codegen_x86.h revision 69dfe51b684dd9d510dbcb63295fe180f998efde
1/* 2 * Copyright (C) 2011 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17#ifndef ART_COMPILER_DEX_QUICK_X86_CODEGEN_X86_H_ 18#define ART_COMPILER_DEX_QUICK_X86_CODEGEN_X86_H_ 19 20#include "dex/compiler_internals.h" 21#include "x86_lir.h" 22 23#include <map> 24 25namespace art { 26 27class X86Mir2Lir : public Mir2Lir { 28 protected: 29 class InToRegStorageMapper { 30 public: 31 virtual RegStorage GetNextReg(bool is_double_or_float, bool is_wide, bool is_ref) = 0; 32 virtual ~InToRegStorageMapper() {} 33 }; 34 35 class InToRegStorageX86_64Mapper : public InToRegStorageMapper { 36 public: 37 explicit InToRegStorageX86_64Mapper(Mir2Lir* ml) : ml_(ml), cur_core_reg_(0), cur_fp_reg_(0) {} 38 virtual ~InToRegStorageX86_64Mapper() {} 39 virtual RegStorage GetNextReg(bool is_double_or_float, bool is_wide, bool is_ref); 40 protected: 41 Mir2Lir* ml_; 42 private: 43 int cur_core_reg_; 44 int cur_fp_reg_; 45 }; 46 47 class InToRegStorageMapping { 48 public: 49 InToRegStorageMapping() : max_mapped_in_(0), is_there_stack_mapped_(false), 50 initialized_(false) {} 51 void Initialize(RegLocation* arg_locs, int count, InToRegStorageMapper* mapper); 52 int GetMaxMappedIn() { return max_mapped_in_; } 53 bool IsThereStackMapped() { return is_there_stack_mapped_; } 54 RegStorage Get(int in_position); 55 bool IsInitialized() { return initialized_; } 56 private: 57 std::map<int, RegStorage> mapping_; 58 int max_mapped_in_; 59 bool is_there_stack_mapped_; 60 bool initialized_; 61 }; 62 63 public: 64 X86Mir2Lir(CompilationUnit* cu, MIRGraph* mir_graph, ArenaAllocator* arena); 65 66 // Required for target - codegen helpers. 67 bool SmallLiteralDivRem(Instruction::Code dalvik_opcode, bool is_div, RegLocation rl_src, 68 RegLocation rl_dest, int lit); 69 bool EasyMultiply(RegLocation rl_src, RegLocation rl_dest, int lit) OVERRIDE; 70 LIR* CheckSuspendUsingLoad() OVERRIDE; 71 RegStorage LoadHelper(ThreadOffset<4> offset) OVERRIDE; 72 RegStorage LoadHelper(ThreadOffset<8> offset) OVERRIDE; 73 LIR* LoadBaseDisp(RegStorage r_base, int displacement, RegStorage r_dest, 74 OpSize size, VolatileKind is_volatile) OVERRIDE; 75 LIR* LoadBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_dest, int scale, 76 OpSize size) OVERRIDE; 77 LIR* LoadBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale, int displacement, 78 RegStorage r_dest, OpSize size) OVERRIDE; 79 LIR* LoadConstantNoClobber(RegStorage r_dest, int value); 80 LIR* LoadConstantWide(RegStorage r_dest, int64_t value); 81 LIR* StoreBaseDisp(RegStorage r_base, int displacement, RegStorage r_src, 82 OpSize size, VolatileKind is_volatile) OVERRIDE; 83 LIR* StoreBaseIndexed(RegStorage r_base, RegStorage r_index, RegStorage r_src, int scale, 84 OpSize size) OVERRIDE; 85 LIR* StoreBaseIndexedDisp(RegStorage r_base, RegStorage r_index, int scale, int displacement, 86 RegStorage r_src, OpSize size) OVERRIDE; 87 void MarkGCCard(RegStorage val_reg, RegStorage tgt_addr_reg); 88 void GenImplicitNullCheck(RegStorage reg, int opt_flags); 89 90 // Required for target - register utilities. 91 RegStorage TargetReg(SpecialTargetRegister reg) OVERRIDE; 92 RegStorage TargetReg32(SpecialTargetRegister reg); 93 RegStorage TargetReg(SpecialTargetRegister symbolic_reg, WideKind wide_kind) OVERRIDE { 94 if (wide_kind == kWide) { 95 if (cu_->target64) { 96 return As64BitReg(TargetReg32(symbolic_reg)); 97 } else { 98 // x86: construct a pair. 99 DCHECK((kArg0 <= symbolic_reg && symbolic_reg < kArg3) || 100 (kFArg0 <= symbolic_reg && symbolic_reg < kFArg3) || 101 (kRet0 == symbolic_reg)); 102 return RegStorage::MakeRegPair(TargetReg32(symbolic_reg), 103 TargetReg32(static_cast<SpecialTargetRegister>(symbolic_reg + 1))); 104 } 105 } else if (wide_kind == kRef && cu_->target64) { 106 return As64BitReg(TargetReg32(symbolic_reg)); 107 } else { 108 return TargetReg32(symbolic_reg); 109 } 110 } 111 RegStorage TargetPtrReg(SpecialTargetRegister symbolic_reg) OVERRIDE { 112 return TargetReg(symbolic_reg, cu_->target64 ? kWide : kNotWide); 113 } 114 RegStorage GetArgMappingToPhysicalReg(int arg_num); 115 RegStorage GetCoreArgMappingToPhysicalReg(int core_arg_num); 116 RegLocation GetReturnAlt(); 117 RegLocation GetReturnWideAlt(); 118 RegLocation LocCReturn(); 119 RegLocation LocCReturnRef(); 120 RegLocation LocCReturnDouble(); 121 RegLocation LocCReturnFloat(); 122 RegLocation LocCReturnWide(); 123 ResourceMask GetRegMaskCommon(const RegStorage& reg) const OVERRIDE; 124 void AdjustSpillMask(); 125 void ClobberCallerSave(); 126 void FreeCallTemps(); 127 void LockCallTemps(); 128 void CompilerInitializeRegAlloc(); 129 int VectorRegisterSize(); 130 int NumReservableVectorRegisters(bool fp_used); 131 132 // Required for target - miscellaneous. 133 void AssembleLIR(); 134 int AssignInsnOffsets(); 135 void AssignOffsets(); 136 AssemblerStatus AssembleInstructions(CodeOffset start_addr); 137 void DumpResourceMask(LIR* lir, const ResourceMask& mask, const char* prefix) OVERRIDE; 138 void SetupTargetResourceMasks(LIR* lir, uint64_t flags, 139 ResourceMask* use_mask, ResourceMask* def_mask) OVERRIDE; 140 const char* GetTargetInstFmt(int opcode); 141 const char* GetTargetInstName(int opcode); 142 std::string BuildInsnString(const char* fmt, LIR* lir, unsigned char* base_addr); 143 ResourceMask GetPCUseDefEncoding() const OVERRIDE; 144 uint64_t GetTargetInstFlags(int opcode); 145 size_t GetInsnSize(LIR* lir) OVERRIDE; 146 bool IsUnconditionalBranch(LIR* lir); 147 148 // Check support for volatile load/store of a given size. 149 bool SupportsVolatileLoadStore(OpSize size) OVERRIDE; 150 // Get the register class for load/store of a field. 151 RegisterClass RegClassForFieldLoadStore(OpSize size, bool is_volatile) OVERRIDE; 152 153 // Required for target - Dalvik-level generators. 154 void GenArithImmOpLong(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1, 155 RegLocation rl_src2); 156 void GenArrayGet(int opt_flags, OpSize size, RegLocation rl_array, RegLocation rl_index, 157 RegLocation rl_dest, int scale); 158 void GenArrayPut(int opt_flags, OpSize size, RegLocation rl_array, 159 RegLocation rl_index, RegLocation rl_src, int scale, bool card_mark); 160 void GenShiftImmOpLong(Instruction::Code opcode, RegLocation rl_dest, 161 RegLocation rl_src1, RegLocation rl_shift); 162 void GenMulLong(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1, 163 RegLocation rl_src2); 164 void GenAddLong(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1, 165 RegLocation rl_src2); 166 void GenAndLong(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1, 167 RegLocation rl_src2); 168 void GenArithOpDouble(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1, 169 RegLocation rl_src2); 170 void GenArithOpFloat(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1, 171 RegLocation rl_src2); 172 void GenRemFP(RegLocation rl_dest, RegLocation rl_src1, RegLocation rl_src2, bool is_double); 173 void GenCmpFP(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1, 174 RegLocation rl_src2); 175 void GenConversion(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src); 176 bool GenInlinedCas(CallInfo* info, bool is_long, bool is_object); 177 bool GenInlinedMinMax(CallInfo* info, bool is_min, bool is_long); 178 bool GenInlinedSqrt(CallInfo* info); 179 bool GenInlinedAbsFloat(CallInfo* info) OVERRIDE; 180 bool GenInlinedAbsDouble(CallInfo* info) OVERRIDE; 181 bool GenInlinedPeek(CallInfo* info, OpSize size); 182 bool GenInlinedPoke(CallInfo* info, OpSize size); 183 void GenNotLong(RegLocation rl_dest, RegLocation rl_src); 184 void GenNegLong(RegLocation rl_dest, RegLocation rl_src); 185 void GenOrLong(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1, 186 RegLocation rl_src2); 187 void GenSubLong(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1, 188 RegLocation rl_src2); 189 void GenXorLong(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1, 190 RegLocation rl_src2); 191 void GenDivRemLong(Instruction::Code, RegLocation rl_dest, RegLocation rl_src1, 192 RegLocation rl_src2, bool is_div); 193 // TODO: collapse reg_lo, reg_hi 194 RegLocation GenDivRem(RegLocation rl_dest, RegStorage reg_lo, RegStorage reg_hi, bool is_div); 195 RegLocation GenDivRemLit(RegLocation rl_dest, RegStorage reg_lo, int lit, bool is_div); 196 void GenCmpLong(RegLocation rl_dest, RegLocation rl_src1, RegLocation rl_src2); 197 void GenDivZeroCheckWide(RegStorage reg); 198 void GenArrayBoundsCheck(RegStorage index, RegStorage array_base, int32_t len_offset); 199 void GenArrayBoundsCheck(int32_t index, RegStorage array_base, int32_t len_offset); 200 void GenEntrySequence(RegLocation* ArgLocs, RegLocation rl_method); 201 void GenExitSequence(); 202 void GenSpecialExitSequence(); 203 void GenFillArrayData(DexOffset table_offset, RegLocation rl_src); 204 void GenFusedFPCmpBranch(BasicBlock* bb, MIR* mir, bool gt_bias, bool is_double); 205 void GenFusedLongCmpBranch(BasicBlock* bb, MIR* mir); 206 void GenSelect(BasicBlock* bb, MIR* mir); 207 bool GenMemBarrier(MemBarrierKind barrier_kind); 208 void GenMoveException(RegLocation rl_dest); 209 void GenMultiplyByTwoBitMultiplier(RegLocation rl_src, RegLocation rl_result, int lit, 210 int first_bit, int second_bit); 211 void GenNegDouble(RegLocation rl_dest, RegLocation rl_src); 212 void GenNegFloat(RegLocation rl_dest, RegLocation rl_src); 213 void GenPackedSwitch(MIR* mir, DexOffset table_offset, RegLocation rl_src); 214 void GenSparseSwitch(MIR* mir, DexOffset table_offset, RegLocation rl_src); 215 void GenIntToLong(RegLocation rl_dest, RegLocation rl_src); 216 217 /* 218 * @brief Generate a two address long operation with a constant value 219 * @param rl_dest location of result 220 * @param rl_src constant source operand 221 * @param op Opcode to be generated 222 * @return success or not 223 */ 224 bool GenLongImm(RegLocation rl_dest, RegLocation rl_src, Instruction::Code op); 225 /* 226 * @brief Generate a three address long operation with a constant value 227 * @param rl_dest location of result 228 * @param rl_src1 source operand 229 * @param rl_src2 constant source operand 230 * @param op Opcode to be generated 231 * @return success or not 232 */ 233 bool GenLongLongImm(RegLocation rl_dest, RegLocation rl_src1, RegLocation rl_src2, 234 Instruction::Code op); 235 236 /** 237 * @brief Generate a long arithmetic operation. 238 * @param rl_dest The destination. 239 * @param rl_src1 First operand. 240 * @param rl_src2 Second operand. 241 * @param op The DEX opcode for the operation. 242 * @param is_commutative The sources can be swapped if needed. 243 */ 244 virtual void GenLongArith(RegLocation rl_dest, RegLocation rl_src1, RegLocation rl_src2, 245 Instruction::Code op, bool is_commutative); 246 247 /** 248 * @brief Generate a two operand long arithmetic operation. 249 * @param rl_dest The destination. 250 * @param rl_src Second operand. 251 * @param op The DEX opcode for the operation. 252 */ 253 void GenLongArith(RegLocation rl_dest, RegLocation rl_src, Instruction::Code op); 254 255 /** 256 * @brief Generate a long operation. 257 * @param rl_dest The destination. Must be in a register 258 * @param rl_src The other operand. May be in a register or in memory. 259 * @param op The DEX opcode for the operation. 260 */ 261 virtual void GenLongRegOrMemOp(RegLocation rl_dest, RegLocation rl_src, Instruction::Code op); 262 263 /** 264 * @brief Implement instanceof a final class with x86 specific code. 265 * @param use_declaring_class 'true' if we can use the class itself. 266 * @param type_idx Type index to use if use_declaring_class is 'false'. 267 * @param rl_dest Result to be set to 0 or 1. 268 * @param rl_src Object to be tested. 269 */ 270 void GenInstanceofFinal(bool use_declaring_class, uint32_t type_idx, RegLocation rl_dest, 271 RegLocation rl_src); 272 /* 273 * 274 * @brief Implement Set up instanceof a class with x86 specific code. 275 * @param needs_access_check 'true' if we must check the access. 276 * @param type_known_final 'true' if the type is known to be a final class. 277 * @param type_known_abstract 'true' if the type is known to be an abstract class. 278 * @param use_declaring_class 'true' if the type can be loaded off the current Method*. 279 * @param can_assume_type_is_in_dex_cache 'true' if the type is known to be in the cache. 280 * @param type_idx Type index to use if use_declaring_class is 'false'. 281 * @param rl_dest Result to be set to 0 or 1. 282 * @param rl_src Object to be tested. 283 */ 284 void GenInstanceofCallingHelper(bool needs_access_check, bool type_known_final, 285 bool type_known_abstract, bool use_declaring_class, 286 bool can_assume_type_is_in_dex_cache, 287 uint32_t type_idx, RegLocation rl_dest, RegLocation rl_src); 288 289 void GenShiftOpLong(Instruction::Code opcode, RegLocation rl_dest, 290 RegLocation rl_src1, RegLocation rl_shift); 291 292 // Single operation generators. 293 LIR* OpUnconditionalBranch(LIR* target); 294 LIR* OpCmpBranch(ConditionCode cond, RegStorage src1, RegStorage src2, LIR* target); 295 LIR* OpCmpImmBranch(ConditionCode cond, RegStorage reg, int check_value, LIR* target); 296 LIR* OpCondBranch(ConditionCode cc, LIR* target); 297 LIR* OpDecAndBranch(ConditionCode c_code, RegStorage reg, LIR* target); 298 LIR* OpFpRegCopy(RegStorage r_dest, RegStorage r_src); 299 LIR* OpIT(ConditionCode cond, const char* guide); 300 void OpEndIT(LIR* it); 301 LIR* OpMem(OpKind op, RegStorage r_base, int disp); 302 LIR* OpPcRelLoad(RegStorage reg, LIR* target); 303 LIR* OpReg(OpKind op, RegStorage r_dest_src); 304 void OpRegCopy(RegStorage r_dest, RegStorage r_src); 305 LIR* OpRegCopyNoInsert(RegStorage r_dest, RegStorage r_src); 306 LIR* OpRegImm(OpKind op, RegStorage r_dest_src1, int value); 307 LIR* OpRegMem(OpKind op, RegStorage r_dest, RegStorage r_base, int offset); 308 LIR* OpRegMem(OpKind op, RegStorage r_dest, RegLocation value); 309 LIR* OpMemReg(OpKind op, RegLocation rl_dest, int value); 310 LIR* OpRegReg(OpKind op, RegStorage r_dest_src1, RegStorage r_src2); 311 LIR* OpMovRegMem(RegStorage r_dest, RegStorage r_base, int offset, MoveType move_type); 312 LIR* OpMovMemReg(RegStorage r_base, int offset, RegStorage r_src, MoveType move_type); 313 LIR* OpCondRegReg(OpKind op, ConditionCode cc, RegStorage r_dest, RegStorage r_src); 314 LIR* OpRegRegImm(OpKind op, RegStorage r_dest, RegStorage r_src1, int value); 315 LIR* OpRegRegReg(OpKind op, RegStorage r_dest, RegStorage r_src1, RegStorage r_src2); 316 LIR* OpTestSuspend(LIR* target); 317 LIR* OpThreadMem(OpKind op, ThreadOffset<4> thread_offset) OVERRIDE; 318 LIR* OpThreadMem(OpKind op, ThreadOffset<8> thread_offset) OVERRIDE; 319 LIR* OpVldm(RegStorage r_base, int count); 320 LIR* OpVstm(RegStorage r_base, int count); 321 void OpLea(RegStorage r_base, RegStorage reg1, RegStorage reg2, int scale, int offset); 322 void OpRegCopyWide(RegStorage dest, RegStorage src); 323 void OpTlsCmp(ThreadOffset<4> offset, int val) OVERRIDE; 324 void OpTlsCmp(ThreadOffset<8> offset, int val) OVERRIDE; 325 326 void OpRegThreadMem(OpKind op, RegStorage r_dest, ThreadOffset<4> thread_offset); 327 void OpRegThreadMem(OpKind op, RegStorage r_dest, ThreadOffset<8> thread_offset); 328 void SpillCoreRegs(); 329 void UnSpillCoreRegs(); 330 void UnSpillFPRegs(); 331 void SpillFPRegs(); 332 static const X86EncodingMap EncodingMap[kX86Last]; 333 bool InexpensiveConstantInt(int32_t value); 334 bool InexpensiveConstantFloat(int32_t value); 335 bool InexpensiveConstantLong(int64_t value); 336 bool InexpensiveConstantDouble(int64_t value); 337 338 /* 339 * @brief Should try to optimize for two address instructions? 340 * @return true if we try to avoid generating three operand instructions. 341 */ 342 virtual bool GenerateTwoOperandInstructions() const { return true; } 343 344 /* 345 * @brief x86 specific codegen for int operations. 346 * @param opcode Operation to perform. 347 * @param rl_dest Destination for the result. 348 * @param rl_lhs Left hand operand. 349 * @param rl_rhs Right hand operand. 350 */ 351 void GenArithOpInt(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_lhs, 352 RegLocation rl_rhs); 353 354 /* 355 * @brief Dump a RegLocation using printf 356 * @param loc Register location to dump 357 */ 358 static void DumpRegLocation(RegLocation loc); 359 360 /* 361 * @brief Load the Method* of a dex method into the register. 362 * @param target_method The MethodReference of the method to be invoked. 363 * @param type How the method will be invoked. 364 * @param register that will contain the code address. 365 * @note register will be passed to TargetReg to get physical register. 366 */ 367 void LoadMethodAddress(const MethodReference& target_method, InvokeType type, 368 SpecialTargetRegister symbolic_reg); 369 370 /* 371 * @brief Load the Class* of a Dex Class type into the register. 372 * @param type How the method will be invoked. 373 * @param register that will contain the code address. 374 * @note register will be passed to TargetReg to get physical register. 375 */ 376 void LoadClassType(uint32_t type_idx, SpecialTargetRegister symbolic_reg); 377 378 void FlushIns(RegLocation* ArgLocs, RegLocation rl_method); 379 380 int GenDalvikArgsNoRange(CallInfo* info, int call_state, LIR** pcrLabel, 381 NextCallInsn next_call_insn, 382 const MethodReference& target_method, 383 uint32_t vtable_idx, 384 uintptr_t direct_code, uintptr_t direct_method, InvokeType type, 385 bool skip_this); 386 387 int GenDalvikArgsRange(CallInfo* info, int call_state, LIR** pcrLabel, 388 NextCallInsn next_call_insn, 389 const MethodReference& target_method, 390 uint32_t vtable_idx, 391 uintptr_t direct_code, uintptr_t direct_method, InvokeType type, 392 bool skip_this); 393 394 /* 395 * @brief Generate a relative call to the method that will be patched at link time. 396 * @param target_method The MethodReference of the method to be invoked. 397 * @param type How the method will be invoked. 398 * @returns Call instruction 399 */ 400 virtual LIR * CallWithLinkerFixup(const MethodReference& target_method, InvokeType type); 401 402 /* 403 * @brief Handle x86 specific literals 404 */ 405 void InstallLiteralPools(); 406 407 /* 408 * @brief Generate the debug_frame CFI information. 409 * @returns pointer to vector containing CFE information 410 */ 411 static std::vector<uint8_t>* ReturnCommonCallFrameInformation(); 412 413 /* 414 * @brief Generate the debug_frame FDE information. 415 * @returns pointer to vector containing CFE information 416 */ 417 std::vector<uint8_t>* ReturnCallFrameInformation(); 418 419 protected: 420 // Casting of RegStorage 421 RegStorage As32BitReg(RegStorage reg) { 422 DCHECK(!reg.IsPair()); 423 if ((kFailOnSizeError || kReportSizeError) && !reg.Is64Bit()) { 424 if (kFailOnSizeError) { 425 LOG(FATAL) << "Expected 64b register " << reg.GetReg(); 426 } else { 427 LOG(WARNING) << "Expected 64b register " << reg.GetReg(); 428 return reg; 429 } 430 } 431 RegStorage ret_val = RegStorage(RegStorage::k32BitSolo, 432 reg.GetRawBits() & RegStorage::kRegTypeMask); 433 DCHECK_EQ(GetRegInfo(reg)->FindMatchingView(RegisterInfo::k32SoloStorageMask) 434 ->GetReg().GetReg(), 435 ret_val.GetReg()); 436 return ret_val; 437 } 438 439 RegStorage As64BitReg(RegStorage reg) { 440 DCHECK(!reg.IsPair()); 441 if ((kFailOnSizeError || kReportSizeError) && !reg.Is32Bit()) { 442 if (kFailOnSizeError) { 443 LOG(FATAL) << "Expected 32b register " << reg.GetReg(); 444 } else { 445 LOG(WARNING) << "Expected 32b register " << reg.GetReg(); 446 return reg; 447 } 448 } 449 RegStorage ret_val = RegStorage(RegStorage::k64BitSolo, 450 reg.GetRawBits() & RegStorage::kRegTypeMask); 451 DCHECK_EQ(GetRegInfo(reg)->FindMatchingView(RegisterInfo::k64SoloStorageMask) 452 ->GetReg().GetReg(), 453 ret_val.GetReg()); 454 return ret_val; 455 } 456 457 size_t ComputeSize(const X86EncodingMap* entry, int32_t raw_reg, int32_t raw_index, 458 int32_t raw_base, int32_t displacement); 459 void CheckValidByteRegister(const X86EncodingMap* entry, int32_t raw_reg); 460 void EmitPrefix(const X86EncodingMap* entry, 461 int32_t raw_reg_r, int32_t raw_reg_x, int32_t raw_reg_b); 462 void EmitOpcode(const X86EncodingMap* entry); 463 void EmitPrefixAndOpcode(const X86EncodingMap* entry, 464 int32_t reg_r, int32_t reg_x, int32_t reg_b); 465 void EmitDisp(uint8_t base, int32_t disp); 466 void EmitModrmThread(uint8_t reg_or_opcode); 467 void EmitModrmDisp(uint8_t reg_or_opcode, uint8_t base, int32_t disp); 468 void EmitModrmSibDisp(uint8_t reg_or_opcode, uint8_t base, uint8_t index, int scale, 469 int32_t disp); 470 void EmitImm(const X86EncodingMap* entry, int64_t imm); 471 void EmitNullary(const X86EncodingMap* entry); 472 void EmitOpRegOpcode(const X86EncodingMap* entry, int32_t raw_reg); 473 void EmitOpReg(const X86EncodingMap* entry, int32_t raw_reg); 474 void EmitOpMem(const X86EncodingMap* entry, int32_t raw_base, int32_t disp); 475 void EmitOpArray(const X86EncodingMap* entry, int32_t raw_base, int32_t raw_index, int scale, 476 int32_t disp); 477 void EmitMemReg(const X86EncodingMap* entry, int32_t raw_base, int32_t disp, int32_t raw_reg); 478 void EmitRegMem(const X86EncodingMap* entry, int32_t raw_reg, int32_t raw_base, int32_t disp); 479 void EmitRegArray(const X86EncodingMap* entry, int32_t raw_reg, int32_t raw_base, 480 int32_t raw_index, int scale, int32_t disp); 481 void EmitArrayReg(const X86EncodingMap* entry, int32_t raw_base, int32_t raw_index, int scale, 482 int32_t disp, int32_t raw_reg); 483 void EmitMemImm(const X86EncodingMap* entry, int32_t raw_base, int32_t disp, int32_t imm); 484 void EmitArrayImm(const X86EncodingMap* entry, int32_t raw_base, int32_t raw_index, int scale, 485 int32_t raw_disp, int32_t imm); 486 void EmitRegThread(const X86EncodingMap* entry, int32_t raw_reg, int32_t disp); 487 void EmitRegReg(const X86EncodingMap* entry, int32_t raw_reg1, int32_t raw_reg2); 488 void EmitRegRegImm(const X86EncodingMap* entry, int32_t raw_reg1, int32_t raw_reg2, int32_t imm); 489 void EmitRegMemImm(const X86EncodingMap* entry, int32_t raw_reg1, int32_t raw_base, int32_t disp, 490 int32_t imm); 491 void EmitMemRegImm(const X86EncodingMap* entry, int32_t base, int32_t disp, int32_t raw_reg1, 492 int32_t imm); 493 void EmitRegImm(const X86EncodingMap* entry, int32_t raw_reg, int32_t imm); 494 void EmitThreadImm(const X86EncodingMap* entry, int32_t disp, int32_t imm); 495 void EmitMovRegImm(const X86EncodingMap* entry, int32_t raw_reg, int64_t imm); 496 void EmitShiftRegImm(const X86EncodingMap* entry, int32_t raw_reg, int32_t imm); 497 void EmitShiftRegCl(const X86EncodingMap* entry, int32_t raw_reg, int32_t raw_cl); 498 void EmitShiftMemCl(const X86EncodingMap* entry, int32_t raw_base, int32_t disp, int32_t raw_cl); 499 void EmitShiftMemImm(const X86EncodingMap* entry, int32_t raw_base, int32_t disp, int32_t imm); 500 void EmitRegCond(const X86EncodingMap* entry, int32_t raw_reg, int32_t cc); 501 void EmitMemCond(const X86EncodingMap* entry, int32_t raw_base, int32_t disp, int32_t cc); 502 void EmitRegRegCond(const X86EncodingMap* entry, int32_t raw_reg1, int32_t raw_reg2, int32_t cc); 503 void EmitRegMemCond(const X86EncodingMap* entry, int32_t raw_reg1, int32_t raw_base, int32_t disp, 504 int32_t cc); 505 506 void EmitJmp(const X86EncodingMap* entry, int32_t rel); 507 void EmitJcc(const X86EncodingMap* entry, int32_t rel, int32_t cc); 508 void EmitCallMem(const X86EncodingMap* entry, int32_t raw_base, int32_t disp); 509 void EmitCallImmediate(const X86EncodingMap* entry, int32_t disp); 510 void EmitCallThread(const X86EncodingMap* entry, int32_t disp); 511 void EmitPcRel(const X86EncodingMap* entry, int32_t raw_reg, int32_t raw_base_or_table, 512 int32_t raw_index, int scale, int32_t table_or_disp); 513 void EmitMacro(const X86EncodingMap* entry, int32_t raw_reg, int32_t offset); 514 void EmitUnimplemented(const X86EncodingMap* entry, LIR* lir); 515 void GenFusedLongCmpImmBranch(BasicBlock* bb, RegLocation rl_src1, 516 int64_t val, ConditionCode ccode); 517 void GenConstWide(RegLocation rl_dest, int64_t value); 518 void GenMultiplyVectorSignedByte(BasicBlock *bb, MIR *mir); 519 void GenShiftByteVector(BasicBlock *bb, MIR *mir); 520 void AndMaskVectorRegister(RegStorage rs_src1, uint32_t m1, uint32_t m2, uint32_t m3, uint32_t m4); 521 void MaskVectorRegister(X86OpCode opcode, RegStorage rs_src1, uint32_t m1, uint32_t m2, uint32_t m3, uint32_t m4); 522 void AppendOpcodeWithConst(X86OpCode opcode, int reg, MIR* mir); 523 524 static bool ProvidesFullMemoryBarrier(X86OpCode opcode); 525 526 /* 527 * @brief Ensure that a temporary register is byte addressable. 528 * @returns a temporary guarenteed to be byte addressable. 529 */ 530 virtual RegStorage AllocateByteRegister(); 531 532 /* 533 * @brief Use a wide temporary as a 128-bit register 534 * @returns a 128-bit temporary register. 535 */ 536 virtual RegStorage Get128BitRegister(RegStorage reg); 537 538 /* 539 * @brief Check if a register is byte addressable. 540 * @returns true if a register is byte addressable. 541 */ 542 bool IsByteRegister(RegStorage reg); 543 bool GenInlinedArrayCopyCharArray(CallInfo* info) OVERRIDE; 544 545 /* 546 * @brief generate inline code for fast case of Strng.indexOf. 547 * @param info Call parameters 548 * @param zero_based 'true' if the index into the string is 0. 549 * @returns 'true' if the call was inlined, 'false' if a regular call needs to be 550 * generated. 551 */ 552 bool GenInlinedIndexOf(CallInfo* info, bool zero_based); 553 554 /** 555 * @brief Reserve a fixed number of vector registers from the register pool 556 * @details The mir->dalvikInsn.vA specifies an N such that vector registers 557 * [0..N-1] are removed from the temporary pool. The caller must call 558 * ReturnVectorRegisters before calling ReserveVectorRegisters again. 559 * Also sets the num_reserved_vector_regs_ to the specified value 560 * @param mir whose vA specifies the number of registers to reserve 561 */ 562 void ReserveVectorRegisters(MIR* mir); 563 564 /** 565 * @brief Return all the reserved vector registers to the temp pool 566 * @details Returns [0..num_reserved_vector_regs_] 567 */ 568 void ReturnVectorRegisters(); 569 570 /* 571 * @brief Load 128 bit constant into vector register. 572 * @param bb The basic block in which the MIR is from. 573 * @param mir The MIR whose opcode is kMirConstVector 574 * @note vA is the TypeSize for the register. 575 * @note vB is the destination XMM register. arg[0..3] are 32 bit constant values. 576 */ 577 void GenConst128(BasicBlock* bb, MIR* mir); 578 579 /* 580 * @brief MIR to move a vectorized register to another. 581 * @param bb The basic block in which the MIR is from. 582 * @param mir The MIR whose opcode is kMirConstVector. 583 * @note vA: TypeSize 584 * @note vB: destination 585 * @note vC: source 586 */ 587 void GenMoveVector(BasicBlock *bb, MIR *mir); 588 589 /* 590 * @brief Packed multiply of units in two vector registers: vB = vB .* @note vC using vA to know the type of the vector. 591 * @param bb The basic block in which the MIR is from. 592 * @param mir The MIR whose opcode is kMirConstVector. 593 * @note vA: TypeSize 594 * @note vB: destination and source 595 * @note vC: source 596 */ 597 void GenMultiplyVector(BasicBlock *bb, MIR *mir); 598 599 /* 600 * @brief Packed addition of units in two vector registers: vB = vB .+ vC using vA to know the type of the vector. 601 * @param bb The basic block in which the MIR is from. 602 * @param mir The MIR whose opcode is kMirConstVector. 603 * @note vA: TypeSize 604 * @note vB: destination and source 605 * @note vC: source 606 */ 607 void GenAddVector(BasicBlock *bb, MIR *mir); 608 609 /* 610 * @brief Packed subtraction of units in two vector registers: vB = vB .- vC using vA to know the type of the vector. 611 * @param bb The basic block in which the MIR is from. 612 * @param mir The MIR whose opcode is kMirConstVector. 613 * @note vA: TypeSize 614 * @note vB: destination and source 615 * @note vC: source 616 */ 617 void GenSubtractVector(BasicBlock *bb, MIR *mir); 618 619 /* 620 * @brief Packed shift left of units in two vector registers: vB = vB .<< vC using vA to know the type of the vector. 621 * @param bb The basic block in which the MIR is from. 622 * @param mir The MIR whose opcode is kMirConstVector. 623 * @note vA: TypeSize 624 * @note vB: destination and source 625 * @note vC: immediate 626 */ 627 void GenShiftLeftVector(BasicBlock *bb, MIR *mir); 628 629 /* 630 * @brief Packed signed shift right of units in two vector registers: vB = vB .>> vC using vA to know the type of the vector. 631 * @param bb The basic block in which the MIR is from. 632 * @param mir The MIR whose opcode is kMirConstVector. 633 * @note vA: TypeSize 634 * @note vB: destination and source 635 * @note vC: immediate 636 */ 637 void GenSignedShiftRightVector(BasicBlock *bb, MIR *mir); 638 639 /* 640 * @brief Packed unsigned shift right of units in two vector registers: vB = vB .>>> vC using vA to know the type of the vector. 641 * @param bb The basic block in which the MIR is from.. 642 * @param mir The MIR whose opcode is kMirConstVector. 643 * @note vA: TypeSize 644 * @note vB: destination and source 645 * @note vC: immediate 646 */ 647 void GenUnsignedShiftRightVector(BasicBlock *bb, MIR *mir); 648 649 /* 650 * @brief Packed bitwise and of units in two vector registers: vB = vB .& vC using vA to know the type of the vector. 651 * @note vA: TypeSize 652 * @note vB: destination and source 653 * @note vC: source 654 */ 655 void GenAndVector(BasicBlock *bb, MIR *mir); 656 657 /* 658 * @brief Packed bitwise or of units in two vector registers: vB = vB .| vC using vA to know the type of the vector. 659 * @param bb The basic block in which the MIR is from. 660 * @param mir The MIR whose opcode is kMirConstVector. 661 * @note vA: TypeSize 662 * @note vB: destination and source 663 * @note vC: source 664 */ 665 void GenOrVector(BasicBlock *bb, MIR *mir); 666 667 /* 668 * @brief Packed bitwise xor of units in two vector registers: vB = vB .^ vC using vA to know the type of the vector. 669 * @param bb The basic block in which the MIR is from. 670 * @param mir The MIR whose opcode is kMirConstVector. 671 * @note vA: TypeSize 672 * @note vB: destination and source 673 * @note vC: source 674 */ 675 void GenXorVector(BasicBlock *bb, MIR *mir); 676 677 /* 678 * @brief Reduce a 128-bit packed element into a single VR by taking lower bits 679 * @param bb The basic block in which the MIR is from. 680 * @param mir The MIR whose opcode is kMirConstVector. 681 * @details Instruction does a horizontal addition of the packed elements and then adds it to VR. 682 * @note vA: TypeSize 683 * @note vB: destination and source VR (not vector register) 684 * @note vC: source (vector register) 685 */ 686 void GenAddReduceVector(BasicBlock *bb, MIR *mir); 687 688 /* 689 * @brief Extract a packed element into a single VR. 690 * @param bb The basic block in which the MIR is from. 691 * @param mir The MIR whose opcode is kMirConstVector. 692 * @note vA: TypeSize 693 * @note vB: destination VR (not vector register) 694 * @note vC: source (vector register) 695 * @note arg[0]: The index to use for extraction from vector register (which packed element). 696 */ 697 void GenReduceVector(BasicBlock *bb, MIR *mir); 698 699 /* 700 * @brief Create a vector value, with all TypeSize values equal to vC 701 * @param bb The basic block in which the MIR is from. 702 * @param mir The MIR whose opcode is kMirConstVector. 703 * @note vA: TypeSize. 704 * @note vB: destination vector register. 705 * @note vC: source VR (not vector register). 706 */ 707 void GenSetVector(BasicBlock *bb, MIR *mir); 708 709 /* 710 * @brief Generate code for a vector opcode. 711 * @param bb The basic block in which the MIR is from. 712 * @param mir The MIR whose opcode is a non-standard opcode. 713 */ 714 void GenMachineSpecificExtendedMethodMIR(BasicBlock* bb, MIR* mir); 715 716 /* 717 * @brief Return the correct x86 opcode for the Dex operation 718 * @param op Dex opcode for the operation 719 * @param loc Register location of the operand 720 * @param is_high_op 'true' if this is an operation on the high word 721 * @param value Immediate value for the operation. Used for byte variants 722 * @returns the correct x86 opcode to perform the operation 723 */ 724 X86OpCode GetOpcode(Instruction::Code op, RegLocation loc, bool is_high_op, int32_t value); 725 726 /* 727 * @brief Return the correct x86 opcode for the Dex operation 728 * @param op Dex opcode for the operation 729 * @param dest location of the destination. May be register or memory. 730 * @param rhs Location for the rhs of the operation. May be in register or memory. 731 * @param is_high_op 'true' if this is an operation on the high word 732 * @returns the correct x86 opcode to perform the operation 733 * @note at most one location may refer to memory 734 */ 735 X86OpCode GetOpcode(Instruction::Code op, RegLocation dest, RegLocation rhs, 736 bool is_high_op); 737 738 /* 739 * @brief Is this operation a no-op for this opcode and value 740 * @param op Dex opcode for the operation 741 * @param value Immediate value for the operation. 742 * @returns 'true' if the operation will have no effect 743 */ 744 bool IsNoOp(Instruction::Code op, int32_t value); 745 746 /** 747 * @brief Calculate magic number and shift for a given divisor 748 * @param divisor divisor number for calculation 749 * @param magic hold calculated magic number 750 * @param shift hold calculated shift 751 */ 752 void CalculateMagicAndShift(int divisor, int& magic, int& shift); 753 754 /* 755 * @brief Generate an integer div or rem operation. 756 * @param rl_dest Destination Location. 757 * @param rl_src1 Numerator Location. 758 * @param rl_src2 Divisor Location. 759 * @param is_div 'true' if this is a division, 'false' for a remainder. 760 * @param check_zero 'true' if an exception should be generated if the divisor is 0. 761 */ 762 RegLocation GenDivRem(RegLocation rl_dest, RegLocation rl_src1, RegLocation rl_src2, 763 bool is_div, bool check_zero); 764 765 /* 766 * @brief Generate an integer div or rem operation by a literal. 767 * @param rl_dest Destination Location. 768 * @param rl_src Numerator Location. 769 * @param lit Divisor. 770 * @param is_div 'true' if this is a division, 'false' for a remainder. 771 */ 772 RegLocation GenDivRemLit(RegLocation rl_dest, RegLocation rl_src, int lit, bool is_div); 773 774 /* 775 * Generate code to implement long shift operations. 776 * @param opcode The DEX opcode to specify the shift type. 777 * @param rl_dest The destination. 778 * @param rl_src The value to be shifted. 779 * @param shift_amount How much to shift. 780 * @returns the RegLocation of the result. 781 */ 782 RegLocation GenShiftImmOpLong(Instruction::Code opcode, RegLocation rl_dest, 783 RegLocation rl_src, int shift_amount); 784 /* 785 * Generate an imul of a register by a constant or a better sequence. 786 * @param dest Destination Register. 787 * @param src Source Register. 788 * @param val Constant multiplier. 789 */ 790 void GenImulRegImm(RegStorage dest, RegStorage src, int val); 791 792 /* 793 * Generate an imul of a memory location by a constant or a better sequence. 794 * @param dest Destination Register. 795 * @param sreg Symbolic register. 796 * @param displacement Displacement on stack of Symbolic Register. 797 * @param val Constant multiplier. 798 */ 799 void GenImulMemImm(RegStorage dest, int sreg, int displacement, int val); 800 801 /* 802 * @brief Compare memory to immediate, and branch if condition true. 803 * @param cond The condition code that when true will branch to the target. 804 * @param temp_reg A temporary register that can be used if compare memory is not 805 * supported by the architecture. 806 * @param base_reg The register holding the base address. 807 * @param offset The offset from the base. 808 * @param check_value The immediate to compare to. 809 * @param target branch target (or nullptr) 810 * @param compare output for getting LIR for comparison (or nullptr) 811 */ 812 LIR* OpCmpMemImmBranch(ConditionCode cond, RegStorage temp_reg, RegStorage base_reg, 813 int offset, int check_value, LIR* target, LIR** compare); 814 815 /* 816 * Can this operation be using core registers without temporaries? 817 * @param rl_lhs Left hand operand. 818 * @param rl_rhs Right hand operand. 819 * @returns 'true' if the operation can proceed without needing temporary regs. 820 */ 821 bool IsOperationSafeWithoutTemps(RegLocation rl_lhs, RegLocation rl_rhs); 822 823 /** 824 * @brief Generates inline code for conversion of long to FP by using x87/ 825 * @param rl_dest The destination of the FP. 826 * @param rl_src The source of the long. 827 * @param is_double 'true' if dealing with double, 'false' for float. 828 */ 829 virtual void GenLongToFP(RegLocation rl_dest, RegLocation rl_src, bool is_double); 830 831 /* 832 * @brief Perform MIR analysis before compiling method. 833 * @note Invokes Mir2LiR::Materialize after analysis. 834 */ 835 void Materialize(); 836 837 /* 838 * Mir2Lir's UpdateLoc() looks to see if the Dalvik value is currently live in any temp register 839 * without regard to data type. In practice, this can result in UpdateLoc returning a 840 * location record for a Dalvik float value in a core register, and vis-versa. For targets 841 * which can inexpensively move data between core and float registers, this can often be a win. 842 * However, for x86 this is generally not a win. These variants of UpdateLoc() 843 * take a register class argument - and will return an in-register location record only if 844 * the value is live in a temp register of the correct class. Additionally, if the value is in 845 * a temp register of the wrong register class, it will be clobbered. 846 */ 847 RegLocation UpdateLocTyped(RegLocation loc, int reg_class); 848 RegLocation UpdateLocWideTyped(RegLocation loc, int reg_class); 849 850 /* 851 * @brief Analyze MIR before generating code, to prepare for the code generation. 852 */ 853 void AnalyzeMIR(); 854 855 /* 856 * @brief Analyze one basic block. 857 * @param bb Basic block to analyze. 858 */ 859 void AnalyzeBB(BasicBlock * bb); 860 861 /* 862 * @brief Analyze one extended MIR instruction 863 * @param opcode MIR instruction opcode. 864 * @param bb Basic block containing instruction. 865 * @param mir Extended instruction to analyze. 866 */ 867 void AnalyzeExtendedMIR(int opcode, BasicBlock * bb, MIR *mir); 868 869 /* 870 * @brief Analyze one MIR instruction 871 * @param opcode MIR instruction opcode. 872 * @param bb Basic block containing instruction. 873 * @param mir Instruction to analyze. 874 */ 875 virtual void AnalyzeMIR(int opcode, BasicBlock * bb, MIR *mir); 876 877 /* 878 * @brief Analyze one MIR float/double instruction 879 * @param opcode MIR instruction opcode. 880 * @param bb Basic block containing instruction. 881 * @param mir Instruction to analyze. 882 */ 883 void AnalyzeFPInstruction(int opcode, BasicBlock * bb, MIR *mir); 884 885 /* 886 * @brief Analyze one use of a double operand. 887 * @param rl_use Double RegLocation for the operand. 888 */ 889 void AnalyzeDoubleUse(RegLocation rl_use); 890 891 /* 892 * @brief Analyze one invoke-static MIR instruction 893 * @param opcode MIR instruction opcode. 894 * @param bb Basic block containing instruction. 895 * @param mir Instruction to analyze. 896 */ 897 void AnalyzeInvokeStatic(int opcode, BasicBlock * bb, MIR *mir); 898 899 // Information derived from analysis of MIR 900 901 // The compiler temporary for the code address of the method. 902 CompilerTemp *base_of_code_; 903 904 // Have we decided to compute a ptr to code and store in temporary VR? 905 bool store_method_addr_; 906 907 // Have we used the stored method address? 908 bool store_method_addr_used_; 909 910 // Instructions to remove if we didn't use the stored method address. 911 LIR* setup_method_address_[2]; 912 913 // Instructions needing patching with Method* values. 914 GrowableArray<LIR*> method_address_insns_; 915 916 // Instructions needing patching with Class Type* values. 917 GrowableArray<LIR*> class_type_address_insns_; 918 919 // Instructions needing patching with PC relative code addresses. 920 GrowableArray<LIR*> call_method_insns_; 921 922 // Prologue decrement of stack pointer. 923 LIR* stack_decrement_; 924 925 // Epilogue increment of stack pointer. 926 LIR* stack_increment_; 927 928 // The list of const vector literals. 929 LIR *const_vectors_; 930 931 /* 932 * @brief Search for a matching vector literal 933 * @param mir A kMirOpConst128b MIR instruction to match. 934 * @returns pointer to matching LIR constant, or nullptr if not found. 935 */ 936 LIR *ScanVectorLiteral(MIR *mir); 937 938 /* 939 * @brief Add a constant vector literal 940 * @param mir A kMirOpConst128b MIR instruction to match. 941 */ 942 LIR *AddVectorLiteral(MIR *mir); 943 944 InToRegStorageMapping in_to_reg_storage_mapping_; 945 946 bool WideGPRsAreAliases() OVERRIDE { 947 return cu_->target64; // On 64b, we have 64b GPRs. 948 } 949 bool WideFPRsAreAliases() OVERRIDE { 950 return true; // xmm registers have 64b views even on x86. 951 } 952 953 private: 954 // The number of vector registers [0..N] reserved by a call to ReserveVectorRegisters 955 int num_reserved_vector_regs_; 956}; 957 958} // namespace art 959 960#endif // ART_COMPILER_DEX_QUICK_X86_CODEGEN_X86_H_ 961