target_x86.cc revision b5874a47a4d2c4d2971116b031b4068021ffda05
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 "dex/reg_storage_eq.h" 24#include "mirror/array.h" 25#include "mirror/string.h" 26#include "x86_lir.h" 27 28namespace art { 29 30static constexpr RegStorage core_regs_arr_32[] = { 31 rs_rAX, rs_rCX, rs_rDX, rs_rBX, rs_rX86_SP_32, rs_rBP, rs_rSI, rs_rDI, 32}; 33static constexpr RegStorage core_regs_arr_64[] = { 34 rs_rAX, rs_rCX, rs_rDX, rs_rBX, rs_rX86_SP_32, rs_rBP, rs_rSI, rs_rDI, 35 rs_r8, rs_r9, rs_r10, rs_r11, rs_r12, rs_r13, rs_r14, rs_r15 36}; 37static constexpr RegStorage core_regs_arr_64q[] = { 38 rs_r0q, rs_r1q, rs_r2q, rs_r3q, rs_rX86_SP_64, rs_r5q, rs_r6q, rs_r7q, 39 rs_r8q, rs_r9q, rs_r10q, rs_r11q, rs_r12q, rs_r13q, rs_r14q, rs_r15q 40}; 41static constexpr RegStorage sp_regs_arr_32[] = { 42 rs_fr0, rs_fr1, rs_fr2, rs_fr3, rs_fr4, rs_fr5, rs_fr6, rs_fr7, 43}; 44static constexpr RegStorage sp_regs_arr_64[] = { 45 rs_fr0, rs_fr1, rs_fr2, rs_fr3, rs_fr4, rs_fr5, rs_fr6, rs_fr7, 46 rs_fr8, rs_fr9, rs_fr10, rs_fr11, rs_fr12, rs_fr13, rs_fr14, rs_fr15 47}; 48static constexpr RegStorage dp_regs_arr_32[] = { 49 rs_dr0, rs_dr1, rs_dr2, rs_dr3, rs_dr4, rs_dr5, rs_dr6, rs_dr7, 50}; 51static constexpr RegStorage dp_regs_arr_64[] = { 52 rs_dr0, rs_dr1, rs_dr2, rs_dr3, rs_dr4, rs_dr5, rs_dr6, rs_dr7, 53 rs_dr8, rs_dr9, rs_dr10, rs_dr11, rs_dr12, rs_dr13, rs_dr14, rs_dr15 54}; 55static constexpr RegStorage xp_regs_arr_32[] = { 56 rs_xr0, rs_xr1, rs_xr2, rs_xr3, rs_xr4, rs_xr5, rs_xr6, rs_xr7, 57}; 58static constexpr RegStorage xp_regs_arr_64[] = { 59 rs_xr0, rs_xr1, rs_xr2, rs_xr3, rs_xr4, rs_xr5, rs_xr6, rs_xr7, 60 rs_xr8, rs_xr9, rs_xr10, rs_xr11, rs_xr12, rs_xr13, rs_xr14, rs_xr15 61}; 62static constexpr RegStorage reserved_regs_arr_32[] = {rs_rX86_SP_32}; 63static constexpr RegStorage reserved_regs_arr_64[] = {rs_rX86_SP_32}; 64static constexpr RegStorage reserved_regs_arr_64q[] = {rs_rX86_SP_64}; 65static constexpr RegStorage core_temps_arr_32[] = {rs_rAX, rs_rCX, rs_rDX, rs_rBX}; 66static constexpr RegStorage core_temps_arr_64[] = { 67 rs_rAX, rs_rCX, rs_rDX, rs_rSI, rs_rDI, 68 rs_r8, rs_r9, rs_r10, rs_r11 69}; 70 71// How to add register to be available for promotion: 72// 1) Remove register from array defining temp 73// 2) Update ClobberCallerSave 74// 3) Update JNI compiler ABI: 75// 3.1) add reg in JniCallingConvention method 76// 3.2) update CoreSpillMask/FpSpillMask 77// 4) Update entrypoints 78// 4.1) Update constants in asm_support_x86_64.h for new frame size 79// 4.2) Remove entry in SmashCallerSaves 80// 4.3) Update jni_entrypoints to spill/unspill new callee save reg 81// 4.4) Update quick_entrypoints to spill/unspill new callee save reg 82// 5) Update runtime ABI 83// 5.1) Update quick_method_frame_info with new required spills 84// 5.2) Update QuickArgumentVisitor with new offsets to gprs and xmms 85// Note that you cannot use register corresponding to incoming args 86// according to ABI and QCG needs one additional XMM temp for 87// bulk copy in preparation to call. 88static constexpr RegStorage core_temps_arr_64q[] = { 89 rs_r0q, rs_r1q, rs_r2q, rs_r6q, rs_r7q, 90 rs_r8q, rs_r9q, rs_r10q, rs_r11q 91}; 92static constexpr RegStorage sp_temps_arr_32[] = { 93 rs_fr0, rs_fr1, rs_fr2, rs_fr3, rs_fr4, rs_fr5, rs_fr6, rs_fr7, 94}; 95static constexpr RegStorage sp_temps_arr_64[] = { 96 rs_fr0, rs_fr1, rs_fr2, rs_fr3, rs_fr4, rs_fr5, rs_fr6, rs_fr7, 97 rs_fr8, rs_fr9, rs_fr10, rs_fr11 98}; 99static constexpr RegStorage dp_temps_arr_32[] = { 100 rs_dr0, rs_dr1, rs_dr2, rs_dr3, rs_dr4, rs_dr5, rs_dr6, rs_dr7, 101}; 102static constexpr RegStorage dp_temps_arr_64[] = { 103 rs_dr0, rs_dr1, rs_dr2, rs_dr3, rs_dr4, rs_dr5, rs_dr6, rs_dr7, 104 rs_dr8, rs_dr9, rs_dr10, rs_dr11 105}; 106 107static constexpr RegStorage xp_temps_arr_32[] = { 108 rs_xr0, rs_xr1, rs_xr2, rs_xr3, rs_xr4, rs_xr5, rs_xr6, rs_xr7, 109}; 110static constexpr RegStorage xp_temps_arr_64[] = { 111 rs_xr0, rs_xr1, rs_xr2, rs_xr3, rs_xr4, rs_xr5, rs_xr6, rs_xr7, 112 rs_xr8, rs_xr9, rs_xr10, rs_xr11 113}; 114 115static constexpr ArrayRef<const RegStorage> empty_pool; 116static constexpr ArrayRef<const RegStorage> core_regs_32(core_regs_arr_32); 117static constexpr ArrayRef<const RegStorage> core_regs_64(core_regs_arr_64); 118static constexpr ArrayRef<const RegStorage> core_regs_64q(core_regs_arr_64q); 119static constexpr ArrayRef<const RegStorage> sp_regs_32(sp_regs_arr_32); 120static constexpr ArrayRef<const RegStorage> sp_regs_64(sp_regs_arr_64); 121static constexpr ArrayRef<const RegStorage> dp_regs_32(dp_regs_arr_32); 122static constexpr ArrayRef<const RegStorage> dp_regs_64(dp_regs_arr_64); 123static constexpr ArrayRef<const RegStorage> xp_regs_32(xp_regs_arr_32); 124static constexpr ArrayRef<const RegStorage> xp_regs_64(xp_regs_arr_64); 125static constexpr ArrayRef<const RegStorage> reserved_regs_32(reserved_regs_arr_32); 126static constexpr ArrayRef<const RegStorage> reserved_regs_64(reserved_regs_arr_64); 127static constexpr ArrayRef<const RegStorage> reserved_regs_64q(reserved_regs_arr_64q); 128static constexpr ArrayRef<const RegStorage> core_temps_32(core_temps_arr_32); 129static constexpr ArrayRef<const RegStorage> core_temps_64(core_temps_arr_64); 130static constexpr ArrayRef<const RegStorage> core_temps_64q(core_temps_arr_64q); 131static constexpr ArrayRef<const RegStorage> sp_temps_32(sp_temps_arr_32); 132static constexpr ArrayRef<const RegStorage> sp_temps_64(sp_temps_arr_64); 133static constexpr ArrayRef<const RegStorage> dp_temps_32(dp_temps_arr_32); 134static constexpr ArrayRef<const RegStorage> dp_temps_64(dp_temps_arr_64); 135 136static constexpr ArrayRef<const RegStorage> xp_temps_32(xp_temps_arr_32); 137static constexpr ArrayRef<const RegStorage> xp_temps_64(xp_temps_arr_64); 138 139RegStorage rs_rX86_SP; 140 141X86NativeRegisterPool rX86_ARG0; 142X86NativeRegisterPool rX86_ARG1; 143X86NativeRegisterPool rX86_ARG2; 144X86NativeRegisterPool rX86_ARG3; 145X86NativeRegisterPool rX86_ARG4; 146X86NativeRegisterPool rX86_ARG5; 147X86NativeRegisterPool rX86_FARG0; 148X86NativeRegisterPool rX86_FARG1; 149X86NativeRegisterPool rX86_FARG2; 150X86NativeRegisterPool rX86_FARG3; 151X86NativeRegisterPool rX86_FARG4; 152X86NativeRegisterPool rX86_FARG5; 153X86NativeRegisterPool rX86_FARG6; 154X86NativeRegisterPool rX86_FARG7; 155X86NativeRegisterPool rX86_RET0; 156X86NativeRegisterPool rX86_RET1; 157X86NativeRegisterPool rX86_INVOKE_TGT; 158X86NativeRegisterPool rX86_COUNT; 159 160RegStorage rs_rX86_ARG0; 161RegStorage rs_rX86_ARG1; 162RegStorage rs_rX86_ARG2; 163RegStorage rs_rX86_ARG3; 164RegStorage rs_rX86_ARG4; 165RegStorage rs_rX86_ARG5; 166RegStorage rs_rX86_FARG0; 167RegStorage rs_rX86_FARG1; 168RegStorage rs_rX86_FARG2; 169RegStorage rs_rX86_FARG3; 170RegStorage rs_rX86_FARG4; 171RegStorage rs_rX86_FARG5; 172RegStorage rs_rX86_FARG6; 173RegStorage rs_rX86_FARG7; 174RegStorage rs_rX86_RET0; 175RegStorage rs_rX86_RET1; 176RegStorage rs_rX86_INVOKE_TGT; 177RegStorage rs_rX86_COUNT; 178 179RegLocation X86Mir2Lir::LocCReturn() { 180 return x86_loc_c_return; 181} 182 183RegLocation X86Mir2Lir::LocCReturnRef() { 184 return cu_->target64 ? x86_64_loc_c_return_ref : x86_loc_c_return_ref; 185} 186 187RegLocation X86Mir2Lir::LocCReturnWide() { 188 return cu_->target64 ? x86_64_loc_c_return_wide : x86_loc_c_return_wide; 189} 190 191RegLocation X86Mir2Lir::LocCReturnFloat() { 192 return x86_loc_c_return_float; 193} 194 195RegLocation X86Mir2Lir::LocCReturnDouble() { 196 return x86_loc_c_return_double; 197} 198 199// Return a target-dependent special register for 32-bit. 200RegStorage X86Mir2Lir::TargetReg32(SpecialTargetRegister reg) { 201 RegStorage res_reg = RegStorage::InvalidReg(); 202 switch (reg) { 203 case kSelf: res_reg = RegStorage::InvalidReg(); break; 204 case kSuspend: res_reg = RegStorage::InvalidReg(); break; 205 case kLr: res_reg = RegStorage::InvalidReg(); break; 206 case kPc: res_reg = RegStorage::InvalidReg(); break; 207 case kSp: res_reg = rs_rX86_SP_32; break; // This must be the concrete one, as _SP is target- 208 // specific size. 209 case kArg0: res_reg = rs_rX86_ARG0; break; 210 case kArg1: res_reg = rs_rX86_ARG1; break; 211 case kArg2: res_reg = rs_rX86_ARG2; break; 212 case kArg3: res_reg = rs_rX86_ARG3; break; 213 case kArg4: res_reg = rs_rX86_ARG4; break; 214 case kArg5: res_reg = rs_rX86_ARG5; break; 215 case kFArg0: res_reg = rs_rX86_FARG0; break; 216 case kFArg1: res_reg = rs_rX86_FARG1; break; 217 case kFArg2: res_reg = rs_rX86_FARG2; break; 218 case kFArg3: res_reg = rs_rX86_FARG3; break; 219 case kFArg4: res_reg = rs_rX86_FARG4; break; 220 case kFArg5: res_reg = rs_rX86_FARG5; break; 221 case kFArg6: res_reg = rs_rX86_FARG6; break; 222 case kFArg7: res_reg = rs_rX86_FARG7; break; 223 case kRet0: res_reg = rs_rX86_RET0; break; 224 case kRet1: res_reg = rs_rX86_RET1; break; 225 case kInvokeTgt: res_reg = rs_rX86_INVOKE_TGT; break; 226 case kHiddenArg: res_reg = rs_rAX; break; 227 case kHiddenFpArg: DCHECK(!cu_->target64); res_reg = rs_fr0; break; 228 case kCount: res_reg = rs_rX86_COUNT; break; 229 default: res_reg = RegStorage::InvalidReg(); 230 } 231 return res_reg; 232} 233 234RegStorage X86Mir2Lir::TargetReg(SpecialTargetRegister reg) { 235 LOG(FATAL) << "Do not use this function!!!"; 236 return RegStorage::InvalidReg(); 237} 238 239/* 240 * Decode the register id. 241 */ 242ResourceMask X86Mir2Lir::GetRegMaskCommon(const RegStorage& reg) const { 243 /* Double registers in x86 are just a single FP register. This is always just a single bit. */ 244 return ResourceMask::Bit( 245 /* FP register starts at bit position 16 */ 246 ((reg.IsFloat() || reg.StorageSize() > 8) ? kX86FPReg0 : 0) + reg.GetRegNum()); 247} 248 249ResourceMask X86Mir2Lir::GetPCUseDefEncoding() const { 250 return kEncodeNone; 251} 252 253void X86Mir2Lir::SetupTargetResourceMasks(LIR* lir, uint64_t flags, 254 ResourceMask* use_mask, ResourceMask* def_mask) { 255 DCHECK(cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64); 256 DCHECK(!lir->flags.use_def_invalid); 257 258 // X86-specific resource map setup here. 259 if (flags & REG_USE_SP) { 260 use_mask->SetBit(kX86RegSP); 261 } 262 263 if (flags & REG_DEF_SP) { 264 def_mask->SetBit(kX86RegSP); 265 } 266 267 if (flags & REG_DEFA) { 268 SetupRegMask(def_mask, rs_rAX.GetReg()); 269 } 270 271 if (flags & REG_DEFD) { 272 SetupRegMask(def_mask, rs_rDX.GetReg()); 273 } 274 if (flags & REG_USEA) { 275 SetupRegMask(use_mask, rs_rAX.GetReg()); 276 } 277 278 if (flags & REG_USEC) { 279 SetupRegMask(use_mask, rs_rCX.GetReg()); 280 } 281 282 if (flags & REG_USED) { 283 SetupRegMask(use_mask, rs_rDX.GetReg()); 284 } 285 286 if (flags & REG_USEB) { 287 SetupRegMask(use_mask, rs_rBX.GetReg()); 288 } 289 290 // Fixup hard to describe instruction: Uses rAX, rCX, rDI; sets rDI. 291 if (lir->opcode == kX86RepneScasw) { 292 SetupRegMask(use_mask, rs_rAX.GetReg()); 293 SetupRegMask(use_mask, rs_rCX.GetReg()); 294 SetupRegMask(use_mask, rs_rDI.GetReg()); 295 SetupRegMask(def_mask, rs_rDI.GetReg()); 296 } 297 298 if (flags & USE_FP_STACK) { 299 use_mask->SetBit(kX86FPStack); 300 def_mask->SetBit(kX86FPStack); 301 } 302} 303 304/* For dumping instructions */ 305static const char* x86RegName[] = { 306 "rax", "rcx", "rdx", "rbx", "rsp", "rbp", "rsi", "rdi", 307 "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15" 308}; 309 310static const char* x86CondName[] = { 311 "O", 312 "NO", 313 "B/NAE/C", 314 "NB/AE/NC", 315 "Z/EQ", 316 "NZ/NE", 317 "BE/NA", 318 "NBE/A", 319 "S", 320 "NS", 321 "P/PE", 322 "NP/PO", 323 "L/NGE", 324 "NL/GE", 325 "LE/NG", 326 "NLE/G" 327}; 328 329/* 330 * Interpret a format string and build a string no longer than size 331 * See format key in Assemble.cc. 332 */ 333std::string X86Mir2Lir::BuildInsnString(const char *fmt, LIR *lir, unsigned char* base_addr) { 334 std::string buf; 335 size_t i = 0; 336 size_t fmt_len = strlen(fmt); 337 while (i < fmt_len) { 338 if (fmt[i] != '!') { 339 buf += fmt[i]; 340 i++; 341 } else { 342 i++; 343 DCHECK_LT(i, fmt_len); 344 char operand_number_ch = fmt[i]; 345 i++; 346 if (operand_number_ch == '!') { 347 buf += "!"; 348 } else { 349 int operand_number = operand_number_ch - '0'; 350 DCHECK_LT(operand_number, 6); // Expect upto 6 LIR operands. 351 DCHECK_LT(i, fmt_len); 352 int operand = lir->operands[operand_number]; 353 switch (fmt[i]) { 354 case 'c': 355 DCHECK_LT(static_cast<size_t>(operand), sizeof(x86CondName)); 356 buf += x86CondName[operand]; 357 break; 358 case 'd': 359 buf += StringPrintf("%d", operand); 360 break; 361 case 'q': { 362 int64_t value = static_cast<int64_t>(static_cast<int64_t>(operand) << 32 | 363 static_cast<uint32_t>(lir->operands[operand_number+1])); 364 buf +=StringPrintf("%" PRId64, value); 365 } 366 case 'p': { 367 EmbeddedData *tab_rec = reinterpret_cast<EmbeddedData*>(UnwrapPointer(operand)); 368 buf += StringPrintf("0x%08x", tab_rec->offset); 369 break; 370 } 371 case 'r': 372 if (RegStorage::IsFloat(operand)) { 373 int fp_reg = RegStorage::RegNum(operand); 374 buf += StringPrintf("xmm%d", fp_reg); 375 } else { 376 int reg_num = RegStorage::RegNum(operand); 377 DCHECK_LT(static_cast<size_t>(reg_num), sizeof(x86RegName)); 378 buf += x86RegName[reg_num]; 379 } 380 break; 381 case 't': 382 buf += StringPrintf("0x%08" PRIxPTR " (L%p)", 383 reinterpret_cast<uintptr_t>(base_addr) + lir->offset + operand, 384 lir->target); 385 break; 386 default: 387 buf += StringPrintf("DecodeError '%c'", fmt[i]); 388 break; 389 } 390 i++; 391 } 392 } 393 } 394 return buf; 395} 396 397void X86Mir2Lir::DumpResourceMask(LIR *x86LIR, const ResourceMask& mask, const char *prefix) { 398 char buf[256]; 399 buf[0] = 0; 400 401 if (mask.Equals(kEncodeAll)) { 402 strcpy(buf, "all"); 403 } else { 404 char num[8]; 405 int i; 406 407 for (i = 0; i < kX86RegEnd; i++) { 408 if (mask.HasBit(i)) { 409 snprintf(num, arraysize(num), "%d ", i); 410 strcat(buf, num); 411 } 412 } 413 414 if (mask.HasBit(ResourceMask::kCCode)) { 415 strcat(buf, "cc "); 416 } 417 /* Memory bits */ 418 if (x86LIR && (mask.HasBit(ResourceMask::kDalvikReg))) { 419 snprintf(buf + strlen(buf), arraysize(buf) - strlen(buf), "dr%d%s", 420 DECODE_ALIAS_INFO_REG(x86LIR->flags.alias_info), 421 (DECODE_ALIAS_INFO_WIDE(x86LIR->flags.alias_info)) ? "(+1)" : ""); 422 } 423 if (mask.HasBit(ResourceMask::kLiteral)) { 424 strcat(buf, "lit "); 425 } 426 427 if (mask.HasBit(ResourceMask::kHeapRef)) { 428 strcat(buf, "heap "); 429 } 430 if (mask.HasBit(ResourceMask::kMustNotAlias)) { 431 strcat(buf, "noalias "); 432 } 433 } 434 if (buf[0]) { 435 LOG(INFO) << prefix << ": " << buf; 436 } 437} 438 439void X86Mir2Lir::AdjustSpillMask() { 440 // Adjustment for LR spilling, x86 has no LR so nothing to do here 441 core_spill_mask_ |= (1 << rs_rRET.GetRegNum()); 442 num_core_spills_++; 443} 444 445RegStorage X86Mir2Lir::AllocateByteRegister() { 446 RegStorage reg = AllocTypedTemp(false, kCoreReg); 447 if (!cu_->target64) { 448 DCHECK_LT(reg.GetRegNum(), rs_rX86_SP.GetRegNum()); 449 } 450 return reg; 451} 452 453RegStorage X86Mir2Lir::Get128BitRegister(RegStorage reg) { 454 return GetRegInfo(reg)->FindMatchingView(RegisterInfo::k128SoloStorageMask)->GetReg(); 455} 456 457bool X86Mir2Lir::IsByteRegister(RegStorage reg) { 458 return cu_->target64 || reg.GetRegNum() < rs_rX86_SP.GetRegNum(); 459} 460 461/* Clobber all regs that might be used by an external C call */ 462void X86Mir2Lir::ClobberCallerSave() { 463 if (cu_->target64) { 464 Clobber(rs_rAX); 465 Clobber(rs_rCX); 466 Clobber(rs_rDX); 467 Clobber(rs_rSI); 468 Clobber(rs_rDI); 469 470 Clobber(rs_r8); 471 Clobber(rs_r9); 472 Clobber(rs_r10); 473 Clobber(rs_r11); 474 475 Clobber(rs_fr8); 476 Clobber(rs_fr9); 477 Clobber(rs_fr10); 478 Clobber(rs_fr11); 479 } else { 480 Clobber(rs_rAX); 481 Clobber(rs_rCX); 482 Clobber(rs_rDX); 483 Clobber(rs_rBX); 484 } 485 486 Clobber(rs_fr0); 487 Clobber(rs_fr1); 488 Clobber(rs_fr2); 489 Clobber(rs_fr3); 490 Clobber(rs_fr4); 491 Clobber(rs_fr5); 492 Clobber(rs_fr6); 493 Clobber(rs_fr7); 494} 495 496RegLocation X86Mir2Lir::GetReturnWideAlt() { 497 RegLocation res = LocCReturnWide(); 498 DCHECK(res.reg.GetLowReg() == rs_rAX.GetReg()); 499 DCHECK(res.reg.GetHighReg() == rs_rDX.GetReg()); 500 Clobber(rs_rAX); 501 Clobber(rs_rDX); 502 MarkInUse(rs_rAX); 503 MarkInUse(rs_rDX); 504 MarkWide(res.reg); 505 return res; 506} 507 508RegLocation X86Mir2Lir::GetReturnAlt() { 509 RegLocation res = LocCReturn(); 510 res.reg.SetReg(rs_rDX.GetReg()); 511 Clobber(rs_rDX); 512 MarkInUse(rs_rDX); 513 return res; 514} 515 516/* To be used when explicitly managing register use */ 517void X86Mir2Lir::LockCallTemps() { 518 LockTemp(rs_rX86_ARG0); 519 LockTemp(rs_rX86_ARG1); 520 LockTemp(rs_rX86_ARG2); 521 LockTemp(rs_rX86_ARG3); 522 if (cu_->target64) { 523 LockTemp(rs_rX86_ARG4); 524 LockTemp(rs_rX86_ARG5); 525 LockTemp(rs_rX86_FARG0); 526 LockTemp(rs_rX86_FARG1); 527 LockTemp(rs_rX86_FARG2); 528 LockTemp(rs_rX86_FARG3); 529 LockTemp(rs_rX86_FARG4); 530 LockTemp(rs_rX86_FARG5); 531 LockTemp(rs_rX86_FARG6); 532 LockTemp(rs_rX86_FARG7); 533 } 534} 535 536/* To be used when explicitly managing register use */ 537void X86Mir2Lir::FreeCallTemps() { 538 FreeTemp(rs_rX86_ARG0); 539 FreeTemp(rs_rX86_ARG1); 540 FreeTemp(rs_rX86_ARG2); 541 FreeTemp(rs_rX86_ARG3); 542 if (cu_->target64) { 543 FreeTemp(rs_rX86_ARG4); 544 FreeTemp(rs_rX86_ARG5); 545 FreeTemp(rs_rX86_FARG0); 546 FreeTemp(rs_rX86_FARG1); 547 FreeTemp(rs_rX86_FARG2); 548 FreeTemp(rs_rX86_FARG3); 549 FreeTemp(rs_rX86_FARG4); 550 FreeTemp(rs_rX86_FARG5); 551 FreeTemp(rs_rX86_FARG6); 552 FreeTemp(rs_rX86_FARG7); 553 } 554} 555 556bool X86Mir2Lir::ProvidesFullMemoryBarrier(X86OpCode opcode) { 557 switch (opcode) { 558 case kX86LockCmpxchgMR: 559 case kX86LockCmpxchgAR: 560 case kX86LockCmpxchg64M: 561 case kX86LockCmpxchg64A: 562 case kX86XchgMR: 563 case kX86Mfence: 564 // Atomic memory instructions provide full barrier. 565 return true; 566 default: 567 break; 568 } 569 570 // Conservative if cannot prove it provides full barrier. 571 return false; 572} 573 574bool X86Mir2Lir::GenMemBarrier(MemBarrierKind barrier_kind) { 575#if ANDROID_SMP != 0 576 // Start off with using the last LIR as the barrier. If it is not enough, then we will update it. 577 LIR* mem_barrier = last_lir_insn_; 578 579 bool ret = false; 580 /* 581 * According to the JSR-133 Cookbook, for x86 only StoreLoad/AnyAny barriers need memory fence. 582 * All other barriers (LoadAny, AnyStore, StoreStore) are nops due to the x86 memory model. 583 * For those cases, all we need to ensure is that there is a scheduling barrier in place. 584 */ 585 if (barrier_kind == kAnyAny) { 586 // If no LIR exists already that can be used a barrier, then generate an mfence. 587 if (mem_barrier == nullptr) { 588 mem_barrier = NewLIR0(kX86Mfence); 589 ret = true; 590 } 591 592 // If last instruction does not provide full barrier, then insert an mfence. 593 if (ProvidesFullMemoryBarrier(static_cast<X86OpCode>(mem_barrier->opcode)) == false) { 594 mem_barrier = NewLIR0(kX86Mfence); 595 ret = true; 596 } 597 } 598 599 // Now ensure that a scheduling barrier is in place. 600 if (mem_barrier == nullptr) { 601 GenBarrier(); 602 } else { 603 // Mark as a scheduling barrier. 604 DCHECK(!mem_barrier->flags.use_def_invalid); 605 mem_barrier->u.m.def_mask = &kEncodeAll; 606 } 607 return ret; 608#else 609 return false; 610#endif 611} 612 613void X86Mir2Lir::CompilerInitializeRegAlloc() { 614 if (cu_->target64) { 615 reg_pool_ = new (arena_) RegisterPool(this, arena_, core_regs_64, core_regs_64q, sp_regs_64, 616 dp_regs_64, reserved_regs_64, reserved_regs_64q, 617 core_temps_64, core_temps_64q, sp_temps_64, dp_temps_64); 618 } else { 619 reg_pool_ = new (arena_) RegisterPool(this, arena_, core_regs_32, empty_pool, sp_regs_32, 620 dp_regs_32, reserved_regs_32, empty_pool, 621 core_temps_32, empty_pool, sp_temps_32, dp_temps_32); 622 } 623 624 // Target-specific adjustments. 625 626 // Add in XMM registers. 627 const ArrayRef<const RegStorage> *xp_regs = cu_->target64 ? &xp_regs_64 : &xp_regs_32; 628 for (RegStorage reg : *xp_regs) { 629 RegisterInfo* info = new (arena_) RegisterInfo(reg, GetRegMaskCommon(reg)); 630 reginfo_map_.Put(reg.GetReg(), info); 631 } 632 const ArrayRef<const RegStorage> *xp_temps = cu_->target64 ? &xp_temps_64 : &xp_temps_32; 633 for (RegStorage reg : *xp_temps) { 634 RegisterInfo* xp_reg_info = GetRegInfo(reg); 635 xp_reg_info->SetIsTemp(true); 636 } 637 638 // Alias single precision xmm to double xmms. 639 // TODO: as needed, add larger vector sizes - alias all to the largest. 640 GrowableArray<RegisterInfo*>::Iterator it(®_pool_->sp_regs_); 641 for (RegisterInfo* info = it.Next(); info != nullptr; info = it.Next()) { 642 int sp_reg_num = info->GetReg().GetRegNum(); 643 RegStorage xp_reg = RegStorage::Solo128(sp_reg_num); 644 RegisterInfo* xp_reg_info = GetRegInfo(xp_reg); 645 // 128-bit xmm vector register's master storage should refer to itself. 646 DCHECK_EQ(xp_reg_info, xp_reg_info->Master()); 647 648 // Redirect 32-bit vector's master storage to 128-bit vector. 649 info->SetMaster(xp_reg_info); 650 651 RegStorage dp_reg = RegStorage::FloatSolo64(sp_reg_num); 652 RegisterInfo* dp_reg_info = GetRegInfo(dp_reg); 653 // Redirect 64-bit vector's master storage to 128-bit vector. 654 dp_reg_info->SetMaster(xp_reg_info); 655 // Singles should show a single 32-bit mask bit, at first referring to the low half. 656 DCHECK_EQ(info->StorageMask(), 0x1U); 657 } 658 659 if (cu_->target64) { 660 // Alias 32bit W registers to corresponding 64bit X registers. 661 GrowableArray<RegisterInfo*>::Iterator w_it(®_pool_->core_regs_); 662 for (RegisterInfo* info = w_it.Next(); info != nullptr; info = w_it.Next()) { 663 int x_reg_num = info->GetReg().GetRegNum(); 664 RegStorage x_reg = RegStorage::Solo64(x_reg_num); 665 RegisterInfo* x_reg_info = GetRegInfo(x_reg); 666 // 64bit X register's master storage should refer to itself. 667 DCHECK_EQ(x_reg_info, x_reg_info->Master()); 668 // Redirect 32bit W master storage to 64bit X. 669 info->SetMaster(x_reg_info); 670 // 32bit W should show a single 32-bit mask bit, at first referring to the low half. 671 DCHECK_EQ(info->StorageMask(), 0x1U); 672 } 673 } 674 675 // Don't start allocating temps at r0/s0/d0 or you may clobber return regs in early-exit methods. 676 // TODO: adjust for x86/hard float calling convention. 677 reg_pool_->next_core_reg_ = 2; 678 reg_pool_->next_sp_reg_ = 2; 679 reg_pool_->next_dp_reg_ = 1; 680} 681 682int X86Mir2Lir::VectorRegisterSize() { 683 return 128; 684} 685 686int X86Mir2Lir::NumReservableVectorRegisters(bool fp_used) { 687 return fp_used ? 5 : 7; 688} 689 690void X86Mir2Lir::SpillCoreRegs() { 691 if (num_core_spills_ == 0) { 692 return; 693 } 694 // Spill mask not including fake return address register 695 uint32_t mask = core_spill_mask_ & ~(1 << rs_rRET.GetRegNum()); 696 int offset = frame_size_ - (GetInstructionSetPointerSize(cu_->instruction_set) * num_core_spills_); 697 OpSize size = cu_->target64 ? k64 : k32; 698 for (int reg = 0; mask; mask >>= 1, reg++) { 699 if (mask & 0x1) { 700 StoreBaseDisp(rs_rX86_SP, offset, cu_->target64 ? RegStorage::Solo64(reg) : RegStorage::Solo32(reg), 701 size, kNotVolatile); 702 offset += GetInstructionSetPointerSize(cu_->instruction_set); 703 } 704 } 705} 706 707void X86Mir2Lir::UnSpillCoreRegs() { 708 if (num_core_spills_ == 0) { 709 return; 710 } 711 // Spill mask not including fake return address register 712 uint32_t mask = core_spill_mask_ & ~(1 << rs_rRET.GetRegNum()); 713 int offset = frame_size_ - (GetInstructionSetPointerSize(cu_->instruction_set) * num_core_spills_); 714 OpSize size = cu_->target64 ? k64 : k32; 715 for (int reg = 0; mask; mask >>= 1, reg++) { 716 if (mask & 0x1) { 717 LoadBaseDisp(rs_rX86_SP, offset, cu_->target64 ? RegStorage::Solo64(reg) : RegStorage::Solo32(reg), 718 size, kNotVolatile); 719 offset += GetInstructionSetPointerSize(cu_->instruction_set); 720 } 721 } 722} 723 724void X86Mir2Lir::SpillFPRegs() { 725 if (num_fp_spills_ == 0) { 726 return; 727 } 728 uint32_t mask = fp_spill_mask_; 729 int offset = frame_size_ - (GetInstructionSetPointerSize(cu_->instruction_set) * (num_fp_spills_ + num_core_spills_)); 730 for (int reg = 0; mask; mask >>= 1, reg++) { 731 if (mask & 0x1) { 732 StoreBaseDisp(rs_rX86_SP, offset, RegStorage::FloatSolo64(reg), 733 k64, kNotVolatile); 734 offset += sizeof(double); 735 } 736 } 737} 738void X86Mir2Lir::UnSpillFPRegs() { 739 if (num_fp_spills_ == 0) { 740 return; 741 } 742 uint32_t mask = fp_spill_mask_; 743 int offset = frame_size_ - (GetInstructionSetPointerSize(cu_->instruction_set) * (num_fp_spills_ + num_core_spills_)); 744 for (int reg = 0; mask; mask >>= 1, reg++) { 745 if (mask & 0x1) { 746 LoadBaseDisp(rs_rX86_SP, offset, RegStorage::FloatSolo64(reg), 747 k64, kNotVolatile); 748 offset += sizeof(double); 749 } 750 } 751} 752 753 754bool X86Mir2Lir::IsUnconditionalBranch(LIR* lir) { 755 return (lir->opcode == kX86Jmp8 || lir->opcode == kX86Jmp32); 756} 757 758RegisterClass X86Mir2Lir::RegClassForFieldLoadStore(OpSize size, bool is_volatile) { 759 // X86_64 can handle any size. 760 if (cu_->target64) { 761 if (size == kReference) { 762 return kRefReg; 763 } 764 return kCoreReg; 765 } 766 767 if (UNLIKELY(is_volatile)) { 768 // On x86, atomic 64-bit load/store requires an fp register. 769 // Smaller aligned load/store is atomic for both core and fp registers. 770 if (size == k64 || size == kDouble) { 771 return kFPReg; 772 } 773 } 774 return RegClassBySize(size); 775} 776 777X86Mir2Lir::X86Mir2Lir(CompilationUnit* cu, MIRGraph* mir_graph, ArenaAllocator* arena) 778 : Mir2Lir(cu, mir_graph, arena), 779 base_of_code_(nullptr), store_method_addr_(false), store_method_addr_used_(false), 780 method_address_insns_(arena, 100, kGrowableArrayMisc), 781 class_type_address_insns_(arena, 100, kGrowableArrayMisc), 782 call_method_insns_(arena, 100, kGrowableArrayMisc), 783 stack_decrement_(nullptr), stack_increment_(nullptr), 784 const_vectors_(nullptr) { 785 store_method_addr_used_ = false; 786 if (kIsDebugBuild) { 787 for (int i = 0; i < kX86Last; i++) { 788 if (X86Mir2Lir::EncodingMap[i].opcode != i) { 789 LOG(FATAL) << "Encoding order for " << X86Mir2Lir::EncodingMap[i].name 790 << " is wrong: expecting " << i << ", seeing " 791 << static_cast<int>(X86Mir2Lir::EncodingMap[i].opcode); 792 } 793 } 794 } 795 if (cu_->target64) { 796 rs_rX86_SP = rs_rX86_SP_64; 797 798 rs_rX86_ARG0 = rs_rDI; 799 rs_rX86_ARG1 = rs_rSI; 800 rs_rX86_ARG2 = rs_rDX; 801 rs_rX86_ARG3 = rs_rCX; 802 rs_rX86_ARG4 = rs_r8; 803 rs_rX86_ARG5 = rs_r9; 804 rs_rX86_FARG0 = rs_fr0; 805 rs_rX86_FARG1 = rs_fr1; 806 rs_rX86_FARG2 = rs_fr2; 807 rs_rX86_FARG3 = rs_fr3; 808 rs_rX86_FARG4 = rs_fr4; 809 rs_rX86_FARG5 = rs_fr5; 810 rs_rX86_FARG6 = rs_fr6; 811 rs_rX86_FARG7 = rs_fr7; 812 rX86_ARG0 = rDI; 813 rX86_ARG1 = rSI; 814 rX86_ARG2 = rDX; 815 rX86_ARG3 = rCX; 816 rX86_ARG4 = r8; 817 rX86_ARG5 = r9; 818 rX86_FARG0 = fr0; 819 rX86_FARG1 = fr1; 820 rX86_FARG2 = fr2; 821 rX86_FARG3 = fr3; 822 rX86_FARG4 = fr4; 823 rX86_FARG5 = fr5; 824 rX86_FARG6 = fr6; 825 rX86_FARG7 = fr7; 826 rs_rX86_INVOKE_TGT = rs_rDI; 827 } else { 828 rs_rX86_SP = rs_rX86_SP_32; 829 830 rs_rX86_ARG0 = rs_rAX; 831 rs_rX86_ARG1 = rs_rCX; 832 rs_rX86_ARG2 = rs_rDX; 833 rs_rX86_ARG3 = rs_rBX; 834 rs_rX86_ARG4 = RegStorage::InvalidReg(); 835 rs_rX86_ARG5 = RegStorage::InvalidReg(); 836 rs_rX86_FARG0 = rs_rAX; 837 rs_rX86_FARG1 = rs_rCX; 838 rs_rX86_FARG2 = rs_rDX; 839 rs_rX86_FARG3 = rs_rBX; 840 rs_rX86_FARG4 = RegStorage::InvalidReg(); 841 rs_rX86_FARG5 = RegStorage::InvalidReg(); 842 rs_rX86_FARG6 = RegStorage::InvalidReg(); 843 rs_rX86_FARG7 = RegStorage::InvalidReg(); 844 rX86_ARG0 = rAX; 845 rX86_ARG1 = rCX; 846 rX86_ARG2 = rDX; 847 rX86_ARG3 = rBX; 848 rX86_FARG0 = rAX; 849 rX86_FARG1 = rCX; 850 rX86_FARG2 = rDX; 851 rX86_FARG3 = rBX; 852 rs_rX86_INVOKE_TGT = rs_rAX; 853 // TODO(64): Initialize with invalid reg 854// rX86_ARG4 = RegStorage::InvalidReg(); 855// rX86_ARG5 = RegStorage::InvalidReg(); 856 } 857 rs_rX86_RET0 = rs_rAX; 858 rs_rX86_RET1 = rs_rDX; 859 rs_rX86_COUNT = rs_rCX; 860 rX86_RET0 = rAX; 861 rX86_RET1 = rDX; 862 rX86_INVOKE_TGT = rAX; 863 rX86_COUNT = rCX; 864 865 // Initialize the number of reserved vector registers 866 num_reserved_vector_regs_ = -1; 867} 868 869Mir2Lir* X86CodeGenerator(CompilationUnit* const cu, MIRGraph* const mir_graph, 870 ArenaAllocator* const arena) { 871 return new X86Mir2Lir(cu, mir_graph, arena); 872} 873 874// Not used in x86(-64) 875RegStorage X86Mir2Lir::LoadHelper(QuickEntrypointEnum trampoline) { 876 LOG(FATAL) << "Unexpected use of LoadHelper in x86"; 877 return RegStorage::InvalidReg(); 878} 879 880LIR* X86Mir2Lir::CheckSuspendUsingLoad() { 881 // First load the pointer in fs:[suspend-trigger] into eax 882 // Then use a test instruction to indirect via that address. 883 if (cu_->target64) { 884 NewLIR2(kX86Mov64RT, rs_rAX.GetReg(), 885 Thread::ThreadSuspendTriggerOffset<8>().Int32Value()); 886 } else { 887 NewLIR2(kX86Mov32RT, rs_rAX.GetReg(), 888 Thread::ThreadSuspendTriggerOffset<4>().Int32Value()); 889 } 890 return NewLIR3(kX86Test32RM, rs_rAX.GetReg(), rs_rAX.GetReg(), 0); 891} 892 893uint64_t X86Mir2Lir::GetTargetInstFlags(int opcode) { 894 DCHECK(!IsPseudoLirOp(opcode)); 895 return X86Mir2Lir::EncodingMap[opcode].flags; 896} 897 898const char* X86Mir2Lir::GetTargetInstName(int opcode) { 899 DCHECK(!IsPseudoLirOp(opcode)); 900 return X86Mir2Lir::EncodingMap[opcode].name; 901} 902 903const char* X86Mir2Lir::GetTargetInstFmt(int opcode) { 904 DCHECK(!IsPseudoLirOp(opcode)); 905 return X86Mir2Lir::EncodingMap[opcode].fmt; 906} 907 908void X86Mir2Lir::GenConstWide(RegLocation rl_dest, int64_t value) { 909 // Can we do this directly to memory? 910 rl_dest = UpdateLocWide(rl_dest); 911 if ((rl_dest.location == kLocDalvikFrame) || 912 (rl_dest.location == kLocCompilerTemp)) { 913 int32_t val_lo = Low32Bits(value); 914 int32_t val_hi = High32Bits(value); 915 int r_base = rs_rX86_SP.GetReg(); 916 int displacement = SRegOffset(rl_dest.s_reg_low); 917 918 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 919 LIR * store = NewLIR3(kX86Mov32MI, r_base, displacement + LOWORD_OFFSET, val_lo); 920 AnnotateDalvikRegAccess(store, (displacement + LOWORD_OFFSET) >> 2, 921 false /* is_load */, true /* is64bit */); 922 store = NewLIR3(kX86Mov32MI, r_base, displacement + HIWORD_OFFSET, val_hi); 923 AnnotateDalvikRegAccess(store, (displacement + HIWORD_OFFSET) >> 2, 924 false /* is_load */, true /* is64bit */); 925 return; 926 } 927 928 // Just use the standard code to do the generation. 929 Mir2Lir::GenConstWide(rl_dest, value); 930} 931 932// TODO: Merge with existing RegLocation dumper in vreg_analysis.cc 933void X86Mir2Lir::DumpRegLocation(RegLocation loc) { 934 LOG(INFO) << "location: " << loc.location << ',' 935 << (loc.wide ? " w" : " ") 936 << (loc.defined ? " D" : " ") 937 << (loc.is_const ? " c" : " ") 938 << (loc.fp ? " F" : " ") 939 << (loc.core ? " C" : " ") 940 << (loc.ref ? " r" : " ") 941 << (loc.high_word ? " h" : " ") 942 << (loc.home ? " H" : " ") 943 << ", low: " << static_cast<int>(loc.reg.GetLowReg()) 944 << ", high: " << static_cast<int>(loc.reg.GetHighReg()) 945 << ", s_reg: " << loc.s_reg_low 946 << ", orig: " << loc.orig_sreg; 947} 948 949void X86Mir2Lir::Materialize() { 950 // A good place to put the analysis before starting. 951 AnalyzeMIR(); 952 953 // Now continue with regular code generation. 954 Mir2Lir::Materialize(); 955} 956 957void X86Mir2Lir::LoadMethodAddress(const MethodReference& target_method, InvokeType type, 958 SpecialTargetRegister symbolic_reg) { 959 /* 960 * For x86, just generate a 32 bit move immediate instruction, that will be filled 961 * in at 'link time'. For now, put a unique value based on target to ensure that 962 * code deduplication works. 963 */ 964 int target_method_idx = target_method.dex_method_index; 965 const DexFile* target_dex_file = target_method.dex_file; 966 const DexFile::MethodId& target_method_id = target_dex_file->GetMethodId(target_method_idx); 967 uintptr_t target_method_id_ptr = reinterpret_cast<uintptr_t>(&target_method_id); 968 969 // Generate the move instruction with the unique pointer and save index, dex_file, and type. 970 LIR *move = RawLIR(current_dalvik_offset_, kX86Mov32RI, 971 TargetReg(symbolic_reg, kNotWide).GetReg(), 972 static_cast<int>(target_method_id_ptr), target_method_idx, 973 WrapPointer(const_cast<DexFile*>(target_dex_file)), type); 974 AppendLIR(move); 975 method_address_insns_.Insert(move); 976} 977 978void X86Mir2Lir::LoadClassType(uint32_t type_idx, SpecialTargetRegister symbolic_reg) { 979 /* 980 * For x86, just generate a 32 bit move immediate instruction, that will be filled 981 * in at 'link time'. For now, put a unique value based on target to ensure that 982 * code deduplication works. 983 */ 984 const DexFile::TypeId& id = cu_->dex_file->GetTypeId(type_idx); 985 uintptr_t ptr = reinterpret_cast<uintptr_t>(&id); 986 987 // Generate the move instruction with the unique pointer and save index and type. 988 LIR *move = RawLIR(current_dalvik_offset_, kX86Mov32RI, 989 TargetReg(symbolic_reg, kNotWide).GetReg(), 990 static_cast<int>(ptr), type_idx); 991 AppendLIR(move); 992 class_type_address_insns_.Insert(move); 993} 994 995LIR *X86Mir2Lir::CallWithLinkerFixup(const MethodReference& target_method, InvokeType type) { 996 /* 997 * For x86, just generate a 32 bit call relative instruction, that will be filled 998 * in at 'link time'. For now, put a unique value based on target to ensure that 999 * code deduplication works. 1000 */ 1001 int target_method_idx = target_method.dex_method_index; 1002 const DexFile* target_dex_file = target_method.dex_file; 1003 const DexFile::MethodId& target_method_id = target_dex_file->GetMethodId(target_method_idx); 1004 uintptr_t target_method_id_ptr = reinterpret_cast<uintptr_t>(&target_method_id); 1005 1006 // Generate the call instruction with the unique pointer and save index, dex_file, and type. 1007 LIR *call = RawLIR(current_dalvik_offset_, kX86CallI, static_cast<int>(target_method_id_ptr), 1008 target_method_idx, WrapPointer(const_cast<DexFile*>(target_dex_file)), type); 1009 AppendLIR(call); 1010 call_method_insns_.Insert(call); 1011 return call; 1012} 1013 1014/* 1015 * @brief Enter a 32 bit quantity into a buffer 1016 * @param buf buffer. 1017 * @param data Data value. 1018 */ 1019 1020static void PushWord(std::vector<uint8_t>&buf, int32_t data) { 1021 buf.push_back(data & 0xff); 1022 buf.push_back((data >> 8) & 0xff); 1023 buf.push_back((data >> 16) & 0xff); 1024 buf.push_back((data >> 24) & 0xff); 1025} 1026 1027void X86Mir2Lir::InstallLiteralPools() { 1028 // These are handled differently for x86. 1029 DCHECK(code_literal_list_ == nullptr); 1030 DCHECK(method_literal_list_ == nullptr); 1031 DCHECK(class_literal_list_ == nullptr); 1032 1033 // Align to 16 byte boundary. We have implicit knowledge that the start of the method is 1034 // on a 4 byte boundary. How can I check this if it changes (other than aligned loads 1035 // will fail at runtime)? 1036 if (const_vectors_ != nullptr) { 1037 int align_size = (16-4) - (code_buffer_.size() & 0xF); 1038 if (align_size < 0) { 1039 align_size += 16; 1040 } 1041 1042 while (align_size > 0) { 1043 code_buffer_.push_back(0); 1044 align_size--; 1045 } 1046 for (LIR *p = const_vectors_; p != nullptr; p = p->next) { 1047 PushWord(code_buffer_, p->operands[0]); 1048 PushWord(code_buffer_, p->operands[1]); 1049 PushWord(code_buffer_, p->operands[2]); 1050 PushWord(code_buffer_, p->operands[3]); 1051 } 1052 } 1053 1054 // Handle the fixups for methods. 1055 for (uint32_t i = 0; i < method_address_insns_.Size(); i++) { 1056 LIR* p = method_address_insns_.Get(i); 1057 DCHECK_EQ(p->opcode, kX86Mov32RI); 1058 uint32_t target_method_idx = p->operands[2]; 1059 const DexFile* target_dex_file = 1060 reinterpret_cast<const DexFile*>(UnwrapPointer(p->operands[3])); 1061 1062 // The offset to patch is the last 4 bytes of the instruction. 1063 int patch_offset = p->offset + p->flags.size - 4; 1064 cu_->compiler_driver->AddMethodPatch(cu_->dex_file, cu_->class_def_idx, 1065 cu_->method_idx, cu_->invoke_type, 1066 target_method_idx, target_dex_file, 1067 static_cast<InvokeType>(p->operands[4]), 1068 patch_offset); 1069 } 1070 1071 // Handle the fixups for class types. 1072 for (uint32_t i = 0; i < class_type_address_insns_.Size(); i++) { 1073 LIR* p = class_type_address_insns_.Get(i); 1074 DCHECK_EQ(p->opcode, kX86Mov32RI); 1075 uint32_t target_method_idx = p->operands[2]; 1076 1077 // The offset to patch is the last 4 bytes of the instruction. 1078 int patch_offset = p->offset + p->flags.size - 4; 1079 cu_->compiler_driver->AddClassPatch(cu_->dex_file, cu_->class_def_idx, 1080 cu_->method_idx, target_method_idx, patch_offset); 1081 } 1082 1083 // And now the PC-relative calls to methods. 1084 for (uint32_t i = 0; i < call_method_insns_.Size(); i++) { 1085 LIR* p = call_method_insns_.Get(i); 1086 DCHECK_EQ(p->opcode, kX86CallI); 1087 uint32_t target_method_idx = p->operands[1]; 1088 const DexFile* target_dex_file = 1089 reinterpret_cast<const DexFile*>(UnwrapPointer(p->operands[2])); 1090 1091 // The offset to patch is the last 4 bytes of the instruction. 1092 int patch_offset = p->offset + p->flags.size - 4; 1093 cu_->compiler_driver->AddRelativeCodePatch(cu_->dex_file, cu_->class_def_idx, 1094 cu_->method_idx, cu_->invoke_type, 1095 target_method_idx, target_dex_file, 1096 static_cast<InvokeType>(p->operands[3]), 1097 patch_offset, -4 /* offset */); 1098 } 1099 1100 // And do the normal processing. 1101 Mir2Lir::InstallLiteralPools(); 1102} 1103 1104bool X86Mir2Lir::GenInlinedArrayCopyCharArray(CallInfo* info) { 1105 RegLocation rl_src = info->args[0]; 1106 RegLocation rl_srcPos = info->args[1]; 1107 RegLocation rl_dst = info->args[2]; 1108 RegLocation rl_dstPos = info->args[3]; 1109 RegLocation rl_length = info->args[4]; 1110 if (rl_srcPos.is_const && (mir_graph_->ConstantValue(rl_srcPos) < 0)) { 1111 return false; 1112 } 1113 if (rl_dstPos.is_const && (mir_graph_->ConstantValue(rl_dstPos) < 0)) { 1114 return false; 1115 } 1116 ClobberCallerSave(); 1117 LockCallTemps(); // Using fixed registers. 1118 RegStorage tmp_reg = cu_->target64 ? rs_r11 : rs_rBX; 1119 LoadValueDirectFixed(rl_src, rs_rAX); 1120 LoadValueDirectFixed(rl_dst, rs_rCX); 1121 LIR* src_dst_same = OpCmpBranch(kCondEq, rs_rAX, rs_rCX, nullptr); 1122 LIR* src_null_branch = OpCmpImmBranch(kCondEq, rs_rAX, 0, nullptr); 1123 LIR* dst_null_branch = OpCmpImmBranch(kCondEq, rs_rCX, 0, nullptr); 1124 LoadValueDirectFixed(rl_length, rs_rDX); 1125 // If the length of the copy is > 128 characters (256 bytes) or negative then go slow path. 1126 LIR* len_too_big = OpCmpImmBranch(kCondHi, rs_rDX, 128, nullptr); 1127 LoadValueDirectFixed(rl_src, rs_rAX); 1128 LoadWordDisp(rs_rAX, mirror::Array::LengthOffset().Int32Value(), rs_rAX); 1129 LIR* src_bad_len = nullptr; 1130 LIR* srcPos_negative = nullptr; 1131 if (!rl_srcPos.is_const) { 1132 LoadValueDirectFixed(rl_srcPos, tmp_reg); 1133 srcPos_negative = OpCmpImmBranch(kCondLt, tmp_reg, 0, nullptr); 1134 OpRegReg(kOpAdd, tmp_reg, rs_rDX); 1135 src_bad_len = OpCmpBranch(kCondLt, rs_rAX, tmp_reg, nullptr); 1136 } else { 1137 int32_t pos_val = mir_graph_->ConstantValue(rl_srcPos.orig_sreg); 1138 if (pos_val == 0) { 1139 src_bad_len = OpCmpBranch(kCondLt, rs_rAX, rs_rDX, nullptr); 1140 } else { 1141 OpRegRegImm(kOpAdd, tmp_reg, rs_rDX, pos_val); 1142 src_bad_len = OpCmpBranch(kCondLt, rs_rAX, tmp_reg, nullptr); 1143 } 1144 } 1145 LIR* dstPos_negative = nullptr; 1146 LIR* dst_bad_len = nullptr; 1147 LoadValueDirectFixed(rl_dst, rs_rAX); 1148 LoadWordDisp(rs_rAX, mirror::Array::LengthOffset().Int32Value(), rs_rAX); 1149 if (!rl_dstPos.is_const) { 1150 LoadValueDirectFixed(rl_dstPos, tmp_reg); 1151 dstPos_negative = OpCmpImmBranch(kCondLt, tmp_reg, 0, nullptr); 1152 OpRegRegReg(kOpAdd, tmp_reg, tmp_reg, rs_rDX); 1153 dst_bad_len = OpCmpBranch(kCondLt, rs_rAX, tmp_reg, nullptr); 1154 } else { 1155 int32_t pos_val = mir_graph_->ConstantValue(rl_dstPos.orig_sreg); 1156 if (pos_val == 0) { 1157 dst_bad_len = OpCmpBranch(kCondLt, rs_rAX, rs_rDX, nullptr); 1158 } else { 1159 OpRegRegImm(kOpAdd, tmp_reg, rs_rDX, pos_val); 1160 dst_bad_len = OpCmpBranch(kCondLt, rs_rAX, tmp_reg, nullptr); 1161 } 1162 } 1163 // Everything is checked now. 1164 LoadValueDirectFixed(rl_src, rs_rAX); 1165 LoadValueDirectFixed(rl_dst, tmp_reg); 1166 LoadValueDirectFixed(rl_srcPos, rs_rCX); 1167 NewLIR5(kX86Lea32RA, rs_rAX.GetReg(), rs_rAX.GetReg(), 1168 rs_rCX.GetReg(), 1, mirror::Array::DataOffset(2).Int32Value()); 1169 // RAX now holds the address of the first src element to be copied. 1170 1171 LoadValueDirectFixed(rl_dstPos, rs_rCX); 1172 NewLIR5(kX86Lea32RA, tmp_reg.GetReg(), tmp_reg.GetReg(), 1173 rs_rCX.GetReg(), 1, mirror::Array::DataOffset(2).Int32Value() ); 1174 // RBX now holds the address of the first dst element to be copied. 1175 1176 // Check if the number of elements to be copied is odd or even. If odd 1177 // then copy the first element (so that the remaining number of elements 1178 // is even). 1179 LoadValueDirectFixed(rl_length, rs_rCX); 1180 OpRegImm(kOpAnd, rs_rCX, 1); 1181 LIR* jmp_to_begin_loop = OpCmpImmBranch(kCondEq, rs_rCX, 0, nullptr); 1182 OpRegImm(kOpSub, rs_rDX, 1); 1183 LoadBaseIndexedDisp(rs_rAX, rs_rDX, 1, 0, rs_rCX, kSignedHalf); 1184 StoreBaseIndexedDisp(tmp_reg, rs_rDX, 1, 0, rs_rCX, kSignedHalf); 1185 1186 // Since the remaining number of elements is even, we will copy by 1187 // two elements at a time. 1188 LIR* beginLoop = NewLIR0(kPseudoTargetLabel); 1189 LIR* jmp_to_ret = OpCmpImmBranch(kCondEq, rs_rDX, 0, nullptr); 1190 OpRegImm(kOpSub, rs_rDX, 2); 1191 LoadBaseIndexedDisp(rs_rAX, rs_rDX, 1, 0, rs_rCX, kSingle); 1192 StoreBaseIndexedDisp(tmp_reg, rs_rDX, 1, 0, rs_rCX, kSingle); 1193 OpUnconditionalBranch(beginLoop); 1194 LIR *check_failed = NewLIR0(kPseudoTargetLabel); 1195 LIR* launchpad_branch = OpUnconditionalBranch(nullptr); 1196 LIR *return_point = NewLIR0(kPseudoTargetLabel); 1197 jmp_to_ret->target = return_point; 1198 jmp_to_begin_loop->target = beginLoop; 1199 src_dst_same->target = check_failed; 1200 len_too_big->target = check_failed; 1201 src_null_branch->target = check_failed; 1202 if (srcPos_negative != nullptr) 1203 srcPos_negative ->target = check_failed; 1204 if (src_bad_len != nullptr) 1205 src_bad_len->target = check_failed; 1206 dst_null_branch->target = check_failed; 1207 if (dstPos_negative != nullptr) 1208 dstPos_negative->target = check_failed; 1209 if (dst_bad_len != nullptr) 1210 dst_bad_len->target = check_failed; 1211 AddIntrinsicSlowPath(info, launchpad_branch, return_point); 1212 return true; 1213} 1214 1215 1216/* 1217 * Fast string.index_of(I) & (II). Inline check for simple case of char <= 0xffff, 1218 * otherwise bails to standard library code. 1219 */ 1220bool X86Mir2Lir::GenInlinedIndexOf(CallInfo* info, bool zero_based) { 1221 RegLocation rl_obj = info->args[0]; 1222 RegLocation rl_char = info->args[1]; 1223 RegLocation rl_start; // Note: only present in III flavor or IndexOf. 1224 // RBX is callee-save register in 64-bit mode. 1225 RegStorage rs_tmp = cu_->target64 ? rs_r11 : rs_rBX; 1226 int start_value = -1; 1227 1228 uint32_t char_value = 1229 rl_char.is_const ? mir_graph_->ConstantValue(rl_char.orig_sreg) : 0; 1230 1231 if (char_value > 0xFFFF) { 1232 // We have to punt to the real String.indexOf. 1233 return false; 1234 } 1235 1236 // Okay, we are commited to inlining this. 1237 // EAX: 16 bit character being searched. 1238 // ECX: count: number of words to be searched. 1239 // EDI: String being searched. 1240 // EDX: temporary during execution. 1241 // EBX or R11: temporary during execution (depending on mode). 1242 // REP SCASW: search instruction. 1243 1244 FlushReg(rs_rAX); 1245 Clobber(rs_rAX); 1246 LockTemp(rs_rAX); 1247 FlushReg(rs_rCX); 1248 Clobber(rs_rCX); 1249 LockTemp(rs_rCX); 1250 FlushReg(rs_rDX); 1251 Clobber(rs_rDX); 1252 LockTemp(rs_rDX); 1253 FlushReg(rs_tmp); 1254 Clobber(rs_tmp); 1255 LockTemp(rs_tmp); 1256 if (cu_->target64) { 1257 FlushReg(rs_rDI); 1258 Clobber(rs_rDI); 1259 LockTemp(rs_rDI); 1260 } 1261 1262 RegLocation rl_return = GetReturn(kCoreReg); 1263 RegLocation rl_dest = InlineTarget(info); 1264 1265 // Is the string non-NULL? 1266 LoadValueDirectFixed(rl_obj, rs_rDX); 1267 GenNullCheck(rs_rDX, info->opt_flags); 1268 info->opt_flags |= MIR_IGNORE_NULL_CHECK; // Record that we've null checked. 1269 1270 LIR *slowpath_branch = nullptr, *length_compare = nullptr; 1271 1272 // We need the value in EAX. 1273 if (rl_char.is_const) { 1274 LoadConstantNoClobber(rs_rAX, char_value); 1275 } else { 1276 // Does the character fit in 16 bits? Compare it at runtime. 1277 LoadValueDirectFixed(rl_char, rs_rAX); 1278 slowpath_branch = OpCmpImmBranch(kCondGt, rs_rAX, 0xFFFF, nullptr); 1279 } 1280 1281 // From here down, we know that we are looking for a char that fits in 16 bits. 1282 // Location of reference to data array within the String object. 1283 int value_offset = mirror::String::ValueOffset().Int32Value(); 1284 // Location of count within the String object. 1285 int count_offset = mirror::String::CountOffset().Int32Value(); 1286 // Starting offset within data array. 1287 int offset_offset = mirror::String::OffsetOffset().Int32Value(); 1288 // Start of char data with array_. 1289 int data_offset = mirror::Array::DataOffset(sizeof(uint16_t)).Int32Value(); 1290 1291 // Compute the number of words to search in to rCX. 1292 Load32Disp(rs_rDX, count_offset, rs_rCX); 1293 1294 // Possible signal here due to null pointer dereference. 1295 // Note that the signal handler will expect the top word of 1296 // the stack to be the ArtMethod*. If the PUSH edi instruction 1297 // below is ahead of the load above then this will not be true 1298 // and the signal handler will not work. 1299 MarkPossibleNullPointerException(0); 1300 1301 if (!cu_->target64) { 1302 // EDI is callee-save register in 32-bit mode. 1303 NewLIR1(kX86Push32R, rs_rDI.GetReg()); 1304 } 1305 1306 if (zero_based) { 1307 // Start index is not present. 1308 // We have to handle an empty string. Use special instruction JECXZ. 1309 length_compare = NewLIR0(kX86Jecxz8); 1310 1311 // Copy the number of words to search in a temporary register. 1312 // We will use the register at the end to calculate result. 1313 OpRegReg(kOpMov, rs_tmp, rs_rCX); 1314 } else { 1315 // Start index is present. 1316 rl_start = info->args[2]; 1317 1318 // We have to offset by the start index. 1319 if (rl_start.is_const) { 1320 start_value = mir_graph_->ConstantValue(rl_start.orig_sreg); 1321 start_value = std::max(start_value, 0); 1322 1323 // Is the start > count? 1324 length_compare = OpCmpImmBranch(kCondLe, rs_rCX, start_value, nullptr); 1325 OpRegImm(kOpMov, rs_rDI, start_value); 1326 1327 // Copy the number of words to search in a temporary register. 1328 // We will use the register at the end to calculate result. 1329 OpRegReg(kOpMov, rs_tmp, rs_rCX); 1330 1331 if (start_value != 0) { 1332 // Decrease the number of words to search by the start index. 1333 OpRegImm(kOpSub, rs_rCX, start_value); 1334 } 1335 } else { 1336 // Handle "start index < 0" case. 1337 if (!cu_->target64 && rl_start.location != kLocPhysReg) { 1338 // Load the start index from stack, remembering that we pushed EDI. 1339 int displacement = SRegOffset(rl_start.s_reg_low) + sizeof(uint32_t); 1340 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 1341 Load32Disp(rs_rX86_SP, displacement, rs_rDI); 1342 // Dalvik register annotation in LoadBaseIndexedDisp() used wrong offset. Fix it. 1343 DCHECK(!DECODE_ALIAS_INFO_WIDE(last_lir_insn_->flags.alias_info)); 1344 int reg_id = DECODE_ALIAS_INFO_REG(last_lir_insn_->flags.alias_info) - 1; 1345 AnnotateDalvikRegAccess(last_lir_insn_, reg_id, true, false); 1346 } else { 1347 LoadValueDirectFixed(rl_start, rs_rDI); 1348 } 1349 OpRegReg(kOpXor, rs_tmp, rs_tmp); 1350 OpRegReg(kOpCmp, rs_rDI, rs_tmp); 1351 OpCondRegReg(kOpCmov, kCondLt, rs_rDI, rs_tmp); 1352 1353 // The length of the string should be greater than the start index. 1354 length_compare = OpCmpBranch(kCondLe, rs_rCX, rs_rDI, nullptr); 1355 1356 // Copy the number of words to search in a temporary register. 1357 // We will use the register at the end to calculate result. 1358 OpRegReg(kOpMov, rs_tmp, rs_rCX); 1359 1360 // Decrease the number of words to search by the start index. 1361 OpRegReg(kOpSub, rs_rCX, rs_rDI); 1362 } 1363 } 1364 1365 // Load the address of the string into EDI. 1366 // In case of start index we have to add the address to existing value in EDI. 1367 // The string starts at VALUE(String) + 2 * OFFSET(String) + DATA_OFFSET. 1368 if (zero_based || (!zero_based && rl_start.is_const && start_value == 0)) { 1369 Load32Disp(rs_rDX, offset_offset, rs_rDI); 1370 } else { 1371 OpRegMem(kOpAdd, rs_rDI, rs_rDX, offset_offset); 1372 } 1373 OpRegImm(kOpLsl, rs_rDI, 1); 1374 OpRegMem(kOpAdd, rs_rDI, rs_rDX, value_offset); 1375 OpRegImm(kOpAdd, rs_rDI, data_offset); 1376 1377 // EDI now contains the start of the string to be searched. 1378 // We are all prepared to do the search for the character. 1379 NewLIR0(kX86RepneScasw); 1380 1381 // Did we find a match? 1382 LIR* failed_branch = OpCondBranch(kCondNe, nullptr); 1383 1384 // yes, we matched. Compute the index of the result. 1385 OpRegReg(kOpSub, rs_tmp, rs_rCX); 1386 NewLIR3(kX86Lea32RM, rl_return.reg.GetReg(), rs_tmp.GetReg(), -1); 1387 1388 LIR *all_done = NewLIR1(kX86Jmp8, 0); 1389 1390 // Failed to match; return -1. 1391 LIR *not_found = NewLIR0(kPseudoTargetLabel); 1392 length_compare->target = not_found; 1393 failed_branch->target = not_found; 1394 LoadConstantNoClobber(rl_return.reg, -1); 1395 1396 // And join up at the end. 1397 all_done->target = NewLIR0(kPseudoTargetLabel); 1398 1399 if (!cu_->target64) 1400 NewLIR1(kX86Pop32R, rs_rDI.GetReg()); 1401 1402 // Out of line code returns here. 1403 if (slowpath_branch != nullptr) { 1404 LIR *return_point = NewLIR0(kPseudoTargetLabel); 1405 AddIntrinsicSlowPath(info, slowpath_branch, return_point); 1406 } 1407 1408 StoreValue(rl_dest, rl_return); 1409 1410 FreeTemp(rs_rAX); 1411 FreeTemp(rs_rCX); 1412 FreeTemp(rs_rDX); 1413 FreeTemp(rs_tmp); 1414 if (cu_->target64) { 1415 FreeTemp(rs_rDI); 1416 } 1417 1418 return true; 1419} 1420 1421/* 1422 * @brief Enter an 'advance LOC' into the FDE buffer 1423 * @param buf FDE buffer. 1424 * @param increment Amount by which to increase the current location. 1425 */ 1426static void AdvanceLoc(std::vector<uint8_t>&buf, uint32_t increment) { 1427 if (increment < 64) { 1428 // Encoding in opcode. 1429 buf.push_back(0x1 << 6 | increment); 1430 } else if (increment < 256) { 1431 // Single byte delta. 1432 buf.push_back(0x02); 1433 buf.push_back(increment); 1434 } else if (increment < 256 * 256) { 1435 // Two byte delta. 1436 buf.push_back(0x03); 1437 buf.push_back(increment & 0xff); 1438 buf.push_back((increment >> 8) & 0xff); 1439 } else { 1440 // Four byte delta. 1441 buf.push_back(0x04); 1442 PushWord(buf, increment); 1443 } 1444} 1445 1446 1447std::vector<uint8_t>* X86CFIInitialization(bool is_x86_64) { 1448 return X86Mir2Lir::ReturnCommonCallFrameInformation(is_x86_64); 1449} 1450 1451static void EncodeUnsignedLeb128(std::vector<uint8_t>& buf, uint32_t value) { 1452 uint8_t buffer[12]; 1453 uint8_t *ptr = EncodeUnsignedLeb128(buffer, value); 1454 for (uint8_t *p = buffer; p < ptr; p++) { 1455 buf.push_back(*p); 1456 } 1457} 1458 1459static void EncodeSignedLeb128(std::vector<uint8_t>& buf, int32_t value) { 1460 uint8_t buffer[12]; 1461 uint8_t *ptr = EncodeSignedLeb128(buffer, value); 1462 for (uint8_t *p = buffer; p < ptr; p++) { 1463 buf.push_back(*p); 1464 } 1465} 1466 1467std::vector<uint8_t>* X86Mir2Lir::ReturnCommonCallFrameInformation(bool is_x86_64) { 1468 std::vector<uint8_t>*cfi_info = new std::vector<uint8_t>; 1469 1470 // Length (will be filled in later in this routine). 1471 PushWord(*cfi_info, 0); 1472 1473 // CIE id: always 0. 1474 PushWord(*cfi_info, 0); 1475 1476 // Version: always 1. 1477 cfi_info->push_back(0x01); 1478 1479 // Augmentation: 'zR\0' 1480 cfi_info->push_back(0x7a); 1481 cfi_info->push_back(0x52); 1482 cfi_info->push_back(0x0); 1483 1484 // Code alignment: 1. 1485 EncodeUnsignedLeb128(*cfi_info, 1); 1486 1487 // Data alignment. 1488 if (is_x86_64) { 1489 EncodeSignedLeb128(*cfi_info, -8); 1490 } else { 1491 EncodeSignedLeb128(*cfi_info, -4); 1492 } 1493 1494 // Return address register. 1495 if (is_x86_64) { 1496 // R16(RIP) 1497 cfi_info->push_back(0x10); 1498 } else { 1499 // R8(EIP) 1500 cfi_info->push_back(0x08); 1501 } 1502 1503 // Augmentation length: 1. 1504 cfi_info->push_back(1); 1505 1506 // Augmentation data: 0x03 ((DW_EH_PE_absptr << 4) | DW_EH_PE_udata4). 1507 cfi_info->push_back(0x03); 1508 1509 // Initial instructions. 1510 if (is_x86_64) { 1511 // DW_CFA_def_cfa R7(RSP) 8. 1512 cfi_info->push_back(0x0c); 1513 cfi_info->push_back(0x07); 1514 cfi_info->push_back(0x08); 1515 1516 // DW_CFA_offset R16(RIP) 1 (* -8). 1517 cfi_info->push_back(0x90); 1518 cfi_info->push_back(0x01); 1519 } else { 1520 // DW_CFA_def_cfa R4(ESP) 4. 1521 cfi_info->push_back(0x0c); 1522 cfi_info->push_back(0x04); 1523 cfi_info->push_back(0x04); 1524 1525 // DW_CFA_offset R8(EIP) 1 (* -4). 1526 cfi_info->push_back(0x88); 1527 cfi_info->push_back(0x01); 1528 } 1529 1530 // Padding to a multiple of 4 1531 while ((cfi_info->size() & 3) != 0) { 1532 // DW_CFA_nop is encoded as 0. 1533 cfi_info->push_back(0); 1534 } 1535 1536 // Set the length of the CIE inside the generated bytes. 1537 uint32_t length = cfi_info->size() - 4; 1538 (*cfi_info)[0] = length; 1539 (*cfi_info)[1] = length >> 8; 1540 (*cfi_info)[2] = length >> 16; 1541 (*cfi_info)[3] = length >> 24; 1542 return cfi_info; 1543} 1544 1545static bool ARTRegIDToDWARFRegID(bool is_x86_64, int art_reg_id, int* dwarf_reg_id) { 1546 if (is_x86_64) { 1547 switch (art_reg_id) { 1548 case 3 : *dwarf_reg_id = 3; return true; // %rbx 1549 // This is the only discrepancy between ART & DWARF register numbering. 1550 case 5 : *dwarf_reg_id = 6; return true; // %rbp 1551 case 12: *dwarf_reg_id = 12; return true; // %r12 1552 case 13: *dwarf_reg_id = 13; return true; // %r13 1553 case 14: *dwarf_reg_id = 14; return true; // %r14 1554 case 15: *dwarf_reg_id = 15; return true; // %r15 1555 default: return false; // Should not get here 1556 } 1557 } else { 1558 switch (art_reg_id) { 1559 case 5: *dwarf_reg_id = 5; return true; // %ebp 1560 case 6: *dwarf_reg_id = 6; return true; // %esi 1561 case 7: *dwarf_reg_id = 7; return true; // %edi 1562 default: return false; // Should not get here 1563 } 1564 } 1565} 1566 1567std::vector<uint8_t>* X86Mir2Lir::ReturnCallFrameInformation() { 1568 std::vector<uint8_t>*cfi_info = new std::vector<uint8_t>; 1569 1570 // Generate the FDE for the method. 1571 DCHECK_NE(data_offset_, 0U); 1572 1573 // Length (will be filled in later in this routine). 1574 PushWord(*cfi_info, 0); 1575 1576 // 'CIE_pointer' (filled in by linker). 1577 PushWord(*cfi_info, 0); 1578 1579 // 'initial_location' (filled in by linker). 1580 PushWord(*cfi_info, 0); 1581 1582 // 'address_range' (number of bytes in the method). 1583 PushWord(*cfi_info, data_offset_); 1584 1585 // Augmentation length: 0 1586 cfi_info->push_back(0); 1587 1588 // The instructions in the FDE. 1589 if (stack_decrement_ != nullptr) { 1590 // Advance LOC to just past the stack decrement. 1591 uint32_t pc = NEXT_LIR(stack_decrement_)->offset; 1592 AdvanceLoc(*cfi_info, pc); 1593 1594 // Now update the offset to the call frame: DW_CFA_def_cfa_offset frame_size. 1595 cfi_info->push_back(0x0e); 1596 EncodeUnsignedLeb128(*cfi_info, frame_size_); 1597 1598 // Handle register spills 1599 const uint32_t kSpillInstLen = (cu_->target64) ? 5 : 4; 1600 const int kDataAlignmentFactor = (cu_->target64) ? -8 : -4; 1601 uint32_t mask = core_spill_mask_ & ~(1 << rs_rRET.GetRegNum()); 1602 int offset = -(GetInstructionSetPointerSize(cu_->instruction_set) * num_core_spills_); 1603 for (int reg = 0; mask; mask >>= 1, reg++) { 1604 if (mask & 0x1) { 1605 pc += kSpillInstLen; 1606 1607 // Advance LOC to pass this instruction 1608 AdvanceLoc(*cfi_info, kSpillInstLen); 1609 1610 int dwarf_reg_id; 1611 if (ARTRegIDToDWARFRegID(cu_->target64, reg, &dwarf_reg_id)) { 1612 // DW_CFA_offset_extended_sf reg_no offset 1613 cfi_info->push_back(0x11); 1614 EncodeUnsignedLeb128(*cfi_info, dwarf_reg_id); 1615 EncodeSignedLeb128(*cfi_info, offset / kDataAlignmentFactor); 1616 } 1617 1618 offset += GetInstructionSetPointerSize(cu_->instruction_set); 1619 } 1620 } 1621 1622 // We continue with that stack until the epilogue. 1623 if (stack_increment_ != nullptr) { 1624 uint32_t new_pc = NEXT_LIR(stack_increment_)->offset; 1625 AdvanceLoc(*cfi_info, new_pc - pc); 1626 1627 // We probably have code snippets after the epilogue, so save the 1628 // current state: DW_CFA_remember_state. 1629 cfi_info->push_back(0x0a); 1630 1631 // We have now popped the stack: DW_CFA_def_cfa_offset 4/8. 1632 // There is only the return PC on the stack now. 1633 cfi_info->push_back(0x0e); 1634 EncodeUnsignedLeb128(*cfi_info, GetInstructionSetPointerSize(cu_->instruction_set)); 1635 1636 // Everything after that is the same as before the epilogue. 1637 // Stack bump was followed by RET instruction. 1638 LIR *post_ret_insn = NEXT_LIR(NEXT_LIR(stack_increment_)); 1639 if (post_ret_insn != nullptr) { 1640 pc = new_pc; 1641 new_pc = post_ret_insn->offset; 1642 AdvanceLoc(*cfi_info, new_pc - pc); 1643 // Restore the state: DW_CFA_restore_state. 1644 cfi_info->push_back(0x0b); 1645 } 1646 } 1647 } 1648 1649 // Padding to a multiple of 4 1650 while ((cfi_info->size() & 3) != 0) { 1651 // DW_CFA_nop is encoded as 0. 1652 cfi_info->push_back(0); 1653 } 1654 1655 // Set the length of the FDE inside the generated bytes. 1656 uint32_t length = cfi_info->size() - 4; 1657 (*cfi_info)[0] = length; 1658 (*cfi_info)[1] = length >> 8; 1659 (*cfi_info)[2] = length >> 16; 1660 (*cfi_info)[3] = length >> 24; 1661 return cfi_info; 1662} 1663 1664void X86Mir2Lir::GenMachineSpecificExtendedMethodMIR(BasicBlock* bb, MIR* mir) { 1665 switch (static_cast<ExtendedMIROpcode>(mir->dalvikInsn.opcode)) { 1666 case kMirOpReserveVectorRegisters: 1667 ReserveVectorRegisters(mir); 1668 break; 1669 case kMirOpReturnVectorRegisters: 1670 ReturnVectorRegisters(); 1671 break; 1672 case kMirOpConstVector: 1673 GenConst128(bb, mir); 1674 break; 1675 case kMirOpMoveVector: 1676 GenMoveVector(bb, mir); 1677 break; 1678 case kMirOpPackedMultiply: 1679 GenMultiplyVector(bb, mir); 1680 break; 1681 case kMirOpPackedAddition: 1682 GenAddVector(bb, mir); 1683 break; 1684 case kMirOpPackedSubtract: 1685 GenSubtractVector(bb, mir); 1686 break; 1687 case kMirOpPackedShiftLeft: 1688 GenShiftLeftVector(bb, mir); 1689 break; 1690 case kMirOpPackedSignedShiftRight: 1691 GenSignedShiftRightVector(bb, mir); 1692 break; 1693 case kMirOpPackedUnsignedShiftRight: 1694 GenUnsignedShiftRightVector(bb, mir); 1695 break; 1696 case kMirOpPackedAnd: 1697 GenAndVector(bb, mir); 1698 break; 1699 case kMirOpPackedOr: 1700 GenOrVector(bb, mir); 1701 break; 1702 case kMirOpPackedXor: 1703 GenXorVector(bb, mir); 1704 break; 1705 case kMirOpPackedAddReduce: 1706 GenAddReduceVector(bb, mir); 1707 break; 1708 case kMirOpPackedReduce: 1709 GenReduceVector(bb, mir); 1710 break; 1711 case kMirOpPackedSet: 1712 GenSetVector(bb, mir); 1713 break; 1714 default: 1715 break; 1716 } 1717} 1718 1719void X86Mir2Lir::ReserveVectorRegisters(MIR* mir) { 1720 // We should not try to reserve twice without returning the registers 1721 DCHECK_NE(num_reserved_vector_regs_, -1); 1722 1723 int num_vector_reg = mir->dalvikInsn.vA; 1724 for (int i = 0; i < num_vector_reg; i++) { 1725 RegStorage xp_reg = RegStorage::Solo128(i); 1726 RegisterInfo *xp_reg_info = GetRegInfo(xp_reg); 1727 Clobber(xp_reg); 1728 1729 for (RegisterInfo *info = xp_reg_info->GetAliasChain(); 1730 info != nullptr; 1731 info = info->GetAliasChain()) { 1732 if (info->GetReg().IsSingle()) { 1733 reg_pool_->sp_regs_.Delete(info); 1734 } else { 1735 reg_pool_->dp_regs_.Delete(info); 1736 } 1737 } 1738 } 1739 1740 num_reserved_vector_regs_ = num_vector_reg; 1741} 1742 1743void X86Mir2Lir::ReturnVectorRegisters() { 1744 // Return all the reserved registers 1745 for (int i = 0; i < num_reserved_vector_regs_; i++) { 1746 RegStorage xp_reg = RegStorage::Solo128(i); 1747 RegisterInfo *xp_reg_info = GetRegInfo(xp_reg); 1748 1749 for (RegisterInfo *info = xp_reg_info->GetAliasChain(); 1750 info != nullptr; 1751 info = info->GetAliasChain()) { 1752 if (info->GetReg().IsSingle()) { 1753 reg_pool_->sp_regs_.Insert(info); 1754 } else { 1755 reg_pool_->dp_regs_.Insert(info); 1756 } 1757 } 1758 } 1759 1760 // We don't have anymore reserved vector registers 1761 num_reserved_vector_regs_ = -1; 1762} 1763 1764void X86Mir2Lir::GenConst128(BasicBlock* bb, MIR* mir) { 1765 store_method_addr_used_ = true; 1766 int type_size = mir->dalvikInsn.vB; 1767 // We support 128 bit vectors. 1768 DCHECK_EQ(type_size & 0xFFFF, 128); 1769 RegStorage rs_dest = RegStorage::Solo128(mir->dalvikInsn.vA); 1770 uint32_t *args = mir->dalvikInsn.arg; 1771 int reg = rs_dest.GetReg(); 1772 // Check for all 0 case. 1773 if (args[0] == 0 && args[1] == 0 && args[2] == 0 && args[3] == 0) { 1774 NewLIR2(kX86XorpsRR, reg, reg); 1775 return; 1776 } 1777 1778 // Append the mov const vector to reg opcode. 1779 AppendOpcodeWithConst(kX86MovupsRM, reg, mir); 1780} 1781 1782void X86Mir2Lir::AppendOpcodeWithConst(X86OpCode opcode, int reg, MIR* mir) { 1783 // Okay, load it from the constant vector area. 1784 LIR *data_target = ScanVectorLiteral(mir); 1785 if (data_target == nullptr) { 1786 data_target = AddVectorLiteral(mir); 1787 } 1788 1789 // Address the start of the method. 1790 RegLocation rl_method = mir_graph_->GetRegLocation(base_of_code_->s_reg_low); 1791 if (rl_method.wide) { 1792 rl_method = LoadValueWide(rl_method, kCoreReg); 1793 } else { 1794 rl_method = LoadValue(rl_method, kCoreReg); 1795 } 1796 1797 // Load the proper value from the literal area. 1798 // We don't know the proper offset for the value, so pick one that will force 1799 // 4 byte offset. We will fix this up in the assembler later to have the right 1800 // value. 1801 ScopedMemRefType mem_ref_type(this, ResourceMask::kLiteral); 1802 LIR *load = NewLIR2(opcode, reg, rl_method.reg.GetReg()); 1803 load->flags.fixup = kFixupLoad; 1804 load->target = data_target; 1805} 1806 1807void X86Mir2Lir::GenMoveVector(BasicBlock *bb, MIR *mir) { 1808 // We only support 128 bit registers. 1809 DCHECK_EQ(mir->dalvikInsn.vC & 0xFFFF, 128U); 1810 RegStorage rs_dest = RegStorage::Solo128(mir->dalvikInsn.vA); 1811 RegStorage rs_src = RegStorage::Solo128(mir->dalvikInsn.vB); 1812 NewLIR2(kX86Mova128RR, rs_dest.GetReg(), rs_src.GetReg()); 1813} 1814 1815void X86Mir2Lir::GenMultiplyVectorSignedByte(BasicBlock *bb, MIR *mir) { 1816 const int BYTE_SIZE = 8; 1817 RegStorage rs_dest_src1 = RegStorage::Solo128(mir->dalvikInsn.vA); 1818 RegStorage rs_src2 = RegStorage::Solo128(mir->dalvikInsn.vB); 1819 RegStorage rs_src1_high_tmp = Get128BitRegister(AllocTempWide()); 1820 1821 /* 1822 * Emulate the behavior of a kSignedByte by separating out the 16 values in the two XMM 1823 * and multiplying 8 at a time before recombining back into one XMM register. 1824 * 1825 * let xmm1, xmm2 be real srcs (keep low bits of 16bit lanes) 1826 * xmm3 is tmp (operate on high bits of 16bit lanes) 1827 * 1828 * xmm3 = xmm1 1829 * xmm1 = xmm1 .* xmm2 1830 * xmm1 = xmm1 & 0x00ff00ff00ff00ff00ff00ff00ff00ff // xmm1 now has low bits 1831 * xmm3 = xmm3 .>> 8 1832 * xmm2 = xmm2 & 0xff00ff00ff00ff00ff00ff00ff00ff00 1833 * xmm2 = xmm2 .* xmm3 // xmm2 now has high bits 1834 * xmm1 = xmm1 | xmm2 // combine results 1835 */ 1836 1837 // Copy xmm1. 1838 NewLIR2(kX86Mova128RR, rs_src1_high_tmp.GetReg(), rs_dest_src1.GetReg()); 1839 1840 // Multiply low bits. 1841 NewLIR2(kX86PmullwRR, rs_dest_src1.GetReg(), rs_src2.GetReg()); 1842 1843 // xmm1 now has low bits. 1844 AndMaskVectorRegister(rs_dest_src1, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF); 1845 1846 // Prepare high bits for multiplication. 1847 NewLIR2(kX86PsrlwRI, rs_src1_high_tmp.GetReg(), BYTE_SIZE); 1848 AndMaskVectorRegister(rs_src2, 0xFF00FF00, 0xFF00FF00, 0xFF00FF00, 0xFF00FF00); 1849 1850 // Multiply high bits and xmm2 now has high bits. 1851 NewLIR2(kX86PmullwRR, rs_src2.GetReg(), rs_src1_high_tmp.GetReg()); 1852 1853 // Combine back into dest XMM register. 1854 NewLIR2(kX86PorRR, rs_dest_src1.GetReg(), rs_src2.GetReg()); 1855} 1856 1857void X86Mir2Lir::GenMultiplyVector(BasicBlock *bb, MIR *mir) { 1858 DCHECK_EQ(mir->dalvikInsn.vC & 0xFFFF, 128U); 1859 OpSize opsize = static_cast<OpSize>(mir->dalvikInsn.vC >> 16); 1860 RegStorage rs_dest_src1 = RegStorage::Solo128(mir->dalvikInsn.vA); 1861 RegStorage rs_src2 = RegStorage::Solo128(mir->dalvikInsn.vB); 1862 int opcode = 0; 1863 switch (opsize) { 1864 case k32: 1865 opcode = kX86PmulldRR; 1866 break; 1867 case kSignedHalf: 1868 opcode = kX86PmullwRR; 1869 break; 1870 case kSingle: 1871 opcode = kX86MulpsRR; 1872 break; 1873 case kDouble: 1874 opcode = kX86MulpdRR; 1875 break; 1876 case kSignedByte: 1877 // HW doesn't support 16x16 byte multiplication so emulate it. 1878 GenMultiplyVectorSignedByte(bb, mir); 1879 return; 1880 default: 1881 LOG(FATAL) << "Unsupported vector multiply " << opsize; 1882 break; 1883 } 1884 NewLIR2(opcode, rs_dest_src1.GetReg(), rs_src2.GetReg()); 1885} 1886 1887void X86Mir2Lir::GenAddVector(BasicBlock *bb, MIR *mir) { 1888 DCHECK_EQ(mir->dalvikInsn.vC & 0xFFFF, 128U); 1889 OpSize opsize = static_cast<OpSize>(mir->dalvikInsn.vC >> 16); 1890 RegStorage rs_dest_src1 = RegStorage::Solo128(mir->dalvikInsn.vA); 1891 RegStorage rs_src2 = RegStorage::Solo128(mir->dalvikInsn.vB); 1892 int opcode = 0; 1893 switch (opsize) { 1894 case k32: 1895 opcode = kX86PadddRR; 1896 break; 1897 case kSignedHalf: 1898 case kUnsignedHalf: 1899 opcode = kX86PaddwRR; 1900 break; 1901 case kUnsignedByte: 1902 case kSignedByte: 1903 opcode = kX86PaddbRR; 1904 break; 1905 case kSingle: 1906 opcode = kX86AddpsRR; 1907 break; 1908 case kDouble: 1909 opcode = kX86AddpdRR; 1910 break; 1911 default: 1912 LOG(FATAL) << "Unsupported vector addition " << opsize; 1913 break; 1914 } 1915 NewLIR2(opcode, rs_dest_src1.GetReg(), rs_src2.GetReg()); 1916} 1917 1918void X86Mir2Lir::GenSubtractVector(BasicBlock *bb, MIR *mir) { 1919 DCHECK_EQ(mir->dalvikInsn.vC & 0xFFFF, 128U); 1920 OpSize opsize = static_cast<OpSize>(mir->dalvikInsn.vC >> 16); 1921 RegStorage rs_dest_src1 = RegStorage::Solo128(mir->dalvikInsn.vA); 1922 RegStorage rs_src2 = RegStorage::Solo128(mir->dalvikInsn.vB); 1923 int opcode = 0; 1924 switch (opsize) { 1925 case k32: 1926 opcode = kX86PsubdRR; 1927 break; 1928 case kSignedHalf: 1929 case kUnsignedHalf: 1930 opcode = kX86PsubwRR; 1931 break; 1932 case kUnsignedByte: 1933 case kSignedByte: 1934 opcode = kX86PsubbRR; 1935 break; 1936 case kSingle: 1937 opcode = kX86SubpsRR; 1938 break; 1939 case kDouble: 1940 opcode = kX86SubpdRR; 1941 break; 1942 default: 1943 LOG(FATAL) << "Unsupported vector subtraction " << opsize; 1944 break; 1945 } 1946 NewLIR2(opcode, rs_dest_src1.GetReg(), rs_src2.GetReg()); 1947} 1948 1949void X86Mir2Lir::GenShiftByteVector(BasicBlock *bb, MIR *mir) { 1950 RegStorage rs_dest_src1 = RegStorage::Solo128(mir->dalvikInsn.vA); 1951 RegStorage rs_tmp = Get128BitRegister(AllocTempWide()); 1952 1953 int opcode = 0; 1954 int imm = mir->dalvikInsn.vB; 1955 1956 switch (static_cast<ExtendedMIROpcode>(mir->dalvikInsn.opcode)) { 1957 case kMirOpPackedShiftLeft: 1958 opcode = kX86PsllwRI; 1959 break; 1960 case kMirOpPackedSignedShiftRight: 1961 opcode = kX86PsrawRI; 1962 break; 1963 case kMirOpPackedUnsignedShiftRight: 1964 opcode = kX86PsrlwRI; 1965 break; 1966 default: 1967 LOG(FATAL) << "Unsupported shift operation on byte vector " << opcode; 1968 break; 1969 } 1970 1971 /* 1972 * xmm1 will have low bits 1973 * xmm2 will have high bits 1974 * 1975 * xmm2 = xmm1 1976 * xmm1 = xmm1 .<< N 1977 * xmm2 = xmm2 && 0xFF00FF00FF00FF00FF00FF00FF00FF00 1978 * xmm2 = xmm2 .<< N 1979 * xmm1 = xmm1 | xmm2 1980 */ 1981 1982 // Copy xmm1. 1983 NewLIR2(kX86Mova128RR, rs_tmp.GetReg(), rs_dest_src1.GetReg()); 1984 1985 // Shift lower values. 1986 NewLIR2(opcode, rs_dest_src1.GetReg(), imm); 1987 1988 // Mask bottom bits. 1989 AndMaskVectorRegister(rs_tmp, 0xFF00FF00, 0xFF00FF00, 0xFF00FF00, 0xFF00FF00); 1990 1991 // Shift higher values. 1992 NewLIR2(opcode, rs_tmp.GetReg(), imm); 1993 1994 // Combine back into dest XMM register. 1995 NewLIR2(kX86PorRR, rs_dest_src1.GetReg(), rs_tmp.GetReg()); 1996} 1997 1998void X86Mir2Lir::GenShiftLeftVector(BasicBlock *bb, MIR *mir) { 1999 DCHECK_EQ(mir->dalvikInsn.vC & 0xFFFF, 128U); 2000 OpSize opsize = static_cast<OpSize>(mir->dalvikInsn.vC >> 16); 2001 RegStorage rs_dest_src1 = RegStorage::Solo128(mir->dalvikInsn.vA); 2002 int imm = mir->dalvikInsn.vB; 2003 int opcode = 0; 2004 switch (opsize) { 2005 case k32: 2006 opcode = kX86PslldRI; 2007 break; 2008 case k64: 2009 opcode = kX86PsllqRI; 2010 break; 2011 case kSignedHalf: 2012 case kUnsignedHalf: 2013 opcode = kX86PsllwRI; 2014 break; 2015 case kSignedByte: 2016 case kUnsignedByte: 2017 GenShiftByteVector(bb, mir); 2018 return; 2019 default: 2020 LOG(FATAL) << "Unsupported vector shift left " << opsize; 2021 break; 2022 } 2023 NewLIR2(opcode, rs_dest_src1.GetReg(), imm); 2024} 2025 2026void X86Mir2Lir::GenSignedShiftRightVector(BasicBlock *bb, MIR *mir) { 2027 DCHECK_EQ(mir->dalvikInsn.vC & 0xFFFF, 128U); 2028 OpSize opsize = static_cast<OpSize>(mir->dalvikInsn.vC >> 16); 2029 RegStorage rs_dest_src1 = RegStorage::Solo128(mir->dalvikInsn.vA); 2030 int imm = mir->dalvikInsn.vB; 2031 int opcode = 0; 2032 switch (opsize) { 2033 case k32: 2034 opcode = kX86PsradRI; 2035 break; 2036 case kSignedHalf: 2037 case kUnsignedHalf: 2038 opcode = kX86PsrawRI; 2039 break; 2040 case kSignedByte: 2041 case kUnsignedByte: 2042 GenShiftByteVector(bb, mir); 2043 return; 2044 default: 2045 LOG(FATAL) << "Unsupported vector signed shift right " << opsize; 2046 break; 2047 } 2048 NewLIR2(opcode, rs_dest_src1.GetReg(), imm); 2049} 2050 2051void X86Mir2Lir::GenUnsignedShiftRightVector(BasicBlock *bb, MIR *mir) { 2052 DCHECK_EQ(mir->dalvikInsn.vC & 0xFFFF, 128U); 2053 OpSize opsize = static_cast<OpSize>(mir->dalvikInsn.vC >> 16); 2054 RegStorage rs_dest_src1 = RegStorage::Solo128(mir->dalvikInsn.vA); 2055 int imm = mir->dalvikInsn.vB; 2056 int opcode = 0; 2057 switch (opsize) { 2058 case k32: 2059 opcode = kX86PsrldRI; 2060 break; 2061 case k64: 2062 opcode = kX86PsrlqRI; 2063 break; 2064 case kSignedHalf: 2065 case kUnsignedHalf: 2066 opcode = kX86PsrlwRI; 2067 break; 2068 case kSignedByte: 2069 case kUnsignedByte: 2070 GenShiftByteVector(bb, mir); 2071 return; 2072 default: 2073 LOG(FATAL) << "Unsupported vector unsigned shift right " << opsize; 2074 break; 2075 } 2076 NewLIR2(opcode, rs_dest_src1.GetReg(), imm); 2077} 2078 2079void X86Mir2Lir::GenAndVector(BasicBlock *bb, MIR *mir) { 2080 // We only support 128 bit registers. 2081 DCHECK_EQ(mir->dalvikInsn.vC & 0xFFFF, 128U); 2082 RegStorage rs_dest_src1 = RegStorage::Solo128(mir->dalvikInsn.vA); 2083 RegStorage rs_src2 = RegStorage::Solo128(mir->dalvikInsn.vB); 2084 NewLIR2(kX86PandRR, rs_dest_src1.GetReg(), rs_src2.GetReg()); 2085} 2086 2087void X86Mir2Lir::GenOrVector(BasicBlock *bb, MIR *mir) { 2088 // We only support 128 bit registers. 2089 DCHECK_EQ(mir->dalvikInsn.vC & 0xFFFF, 128U); 2090 RegStorage rs_dest_src1 = RegStorage::Solo128(mir->dalvikInsn.vA); 2091 RegStorage rs_src2 = RegStorage::Solo128(mir->dalvikInsn.vB); 2092 NewLIR2(kX86PorRR, rs_dest_src1.GetReg(), rs_src2.GetReg()); 2093} 2094 2095void X86Mir2Lir::GenXorVector(BasicBlock *bb, MIR *mir) { 2096 // We only support 128 bit registers. 2097 DCHECK_EQ(mir->dalvikInsn.vC & 0xFFFF, 128U); 2098 RegStorage rs_dest_src1 = RegStorage::Solo128(mir->dalvikInsn.vA); 2099 RegStorage rs_src2 = RegStorage::Solo128(mir->dalvikInsn.vB); 2100 NewLIR2(kX86PxorRR, rs_dest_src1.GetReg(), rs_src2.GetReg()); 2101} 2102 2103void X86Mir2Lir::AndMaskVectorRegister(RegStorage rs_src1, uint32_t m1, uint32_t m2, uint32_t m3, uint32_t m4) { 2104 MaskVectorRegister(kX86PandRM, rs_src1, m1, m2, m3, m4); 2105} 2106 2107void X86Mir2Lir::MaskVectorRegister(X86OpCode opcode, RegStorage rs_src1, uint32_t m0, uint32_t m1, uint32_t m2, uint32_t m3) { 2108 // Create temporary MIR as container for 128-bit binary mask. 2109 MIR const_mir; 2110 MIR* const_mirp = &const_mir; 2111 const_mirp->dalvikInsn.opcode = static_cast<Instruction::Code>(kMirOpConstVector); 2112 const_mirp->dalvikInsn.arg[0] = m0; 2113 const_mirp->dalvikInsn.arg[1] = m1; 2114 const_mirp->dalvikInsn.arg[2] = m2; 2115 const_mirp->dalvikInsn.arg[3] = m3; 2116 2117 // Mask vector with const from literal pool. 2118 AppendOpcodeWithConst(opcode, rs_src1.GetReg(), const_mirp); 2119} 2120 2121void X86Mir2Lir::GenAddReduceVector(BasicBlock *bb, MIR *mir) { 2122 OpSize opsize = static_cast<OpSize>(mir->dalvikInsn.vC >> 16); 2123 RegStorage rs_src1 = RegStorage::Solo128(mir->dalvikInsn.vB); 2124 RegLocation rl_dest = mir_graph_->GetDest(mir); 2125 RegStorage rs_tmp; 2126 2127 int vec_bytes = (mir->dalvikInsn.vC & 0xFFFF) / 8; 2128 int vec_unit_size = 0; 2129 int opcode = 0; 2130 int extr_opcode = 0; 2131 RegLocation rl_result; 2132 2133 switch (opsize) { 2134 case k32: 2135 extr_opcode = kX86PextrdRRI; 2136 opcode = kX86PhadddRR; 2137 vec_unit_size = 4; 2138 break; 2139 case kSignedByte: 2140 case kUnsignedByte: 2141 extr_opcode = kX86PextrbRRI; 2142 opcode = kX86PhaddwRR; 2143 vec_unit_size = 2; 2144 break; 2145 case kSignedHalf: 2146 case kUnsignedHalf: 2147 extr_opcode = kX86PextrwRRI; 2148 opcode = kX86PhaddwRR; 2149 vec_unit_size = 2; 2150 break; 2151 case kSingle: 2152 rl_result = EvalLoc(rl_dest, kFPReg, true); 2153 vec_unit_size = 4; 2154 for (int i = 0; i < 3; i++) { 2155 NewLIR2(kX86AddssRR, rl_result.reg.GetReg(), rs_src1.GetReg()); 2156 NewLIR3(kX86ShufpsRRI, rs_src1.GetReg(), rs_src1.GetReg(), 0x39); 2157 } 2158 NewLIR2(kX86AddssRR, rl_result.reg.GetReg(), rs_src1.GetReg()); 2159 StoreValue(rl_dest, rl_result); 2160 2161 // For single-precision floats, we are done here 2162 return; 2163 default: 2164 LOG(FATAL) << "Unsupported vector add reduce " << opsize; 2165 break; 2166 } 2167 2168 int elems = vec_bytes / vec_unit_size; 2169 2170 // Emulate horizontal add instruction by reducing 2 vectors with 8 values before adding them again 2171 // TODO is overflow handled correctly? 2172 if (opsize == kSignedByte || opsize == kUnsignedByte) { 2173 rs_tmp = Get128BitRegister(AllocTempWide()); 2174 2175 // tmp = xmm1 .>> 8. 2176 NewLIR2(kX86Mova128RR, rs_tmp.GetReg(), rs_src1.GetReg()); 2177 NewLIR2(kX86PsrlwRI, rs_tmp.GetReg(), 8); 2178 2179 // Zero extend low bits in xmm1. 2180 AndMaskVectorRegister(rs_src1, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF); 2181 } 2182 2183 while (elems > 1) { 2184 if (opsize == kSignedByte || opsize == kUnsignedByte) { 2185 NewLIR2(opcode, rs_tmp.GetReg(), rs_tmp.GetReg()); 2186 } 2187 NewLIR2(opcode, rs_src1.GetReg(), rs_src1.GetReg()); 2188 elems >>= 1; 2189 } 2190 2191 // Combine the results if we separated them. 2192 if (opsize == kSignedByte || opsize == kUnsignedByte) { 2193 NewLIR2(kX86PaddbRR, rs_src1.GetReg(), rs_tmp.GetReg()); 2194 } 2195 2196 // We need to extract to a GPR. 2197 RegStorage temp = AllocTemp(); 2198 NewLIR3(extr_opcode, temp.GetReg(), rs_src1.GetReg(), 0); 2199 2200 // Can we do this directly into memory? 2201 rl_result = UpdateLocTyped(rl_dest, kCoreReg); 2202 if (rl_result.location == kLocPhysReg) { 2203 // Ensure res is in a core reg 2204 rl_result = EvalLoc(rl_dest, kCoreReg, true); 2205 OpRegReg(kOpAdd, rl_result.reg, temp); 2206 StoreFinalValue(rl_dest, rl_result); 2207 } else { 2208 OpMemReg(kOpAdd, rl_result, temp.GetReg()); 2209 } 2210 2211 FreeTemp(temp); 2212} 2213 2214void X86Mir2Lir::GenReduceVector(BasicBlock *bb, MIR *mir) { 2215 OpSize opsize = static_cast<OpSize>(mir->dalvikInsn.vC >> 16); 2216 RegLocation rl_dest = mir_graph_->GetDest(mir); 2217 RegStorage rs_src1 = RegStorage::Solo128(mir->dalvikInsn.vB); 2218 int extract_index = mir->dalvikInsn.arg[0]; 2219 int extr_opcode = 0; 2220 RegLocation rl_result; 2221 bool is_wide = false; 2222 2223 switch (opsize) { 2224 case k32: 2225 rl_result = UpdateLocTyped(rl_dest, kCoreReg); 2226 extr_opcode = (rl_result.location == kLocPhysReg) ? kX86PextrdMRI : kX86PextrdRRI; 2227 break; 2228 case kSignedHalf: 2229 case kUnsignedHalf: 2230 rl_result= UpdateLocTyped(rl_dest, kCoreReg); 2231 extr_opcode = (rl_result.location == kLocPhysReg) ? kX86PextrwMRI : kX86PextrwRRI; 2232 break; 2233 default: 2234 LOG(FATAL) << "Unsupported vector add reduce " << opsize; 2235 return; 2236 break; 2237 } 2238 2239 if (rl_result.location == kLocPhysReg) { 2240 NewLIR3(extr_opcode, rl_result.reg.GetReg(), rs_src1.GetReg(), extract_index); 2241 if (is_wide == true) { 2242 StoreFinalValue(rl_dest, rl_result); 2243 } else { 2244 StoreFinalValueWide(rl_dest, rl_result); 2245 } 2246 } else { 2247 int displacement = SRegOffset(rl_result.s_reg_low); 2248 LIR *l = NewLIR3(extr_opcode, rs_rX86_SP.GetReg(), displacement, rs_src1.GetReg()); 2249 AnnotateDalvikRegAccess(l, displacement >> 2, true /* is_load */, is_wide /* is_64bit */); 2250 AnnotateDalvikRegAccess(l, displacement >> 2, false /* is_load */, is_wide /* is_64bit */); 2251 } 2252} 2253 2254void X86Mir2Lir::GenSetVector(BasicBlock *bb, MIR *mir) { 2255 DCHECK_EQ(mir->dalvikInsn.vC & 0xFFFF, 128U); 2256 OpSize opsize = static_cast<OpSize>(mir->dalvikInsn.vC >> 16); 2257 RegStorage rs_dest = RegStorage::Solo128(mir->dalvikInsn.vA); 2258 int op_low = 0, op_high = 0, imm = 0, op_mov = kX86MovdxrRR; 2259 RegisterClass reg_type = kCoreReg; 2260 2261 switch (opsize) { 2262 case k32: 2263 op_low = kX86PshufdRRI; 2264 break; 2265 case kSingle: 2266 op_low = kX86PshufdRRI; 2267 op_mov = kX86Mova128RR; 2268 reg_type = kFPReg; 2269 break; 2270 case k64: 2271 op_low = kX86PshufdRRI; 2272 imm = 0x44; 2273 break; 2274 case kDouble: 2275 op_low = kX86PshufdRRI; 2276 op_mov = kX86Mova128RR; 2277 reg_type = kFPReg; 2278 imm = 0x44; 2279 break; 2280 case kSignedByte: 2281 case kUnsignedByte: 2282 // Shuffle 8 bit value into 16 bit word. 2283 // We set val = val + (val << 8) below and use 16 bit shuffle. 2284 case kSignedHalf: 2285 case kUnsignedHalf: 2286 // Handles low quadword. 2287 op_low = kX86PshuflwRRI; 2288 // Handles upper quadword. 2289 op_high = kX86PshufdRRI; 2290 break; 2291 default: 2292 LOG(FATAL) << "Unsupported vector set " << opsize; 2293 break; 2294 } 2295 2296 RegLocation rl_src = mir_graph_->GetSrc(mir, 0); 2297 2298 // Load the value from the VR into the reg. 2299 if (rl_src.wide == 0) { 2300 rl_src = LoadValue(rl_src, reg_type); 2301 } else { 2302 rl_src = LoadValueWide(rl_src, reg_type); 2303 } 2304 2305 // If opsize is 8 bits wide then double value and use 16 bit shuffle instead. 2306 if (opsize == kSignedByte || opsize == kUnsignedByte) { 2307 RegStorage temp = AllocTemp(); 2308 // val = val + (val << 8). 2309 NewLIR2(kX86Mov32RR, temp.GetReg(), rl_src.reg.GetReg()); 2310 NewLIR2(kX86Sal32RI, temp.GetReg(), 8); 2311 NewLIR2(kX86Or32RR, rl_src.reg.GetReg(), temp.GetReg()); 2312 FreeTemp(temp); 2313 } 2314 2315 // Load the value into the XMM register. 2316 NewLIR2(op_mov, rs_dest.GetReg(), rl_src.reg.GetReg()); 2317 2318 // Now shuffle the value across the destination. 2319 NewLIR3(op_low, rs_dest.GetReg(), rs_dest.GetReg(), imm); 2320 2321 // And then repeat as needed. 2322 if (op_high != 0) { 2323 NewLIR3(op_high, rs_dest.GetReg(), rs_dest.GetReg(), imm); 2324 } 2325} 2326 2327LIR *X86Mir2Lir::ScanVectorLiteral(MIR *mir) { 2328 int *args = reinterpret_cast<int*>(mir->dalvikInsn.arg); 2329 for (LIR *p = const_vectors_; p != nullptr; p = p->next) { 2330 if (args[0] == p->operands[0] && args[1] == p->operands[1] && 2331 args[2] == p->operands[2] && args[3] == p->operands[3]) { 2332 return p; 2333 } 2334 } 2335 return nullptr; 2336} 2337 2338LIR *X86Mir2Lir::AddVectorLiteral(MIR *mir) { 2339 LIR* new_value = static_cast<LIR*>(arena_->Alloc(sizeof(LIR), kArenaAllocData)); 2340 int *args = reinterpret_cast<int*>(mir->dalvikInsn.arg); 2341 new_value->operands[0] = args[0]; 2342 new_value->operands[1] = args[1]; 2343 new_value->operands[2] = args[2]; 2344 new_value->operands[3] = args[3]; 2345 new_value->next = const_vectors_; 2346 if (const_vectors_ == nullptr) { 2347 estimated_native_code_size_ += 12; // Amount needed to align to 16 byte boundary. 2348 } 2349 estimated_native_code_size_ += 16; // Space for one vector. 2350 const_vectors_ = new_value; 2351 return new_value; 2352} 2353 2354// ------------ ABI support: mapping of args to physical registers ------------- 2355RegStorage X86Mir2Lir::InToRegStorageX86_64Mapper::GetNextReg(bool is_double_or_float, bool is_wide, 2356 bool is_ref) { 2357 const SpecialTargetRegister coreArgMappingToPhysicalReg[] = {kArg1, kArg2, kArg3, kArg4, kArg5}; 2358 const int coreArgMappingToPhysicalRegSize = sizeof(coreArgMappingToPhysicalReg) / 2359 sizeof(SpecialTargetRegister); 2360 const SpecialTargetRegister fpArgMappingToPhysicalReg[] = {kFArg0, kFArg1, kFArg2, kFArg3, 2361 kFArg4, kFArg5, kFArg6, kFArg7}; 2362 const int fpArgMappingToPhysicalRegSize = sizeof(fpArgMappingToPhysicalReg) / 2363 sizeof(SpecialTargetRegister); 2364 2365 if (is_double_or_float) { 2366 if (cur_fp_reg_ < fpArgMappingToPhysicalRegSize) { 2367 return ml_->TargetReg(fpArgMappingToPhysicalReg[cur_fp_reg_++], is_wide ? kWide : kNotWide); 2368 } 2369 } else { 2370 if (cur_core_reg_ < coreArgMappingToPhysicalRegSize) { 2371 return ml_->TargetReg(coreArgMappingToPhysicalReg[cur_core_reg_++], 2372 is_ref ? kRef : (is_wide ? kWide : kNotWide)); 2373 } 2374 } 2375 return RegStorage::InvalidReg(); 2376} 2377 2378RegStorage X86Mir2Lir::InToRegStorageMapping::Get(int in_position) { 2379 DCHECK(IsInitialized()); 2380 auto res = mapping_.find(in_position); 2381 return res != mapping_.end() ? res->second : RegStorage::InvalidReg(); 2382} 2383 2384void X86Mir2Lir::InToRegStorageMapping::Initialize(RegLocation* arg_locs, int count, 2385 InToRegStorageMapper* mapper) { 2386 DCHECK(mapper != nullptr); 2387 max_mapped_in_ = -1; 2388 is_there_stack_mapped_ = false; 2389 for (int in_position = 0; in_position < count; in_position++) { 2390 RegStorage reg = mapper->GetNextReg(arg_locs[in_position].fp, 2391 arg_locs[in_position].wide, arg_locs[in_position].ref); 2392 if (reg.Valid()) { 2393 mapping_[in_position] = reg; 2394 max_mapped_in_ = std::max(max_mapped_in_, in_position); 2395 if (arg_locs[in_position].wide) { 2396 // We covered 2 args, so skip the next one 2397 in_position++; 2398 } 2399 } else { 2400 is_there_stack_mapped_ = true; 2401 } 2402 } 2403 initialized_ = true; 2404} 2405 2406RegStorage X86Mir2Lir::GetArgMappingToPhysicalReg(int arg_num) { 2407 if (!cu_->target64) { 2408 return GetCoreArgMappingToPhysicalReg(arg_num); 2409 } 2410 2411 if (!in_to_reg_storage_mapping_.IsInitialized()) { 2412 int start_vreg = cu_->num_dalvik_registers - cu_->num_ins; 2413 RegLocation* arg_locs = &mir_graph_->reg_location_[start_vreg]; 2414 2415 InToRegStorageX86_64Mapper mapper(this); 2416 in_to_reg_storage_mapping_.Initialize(arg_locs, cu_->num_ins, &mapper); 2417 } 2418 return in_to_reg_storage_mapping_.Get(arg_num); 2419} 2420 2421RegStorage X86Mir2Lir::GetCoreArgMappingToPhysicalReg(int core_arg_num) { 2422 // For the 32-bit internal ABI, the first 3 arguments are passed in registers. 2423 // Not used for 64-bit, TODO: Move X86_32 to the same framework 2424 switch (core_arg_num) { 2425 case 0: 2426 return rs_rX86_ARG1; 2427 case 1: 2428 return rs_rX86_ARG2; 2429 case 2: 2430 return rs_rX86_ARG3; 2431 default: 2432 return RegStorage::InvalidReg(); 2433 } 2434} 2435 2436// ---------End of ABI support: mapping of args to physical registers ------------- 2437 2438/* 2439 * If there are any ins passed in registers that have not been promoted 2440 * to a callee-save register, flush them to the frame. Perform initial 2441 * assignment of promoted arguments. 2442 * 2443 * ArgLocs is an array of location records describing the incoming arguments 2444 * with one location record per word of argument. 2445 */ 2446void X86Mir2Lir::FlushIns(RegLocation* ArgLocs, RegLocation rl_method) { 2447 if (!cu_->target64) return Mir2Lir::FlushIns(ArgLocs, rl_method); 2448 /* 2449 * Dummy up a RegLocation for the incoming Method* 2450 * It will attempt to keep kArg0 live (or copy it to home location 2451 * if promoted). 2452 */ 2453 2454 RegLocation rl_src = rl_method; 2455 rl_src.location = kLocPhysReg; 2456 rl_src.reg = TargetReg(kArg0, kRef); 2457 rl_src.home = false; 2458 MarkLive(rl_src); 2459 StoreValue(rl_method, rl_src); 2460 // If Method* has been promoted, explicitly flush 2461 if (rl_method.location == kLocPhysReg) { 2462 StoreRefDisp(rs_rX86_SP, 0, As32BitReg(TargetReg(kArg0, kRef)), kNotVolatile); 2463 } 2464 2465 if (cu_->num_ins == 0) { 2466 return; 2467 } 2468 2469 int start_vreg = cu_->num_dalvik_registers - cu_->num_ins; 2470 /* 2471 * Copy incoming arguments to their proper home locations. 2472 * NOTE: an older version of dx had an issue in which 2473 * it would reuse static method argument registers. 2474 * This could result in the same Dalvik virtual register 2475 * being promoted to both core and fp regs. To account for this, 2476 * we only copy to the corresponding promoted physical register 2477 * if it matches the type of the SSA name for the incoming 2478 * argument. It is also possible that long and double arguments 2479 * end up half-promoted. In those cases, we must flush the promoted 2480 * half to memory as well. 2481 */ 2482 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 2483 for (int i = 0; i < cu_->num_ins; i++) { 2484 // get reg corresponding to input 2485 RegStorage reg = GetArgMappingToPhysicalReg(i); 2486 2487 RegLocation* t_loc = &ArgLocs[i]; 2488 if (reg.Valid()) { 2489 // If arriving in register. 2490 2491 // We have already updated the arg location with promoted info 2492 // so we can be based on it. 2493 if (t_loc->location == kLocPhysReg) { 2494 // Just copy it. 2495 OpRegCopy(t_loc->reg, reg); 2496 } else { 2497 // Needs flush. 2498 if (t_loc->ref) { 2499 StoreRefDisp(rs_rX86_SP, SRegOffset(start_vreg + i), reg, kNotVolatile); 2500 } else { 2501 StoreBaseDisp(rs_rX86_SP, SRegOffset(start_vreg + i), reg, t_loc->wide ? k64 : k32, 2502 kNotVolatile); 2503 } 2504 } 2505 } else { 2506 // If arriving in frame & promoted. 2507 if (t_loc->location == kLocPhysReg) { 2508 if (t_loc->ref) { 2509 LoadRefDisp(rs_rX86_SP, SRegOffset(start_vreg + i), t_loc->reg, kNotVolatile); 2510 } else { 2511 LoadBaseDisp(rs_rX86_SP, SRegOffset(start_vreg + i), t_loc->reg, 2512 t_loc->wide ? k64 : k32, kNotVolatile); 2513 } 2514 } 2515 } 2516 if (t_loc->wide) { 2517 // Increment i to skip the next one. 2518 i++; 2519 } 2520 } 2521} 2522 2523/* 2524 * Load up to 5 arguments, the first three of which will be in 2525 * kArg1 .. kArg3. On entry kArg0 contains the current method pointer, 2526 * and as part of the load sequence, it must be replaced with 2527 * the target method pointer. Note, this may also be called 2528 * for "range" variants if the number of arguments is 5 or fewer. 2529 */ 2530int X86Mir2Lir::GenDalvikArgsNoRange(CallInfo* info, 2531 int call_state, LIR** pcrLabel, NextCallInsn next_call_insn, 2532 const MethodReference& target_method, 2533 uint32_t vtable_idx, uintptr_t direct_code, 2534 uintptr_t direct_method, InvokeType type, bool skip_this) { 2535 if (!cu_->target64) { 2536 return Mir2Lir::GenDalvikArgsNoRange(info, 2537 call_state, pcrLabel, next_call_insn, 2538 target_method, 2539 vtable_idx, direct_code, 2540 direct_method, type, skip_this); 2541 } 2542 return GenDalvikArgsRange(info, 2543 call_state, pcrLabel, next_call_insn, 2544 target_method, 2545 vtable_idx, direct_code, 2546 direct_method, type, skip_this); 2547} 2548 2549/* 2550 * May have 0+ arguments (also used for jumbo). Note that 2551 * source virtual registers may be in physical registers, so may 2552 * need to be flushed to home location before copying. This 2553 * applies to arg3 and above (see below). 2554 * 2555 * Two general strategies: 2556 * If < 20 arguments 2557 * Pass args 3-18 using vldm/vstm block copy 2558 * Pass arg0, arg1 & arg2 in kArg1-kArg3 2559 * If 20+ arguments 2560 * Pass args arg19+ using memcpy block copy 2561 * Pass arg0, arg1 & arg2 in kArg1-kArg3 2562 * 2563 */ 2564int X86Mir2Lir::GenDalvikArgsRange(CallInfo* info, int call_state, 2565 LIR** pcrLabel, NextCallInsn next_call_insn, 2566 const MethodReference& target_method, 2567 uint32_t vtable_idx, uintptr_t direct_code, uintptr_t direct_method, 2568 InvokeType type, bool skip_this) { 2569 if (!cu_->target64) { 2570 return Mir2Lir::GenDalvikArgsRange(info, call_state, 2571 pcrLabel, next_call_insn, 2572 target_method, 2573 vtable_idx, direct_code, direct_method, 2574 type, skip_this); 2575 } 2576 2577 /* If no arguments, just return */ 2578 if (info->num_arg_words == 0) 2579 return call_state; 2580 2581 const int start_index = skip_this ? 1 : 0; 2582 2583 InToRegStorageX86_64Mapper mapper(this); 2584 InToRegStorageMapping in_to_reg_storage_mapping; 2585 in_to_reg_storage_mapping.Initialize(info->args, info->num_arg_words, &mapper); 2586 const int last_mapped_in = in_to_reg_storage_mapping.GetMaxMappedIn(); 2587 const int size_of_the_last_mapped = last_mapped_in == -1 ? 1 : 2588 info->args[last_mapped_in].wide ? 2 : 1; 2589 int regs_left_to_pass_via_stack = info->num_arg_words - (last_mapped_in + size_of_the_last_mapped); 2590 2591 // Fisrt of all, check whether it make sense to use bulk copying 2592 // Optimization is aplicable only for range case 2593 // TODO: make a constant instead of 2 2594 if (info->is_range && regs_left_to_pass_via_stack >= 2) { 2595 // Scan the rest of the args - if in phys_reg flush to memory 2596 for (int next_arg = last_mapped_in + size_of_the_last_mapped; next_arg < info->num_arg_words;) { 2597 RegLocation loc = info->args[next_arg]; 2598 if (loc.wide) { 2599 loc = UpdateLocWide(loc); 2600 if (loc.location == kLocPhysReg) { 2601 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 2602 StoreBaseDisp(rs_rX86_SP, SRegOffset(loc.s_reg_low), loc.reg, k64, kNotVolatile); 2603 } 2604 next_arg += 2; 2605 } else { 2606 loc = UpdateLoc(loc); 2607 if (loc.location == kLocPhysReg) { 2608 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 2609 StoreBaseDisp(rs_rX86_SP, SRegOffset(loc.s_reg_low), loc.reg, k32, kNotVolatile); 2610 } 2611 next_arg++; 2612 } 2613 } 2614 2615 // Logic below assumes that Method pointer is at offset zero from SP. 2616 DCHECK_EQ(VRegOffset(static_cast<int>(kVRegMethodPtrBaseReg)), 0); 2617 2618 // The rest can be copied together 2619 int start_offset = SRegOffset(info->args[last_mapped_in + size_of_the_last_mapped].s_reg_low); 2620 int outs_offset = StackVisitor::GetOutVROffset(last_mapped_in + size_of_the_last_mapped, 2621 cu_->instruction_set); 2622 2623 int current_src_offset = start_offset; 2624 int current_dest_offset = outs_offset; 2625 2626 // Only davik regs are accessed in this loop; no next_call_insn() calls. 2627 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 2628 while (regs_left_to_pass_via_stack > 0) { 2629 // This is based on the knowledge that the stack itself is 16-byte aligned. 2630 bool src_is_16b_aligned = (current_src_offset & 0xF) == 0; 2631 bool dest_is_16b_aligned = (current_dest_offset & 0xF) == 0; 2632 size_t bytes_to_move; 2633 2634 /* 2635 * The amount to move defaults to 32-bit. If there are 4 registers left to move, then do a 2636 * a 128-bit move because we won't get the chance to try to aligned. If there are more than 2637 * 4 registers left to move, consider doing a 128-bit only if either src or dest are aligned. 2638 * We do this because we could potentially do a smaller move to align. 2639 */ 2640 if (regs_left_to_pass_via_stack == 4 || 2641 (regs_left_to_pass_via_stack > 4 && (src_is_16b_aligned || dest_is_16b_aligned))) { 2642 // Moving 128-bits via xmm register. 2643 bytes_to_move = sizeof(uint32_t) * 4; 2644 2645 // Allocate a free xmm temp. Since we are working through the calling sequence, 2646 // we expect to have an xmm temporary available. AllocTempDouble will abort if 2647 // there are no free registers. 2648 RegStorage temp = AllocTempDouble(); 2649 2650 LIR* ld1 = nullptr; 2651 LIR* ld2 = nullptr; 2652 LIR* st1 = nullptr; 2653 LIR* st2 = nullptr; 2654 2655 /* 2656 * The logic is similar for both loads and stores. If we have 16-byte alignment, 2657 * do an aligned move. If we have 8-byte alignment, then do the move in two 2658 * parts. This approach prevents possible cache line splits. Finally, fall back 2659 * to doing an unaligned move. In most cases we likely won't split the cache 2660 * line but we cannot prove it and thus take a conservative approach. 2661 */ 2662 bool src_is_8b_aligned = (current_src_offset & 0x7) == 0; 2663 bool dest_is_8b_aligned = (current_dest_offset & 0x7) == 0; 2664 2665 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 2666 if (src_is_16b_aligned) { 2667 ld1 = OpMovRegMem(temp, rs_rX86_SP, current_src_offset, kMovA128FP); 2668 } else if (src_is_8b_aligned) { 2669 ld1 = OpMovRegMem(temp, rs_rX86_SP, current_src_offset, kMovLo128FP); 2670 ld2 = OpMovRegMem(temp, rs_rX86_SP, current_src_offset + (bytes_to_move >> 1), 2671 kMovHi128FP); 2672 } else { 2673 ld1 = OpMovRegMem(temp, rs_rX86_SP, current_src_offset, kMovU128FP); 2674 } 2675 2676 if (dest_is_16b_aligned) { 2677 st1 = OpMovMemReg(rs_rX86_SP, current_dest_offset, temp, kMovA128FP); 2678 } else if (dest_is_8b_aligned) { 2679 st1 = OpMovMemReg(rs_rX86_SP, current_dest_offset, temp, kMovLo128FP); 2680 st2 = OpMovMemReg(rs_rX86_SP, current_dest_offset + (bytes_to_move >> 1), 2681 temp, kMovHi128FP); 2682 } else { 2683 st1 = OpMovMemReg(rs_rX86_SP, current_dest_offset, temp, kMovU128FP); 2684 } 2685 2686 // TODO If we could keep track of aliasing information for memory accesses that are wider 2687 // than 64-bit, we wouldn't need to set up a barrier. 2688 if (ld1 != nullptr) { 2689 if (ld2 != nullptr) { 2690 // For 64-bit load we can actually set up the aliasing information. 2691 AnnotateDalvikRegAccess(ld1, current_src_offset >> 2, true, true); 2692 AnnotateDalvikRegAccess(ld2, (current_src_offset + (bytes_to_move >> 1)) >> 2, true, true); 2693 } else { 2694 // Set barrier for 128-bit load. 2695 ld1->u.m.def_mask = &kEncodeAll; 2696 } 2697 } 2698 if (st1 != nullptr) { 2699 if (st2 != nullptr) { 2700 // For 64-bit store we can actually set up the aliasing information. 2701 AnnotateDalvikRegAccess(st1, current_dest_offset >> 2, false, true); 2702 AnnotateDalvikRegAccess(st2, (current_dest_offset + (bytes_to_move >> 1)) >> 2, false, true); 2703 } else { 2704 // Set barrier for 128-bit store. 2705 st1->u.m.def_mask = &kEncodeAll; 2706 } 2707 } 2708 2709 // Free the temporary used for the data movement. 2710 FreeTemp(temp); 2711 } else { 2712 // Moving 32-bits via general purpose register. 2713 bytes_to_move = sizeof(uint32_t); 2714 2715 // Instead of allocating a new temp, simply reuse one of the registers being used 2716 // for argument passing. 2717 RegStorage temp = TargetReg(kArg3, kNotWide); 2718 2719 // Now load the argument VR and store to the outs. 2720 Load32Disp(rs_rX86_SP, current_src_offset, temp); 2721 Store32Disp(rs_rX86_SP, current_dest_offset, temp); 2722 } 2723 2724 current_src_offset += bytes_to_move; 2725 current_dest_offset += bytes_to_move; 2726 regs_left_to_pass_via_stack -= (bytes_to_move >> 2); 2727 } 2728 DCHECK_EQ(regs_left_to_pass_via_stack, 0); 2729 } 2730 2731 // Now handle rest not registers if they are 2732 if (in_to_reg_storage_mapping.IsThereStackMapped()) { 2733 RegStorage regSingle = TargetReg(kArg2, kNotWide); 2734 RegStorage regWide = TargetReg(kArg3, kWide); 2735 for (int i = start_index; 2736 i < last_mapped_in + size_of_the_last_mapped + regs_left_to_pass_via_stack; i++) { 2737 RegLocation rl_arg = info->args[i]; 2738 rl_arg = UpdateRawLoc(rl_arg); 2739 RegStorage reg = in_to_reg_storage_mapping.Get(i); 2740 if (!reg.Valid()) { 2741 int out_offset = StackVisitor::GetOutVROffset(i, cu_->instruction_set); 2742 2743 { 2744 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 2745 if (rl_arg.wide) { 2746 if (rl_arg.location == kLocPhysReg) { 2747 StoreBaseDisp(rs_rX86_SP, out_offset, rl_arg.reg, k64, kNotVolatile); 2748 } else { 2749 LoadValueDirectWideFixed(rl_arg, regWide); 2750 StoreBaseDisp(rs_rX86_SP, out_offset, regWide, k64, kNotVolatile); 2751 } 2752 } else { 2753 if (rl_arg.location == kLocPhysReg) { 2754 StoreBaseDisp(rs_rX86_SP, out_offset, rl_arg.reg, k32, kNotVolatile); 2755 } else { 2756 LoadValueDirectFixed(rl_arg, regSingle); 2757 StoreBaseDisp(rs_rX86_SP, out_offset, regSingle, k32, kNotVolatile); 2758 } 2759 } 2760 } 2761 call_state = next_call_insn(cu_, info, call_state, target_method, 2762 vtable_idx, direct_code, direct_method, type); 2763 } 2764 if (rl_arg.wide) { 2765 i++; 2766 } 2767 } 2768 } 2769 2770 // Finish with mapped registers 2771 for (int i = start_index; i <= last_mapped_in; i++) { 2772 RegLocation rl_arg = info->args[i]; 2773 rl_arg = UpdateRawLoc(rl_arg); 2774 RegStorage reg = in_to_reg_storage_mapping.Get(i); 2775 if (reg.Valid()) { 2776 if (rl_arg.wide) { 2777 LoadValueDirectWideFixed(rl_arg, reg); 2778 } else { 2779 LoadValueDirectFixed(rl_arg, reg); 2780 } 2781 call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx, 2782 direct_code, direct_method, type); 2783 } 2784 if (rl_arg.wide) { 2785 i++; 2786 } 2787 } 2788 2789 call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx, 2790 direct_code, direct_method, type); 2791 if (pcrLabel) { 2792 if (!cu_->compiler_driver->GetCompilerOptions().GetImplicitNullChecks()) { 2793 *pcrLabel = GenExplicitNullCheck(TargetReg(kArg1, kRef), info->opt_flags); 2794 } else { 2795 *pcrLabel = nullptr; 2796 // In lieu of generating a check for kArg1 being null, we need to 2797 // perform a load when doing implicit checks. 2798 RegStorage tmp = AllocTemp(); 2799 Load32Disp(TargetReg(kArg1, kRef), 0, tmp); 2800 MarkPossibleNullPointerException(info->opt_flags); 2801 FreeTemp(tmp); 2802 } 2803 } 2804 return call_state; 2805} 2806 2807bool X86Mir2Lir::GenInlinedCharAt(CallInfo* info) { 2808 // Location of reference to data array 2809 int value_offset = mirror::String::ValueOffset().Int32Value(); 2810 // Location of count 2811 int count_offset = mirror::String::CountOffset().Int32Value(); 2812 // Starting offset within data array 2813 int offset_offset = mirror::String::OffsetOffset().Int32Value(); 2814 // Start of char data with array_ 2815 int data_offset = mirror::Array::DataOffset(sizeof(uint16_t)).Int32Value(); 2816 2817 RegLocation rl_obj = info->args[0]; 2818 RegLocation rl_idx = info->args[1]; 2819 rl_obj = LoadValue(rl_obj, kRefReg); 2820 // X86 wants to avoid putting a constant index into a register. 2821 if (!rl_idx.is_const) { 2822 rl_idx = LoadValue(rl_idx, kCoreReg); 2823 } 2824 RegStorage reg_max; 2825 GenNullCheck(rl_obj.reg, info->opt_flags); 2826 bool range_check = (!(info->opt_flags & MIR_IGNORE_RANGE_CHECK)); 2827 LIR* range_check_branch = nullptr; 2828 RegStorage reg_off; 2829 RegStorage reg_ptr; 2830 if (range_check) { 2831 // On x86, we can compare to memory directly 2832 // Set up a launch pad to allow retry in case of bounds violation */ 2833 if (rl_idx.is_const) { 2834 LIR* comparison; 2835 range_check_branch = OpCmpMemImmBranch( 2836 kCondUlt, RegStorage::InvalidReg(), rl_obj.reg, count_offset, 2837 mir_graph_->ConstantValue(rl_idx.orig_sreg), nullptr, &comparison); 2838 MarkPossibleNullPointerExceptionAfter(0, comparison); 2839 } else { 2840 OpRegMem(kOpCmp, rl_idx.reg, rl_obj.reg, count_offset); 2841 MarkPossibleNullPointerException(0); 2842 range_check_branch = OpCondBranch(kCondUge, nullptr); 2843 } 2844 } 2845 reg_off = AllocTemp(); 2846 reg_ptr = AllocTempRef(); 2847 Load32Disp(rl_obj.reg, offset_offset, reg_off); 2848 LoadRefDisp(rl_obj.reg, value_offset, reg_ptr, kNotVolatile); 2849 if (rl_idx.is_const) { 2850 OpRegImm(kOpAdd, reg_off, mir_graph_->ConstantValue(rl_idx.orig_sreg)); 2851 } else { 2852 OpRegReg(kOpAdd, reg_off, rl_idx.reg); 2853 } 2854 FreeTemp(rl_obj.reg); 2855 if (rl_idx.location == kLocPhysReg) { 2856 FreeTemp(rl_idx.reg); 2857 } 2858 RegLocation rl_dest = InlineTarget(info); 2859 RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true); 2860 LoadBaseIndexedDisp(reg_ptr, reg_off, 1, data_offset, rl_result.reg, kUnsignedHalf); 2861 FreeTemp(reg_off); 2862 FreeTemp(reg_ptr); 2863 StoreValue(rl_dest, rl_result); 2864 if (range_check) { 2865 DCHECK(range_check_branch != nullptr); 2866 info->opt_flags |= MIR_IGNORE_NULL_CHECK; // Record that we've already null checked. 2867 AddIntrinsicSlowPath(info, range_check_branch); 2868 } 2869 return true; 2870} 2871 2872bool X86Mir2Lir::GenInlinedCurrentThread(CallInfo* info) { 2873 RegLocation rl_dest = InlineTarget(info); 2874 2875 // Early exit if the result is unused. 2876 if (rl_dest.orig_sreg < 0) { 2877 return true; 2878 } 2879 2880 RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true); 2881 2882 if (cu_->target64) { 2883 OpRegThreadMem(kOpMov, rl_result.reg, Thread::PeerOffset<8>()); 2884 } else { 2885 OpRegThreadMem(kOpMov, rl_result.reg, Thread::PeerOffset<4>()); 2886 } 2887 2888 StoreValue(rl_dest, rl_result); 2889 return true; 2890} 2891 2892} // namespace art 2893