop, varEnd) 7211 // srcPhi = PHI(srcLoop, src) 7212 // destPhi = PHI(destLoop, dst) 7213 MachineBasicBlock *entryBB = BB; 7214 BB = loopMBB; 7215 unsigned varLoop = MRI.createVirtualRegister(TRC); 7216 unsigned varPhi = MRI.createVirtualRegister(TRC); 7217 unsigned srcLoop = MRI.createVirtualRegister(TRC); 7218 unsigned srcPhi = MRI.createVirtualRegister(TRC); 7219 unsigned destLoop = MRI.createVirtualRegister(TRC); 7220 unsigned destPhi = MRI.createVirtualRegister(TRC); 7221 7222 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi) 7223 .addReg(varLoop).addMBB(loopMBB) 7224 .addReg(varEnd).addMBB(entryBB); 7225 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi) 7226 .addReg(srcLoop).addMBB(loopMBB) 7227 .addReg(src).addMBB(entryBB); 7228 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi) 7229 .addReg(destLoop).addMBB(loopMBB) 7230 .addReg(dest).addMBB(entryBB); 7231 7232 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize) 7233 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz) 7234 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC); 7235 if (UnitSize >= 8) { 7236 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch) 7237 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0)); 7238 7239 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop) 7240 .addReg(destPhi).addImm(0).addReg(scratch)); 7241 } else if (isThumb2) { 7242 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch) 7243 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize)); 7244 7245 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop) 7246 .addReg(scratch).addReg(destPhi) 7247 .addImm(UnitSize)); 7248 } else { 7249 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch) 7250 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0) 7251 .addImm(UnitSize)); 7252 7253 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop) 7254 .addReg(scratch).addReg(destPhi) 7255 .addReg(0).addImm(UnitSize)); 7256 } 7257 7258 // Decrement loop variable by UnitSize. 7259 MachineInstrBuilder MIB = BuildMI(BB, dl, 7260 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop); 7261 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize))); 7262 MIB->getOperand(5).setReg(ARM::CPSR); 7263 MIB->getOperand(5).setIsDef(true); 7264 7265 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 7266 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR); 7267 7268 // loopMBB can loop back to loopMBB or fall through to exitMBB. 7269 BB->addSuccessor(loopMBB); 7270 BB->addSuccessor(exitMBB); 7271 7272 // Add epilogue to handle BytesLeft. 7273 BB = exitMBB; 7274 MachineInstr *StartOfExit = exitMBB->begin(); 7275 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM; 7276 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM; 7277 7278 // [scratch, srcOut] = LDRB_POST(srcLoop, 1) 7279 // [destOut] = STRB_POST(scratch, destLoop, 1) 7280 unsigned srcIn = srcLoop; 7281 unsigned destIn = destLoop; 7282 for (unsigned i = 0; i < BytesLeft; i++) { 7283 unsigned scratch = MRI.createVirtualRegister(TRC); 7284 unsigned srcOut = MRI.createVirtualRegister(TRC); 7285 unsigned destOut = MRI.createVirtualRegister(TRC); 7286 if (isThumb2) { 7287 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, 7288 TII->get(ldrOpc),scratch) 7289 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1)); 7290 7291 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut) 7292 .addReg(scratch).addReg(destIn) 7293 .addImm(1)); 7294 } else { 7295 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, 7296 TII->get(ldrOpc),scratch) 7297 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1)); 7298 7299 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut) 7300 .addReg(scratch).addReg(destIn) 7301 .addReg(0).addImm(1)); 7302 } 7303 srcIn = srcOut; 7304 destIn = destOut; 7305 } 7306 7307 MI->eraseFromParent(); // The instruction is gone now. 7308 return BB; 7309 } 7310 7311 MachineBasicBlock * 7312 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 7313 MachineBasicBlock *BB) const { 7314 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 7315 DebugLoc dl = MI->getDebugLoc(); 7316 bool isThumb2 = Subtarget->isThumb2(); 7317 switch (MI->getOpcode()) { 7318 default: { 7319 MI->dump(); 7320 llvm_unreachable("Unexpected instr type to insert"); 7321 } 7322 // The Thumb2 pre-indexed stores have the same MI operands, they just 7323 // define them differently in the .td files from the isel patterns, so 7324 // they need pseudos. 7325 case ARM::t2STR_preidx: 7326 MI->setDesc(TII->get(ARM::t2STR_PRE)); 7327 return BB; 7328 case ARM::t2STRB_preidx: 7329 MI->setDesc(TII->get(ARM::t2STRB_PRE)); 7330 return BB; 7331 case ARM::t2STRH_preidx: 7332 MI->setDesc(TII->get(ARM::t2STRH_PRE)); 7333 return BB; 7334 7335 case ARM::STRi_preidx: 7336 case ARM::STRBi_preidx: { 7337 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ? 7338 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM; 7339 // Decode the offset. 7340 unsigned Offset = MI->getOperand(4).getImm(); 7341 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub; 7342 Offset = ARM_AM::getAM2Offset(Offset); 7343 if (isSub) 7344 Offset = -Offset; 7345 7346 MachineMemOperand *MMO = *MI->memoperands_begin(); 7347 BuildMI(*BB, MI, dl, TII->get(NewOpc)) 7348 .addOperand(MI->getOperand(0)) // Rn_wb 7349 .addOperand(MI->getOperand(1)) // Rt 7350 .addOperand(MI->getOperand(2)) // Rn 7351 .addImm(Offset) // offset (skip GPR==zero_reg) 7352 .addOperand(MI->getOperand(5)) // pred 7353 .addOperand(MI->getOperand(6)) 7354 .addMemOperand(MMO); 7355 MI->eraseFromParent(); 7356 return BB; 7357 } 7358 case ARM::STRr_preidx: 7359 case ARM::STRBr_preidx: 7360 case ARM::STRH_preidx: { 7361 unsigned NewOpc; 7362 switch (MI->getOpcode()) { 7363 default: llvm_unreachable("unexpected opcode!"); 7364 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break; 7365 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break; 7366 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break; 7367 } 7368 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc)); 7369 for (unsigned i = 0; i < MI->getNumOperands(); ++i) 7370 MIB.addOperand(MI->getOperand(i)); 7371 MI->eraseFromParent(); 7372 return BB; 7373 } 7374 case ARM::ATOMIC_LOAD_ADD_I8: 7375 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr); 7376 case ARM::ATOMIC_LOAD_ADD_I16: 7377 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr); 7378 case ARM::ATOMIC_LOAD_ADD_I32: 7379 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr); 7380 7381 case ARM::ATOMIC_LOAD_AND_I8: 7382 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr); 7383 case ARM::ATOMIC_LOAD_AND_I16: 7384 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr); 7385 case ARM::ATOMIC_LOAD_AND_I32: 7386 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr); 7387 7388 case ARM::ATOMIC_LOAD_OR_I8: 7389 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr); 7390 case ARM::ATOMIC_LOAD_OR_I16: 7391 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr); 7392 case ARM::ATOMIC_LOAD_OR_I32: 7393 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr); 7394 7395 case ARM::ATOMIC_LOAD_XOR_I8: 7396 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr); 7397 case ARM::ATOMIC_LOAD_XOR_I16: 7398 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr); 7399 case ARM::ATOMIC_LOAD_XOR_I32: 7400 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr); 7401 7402 case ARM::ATOMIC_LOAD_NAND_I8: 7403 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr); 7404 case ARM::ATOMIC_LOAD_NAND_I16: 7405 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr); 7406 case ARM::ATOMIC_LOAD_NAND_I32: 7407 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr); 7408 7409 case ARM::ATOMIC_LOAD_SUB_I8: 7410 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr); 7411 case ARM::ATOMIC_LOAD_SUB_I16: 7412 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr); 7413 case ARM::ATOMIC_LOAD_SUB_I32: 7414 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr); 7415 7416 case ARM::ATOMIC_LOAD_MIN_I8: 7417 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT); 7418 case ARM::ATOMIC_LOAD_MIN_I16: 7419 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT); 7420 case ARM::ATOMIC_LOAD_MIN_I32: 7421 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT); 7422 7423 case ARM::ATOMIC_LOAD_MAX_I8: 7424 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT); 7425 case ARM::ATOMIC_LOAD_MAX_I16: 7426 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT); 7427 case ARM::ATOMIC_LOAD_MAX_I32: 7428 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT); 7429 7430 case ARM::ATOMIC_LOAD_UMIN_I8: 7431 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO); 7432 case ARM::ATOMIC_LOAD_UMIN_I16: 7433 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO); 7434 case ARM::ATOMIC_LOAD_UMIN_I32: 7435 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO); 7436 7437 case ARM::ATOMIC_LOAD_UMAX_I8: 7438 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI); 7439 case ARM::ATOMIC_LOAD_UMAX_I16: 7440 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI); 7441 case ARM::ATOMIC_LOAD_UMAX_I32: 7442 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI); 7443 7444 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0); 7445 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0); 7446 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0); 7447 7448 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1); 7449 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2); 7450 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4); 7451 7452 7453 case ARM::ATOMADD6432: 7454 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr, 7455 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr, 7456 /*NeedsCarry*/ true); 7457 case ARM::ATOMSUB6432: 7458 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr, 7459 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr, 7460 /*NeedsCarry*/ true); 7461 case ARM::ATOMOR6432: 7462 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr, 7463 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr); 7464 case ARM::ATOMXOR6432: 7465 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr, 7466 isThumb2 ? ARM::t2EORrr : ARM::EORrr); 7467 case ARM::ATOMAND6432: 7468 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr, 7469 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr); 7470 case ARM::ATOMSWAP6432: 7471 return EmitAtomicBinary64(MI, BB, 0, 0, false); 7472 case ARM::ATOMCMPXCHG6432: 7473 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr, 7474 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr, 7475 /*NeedsCarry*/ false, /*IsCmpxchg*/true); 7476 case ARM::ATOMMIN6432: 7477 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr, 7478 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr, 7479 /*NeedsCarry*/ true, /*IsCmpxchg*/false, 7480 /*IsMinMax*/ true, ARMCC::LT); 7481 case ARM::ATOMMAX6432: 7482 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr, 7483 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr, 7484 /*NeedsCarry*/ true, /*IsCmpxchg*/false, 7485 /*IsMinMax*/ true, ARMCC::GE); 7486 case ARM::ATOMUMIN6432: 7487 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr, 7488 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr, 7489 /*NeedsCarry*/ true, /*IsCmpxchg*/false, 7490 /*IsMinMax*/ true, ARMCC::LO); 7491 case ARM::ATOMUMAX6432: 7492 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr, 7493 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr, 7494 /*NeedsCarry*/ true, /*IsCmpxchg*/false, 7495 /*IsMinMax*/ true, ARMCC::HS); 7496 7497 case ARM::tMOVCCr_pseudo: { 7498 // To "insert" a SELECT_CC instruction, we actually have to insert the 7499 // diamond control-flow pattern. The incoming instruction knows the 7500 // destination vreg to set, the condition code register to branch on, the 7501 // true/false values to select between, and a branch opcode to use. 7502 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 7503 MachineFunction::iterator It = BB; 7504 ++It; 7505 7506 // thisMBB: 7507 // ... 7508 // TrueVal = ... 7509 // cmpTY ccX, r1, r2 7510 // bCC copy1MBB 7511 // fallthrough --> copy0MBB 7512 MachineBasicBlock *thisMBB = BB; 7513 MachineFunction *F = BB->getParent(); 7514 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 7515 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); 7516 F->insert(It, copy0MBB); 7517 F->insert(It, sinkMBB); 7518 7519 // Transfer the remainder of BB and its successor edges to sinkMBB. 7520 sinkMBB->splice(sinkMBB->begin(), BB, 7521 llvm::next(MachineBasicBlock::iterator(MI)), 7522 BB->end()); 7523 sinkMBB->transferSuccessorsAndUpdatePHIs(BB); 7524 7525 BB->addSuccessor(copy0MBB); 7526 BB->addSuccessor(sinkMBB); 7527 7528 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB) 7529 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg()); 7530 7531 // copy0MBB: 7532 // %FalseValue = ... 7533 // # fallthrough to sinkMBB 7534 BB = copy0MBB; 7535 7536 // Update machine-CFG edges 7537 BB->addSuccessor(sinkMBB); 7538 7539 // sinkMBB: 7540 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] 7541 // ... 7542 BB = sinkMBB; 7543 BuildMI(*BB, BB->begin(), dl, 7544 TII->get(ARM::PHI), MI->getOperand(0).getReg()) 7545 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB) 7546 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB); 7547 7548 MI->eraseFromParent(); // The pseudo instruction is gone now. 7549 return BB; 7550 } 7551 7552 case ARM::BCCi64: 7553 case ARM::BCCZi64: { 7554 // If there is an unconditional branch to the other successor, remove it. 7555 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end()); 7556 7557 // Compare both parts that make up the double comparison separately for 7558 // equality. 7559 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64; 7560 7561 unsigned LHS1 = MI->getOperand(1).getReg(); 7562 unsigned LHS2 = MI->getOperand(2).getReg(); 7563 if (RHSisZero) { 7564 AddDefaultPred(BuildMI(BB, dl, 7565 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 7566 .addReg(LHS1).addImm(0)); 7567 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 7568 .addReg(LHS2).addImm(0) 7569 .addImm(ARMCC::EQ).addReg(ARM::CPSR); 7570 } else { 7571 unsigned RHS1 = MI->getOperand(3).getReg(); 7572 unsigned RHS2 = MI->getOperand(4).getReg(); 7573 AddDefaultPred(BuildMI(BB, dl, 7574 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) 7575 .addReg(LHS1).addReg(RHS1)); 7576 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr)) 7577 .addReg(LHS2).addReg(RHS2) 7578 .addImm(ARMCC::EQ).addReg(ARM::CPSR); 7579 } 7580 7581 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB(); 7582 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB); 7583 if (MI->getOperand(0).getImm() == ARMCC::NE) 7584 std::swap(destMBB, exitMBB); 7585 7586 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)) 7587 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR); 7588 if (isThumb2) 7589 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB)); 7590 else 7591 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB); 7592 7593 MI->eraseFromParent(); // The pseudo instruction is gone now. 7594 return BB; 7595 } 7596 7597 case ARM::Int_eh_sjlj_setjmp: 7598 case ARM::Int_eh_sjlj_setjmp_nofp: 7599 case ARM::tInt_eh_sjlj_setjmp: 7600 case ARM::t2Int_eh_sjlj_setjmp: 7601 case ARM::t2Int_eh_sjlj_setjmp_nofp: 7602 EmitSjLjDispatchBlock(MI, BB); 7603 return BB; 7604 7605 case ARM::ABS: 7606 case ARM::t2ABS: { 7607 // To insert an ABS instruction, we have to insert the 7608 // diamond control-flow pattern. The incoming instruction knows the 7609 // source vreg to test against 0, the destination vreg to set, 7610 // the condition code register to branch on, the 7611 // true/false values to select between, and a branch opcode to use. 7612 // It transforms 7613 // V1 = ABS V0 7614 // into 7615 // V2 = MOVS V0 7616 // BCC (branch to SinkBB if V0 >= 0) 7617 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0) 7618 // SinkBB: V1 = PHI(V2, V3) 7619 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 7620 MachineFunction::iterator BBI = BB; 7621 ++BBI; 7622 MachineFunction *Fn = BB->getParent(); 7623 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB); 7624 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB); 7625 Fn->insert(BBI, RSBBB); 7626 Fn->insert(BBI, SinkBB); 7627 7628 unsigned int ABSSrcReg = MI->getOperand(1).getReg(); 7629 unsigned int ABSDstReg = MI->getOperand(0).getReg(); 7630 bool isThumb2 = Subtarget->isThumb2(); 7631 MachineRegisterInfo &MRI = Fn->getRegInfo(); 7632 // In Thumb mode S must not be specified if source register is the SP or 7633 // PC and if destination register is the SP, so restrict register class 7634 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ? 7635 (const TargetRegisterClass*)&ARM::rGPRRegClass : 7636 (const TargetRegisterClass*)&ARM::GPRRegClass); 7637 7638 // Transfer the remainder of BB and its successor edges to sinkMBB. 7639 SinkBB->splice(SinkBB->begin(), BB, 7640 llvm::next(MachineBasicBlock::iterator(MI)), 7641 BB->end()); 7642 SinkBB->transferSuccessorsAndUpdatePHIs(BB); 7643 7644 BB->addSuccessor(RSBBB); 7645 BB->addSuccessor(SinkBB); 7646 7647 // fall through to SinkMBB 7648 RSBBB->addSuccessor(SinkBB); 7649 7650 // insert a cmp at the end of BB 7651 AddDefaultPred(BuildMI(BB, dl, 7652 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri)) 7653 .addReg(ABSSrcReg).addImm(0)); 7654 7655 // insert a bcc with opposite CC to ARMCC::MI at the end of BB 7656 BuildMI(BB, dl, 7657 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB) 7658 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR); 7659 7660 // insert rsbri in RSBBB 7661 // Note: BCC and rsbri will be converted into predicated rsbmi 7662 // by if-conversion pass 7663 BuildMI(*RSBBB, RSBBB->begin(), dl, 7664 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg) 7665 .addReg(ABSSrcReg, RegState::Kill) 7666 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0); 7667 7668 // insert PHI in SinkBB, 7669 // reuse ABSDstReg to not change uses of ABS instruction 7670 BuildMI(*SinkBB, SinkBB->begin(), dl, 7671 TII->get(ARM::PHI), ABSDstReg) 7672 .addReg(NewRsbDstReg).addMBB(RSBBB) 7673 .addReg(ABSSrcReg).addMBB(BB); 7674 7675 // remove ABS instruction 7676 MI->eraseFromParent(); 7677 7678 // return last added BB 7679 return SinkBB; 7680 } 7681 case ARM::COPY_STRUCT_BYVAL_I32: 7682 ++NumLoopByVals; 7683 return EmitStructByval(MI, BB); 7684 } 7685 } 7686 7687 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI, 7688 SDNode *Node) const { 7689 if (!MI->hasPostISelHook()) { 7690 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) && 7691 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'"); 7692 return; 7693 } 7694 7695 const MCInstrDesc *MCID = &MI->getDesc(); 7696 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB, 7697 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional 7698 // operand is still set to noreg. If needed, set the optional operand's 7699 // register to CPSR, and remove the redundant implicit def. 7700 // 7701 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>). 7702 7703 // Rename pseudo opcodes. 7704 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode()); 7705 if (NewOpc) { 7706 const ARMBaseInstrInfo *TII = 7707 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo()); 7708 MCID = &TII->get(NewOpc); 7709 7710 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 && 7711 "converted opcode should be the same except for cc_out"); 7712 7713 MI->setDesc(*MCID); 7714 7715 // Add the optional cc_out operand 7716 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true)); 7717 } 7718 unsigned ccOutIdx = MCID->getNumOperands() - 1; 7719 7720 // Any ARM instruction that sets the 's' bit should specify an optional 7721 // "cc_out" operand in the last operand position. 7722 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) { 7723 assert(!NewOpc && "Optional cc_out operand required"); 7724 return; 7725 } 7726 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it 7727 // since we already have an optional CPSR def. 7728 bool definesCPSR = false; 7729 bool deadCPSR = false; 7730 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands(); 7731 i != e; ++i) { 7732 const MachineOperand &MO = MI->getOperand(i); 7733 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) { 7734 definesCPSR = true; 7735 if (MO.isDead()) 7736 deadCPSR = true; 7737 MI->RemoveOperand(i); 7738 break; 7739 } 7740 } 7741 if (!definesCPSR) { 7742 assert(!NewOpc && "Optional cc_out operand required"); 7743 return; 7744 } 7745 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag"); 7746 if (deadCPSR) { 7747 assert(!MI->getOperand(ccOutIdx).getReg() && 7748 "expect uninitialized optional cc_out operand"); 7749 return; 7750 } 7751 7752 // If this instruction was defined with an optional CPSR def and its dag node 7753 // had a live implicit CPSR def, then activate the optional CPSR def. 7754 MachineOperand &MO = MI->getOperand(ccOutIdx); 7755 MO.setReg(ARM::CPSR); 7756 MO.setIsDef(true); 7757 } 7758 7759 //===----------------------------------------------------------------------===// 7760 // ARM Optimization Hooks 7761 //===----------------------------------------------------------------------===// 7762 7763 // Helper function that checks if N is a null or all ones constant. 7764 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) { 7765 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N); 7766 if (!C) 7767 return false; 7768 return AllOnes ? C->isAllOnesValue() : C->isNullValue(); 7769 } 7770 7771 // Return true if N is conditionally 0 or all ones. 7772 // Detects these expressions where cc is an i1 value: 7773 // 7774 // (select cc 0, y) [AllOnes=0] 7775 // (select cc y, 0) [AllOnes=0] 7776 // (zext cc) [AllOnes=0] 7777 // (sext cc) [AllOnes=0/1] 7778 // (select cc -1, y) [AllOnes=1] 7779 // (select cc y, -1) [AllOnes=1] 7780 // 7781 // Invert is set when N is the null/all ones constant when CC is false. 7782 // OtherOp is set to the alternative value of N. 7783 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes, 7784 SDValue &CC, bool &Invert, 7785 SDValue &OtherOp, 7786 SelectionDAG &DAG) { 7787 switch (N->getOpcode()) { 7788 default: return false; 7789 case ISD::SELECT: { 7790 CC = N->getOperand(0); 7791 SDValue N1 = N->getOperand(1); 7792 SDValue N2 = N->getOperand(2); 7793 if (isZeroOrAllOnes(N1, AllOnes)) { 7794 Invert = false; 7795 OtherOp = N2; 7796 return true; 7797 } 7798 if (isZeroOrAllOnes(N2, AllOnes)) { 7799 Invert = true; 7800 OtherOp = N1; 7801 return true; 7802 } 7803 return false; 7804 } 7805 case ISD::ZERO_EXTEND: 7806 // (zext cc) can never be the all ones value. 7807 if (AllOnes) 7808 return false; 7809 // Fall through. 7810 case ISD::SIGN_EXTEND: { 7811 EVT VT = N->getValueType(0); 7812 CC = N->getOperand(0); 7813 if (CC.getValueType() != MVT::i1) 7814 return false; 7815 Invert = !AllOnes; 7816 if (AllOnes) 7817 // When looking for an AllOnes constant, N is an sext, and the 'other' 7818 // value is 0. 7819 OtherOp = DAG.getConstant(0, VT); 7820 else if (N->getOpcode() == ISD::ZERO_EXTEND) 7821 // When looking for a 0 constant, N can be zext or sext. 7822 OtherOp = DAG.getConstant(1, VT); 7823 else 7824 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT); 7825 return true; 7826 } 7827 } 7828 } 7829 7830 // Combine a constant select operand into its use: 7831 // 7832 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c)) 7833 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c)) 7834 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1] 7835 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c)) 7836 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c)) 7837 // 7838 // The transform is rejected if the select doesn't have a constant operand that 7839 // is null, or all ones when AllOnes is set. 7840 // 7841 // Also recognize sext/zext from i1: 7842 // 7843 // (add (zext cc), x) -> (select cc (add x, 1), x) 7844 // (add (sext cc), x) -> (select cc (add x, -1), x) 7845 // 7846 // These transformations eventually create predicated instructions. 7847 // 7848 // @param N The node to transform. 7849 // @param Slct The N operand that is a select. 7850 // @param OtherOp The other N operand (x above). 7851 // @param DCI Context. 7852 // @param AllOnes Require the select constant to be all ones instead of null. 7853 // @returns The new node, or SDValue() on failure. 7854 static 7855 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp, 7856 TargetLowering::DAGCombinerInfo &DCI, 7857 bool AllOnes = false) { 7858 SelectionDAG &DAG = DCI.DAG; 7859 EVT VT = N->getValueType(0); 7860 SDValue NonConstantVal; 7861 SDValue CCOp; 7862 bool SwapSelectOps; 7863 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps, 7864 NonConstantVal, DAG)) 7865 return SDValue(); 7866 7867 // Slct is now know to be the desired identity constant when CC is true. 7868 SDValue TrueVal = OtherOp; 7869 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT, 7870 OtherOp, NonConstantVal); 7871 // Unless SwapSelectOps says CC should be false. 7872 if (SwapSelectOps) 7873 std::swap(TrueVal, FalseVal); 7874 7875 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, 7876 CCOp, TrueVal, FalseVal); 7877 } 7878 7879 // Attempt combineSelectAndUse on each operand of a commutative operator N. 7880 static 7881 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes, 7882 TargetLowering::DAGCombinerInfo &DCI) { 7883 SDValue N0 = N->getOperand(0); 7884 SDValue N1 = N->getOperand(1); 7885 if (N0.getNode()->hasOneUse()) { 7886 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes); 7887 if (Result.getNode()) 7888 return Result; 7889 } 7890 if (N1.getNode()->hasOneUse()) { 7891 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes); 7892 if (Result.getNode()) 7893 return Result; 7894 } 7895 return SDValue(); 7896 } 7897 7898 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction 7899 // (only after legalization). 7900 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1, 7901 TargetLowering::DAGCombinerInfo &DCI, 7902 const ARMSubtarget *Subtarget) { 7903 7904 // Only perform optimization if after legalize, and if NEON is available. We 7905 // also expected both operands to be BUILD_VECTORs. 7906 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON() 7907 || N0.getOpcode() != ISD::BUILD_VECTOR 7908 || N1.getOpcode() != ISD::BUILD_VECTOR) 7909 return SDValue(); 7910 7911 // Check output type since VPADDL operand elements can only be 8, 16, or 32. 7912 EVT VT = N->getValueType(0); 7913 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64) 7914 return SDValue(); 7915 7916 // Check that the vector operands are of the right form. 7917 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR 7918 // operands, where N is the size of the formed vector. 7919 // Each EXTRACT_VECTOR should have the same input vector and odd or even 7920 // index such that we have a pair wise add pattern. 7921 7922 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing. 7923 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT) 7924 return SDValue(); 7925 SDValue Vec = N0->getOperand(0)->getOperand(0); 7926 SDNode *V = Vec.getNode(); 7927 unsigned nextIndex = 0; 7928 7929 // For each operands to the ADD which are BUILD_VECTORs, 7930 // check to see if each of their operands are an EXTRACT_VECTOR with 7931 // the same vector and appropriate index. 7932 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) { 7933 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT 7934 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) { 7935 7936 SDValue ExtVec0 = N0->getOperand(i); 7937 SDValue ExtVec1 = N1->getOperand(i); 7938 7939 // First operand is the vector, verify its the same. 7940 if (V != ExtVec0->getOperand(0).getNode() || 7941 V != ExtVec1->getOperand(0).getNode()) 7942 return SDValue(); 7943 7944 // Second is the constant, verify its correct. 7945 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1)); 7946 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1)); 7947 7948 // For the constant, we want to see all the even or all the odd. 7949 if (!C0 || !C1 || C0->getZExtValue() != nextIndex 7950 || C1->getZExtValue() != nextIndex+1) 7951 return SDValue(); 7952 7953 // Increment index. 7954 nextIndex+=2; 7955 } else 7956 return SDValue(); 7957 } 7958 7959 // Create VPADDL node. 7960 SelectionDAG &DAG = DCI.DAG; 7961 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 7962 7963 // Build operand list. 7964 SmallVector<SDValue, 8> Ops; 7965 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, 7966 TLI.getPointerTy())); 7967 7968 // Input is the vector. 7969 Ops.push_back(Vec); 7970 7971 // Get widened type and narrowed type. 7972 MVT widenType; 7973 unsigned numElem = VT.getVectorNumElements(); 7974 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) { 7975 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break; 7976 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break; 7977 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break; 7978 default: 7979 llvm_unreachable("Invalid vector element type for padd optimization."); 7980 } 7981 7982 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), 7983 widenType, &Ops[0], Ops.size()); 7984 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, tmp); 7985 } 7986 7987 static SDValue findMUL_LOHI(SDValue V) { 7988 if (V->getOpcode() == ISD::UMUL_LOHI || 7989 V->getOpcode() == ISD::SMUL_LOHI) 7990 return V; 7991 return SDValue(); 7992 } 7993 7994 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode, 7995 TargetLowering::DAGCombinerInfo &DCI, 7996 const ARMSubtarget *Subtarget) { 7997 7998 if (Subtarget->isThumb1Only()) return SDValue(); 7999 8000 // Only perform the checks after legalize when the pattern is available. 8001 if (DCI.isBeforeLegalize()) return SDValue(); 8002 8003 // Look for multiply add opportunities. 8004 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where 8005 // each add nodes consumes a value from ISD::UMUL_LOHI and there is 8006 // a glue link from the first add to the second add. 8007 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by 8008 // a S/UMLAL instruction. 8009 // loAdd UMUL_LOHI 8010 // \ / :lo \ :hi 8011 // \ / \ [no multiline comment] 8012 // ADDC | hiAdd 8013 // \ :glue / / 8014 // \ / / 8015 // ADDE 8016 // 8017 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC"); 8018 SDValue AddcOp0 = AddcNode->getOperand(0); 8019 SDValue AddcOp1 = AddcNode->getOperand(1); 8020 8021 // Check if the two operands are from the same mul_lohi node. 8022 if (AddcOp0.getNode() == AddcOp1.getNode()) 8023 return SDValue(); 8024 8025 assert(AddcNode->getNumValues() == 2 && 8026 AddcNode->getValueType(0) == MVT::i32 && 8027 "Expect ADDC with two result values. First: i32"); 8028 8029 // Check that we have a glued ADDC node. 8030 if (AddcNode->getValueType(1) != MVT::Glue) 8031 return SDValue(); 8032 8033 // Check that the ADDC adds the low result of the S/UMUL_LOHI. 8034 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI && 8035 AddcOp0->getOpcode() != ISD::SMUL_LOHI && 8036 AddcOp1->getOpcode() != ISD::UMUL_LOHI && 8037 AddcOp1->getOpcode() != ISD::SMUL_LOHI) 8038 return SDValue(); 8039 8040 // Look for the glued ADDE. 8041 SDNode* AddeNode = AddcNode->getGluedUser(); 8042 if (AddeNode == NULL) 8043 return SDValue(); 8044 8045 // Make sure it is really an ADDE. 8046 if (AddeNode->getOpcode() != ISD::ADDE) 8047 return SDValue(); 8048 8049 assert(AddeNode->getNumOperands() == 3 && 8050 AddeNode->getOperand(2).getValueType() == MVT::Glue && 8051 "ADDE node has the wrong inputs"); 8052 8053 // Check for the triangle shape. 8054 SDValue AddeOp0 = AddeNode->getOperand(0); 8055 SDValue AddeOp1 = AddeNode->getOperand(1); 8056 8057 // Make sure that the ADDE operands are not coming from the same node. 8058 if (AddeOp0.getNode() == AddeOp1.getNode()) 8059 return SDValue(); 8060 8061 // Find the MUL_LOHI node walking up ADDE's operands. 8062 bool IsLeftOperandMUL = false; 8063 SDValue MULOp = findMUL_LOHI(AddeOp0); 8064 if (MULOp == SDValue()) 8065 MULOp = findMUL_LOHI(AddeOp1); 8066 else 8067 IsLeftOperandMUL = true; 8068 if (MULOp == SDValue()) 8069 return SDValue(); 8070 8071 // Figure out the right opcode. 8072 unsigned Opc = MULOp->getOpcode(); 8073 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL; 8074 8075 // Figure out the high and low input values to the MLAL node. 8076 SDValue* HiMul = &MULOp; 8077 SDValue* HiAdd = NULL; 8078 SDValue* LoMul = NULL; 8079 SDValue* LowAdd = NULL; 8080 8081 if (IsLeftOperandMUL) 8082 HiAdd = &AddeOp1; 8083 else 8084 HiAdd = &AddeOp0; 8085 8086 8087 if (AddcOp0->getOpcode() == Opc) { 8088 LoMul = &AddcOp0; 8089 LowAdd = &AddcOp1; 8090 } 8091 if (AddcOp1->getOpcode() == Opc) { 8092 LoMul = &AddcOp1; 8093 LowAdd = &AddcOp0; 8094 } 8095 8096 if (LoMul == NULL) 8097 return SDValue(); 8098 8099 if (LoMul->getNode() != HiMul->getNode()) 8100 return SDValue(); 8101 8102 // Create the merged node. 8103 SelectionDAG &DAG = DCI.DAG; 8104 8105 // Build operand list. 8106 SmallVector<SDValue, 8> Ops; 8107 Ops.push_back(LoMul->getOperand(0)); 8108 Ops.push_back(LoMul->getOperand(1)); 8109 Ops.push_back(*LowAdd); 8110 Ops.push_back(*HiAdd); 8111 8112 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode), 8113 DAG.getVTList(MVT::i32, MVT::i32), 8114 &Ops[0], Ops.size()); 8115 8116 // Replace the ADDs' nodes uses by the MLA node's values. 8117 SDValue HiMLALResult(MLALNode.getNode(), 1); 8118 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult); 8119 8120 SDValue LoMLALResult(MLALNode.getNode(), 0); 8121 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult); 8122 8123 // Return original node to notify the driver to stop replacing. 8124 SDValue resNode(AddcNode, 0); 8125 return resNode; 8126 } 8127 8128 /// PerformADDCCombine - Target-specific dag combine transform from 8129 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL. 8130 static SDValue PerformADDCCombine(SDNode *N, 8131 TargetLowering::DAGCombinerInfo &DCI, 8132 const ARMSubtarget *Subtarget) { 8133 8134 return AddCombineTo64bitMLAL(N, DCI, Subtarget); 8135 8136 } 8137 8138 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with 8139 /// operands N0 and N1. This is a helper for PerformADDCombine that is 8140 /// called with the default operands, and if that fails, with commuted 8141 /// operands. 8142 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1, 8143 TargetLowering::DAGCombinerInfo &DCI, 8144 const ARMSubtarget *Subtarget){ 8145 8146 // Attempt to create vpaddl for this add. 8147 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget); 8148 if (Result.getNode()) 8149 return Result; 8150 8151 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c)) 8152 if (N0.getNode()->hasOneUse()) { 8153 SDValue Result = combineSelectAndUse(N, N0, N1, DCI); 8154 if (Result.getNode()) return Result; 8155 } 8156 return SDValue(); 8157 } 8158 8159 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD. 8160 /// 8161 static SDValue PerformADDCombine(SDNode *N, 8162 TargetLowering::DAGCombinerInfo &DCI, 8163 const ARMSubtarget *Subtarget) { 8164 SDValue N0 = N->getOperand(0); 8165 SDValue N1 = N->getOperand(1); 8166 8167 // First try with the default operand order. 8168 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget); 8169 if (Result.getNode()) 8170 return Result; 8171 8172 // If that didn't work, try again with the operands commuted. 8173 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget); 8174 } 8175 8176 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB. 8177 /// 8178 static SDValue PerformSUBCombine(SDNode *N, 8179 TargetLowering::DAGCombinerInfo &DCI) { 8180 SDValue N0 = N->getOperand(0); 8181 SDValue N1 = N->getOperand(1); 8182 8183 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c)) 8184 if (N1.getNode()->hasOneUse()) { 8185 SDValue Result = combineSelectAndUse(N, N1, N0, DCI); 8186 if (Result.getNode()) return Result; 8187 } 8188 8189 return SDValue(); 8190 } 8191 8192 /// PerformVMULCombine 8193 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the 8194 /// special multiplier accumulator forwarding. 8195 /// vmul d3, d0, d2 8196 /// vmla d3, d1, d2 8197 /// is faster than 8198 /// vadd d3, d0, d1 8199 /// vmul d3, d3, d2 8200 static SDValue PerformVMULCombine(SDNode *N, 8201 TargetLowering::DAGCombinerInfo &DCI, 8202 const ARMSubtarget *Subtarget) { 8203 if (!Subtarget->hasVMLxForwarding()) 8204 return SDValue(); 8205 8206 SelectionDAG &DAG = DCI.DAG; 8207 SDValue N0 = N->getOperand(0); 8208 SDValue N1 = N->getOperand(1); 8209 unsigned Opcode = N0.getOpcode(); 8210 if (Opcode != ISD::ADD && Opcode != ISD::SUB && 8211 Opcode != ISD::FADD && Opcode != ISD::FSUB) { 8212 Opcode = N1.getOpcode(); 8213 if (Opcode != ISD::ADD && Opcode != ISD::SUB && 8214 Opcode != ISD::FADD && Opcode != ISD::FSUB) 8215 return SDValue(); 8216 std::swap(N0, N1); 8217 } 8218 8219 EVT VT = N->getValueType(0); 8220 SDLoc DL(N); 8221 SDValue N00 = N0->getOperand(0); 8222 SDValue N01 = N0->getOperand(1); 8223 return DAG.getNode(Opcode, DL, VT, 8224 DAG.getNode(ISD::MUL, DL, VT, N00, N1), 8225 DAG.getNode(ISD::MUL, DL, VT, N01, N1)); 8226 } 8227 8228 static SDValue PerformMULCombine(SDNode *N, 8229 TargetLowering::DAGCombinerInfo &DCI, 8230 const ARMSubtarget *Subtarget) { 8231 SelectionDAG &DAG = DCI.DAG; 8232 8233 if (Subtarget->isThumb1Only()) 8234 return SDValue(); 8235 8236 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer()) 8237 return SDValue(); 8238 8239 EVT VT = N->getValueType(0); 8240 if (VT.is64BitVector() || VT.is128BitVector()) 8241 return PerformVMULCombine(N, DCI, Subtarget); 8242 if (VT != MVT::i32) 8243 return SDValue(); 8244 8245 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)); 8246 if (!C) 8247 return SDValue(); 8248 8249 int64_t MulAmt = C->getSExtValue(); 8250 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt); 8251 8252 ShiftAmt = ShiftAmt & (32 - 1); 8253 SDValue V = N->getOperand(0); 8254 SDLoc DL(N); 8255 8256 SDValue Res; 8257 MulAmt >>= ShiftAmt; 8258 8259 if (MulAmt >= 0) { 8260 if (isPowerOf2_32(MulAmt - 1)) { 8261 // (mul x, 2^N + 1) => (add (shl x, N), x) 8262 Res = DAG.getNode(ISD::ADD, DL, VT, 8263 V, 8264 DAG.getNode(ISD::SHL, DL, VT, 8265 V, 8266 DAG.getConstant(Log2_32(MulAmt - 1), 8267 MVT::i32))); 8268 } else if (isPowerOf2_32(MulAmt + 1)) { 8269 // (mul x, 2^N - 1) => (sub (shl x, N), x) 8270 Res = DAG.getNode(ISD::SUB, DL, VT, 8271 DAG.getNode(ISD::SHL, DL, VT, 8272 V, 8273 DAG.getConstant(Log2_32(MulAmt + 1), 8274 MVT::i32)), 8275 V); 8276 } else 8277 return SDValue(); 8278 } else { 8279 uint64_t MulAmtAbs = -MulAmt; 8280 if (isPowerOf2_32(MulAmtAbs + 1)) { 8281 // (mul x, -(2^N - 1)) => (sub x, (shl x, N)) 8282 Res = DAG.getNode(ISD::SUB, DL, VT, 8283 V, 8284 DAG.getNode(ISD::SHL, DL, VT, 8285 V, 8286 DAG.getConstant(Log2_32(MulAmtAbs + 1), 8287 MVT::i32))); 8288 } else if (isPowerOf2_32(MulAmtAbs - 1)) { 8289 // (mul x, -(2^N + 1)) => - (add (shl x, N), x) 8290 Res = DAG.getNode(ISD::ADD, DL, VT, 8291 V, 8292 DAG.getNode(ISD::SHL, DL, VT, 8293 V, 8294 DAG.getConstant(Log2_32(MulAmtAbs-1), 8295 MVT::i32))); 8296 Res = DAG.getNode(ISD::SUB, DL, VT, 8297 DAG.getConstant(0, MVT::i32),Res); 8298 8299 } else 8300 return SDValue(); 8301 } 8302 8303 if (ShiftAmt != 0) 8304 Res = DAG.getNode(ISD::SHL, DL, VT, 8305 Res, DAG.getConstant(ShiftAmt, MVT::i32)); 8306 8307 // Do not add new nodes to DAG combiner worklist. 8308 DCI.CombineTo(N, Res, false); 8309 return SDValue(); 8310 } 8311 8312 static SDValue PerformANDCombine(SDNode *N, 8313 TargetLowering::DAGCombinerInfo &DCI, 8314 const ARMSubtarget *Subtarget) { 8315 8316 // Attempt to use immediate-form VBIC 8317 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1)); 8318 SDLoc dl(N); 8319 EVT VT = N->getValueType(0); 8320 SelectionDAG &DAG = DCI.DAG; 8321 8322 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT)) 8323 return SDValue(); 8324 8325 APInt SplatBits, SplatUndef; 8326 unsigned SplatBitSize; 8327 bool HasAnyUndefs; 8328 if (BVN && 8329 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) { 8330 if (SplatBitSize <= 64) { 8331 EVT VbicVT; 8332 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(), 8333 SplatUndef.getZExtValue(), SplatBitSize, 8334 DAG, VbicVT, VT.is128BitVector(), 8335 OtherModImm); 8336 if (Val.getNode()) { 8337 SDValue Input = 8338 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0)); 8339 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val); 8340 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic); 8341 } 8342 } 8343 } 8344 8345 if (!Subtarget->isThumb1Only()) { 8346 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) 8347 SDValue Result = combineSelectAndUseCommutative(N, true, DCI); 8348 if (Result.getNode()) 8349 return Result; 8350 } 8351 8352 return SDValue(); 8353 } 8354 8355 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR 8356 static SDValue PerformORCombine(SDNode *N, 8357 TargetLowering::DAGCombinerInfo &DCI, 8358 const ARMSubtarget *Subtarget) { 8359 // Attempt to use immediate-form VORR 8360 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1)); 8361 SDLoc dl(N); 8362 EVT VT = N->getValueType(0); 8363 SelectionDAG &DAG = DCI.DAG; 8364 8365 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT)) 8366 return SDValue(); 8367 8368 APInt SplatBits, SplatUndef; 8369 unsigned SplatBitSize; 8370 bool HasAnyUndefs; 8371 if (BVN && Subtarget->hasNEON() && 8372 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) { 8373 if (SplatBitSize <= 64) { 8374 EVT VorrVT; 8375 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(), 8376 SplatUndef.getZExtValue(), SplatBitSize, 8377 DAG, VorrVT, VT.is128BitVector(), 8378 OtherModImm); 8379 if (Val.getNode()) { 8380 SDValue Input = 8381 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0)); 8382 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val); 8383 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr); 8384 } 8385 } 8386 } 8387 8388 if (!Subtarget->isThumb1Only()) { 8389 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c)) 8390 SDValue Result = combineSelectAndUseCommutative(N, false, DCI); 8391 if (Result.getNode()) 8392 return Result; 8393 } 8394 8395 // The code below optimizes (or (and X, Y), Z). 8396 // The AND operand needs to have a single user to make these optimizations 8397 // profitable. 8398 SDValue N0 = N->getOperand(0); 8399 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse()) 8400 return SDValue(); 8401 SDValue N1 = N->getOperand(1); 8402 8403 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant. 8404 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() && 8405 DAG.getTargetLoweringInfo().isTypeLegal(VT)) { 8406 APInt SplatUndef; 8407 unsigned SplatBitSize; 8408 bool HasAnyUndefs; 8409 8410 APInt SplatBits0, SplatBits1; 8411 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1)); 8412 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1)); 8413 // Ensure that the second operand of both ands are constants 8414 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize, 8415 HasAnyUndefs) && !HasAnyUndefs) { 8416 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize, 8417 HasAnyUndefs) && !HasAnyUndefs) { 8418 // Ensure that the bit width of the constants are the same and that 8419 // the splat arguments are logical inverses as per the pattern we 8420 // are trying to simplify. 8421 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() && 8422 SplatBits0 == ~SplatBits1) { 8423 // Canonicalize the vector type to make instruction selection 8424 // simpler. 8425 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32; 8426 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT, 8427 N0->getOperand(1), 8428 N0->getOperand(0), 8429 N1->getOperand(0)); 8430 return DAG.getNode(ISD::BITCAST, dl, VT, Result); 8431 } 8432 } 8433 } 8434 } 8435 8436 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when 8437 // reasonable. 8438 8439 // BFI is only available on V6T2+ 8440 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops()) 8441 return SDValue(); 8442 8443 SDLoc DL(N); 8444 // 1) or (and A, mask), val => ARMbfi A, val, mask 8445 // iff (val & mask) == val 8446 // 8447 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask 8448 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2) 8449 // && mask == ~mask2 8450 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2) 8451 // && ~mask == mask2 8452 // (i.e., copy a bitfield value into another bitfield of the same width) 8453 8454 if (VT != MVT::i32) 8455 return SDValue(); 8456 8457 SDValue N00 = N0.getOperand(0); 8458 8459 // The value and the mask need to be constants so we can verify this is 8460 // actually a bitfield set. If the mask is 0xffff, we can do better 8461 // via a movt instruction, so don't use BFI in that case. 8462 SDValue MaskOp = N0.getOperand(1); 8463 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp); 8464 if (!MaskC) 8465 return SDValue(); 8466 unsigned Mask = MaskC->getZExtValue(); 8467 if (Mask == 0xffff) 8468 return SDValue(); 8469 SDValue Res; 8470 // Case (1): or (and A, mask), val => ARMbfi A, val, mask 8471 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 8472 if (N1C) { 8473 unsigned Val = N1C->getZExtValue(); 8474 if ((Val & ~Mask) != Val) 8475 return SDValue(); 8476 8477 if (ARM::isBitFieldInvertedMask(Mask)) { 8478 Val >>= countTrailingZeros(~Mask); 8479 8480 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, 8481 DAG.getConstant(Val, MVT::i32), 8482 DAG.getConstant(Mask, MVT::i32)); 8483 8484 // Do not add new nodes to DAG combiner worklist. 8485 DCI.CombineTo(N, Res, false); 8486 return SDValue(); 8487 } 8488 } else if (N1.getOpcode() == ISD::AND) { 8489 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask 8490 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); 8491 if (!N11C) 8492 return SDValue(); 8493 unsigned Mask2 = N11C->getZExtValue(); 8494 8495 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern 8496 // as is to match. 8497 if (ARM::isBitFieldInvertedMask(Mask) && 8498 (Mask == ~Mask2)) { 8499 // The pack halfword instruction works better for masks that fit it, 8500 // so use that when it's available. 8501 if (Subtarget->hasT2ExtractPack() && 8502 (Mask == 0xffff || Mask == 0xffff0000)) 8503 return SDValue(); 8504 // 2a 8505 unsigned amt = countTrailingZeros(Mask2); 8506 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0), 8507 DAG.getConstant(amt, MVT::i32)); 8508 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res, 8509 DAG.getConstant(Mask, MVT::i32)); 8510 // Do not add new nodes to DAG combiner worklist. 8511 DCI.CombineTo(N, Res, false); 8512 return SDValue(); 8513 } else if (ARM::isBitFieldInvertedMask(~Mask) && 8514 (~Mask == Mask2)) { 8515 // The pack halfword instruction works better for masks that fit it, 8516 // so use that when it's available. 8517 if (Subtarget->hasT2ExtractPack() && 8518 (Mask2 == 0xffff || Mask2 == 0xffff0000)) 8519 return SDValue(); 8520 // 2b 8521 unsigned lsb = countTrailingZeros(Mask); 8522 Res = DAG.getNode(ISD::SRL, DL, VT, N00, 8523 DAG.getConstant(lsb, MVT::i32)); 8524 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res, 8525 DAG.getConstant(Mask2, MVT::i32)); 8526 // Do not add new nodes to DAG combiner worklist. 8527 DCI.CombineTo(N, Res, false); 8528 return SDValue(); 8529 } 8530 } 8531 8532 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) && 8533 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) && 8534 ARM::isBitFieldInvertedMask(~Mask)) { 8535 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask 8536 // where lsb(mask) == #shamt and masked bits of B are known zero. 8537 SDValue ShAmt = N00.getOperand(1); 8538 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue(); 8539 unsigned LSB = countTrailingZeros(Mask); 8540 if (ShAmtC != LSB) 8541 return SDValue(); 8542 8543 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0), 8544 DAG.getConstant(~Mask, MVT::i32)); 8545 8546 // Do not add new nodes to DAG combiner worklist. 8547 DCI.CombineTo(N, Res, false); 8548 } 8549 8550 return SDValue(); 8551 } 8552 8553 static SDValue PerformXORCombine(SDNode *N, 8554 TargetLowering::DAGCombinerInfo &DCI, 8555 const ARMSubtarget *Subtarget) { 8556 EVT VT = N->getValueType(0); 8557 SelectionDAG &DAG = DCI.DAG; 8558 8559 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT)) 8560 return SDValue(); 8561 8562 if (!Subtarget->isThumb1Only()) { 8563 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c)) 8564 SDValue Result = combineSelectAndUseCommutative(N, false, DCI); 8565 if (Result.getNode()) 8566 return Result; 8567 } 8568 8569 return SDValue(); 8570 } 8571 8572 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff 8573 /// the bits being cleared by the AND are not demanded by the BFI. 8574 static SDValue PerformBFICombine(SDNode *N, 8575 TargetLowering::DAGCombinerInfo &DCI) { 8576 SDValue N1 = N->getOperand(1); 8577 if (N1.getOpcode() == ISD::AND) { 8578 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); 8579 if (!N11C) 8580 return SDValue(); 8581 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue(); 8582 unsigned LSB = countTrailingZeros(~InvMask); 8583 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB; 8584 unsigned Mask = (1 << Width)-1; 8585 unsigned Mask2 = N11C->getZExtValue(); 8586 if ((Mask & (~Mask2)) == 0) 8587 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0), 8588 N->getOperand(0), N1.getOperand(0), 8589 N->getOperand(2)); 8590 } 8591 return SDValue(); 8592 } 8593 8594 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for 8595 /// ARMISD::VMOVRRD. 8596 static SDValue PerformVMOVRRDCombine(SDNode *N, 8597 TargetLowering::DAGCombinerInfo &DCI) { 8598 // vmovrrd(vmovdrr x, y) -> x,y 8599 SDValue InDouble = N->getOperand(0); 8600 if (InDouble.getOpcode() == ARMISD::VMOVDRR) 8601 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1)); 8602 8603 // vmovrrd(load f64) -> (load i32), (load i32) 8604 SDNode *InNode = InDouble.getNode(); 8605 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() && 8606 InNode->getValueType(0) == MVT::f64 && 8607 InNode->getOperand(1).getOpcode() == ISD::FrameIndex && 8608 !cast<LoadSDNode>(InNode)->isVolatile()) { 8609 // TODO: Should this be done for non-FrameIndex operands? 8610 LoadSDNode *LD = cast<LoadSDNode>(InNode); 8611 8612 SelectionDAG &DAG = DCI.DAG; 8613 SDLoc DL(LD); 8614 SDValue BasePtr = LD->getBasePtr(); 8615 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr, 8616 LD->getPointerInfo(), LD->isVolatile(), 8617 LD->isNonTemporal(), LD->isInvariant(), 8618 LD->getAlignment()); 8619 8620 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr, 8621 DAG.getConstant(4, MVT::i32)); 8622 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr, 8623 LD->getPointerInfo(), LD->isVolatile(), 8624 LD->isNonTemporal(), LD->isInvariant(), 8625 std::min(4U, LD->getAlignment() / 2)); 8626 8627 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1)); 8628 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2); 8629 DCI.RemoveFromWorklist(LD); 8630 DAG.DeleteNode(LD); 8631 return Result; 8632 } 8633 8634 return SDValue(); 8635 } 8636 8637 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for 8638 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands. 8639 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) { 8640 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X) 8641 SDValue Op0 = N->getOperand(0); 8642 SDValue Op1 = N->getOperand(1); 8643 if (Op0.getOpcode() == ISD::BITCAST) 8644 Op0 = Op0.getOperand(0); 8645 if (Op1.getOpcode() == ISD::BITCAST) 8646 Op1 = Op1.getOperand(0); 8647 if (Op0.getOpcode() == ARMISD::VMOVRRD && 8648 Op0.getNode() == Op1.getNode() && 8649 Op0.getResNo() == 0 && Op1.getResNo() == 1) 8650 return DAG.getNode(ISD::BITCAST, SDLoc(N), 8651 N->getValueType(0), Op0.getOperand(0)); 8652 return SDValue(); 8653 } 8654 8655 /// PerformSTORECombine - Target-specific dag combine xforms for 8656 /// ISD::STORE. 8657 static SDValue PerformSTORECombine(SDNode *N, 8658 TargetLowering::DAGCombinerInfo &DCI) { 8659 StoreSDNode *St = cast<StoreSDNode>(N); 8660 if (St->isVolatile()) 8661 return SDValue(); 8662 8663 // Optimize trunc store (of multiple scalars) to shuffle and store. First, 8664 // pack all of the elements in one place. Next, store to memory in fewer 8665 // chunks. 8666 SDValue StVal = St->getValue(); 8667 EVT VT = StVal.getValueType(); 8668 if (St->isTruncatingStore() && VT.isVector()) { 8669 SelectionDAG &DAG = DCI.DAG; 8670 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 8671 EVT StVT = St->getMemoryVT(); 8672 unsigned NumElems = VT.getVectorNumElements(); 8673 assert(StVT != VT && "Cannot truncate to the same type"); 8674 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits(); 8675 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits(); 8676 8677 // From, To sizes and ElemCount must be pow of two 8678 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue(); 8679 8680 // We are going to use the original vector elt for storing. 8681 // Accumulated smaller vector elements must be a multiple of the store size. 8682 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue(); 8683 8684 unsigned SizeRatio = FromEltSz / ToEltSz; 8685 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits()); 8686 8687 // Create a type on which we perform the shuffle. 8688 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(), 8689 NumElems*SizeRatio); 8690 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits()); 8691 8692 SDLoc DL(St); 8693 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal); 8694 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1); 8695 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio; 8696 8697 // Can't shuffle using an illegal type. 8698 if (!TLI.isTypeLegal(WideVecVT)) return SDValue(); 8699 8700 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec, 8701 DAG.getUNDEF(WideVec.getValueType()), 8702 ShuffleVec.data()); 8703 // At this point all of the data is stored at the bottom of the 8704 // register. We now need to save it to mem. 8705 8706 // Find the largest store unit 8707 MVT StoreType = MVT::i8; 8708 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE; 8709 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) { 8710 MVT Tp = (MVT::SimpleValueType)tp; 8711 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz) 8712 StoreType = Tp; 8713 } 8714 // Didn't find a legal store type. 8715 if (!TLI.isTypeLegal(StoreType)) 8716 return SDValue(); 8717 8718 // Bitcast the original vector into a vector of store-size units 8719 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(), 8720 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits()); 8721 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits()); 8722 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff); 8723 SmallVector<SDValue, 8> Chains; 8724 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8, 8725 TLI.getPointerTy()); 8726 SDValue BasePtr = St->getBasePtr(); 8727 8728 // Perform one or more big stores into memory. 8729 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits(); 8730 for (unsigned I = 0; I < E; I++) { 8731 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, 8732 StoreType, ShuffWide, 8733 DAG.getIntPtrConstant(I)); 8734 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr, 8735 St->getPointerInfo(), St->isVolatile(), 8736 St->isNonTemporal(), St->getAlignment()); 8737 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr, 8738 Increment); 8739 Chains.push_back(Ch); 8740 } 8741 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0], 8742 Chains.size()); 8743 } 8744 8745 if (!ISD::isNormalStore(St)) 8746 return SDValue(); 8747 8748 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and 8749 // ARM stores of arguments in the same cache line. 8750 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR && 8751 StVal.getNode()->hasOneUse()) { 8752 SelectionDAG &DAG = DCI.DAG; 8753 SDLoc DL(St); 8754 SDValue BasePtr = St->getBasePtr(); 8755 SDValue NewST1 = DAG.getStore(St->getChain(), DL, 8756 StVal.getNode()->getOperand(0), BasePtr, 8757 St->getPointerInfo(), St->isVolatile(), 8758 St->isNonTemporal(), St->getAlignment()); 8759 8760 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr, 8761 DAG.getConstant(4, MVT::i32)); 8762 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1), 8763 OffsetPtr, St->getPointerInfo(), St->isVolatile(), 8764 St->isNonTemporal(), 8765 std::min(4U, St->getAlignment() / 2)); 8766 } 8767 8768 if (StVal.getValueType() != MVT::i64 || 8769 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT) 8770 return SDValue(); 8771 8772 // Bitcast an i64 store extracted from a vector to f64. 8773 // Otherwise, the i64 value will be legalized to a pair of i32 values. 8774 SelectionDAG &DAG = DCI.DAG; 8775 SDLoc dl(StVal); 8776 SDValue IntVec = StVal.getOperand(0); 8777 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, 8778 IntVec.getValueType().getVectorNumElements()); 8779 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec); 8780 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, 8781 Vec, StVal.getOperand(1)); 8782 dl = SDLoc(N); 8783 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt); 8784 // Make the DAGCombiner fold the bitcasts. 8785 DCI.AddToWorklist(Vec.getNode()); 8786 DCI.AddToWorklist(ExtElt.getNode()); 8787 DCI.AddToWorklist(V.getNode()); 8788 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(), 8789 St->getPointerInfo(), St->isVolatile(), 8790 St->isNonTemporal(), St->getAlignment(), 8791 St->getTBAAInfo()); 8792 } 8793 8794 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node 8795 /// are normal, non-volatile loads. If so, it is profitable to bitcast an 8796 /// i64 vector to have f64 elements, since the value can then be loaded 8797 /// directly into a VFP register. 8798 static bool hasNormalLoadOperand(SDNode *N) { 8799 unsigned NumElts = N->getValueType(0).getVectorNumElements(); 8800 for (unsigned i = 0; i < NumElts; ++i) { 8801 SDNode *Elt = N->getOperand(i).getNode(); 8802 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile()) 8803 return true; 8804 } 8805 return false; 8806 } 8807 8808 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for 8809 /// ISD::BUILD_VECTOR. 8810 static SDValue PerformBUILD_VECTORCombine(SDNode *N, 8811 TargetLowering::DAGCombinerInfo &DCI){ 8812 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X): 8813 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value 8814 // into a pair of GPRs, which is fine when the value is used as a scalar, 8815 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD. 8816 SelectionDAG &DAG = DCI.DAG; 8817 if (N->getNumOperands() == 2) { 8818 SDValue RV = PerformVMOVDRRCombine(N, DAG); 8819 if (RV.getNode()) 8820 return RV; 8821 } 8822 8823 // Load i64 elements as f64 values so that type legalization does not split 8824 // them up into i32 values. 8825 EVT VT = N->getValueType(0); 8826 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N)) 8827 return SDValue(); 8828 SDLoc dl(N); 8829 SmallVector<SDValue, 8> Ops; 8830 unsigned NumElts = VT.getVectorNumElements(); 8831 for (unsigned i = 0; i < NumElts; ++i) { 8832 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i)); 8833 Ops.push_back(V); 8834 // Make the DAGCombiner fold the bitcast. 8835 DCI.AddToWorklist(V.getNode()); 8836 } 8837 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts); 8838 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts); 8839 return DAG.getNode(ISD::BITCAST, dl, VT, BV); 8840 } 8841 8842 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR. 8843 static SDValue 8844 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { 8845 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR. 8846 // At that time, we may have inserted bitcasts from integer to float. 8847 // If these bitcasts have survived DAGCombine, change the lowering of this 8848 // BUILD_VECTOR in something more vector friendly, i.e., that does not 8849 // force to use floating point types. 8850 8851 // Make sure we can change the type of the vector. 8852 // This is possible iff: 8853 // 1. The vector is only used in a bitcast to a integer type. I.e., 8854 // 1.1. Vector is used only once. 8855 // 1.2. Use is a bit convert to an integer type. 8856 // 2. The size of its operands are 32-bits (64-bits are not legal). 8857 EVT VT = N->getValueType(0); 8858 EVT EltVT = VT.getVectorElementType(); 8859 8860 // Check 1.1. and 2. 8861 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse()) 8862 return SDValue(); 8863 8864 // By construction, the input type must be float. 8865 assert(EltVT == MVT::f32 && "Unexpected type!"); 8866 8867 // Check 1.2. 8868 SDNode *Use = *N->use_begin(); 8869 if (Use->getOpcode() != ISD::BITCAST || 8870 Use->getValueType(0).isFloatingPoint()) 8871 return SDValue(); 8872 8873 // Check profitability. 8874 // Model is, if more than half of the relevant operands are bitcast from 8875 // i32, turn the build_vector into a sequence of insert_vector_elt. 8876 // Relevant operands are everything that is not statically 8877 // (i.e., at compile time) bitcasted. 8878 unsigned NumOfBitCastedElts = 0; 8879 unsigned NumElts = VT.getVectorNumElements(); 8880 unsigned NumOfRelevantElts = NumElts; 8881 for (unsigned Idx = 0; Idx < NumElts; ++Idx) { 8882 SDValue Elt = N->getOperand(Idx); 8883 if (Elt->getOpcode() == ISD::BITCAST) { 8884 // Assume only bit cast to i32 will go away. 8885 if (Elt->getOperand(0).getValueType() == MVT::i32) 8886 ++NumOfBitCastedElts; 8887 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt)) 8888 // Constants are statically casted, thus do not count them as 8889 // relevant operands. 8890 --NumOfRelevantElts; 8891 } 8892 8893 // Check if more than half of the elements require a non-free bitcast. 8894 if (NumOfBitCastedElts <= NumOfRelevantElts / 2) 8895 return SDValue(); 8896 8897 SelectionDAG &DAG = DCI.DAG; 8898 // Create the new vector type. 8899 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts); 8900 // Check if the type is legal. 8901 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 8902 if (!TLI.isTypeLegal(VecVT)) 8903 return SDValue(); 8904 8905 // Combine: 8906 // ARMISD::BUILD_VECTOR E1, E2, ..., EN. 8907 // => BITCAST INSERT_VECTOR_ELT 8908 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1), 8909 // (BITCAST EN), N. 8910 SDValue Vec = DAG.getUNDEF(VecVT); 8911 SDLoc dl(N); 8912 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) { 8913 SDValue V = N->getOperand(Idx); 8914 if (V.getOpcode() == ISD::UNDEF) 8915 continue; 8916 if (V.getOpcode() == ISD::BITCAST && 8917 V->getOperand(0).getValueType() == MVT::i32) 8918 // Fold obvious case. 8919 V = V.getOperand(0); 8920 else { 8921 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V); 8922 // Make the DAGCombiner fold the bitcasts. 8923 DCI.AddToWorklist(V.getNode()); 8924 } 8925 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32); 8926 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx); 8927 } 8928 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec); 8929 // Make the DAGCombiner fold the bitcasts. 8930 DCI.AddToWorklist(Vec.getNode()); 8931 return Vec; 8932 } 8933 8934 /// PerformInsertEltCombine - Target-specific dag combine xforms for 8935 /// ISD::INSERT_VECTOR_ELT. 8936 static SDValue PerformInsertEltCombine(SDNode *N, 8937 TargetLowering::DAGCombinerInfo &DCI) { 8938 // Bitcast an i64 load inserted into a vector to f64. 8939 // Otherwise, the i64 value will be legalized to a pair of i32 values. 8940 EVT VT = N->getValueType(0); 8941 SDNode *Elt = N->getOperand(1).getNode(); 8942 if (VT.getVectorElementType() != MVT::i64 || 8943 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile()) 8944 return SDValue(); 8945 8946 SelectionDAG &DAG = DCI.DAG; 8947 SDLoc dl(N); 8948 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, 8949 VT.getVectorNumElements()); 8950 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0)); 8951 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1)); 8952 // Make the DAGCombiner fold the bitcasts. 8953 DCI.AddToWorklist(Vec.getNode()); 8954 DCI.AddToWorklist(V.getNode()); 8955 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT, 8956 Vec, V, N->getOperand(2)); 8957 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt); 8958 } 8959 8960 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for 8961 /// ISD::VECTOR_SHUFFLE. 8962 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) { 8963 // The LLVM shufflevector instruction does not require the shuffle mask 8964 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does 8965 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the 8966 // operands do not match the mask length, they are extended by concatenating 8967 // them with undef vectors. That is probably the right thing for other 8968 // targets, but for NEON it is better to concatenate two double-register 8969 // size vector operands into a single quad-register size vector. Do that 8970 // transformation here: 8971 // shuffle(concat(v1, undef), concat(v2, undef)) -> 8972 // shuffle(concat(v1, v2), undef) 8973 SDValue Op0 = N->getOperand(0); 8974 SDValue Op1 = N->getOperand(1); 8975 if (Op0.getOpcode() != ISD::CONCAT_VECTORS || 8976 Op1.getOpcode() != ISD::CONCAT_VECTORS || 8977 Op0.getNumOperands() != 2 || 8978 Op1.getNumOperands() != 2) 8979 return SDValue(); 8980 SDValue Concat0Op1 = Op0.getOperand(1); 8981 SDValue Concat1Op1 = Op1.getOperand(1); 8982 if (Concat0Op1.getOpcode() != ISD::UNDEF || 8983 Concat1Op1.getOpcode() != ISD::UNDEF) 8984 return SDValue(); 8985 // Skip the transformation if any of the types are illegal. 8986 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 8987 EVT VT = N->getValueType(0); 8988 if (!TLI.isTypeLegal(VT) || 8989 !TLI.isTypeLegal(Concat0Op1.getValueType()) || 8990 !TLI.isTypeLegal(Concat1Op1.getValueType())) 8991 return SDValue(); 8992 8993 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, 8994 Op0.getOperand(0), Op1.getOperand(0)); 8995 // Translate the shuffle mask. 8996 SmallVector<int, 16> NewMask; 8997 unsigned NumElts = VT.getVectorNumElements(); 8998 unsigned HalfElts = NumElts/2; 8999 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 9000 for (unsigned n = 0; n < NumElts; ++n) { 9001 int MaskElt = SVN->getMaskElt(n); 9002 int NewElt = -1; 9003 if (MaskElt < (int)HalfElts) 9004 NewElt = MaskElt; 9005 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts)) 9006 NewElt = HalfElts + MaskElt - NumElts; 9007 NewMask.push_back(NewElt); 9008 } 9009 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat, 9010 DAG.getUNDEF(VT), NewMask.data()); 9011 } 9012 9013 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and 9014 /// NEON load/store intrinsics to merge base address updates. 9015 static SDValue CombineBaseUpdate(SDNode *N, 9016 TargetLowering::DAGCombinerInfo &DCI) { 9017 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer()) 9018 return SDValue(); 9019 9020 SelectionDAG &DAG = DCI.DAG; 9021 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID || 9022 N->getOpcode() == ISD::INTRINSIC_W_CHAIN); 9023 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1); 9024 SDValue Addr = N->getOperand(AddrOpIdx); 9025 9026 // Search for a use of the address operand that is an increment. 9027 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), 9028 UE = Addr.getNode()->use_end(); UI != UE; ++UI) { 9029 SDNode *User = *UI; 9030 if (User->getOpcode() != ISD::ADD || 9031 UI.getUse().getResNo() != Addr.getResNo()) 9032 continue; 9033 9034 // Check that the add is independent of the load/store. Otherwise, folding 9035 // it would create a cycle. 9036 if (User->isPredecessorOf(N) || N->isPredecessorOf(User)) 9037 continue; 9038 9039 // Find the new opcode for the updating load/store. 9040 bool isLoad = true; 9041 bool isLaneOp = false; 9042 unsigned NewOpc = 0; 9043 unsigned NumVecs = 0; 9044 if (isIntrinsic) { 9045 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); 9046 switch (IntNo) { 9047 default: llvm_unreachable("unexpected intrinsic for Neon base update"); 9048 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD; 9049 NumVecs = 1; break; 9050 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD; 9051 NumVecs = 2; break; 9052 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD; 9053 NumVecs = 3; break; 9054 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD; 9055 NumVecs = 4; break; 9056 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD; 9057 NumVecs = 2; isLaneOp = true; break; 9058 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD; 9059 NumVecs = 3; isLaneOp = true; break; 9060 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD; 9061 NumVecs = 4; isLaneOp = true; break; 9062 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD; 9063 NumVecs = 1; isLoad = false; break; 9064 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD; 9065 NumVecs = 2; isLoad = false; break; 9066 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD; 9067 NumVecs = 3; isLoad = false; break; 9068 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD; 9069 NumVecs = 4; isLoad = false; break; 9070 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD; 9071 NumVecs = 2; isLoad = false; isLaneOp = true; break; 9072 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD; 9073 NumVecs = 3; isLoad = false; isLaneOp = true; break; 9074 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD; 9075 NumVecs = 4; isLoad = false; isLaneOp = true; break; 9076 } 9077 } else { 9078 isLaneOp = true; 9079 switch (N->getOpcode()) { 9080 default: llvm_unreachable("unexpected opcode for Neon base update"); 9081 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break; 9082 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break; 9083 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break; 9084 } 9085 } 9086 9087 // Find the size of memory referenced by the load/store. 9088 EVT VecTy; 9089 if (isLoad) 9090 VecTy = N->getValueType(0); 9091 else 9092 VecTy = N->getOperand(AddrOpIdx+1).getValueType(); 9093 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8; 9094 if (isLaneOp) 9095 NumBytes /= VecTy.getVectorNumElements(); 9096 9097 // If the increment is a constant, it must match the memory ref size. 9098 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0); 9099 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) { 9100 uint64_t IncVal = CInc->getZExtValue(); 9101 if (IncVal != NumBytes) 9102 continue; 9103 } else if (NumBytes >= 3 * 16) { 9104 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two 9105 // separate instructions that make it harder to use a non-constant update. 9106 continue; 9107 } 9108 9109 // Create the new updating load/store node. 9110 EVT Tys[6]; 9111 unsigned NumResultVecs = (isLoad ? NumVecs : 0); 9112 unsigned n; 9113 for (n = 0; n < NumResultVecs; ++n) 9114 Tys[n] = VecTy; 9115 Tys[n++] = MVT::i32; 9116 Tys[n] = MVT::Other; 9117 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2); 9118 SmallVector<SDValue, 8> Ops; 9119 Ops.push_back(N->getOperand(0)); // incoming chain 9120 Ops.push_back(N->getOperand(AddrOpIdx)); 9121 Ops.push_back(Inc); 9122 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) { 9123 Ops.push_back(N->getOperand(i)); 9124 } 9125 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N); 9126 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, 9127 Ops.data(), Ops.size(), 9128 MemInt->getMemoryVT(), 9129 MemInt->getMemOperand()); 9130 9131 // Update the uses. 9132 std::vector<SDValue> NewResults; 9133 for (unsigned i = 0; i < NumResultVecs; ++i) { 9134 NewResults.push_back(SDValue(UpdN.getNode(), i)); 9135 } 9136 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain 9137 DCI.CombineTo(N, NewResults); 9138 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs)); 9139 9140 break; 9141 } 9142 return SDValue(); 9143 } 9144 9145 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a 9146 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic 9147 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and 9148 /// return true. 9149 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { 9150 SelectionDAG &DAG = DCI.DAG; 9151 EVT VT = N->getValueType(0); 9152 // vldN-dup instructions only support 64-bit vectors for N > 1. 9153 if (!VT.is64BitVector()) 9154 return false; 9155 9156 // Check if the VDUPLANE operand is a vldN-dup intrinsic. 9157 SDNode *VLD = N->getOperand(0).getNode(); 9158 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN) 9159 return false; 9160 unsigned NumVecs = 0; 9161 unsigned NewOpc = 0; 9162 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue(); 9163 if (IntNo == Intrinsic::arm_neon_vld2lane) { 9164 NumVecs = 2; 9165 NewOpc = ARMISD::VLD2DUP; 9166 } else if (IntNo == Intrinsic::arm_neon_vld3lane) { 9167 NumVecs = 3; 9168 NewOpc = ARMISD::VLD3DUP; 9169 } else if (IntNo == Intrinsic::arm_neon_vld4lane) { 9170 NumVecs = 4; 9171 NewOpc = ARMISD::VLD4DUP; 9172 } else { 9173 return false; 9174 } 9175 9176 // First check that all the vldN-lane uses are VDUPLANEs and that the lane 9177 // numbers match the load. 9178 unsigned VLDLaneNo = 9179 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue(); 9180 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end(); 9181 UI != UE; ++UI) { 9182 // Ignore uses of the chain result. 9183 if (UI.getUse().getResNo() == NumVecs) 9184 continue; 9185 SDNode *User = *UI; 9186 if (User->getOpcode() != ARMISD::VDUPLANE || 9187 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue()) 9188 return false; 9189 } 9190 9191 // Create the vldN-dup node. 9192 EVT Tys[5]; 9193 unsigned n; 9194 for (n = 0; n < NumVecs; ++n) 9195 Tys[n] = VT; 9196 Tys[n] = MVT::Other; 9197 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1); 9198 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) }; 9199 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD); 9200 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys, 9201 Ops, 2, VLDMemInt->getMemoryVT(), 9202 VLDMemInt->getMemOperand()); 9203 9204 // Update the uses. 9205 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end(); 9206 UI != UE; ++UI) { 9207 unsigned ResNo = UI.getUse().getResNo(); 9208 // Ignore uses of the chain result. 9209 if (ResNo == NumVecs) 9210 continue; 9211 SDNode *User = *UI; 9212 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo)); 9213 } 9214 9215 // Now the vldN-lane intrinsic is dead except for its chain result. 9216 // Update uses of the chain. 9217 std::vector<SDValue> VLDDupResults; 9218 for (unsigned n = 0; n < NumVecs; ++n) 9219 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n)); 9220 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs)); 9221 DCI.CombineTo(VLD, VLDDupResults); 9222 9223 return true; 9224 } 9225 9226 /// PerformVDUPLANECombine - Target-specific dag combine xforms for 9227 /// ARMISD::VDUPLANE. 9228 static SDValue PerformVDUPLANECombine(SDNode *N, 9229 TargetLowering::DAGCombinerInfo &DCI) { 9230 SDValue Op = N->getOperand(0); 9231 9232 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses 9233 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation. 9234 if (CombineVLDDUP(N, DCI)) 9235 return SDValue(N, 0); 9236 9237 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is 9238 // redundant. Ignore bit_converts for now; element sizes are checked below. 9239 while (Op.getOpcode() == ISD::BITCAST) 9240 Op = Op.getOperand(0); 9241 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM) 9242 return SDValue(); 9243 9244 // Make sure the VMOV element size is not bigger than the VDUPLANE elements. 9245 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits(); 9246 // The canonical VMOV for a zero vector uses a 32-bit element size. 9247 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 9248 unsigned EltBits; 9249 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0) 9250 EltSize = 8; 9251 EVT VT = N->getValueType(0); 9252 if (EltSize > VT.getVectorElementType().getSizeInBits()) 9253 return SDValue(); 9254 9255 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op); 9256 } 9257 9258 // isConstVecPow2 - Return true if each vector element is a power of 2, all 9259 // elements are the same constant, C, and Log2(C) ranges from 1 to 32. 9260 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C) 9261 { 9262 integerPart cN; 9263 integerPart c0 = 0; 9264 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements(); 9265 I != E; I++) { 9266 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I)); 9267 if (!C) 9268 return false; 9269 9270 bool isExact; 9271 APFloat APF = C->getValueAPF(); 9272 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact) 9273 != APFloat::opOK || !isExact) 9274 return false; 9275 9276 c0 = (I == 0) ? cN : c0; 9277 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32) 9278 return false; 9279 } 9280 C = c0; 9281 return true; 9282 } 9283 9284 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD) 9285 /// can replace combinations of VMUL and VCVT (floating-point to integer) 9286 /// when the VMUL has a constant operand that is a power of 2. 9287 /// 9288 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>): 9289 /// vmul.f32 d16, d17, d16 9290 /// vcvt.s32.f32 d16, d16 9291 /// becomes: 9292 /// vcvt.s32.f32 d16, d16, #3 9293 static SDValue PerformVCVTCombine(SDNode *N, 9294 TargetLowering::DAGCombinerInfo &DCI, 9295 const ARMSubtarget *Subtarget) { 9296 SelectionDAG &DAG = DCI.DAG; 9297 SDValue Op = N->getOperand(0); 9298 9299 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() || 9300 Op.getOpcode() != ISD::FMUL) 9301 return SDValue(); 9302 9303 uint64_t C; 9304 SDValue N0 = Op->getOperand(0); 9305 SDValue ConstVec = Op->getOperand(1); 9306 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT; 9307 9308 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR || 9309 !isConstVecPow2(ConstVec, isSigned, C)) 9310 return SDValue(); 9311 9312 MVT FloatTy = Op.getSimpleValueType().getVectorElementType(); 9313 MVT IntTy = N->getSimpleValueType(0).getVectorElementType(); 9314 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) { 9315 // These instructions only exist converting from f32 to i32. We can handle 9316 // smaller integers by generating an extra truncate, but larger ones would 9317 // be lossy. 9318 return SDValue(); 9319 } 9320 9321 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs : 9322 Intrinsic::arm_neon_vcvtfp2fxu; 9323 unsigned NumLanes = Op.getValueType().getVectorNumElements(); 9324 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), 9325 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32, 9326 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0, 9327 DAG.getConstant(Log2_64(C), MVT::i32)); 9328 9329 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits()) 9330 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv); 9331 9332 return FixConv; 9333 } 9334 9335 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD) 9336 /// can replace combinations of VCVT (integer to floating-point) and VDIV 9337 /// when the VDIV has a constant operand that is a power of 2. 9338 /// 9339 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>): 9340 /// vcvt.f32.s32 d16, d16 9341 /// vdiv.f32 d16, d17, d16 9342 /// becomes: 9343 /// vcvt.f32.s32 d16, d16, #3 9344 static SDValue PerformVDIVCombine(SDNode *N, 9345 TargetLowering::DAGCombinerInfo &DCI, 9346 const ARMSubtarget *Subtarget) { 9347 SelectionDAG &DAG = DCI.DAG; 9348 SDValue Op = N->getOperand(0); 9349 unsigned OpOpcode = Op.getNode()->getOpcode(); 9350 9351 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() || 9352 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP)) 9353 return SDValue(); 9354 9355 uint64_t C; 9356 SDValue ConstVec = N->getOperand(1); 9357 bool isSigned = OpOpcode == ISD::SINT_TO_FP; 9358 9359 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR || 9360 !isConstVecPow2(ConstVec, isSigned, C)) 9361 return SDValue(); 9362 9363 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType(); 9364 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType(); 9365 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) { 9366 // These instructions only exist converting from i32 to f32. We can handle 9367 // smaller integers by generating an extra extend, but larger ones would 9368 // be lossy. 9369 return SDValue(); 9370 } 9371 9372 SDValue ConvInput = Op.getOperand(0); 9373 unsigned NumLanes = Op.getValueType().getVectorNumElements(); 9374 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits()) 9375 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, 9376 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32, 9377 ConvInput); 9378 9379 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp : 9380 Intrinsic::arm_neon_vcvtfxu2fp; 9381 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), 9382 Op.getValueType(), 9383 DAG.getConstant(IntrinsicOpcode, MVT::i32), 9384 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32)); 9385 } 9386 9387 /// Getvshiftimm - Check if this is a valid build_vector for the immediate 9388 /// operand of a vector shift operation, where all the elements of the 9389 /// build_vector must have the same constant integer value. 9390 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) { 9391 // Ignore bit_converts. 9392 while (Op.getOpcode() == ISD::BITCAST) 9393 Op = Op.getOperand(0); 9394 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode()); 9395 APInt SplatBits, SplatUndef; 9396 unsigned SplatBitSize; 9397 bool HasAnyUndefs; 9398 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, 9399 HasAnyUndefs, ElementBits) || 9400 SplatBitSize > ElementBits) 9401 return false; 9402 Cnt = SplatBits.getSExtValue(); 9403 return true; 9404 } 9405 9406 /// isVShiftLImm - Check if this is a valid build_vector for the immediate 9407 /// operand of a vector shift left operation. That value must be in the range: 9408 /// 0 <= Value < ElementBits for a left shift; or 9409 /// 0 <= Value <= ElementBits for a long left shift. 9410 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) { 9411 assert(VT.isVector() && "vector shift count is not a vector type"); 9412 unsigned ElementBits = VT.getVectorElementType().getSizeInBits(); 9413 if (! getVShiftImm(Op, ElementBits, Cnt)) 9414 return false; 9415 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits); 9416 } 9417 9418 /// isVShiftRImm - Check if this is a valid build_vector for the immediate 9419 /// operand of a vector shift right operation. For a shift opcode, the value 9420 /// is positive, but for an intrinsic the value count must be negative. The 9421 /// absolute value must be in the range: 9422 /// 1 <= |Value| <= ElementBits for a right shift; or 9423 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift. 9424 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic, 9425 int64_t &Cnt) { 9426 assert(VT.isVector() && "vector shift count is not a vector type"); 9427 unsigned ElementBits = VT.getVectorElementType().getSizeInBits(); 9428 if (! getVShiftImm(Op, ElementBits, Cnt)) 9429 return false; 9430 if (isIntrinsic) 9431 Cnt = -Cnt; 9432 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits)); 9433 } 9434 9435 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics. 9436 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) { 9437 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); 9438 switch (IntNo) { 9439 default: 9440 // Don't do anything for most intrinsics. 9441 break; 9442 9443 // Vector shifts: check for immediate versions and lower them. 9444 // Note: This is done during DAG combining instead of DAG legalizing because 9445 // the build_vectors for 64-bit vector element shift counts are generally 9446 // not legal, and it is hard to see their values after they get legalized to 9447 // loads from a constant pool. 9448 case Intrinsic::arm_neon_vshifts: 9449 case Intrinsic::arm_neon_vshiftu: 9450 case Intrinsic::arm_neon_vshiftls: 9451 case Intrinsic::arm_neon_vshiftlu: 9452 case Intrinsic::arm_neon_vshiftn: 9453 case Intrinsic::arm_neon_vrshifts: 9454 case Intrinsic::arm_neon_vrshiftu: 9455 case Intrinsic::arm_neon_vrshiftn: 9456 case Intrinsic::arm_neon_vqshifts: 9457 case Intrinsic::arm_neon_vqshiftu: 9458 case Intrinsic::arm_neon_vqshiftsu: 9459 case Intrinsic::arm_neon_vqshiftns: 9460 case Intrinsic::arm_neon_vqshiftnu: 9461 case Intrinsic::arm_neon_vqshiftnsu: 9462 case Intrinsic::arm_neon_vqrshiftns: 9463 case Intrinsic::arm_neon_vqrshiftnu: 9464 case Intrinsic::arm_neon_vqrshiftnsu: { 9465 EVT VT = N->getOperand(1).getValueType(); 9466 int64_t Cnt; 9467 unsigned VShiftOpc = 0; 9468 9469 switch (IntNo) { 9470 case Intrinsic::arm_neon_vshifts: 9471 case Intrinsic::arm_neon_vshiftu: 9472 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) { 9473 VShiftOpc = ARMISD::VSHL; 9474 break; 9475 } 9476 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) { 9477 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ? 9478 ARMISD::VSHRs : ARMISD::VSHRu); 9479 break; 9480 } 9481 return SDValue(); 9482 9483 case Intrinsic::arm_neon_vshiftls: 9484 case Intrinsic::arm_neon_vshiftlu: 9485 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt)) 9486 break; 9487 llvm_unreachable("invalid shift count for vshll intrinsic"); 9488 9489 case Intrinsic::arm_neon_vrshifts: 9490 case Intrinsic::arm_neon_vrshiftu: 9491 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) 9492 break; 9493 return SDValue(); 9494 9495 case Intrinsic::arm_neon_vqshifts: 9496 case Intrinsic::arm_neon_vqshiftu: 9497 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) 9498 break; 9499 return SDValue(); 9500 9501 case Intrinsic::arm_neon_vqshiftsu: 9502 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) 9503 break; 9504 llvm_unreachable("invalid shift count for vqshlu intrinsic"); 9505 9506 case Intrinsic::arm_neon_vshiftn: 9507 case Intrinsic::arm_neon_vrshiftn: 9508 case Intrinsic::arm_neon_vqshiftns: 9509 case Intrinsic::arm_neon_vqshiftnu: 9510 case Intrinsic::arm_neon_vqshiftnsu: 9511 case Intrinsic::arm_neon_vqrshiftns: 9512 case Intrinsic::arm_neon_vqrshiftnu: 9513 case Intrinsic::arm_neon_vqrshiftnsu: 9514 // Narrowing shifts require an immediate right shift. 9515 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt)) 9516 break; 9517 llvm_unreachable("invalid shift count for narrowing vector shift " 9518 "intrinsic"); 9519 9520 default: 9521 llvm_unreachable("unhandled vector shift"); 9522 } 9523 9524 switch (IntNo) { 9525 case Intrinsic::arm_neon_vshifts: 9526 case Intrinsic::arm_neon_vshiftu: 9527 // Opcode already set above. 9528 break; 9529 case Intrinsic::arm_neon_vshiftls: 9530 case Intrinsic::arm_neon_vshiftlu: 9531 if (Cnt == VT.getVectorElementType().getSizeInBits()) 9532 VShiftOpc = ARMISD::VSHLLi; 9533 else 9534 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ? 9535 ARMISD::VSHLLs : ARMISD::VSHLLu); 9536 break; 9537 case Intrinsic::arm_neon_vshiftn: 9538 VShiftOpc = ARMISD::VSHRN; break; 9539 case Intrinsic::arm_neon_vrshifts: 9540 VShiftOpc = ARMISD::VRSHRs; break; 9541 case Intrinsic::arm_neon_vrshiftu: 9542 VShiftOpc = ARMISD::VRSHRu; break; 9543 case Intrinsic::arm_neon_vrshiftn: 9544 VShiftOpc = ARMISD::VRSHRN; break; 9545 case Intrinsic::arm_neon_vqshifts: 9546 VShiftOpc = ARMISD::VQSHLs; break; 9547 case Intrinsic::arm_neon_vqshiftu: 9548 VShiftOpc = ARMISD::VQSHLu; break; 9549 case Intrinsic::arm_neon_vqshiftsu: 9550 VShiftOpc = ARMISD::VQSHLsu; break; 9551 case Intrinsic::arm_neon_vqshiftns: 9552 VShiftOpc = ARMISD::VQSHRNs; break; 9553 case Intrinsic::arm_neon_vqshiftnu: 9554 VShiftOpc = ARMISD::VQSHRNu; break; 9555 case Intrinsic::arm_neon_vqshiftnsu: 9556 VShiftOpc = ARMISD::VQSHRNsu; break; 9557 case Intrinsic::arm_neon_vqrshiftns: 9558 VShiftOpc = ARMISD::VQRSHRNs; break; 9559 case Intrinsic::arm_neon_vqrshiftnu: 9560 VShiftOpc = ARMISD::VQRSHRNu; break; 9561 case Intrinsic::arm_neon_vqrshiftnsu: 9562 VShiftOpc = ARMISD::VQRSHRNsu; break; 9563 } 9564 9565 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0), 9566 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32)); 9567 } 9568 9569 case Intrinsic::arm_neon_vshiftins: { 9570 EVT VT = N->getOperand(1).getValueType(); 9571 int64_t Cnt; 9572 unsigned VShiftOpc = 0; 9573 9574 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt)) 9575 VShiftOpc = ARMISD::VSLI; 9576 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt)) 9577 VShiftOpc = ARMISD::VSRI; 9578 else { 9579 llvm_unreachable("invalid shift count for vsli/vsri intrinsic"); 9580 } 9581 9582 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0), 9583 N->getOperand(1), N->getOperand(2), 9584 DAG.getConstant(Cnt, MVT::i32)); 9585 } 9586 9587 case Intrinsic::arm_neon_vqrshifts: 9588 case Intrinsic::arm_neon_vqrshiftu: 9589 // No immediate versions of these to check for. 9590 break; 9591 } 9592 9593 return SDValue(); 9594 } 9595 9596 /// PerformShiftCombine - Checks for immediate versions of vector shifts and 9597 /// lowers them. As with the vector shift intrinsics, this is done during DAG 9598 /// combining instead of DAG legalizing because the build_vectors for 64-bit 9599 /// vector element shift counts are generally not legal, and it is hard to see 9600 /// their values after they get legalized to loads from a constant pool. 9601 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG, 9602 const ARMSubtarget *ST) { 9603 EVT VT = N->getValueType(0); 9604 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) { 9605 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high 9606 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16. 9607 SDValue N1 = N->getOperand(1); 9608 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) { 9609 SDValue N0 = N->getOperand(0); 9610 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP && 9611 DAG.MaskedValueIsZero(N0.getOperand(0), 9612 APInt::getHighBitsSet(32, 16))) 9613 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1); 9614 } 9615 } 9616 9617 // Nothing to be done for scalar shifts. 9618 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 9619 if (!VT.isVector() || !TLI.isTypeLegal(VT)) 9620 return SDValue(); 9621 9622 assert(ST->hasNEON() && "unexpected vector shift"); 9623 int64_t Cnt; 9624 9625 switch (N->getOpcode()) { 9626 default: llvm_unreachable("unexpected shift opcode"); 9627 9628 case ISD::SHL: 9629 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) 9630 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0), 9631 DAG.getConstant(Cnt, MVT::i32)); 9632 break; 9633 9634 case ISD::SRA: 9635 case ISD::SRL: 9636 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) { 9637 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ? 9638 ARMISD::VSHRs : ARMISD::VSHRu); 9639 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0), 9640 DAG.getConstant(Cnt, MVT::i32)); 9641 } 9642 } 9643 return SDValue(); 9644 } 9645 9646 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND, 9647 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND. 9648 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG, 9649 const ARMSubtarget *ST) { 9650 SDValue N0 = N->getOperand(0); 9651 9652 // Check for sign- and zero-extensions of vector extract operations of 8- 9653 // and 16-bit vector elements. NEON supports these directly. They are 9654 // handled during DAG combining because type legalization will promote them 9655 // to 32-bit types and it is messy to recognize the operations after that. 9656 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) { 9657 SDValue Vec = N0.getOperand(0); 9658 SDValue Lane = N0.getOperand(1); 9659 EVT VT = N->getValueType(0); 9660 EVT EltVT = N0.getValueType(); 9661 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 9662 9663 if (VT == MVT::i32 && 9664 (EltVT == MVT::i8 || EltVT == MVT::i16) && 9665 TLI.isTypeLegal(Vec.getValueType()) && 9666 isa<ConstantSDNode>(Lane)) { 9667 9668 unsigned Opc = 0; 9669 switch (N->getOpcode()) { 9670 default: llvm_unreachable("unexpected opcode"); 9671 case ISD::SIGN_EXTEND: 9672 Opc = ARMISD::VGETLANEs; 9673 break; 9674 case ISD::ZERO_EXTEND: 9675 case ISD::ANY_EXTEND: 9676 Opc = ARMISD::VGETLANEu; 9677 break; 9678 } 9679 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane); 9680 } 9681 } 9682 9683 return SDValue(); 9684 } 9685 9686 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC 9687 /// to match f32 max/min patterns to use NEON vmax/vmin instructions. 9688 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG, 9689 const ARMSubtarget *ST) { 9690 // If the target supports NEON, try to use vmax/vmin instructions for f32 9691 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set, 9692 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is 9693 // a NaN; only do the transformation when it matches that behavior. 9694 9695 // For now only do this when using NEON for FP operations; if using VFP, it 9696 // is not obvious that the benefit outweighs the cost of switching to the 9697 // NEON pipeline. 9698 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() || 9699 N->getValueType(0) != MVT::f32) 9700 return SDValue(); 9701 9702 SDValue CondLHS = N->getOperand(0); 9703 SDValue CondRHS = N->getOperand(1); 9704 SDValue LHS = N->getOperand(2); 9705 SDValue RHS = N->getOperand(3); 9706 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get(); 9707 9708 unsigned Opcode = 0; 9709 bool IsReversed; 9710 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) { 9711 IsReversed = false; // x CC y ? x : y 9712 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) { 9713 IsReversed = true ; // x CC y ? y : x 9714 } else { 9715 return SDValue(); 9716 } 9717 9718 bool IsUnordered; 9719 switch (CC) { 9720 default: break; 9721 case ISD::SETOLT: 9722 case ISD::SETOLE: 9723 case ISD::SETLT: 9724 case ISD::SETLE: 9725 case ISD::SETULT: 9726 case ISD::SETULE: 9727 // If LHS is NaN, an ordered comparison will be false and the result will 9728 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS 9729 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN. 9730 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE); 9731 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS)) 9732 break; 9733 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin 9734 // will return -0, so vmin can only be used for unsafe math or if one of 9735 // the operands is known to be nonzero. 9736 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) && 9737 !DAG.getTarget().Options.UnsafeFPMath && 9738 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) 9739 break; 9740 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN; 9741 break; 9742 9743 case ISD::SETOGT: 9744 case ISD::SETOGE: 9745 case ISD::SETGT: 9746 case ISD::SETGE: 9747 case ISD::SETUGT: 9748 case ISD::SETUGE: 9749 // If LHS is NaN, an ordered comparison will be false and the result will 9750 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS 9751 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN. 9752 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE); 9753 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS)) 9754 break; 9755 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax 9756 // will return +0, so vmax can only be used for unsafe math or if one of 9757 // the operands is known to be nonzero. 9758 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) && 9759 !DAG.getTarget().Options.UnsafeFPMath && 9760 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) 9761 break; 9762 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX; 9763 break; 9764 } 9765 9766 if (!Opcode) 9767 return SDValue(); 9768 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS); 9769 } 9770 9771 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV. 9772 SDValue 9773 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const { 9774 SDValue Cmp = N->getOperand(4); 9775 if (Cmp.getOpcode() != ARMISD::CMPZ) 9776 // Only looking at EQ and NE cases. 9777 return SDValue(); 9778 9779 EVT VT = N->getValueType(0); 9780 SDLoc dl(N); 9781 SDValue LHS = Cmp.getOperand(0); 9782 SDValue RHS = Cmp.getOperand(1); 9783 SDValue FalseVal = N->getOperand(0); 9784 SDValue TrueVal = N->getOperand(1); 9785 SDValue ARMcc = N->getOperand(2); 9786 ARMCC::CondCodes CC = 9787 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue(); 9788 9789 // Simplify 9790 // mov r1, r0 9791 // cmp r1, x 9792 // mov r0, y 9793 // moveq r0, x 9794 // to 9795 // cmp r0, x 9796 // movne r0, y 9797 // 9798 // mov r1, r0 9799 // cmp r1, x 9800 // mov r0, x 9801 // movne r0, y 9802 // to 9803 // cmp r0, x 9804 // movne r0, y 9805 /// FIXME: Turn this into a target neutral optimization? 9806 SDValue Res; 9807 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) { 9808 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc, 9809 N->getOperand(3), Cmp); 9810 } else if (CC == ARMCC::EQ && TrueVal == RHS) { 9811 SDValue ARMcc; 9812 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl); 9813 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc, 9814 N->getOperand(3), NewCmp); 9815 } 9816 9817 if (Res.getNode()) { 9818 APInt KnownZero, KnownOne; 9819 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne); 9820 // Capture demanded bits information that would be otherwise lost. 9821 if (KnownZero == 0xfffffffe) 9822 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res, 9823 DAG.getValueType(MVT::i1)); 9824 else if (KnownZero == 0xffffff00) 9825 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res, 9826 DAG.getValueType(MVT::i8)); 9827 else if (KnownZero == 0xffff0000) 9828 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res, 9829 DAG.getValueType(MVT::i16)); 9830 } 9831 9832 return Res; 9833 } 9834 9835 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N, 9836 DAGCombinerInfo &DCI) const { 9837 switch (N->getOpcode()) { 9838 default: break; 9839 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget); 9840 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget); 9841 case ISD::SUB: return PerformSUBCombine(N, DCI); 9842 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget); 9843 case ISD::OR: return PerformORCombine(N, DCI, Subtarget); 9844 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget); 9845 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget); 9846 case ARMISD::BFI: return PerformBFICombine(N, DCI); 9847 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI); 9848 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG); 9849 case ISD::STORE: return PerformSTORECombine(N, DCI); 9850 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI); 9851 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI); 9852 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG); 9853 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI); 9854 case ISD::FP_TO_SINT: 9855 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget); 9856 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget); 9857 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG); 9858 case ISD::SHL: 9859 case ISD::SRA: 9860 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget); 9861 case ISD::SIGN_EXTEND: 9862 case ISD::ZERO_EXTEND: 9863 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget); 9864 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget); 9865 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG); 9866 case ARMISD::VLD2DUP: 9867 case ARMISD::VLD3DUP: 9868 case ARMISD::VLD4DUP: 9869 return CombineBaseUpdate(N, DCI); 9870 case ARMISD::BUILD_VECTOR: 9871 return PerformARMBUILD_VECTORCombine(N, DCI); 9872 case ISD::INTRINSIC_VOID: 9873 case ISD::INTRINSIC_W_CHAIN: 9874 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) { 9875 case Intrinsic::arm_neon_vld1: 9876 case Intrinsic::arm_neon_vld2: 9877 case Intrinsic::arm_neon_vld3: 9878 case Intrinsic::arm_neon_vld4: 9879 case Intrinsic::arm_neon_vld2lane: 9880 case Intrinsic::arm_neon_vld3lane: 9881 case Intrinsic::arm_neon_vld4lane: 9882 case Intrinsic::arm_neon_vst1: 9883 case Intrinsic::arm_neon_vst2: 9884 case Intrinsic::arm_neon_vst3: 9885 case Intrinsic::arm_neon_vst4: 9886 case Intrinsic::arm_neon_vst2lane: 9887 case Intrinsic::arm_neon_vst3lane: 9888 case Intrinsic::arm_neon_vst4lane: 9889 return CombineBaseUpdate(N, DCI); 9890 default: break; 9891 } 9892 break; 9893 } 9894 return SDValue(); 9895 } 9896 9897 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc, 9898 EVT VT) const { 9899 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE); 9900 } 9901 9902 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const { 9903 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus 9904 bool AllowsUnaligned = Subtarget->allowsUnalignedMem(); 9905 9906 switch (VT.getSimpleVT().SimpleTy) { 9907 default: 9908 return false; 9909 case MVT::i8: 9910 case MVT::i16: 9911 case MVT::i32: { 9912 // Unaligned access can use (for example) LRDB, LRDH, LDR 9913 if (AllowsUnaligned) { 9914 if (Fast) 9915 *Fast = Subtarget->hasV7Ops(); 9916 return true; 9917 } 9918 return false; 9919 } 9920 case MVT::f64: 9921 case MVT::v2f64: { 9922 // For any little-endian targets with neon, we can support unaligned ld/st 9923 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8. 9924 // A big-endian target may also explictly support unaligned accesses 9925 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) { 9926 if (Fast) 9927 *Fast = true; 9928 return true; 9929 } 9930 return false; 9931 } 9932 } 9933 } 9934 9935 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign, 9936 unsigned AlignCheck) { 9937 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) && 9938 (DstAlign == 0 || DstAlign % AlignCheck == 0)); 9939 } 9940 9941 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size, 9942 unsigned DstAlign, unsigned SrcAlign, 9943 bool IsMemset, bool ZeroMemset, 9944 bool MemcpyStrSrc, 9945 MachineFunction &MF) const { 9946 const Function *F = MF.getFunction(); 9947 9948 // See if we can use NEON instructions for this... 9949 if ((!IsMemset || ZeroMemset) && 9950 Subtarget->hasNEON() && 9951 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex, 9952 Attribute::NoImplicitFloat)) { 9953 bool Fast; 9954 if (Size >= 16 && 9955 (memOpAlign(SrcAlign, DstAlign, 16) || 9956 (allowsUnalignedMemoryAccesses(MVT::v2f64, &Fast) && Fast))) { 9957 return MVT::v2f64; 9958 } else if (Size >= 8 && 9959 (memOpAlign(SrcAlign, DstAlign, 8) || 9960 (allowsUnalignedMemoryAccesses(MVT::f64, &Fast) && Fast))) { 9961 return MVT::f64; 9962 } 9963 } 9964 9965 // Lowering to i32/i16 if the size permits. 9966 if (Size >= 4) 9967 return MVT::i32; 9968 else if (Size >= 2) 9969 return MVT::i16; 9970 9971 // Let the target-independent logic figure it out. 9972 return MVT::Other; 9973 } 9974 9975 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const { 9976 if (Val.getOpcode() != ISD::LOAD) 9977 return false; 9978 9979 EVT VT1 = Val.getValueType(); 9980 if (!VT1.isSimple() || !VT1.isInteger() || 9981 !VT2.isSimple() || !VT2.isInteger()) 9982 return false; 9983 9984 switch (VT1.getSimpleVT().SimpleTy) { 9985 default: break; 9986 case MVT::i1: 9987 case MVT::i8: 9988 case MVT::i16: 9989 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits. 9990 return true; 9991 } 9992 9993 return false; 9994 } 9995 9996 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const { 9997 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy()) 9998 return false; 9999 10000 if (!isTypeLegal(EVT::getEVT(Ty1))) 10001 return false; 10002 10003 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop"); 10004 10005 // Assuming the caller doesn't have a zeroext or signext return parameter, 10006 // truncation all the way down to i1 is valid. 10007 return true; 10008 } 10009 10010 10011 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) { 10012 if (V < 0) 10013 return false; 10014 10015 unsigned Scale = 1; 10016 switch (VT.getSimpleVT().SimpleTy) { 10017 default: return false; 10018 case MVT::i1: 10019 case MVT::i8: 10020 // Scale == 1; 10021 break; 10022 case MVT::i16: 10023 // Scale == 2; 10024 Scale = 2; 10025 break; 10026 case MVT::i32: 10027 // Scale == 4; 10028 Scale = 4; 10029 break; 10030 } 10031 10032 if ((V & (Scale - 1)) != 0) 10033 return false; 10034 V /= Scale; 10035 return V == (V & ((1LL << 5) - 1)); 10036 } 10037 10038 static bool isLegalT2AddressImmediate(int64_t V, EVT VT, 10039 const ARMSubtarget *Subtarget) { 10040 bool isNeg = false; 10041 if (V < 0) { 10042 isNeg = true; 10043 V = - V; 10044 } 10045 10046 switch (VT.getSimpleVT().SimpleTy) { 10047 default: return false; 10048 case MVT::i1: 10049 case MVT::i8: 10050 case MVT::i16: 10051 case MVT::i32: 10052 // + imm12 or - imm8 10053 if (isNeg) 10054 return V == (V & ((1LL << 8) - 1)); 10055 return V == (V & ((1LL << 12) - 1)); 10056 case MVT::f32: 10057 case MVT::f64: 10058 // Same as ARM mode. FIXME: NEON? 10059 if (!Subtarget->hasVFP2()) 10060 return false; 10061 if ((V & 3) != 0) 10062 return false; 10063 V >>= 2; 10064 return V == (V & ((1LL << 8) - 1)); 10065 } 10066 } 10067 10068 /// isLegalAddressImmediate - Return true if the integer value can be used 10069 /// as the offset of the target addressing mode for load / store of the 10070 /// given type. 10071 static bool isLegalAddressImmediate(int64_t V, EVT VT, 10072 const ARMSubtarget *Subtarget) { 10073 if (V == 0) 10074 return true; 10075 10076 if (!VT.isSimple()) 10077 return false; 10078 10079 if (Subtarget->isThumb1Only()) 10080 return isLegalT1AddressImmediate(V, VT); 10081 else if (Subtarget->isThumb2()) 10082 return isLegalT2AddressImmediate(V, VT, Subtarget); 10083 10084 // ARM mode. 10085 if (V < 0) 10086 V = - V; 10087 switch (VT.getSimpleVT().SimpleTy) { 10088 default: return false; 10089 case MVT::i1: 10090 case MVT::i8: 10091 case MVT::i32: 10092 // +- imm12 10093 return V == (V & ((1LL << 12) - 1)); 10094 case MVT::i16: 10095 // +- imm8 10096 return V == (V & ((1LL << 8) - 1)); 10097 case MVT::f32: 10098 case MVT::f64: 10099 if (!Subtarget->hasVFP2()) // FIXME: NEON? 10100 return false; 10101 if ((V & 3) != 0) 10102 return false; 10103 V >>= 2; 10104 return V == (V & ((1LL << 8) - 1)); 10105 } 10106 } 10107 10108 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM, 10109 EVT VT) const { 10110 int Scale = AM.Scale; 10111 if (Scale < 0) 10112 return false; 10113 10114 switch (VT.getSimpleVT().SimpleTy) { 10115 default: return false; 10116 case MVT::i1: 10117 case MVT::i8: 10118 case MVT::i16: 10119 case MVT::i32: 10120 if (Scale == 1) 10121 return true; 10122 // r + r << imm 10123 Scale = Scale & ~1; 10124 return Scale == 2 || Scale == 4 || Scale == 8; 10125 case MVT::i64: 10126 // r + r 10127 if (((unsigned)AM.HasBaseReg + Scale) <= 2) 10128 return true; 10129 return false; 10130 case MVT::isVoid: 10131 // Note, we allow "void" uses (basically, uses that aren't loads or 10132 // stores), because arm allows folding a scale into many arithmetic 10133 // operations. This should be made more precise and revisited later. 10134 10135 // Allow r << imm, but the imm has to be a multiple of two. 10136 if (Scale & 1) return false; 10137 return isPowerOf2_32(Scale); 10138 } 10139 } 10140 10141 /// isLegalAddressingMode - Return true if the addressing mode represented 10142 /// by AM is legal for this target, for a load/store of the specified type. 10143 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM, 10144 Type *Ty) const { 10145 EVT VT = getValueType(Ty, true); 10146 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget)) 10147 return false; 10148 10149 // Can never fold addr of global into load/store. 10150 if (AM.BaseGV) 10151 return false; 10152 10153 switch (AM.Scale) { 10154 case 0: // no scale reg, must be "r+i" or "r", or "i". 10155 break; 10156 case 1: 10157 if (Subtarget->isThumb1Only()) 10158 return false; 10159 // FALL THROUGH. 10160 default: 10161 // ARM doesn't support any R+R*scale+imm addr modes. 10162 if (AM.BaseOffs) 10163 return false; 10164 10165 if (!VT.isSimple()) 10166 return false; 10167 10168 if (Subtarget->isThumb2()) 10169 return isLegalT2ScaledAddressingMode(AM, VT); 10170 10171 int Scale = AM.Scale; 10172 switch (VT.getSimpleVT().SimpleTy) { 10173 default: return false; 10174 case MVT::i1: 10175 case MVT::i8: 10176 case MVT::i32: 10177 if (Scale < 0) Scale = -Scale; 10178 if (Scale == 1) 10179 return true; 10180 // r + r << imm 10181 return isPowerOf2_32(Scale & ~1); 10182 case MVT::i16: 10183 case MVT::i64: 10184 // r + r 10185 if (((unsigned)AM.HasBaseReg + Scale) <= 2) 10186 return true; 10187 return false; 10188 10189 case MVT::isVoid: 10190 // Note, we allow "void" uses (basically, uses that aren't loads or 10191 // stores), because arm allows folding a scale into many arithmetic 10192 // operations. This should be made more precise and revisited later. 10193 10194 // Allow r << imm, but the imm has to be a multiple of two. 10195 if (Scale & 1) return false; 10196 return isPowerOf2_32(Scale); 10197 } 10198 } 10199 return true; 10200 } 10201 10202 /// isLegalICmpImmediate - Return true if the specified immediate is legal 10203 /// icmp immediate, that is the target has icmp instructions which can compare 10204 /// a register against the immediate without having to materialize the 10205 /// immediate into a register. 10206 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const { 10207 // Thumb2 and ARM modes can use cmn for negative immediates. 10208 if (!Subtarget->isThumb()) 10209 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1; 10210 if (Subtarget->isThumb2()) 10211 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1; 10212 // Thumb1 doesn't have cmn, and only 8-bit immediates. 10213 return Imm >= 0 && Imm <= 255; 10214 } 10215 10216 /// isLegalAddImmediate - Return true if the specified immediate is a legal add 10217 /// *or sub* immediate, that is the target has add or sub instructions which can 10218 /// add a register with the immediate without having to materialize the 10219 /// immediate into a register. 10220 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const { 10221 // Same encoding for add/sub, just flip the sign. 10222 int64_t AbsImm = llvm::abs64(Imm); 10223 if (!Subtarget->isThumb()) 10224 return ARM_AM::getSOImmVal(AbsImm) != -1; 10225 if (Subtarget->isThumb2()) 10226 return ARM_AM::getT2SOImmVal(AbsImm) != -1; 10227 // Thumb1 only has 8-bit unsigned immediate. 10228 return AbsImm >= 0 && AbsImm <= 255; 10229 } 10230 10231 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT, 10232 bool isSEXTLoad, SDValue &Base, 10233 SDValue &Offset, bool &isInc, 10234 SelectionDAG &DAG) { 10235 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB) 10236 return false; 10237 10238 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) { 10239 // AddressingMode 3 10240 Base = Ptr->getOperand(0); 10241 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 10242 int RHSC = (int)RHS->getZExtValue(); 10243 if (RHSC < 0 && RHSC > -256) { 10244 assert(Ptr->getOpcode() == ISD::ADD); 10245 isInc = false; 10246 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 10247 return true; 10248 } 10249 } 10250 isInc = (Ptr->getOpcode() == ISD::ADD); 10251 Offset = Ptr->getOperand(1); 10252 return true; 10253 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) { 10254 // AddressingMode 2 10255 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 10256 int RHSC = (int)RHS->getZExtValue(); 10257 if (RHSC < 0 && RHSC > -0x1000) { 10258 assert(Ptr->getOpcode() == ISD::ADD); 10259 isInc = false; 10260 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 10261 Base = Ptr->getOperand(0); 10262 return true; 10263 } 10264 } 10265 10266 if (Ptr->getOpcode() == ISD::ADD) { 10267 isInc = true; 10268 ARM_AM::ShiftOpc ShOpcVal= 10269 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode()); 10270 if (ShOpcVal != ARM_AM::no_shift) { 10271 Base = Ptr->getOperand(1); 10272 Offset = Ptr->getOperand(0); 10273 } else { 10274 Base = Ptr->getOperand(0); 10275 Offset = Ptr->getOperand(1); 10276 } 10277 return true; 10278 } 10279 10280 isInc = (Ptr->getOpcode() == ISD::ADD); 10281 Base = Ptr->getOperand(0); 10282 Offset = Ptr->getOperand(1); 10283 return true; 10284 } 10285 10286 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store. 10287 return false; 10288 } 10289 10290 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT, 10291 bool isSEXTLoad, SDValue &Base, 10292 SDValue &Offset, bool &isInc, 10293 SelectionDAG &DAG) { 10294 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB) 10295 return false; 10296 10297 Base = Ptr->getOperand(0); 10298 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 10299 int RHSC = (int)RHS->getZExtValue(); 10300 if (RHSC < 0 && RHSC > -0x100) { // 8 bits. 10301 assert(Ptr->getOpcode() == ISD::ADD); 10302 isInc = false; 10303 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 10304 return true; 10305 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero. 10306 isInc = Ptr->getOpcode() == ISD::ADD; 10307 Offset = DAG.getConstant(RHSC, RHS->getValueType(0)); 10308 return true; 10309 } 10310 } 10311 10312 return false; 10313 } 10314 10315 /// getPreIndexedAddressParts - returns true by value, base pointer and 10316 /// offset pointer and addressing mode by reference if the node's address 10317 /// can be legally represented as pre-indexed load / store address. 10318 bool 10319 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base, 10320 SDValue &Offset, 10321 ISD::MemIndexedMode &AM, 10322 SelectionDAG &DAG) const { 10323 if (Subtarget->isThumb1Only()) 10324 return false; 10325 10326 EVT VT; 10327 SDValue Ptr; 10328 bool isSEXTLoad = false; 10329 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 10330 Ptr = LD->getBasePtr(); 10331 VT = LD->getMemoryVT(); 10332 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD; 10333 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 10334 Ptr = ST->getBasePtr(); 10335 VT = ST->getMemoryVT(); 10336 } else 10337 return false; 10338 10339 bool isInc; 10340 bool isLegal = false; 10341 if (Subtarget->isThumb2()) 10342 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base, 10343 Offset, isInc, DAG); 10344 else 10345 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base, 10346 Offset, isInc, DAG); 10347 if (!isLegal) 10348 return false; 10349 10350 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC; 10351 return true; 10352 } 10353 10354 /// getPostIndexedAddressParts - returns true by value, base pointer and 10355 /// offset pointer and addressing mode by reference if this node can be 10356 /// combined with a load / store to form a post-indexed load / store. 10357 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op, 10358 SDValue &Base, 10359 SDValue &Offset, 10360 ISD::MemIndexedMode &AM, 10361 SelectionDAG &DAG) const { 10362 if (Subtarget->isThumb1Only()) 10363 return false; 10364 10365 EVT VT; 10366 SDValue Ptr; 10367 bool isSEXTLoad = false; 10368 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 10369 VT = LD->getMemoryVT(); 10370 Ptr = LD->getBasePtr(); 10371 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD; 10372 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 10373 VT = ST->getMemoryVT(); 10374 Ptr = ST->getBasePtr(); 10375 } else 10376 return false; 10377 10378 bool isInc; 10379 bool isLegal = false; 10380 if (Subtarget->isThumb2()) 10381 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset, 10382 isInc, DAG); 10383 else 10384 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset, 10385 isInc, DAG); 10386 if (!isLegal) 10387 return false; 10388 10389 if (Ptr != Base) { 10390 // Swap base ptr and offset to catch more post-index load / store when 10391 // it's legal. In Thumb2 mode, offset must be an immediate. 10392 if (Ptr == Offset && Op->getOpcode() == ISD::ADD && 10393 !Subtarget->isThumb2()) 10394 std::swap(Base, Offset); 10395 10396 // Post-indexed load / store update the base pointer. 10397 if (Ptr != Base) 10398 return false; 10399 } 10400 10401 AM = isInc ? ISD::POST_INC : ISD::POST_DEC; 10402 return true; 10403 } 10404 10405 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 10406 APInt &KnownZero, 10407 APInt &KnownOne, 10408 const SelectionDAG &DAG, 10409 unsigned Depth) const { 10410 unsigned BitWidth = KnownOne.getBitWidth(); 10411 KnownZero = KnownOne = APInt(BitWidth, 0); 10412 switch (Op.getOpcode()) { 10413 default: break; 10414 case ARMISD::ADDC: 10415 case ARMISD::ADDE: 10416 case ARMISD::SUBC: 10417 case ARMISD::SUBE: 10418 // These nodes' second result is a boolean 10419 if (Op.getResNo() == 0) 10420 break; 10421 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1); 10422 break; 10423 case ARMISD::CMOV: { 10424 // Bits are known zero/one if known on the LHS and RHS. 10425 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1); 10426 if (KnownZero == 0 && KnownOne == 0) return; 10427 10428 APInt KnownZeroRHS, KnownOneRHS; 10429 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1); 10430 KnownZero &= KnownZeroRHS; 10431 KnownOne &= KnownOneRHS; 10432 return; 10433 } 10434 } 10435 } 10436 10437 //===----------------------------------