target_x86.cc revision 674744e635ddbdfb311fbd25b5a27356560d30c3
1/* 2 * Copyright (C) 2012 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#include <string> 18#include <inttypes.h> 19 20#include "codegen_x86.h" 21#include "dex/compiler_internals.h" 22#include "dex/quick/mir_to_lir-inl.h" 23#include "mirror/array.h" 24#include "mirror/string.h" 25#include "x86_lir.h" 26 27namespace art { 28 29static const RegStorage core_regs_arr[] = { 30 rs_rAX, rs_rCX, rs_rDX, rs_rBX, rs_rX86_SP, rs_rBP, rs_rSI, rs_rDI 31#ifdef TARGET_REX_SUPPORT 32 rs_r8, rs_r9, rs_r10, rs_r11, rs_r12, rs_r13, rs_r14, rs_r15 33#endif 34}; 35static const RegStorage sp_regs_arr[] = { 36 rs_fr0, rs_fr1, rs_fr2, rs_fr3, rs_fr4, rs_fr5, rs_fr6, rs_fr7, 37#ifdef TARGET_REX_SUPPORT 38 rs_fr8, rs_fr9, rs_fr10, rs_fr11, rs_fr12, rs_fr13, rs_fr14, rs_fr15 39#endif 40}; 41static const RegStorage dp_regs_arr[] = { 42 rs_dr0, rs_dr1, rs_dr2, rs_dr3, rs_dr4, rs_dr5, rs_dr6, rs_dr7, 43#ifdef TARGET_REX_SUPPORT 44 rs_dr8, rs_dr9, rs_dr10, rs_dr11, rs_dr12, rs_dr13, rs_dr14, rs_dr15 45#endif 46}; 47static const RegStorage reserved_regs_arr[] = {rs_rX86_SP}; 48static const RegStorage core_temps_arr[] = {rs_rAX, rs_rCX, rs_rDX, rs_rBX}; 49static const RegStorage sp_temps_arr[] = { 50 rs_fr0, rs_fr1, rs_fr2, rs_fr3, rs_fr4, rs_fr5, rs_fr6, rs_fr7, 51#ifdef TARGET_REX_SUPPORT 52 rs_fr8, rs_fr9, rs_fr10, rs_fr11, rs_fr12, rs_fr13, rs_fr14, rs_fr15 53#endif 54}; 55static const RegStorage dp_temps_arr[] = { 56 rs_dr0, rs_dr1, rs_dr2, rs_dr3, rs_dr4, rs_dr5, rs_dr6, rs_dr7, 57#ifdef TARGET_REX_SUPPORT 58 rs_dr8, rs_dr9, rs_dr10, rs_dr11, rs_dr12, rs_dr13, rs_dr14, rs_dr15 59#endif 60}; 61 62static const std::vector<RegStorage> core_regs(core_regs_arr, 63 core_regs_arr + sizeof(core_regs_arr) / sizeof(core_regs_arr[0])); 64static const std::vector<RegStorage> sp_regs(sp_regs_arr, 65 sp_regs_arr + sizeof(sp_regs_arr) / sizeof(sp_regs_arr[0])); 66static const std::vector<RegStorage> dp_regs(dp_regs_arr, 67 dp_regs_arr + sizeof(dp_regs_arr) / sizeof(dp_regs_arr[0])); 68static const std::vector<RegStorage> reserved_regs(reserved_regs_arr, 69 reserved_regs_arr + sizeof(reserved_regs_arr) / sizeof(reserved_regs_arr[0])); 70static const std::vector<RegStorage> core_temps(core_temps_arr, 71 core_temps_arr + sizeof(core_temps_arr) / sizeof(core_temps_arr[0])); 72static const std::vector<RegStorage> sp_temps(sp_temps_arr, 73 sp_temps_arr + sizeof(sp_temps_arr) / sizeof(sp_temps_arr[0])); 74static const std::vector<RegStorage> dp_temps(dp_temps_arr, 75 dp_temps_arr + sizeof(dp_temps_arr) / sizeof(dp_temps_arr[0])); 76 77RegLocation X86Mir2Lir::LocCReturn() { 78 return x86_loc_c_return; 79} 80 81RegLocation X86Mir2Lir::LocCReturnWide() { 82 return x86_loc_c_return_wide; 83} 84 85RegLocation X86Mir2Lir::LocCReturnFloat() { 86 return x86_loc_c_return_float; 87} 88 89RegLocation X86Mir2Lir::LocCReturnDouble() { 90 return x86_loc_c_return_double; 91} 92 93// Return a target-dependent special register. 94RegStorage X86Mir2Lir::TargetReg(SpecialTargetRegister reg) { 95 RegStorage res_reg = RegStorage::InvalidReg(); 96 switch (reg) { 97 case kSelf: res_reg = RegStorage::InvalidReg(); break; 98 case kSuspend: res_reg = RegStorage::InvalidReg(); break; 99 case kLr: res_reg = RegStorage::InvalidReg(); break; 100 case kPc: res_reg = RegStorage::InvalidReg(); break; 101 case kSp: res_reg = rs_rX86_SP; break; 102 case kArg0: res_reg = rs_rX86_ARG0; break; 103 case kArg1: res_reg = rs_rX86_ARG1; break; 104 case kArg2: res_reg = rs_rX86_ARG2; break; 105 case kArg3: res_reg = rs_rX86_ARG3; break; 106 case kFArg0: res_reg = rs_rX86_FARG0; break; 107 case kFArg1: res_reg = rs_rX86_FARG1; break; 108 case kFArg2: res_reg = rs_rX86_FARG2; break; 109 case kFArg3: res_reg = rs_rX86_FARG3; break; 110 case kRet0: res_reg = rs_rX86_RET0; break; 111 case kRet1: res_reg = rs_rX86_RET1; break; 112 case kInvokeTgt: res_reg = rs_rX86_INVOKE_TGT; break; 113 case kHiddenArg: res_reg = rs_rAX; break; 114 case kHiddenFpArg: res_reg = rs_fr0; break; 115 case kCount: res_reg = rs_rX86_COUNT; break; 116 } 117 return res_reg; 118} 119 120RegStorage X86Mir2Lir::GetArgMappingToPhysicalReg(int arg_num) { 121 // For the 32-bit internal ABI, the first 3 arguments are passed in registers. 122 // TODO: This is not 64-bit compliant and depends on new internal ABI. 123 switch (arg_num) { 124 case 0: 125 return rs_rX86_ARG1; 126 case 1: 127 return rs_rX86_ARG2; 128 case 2: 129 return rs_rX86_ARG3; 130 default: 131 return RegStorage::InvalidReg(); 132 } 133} 134 135/* 136 * Decode the register id. 137 */ 138uint64_t X86Mir2Lir::GetRegMaskCommon(RegStorage reg) { 139 uint64_t seed; 140 int shift; 141 int reg_id; 142 143 reg_id = reg.GetRegNum(); 144 /* Double registers in x86 are just a single FP register */ 145 seed = 1; 146 /* FP register starts at bit position 16 */ 147 shift = reg.IsFloat() ? kX86FPReg0 : 0; 148 /* Expand the double register id into single offset */ 149 shift += reg_id; 150 return (seed << shift); 151} 152 153uint64_t X86Mir2Lir::GetPCUseDefEncoding() { 154 /* 155 * FIXME: might make sense to use a virtual resource encoding bit for pc. Might be 156 * able to clean up some of the x86/Arm_Mips differences 157 */ 158 LOG(FATAL) << "Unexpected call to GetPCUseDefEncoding for x86"; 159 return 0ULL; 160} 161 162void X86Mir2Lir::SetupTargetResourceMasks(LIR* lir, uint64_t flags) { 163 DCHECK(cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64); 164 DCHECK(!lir->flags.use_def_invalid); 165 166 // X86-specific resource map setup here. 167 if (flags & REG_USE_SP) { 168 lir->u.m.use_mask |= ENCODE_X86_REG_SP; 169 } 170 171 if (flags & REG_DEF_SP) { 172 lir->u.m.def_mask |= ENCODE_X86_REG_SP; 173 } 174 175 if (flags & REG_DEFA) { 176 SetupRegMask(&lir->u.m.def_mask, rs_rAX.GetReg()); 177 } 178 179 if (flags & REG_DEFD) { 180 SetupRegMask(&lir->u.m.def_mask, rs_rDX.GetReg()); 181 } 182 if (flags & REG_USEA) { 183 SetupRegMask(&lir->u.m.use_mask, rs_rAX.GetReg()); 184 } 185 186 if (flags & REG_USEC) { 187 SetupRegMask(&lir->u.m.use_mask, rs_rCX.GetReg()); 188 } 189 190 if (flags & REG_USED) { 191 SetupRegMask(&lir->u.m.use_mask, rs_rDX.GetReg()); 192 } 193 194 if (flags & REG_USEB) { 195 SetupRegMask(&lir->u.m.use_mask, rs_rBX.GetReg()); 196 } 197 198 // Fixup hard to describe instruction: Uses rAX, rCX, rDI; sets rDI. 199 if (lir->opcode == kX86RepneScasw) { 200 SetupRegMask(&lir->u.m.use_mask, rs_rAX.GetReg()); 201 SetupRegMask(&lir->u.m.use_mask, rs_rCX.GetReg()); 202 SetupRegMask(&lir->u.m.use_mask, rs_rDI.GetReg()); 203 SetupRegMask(&lir->u.m.def_mask, rs_rDI.GetReg()); 204 } 205 206 if (flags & USE_FP_STACK) { 207 lir->u.m.use_mask |= ENCODE_X86_FP_STACK; 208 lir->u.m.def_mask |= ENCODE_X86_FP_STACK; 209 } 210} 211 212/* For dumping instructions */ 213static const char* x86RegName[] = { 214 "rax", "rcx", "rdx", "rbx", "rsp", "rbp", "rsi", "rdi", 215 "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15" 216}; 217 218static const char* x86CondName[] = { 219 "O", 220 "NO", 221 "B/NAE/C", 222 "NB/AE/NC", 223 "Z/EQ", 224 "NZ/NE", 225 "BE/NA", 226 "NBE/A", 227 "S", 228 "NS", 229 "P/PE", 230 "NP/PO", 231 "L/NGE", 232 "NL/GE", 233 "LE/NG", 234 "NLE/G" 235}; 236 237/* 238 * Interpret a format string and build a string no longer than size 239 * See format key in Assemble.cc. 240 */ 241std::string X86Mir2Lir::BuildInsnString(const char *fmt, LIR *lir, unsigned char* base_addr) { 242 std::string buf; 243 size_t i = 0; 244 size_t fmt_len = strlen(fmt); 245 while (i < fmt_len) { 246 if (fmt[i] != '!') { 247 buf += fmt[i]; 248 i++; 249 } else { 250 i++; 251 DCHECK_LT(i, fmt_len); 252 char operand_number_ch = fmt[i]; 253 i++; 254 if (operand_number_ch == '!') { 255 buf += "!"; 256 } else { 257 int operand_number = operand_number_ch - '0'; 258 DCHECK_LT(operand_number, 6); // Expect upto 6 LIR operands. 259 DCHECK_LT(i, fmt_len); 260 int operand = lir->operands[operand_number]; 261 switch (fmt[i]) { 262 case 'c': 263 DCHECK_LT(static_cast<size_t>(operand), sizeof(x86CondName)); 264 buf += x86CondName[operand]; 265 break; 266 case 'd': 267 buf += StringPrintf("%d", operand); 268 break; 269 case 'p': { 270 EmbeddedData *tab_rec = reinterpret_cast<EmbeddedData*>(UnwrapPointer(operand)); 271 buf += StringPrintf("0x%08x", tab_rec->offset); 272 break; 273 } 274 case 'r': 275 if (RegStorage::IsFloat(operand)) { 276 int fp_reg = RegStorage::RegNum(operand); 277 buf += StringPrintf("xmm%d", fp_reg); 278 } else { 279 int reg_num = RegStorage::RegNum(operand); 280 DCHECK_LT(static_cast<size_t>(reg_num), sizeof(x86RegName)); 281 buf += x86RegName[reg_num]; 282 } 283 break; 284 case 't': 285 buf += StringPrintf("0x%08" PRIxPTR " (L%p)", 286 reinterpret_cast<uintptr_t>(base_addr) + lir->offset + operand, 287 lir->target); 288 break; 289 default: 290 buf += StringPrintf("DecodeError '%c'", fmt[i]); 291 break; 292 } 293 i++; 294 } 295 } 296 } 297 return buf; 298} 299 300void X86Mir2Lir::DumpResourceMask(LIR *x86LIR, uint64_t mask, const char *prefix) { 301 char buf[256]; 302 buf[0] = 0; 303 304 if (mask == ENCODE_ALL) { 305 strcpy(buf, "all"); 306 } else { 307 char num[8]; 308 int i; 309 310 for (i = 0; i < kX86RegEnd; i++) { 311 if (mask & (1ULL << i)) { 312 snprintf(num, arraysize(num), "%d ", i); 313 strcat(buf, num); 314 } 315 } 316 317 if (mask & ENCODE_CCODE) { 318 strcat(buf, "cc "); 319 } 320 /* Memory bits */ 321 if (x86LIR && (mask & ENCODE_DALVIK_REG)) { 322 snprintf(buf + strlen(buf), arraysize(buf) - strlen(buf), "dr%d%s", 323 DECODE_ALIAS_INFO_REG(x86LIR->flags.alias_info), 324 (DECODE_ALIAS_INFO_WIDE(x86LIR->flags.alias_info)) ? "(+1)" : ""); 325 } 326 if (mask & ENCODE_LITERAL) { 327 strcat(buf, "lit "); 328 } 329 330 if (mask & ENCODE_HEAP_REF) { 331 strcat(buf, "heap "); 332 } 333 if (mask & ENCODE_MUST_NOT_ALIAS) { 334 strcat(buf, "noalias "); 335 } 336 } 337 if (buf[0]) { 338 LOG(INFO) << prefix << ": " << buf; 339 } 340} 341 342void X86Mir2Lir::AdjustSpillMask() { 343 // Adjustment for LR spilling, x86 has no LR so nothing to do here 344 core_spill_mask_ |= (1 << rs_rRET.GetRegNum()); 345 num_core_spills_++; 346} 347 348/* 349 * Mark a callee-save fp register as promoted. Note that 350 * vpush/vpop uses contiguous register lists so we must 351 * include any holes in the mask. Associate holes with 352 * Dalvik register INVALID_VREG (0xFFFFU). 353 */ 354void X86Mir2Lir::MarkPreservedSingle(int v_reg, RegStorage reg) { 355 UNIMPLEMENTED(FATAL) << "MarkPreservedSingle"; 356} 357 358void X86Mir2Lir::MarkPreservedDouble(int v_reg, RegStorage reg) { 359 UNIMPLEMENTED(FATAL) << "MarkPreservedDouble"; 360} 361 362/* Clobber all regs that might be used by an external C call */ 363void X86Mir2Lir::ClobberCallerSave() { 364 Clobber(rs_rAX); 365 Clobber(rs_rCX); 366 Clobber(rs_rDX); 367 Clobber(rs_rBX); 368} 369 370RegLocation X86Mir2Lir::GetReturnWideAlt() { 371 RegLocation res = LocCReturnWide(); 372 DCHECK(res.reg.GetLowReg() == rs_rAX.GetReg()); 373 DCHECK(res.reg.GetHighReg() == rs_rDX.GetReg()); 374 Clobber(rs_rAX); 375 Clobber(rs_rDX); 376 MarkInUse(rs_rAX); 377 MarkInUse(rs_rDX); 378 MarkWide(res.reg); 379 return res; 380} 381 382RegLocation X86Mir2Lir::GetReturnAlt() { 383 RegLocation res = LocCReturn(); 384 res.reg.SetReg(rs_rDX.GetReg()); 385 Clobber(rs_rDX); 386 MarkInUse(rs_rDX); 387 return res; 388} 389 390/* To be used when explicitly managing register use */ 391void X86Mir2Lir::LockCallTemps() { 392 LockTemp(rs_rX86_ARG0); 393 LockTemp(rs_rX86_ARG1); 394 LockTemp(rs_rX86_ARG2); 395 LockTemp(rs_rX86_ARG3); 396} 397 398/* To be used when explicitly managing register use */ 399void X86Mir2Lir::FreeCallTemps() { 400 FreeTemp(rs_rX86_ARG0); 401 FreeTemp(rs_rX86_ARG1); 402 FreeTemp(rs_rX86_ARG2); 403 FreeTemp(rs_rX86_ARG3); 404} 405 406bool X86Mir2Lir::ProvidesFullMemoryBarrier(X86OpCode opcode) { 407 switch (opcode) { 408 case kX86LockCmpxchgMR: 409 case kX86LockCmpxchgAR: 410 case kX86LockCmpxchg8bM: 411 case kX86LockCmpxchg8bA: 412 case kX86XchgMR: 413 case kX86Mfence: 414 // Atomic memory instructions provide full barrier. 415 return true; 416 default: 417 break; 418 } 419 420 // Conservative if cannot prove it provides full barrier. 421 return false; 422} 423 424void X86Mir2Lir::GenMemBarrier(MemBarrierKind barrier_kind) { 425#if ANDROID_SMP != 0 426 // Start off with using the last LIR as the barrier. If it is not enough, then we will update it. 427 LIR* mem_barrier = last_lir_insn_; 428 429 /* 430 * According to the JSR-133 Cookbook, for x86 only StoreLoad barriers need memory fence. All other barriers 431 * (LoadLoad, LoadStore, StoreStore) are nops due to the x86 memory model. For those cases, all we need 432 * to ensure is that there is a scheduling barrier in place. 433 */ 434 if (barrier_kind == kStoreLoad) { 435 // If no LIR exists already that can be used a barrier, then generate an mfence. 436 if (mem_barrier == nullptr) { 437 mem_barrier = NewLIR0(kX86Mfence); 438 } 439 440 // If last instruction does not provide full barrier, then insert an mfence. 441 if (ProvidesFullMemoryBarrier(static_cast<X86OpCode>(mem_barrier->opcode)) == false) { 442 mem_barrier = NewLIR0(kX86Mfence); 443 } 444 } 445 446 // Now ensure that a scheduling barrier is in place. 447 if (mem_barrier == nullptr) { 448 GenBarrier(); 449 } else { 450 // Mark as a scheduling barrier. 451 DCHECK(!mem_barrier->flags.use_def_invalid); 452 mem_barrier->u.m.def_mask = ENCODE_ALL; 453 } 454#endif 455} 456 457// Alloc a pair of core registers, or a double. 458RegStorage X86Mir2Lir::AllocTypedTempWide(bool fp_hint, int reg_class) { 459 if (((reg_class == kAnyReg) && fp_hint) || (reg_class == kFPReg)) { 460 return AllocTempDouble(); 461 } 462 RegStorage low_reg = AllocTemp(); 463 RegStorage high_reg = AllocTemp(); 464 return RegStorage::MakeRegPair(low_reg, high_reg); 465} 466 467RegStorage X86Mir2Lir::AllocTypedTemp(bool fp_hint, int reg_class) { 468 if (((reg_class == kAnyReg) && fp_hint) || (reg_class == kFPReg)) { 469 return AllocTempSingle(); 470 } 471 return AllocTemp(); 472} 473 474void X86Mir2Lir::CompilerInitializeRegAlloc() { 475 reg_pool_ = new (arena_) RegisterPool(this, arena_, core_regs, sp_regs, dp_regs, reserved_regs, 476 core_temps, sp_temps, dp_temps); 477 478 // Target-specific adjustments. 479 480 // Alias single precision xmm to double xmms. 481 // TODO: as needed, add larger vector sizes - alias all to the largest. 482 GrowableArray<RegisterInfo*>::Iterator it(®_pool_->sp_regs_); 483 for (RegisterInfo* info = it.Next(); info != nullptr; info = it.Next()) { 484 int sp_reg_num = info->GetReg().GetRegNum(); 485 RegStorage dp_reg = RegStorage::Solo64(RegStorage::kFloatingPoint | sp_reg_num); 486 RegisterInfo* dp_reg_info = GetRegInfo(dp_reg); 487 // 64-bit xmm vector register's master storage should refer to itself. 488 DCHECK_EQ(dp_reg_info, dp_reg_info->Master()); 489 // Redirect 32-bit vector's master storage to 64-bit vector. 490 info->SetMaster(dp_reg_info); 491 } 492 493 // Don't start allocating temps at r0/s0/d0 or you may clobber return regs in early-exit methods. 494 // TODO: adjust for x86/hard float calling convention. 495 reg_pool_->next_core_reg_ = 2; 496 reg_pool_->next_sp_reg_ = 2; 497 reg_pool_->next_dp_reg_ = 1; 498} 499 500void X86Mir2Lir::FreeRegLocTemps(RegLocation rl_keep, RegLocation rl_free) { 501 DCHECK(rl_keep.wide); 502 DCHECK(rl_free.wide); 503 int free_low = rl_free.reg.GetLowReg(); 504 int free_high = rl_free.reg.GetHighReg(); 505 int keep_low = rl_keep.reg.GetLowReg(); 506 int keep_high = rl_keep.reg.GetHighReg(); 507 if ((free_low != keep_low) && (free_low != keep_high) && 508 (free_high != keep_low) && (free_high != keep_high)) { 509 // No overlap, free both 510 FreeTemp(rl_free.reg); 511 } 512} 513 514void X86Mir2Lir::SpillCoreRegs() { 515 if (num_core_spills_ == 0) { 516 return; 517 } 518 // Spill mask not including fake return address register 519 uint32_t mask = core_spill_mask_ & ~(1 << rs_rRET.GetRegNum()); 520 int offset = frame_size_ - (4 * num_core_spills_); 521 for (int reg = 0; mask; mask >>= 1, reg++) { 522 if (mask & 0x1) { 523 StoreWordDisp(rs_rX86_SP, offset, RegStorage::Solo32(reg)); 524 offset += 4; 525 } 526 } 527} 528 529void X86Mir2Lir::UnSpillCoreRegs() { 530 if (num_core_spills_ == 0) { 531 return; 532 } 533 // Spill mask not including fake return address register 534 uint32_t mask = core_spill_mask_ & ~(1 << rs_rRET.GetRegNum()); 535 int offset = frame_size_ - (4 * num_core_spills_); 536 for (int reg = 0; mask; mask >>= 1, reg++) { 537 if (mask & 0x1) { 538 LoadWordDisp(rs_rX86_SP, offset, RegStorage::Solo32(reg)); 539 offset += 4; 540 } 541 } 542} 543 544bool X86Mir2Lir::IsUnconditionalBranch(LIR* lir) { 545 return (lir->opcode == kX86Jmp8 || lir->opcode == kX86Jmp32); 546} 547 548bool X86Mir2Lir::SupportsVolatileLoadStore(OpSize size) { 549 return true; 550} 551 552RegisterClass X86Mir2Lir::RegClassForFieldLoadStore(OpSize size, bool is_volatile) { 553 if (UNLIKELY(is_volatile)) { 554 // On x86, atomic 64-bit load/store requires an fp register. 555 // Smaller aligned load/store is atomic for both core and fp registers. 556 if (size == k64 || size == kDouble) { 557 return kFPReg; 558 } 559 } 560 return RegClassBySize(size); 561} 562 563X86Mir2Lir::X86Mir2Lir(CompilationUnit* cu, MIRGraph* mir_graph, ArenaAllocator* arena) 564 : Mir2Lir(cu, mir_graph, arena), 565 base_of_code_(nullptr), store_method_addr_(false), store_method_addr_used_(false), 566 method_address_insns_(arena, 100, kGrowableArrayMisc), 567 class_type_address_insns_(arena, 100, kGrowableArrayMisc), 568 call_method_insns_(arena, 100, kGrowableArrayMisc), 569 stack_decrement_(nullptr), stack_increment_(nullptr) { 570 if (kIsDebugBuild) { 571 for (int i = 0; i < kX86Last; i++) { 572 if (X86Mir2Lir::EncodingMap[i].opcode != i) { 573 LOG(FATAL) << "Encoding order for " << X86Mir2Lir::EncodingMap[i].name 574 << " is wrong: expecting " << i << ", seeing " 575 << static_cast<int>(X86Mir2Lir::EncodingMap[i].opcode); 576 } 577 } 578 } 579} 580 581Mir2Lir* X86CodeGenerator(CompilationUnit* const cu, MIRGraph* const mir_graph, 582 ArenaAllocator* const arena) { 583 return new X86Mir2Lir(cu, mir_graph, arena); 584} 585 586// Not used in x86 587RegStorage X86Mir2Lir::LoadHelper(ThreadOffset<4> offset) { 588 LOG(FATAL) << "Unexpected use of LoadHelper in x86"; 589 return RegStorage::InvalidReg(); 590} 591 592LIR* X86Mir2Lir::CheckSuspendUsingLoad() { 593 LOG(FATAL) << "Unexpected use of CheckSuspendUsingLoad in x86"; 594 return nullptr; 595} 596 597uint64_t X86Mir2Lir::GetTargetInstFlags(int opcode) { 598 DCHECK(!IsPseudoLirOp(opcode)); 599 return X86Mir2Lir::EncodingMap[opcode].flags; 600} 601 602const char* X86Mir2Lir::GetTargetInstName(int opcode) { 603 DCHECK(!IsPseudoLirOp(opcode)); 604 return X86Mir2Lir::EncodingMap[opcode].name; 605} 606 607const char* X86Mir2Lir::GetTargetInstFmt(int opcode) { 608 DCHECK(!IsPseudoLirOp(opcode)); 609 return X86Mir2Lir::EncodingMap[opcode].fmt; 610} 611 612void X86Mir2Lir::GenConstWide(RegLocation rl_dest, int64_t value) { 613 // Can we do this directly to memory? 614 rl_dest = UpdateLocWide(rl_dest); 615 if ((rl_dest.location == kLocDalvikFrame) || 616 (rl_dest.location == kLocCompilerTemp)) { 617 int32_t val_lo = Low32Bits(value); 618 int32_t val_hi = High32Bits(value); 619 int r_base = TargetReg(kSp).GetReg(); 620 int displacement = SRegOffset(rl_dest.s_reg_low); 621 622 LIR * store = NewLIR3(kX86Mov32MI, r_base, displacement + LOWORD_OFFSET, val_lo); 623 AnnotateDalvikRegAccess(store, (displacement + LOWORD_OFFSET) >> 2, 624 false /* is_load */, true /* is64bit */); 625 store = NewLIR3(kX86Mov32MI, r_base, displacement + HIWORD_OFFSET, val_hi); 626 AnnotateDalvikRegAccess(store, (displacement + HIWORD_OFFSET) >> 2, 627 false /* is_load */, true /* is64bit */); 628 return; 629 } 630 631 // Just use the standard code to do the generation. 632 Mir2Lir::GenConstWide(rl_dest, value); 633} 634 635// TODO: Merge with existing RegLocation dumper in vreg_analysis.cc 636void X86Mir2Lir::DumpRegLocation(RegLocation loc) { 637 LOG(INFO) << "location: " << loc.location << ',' 638 << (loc.wide ? " w" : " ") 639 << (loc.defined ? " D" : " ") 640 << (loc.is_const ? " c" : " ") 641 << (loc.fp ? " F" : " ") 642 << (loc.core ? " C" : " ") 643 << (loc.ref ? " r" : " ") 644 << (loc.high_word ? " h" : " ") 645 << (loc.home ? " H" : " ") 646 << ", low: " << static_cast<int>(loc.reg.GetLowReg()) 647 << ", high: " << static_cast<int>(loc.reg.GetHighReg()) 648 << ", s_reg: " << loc.s_reg_low 649 << ", orig: " << loc.orig_sreg; 650} 651 652void X86Mir2Lir::Materialize() { 653 // A good place to put the analysis before starting. 654 AnalyzeMIR(); 655 656 // Now continue with regular code generation. 657 Mir2Lir::Materialize(); 658} 659 660void X86Mir2Lir::LoadMethodAddress(const MethodReference& target_method, InvokeType type, 661 SpecialTargetRegister symbolic_reg) { 662 /* 663 * For x86, just generate a 32 bit move immediate instruction, that will be filled 664 * in at 'link time'. For now, put a unique value based on target to ensure that 665 * code deduplication works. 666 */ 667 int target_method_idx = target_method.dex_method_index; 668 const DexFile* target_dex_file = target_method.dex_file; 669 const DexFile::MethodId& target_method_id = target_dex_file->GetMethodId(target_method_idx); 670 uintptr_t target_method_id_ptr = reinterpret_cast<uintptr_t>(&target_method_id); 671 672 // Generate the move instruction with the unique pointer and save index, dex_file, and type. 673 LIR *move = RawLIR(current_dalvik_offset_, kX86Mov32RI, TargetReg(symbolic_reg).GetReg(), 674 static_cast<int>(target_method_id_ptr), target_method_idx, 675 WrapPointer(const_cast<DexFile*>(target_dex_file)), type); 676 AppendLIR(move); 677 method_address_insns_.Insert(move); 678} 679 680void X86Mir2Lir::LoadClassType(uint32_t type_idx, SpecialTargetRegister symbolic_reg) { 681 /* 682 * For x86, just generate a 32 bit move immediate instruction, that will be filled 683 * in at 'link time'. For now, put a unique value based on target to ensure that 684 * code deduplication works. 685 */ 686 const DexFile::TypeId& id = cu_->dex_file->GetTypeId(type_idx); 687 uintptr_t ptr = reinterpret_cast<uintptr_t>(&id); 688 689 // Generate the move instruction with the unique pointer and save index and type. 690 LIR *move = RawLIR(current_dalvik_offset_, kX86Mov32RI, TargetReg(symbolic_reg).GetReg(), 691 static_cast<int>(ptr), type_idx); 692 AppendLIR(move); 693 class_type_address_insns_.Insert(move); 694} 695 696LIR *X86Mir2Lir::CallWithLinkerFixup(const MethodReference& target_method, InvokeType type) { 697 /* 698 * For x86, just generate a 32 bit call relative instruction, that will be filled 699 * in at 'link time'. For now, put a unique value based on target to ensure that 700 * code deduplication works. 701 */ 702 int target_method_idx = target_method.dex_method_index; 703 const DexFile* target_dex_file = target_method.dex_file; 704 const DexFile::MethodId& target_method_id = target_dex_file->GetMethodId(target_method_idx); 705 uintptr_t target_method_id_ptr = reinterpret_cast<uintptr_t>(&target_method_id); 706 707 // Generate the call instruction with the unique pointer and save index, dex_file, and type. 708 LIR *call = RawLIR(current_dalvik_offset_, kX86CallI, static_cast<int>(target_method_id_ptr), 709 target_method_idx, WrapPointer(const_cast<DexFile*>(target_dex_file)), type); 710 AppendLIR(call); 711 call_method_insns_.Insert(call); 712 return call; 713} 714 715void X86Mir2Lir::InstallLiteralPools() { 716 // These are handled differently for x86. 717 DCHECK(code_literal_list_ == nullptr); 718 DCHECK(method_literal_list_ == nullptr); 719 DCHECK(class_literal_list_ == nullptr); 720 721 // Handle the fixups for methods. 722 for (uint32_t i = 0; i < method_address_insns_.Size(); i++) { 723 LIR* p = method_address_insns_.Get(i); 724 DCHECK_EQ(p->opcode, kX86Mov32RI); 725 uint32_t target_method_idx = p->operands[2]; 726 const DexFile* target_dex_file = 727 reinterpret_cast<const DexFile*>(UnwrapPointer(p->operands[3])); 728 729 // The offset to patch is the last 4 bytes of the instruction. 730 int patch_offset = p->offset + p->flags.size - 4; 731 cu_->compiler_driver->AddMethodPatch(cu_->dex_file, cu_->class_def_idx, 732 cu_->method_idx, cu_->invoke_type, 733 target_method_idx, target_dex_file, 734 static_cast<InvokeType>(p->operands[4]), 735 patch_offset); 736 } 737 738 // Handle the fixups for class types. 739 for (uint32_t i = 0; i < class_type_address_insns_.Size(); i++) { 740 LIR* p = class_type_address_insns_.Get(i); 741 DCHECK_EQ(p->opcode, kX86Mov32RI); 742 uint32_t target_method_idx = p->operands[2]; 743 744 // The offset to patch is the last 4 bytes of the instruction. 745 int patch_offset = p->offset + p->flags.size - 4; 746 cu_->compiler_driver->AddClassPatch(cu_->dex_file, cu_->class_def_idx, 747 cu_->method_idx, target_method_idx, patch_offset); 748 } 749 750 // And now the PC-relative calls to methods. 751 for (uint32_t i = 0; i < call_method_insns_.Size(); i++) { 752 LIR* p = call_method_insns_.Get(i); 753 DCHECK_EQ(p->opcode, kX86CallI); 754 uint32_t target_method_idx = p->operands[1]; 755 const DexFile* target_dex_file = 756 reinterpret_cast<const DexFile*>(UnwrapPointer(p->operands[2])); 757 758 // The offset to patch is the last 4 bytes of the instruction. 759 int patch_offset = p->offset + p->flags.size - 4; 760 cu_->compiler_driver->AddRelativeCodePatch(cu_->dex_file, cu_->class_def_idx, 761 cu_->method_idx, cu_->invoke_type, 762 target_method_idx, target_dex_file, 763 static_cast<InvokeType>(p->operands[3]), 764 patch_offset, -4 /* offset */); 765 } 766 767 // And do the normal processing. 768 Mir2Lir::InstallLiteralPools(); 769} 770 771/* 772 * Fast string.index_of(I) & (II). Inline check for simple case of char <= 0xffff, 773 * otherwise bails to standard library code. 774 */ 775bool X86Mir2Lir::GenInlinedIndexOf(CallInfo* info, bool zero_based) { 776 ClobberCallerSave(); 777 LockCallTemps(); // Using fixed registers 778 779 // EAX: 16 bit character being searched. 780 // ECX: count: number of words to be searched. 781 // EDI: String being searched. 782 // EDX: temporary during execution. 783 // EBX: temporary during execution. 784 785 RegLocation rl_obj = info->args[0]; 786 RegLocation rl_char = info->args[1]; 787 RegLocation rl_start; // Note: only present in III flavor or IndexOf. 788 789 uint32_t char_value = 790 rl_char.is_const ? mir_graph_->ConstantValue(rl_char.orig_sreg) : 0; 791 792 if (char_value > 0xFFFF) { 793 // We have to punt to the real String.indexOf. 794 return false; 795 } 796 797 // Okay, we are commited to inlining this. 798 RegLocation rl_return = GetReturn(false); 799 RegLocation rl_dest = InlineTarget(info); 800 801 // Is the string non-NULL? 802 LoadValueDirectFixed(rl_obj, rs_rDX); 803 GenNullCheck(rs_rDX, info->opt_flags); 804 info->opt_flags |= MIR_IGNORE_NULL_CHECK; // Record that we've null checked. 805 806 // Does the character fit in 16 bits? 807 LIR* slowpath_branch = nullptr; 808 if (rl_char.is_const) { 809 // We need the value in EAX. 810 LoadConstantNoClobber(rs_rAX, char_value); 811 } else { 812 // Character is not a constant; compare at runtime. 813 LoadValueDirectFixed(rl_char, rs_rAX); 814 slowpath_branch = OpCmpImmBranch(kCondGt, rs_rAX, 0xFFFF, nullptr); 815 } 816 817 // From here down, we know that we are looking for a char that fits in 16 bits. 818 // Location of reference to data array within the String object. 819 int value_offset = mirror::String::ValueOffset().Int32Value(); 820 // Location of count within the String object. 821 int count_offset = mirror::String::CountOffset().Int32Value(); 822 // Starting offset within data array. 823 int offset_offset = mirror::String::OffsetOffset().Int32Value(); 824 // Start of char data with array_. 825 int data_offset = mirror::Array::DataOffset(sizeof(uint16_t)).Int32Value(); 826 827 // Character is in EAX. 828 // Object pointer is in EDX. 829 830 // We need to preserve EDI, but have no spare registers, so push it on the stack. 831 // We have to remember that all stack addresses after this are offset by sizeof(EDI). 832 NewLIR1(kX86Push32R, rs_rDI.GetReg()); 833 834 // Compute the number of words to search in to rCX. 835 Load32Disp(rs_rDX, count_offset, rs_rCX); 836 LIR *length_compare = nullptr; 837 int start_value = 0; 838 bool is_index_on_stack = false; 839 if (zero_based) { 840 // We have to handle an empty string. Use special instruction JECXZ. 841 length_compare = NewLIR0(kX86Jecxz8); 842 } else { 843 rl_start = info->args[2]; 844 // We have to offset by the start index. 845 if (rl_start.is_const) { 846 start_value = mir_graph_->ConstantValue(rl_start.orig_sreg); 847 start_value = std::max(start_value, 0); 848 849 // Is the start > count? 850 length_compare = OpCmpImmBranch(kCondLe, rs_rCX, start_value, nullptr); 851 852 if (start_value != 0) { 853 OpRegImm(kOpSub, rs_rCX, start_value); 854 } 855 } else { 856 // Runtime start index. 857 rl_start = UpdateLoc(rl_start); 858 if (rl_start.location == kLocPhysReg) { 859 // Handle "start index < 0" case. 860 OpRegReg(kOpXor, rs_rBX, rs_rBX); 861 OpRegReg(kOpCmp, rl_start.reg, rs_rBX); 862 OpCondRegReg(kOpCmov, kCondLt, rl_start.reg, rs_rBX); 863 864 // The length of the string should be greater than the start index. 865 length_compare = OpCmpBranch(kCondLe, rs_rCX, rl_start.reg, nullptr); 866 OpRegReg(kOpSub, rs_rCX, rl_start.reg); 867 if (rl_start.reg == rs_rDI) { 868 // The special case. We will use EDI further, so lets put start index to stack. 869 NewLIR1(kX86Push32R, rs_rDI.GetReg()); 870 is_index_on_stack = true; 871 } 872 } else { 873 // Load the start index from stack, remembering that we pushed EDI. 874 int displacement = SRegOffset(rl_start.s_reg_low) + sizeof(uint32_t); 875 Load32Disp(rs_rX86_SP, displacement, rs_rBX); 876 OpRegReg(kOpXor, rs_rDI, rs_rDI); 877 OpRegReg(kOpCmp, rs_rBX, rs_rDI); 878 OpCondRegReg(kOpCmov, kCondLt, rs_rBX, rs_rDI); 879 880 length_compare = OpCmpBranch(kCondLe, rs_rCX, rs_rBX, nullptr); 881 OpRegReg(kOpSub, rs_rCX, rs_rBX); 882 // Put the start index to stack. 883 NewLIR1(kX86Push32R, rs_rBX.GetReg()); 884 is_index_on_stack = true; 885 } 886 } 887 } 888 DCHECK(length_compare != nullptr); 889 890 // ECX now contains the count in words to be searched. 891 892 // Load the address of the string into EBX. 893 // The string starts at VALUE(String) + 2 * OFFSET(String) + DATA_OFFSET. 894 Load32Disp(rs_rDX, value_offset, rs_rDI); 895 Load32Disp(rs_rDX, offset_offset, rs_rBX); 896 OpLea(rs_rBX, rs_rDI, rs_rBX, 1, data_offset); 897 898 // Now compute into EDI where the search will start. 899 if (zero_based || rl_start.is_const) { 900 if (start_value == 0) { 901 OpRegCopy(rs_rDI, rs_rBX); 902 } else { 903 NewLIR3(kX86Lea32RM, rs_rDI.GetReg(), rs_rBX.GetReg(), 2 * start_value); 904 } 905 } else { 906 if (is_index_on_stack == true) { 907 // Load the start index from stack. 908 NewLIR1(kX86Pop32R, rs_rDX.GetReg()); 909 OpLea(rs_rDI, rs_rBX, rs_rDX, 1, 0); 910 } else { 911 OpLea(rs_rDI, rs_rBX, rl_start.reg, 1, 0); 912 } 913 } 914 915 // EDI now contains the start of the string to be searched. 916 // We are all prepared to do the search for the character. 917 NewLIR0(kX86RepneScasw); 918 919 // Did we find a match? 920 LIR* failed_branch = OpCondBranch(kCondNe, nullptr); 921 922 // yes, we matched. Compute the index of the result. 923 // index = ((curr_ptr - orig_ptr) / 2) - 1. 924 OpRegReg(kOpSub, rs_rDI, rs_rBX); 925 OpRegImm(kOpAsr, rs_rDI, 1); 926 NewLIR3(kX86Lea32RM, rl_return.reg.GetReg(), rs_rDI.GetReg(), -1); 927 LIR *all_done = NewLIR1(kX86Jmp8, 0); 928 929 // Failed to match; return -1. 930 LIR *not_found = NewLIR0(kPseudoTargetLabel); 931 length_compare->target = not_found; 932 failed_branch->target = not_found; 933 LoadConstantNoClobber(rl_return.reg, -1); 934 935 // And join up at the end. 936 all_done->target = NewLIR0(kPseudoTargetLabel); 937 // Restore EDI from the stack. 938 NewLIR1(kX86Pop32R, rs_rDI.GetReg()); 939 940 // Out of line code returns here. 941 if (slowpath_branch != nullptr) { 942 LIR *return_point = NewLIR0(kPseudoTargetLabel); 943 AddIntrinsicSlowPath(info, slowpath_branch, return_point); 944 } 945 946 StoreValue(rl_dest, rl_return); 947 return true; 948} 949 950/* 951 * @brief Enter a 32 bit quantity into the FDE buffer 952 * @param buf FDE buffer. 953 * @param data Data value. 954 */ 955static void PushWord(std::vector<uint8_t>&buf, int data) { 956 buf.push_back(data & 0xff); 957 buf.push_back((data >> 8) & 0xff); 958 buf.push_back((data >> 16) & 0xff); 959 buf.push_back((data >> 24) & 0xff); 960} 961 962/* 963 * @brief Enter an 'advance LOC' into the FDE buffer 964 * @param buf FDE buffer. 965 * @param increment Amount by which to increase the current location. 966 */ 967static void AdvanceLoc(std::vector<uint8_t>&buf, uint32_t increment) { 968 if (increment < 64) { 969 // Encoding in opcode. 970 buf.push_back(0x1 << 6 | increment); 971 } else if (increment < 256) { 972 // Single byte delta. 973 buf.push_back(0x02); 974 buf.push_back(increment); 975 } else if (increment < 256 * 256) { 976 // Two byte delta. 977 buf.push_back(0x03); 978 buf.push_back(increment & 0xff); 979 buf.push_back((increment >> 8) & 0xff); 980 } else { 981 // Four byte delta. 982 buf.push_back(0x04); 983 PushWord(buf, increment); 984 } 985} 986 987 988std::vector<uint8_t>* X86CFIInitialization() { 989 return X86Mir2Lir::ReturnCommonCallFrameInformation(); 990} 991 992std::vector<uint8_t>* X86Mir2Lir::ReturnCommonCallFrameInformation() { 993 std::vector<uint8_t>*cfi_info = new std::vector<uint8_t>; 994 995 // Length of the CIE (except for this field). 996 PushWord(*cfi_info, 16); 997 998 // CIE id. 999 PushWord(*cfi_info, 0xFFFFFFFFU); 1000 1001 // Version: 3. 1002 cfi_info->push_back(0x03); 1003 1004 // Augmentation: empty string. 1005 cfi_info->push_back(0x0); 1006 1007 // Code alignment: 1. 1008 cfi_info->push_back(0x01); 1009 1010 // Data alignment: -4. 1011 cfi_info->push_back(0x7C); 1012 1013 // Return address register (R8). 1014 cfi_info->push_back(0x08); 1015 1016 // Initial return PC is 4(ESP): DW_CFA_def_cfa R4 4. 1017 cfi_info->push_back(0x0C); 1018 cfi_info->push_back(0x04); 1019 cfi_info->push_back(0x04); 1020 1021 // Return address location: 0(SP): DW_CFA_offset R8 1 (* -4);. 1022 cfi_info->push_back(0x2 << 6 | 0x08); 1023 cfi_info->push_back(0x01); 1024 1025 // And 2 Noops to align to 4 byte boundary. 1026 cfi_info->push_back(0x0); 1027 cfi_info->push_back(0x0); 1028 1029 DCHECK_EQ(cfi_info->size() & 3, 0U); 1030 return cfi_info; 1031} 1032 1033static void EncodeUnsignedLeb128(std::vector<uint8_t>& buf, uint32_t value) { 1034 uint8_t buffer[12]; 1035 uint8_t *ptr = EncodeUnsignedLeb128(buffer, value); 1036 for (uint8_t *p = buffer; p < ptr; p++) { 1037 buf.push_back(*p); 1038 } 1039} 1040 1041std::vector<uint8_t>* X86Mir2Lir::ReturnCallFrameInformation() { 1042 std::vector<uint8_t>*cfi_info = new std::vector<uint8_t>; 1043 1044 // Generate the FDE for the method. 1045 DCHECK_NE(data_offset_, 0U); 1046 1047 // Length (will be filled in later in this routine). 1048 PushWord(*cfi_info, 0); 1049 1050 // CIE_pointer (can be filled in by linker); might be left at 0 if there is only 1051 // one CIE for the whole debug_frame section. 1052 PushWord(*cfi_info, 0); 1053 1054 // 'initial_location' (filled in by linker). 1055 PushWord(*cfi_info, 0); 1056 1057 // 'address_range' (number of bytes in the method). 1058 PushWord(*cfi_info, data_offset_); 1059 1060 // The instructions in the FDE. 1061 if (stack_decrement_ != nullptr) { 1062 // Advance LOC to just past the stack decrement. 1063 uint32_t pc = NEXT_LIR(stack_decrement_)->offset; 1064 AdvanceLoc(*cfi_info, pc); 1065 1066 // Now update the offset to the call frame: DW_CFA_def_cfa_offset frame_size. 1067 cfi_info->push_back(0x0e); 1068 EncodeUnsignedLeb128(*cfi_info, frame_size_); 1069 1070 // We continue with that stack until the epilogue. 1071 if (stack_increment_ != nullptr) { 1072 uint32_t new_pc = NEXT_LIR(stack_increment_)->offset; 1073 AdvanceLoc(*cfi_info, new_pc - pc); 1074 1075 // We probably have code snippets after the epilogue, so save the 1076 // current state: DW_CFA_remember_state. 1077 cfi_info->push_back(0x0a); 1078 1079 // We have now popped the stack: DW_CFA_def_cfa_offset 4. There is only the return 1080 // PC on the stack now. 1081 cfi_info->push_back(0x0e); 1082 EncodeUnsignedLeb128(*cfi_info, 4); 1083 1084 // Everything after that is the same as before the epilogue. 1085 // Stack bump was followed by RET instruction. 1086 LIR *post_ret_insn = NEXT_LIR(NEXT_LIR(stack_increment_)); 1087 if (post_ret_insn != nullptr) { 1088 pc = new_pc; 1089 new_pc = post_ret_insn->offset; 1090 AdvanceLoc(*cfi_info, new_pc - pc); 1091 // Restore the state: DW_CFA_restore_state. 1092 cfi_info->push_back(0x0b); 1093 } 1094 } 1095 } 1096 1097 // Padding to a multiple of 4 1098 while ((cfi_info->size() & 3) != 0) { 1099 // DW_CFA_nop is encoded as 0. 1100 cfi_info->push_back(0); 1101 } 1102 1103 // Set the length of the FDE inside the generated bytes. 1104 uint32_t length = cfi_info->size() - 4; 1105 (*cfi_info)[0] = length; 1106 (*cfi_info)[1] = length >> 8; 1107 (*cfi_info)[2] = length >> 16; 1108 (*cfi_info)[3] = length >> 24; 1109 return cfi_info; 1110} 1111 1112} // namespace art 1113