target_x86.cc revision 55d0eac918321e0525f6e6491f36a80977e0d416
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 "codegen_x86.h" 18#include "dex/compiler_internals.h" 19#include "dex/quick/mir_to_lir-inl.h" 20#include "x86_lir.h" 21 22#include <string> 23 24namespace art { 25 26// FIXME: restore "static" when usage uncovered 27/*static*/ int core_regs[] = { 28 rAX, rCX, rDX, rBX, rX86_SP, rBP, rSI, rDI 29#ifdef TARGET_REX_SUPPORT 30 r8, r9, r10, r11, r12, r13, r14, 15 31#endif 32}; 33/*static*/ int ReservedRegs[] = {rX86_SP}; 34/*static*/ int core_temps[] = {rAX, rCX, rDX, rBX}; 35/*static*/ int FpRegs[] = { 36 fr0, fr1, fr2, fr3, fr4, fr5, fr6, fr7, 37#ifdef TARGET_REX_SUPPORT 38 fr8, fr9, fr10, fr11, fr12, fr13, fr14, fr15 39#endif 40}; 41/*static*/ int fp_temps[] = { 42 fr0, fr1, fr2, fr3, fr4, fr5, fr6, fr7, 43#ifdef TARGET_REX_SUPPORT 44 fr8, fr9, fr10, fr11, fr12, fr13, fr14, fr15 45#endif 46}; 47 48RegLocation X86Mir2Lir::LocCReturn() { 49 RegLocation res = X86_LOC_C_RETURN; 50 return res; 51} 52 53RegLocation X86Mir2Lir::LocCReturnWide() { 54 RegLocation res = X86_LOC_C_RETURN_WIDE; 55 return res; 56} 57 58RegLocation X86Mir2Lir::LocCReturnFloat() { 59 RegLocation res = X86_LOC_C_RETURN_FLOAT; 60 return res; 61} 62 63RegLocation X86Mir2Lir::LocCReturnDouble() { 64 RegLocation res = X86_LOC_C_RETURN_DOUBLE; 65 return res; 66} 67 68// Return a target-dependent special register. 69int X86Mir2Lir::TargetReg(SpecialTargetRegister reg) { 70 int res = INVALID_REG; 71 switch (reg) { 72 case kSelf: res = rX86_SELF; break; 73 case kSuspend: res = rX86_SUSPEND; break; 74 case kLr: res = rX86_LR; break; 75 case kPc: res = rX86_PC; break; 76 case kSp: res = rX86_SP; break; 77 case kArg0: res = rX86_ARG0; break; 78 case kArg1: res = rX86_ARG1; break; 79 case kArg2: res = rX86_ARG2; break; 80 case kArg3: res = rX86_ARG3; break; 81 case kFArg0: res = rX86_FARG0; break; 82 case kFArg1: res = rX86_FARG1; break; 83 case kFArg2: res = rX86_FARG2; break; 84 case kFArg3: res = rX86_FARG3; break; 85 case kRet0: res = rX86_RET0; break; 86 case kRet1: res = rX86_RET1; break; 87 case kInvokeTgt: res = rX86_INVOKE_TGT; break; 88 case kHiddenArg: res = rAX; break; 89 case kHiddenFpArg: res = fr0; break; 90 case kCount: res = rX86_COUNT; break; 91 } 92 return res; 93} 94 95// Create a double from a pair of singles. 96int X86Mir2Lir::S2d(int low_reg, int high_reg) { 97 return X86_S2D(low_reg, high_reg); 98} 99 100// Return mask to strip off fp reg flags and bias. 101uint32_t X86Mir2Lir::FpRegMask() { 102 return X86_FP_REG_MASK; 103} 104 105// True if both regs single, both core or both double. 106bool X86Mir2Lir::SameRegType(int reg1, int reg2) { 107 return (X86_REGTYPE(reg1) == X86_REGTYPE(reg2)); 108} 109 110/* 111 * Decode the register id. 112 */ 113uint64_t X86Mir2Lir::GetRegMaskCommon(int reg) { 114 uint64_t seed; 115 int shift; 116 int reg_id; 117 118 reg_id = reg & 0xf; 119 /* Double registers in x86 are just a single FP register */ 120 seed = 1; 121 /* FP register starts at bit position 16 */ 122 shift = X86_FPREG(reg) ? kX86FPReg0 : 0; 123 /* Expand the double register id into single offset */ 124 shift += reg_id; 125 return (seed << shift); 126} 127 128uint64_t X86Mir2Lir::GetPCUseDefEncoding() { 129 /* 130 * FIXME: might make sense to use a virtual resource encoding bit for pc. Might be 131 * able to clean up some of the x86/Arm_Mips differences 132 */ 133 LOG(FATAL) << "Unexpected call to GetPCUseDefEncoding for x86"; 134 return 0ULL; 135} 136 137void X86Mir2Lir::SetupTargetResourceMasks(LIR* lir, uint64_t flags) { 138 DCHECK_EQ(cu_->instruction_set, kX86); 139 DCHECK(!lir->flags.use_def_invalid); 140 141 // X86-specific resource map setup here. 142 if (flags & REG_USE_SP) { 143 lir->u.m.use_mask |= ENCODE_X86_REG_SP; 144 } 145 146 if (flags & REG_DEF_SP) { 147 lir->u.m.def_mask |= ENCODE_X86_REG_SP; 148 } 149 150 if (flags & REG_DEFA) { 151 SetupRegMask(&lir->u.m.def_mask, rAX); 152 } 153 154 if (flags & REG_DEFD) { 155 SetupRegMask(&lir->u.m.def_mask, rDX); 156 } 157 if (flags & REG_USEA) { 158 SetupRegMask(&lir->u.m.use_mask, rAX); 159 } 160 161 if (flags & REG_USEC) { 162 SetupRegMask(&lir->u.m.use_mask, rCX); 163 } 164 165 if (flags & REG_USED) { 166 SetupRegMask(&lir->u.m.use_mask, rDX); 167 } 168 169 if (flags & REG_USEB) { 170 SetupRegMask(&lir->u.m.use_mask, rBX); 171 } 172} 173 174/* For dumping instructions */ 175static const char* x86RegName[] = { 176 "rax", "rcx", "rdx", "rbx", "rsp", "rbp", "rsi", "rdi", 177 "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15" 178}; 179 180static const char* x86CondName[] = { 181 "O", 182 "NO", 183 "B/NAE/C", 184 "NB/AE/NC", 185 "Z/EQ", 186 "NZ/NE", 187 "BE/NA", 188 "NBE/A", 189 "S", 190 "NS", 191 "P/PE", 192 "NP/PO", 193 "L/NGE", 194 "NL/GE", 195 "LE/NG", 196 "NLE/G" 197}; 198 199/* 200 * Interpret a format string and build a string no longer than size 201 * See format key in Assemble.cc. 202 */ 203std::string X86Mir2Lir::BuildInsnString(const char *fmt, LIR *lir, unsigned char* base_addr) { 204 std::string buf; 205 size_t i = 0; 206 size_t fmt_len = strlen(fmt); 207 while (i < fmt_len) { 208 if (fmt[i] != '!') { 209 buf += fmt[i]; 210 i++; 211 } else { 212 i++; 213 DCHECK_LT(i, fmt_len); 214 char operand_number_ch = fmt[i]; 215 i++; 216 if (operand_number_ch == '!') { 217 buf += "!"; 218 } else { 219 int operand_number = operand_number_ch - '0'; 220 DCHECK_LT(operand_number, 6); // Expect upto 6 LIR operands. 221 DCHECK_LT(i, fmt_len); 222 int operand = lir->operands[operand_number]; 223 switch (fmt[i]) { 224 case 'c': 225 DCHECK_LT(static_cast<size_t>(operand), sizeof(x86CondName)); 226 buf += x86CondName[operand]; 227 break; 228 case 'd': 229 buf += StringPrintf("%d", operand); 230 break; 231 case 'p': { 232 EmbeddedData *tab_rec = reinterpret_cast<EmbeddedData*>(UnwrapPointer(operand)); 233 buf += StringPrintf("0x%08x", tab_rec->offset); 234 break; 235 } 236 case 'r': 237 if (X86_FPREG(operand) || X86_DOUBLEREG(operand)) { 238 int fp_reg = operand & X86_FP_REG_MASK; 239 buf += StringPrintf("xmm%d", fp_reg); 240 } else { 241 DCHECK_LT(static_cast<size_t>(operand), sizeof(x86RegName)); 242 buf += x86RegName[operand]; 243 } 244 break; 245 case 't': 246 buf += StringPrintf("0x%08" PRIxPTR " (L%p)", 247 reinterpret_cast<uintptr_t>(base_addr) + lir->offset + operand, 248 lir->target); 249 break; 250 default: 251 buf += StringPrintf("DecodeError '%c'", fmt[i]); 252 break; 253 } 254 i++; 255 } 256 } 257 } 258 return buf; 259} 260 261void X86Mir2Lir::DumpResourceMask(LIR *x86LIR, uint64_t mask, const char *prefix) { 262 char buf[256]; 263 buf[0] = 0; 264 265 if (mask == ENCODE_ALL) { 266 strcpy(buf, "all"); 267 } else { 268 char num[8]; 269 int i; 270 271 for (i = 0; i < kX86RegEnd; i++) { 272 if (mask & (1ULL << i)) { 273 snprintf(num, arraysize(num), "%d ", i); 274 strcat(buf, num); 275 } 276 } 277 278 if (mask & ENCODE_CCODE) { 279 strcat(buf, "cc "); 280 } 281 /* Memory bits */ 282 if (x86LIR && (mask & ENCODE_DALVIK_REG)) { 283 snprintf(buf + strlen(buf), arraysize(buf) - strlen(buf), "dr%d%s", 284 DECODE_ALIAS_INFO_REG(x86LIR->flags.alias_info), 285 (DECODE_ALIAS_INFO_WIDE(x86LIR->flags.alias_info)) ? "(+1)" : ""); 286 } 287 if (mask & ENCODE_LITERAL) { 288 strcat(buf, "lit "); 289 } 290 291 if (mask & ENCODE_HEAP_REF) { 292 strcat(buf, "heap "); 293 } 294 if (mask & ENCODE_MUST_NOT_ALIAS) { 295 strcat(buf, "noalias "); 296 } 297 } 298 if (buf[0]) { 299 LOG(INFO) << prefix << ": " << buf; 300 } 301} 302 303void X86Mir2Lir::AdjustSpillMask() { 304 // Adjustment for LR spilling, x86 has no LR so nothing to do here 305 core_spill_mask_ |= (1 << rRET); 306 num_core_spills_++; 307} 308 309/* 310 * Mark a callee-save fp register as promoted. Note that 311 * vpush/vpop uses contiguous register lists so we must 312 * include any holes in the mask. Associate holes with 313 * Dalvik register INVALID_VREG (0xFFFFU). 314 */ 315void X86Mir2Lir::MarkPreservedSingle(int v_reg, int reg) { 316 UNIMPLEMENTED(WARNING) << "MarkPreservedSingle"; 317#if 0 318 LOG(FATAL) << "No support yet for promoted FP regs"; 319#endif 320} 321 322void X86Mir2Lir::FlushRegWide(int reg1, int reg2) { 323 RegisterInfo* info1 = GetRegInfo(reg1); 324 RegisterInfo* info2 = GetRegInfo(reg2); 325 DCHECK(info1 && info2 && info1->pair && info2->pair && 326 (info1->partner == info2->reg) && 327 (info2->partner == info1->reg)); 328 if ((info1->live && info1->dirty) || (info2->live && info2->dirty)) { 329 if (!(info1->is_temp && info2->is_temp)) { 330 /* Should not happen. If it does, there's a problem in eval_loc */ 331 LOG(FATAL) << "Long half-temp, half-promoted"; 332 } 333 334 info1->dirty = false; 335 info2->dirty = false; 336 if (mir_graph_->SRegToVReg(info2->s_reg) < mir_graph_->SRegToVReg(info1->s_reg)) 337 info1 = info2; 338 int v_reg = mir_graph_->SRegToVReg(info1->s_reg); 339 StoreBaseDispWide(rX86_SP, VRegOffset(v_reg), info1->reg, info1->partner); 340 } 341} 342 343void X86Mir2Lir::FlushReg(int reg) { 344 RegisterInfo* info = GetRegInfo(reg); 345 if (info->live && info->dirty) { 346 info->dirty = false; 347 int v_reg = mir_graph_->SRegToVReg(info->s_reg); 348 StoreBaseDisp(rX86_SP, VRegOffset(v_reg), reg, kWord); 349 } 350} 351 352/* Give access to the target-dependent FP register encoding to common code */ 353bool X86Mir2Lir::IsFpReg(int reg) { 354 return X86_FPREG(reg); 355} 356 357/* Clobber all regs that might be used by an external C call */ 358void X86Mir2Lir::ClobberCallerSave() { 359 Clobber(rAX); 360 Clobber(rCX); 361 Clobber(rDX); 362 Clobber(rBX); 363} 364 365RegLocation X86Mir2Lir::GetReturnWideAlt() { 366 RegLocation res = LocCReturnWide(); 367 CHECK(res.low_reg == rAX); 368 CHECK(res.high_reg == rDX); 369 Clobber(rAX); 370 Clobber(rDX); 371 MarkInUse(rAX); 372 MarkInUse(rDX); 373 MarkPair(res.low_reg, res.high_reg); 374 return res; 375} 376 377RegLocation X86Mir2Lir::GetReturnAlt() { 378 RegLocation res = LocCReturn(); 379 res.low_reg = rDX; 380 Clobber(rDX); 381 MarkInUse(rDX); 382 return res; 383} 384 385/* To be used when explicitly managing register use */ 386void X86Mir2Lir::LockCallTemps() { 387 LockTemp(rX86_ARG0); 388 LockTemp(rX86_ARG1); 389 LockTemp(rX86_ARG2); 390 LockTemp(rX86_ARG3); 391} 392 393/* To be used when explicitly managing register use */ 394void X86Mir2Lir::FreeCallTemps() { 395 FreeTemp(rX86_ARG0); 396 FreeTemp(rX86_ARG1); 397 FreeTemp(rX86_ARG2); 398 FreeTemp(rX86_ARG3); 399} 400 401void X86Mir2Lir::GenMemBarrier(MemBarrierKind barrier_kind) { 402#if ANDROID_SMP != 0 403 // TODO: optimize fences 404 NewLIR0(kX86Mfence); 405#endif 406} 407/* 408 * Alloc a pair of core registers, or a double. Low reg in low byte, 409 * high reg in next byte. 410 */ 411int X86Mir2Lir::AllocTypedTempPair(bool fp_hint, 412 int reg_class) { 413 int high_reg; 414 int low_reg; 415 int res = 0; 416 417 if (((reg_class == kAnyReg) && fp_hint) || (reg_class == kFPReg)) { 418 low_reg = AllocTempDouble(); 419 high_reg = low_reg; // only one allocated! 420 res = (low_reg & 0xff) | ((high_reg & 0xff) << 8); 421 return res; 422 } 423 424 low_reg = AllocTemp(); 425 high_reg = AllocTemp(); 426 res = (low_reg & 0xff) | ((high_reg & 0xff) << 8); 427 return res; 428} 429 430int X86Mir2Lir::AllocTypedTemp(bool fp_hint, int reg_class) { 431 if (((reg_class == kAnyReg) && fp_hint) || (reg_class == kFPReg)) { 432 return AllocTempFloat(); 433 } 434 return AllocTemp(); 435} 436 437void X86Mir2Lir::CompilerInitializeRegAlloc() { 438 int num_regs = sizeof(core_regs)/sizeof(*core_regs); 439 int num_reserved = sizeof(ReservedRegs)/sizeof(*ReservedRegs); 440 int num_temps = sizeof(core_temps)/sizeof(*core_temps); 441 int num_fp_regs = sizeof(FpRegs)/sizeof(*FpRegs); 442 int num_fp_temps = sizeof(fp_temps)/sizeof(*fp_temps); 443 reg_pool_ = static_cast<RegisterPool*>(arena_->Alloc(sizeof(*reg_pool_), 444 ArenaAllocator::kAllocRegAlloc)); 445 reg_pool_->num_core_regs = num_regs; 446 reg_pool_->core_regs = 447 static_cast<RegisterInfo*>(arena_->Alloc(num_regs * sizeof(*reg_pool_->core_regs), 448 ArenaAllocator::kAllocRegAlloc)); 449 reg_pool_->num_fp_regs = num_fp_regs; 450 reg_pool_->FPRegs = 451 static_cast<RegisterInfo *>(arena_->Alloc(num_fp_regs * sizeof(*reg_pool_->FPRegs), 452 ArenaAllocator::kAllocRegAlloc)); 453 CompilerInitPool(reg_pool_->core_regs, core_regs, reg_pool_->num_core_regs); 454 CompilerInitPool(reg_pool_->FPRegs, FpRegs, reg_pool_->num_fp_regs); 455 // Keep special registers from being allocated 456 for (int i = 0; i < num_reserved; i++) { 457 MarkInUse(ReservedRegs[i]); 458 } 459 // Mark temp regs - all others not in use can be used for promotion 460 for (int i = 0; i < num_temps; i++) { 461 MarkTemp(core_temps[i]); 462 } 463 for (int i = 0; i < num_fp_temps; i++) { 464 MarkTemp(fp_temps[i]); 465 } 466} 467 468void X86Mir2Lir::FreeRegLocTemps(RegLocation rl_keep, 469 RegLocation rl_free) { 470 if ((rl_free.low_reg != rl_keep.low_reg) && (rl_free.low_reg != rl_keep.high_reg) && 471 (rl_free.high_reg != rl_keep.low_reg) && (rl_free.high_reg != rl_keep.high_reg)) { 472 // No overlap, free both 473 FreeTemp(rl_free.low_reg); 474 FreeTemp(rl_free.high_reg); 475 } 476} 477 478void X86Mir2Lir::SpillCoreRegs() { 479 if (num_core_spills_ == 0) { 480 return; 481 } 482 // Spill mask not including fake return address register 483 uint32_t mask = core_spill_mask_ & ~(1 << rRET); 484 int offset = frame_size_ - (4 * num_core_spills_); 485 for (int reg = 0; mask; mask >>= 1, reg++) { 486 if (mask & 0x1) { 487 StoreWordDisp(rX86_SP, offset, reg); 488 offset += 4; 489 } 490 } 491} 492 493void X86Mir2Lir::UnSpillCoreRegs() { 494 if (num_core_spills_ == 0) { 495 return; 496 } 497 // Spill mask not including fake return address register 498 uint32_t mask = core_spill_mask_ & ~(1 << rRET); 499 int offset = frame_size_ - (4 * num_core_spills_); 500 for (int reg = 0; mask; mask >>= 1, reg++) { 501 if (mask & 0x1) { 502 LoadWordDisp(rX86_SP, offset, reg); 503 offset += 4; 504 } 505 } 506} 507 508bool X86Mir2Lir::IsUnconditionalBranch(LIR* lir) { 509 return (lir->opcode == kX86Jmp8 || lir->opcode == kX86Jmp32); 510} 511 512X86Mir2Lir::X86Mir2Lir(CompilationUnit* cu, MIRGraph* mir_graph, ArenaAllocator* arena) 513 : Mir2Lir(cu, mir_graph, arena), 514 method_address_insns_(arena, 100, kGrowableArrayMisc), 515 class_type_address_insns_(arena, 100, kGrowableArrayMisc), 516 call_method_insns_(arena, 100, kGrowableArrayMisc) { 517 store_method_addr_used_ = false; 518 for (int i = 0; i < kX86Last; i++) { 519 if (X86Mir2Lir::EncodingMap[i].opcode != i) { 520 LOG(FATAL) << "Encoding order for " << X86Mir2Lir::EncodingMap[i].name 521 << " is wrong: expecting " << i << ", seeing " 522 << static_cast<int>(X86Mir2Lir::EncodingMap[i].opcode); 523 } 524 } 525} 526 527Mir2Lir* X86CodeGenerator(CompilationUnit* const cu, MIRGraph* const mir_graph, 528 ArenaAllocator* const arena) { 529 return new X86Mir2Lir(cu, mir_graph, arena); 530} 531 532// Not used in x86 533int X86Mir2Lir::LoadHelper(ThreadOffset offset) { 534 LOG(FATAL) << "Unexpected use of LoadHelper in x86"; 535 return INVALID_REG; 536} 537 538uint64_t X86Mir2Lir::GetTargetInstFlags(int opcode) { 539 DCHECK(!IsPseudoLirOp(opcode)); 540 return X86Mir2Lir::EncodingMap[opcode].flags; 541} 542 543const char* X86Mir2Lir::GetTargetInstName(int opcode) { 544 DCHECK(!IsPseudoLirOp(opcode)); 545 return X86Mir2Lir::EncodingMap[opcode].name; 546} 547 548const char* X86Mir2Lir::GetTargetInstFmt(int opcode) { 549 DCHECK(!IsPseudoLirOp(opcode)); 550 return X86Mir2Lir::EncodingMap[opcode].fmt; 551} 552 553/* 554 * Return an updated location record with current in-register status. 555 * If the value lives in live temps, reflect that fact. No code 556 * is generated. If the live value is part of an older pair, 557 * clobber both low and high. 558 */ 559// TODO: Reunify with common code after 'pair mess' has been fixed 560RegLocation X86Mir2Lir::UpdateLocWide(RegLocation loc) { 561 DCHECK(loc.wide); 562 DCHECK(CheckCorePoolSanity()); 563 if (loc.location != kLocPhysReg) { 564 DCHECK((loc.location == kLocDalvikFrame) || 565 (loc.location == kLocCompilerTemp)); 566 // Are the dalvik regs already live in physical registers? 567 RegisterInfo* info_lo = AllocLive(loc.s_reg_low, kAnyReg); 568 569 // Handle FP registers specially on x86. 570 if (info_lo && IsFpReg(info_lo->reg)) { 571 bool match = true; 572 573 // We can't match a FP register with a pair of Core registers. 574 match = match && (info_lo->pair == 0); 575 576 if (match) { 577 // We can reuse;update the register usage info. 578 loc.low_reg = info_lo->reg; 579 loc.high_reg = info_lo->reg; // Play nice with existing code. 580 loc.location = kLocPhysReg; 581 loc.vec_len = kVectorLength8; 582 DCHECK(IsFpReg(loc.low_reg)); 583 return loc; 584 } 585 // We can't easily reuse; clobber and free any overlaps. 586 if (info_lo) { 587 Clobber(info_lo->reg); 588 FreeTemp(info_lo->reg); 589 if (info_lo->pair) 590 Clobber(info_lo->partner); 591 } 592 } else { 593 RegisterInfo* info_hi = AllocLive(GetSRegHi(loc.s_reg_low), kAnyReg); 594 bool match = true; 595 match = match && (info_lo != NULL); 596 match = match && (info_hi != NULL); 597 // Are they both core or both FP? 598 match = match && (IsFpReg(info_lo->reg) == IsFpReg(info_hi->reg)); 599 // If a pair of floating point singles, are they properly aligned? 600 if (match && IsFpReg(info_lo->reg)) { 601 match &= ((info_lo->reg & 0x1) == 0); 602 match &= ((info_hi->reg - info_lo->reg) == 1); 603 } 604 // If previously used as a pair, it is the same pair? 605 if (match && (info_lo->pair || info_hi->pair)) { 606 match = (info_lo->pair == info_hi->pair); 607 match &= ((info_lo->reg == info_hi->partner) && 608 (info_hi->reg == info_lo->partner)); 609 } 610 if (match) { 611 // Can reuse - update the register usage info 612 loc.low_reg = info_lo->reg; 613 loc.high_reg = info_hi->reg; 614 loc.location = kLocPhysReg; 615 MarkPair(loc.low_reg, loc.high_reg); 616 DCHECK(!IsFpReg(loc.low_reg) || ((loc.low_reg & 0x1) == 0)); 617 return loc; 618 } 619 // Can't easily reuse - clobber and free any overlaps 620 if (info_lo) { 621 Clobber(info_lo->reg); 622 FreeTemp(info_lo->reg); 623 if (info_lo->pair) 624 Clobber(info_lo->partner); 625 } 626 if (info_hi) { 627 Clobber(info_hi->reg); 628 FreeTemp(info_hi->reg); 629 if (info_hi->pair) 630 Clobber(info_hi->partner); 631 } 632 } 633 } 634 return loc; 635} 636 637// TODO: Reunify with common code after 'pair mess' has been fixed 638RegLocation X86Mir2Lir::EvalLocWide(RegLocation loc, int reg_class, bool update) { 639 DCHECK(loc.wide); 640 int32_t new_regs; 641 int32_t low_reg; 642 int32_t high_reg; 643 644 loc = UpdateLocWide(loc); 645 646 /* If it is already in a register, we can assume proper form. Is it the right reg class? */ 647 if (loc.location == kLocPhysReg) { 648 DCHECK_EQ(IsFpReg(loc.low_reg), loc.IsVectorScalar()); 649 if (!RegClassMatches(reg_class, loc.low_reg)) { 650 /* It is the wrong register class. Reallocate and copy. */ 651 if (!IsFpReg(loc.low_reg)) { 652 // We want this in a FP reg, and it is in core registers. 653 DCHECK(reg_class != kCoreReg); 654 // Allocate this into any FP reg, and mark it with the right size. 655 low_reg = AllocTypedTemp(true, reg_class); 656 OpVectorRegCopyWide(low_reg, loc.low_reg, loc.high_reg); 657 CopyRegInfo(low_reg, loc.low_reg); 658 Clobber(loc.low_reg); 659 Clobber(loc.high_reg); 660 loc.low_reg = low_reg; 661 loc.high_reg = low_reg; // Play nice with existing code. 662 loc.vec_len = kVectorLength8; 663 } else { 664 // The value is in a FP register, and we want it in a pair of core registers. 665 DCHECK_EQ(reg_class, kCoreReg); 666 DCHECK_EQ(loc.low_reg, loc.high_reg); 667 new_regs = AllocTypedTempPair(false, kCoreReg); // Force to core registers. 668 low_reg = new_regs & 0xff; 669 high_reg = (new_regs >> 8) & 0xff; 670 DCHECK_NE(low_reg, high_reg); 671 OpRegCopyWide(low_reg, high_reg, loc.low_reg, loc.high_reg); 672 CopyRegInfo(low_reg, loc.low_reg); 673 CopyRegInfo(high_reg, loc.high_reg); 674 Clobber(loc.low_reg); 675 Clobber(loc.high_reg); 676 loc.low_reg = low_reg; 677 loc.high_reg = high_reg; 678 MarkPair(loc.low_reg, loc.high_reg); 679 DCHECK(!IsFpReg(loc.low_reg) || ((loc.low_reg & 0x1) == 0)); 680 } 681 } 682 return loc; 683 } 684 685 DCHECK_NE(loc.s_reg_low, INVALID_SREG); 686 DCHECK_NE(GetSRegHi(loc.s_reg_low), INVALID_SREG); 687 688 new_regs = AllocTypedTempPair(loc.fp, reg_class); 689 loc.low_reg = new_regs & 0xff; 690 loc.high_reg = (new_regs >> 8) & 0xff; 691 692 if (loc.low_reg == loc.high_reg) { 693 DCHECK(IsFpReg(loc.low_reg)); 694 loc.vec_len = kVectorLength8; 695 } else { 696 MarkPair(loc.low_reg, loc.high_reg); 697 } 698 if (update) { 699 loc.location = kLocPhysReg; 700 MarkLive(loc.low_reg, loc.s_reg_low); 701 if (loc.low_reg != loc.high_reg) { 702 MarkLive(loc.high_reg, GetSRegHi(loc.s_reg_low)); 703 } 704 } 705 return loc; 706} 707 708// TODO: Reunify with common code after 'pair mess' has been fixed 709RegLocation X86Mir2Lir::EvalLoc(RegLocation loc, int reg_class, bool update) { 710 int new_reg; 711 712 if (loc.wide) 713 return EvalLocWide(loc, reg_class, update); 714 715 loc = UpdateLoc(loc); 716 717 if (loc.location == kLocPhysReg) { 718 if (!RegClassMatches(reg_class, loc.low_reg)) { 719 /* Wrong register class. Realloc, copy and transfer ownership. */ 720 new_reg = AllocTypedTemp(loc.fp, reg_class); 721 OpRegCopy(new_reg, loc.low_reg); 722 CopyRegInfo(new_reg, loc.low_reg); 723 Clobber(loc.low_reg); 724 loc.low_reg = new_reg; 725 if (IsFpReg(loc.low_reg) && reg_class != kCoreReg) 726 loc.vec_len = kVectorLength4; 727 } 728 return loc; 729 } 730 731 DCHECK_NE(loc.s_reg_low, INVALID_SREG); 732 733 new_reg = AllocTypedTemp(loc.fp, reg_class); 734 loc.low_reg = new_reg; 735 if (IsFpReg(loc.low_reg) && reg_class != kCoreReg) 736 loc.vec_len = kVectorLength4; 737 738 if (update) { 739 loc.location = kLocPhysReg; 740 MarkLive(loc.low_reg, loc.s_reg_low); 741 } 742 return loc; 743} 744 745int X86Mir2Lir::AllocTempDouble() { 746 // We really don't need a pair of registers. 747 return AllocTempFloat(); 748} 749 750// TODO: Reunify with common code after 'pair mess' has been fixed 751void X86Mir2Lir::ResetDefLocWide(RegLocation rl) { 752 DCHECK(rl.wide); 753 RegisterInfo* p_low = IsTemp(rl.low_reg); 754 if (IsFpReg(rl.low_reg)) { 755 // We are using only the low register. 756 if (p_low && !(cu_->disable_opt & (1 << kSuppressLoads))) { 757 NullifyRange(p_low->def_start, p_low->def_end, p_low->s_reg, rl.s_reg_low); 758 } 759 ResetDef(rl.low_reg); 760 } else { 761 RegisterInfo* p_high = IsTemp(rl.high_reg); 762 if (p_low && !(cu_->disable_opt & (1 << kSuppressLoads))) { 763 DCHECK(p_low->pair); 764 NullifyRange(p_low->def_start, p_low->def_end, p_low->s_reg, rl.s_reg_low); 765 } 766 if (p_high && !(cu_->disable_opt & (1 << kSuppressLoads))) { 767 DCHECK(p_high->pair); 768 } 769 ResetDef(rl.low_reg); 770 ResetDef(rl.high_reg); 771 } 772} 773 774void X86Mir2Lir::GenConstWide(RegLocation rl_dest, int64_t value) { 775 // Can we do this directly to memory? 776 rl_dest = UpdateLocWide(rl_dest); 777 if ((rl_dest.location == kLocDalvikFrame) || 778 (rl_dest.location == kLocCompilerTemp)) { 779 int32_t val_lo = Low32Bits(value); 780 int32_t val_hi = High32Bits(value); 781 int rBase = TargetReg(kSp); 782 int displacement = SRegOffset(rl_dest.s_reg_low); 783 784 LIR * store = NewLIR3(kX86Mov32MI, rBase, displacement + LOWORD_OFFSET, val_lo); 785 AnnotateDalvikRegAccess(store, (displacement + LOWORD_OFFSET) >> 2, 786 false /* is_load */, true /* is64bit */); 787 store = NewLIR3(kX86Mov32MI, rBase, displacement + HIWORD_OFFSET, val_hi); 788 AnnotateDalvikRegAccess(store, (displacement + HIWORD_OFFSET) >> 2, 789 false /* is_load */, true /* is64bit */); 790 return; 791 } 792 793 // Just use the standard code to do the generation. 794 Mir2Lir::GenConstWide(rl_dest, value); 795} 796 797// TODO: Merge with existing RegLocation dumper in vreg_analysis.cc 798void X86Mir2Lir::DumpRegLocation(RegLocation loc) { 799 LOG(INFO) << "location: " << loc.location << ',' 800 << (loc.wide ? " w" : " ") 801 << (loc.defined ? " D" : " ") 802 << (loc.is_const ? " c" : " ") 803 << (loc.fp ? " F" : " ") 804 << (loc.core ? " C" : " ") 805 << (loc.ref ? " r" : " ") 806 << (loc.high_word ? " h" : " ") 807 << (loc.home ? " H" : " ") 808 << " vec_len: " << loc.vec_len 809 << ", low: " << static_cast<int>(loc.low_reg) 810 << ", high: " << static_cast<int>(loc.high_reg) 811 << ", s_reg: " << loc.s_reg_low 812 << ", orig: " << loc.orig_sreg; 813} 814 815void X86Mir2Lir::Materialize() { 816 // A good place to put the analysis before starting. 817 AnalyzeMIR(); 818 819 // Now continue with regular code generation. 820 Mir2Lir::Materialize(); 821} 822 823void X86Mir2Lir::LoadMethodAddress(int dex_method_index, InvokeType type, 824 SpecialTargetRegister symbolic_reg) { 825 /* 826 * For x86, just generate a 32 bit move immediate instruction, that will be filled 827 * in at 'link time'. For now, put a unique value based on target to ensure that 828 * code deduplication works. 829 */ 830 const DexFile::MethodId& id = cu_->dex_file->GetMethodId(dex_method_index); 831 uintptr_t ptr = reinterpret_cast<uintptr_t>(&id); 832 833 // Generate the move instruction with the unique pointer and save index and type. 834 LIR *move = RawLIR(current_dalvik_offset_, kX86Mov32RI, TargetReg(symbolic_reg), 835 static_cast<int>(ptr), dex_method_index, type); 836 AppendLIR(move); 837 method_address_insns_.Insert(move); 838} 839 840void X86Mir2Lir::LoadClassType(uint32_t type_idx, SpecialTargetRegister symbolic_reg) { 841 /* 842 * For x86, just generate a 32 bit move immediate instruction, that will be filled 843 * in at 'link time'. For now, put a unique value based on target to ensure that 844 * code deduplication works. 845 */ 846 const DexFile::TypeId& id = cu_->dex_file->GetTypeId(type_idx); 847 uintptr_t ptr = reinterpret_cast<uintptr_t>(&id); 848 849 // Generate the move instruction with the unique pointer and save index and type. 850 LIR *move = RawLIR(current_dalvik_offset_, kX86Mov32RI, TargetReg(symbolic_reg), 851 static_cast<int>(ptr), type_idx); 852 AppendLIR(move); 853 class_type_address_insns_.Insert(move); 854} 855 856LIR *X86Mir2Lir::CallWithLinkerFixup(int dex_method_index, InvokeType type) { 857 /* 858 * For x86, just generate a 32 bit call relative instruction, that will be filled 859 * in at 'link time'. For now, put a unique value based on target to ensure that 860 * code deduplication works. 861 */ 862 const DexFile::MethodId& id = cu_->dex_file->GetMethodId(dex_method_index); 863 uintptr_t ptr = reinterpret_cast<uintptr_t>(&id); 864 865 // Generate the call instruction with the unique pointer and save index and type. 866 LIR *call = RawLIR(current_dalvik_offset_, kX86CallI, static_cast<int>(ptr), dex_method_index, 867 type); 868 AppendLIR(call); 869 call_method_insns_.Insert(call); 870 return call; 871} 872 873void X86Mir2Lir::InstallLiteralPools() { 874 // These are handled differently for x86. 875 DCHECK(code_literal_list_ == nullptr); 876 DCHECK(method_literal_list_ == nullptr); 877 DCHECK(class_literal_list_ == nullptr); 878 879 // Handle the fixups for methods. 880 for (uint32_t i = 0; i < method_address_insns_.Size(); i++) { 881 LIR* p = method_address_insns_.Get(i); 882 DCHECK_EQ(p->opcode, kX86Mov32RI); 883 uint32_t target = p->operands[2]; 884 885 // The offset to patch is the last 4 bytes of the instruction. 886 int patch_offset = p->offset + p->flags.size - 4; 887 cu_->compiler_driver->AddMethodPatch(cu_->dex_file, cu_->class_def_idx, 888 cu_->method_idx, cu_->invoke_type, 889 target, static_cast<InvokeType>(p->operands[3]), 890 patch_offset); 891 } 892 893 // Handle the fixups for class types. 894 for (uint32_t i = 0; i < class_type_address_insns_.Size(); i++) { 895 LIR* p = class_type_address_insns_.Get(i); 896 DCHECK_EQ(p->opcode, kX86Mov32RI); 897 uint32_t target = p->operands[2]; 898 899 // The offset to patch is the last 4 bytes of the instruction. 900 int patch_offset = p->offset + p->flags.size - 4; 901 cu_->compiler_driver->AddClassPatch(cu_->dex_file, cu_->class_def_idx, 902 cu_->method_idx, target, patch_offset); 903 } 904 905 // And now the PC-relative calls to methods. 906 for (uint32_t i = 0; i < call_method_insns_.Size(); i++) { 907 LIR* p = call_method_insns_.Get(i); 908 DCHECK_EQ(p->opcode, kX86CallI); 909 uint32_t target = p->operands[1]; 910 911 // The offset to patch is the last 4 bytes of the instruction. 912 int patch_offset = p->offset + p->flags.size - 4; 913 cu_->compiler_driver->AddRelativeCodePatch(cu_->dex_file, cu_->class_def_idx, 914 cu_->method_idx, cu_->invoke_type, target, 915 static_cast<InvokeType>(p->operands[2]), 916 patch_offset, -4 /* offset */); 917 } 918 919 // And do the normal processing. 920 Mir2Lir::InstallLiteralPools(); 921} 922 923} // namespace art 924