gen_invoke.cc revision d85614222fa062ec809af9d65f04ab6b7dc1c248
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 "dex/compiler_ir.h" 18#include "dex/frontend.h" 19#include "dex/quick/dex_file_method_inliner.h" 20#include "dex/quick/dex_file_to_method_inliner_map.h" 21#include "dex_file-inl.h" 22#include "entrypoints/quick/quick_entrypoints.h" 23#include "invoke_type.h" 24#include "mirror/array.h" 25#include "mirror/class-inl.h" 26#include "mirror/dex_cache.h" 27#include "mirror/object_array-inl.h" 28#include "mirror/reference-inl.h" 29#include "mirror/string.h" 30#include "mir_to_lir-inl.h" 31#include "scoped_thread_state_change.h" 32#include "x86/codegen_x86.h" 33 34namespace art { 35 36// Shortcuts to repeatedly used long types. 37typedef mirror::ObjectArray<mirror::Object> ObjArray; 38 39/* 40 * This source files contains "gen" codegen routines that should 41 * be applicable to most targets. Only mid-level support utilities 42 * and "op" calls may be used here. 43 */ 44 45void Mir2Lir::AddIntrinsicSlowPath(CallInfo* info, LIR* branch, LIR* resume) { 46 class IntrinsicSlowPathPath : public Mir2Lir::LIRSlowPath { 47 public: 48 IntrinsicSlowPathPath(Mir2Lir* m2l, CallInfo* info, LIR* branch, LIR* resume = nullptr) 49 : LIRSlowPath(m2l, info->offset, branch, resume), info_(info) { 50 } 51 52 void Compile() { 53 m2l_->ResetRegPool(); 54 m2l_->ResetDefTracking(); 55 GenerateTargetLabel(kPseudoIntrinsicRetry); 56 // NOTE: GenInvokeNoInline() handles MarkSafepointPC. 57 m2l_->GenInvokeNoInline(info_); 58 if (cont_ != nullptr) { 59 m2l_->OpUnconditionalBranch(cont_); 60 } 61 } 62 63 private: 64 CallInfo* const info_; 65 }; 66 67 AddSlowPath(new (arena_) IntrinsicSlowPathPath(this, info, branch, resume)); 68} 69 70// Macro to help instantiate. 71// TODO: This might be used to only instantiate <4> on pure 32b systems. 72#define INSTANTIATE(sig_part1, ...) \ 73 template sig_part1(ThreadOffset<4>, __VA_ARGS__); \ 74 template sig_part1(ThreadOffset<8>, __VA_ARGS__); \ 75 76 77/* 78 * To save scheduling time, helper calls are broken into two parts: generation of 79 * the helper target address, and the actual call to the helper. Because x86 80 * has a memory call operation, part 1 is a NOP for x86. For other targets, 81 * load arguments between the two parts. 82 */ 83// template <size_t pointer_size> 84RegStorage Mir2Lir::CallHelperSetup(ThreadOffset<4> helper_offset) { 85 // All CallRuntimeHelperXXX call this first. So make a central check here. 86 DCHECK_EQ(4U, GetInstructionSetPointerSize(cu_->instruction_set)); 87 88 if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) { 89 return RegStorage::InvalidReg(); 90 } else { 91 return LoadHelper(helper_offset); 92 } 93} 94 95RegStorage Mir2Lir::CallHelperSetup(ThreadOffset<8> helper_offset) { 96 // All CallRuntimeHelperXXX call this first. So make a central check here. 97 DCHECK_EQ(8U, GetInstructionSetPointerSize(cu_->instruction_set)); 98 99 if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) { 100 return RegStorage::InvalidReg(); 101 } else { 102 return LoadHelper(helper_offset); 103 } 104} 105 106/* NOTE: if r_tgt is a temp, it will be freed following use */ 107template <size_t pointer_size> 108LIR* Mir2Lir::CallHelper(RegStorage r_tgt, ThreadOffset<pointer_size> helper_offset, 109 bool safepoint_pc, bool use_link) { 110 LIR* call_inst; 111 OpKind op = use_link ? kOpBlx : kOpBx; 112 if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) { 113 call_inst = OpThreadMem(op, helper_offset); 114 } else { 115 call_inst = OpReg(op, r_tgt); 116 FreeTemp(r_tgt); 117 } 118 if (safepoint_pc) { 119 MarkSafepointPC(call_inst); 120 } 121 return call_inst; 122} 123template LIR* Mir2Lir::CallHelper(RegStorage r_tgt, ThreadOffset<4> helper_offset, 124 bool safepoint_pc, bool use_link); 125template LIR* Mir2Lir::CallHelper(RegStorage r_tgt, ThreadOffset<8> helper_offset, 126 bool safepoint_pc, bool use_link); 127 128template <size_t pointer_size> 129void Mir2Lir::CallRuntimeHelper(ThreadOffset<pointer_size> helper_offset, bool safepoint_pc) { 130 RegStorage r_tgt = CallHelperSetup(helper_offset); 131 ClobberCallerSave(); 132 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 133} 134INSTANTIATE(void Mir2Lir::CallRuntimeHelper, bool safepoint_pc) 135 136template <size_t pointer_size> 137void Mir2Lir::CallRuntimeHelperImm(ThreadOffset<pointer_size> helper_offset, int arg0, 138 bool safepoint_pc) { 139 RegStorage r_tgt = CallHelperSetup(helper_offset); 140 LoadConstant(TargetReg(kArg0, kNotWide), arg0); 141 ClobberCallerSave(); 142 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 143} 144INSTANTIATE(void Mir2Lir::CallRuntimeHelperImm, int arg0, bool safepoint_pc) 145 146template <size_t pointer_size> 147void Mir2Lir::CallRuntimeHelperReg(ThreadOffset<pointer_size> helper_offset, RegStorage arg0, 148 bool safepoint_pc) { 149 RegStorage r_tgt = CallHelperSetup(helper_offset); 150 OpRegCopy(TargetReg(kArg0, arg0.GetWideKind()), arg0); 151 ClobberCallerSave(); 152 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 153} 154INSTANTIATE(void Mir2Lir::CallRuntimeHelperReg, RegStorage arg0, bool safepoint_pc) 155 156template <size_t pointer_size> 157void Mir2Lir::CallRuntimeHelperRegLocation(ThreadOffset<pointer_size> helper_offset, 158 RegLocation arg0, bool safepoint_pc) { 159 RegStorage r_tgt = CallHelperSetup(helper_offset); 160 if (arg0.wide == 0) { 161 LoadValueDirectFixed(arg0, TargetReg(arg0.fp ? kFArg0 : kArg0, arg0)); 162 } else { 163 LoadValueDirectWideFixed(arg0, TargetReg(arg0.fp ? kFArg0 : kArg0, kWide)); 164 } 165 ClobberCallerSave(); 166 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 167} 168INSTANTIATE(void Mir2Lir::CallRuntimeHelperRegLocation, RegLocation arg0, bool safepoint_pc) 169 170template <size_t pointer_size> 171void Mir2Lir::CallRuntimeHelperImmImm(ThreadOffset<pointer_size> helper_offset, int arg0, int arg1, 172 bool safepoint_pc) { 173 RegStorage r_tgt = CallHelperSetup(helper_offset); 174 LoadConstant(TargetReg(kArg0, kNotWide), arg0); 175 LoadConstant(TargetReg(kArg1, kNotWide), arg1); 176 ClobberCallerSave(); 177 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 178} 179INSTANTIATE(void Mir2Lir::CallRuntimeHelperImmImm, int arg0, int arg1, bool safepoint_pc) 180 181template <size_t pointer_size> 182void Mir2Lir::CallRuntimeHelperImmRegLocation(ThreadOffset<pointer_size> helper_offset, int arg0, 183 RegLocation arg1, bool safepoint_pc) { 184 RegStorage r_tgt = CallHelperSetup(helper_offset); 185 DCHECK(!arg1.fp); 186 if (arg1.wide == 0) { 187 LoadValueDirectFixed(arg1, TargetReg(kArg1, arg1)); 188 } else { 189 RegStorage r_tmp = TargetReg(cu_->instruction_set == kMips ? kArg2 : kArg1, kWide); 190 LoadValueDirectWideFixed(arg1, r_tmp); 191 } 192 LoadConstant(TargetReg(kArg0, kNotWide), arg0); 193 ClobberCallerSave(); 194 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 195} 196INSTANTIATE(void Mir2Lir::CallRuntimeHelperImmRegLocation, int arg0, RegLocation arg1, 197 bool safepoint_pc) 198 199template <size_t pointer_size> 200void Mir2Lir::CallRuntimeHelperRegLocationImm(ThreadOffset<pointer_size> helper_offset, 201 RegLocation arg0, int arg1, bool safepoint_pc) { 202 RegStorage r_tgt = CallHelperSetup(helper_offset); 203 DCHECK(!arg0.wide); 204 LoadValueDirectFixed(arg0, TargetReg(kArg0, arg0)); 205 LoadConstant(TargetReg(kArg1, kNotWide), arg1); 206 ClobberCallerSave(); 207 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 208} 209INSTANTIATE(void Mir2Lir::CallRuntimeHelperRegLocationImm, RegLocation arg0, int arg1, 210 bool safepoint_pc) 211 212template <size_t pointer_size> 213void Mir2Lir::CallRuntimeHelperImmReg(ThreadOffset<pointer_size> helper_offset, int arg0, 214 RegStorage arg1, bool safepoint_pc) { 215 RegStorage r_tgt = CallHelperSetup(helper_offset); 216 OpRegCopy(TargetReg(kArg1, arg1.GetWideKind()), arg1); 217 LoadConstant(TargetReg(kArg0, kNotWide), arg0); 218 ClobberCallerSave(); 219 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 220} 221INSTANTIATE(void Mir2Lir::CallRuntimeHelperImmReg, int arg0, RegStorage arg1, bool safepoint_pc) 222 223template <size_t pointer_size> 224void Mir2Lir::CallRuntimeHelperRegImm(ThreadOffset<pointer_size> helper_offset, RegStorage arg0, 225 int arg1, bool safepoint_pc) { 226 RegStorage r_tgt = CallHelperSetup(helper_offset); 227 OpRegCopy(TargetReg(kArg0, arg0.GetWideKind()), arg0); 228 LoadConstant(TargetReg(kArg1, kNotWide), arg1); 229 ClobberCallerSave(); 230 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 231} 232INSTANTIATE(void Mir2Lir::CallRuntimeHelperRegImm, RegStorage arg0, int arg1, bool safepoint_pc) 233 234template <size_t pointer_size> 235void Mir2Lir::CallRuntimeHelperImmMethod(ThreadOffset<pointer_size> helper_offset, int arg0, 236 bool safepoint_pc) { 237 RegStorage r_tgt = CallHelperSetup(helper_offset); 238 LoadCurrMethodDirect(TargetReg(kArg1, kRef)); 239 LoadConstant(TargetReg(kArg0, kNotWide), arg0); 240 ClobberCallerSave(); 241 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 242} 243INSTANTIATE(void Mir2Lir::CallRuntimeHelperImmMethod, int arg0, bool safepoint_pc) 244 245template <size_t pointer_size> 246void Mir2Lir::CallRuntimeHelperRegMethod(ThreadOffset<pointer_size> helper_offset, RegStorage arg0, 247 bool safepoint_pc) { 248 RegStorage r_tgt = CallHelperSetup(helper_offset); 249 DCHECK(!IsSameReg(TargetReg(kArg1, arg0.GetWideKind()), arg0)); 250 RegStorage r_tmp = TargetReg(kArg0, arg0.GetWideKind()); 251 if (r_tmp.NotExactlyEquals(arg0)) { 252 OpRegCopy(r_tmp, arg0); 253 } 254 LoadCurrMethodDirect(TargetReg(kArg1, kRef)); 255 ClobberCallerSave(); 256 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 257} 258INSTANTIATE(void Mir2Lir::CallRuntimeHelperRegMethod, RegStorage arg0, bool safepoint_pc) 259 260template <size_t pointer_size> 261void Mir2Lir::CallRuntimeHelperRegMethodRegLocation(ThreadOffset<pointer_size> helper_offset, 262 RegStorage arg0, RegLocation arg2, 263 bool safepoint_pc) { 264 RegStorage r_tgt = CallHelperSetup(helper_offset); 265 DCHECK(!IsSameReg(TargetReg(kArg1, arg0.GetWideKind()), arg0)); 266 RegStorage r_tmp = TargetReg(kArg0, arg0.GetWideKind()); 267 if (r_tmp.NotExactlyEquals(arg0)) { 268 OpRegCopy(r_tmp, arg0); 269 } 270 LoadCurrMethodDirect(TargetReg(kArg1, kRef)); 271 LoadValueDirectFixed(arg2, TargetReg(kArg2, arg2)); 272 ClobberCallerSave(); 273 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 274} 275INSTANTIATE(void Mir2Lir::CallRuntimeHelperRegMethodRegLocation, RegStorage arg0, RegLocation arg2, 276 bool safepoint_pc) 277 278template <size_t pointer_size> 279void Mir2Lir::CallRuntimeHelperRegLocationRegLocation(ThreadOffset<pointer_size> helper_offset, 280 RegLocation arg0, RegLocation arg1, 281 bool safepoint_pc) { 282 RegStorage r_tgt = CallHelperSetup(helper_offset); 283 if (cu_->instruction_set == kArm64 || cu_->instruction_set == kX86_64) { 284 RegStorage arg0_reg = TargetReg((arg0.fp) ? kFArg0 : kArg0, arg0); 285 286 RegStorage arg1_reg; 287 if (arg1.fp == arg0.fp) { 288 arg1_reg = TargetReg((arg1.fp) ? kFArg1 : kArg1, arg1); 289 } else { 290 arg1_reg = TargetReg((arg1.fp) ? kFArg0 : kArg0, arg1); 291 } 292 293 if (arg0.wide == 0) { 294 LoadValueDirectFixed(arg0, arg0_reg); 295 } else { 296 LoadValueDirectWideFixed(arg0, arg0_reg); 297 } 298 299 if (arg1.wide == 0) { 300 LoadValueDirectFixed(arg1, arg1_reg); 301 } else { 302 LoadValueDirectWideFixed(arg1, arg1_reg); 303 } 304 } else { 305 DCHECK(!cu_->target64); 306 if (arg0.wide == 0) { 307 LoadValueDirectFixed(arg0, TargetReg(arg0.fp ? kFArg0 : kArg0, kNotWide)); 308 if (arg1.wide == 0) { 309 if (cu_->instruction_set == kMips) { 310 LoadValueDirectFixed(arg1, TargetReg(arg1.fp ? kFArg2 : kArg1, kNotWide)); 311 } else { 312 LoadValueDirectFixed(arg1, TargetReg(kArg1, kNotWide)); 313 } 314 } else { 315 if (cu_->instruction_set == kMips) { 316 LoadValueDirectWideFixed(arg1, TargetReg(arg1.fp ? kFArg2 : kArg2, kWide)); 317 } else { 318 LoadValueDirectWideFixed(arg1, TargetReg(kArg1, kWide)); 319 } 320 } 321 } else { 322 LoadValueDirectWideFixed(arg0, TargetReg(arg0.fp ? kFArg0 : kArg0, kWide)); 323 if (arg1.wide == 0) { 324 LoadValueDirectFixed(arg1, TargetReg(arg1.fp ? kFArg2 : kArg2, kNotWide)); 325 } else { 326 LoadValueDirectWideFixed(arg1, TargetReg(arg1.fp ? kFArg2 : kArg2, kWide)); 327 } 328 } 329 } 330 ClobberCallerSave(); 331 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 332} 333INSTANTIATE(void Mir2Lir::CallRuntimeHelperRegLocationRegLocation, RegLocation arg0, 334 RegLocation arg1, bool safepoint_pc) 335 336void Mir2Lir::CopyToArgumentRegs(RegStorage arg0, RegStorage arg1) { 337 WideKind arg0_kind = arg0.GetWideKind(); 338 WideKind arg1_kind = arg1.GetWideKind(); 339 if (IsSameReg(arg1, TargetReg(kArg0, arg1_kind))) { 340 if (IsSameReg(arg0, TargetReg(kArg1, arg0_kind))) { 341 // Swap kArg0 and kArg1 with kArg2 as temp. 342 OpRegCopy(TargetReg(kArg2, arg1_kind), arg1); 343 OpRegCopy(TargetReg(kArg0, arg0_kind), arg0); 344 OpRegCopy(TargetReg(kArg1, arg1_kind), TargetReg(kArg2, arg1_kind)); 345 } else { 346 OpRegCopy(TargetReg(kArg1, arg1_kind), arg1); 347 OpRegCopy(TargetReg(kArg0, arg0_kind), arg0); 348 } 349 } else { 350 OpRegCopy(TargetReg(kArg0, arg0_kind), arg0); 351 OpRegCopy(TargetReg(kArg1, arg1_kind), arg1); 352 } 353} 354 355template <size_t pointer_size> 356void Mir2Lir::CallRuntimeHelperRegReg(ThreadOffset<pointer_size> helper_offset, RegStorage arg0, 357 RegStorage arg1, bool safepoint_pc) { 358 RegStorage r_tgt = CallHelperSetup(helper_offset); 359 CopyToArgumentRegs(arg0, arg1); 360 ClobberCallerSave(); 361 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 362} 363INSTANTIATE(void Mir2Lir::CallRuntimeHelperRegReg, RegStorage arg0, RegStorage arg1, 364 bool safepoint_pc) 365 366template <size_t pointer_size> 367void Mir2Lir::CallRuntimeHelperRegRegImm(ThreadOffset<pointer_size> helper_offset, RegStorage arg0, 368 RegStorage arg1, int arg2, bool safepoint_pc) { 369 RegStorage r_tgt = CallHelperSetup(helper_offset); 370 CopyToArgumentRegs(arg0, arg1); 371 LoadConstant(TargetReg(kArg2, kNotWide), arg2); 372 ClobberCallerSave(); 373 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 374} 375INSTANTIATE(void Mir2Lir::CallRuntimeHelperRegRegImm, RegStorage arg0, RegStorage arg1, int arg2, 376 bool safepoint_pc) 377 378template <size_t pointer_size> 379void Mir2Lir::CallRuntimeHelperImmMethodRegLocation(ThreadOffset<pointer_size> helper_offset, 380 int arg0, RegLocation arg2, bool safepoint_pc) { 381 RegStorage r_tgt = CallHelperSetup(helper_offset); 382 LoadValueDirectFixed(arg2, TargetReg(kArg2, arg2)); 383 LoadCurrMethodDirect(TargetReg(kArg1, kRef)); 384 LoadConstant(TargetReg(kArg0, kNotWide), arg0); 385 ClobberCallerSave(); 386 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 387} 388INSTANTIATE(void Mir2Lir::CallRuntimeHelperImmMethodRegLocation, int arg0, RegLocation arg2, 389 bool safepoint_pc) 390 391template <size_t pointer_size> 392void Mir2Lir::CallRuntimeHelperImmMethodImm(ThreadOffset<pointer_size> helper_offset, int arg0, 393 int arg2, bool safepoint_pc) { 394 RegStorage r_tgt = CallHelperSetup(helper_offset); 395 LoadCurrMethodDirect(TargetReg(kArg1, kRef)); 396 LoadConstant(TargetReg(kArg2, kNotWide), arg2); 397 LoadConstant(TargetReg(kArg0, kNotWide), arg0); 398 ClobberCallerSave(); 399 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 400} 401INSTANTIATE(void Mir2Lir::CallRuntimeHelperImmMethodImm, int arg0, int arg2, bool safepoint_pc) 402 403template <size_t pointer_size> 404void Mir2Lir::CallRuntimeHelperImmRegLocationRegLocation(ThreadOffset<pointer_size> helper_offset, 405 int arg0, RegLocation arg1, 406 RegLocation arg2, bool safepoint_pc) { 407 RegStorage r_tgt = CallHelperSetup(helper_offset); 408 DCHECK_EQ(static_cast<unsigned int>(arg1.wide), 0U); // The static_cast works around an 409 // instantiation bug in GCC. 410 LoadValueDirectFixed(arg1, TargetReg(kArg1, arg1)); 411 if (arg2.wide == 0) { 412 LoadValueDirectFixed(arg2, TargetReg(kArg2, arg2)); 413 } else { 414 LoadValueDirectWideFixed(arg2, TargetReg(kArg2, kWide)); 415 } 416 LoadConstant(TargetReg(kArg0, kNotWide), arg0); 417 ClobberCallerSave(); 418 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 419} 420INSTANTIATE(void Mir2Lir::CallRuntimeHelperImmRegLocationRegLocation, int arg0, RegLocation arg1, 421 RegLocation arg2, bool safepoint_pc) 422 423template <size_t pointer_size> 424void Mir2Lir::CallRuntimeHelperRegLocationRegLocationRegLocation( 425 ThreadOffset<pointer_size> helper_offset, 426 RegLocation arg0, 427 RegLocation arg1, 428 RegLocation arg2, 429 bool safepoint_pc) { 430 RegStorage r_tgt = CallHelperSetup(helper_offset); 431 LoadValueDirectFixed(arg0, TargetReg(kArg0, arg0)); 432 LoadValueDirectFixed(arg1, TargetReg(kArg1, arg1)); 433 LoadValueDirectFixed(arg2, TargetReg(kArg2, arg2)); 434 ClobberCallerSave(); 435 CallHelper<pointer_size>(r_tgt, helper_offset, safepoint_pc); 436} 437INSTANTIATE(void Mir2Lir::CallRuntimeHelperRegLocationRegLocationRegLocation, RegLocation arg0, 438 RegLocation arg1, RegLocation arg2, bool safepoint_pc) 439 440/* 441 * If there are any ins passed in registers that have not been promoted 442 * to a callee-save register, flush them to the frame. Perform initial 443 * assignment of promoted arguments. 444 * 445 * ArgLocs is an array of location records describing the incoming arguments 446 * with one location record per word of argument. 447 */ 448void Mir2Lir::FlushIns(RegLocation* ArgLocs, RegLocation rl_method) { 449 /* 450 * Dummy up a RegLocation for the incoming StackReference<mirror::ArtMethod> 451 * It will attempt to keep kArg0 live (or copy it to home location 452 * if promoted). 453 */ 454 RegLocation rl_src = rl_method; 455 rl_src.location = kLocPhysReg; 456 rl_src.reg = TargetReg(kArg0, kRef); 457 rl_src.home = false; 458 MarkLive(rl_src); 459 StoreValue(rl_method, rl_src); 460 // If Method* has been promoted, explicitly flush 461 if (rl_method.location == kLocPhysReg) { 462 StoreRefDisp(TargetPtrReg(kSp), 0, rl_src.reg, kNotVolatile); 463 } 464 465 if (cu_->num_ins == 0) { 466 return; 467 } 468 469 int start_vreg = cu_->num_dalvik_registers - cu_->num_ins; 470 /* 471 * Copy incoming arguments to their proper home locations. 472 * NOTE: an older version of dx had an issue in which 473 * it would reuse static method argument registers. 474 * This could result in the same Dalvik virtual register 475 * being promoted to both core and fp regs. To account for this, 476 * we only copy to the corresponding promoted physical register 477 * if it matches the type of the SSA name for the incoming 478 * argument. It is also possible that long and double arguments 479 * end up half-promoted. In those cases, we must flush the promoted 480 * half to memory as well. 481 */ 482 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 483 for (int i = 0; i < cu_->num_ins; i++) { 484 PromotionMap* v_map = &promotion_map_[start_vreg + i]; 485 RegStorage reg = GetArgMappingToPhysicalReg(i); 486 487 if (reg.Valid()) { 488 // If arriving in register 489 bool need_flush = true; 490 RegLocation* t_loc = &ArgLocs[i]; 491 if ((v_map->core_location == kLocPhysReg) && !t_loc->fp) { 492 OpRegCopy(RegStorage::Solo32(v_map->core_reg), reg); 493 need_flush = false; 494 } else if ((v_map->fp_location == kLocPhysReg) && t_loc->fp) { 495 OpRegCopy(RegStorage::Solo32(v_map->fp_reg), reg); 496 need_flush = false; 497 } else { 498 need_flush = true; 499 } 500 501 // For wide args, force flush if not fully promoted 502 if (t_loc->wide) { 503 PromotionMap* p_map = v_map + (t_loc->high_word ? -1 : +1); 504 // Is only half promoted? 505 need_flush |= (p_map->core_location != v_map->core_location) || 506 (p_map->fp_location != v_map->fp_location); 507 if ((cu_->instruction_set == kThumb2) && t_loc->fp && !need_flush) { 508 /* 509 * In Arm, a double is represented as a pair of consecutive single float 510 * registers starting at an even number. It's possible that both Dalvik vRegs 511 * representing the incoming double were independently promoted as singles - but 512 * not in a form usable as a double. If so, we need to flush - even though the 513 * incoming arg appears fully in register. At this point in the code, both 514 * halves of the double are promoted. Make sure they are in a usable form. 515 */ 516 int lowreg_index = start_vreg + i + (t_loc->high_word ? -1 : 0); 517 int low_reg = promotion_map_[lowreg_index].fp_reg; 518 int high_reg = promotion_map_[lowreg_index + 1].fp_reg; 519 if (((low_reg & 0x1) != 0) || (high_reg != (low_reg + 1))) { 520 need_flush = true; 521 } 522 } 523 } 524 if (need_flush) { 525 Store32Disp(TargetPtrReg(kSp), SRegOffset(start_vreg + i), reg); 526 } 527 } else { 528 // If arriving in frame & promoted 529 if (v_map->core_location == kLocPhysReg) { 530 Load32Disp(TargetPtrReg(kSp), SRegOffset(start_vreg + i), 531 RegStorage::Solo32(v_map->core_reg)); 532 } 533 if (v_map->fp_location == kLocPhysReg) { 534 Load32Disp(TargetPtrReg(kSp), SRegOffset(start_vreg + i), 535 RegStorage::Solo32(v_map->fp_reg)); 536 } 537 } 538 } 539} 540 541static void CommonCallCodeLoadThisIntoArg1(const CallInfo* info, Mir2Lir* cg) { 542 RegLocation rl_arg = info->args[0]; 543 cg->LoadValueDirectFixed(rl_arg, cg->TargetReg(kArg1, kRef)); 544} 545 546static void CommonCallCodeLoadClassIntoArg0(const CallInfo* info, Mir2Lir* cg) { 547 cg->GenNullCheck(cg->TargetReg(kArg1, kRef), info->opt_flags); 548 // get this->klass_ [use kArg1, set kArg0] 549 cg->LoadRefDisp(cg->TargetReg(kArg1, kRef), mirror::Object::ClassOffset().Int32Value(), 550 cg->TargetReg(kArg0, kRef), 551 kNotVolatile); 552 cg->MarkPossibleNullPointerException(info->opt_flags); 553} 554 555static bool CommonCallCodeLoadCodePointerIntoInvokeTgt(const CallInfo* info, 556 const RegStorage* alt_from, 557 const CompilationUnit* cu, Mir2Lir* cg) { 558 if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) { 559 // Get the compiled code address [use *alt_from or kArg0, set kInvokeTgt] 560 cg->LoadWordDisp(alt_from == nullptr ? cg->TargetReg(kArg0, kRef) : *alt_from, 561 mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset().Int32Value(), 562 cg->TargetPtrReg(kInvokeTgt)); 563 return true; 564 } 565 return false; 566} 567 568/* 569 * Bit of a hack here - in the absence of a real scheduling pass, 570 * emit the next instruction in static & direct invoke sequences. 571 */ 572static int NextSDCallInsn(CompilationUnit* cu, CallInfo* info, 573 int state, const MethodReference& target_method, 574 uint32_t unused, 575 uintptr_t direct_code, uintptr_t direct_method, 576 InvokeType type) { 577 Mir2Lir* cg = static_cast<Mir2Lir*>(cu->cg.get()); 578 if (direct_code != 0 && direct_method != 0) { 579 switch (state) { 580 case 0: // Get the current Method* [sets kArg0] 581 if (direct_code != static_cast<uintptr_t>(-1)) { 582 if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) { 583 cg->LoadConstant(cg->TargetPtrReg(kInvokeTgt), direct_code); 584 } 585 } else if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) { 586 cg->LoadCodeAddress(target_method, type, kInvokeTgt); 587 } 588 if (direct_method != static_cast<uintptr_t>(-1)) { 589 cg->LoadConstant(cg->TargetReg(kArg0, kRef), direct_method); 590 } else { 591 cg->LoadMethodAddress(target_method, type, kArg0); 592 } 593 break; 594 default: 595 return -1; 596 } 597 } else { 598 RegStorage arg0_ref = cg->TargetReg(kArg0, kRef); 599 switch (state) { 600 case 0: // Get the current Method* [sets kArg0] 601 // TUNING: we can save a reg copy if Method* has been promoted. 602 cg->LoadCurrMethodDirect(arg0_ref); 603 break; 604 case 1: // Get method->dex_cache_resolved_methods_ 605 cg->LoadRefDisp(arg0_ref, 606 mirror::ArtMethod::DexCacheResolvedMethodsOffset().Int32Value(), 607 arg0_ref, 608 kNotVolatile); 609 // Set up direct code if known. 610 if (direct_code != 0) { 611 if (direct_code != static_cast<uintptr_t>(-1)) { 612 cg->LoadConstant(cg->TargetPtrReg(kInvokeTgt), direct_code); 613 } else if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) { 614 CHECK_LT(target_method.dex_method_index, target_method.dex_file->NumMethodIds()); 615 cg->LoadCodeAddress(target_method, type, kInvokeTgt); 616 } 617 } 618 break; 619 case 2: // Grab target method* 620 CHECK_EQ(cu->dex_file, target_method.dex_file); 621 cg->LoadRefDisp(arg0_ref, 622 ObjArray::OffsetOfElement(target_method.dex_method_index).Int32Value(), 623 arg0_ref, 624 kNotVolatile); 625 break; 626 case 3: // Grab the code from the method* 627 if (direct_code == 0) { 628 if (CommonCallCodeLoadCodePointerIntoInvokeTgt(info, &arg0_ref, cu, cg)) { 629 break; // kInvokeTgt := arg0_ref->entrypoint 630 } 631 } else if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) { 632 break; 633 } 634 // Intentional fallthrough for x86 635 default: 636 return -1; 637 } 638 } 639 return state + 1; 640} 641 642/* 643 * Bit of a hack here - in the absence of a real scheduling pass, 644 * emit the next instruction in a virtual invoke sequence. 645 * We can use kLr as a temp prior to target address loading 646 * Note also that we'll load the first argument ("this") into 647 * kArg1 here rather than the standard LoadArgRegs. 648 */ 649static int NextVCallInsn(CompilationUnit* cu, CallInfo* info, 650 int state, const MethodReference& target_method, 651 uint32_t method_idx, uintptr_t unused, uintptr_t unused2, 652 InvokeType unused3) { 653 Mir2Lir* cg = static_cast<Mir2Lir*>(cu->cg.get()); 654 /* 655 * This is the fast path in which the target virtual method is 656 * fully resolved at compile time. 657 */ 658 switch (state) { 659 case 0: 660 CommonCallCodeLoadThisIntoArg1(info, cg); // kArg1 := this 661 break; 662 case 1: 663 CommonCallCodeLoadClassIntoArg0(info, cg); // kArg0 := kArg1->class 664 // Includes a null-check. 665 break; 666 case 2: { 667 // Get this->klass_.embedded_vtable[method_idx] [usr kArg0, set kArg0] 668 int32_t offset = mirror::Class::EmbeddedVTableOffset().Uint32Value() + 669 method_idx * sizeof(mirror::Class::VTableEntry); 670 // Load target method from embedded vtable to kArg0 [use kArg0, set kArg0] 671 cg->LoadRefDisp(cg->TargetReg(kArg0, kRef), offset, cg->TargetReg(kArg0, kRef), kNotVolatile); 672 break; 673 } 674 case 3: 675 if (CommonCallCodeLoadCodePointerIntoInvokeTgt(info, nullptr, cu, cg)) { 676 break; // kInvokeTgt := kArg0->entrypoint 677 } 678 // Intentional fallthrough for X86 679 default: 680 return -1; 681 } 682 return state + 1; 683} 684 685/* 686 * Emit the next instruction in an invoke interface sequence. This will do a lookup in the 687 * class's IMT, calling either the actual method or art_quick_imt_conflict_trampoline if 688 * more than one interface method map to the same index. Note also that we'll load the first 689 * argument ("this") into kArg1 here rather than the standard LoadArgRegs. 690 */ 691static int NextInterfaceCallInsn(CompilationUnit* cu, CallInfo* info, int state, 692 const MethodReference& target_method, 693 uint32_t method_idx, uintptr_t unused, 694 uintptr_t direct_method, InvokeType unused2) { 695 Mir2Lir* cg = static_cast<Mir2Lir*>(cu->cg.get()); 696 697 switch (state) { 698 case 0: // Set target method index in case of conflict [set kHiddenArg, kHiddenFpArg (x86)] 699 CHECK_LT(target_method.dex_method_index, target_method.dex_file->NumMethodIds()); 700 cg->LoadConstant(cg->TargetReg(kHiddenArg, kNotWide), target_method.dex_method_index); 701 if (cu->instruction_set == kX86) { 702 cg->OpRegCopy(cg->TargetReg(kHiddenFpArg, kNotWide), cg->TargetReg(kHiddenArg, kNotWide)); 703 } 704 break; 705 case 1: 706 CommonCallCodeLoadThisIntoArg1(info, cg); // kArg1 := this 707 break; 708 case 2: 709 CommonCallCodeLoadClassIntoArg0(info, cg); // kArg0 := kArg1->class 710 // Includes a null-check. 711 break; 712 case 3: { // Get target method [use kInvokeTgt, set kArg0] 713 int32_t offset = mirror::Class::EmbeddedImTableOffset().Uint32Value() + 714 (method_idx % mirror::Class::kImtSize) * sizeof(mirror::Class::ImTableEntry); 715 // Load target method from embedded imtable to kArg0 [use kArg0, set kArg0] 716 cg->LoadRefDisp(cg->TargetReg(kArg0, kRef), offset, cg->TargetReg(kArg0, kRef), kNotVolatile); 717 break; 718 } 719 case 4: 720 if (CommonCallCodeLoadCodePointerIntoInvokeTgt(info, nullptr, cu, cg)) { 721 break; // kInvokeTgt := kArg0->entrypoint 722 } 723 // Intentional fallthrough for X86 724 default: 725 return -1; 726 } 727 return state + 1; 728} 729 730template <size_t pointer_size> 731static int NextInvokeInsnSP(CompilationUnit* cu, CallInfo* info, 732 ThreadOffset<pointer_size> trampoline, int state, 733 const MethodReference& target_method, uint32_t method_idx) { 734 Mir2Lir* cg = static_cast<Mir2Lir*>(cu->cg.get()); 735 /* 736 * This handles the case in which the base method is not fully 737 * resolved at compile time, we bail to a runtime helper. 738 */ 739 if (state == 0) { 740 if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) { 741 // Load trampoline target 742 cg->LoadWordDisp(cg->TargetPtrReg(kSelf), trampoline.Int32Value(), 743 cg->TargetPtrReg(kInvokeTgt)); 744 } 745 // Load kArg0 with method index 746 CHECK_EQ(cu->dex_file, target_method.dex_file); 747 cg->LoadConstant(cg->TargetReg(kArg0, kNotWide), target_method.dex_method_index); 748 return 1; 749 } 750 return -1; 751} 752 753static int NextStaticCallInsnSP(CompilationUnit* cu, CallInfo* info, 754 int state, 755 const MethodReference& target_method, 756 uint32_t unused, uintptr_t unused2, 757 uintptr_t unused3, InvokeType unused4) { 758 if (cu->target64) { 759 ThreadOffset<8> trampoline = QUICK_ENTRYPOINT_OFFSET(8, pInvokeStaticTrampolineWithAccessCheck); 760 return NextInvokeInsnSP<8>(cu, info, trampoline, state, target_method, 0); 761 } else { 762 ThreadOffset<4> trampoline = QUICK_ENTRYPOINT_OFFSET(4, pInvokeStaticTrampolineWithAccessCheck); 763 return NextInvokeInsnSP<4>(cu, info, trampoline, state, target_method, 0); 764 } 765} 766 767static int NextDirectCallInsnSP(CompilationUnit* cu, CallInfo* info, int state, 768 const MethodReference& target_method, 769 uint32_t unused, uintptr_t unused2, 770 uintptr_t unused3, InvokeType unused4) { 771 if (cu->target64) { 772 ThreadOffset<8> trampoline = QUICK_ENTRYPOINT_OFFSET(8, pInvokeDirectTrampolineWithAccessCheck); 773 return NextInvokeInsnSP<8>(cu, info, trampoline, state, target_method, 0); 774 } else { 775 ThreadOffset<4> trampoline = QUICK_ENTRYPOINT_OFFSET(4, pInvokeDirectTrampolineWithAccessCheck); 776 return NextInvokeInsnSP<4>(cu, info, trampoline, state, target_method, 0); 777 } 778} 779 780static int NextSuperCallInsnSP(CompilationUnit* cu, CallInfo* info, int state, 781 const MethodReference& target_method, 782 uint32_t unused, uintptr_t unused2, 783 uintptr_t unused3, InvokeType unused4) { 784 if (cu->target64) { 785 ThreadOffset<8> trampoline = QUICK_ENTRYPOINT_OFFSET(8, pInvokeSuperTrampolineWithAccessCheck); 786 return NextInvokeInsnSP<8>(cu, info, trampoline, state, target_method, 0); 787 } else { 788 ThreadOffset<4> trampoline = QUICK_ENTRYPOINT_OFFSET(4, pInvokeSuperTrampolineWithAccessCheck); 789 return NextInvokeInsnSP<4>(cu, info, trampoline, state, target_method, 0); 790 } 791} 792 793static int NextVCallInsnSP(CompilationUnit* cu, CallInfo* info, int state, 794 const MethodReference& target_method, 795 uint32_t unused, uintptr_t unused2, 796 uintptr_t unused3, InvokeType unused4) { 797 if (cu->target64) { 798 ThreadOffset<8> trampoline = QUICK_ENTRYPOINT_OFFSET(8, 799 pInvokeVirtualTrampolineWithAccessCheck); 800 return NextInvokeInsnSP<8>(cu, info, trampoline, state, target_method, 0); 801 } else { 802 ThreadOffset<4> trampoline = QUICK_ENTRYPOINT_OFFSET(4, 803 pInvokeVirtualTrampolineWithAccessCheck); 804 return NextInvokeInsnSP<4>(cu, info, trampoline, state, target_method, 0); 805 } 806} 807 808static int NextInterfaceCallInsnWithAccessCheck(CompilationUnit* cu, 809 CallInfo* info, int state, 810 const MethodReference& target_method, 811 uint32_t unused, uintptr_t unused2, 812 uintptr_t unused3, InvokeType unused4) { 813 if (cu->target64) { 814 ThreadOffset<8> trampoline = QUICK_ENTRYPOINT_OFFSET(8, 815 pInvokeInterfaceTrampolineWithAccessCheck); 816 return NextInvokeInsnSP<8>(cu, info, trampoline, state, target_method, 0); 817 } else { 818 ThreadOffset<4> trampoline = QUICK_ENTRYPOINT_OFFSET(4, 819 pInvokeInterfaceTrampolineWithAccessCheck); 820 return NextInvokeInsnSP<4>(cu, info, trampoline, state, target_method, 0); 821 } 822} 823 824int Mir2Lir::LoadArgRegs(CallInfo* info, int call_state, 825 NextCallInsn next_call_insn, 826 const MethodReference& target_method, 827 uint32_t vtable_idx, uintptr_t direct_code, 828 uintptr_t direct_method, InvokeType type, bool skip_this) { 829 int last_arg_reg = 3 - 1; 830 int arg_regs[3] = {TargetReg(kArg1, kNotWide).GetReg(), TargetReg(kArg2, kNotWide).GetReg(), 831 TargetReg(kArg3, kNotWide).GetReg()}; 832 833 int next_reg = 0; 834 int next_arg = 0; 835 if (skip_this) { 836 next_reg++; 837 next_arg++; 838 } 839 for (; (next_reg <= last_arg_reg) && (next_arg < info->num_arg_words); next_reg++) { 840 RegLocation rl_arg = info->args[next_arg++]; 841 rl_arg = UpdateRawLoc(rl_arg); 842 if (rl_arg.wide && (next_reg <= last_arg_reg - 1)) { 843 RegStorage r_tmp(RegStorage::k64BitPair, arg_regs[next_reg], arg_regs[next_reg + 1]); 844 LoadValueDirectWideFixed(rl_arg, r_tmp); 845 next_reg++; 846 next_arg++; 847 } else { 848 if (rl_arg.wide) { 849 rl_arg = NarrowRegLoc(rl_arg); 850 rl_arg.is_const = false; 851 } 852 LoadValueDirectFixed(rl_arg, RegStorage::Solo32(arg_regs[next_reg])); 853 } 854 call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx, 855 direct_code, direct_method, type); 856 } 857 return call_state; 858} 859 860/* 861 * Load up to 5 arguments, the first three of which will be in 862 * kArg1 .. kArg3. On entry kArg0 contains the current method pointer, 863 * and as part of the load sequence, it must be replaced with 864 * the target method pointer. Note, this may also be called 865 * for "range" variants if the number of arguments is 5 or fewer. 866 */ 867int Mir2Lir::GenDalvikArgsNoRange(CallInfo* info, 868 int call_state, LIR** pcrLabel, NextCallInsn next_call_insn, 869 const MethodReference& target_method, 870 uint32_t vtable_idx, uintptr_t direct_code, 871 uintptr_t direct_method, InvokeType type, bool skip_this) { 872 RegLocation rl_arg; 873 874 /* If no arguments, just return */ 875 if (info->num_arg_words == 0) 876 return call_state; 877 878 call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx, 879 direct_code, direct_method, type); 880 881 DCHECK_LE(info->num_arg_words, 5); 882 if (info->num_arg_words > 3) { 883 int32_t next_use = 3; 884 // Detect special case of wide arg spanning arg3/arg4 885 RegLocation rl_use0 = info->args[0]; 886 RegLocation rl_use1 = info->args[1]; 887 RegLocation rl_use2 = info->args[2]; 888 if (((!rl_use0.wide && !rl_use1.wide) || rl_use0.wide) && rl_use2.wide) { 889 RegStorage reg; 890 // Wide spans, we need the 2nd half of uses[2]. 891 rl_arg = UpdateLocWide(rl_use2); 892 if (rl_arg.location == kLocPhysReg) { 893 if (rl_arg.reg.IsPair()) { 894 reg = rl_arg.reg.GetHigh(); 895 } else { 896 RegisterInfo* info = GetRegInfo(rl_arg.reg); 897 info = info->FindMatchingView(RegisterInfo::kHighSingleStorageMask); 898 if (info == nullptr) { 899 // NOTE: For hard float convention we won't split arguments across reg/mem. 900 UNIMPLEMENTED(FATAL) << "Needs hard float api."; 901 } 902 reg = info->GetReg(); 903 } 904 } else { 905 // kArg2 & rArg3 can safely be used here 906 reg = TargetReg(kArg3, kNotWide); 907 { 908 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 909 Load32Disp(TargetPtrReg(kSp), SRegOffset(rl_arg.s_reg_low) + 4, reg); 910 } 911 call_state = next_call_insn(cu_, info, call_state, target_method, 912 vtable_idx, direct_code, direct_method, type); 913 } 914 { 915 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 916 Store32Disp(TargetPtrReg(kSp), (next_use + 1) * 4, reg); 917 } 918 call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx, 919 direct_code, direct_method, type); 920 next_use++; 921 } 922 // Loop through the rest 923 while (next_use < info->num_arg_words) { 924 RegStorage arg_reg; 925 rl_arg = info->args[next_use]; 926 rl_arg = UpdateRawLoc(rl_arg); 927 if (rl_arg.location == kLocPhysReg) { 928 arg_reg = rl_arg.reg; 929 } else { 930 arg_reg = TargetReg(kArg2, rl_arg.wide ? kWide : kNotWide); 931 if (rl_arg.wide) { 932 LoadValueDirectWideFixed(rl_arg, arg_reg); 933 } else { 934 LoadValueDirectFixed(rl_arg, arg_reg); 935 } 936 call_state = next_call_insn(cu_, info, call_state, target_method, 937 vtable_idx, direct_code, direct_method, type); 938 } 939 int outs_offset = (next_use + 1) * 4; 940 { 941 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 942 if (rl_arg.wide) { 943 StoreBaseDisp(TargetPtrReg(kSp), outs_offset, arg_reg, k64, kNotVolatile); 944 next_use += 2; 945 } else { 946 Store32Disp(TargetPtrReg(kSp), outs_offset, arg_reg); 947 next_use++; 948 } 949 } 950 call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx, 951 direct_code, direct_method, type); 952 } 953 } 954 955 call_state = LoadArgRegs(info, call_state, next_call_insn, 956 target_method, vtable_idx, direct_code, direct_method, 957 type, skip_this); 958 959 if (pcrLabel) { 960 if (cu_->compiler_driver->GetCompilerOptions().GetExplicitNullChecks()) { 961 *pcrLabel = GenExplicitNullCheck(TargetReg(kArg1, kRef), info->opt_flags); 962 } else { 963 *pcrLabel = nullptr; 964 // In lieu of generating a check for kArg1 being null, we need to 965 // perform a load when doing implicit checks. 966 RegStorage tmp = AllocTemp(); 967 Load32Disp(TargetReg(kArg1, kRef), 0, tmp); 968 MarkPossibleNullPointerException(info->opt_flags); 969 FreeTemp(tmp); 970 } 971 } 972 return call_state; 973} 974 975/* 976 * May have 0+ arguments (also used for jumbo). Note that 977 * source virtual registers may be in physical registers, so may 978 * need to be flushed to home location before copying. This 979 * applies to arg3 and above (see below). 980 * 981 * Two general strategies: 982 * If < 20 arguments 983 * Pass args 3-18 using vldm/vstm block copy 984 * Pass arg0, arg1 & arg2 in kArg1-kArg3 985 * If 20+ arguments 986 * Pass args arg19+ using memcpy block copy 987 * Pass arg0, arg1 & arg2 in kArg1-kArg3 988 * 989 */ 990int Mir2Lir::GenDalvikArgsRange(CallInfo* info, int call_state, 991 LIR** pcrLabel, NextCallInsn next_call_insn, 992 const MethodReference& target_method, 993 uint32_t vtable_idx, uintptr_t direct_code, uintptr_t direct_method, 994 InvokeType type, bool skip_this) { 995 // If we can treat it as non-range (Jumbo ops will use range form) 996 if (info->num_arg_words <= 5) 997 return GenDalvikArgsNoRange(info, call_state, pcrLabel, 998 next_call_insn, target_method, vtable_idx, 999 direct_code, direct_method, type, skip_this); 1000 /* 1001 * First load the non-register arguments. Both forms expect all 1002 * of the source arguments to be in their home frame location, so 1003 * scan the s_reg names and flush any that have been promoted to 1004 * frame backing storage. 1005 */ 1006 // Scan the rest of the args - if in phys_reg flush to memory 1007 for (int next_arg = 0; next_arg < info->num_arg_words;) { 1008 RegLocation loc = info->args[next_arg]; 1009 if (loc.wide) { 1010 loc = UpdateLocWide(loc); 1011 if ((next_arg >= 2) && (loc.location == kLocPhysReg)) { 1012 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 1013 StoreBaseDisp(TargetPtrReg(kSp), SRegOffset(loc.s_reg_low), loc.reg, k64, kNotVolatile); 1014 } 1015 next_arg += 2; 1016 } else { 1017 loc = UpdateLoc(loc); 1018 if ((next_arg >= 3) && (loc.location == kLocPhysReg)) { 1019 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 1020 Store32Disp(TargetPtrReg(kSp), SRegOffset(loc.s_reg_low), loc.reg); 1021 } 1022 next_arg++; 1023 } 1024 } 1025 1026 // Logic below assumes that Method pointer is at offset zero from SP. 1027 DCHECK_EQ(VRegOffset(static_cast<int>(kVRegMethodPtrBaseReg)), 0); 1028 1029 // The first 3 arguments are passed via registers. 1030 // TODO: For 64-bit, instead of hardcoding 4 for Method* size, we should either 1031 // get size of uintptr_t or size of object reference according to model being used. 1032 int outs_offset = 4 /* Method* */ + (3 * sizeof(uint32_t)); 1033 int start_offset = SRegOffset(info->args[3].s_reg_low); 1034 int regs_left_to_pass_via_stack = info->num_arg_words - 3; 1035 DCHECK_GT(regs_left_to_pass_via_stack, 0); 1036 1037 if (cu_->instruction_set == kThumb2 && regs_left_to_pass_via_stack <= 16) { 1038 // Use vldm/vstm pair using kArg3 as a temp 1039 call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx, 1040 direct_code, direct_method, type); 1041 OpRegRegImm(kOpAdd, TargetReg(kArg3, kRef), TargetPtrReg(kSp), start_offset); 1042 LIR* ld = nullptr; 1043 { 1044 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 1045 ld = OpVldm(TargetReg(kArg3, kRef), regs_left_to_pass_via_stack); 1046 } 1047 // TUNING: loosen barrier 1048 ld->u.m.def_mask = &kEncodeAll; 1049 call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx, 1050 direct_code, direct_method, type); 1051 OpRegRegImm(kOpAdd, TargetReg(kArg3, kRef), TargetPtrReg(kSp), 4 /* Method* */ + (3 * 4)); 1052 call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx, 1053 direct_code, direct_method, type); 1054 LIR* st = nullptr; 1055 { 1056 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 1057 st = OpVstm(TargetReg(kArg3, kRef), regs_left_to_pass_via_stack); 1058 } 1059 st->u.m.def_mask = &kEncodeAll; 1060 call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx, 1061 direct_code, direct_method, type); 1062 } else if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) { 1063 int current_src_offset = start_offset; 1064 int current_dest_offset = outs_offset; 1065 1066 // Only davik regs are accessed in this loop; no next_call_insn() calls. 1067 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg); 1068 while (regs_left_to_pass_via_stack > 0) { 1069 // This is based on the knowledge that the stack itself is 16-byte aligned. 1070 bool src_is_16b_aligned = (current_src_offset & 0xF) == 0; 1071 bool dest_is_16b_aligned = (current_dest_offset & 0xF) == 0; 1072 size_t bytes_to_move; 1073 1074 /* 1075 * The amount to move defaults to 32-bit. If there are 4 registers left to move, then do a 1076 * a 128-bit move because we won't get the chance to try to aligned. If there are more than 1077 * 4 registers left to move, consider doing a 128-bit only if either src or dest are aligned. 1078 * We do this because we could potentially do a smaller move to align. 1079 */ 1080 if (regs_left_to_pass_via_stack == 4 || 1081 (regs_left_to_pass_via_stack > 4 && (src_is_16b_aligned || dest_is_16b_aligned))) { 1082 // Moving 128-bits via xmm register. 1083 bytes_to_move = sizeof(uint32_t) * 4; 1084 1085 // Allocate a free xmm temp. Since we are working through the calling sequence, 1086 // we expect to have an xmm temporary available. AllocTempDouble will abort if 1087 // there are no free registers. 1088 RegStorage temp = AllocTempDouble(); 1089 1090 LIR* ld1 = nullptr; 1091 LIR* ld2 = nullptr; 1092 LIR* st1 = nullptr; 1093 LIR* st2 = nullptr; 1094 1095 /* 1096 * The logic is similar for both loads and stores. If we have 16-byte alignment, 1097 * do an aligned move. If we have 8-byte alignment, then do the move in two 1098 * parts. This approach prevents possible cache line splits. Finally, fall back 1099 * to doing an unaligned move. In most cases we likely won't split the cache 1100 * line but we cannot prove it and thus take a conservative approach. 1101 */ 1102 bool src_is_8b_aligned = (current_src_offset & 0x7) == 0; 1103 bool dest_is_8b_aligned = (current_dest_offset & 0x7) == 0; 1104 1105 if (src_is_16b_aligned) { 1106 ld1 = OpMovRegMem(temp, TargetPtrReg(kSp), current_src_offset, kMovA128FP); 1107 } else if (src_is_8b_aligned) { 1108 ld1 = OpMovRegMem(temp, TargetPtrReg(kSp), current_src_offset, kMovLo128FP); 1109 ld2 = OpMovRegMem(temp, TargetPtrReg(kSp), current_src_offset + (bytes_to_move >> 1), 1110 kMovHi128FP); 1111 } else { 1112 ld1 = OpMovRegMem(temp, TargetPtrReg(kSp), current_src_offset, kMovU128FP); 1113 } 1114 1115 if (dest_is_16b_aligned) { 1116 st1 = OpMovMemReg(TargetPtrReg(kSp), current_dest_offset, temp, kMovA128FP); 1117 } else if (dest_is_8b_aligned) { 1118 st1 = OpMovMemReg(TargetPtrReg(kSp), current_dest_offset, temp, kMovLo128FP); 1119 st2 = OpMovMemReg(TargetPtrReg(kSp), current_dest_offset + (bytes_to_move >> 1), 1120 temp, kMovHi128FP); 1121 } else { 1122 st1 = OpMovMemReg(TargetPtrReg(kSp), current_dest_offset, temp, kMovU128FP); 1123 } 1124 1125 // TODO If we could keep track of aliasing information for memory accesses that are wider 1126 // than 64-bit, we wouldn't need to set up a barrier. 1127 if (ld1 != nullptr) { 1128 if (ld2 != nullptr) { 1129 // For 64-bit load we can actually set up the aliasing information. 1130 AnnotateDalvikRegAccess(ld1, current_src_offset >> 2, true, true); 1131 AnnotateDalvikRegAccess(ld2, (current_src_offset + (bytes_to_move >> 1)) >> 2, true, 1132 true); 1133 } else { 1134 // Set barrier for 128-bit load. 1135 ld1->u.m.def_mask = &kEncodeAll; 1136 } 1137 } 1138 if (st1 != nullptr) { 1139 if (st2 != nullptr) { 1140 // For 64-bit store we can actually set up the aliasing information. 1141 AnnotateDalvikRegAccess(st1, current_dest_offset >> 2, false, true); 1142 AnnotateDalvikRegAccess(st2, (current_dest_offset + (bytes_to_move >> 1)) >> 2, false, 1143 true); 1144 } else { 1145 // Set barrier for 128-bit store. 1146 st1->u.m.def_mask = &kEncodeAll; 1147 } 1148 } 1149 1150 // Free the temporary used for the data movement. 1151 FreeTemp(temp); 1152 } else { 1153 // Moving 32-bits via general purpose register. 1154 bytes_to_move = sizeof(uint32_t); 1155 1156 // Instead of allocating a new temp, simply reuse one of the registers being used 1157 // for argument passing. 1158 RegStorage temp = TargetReg(kArg3, kNotWide); 1159 1160 // Now load the argument VR and store to the outs. 1161 Load32Disp(TargetPtrReg(kSp), current_src_offset, temp); 1162 Store32Disp(TargetPtrReg(kSp), current_dest_offset, temp); 1163 } 1164 1165 current_src_offset += bytes_to_move; 1166 current_dest_offset += bytes_to_move; 1167 regs_left_to_pass_via_stack -= (bytes_to_move >> 2); 1168 } 1169 } else { 1170 // Generate memcpy 1171 OpRegRegImm(kOpAdd, TargetReg(kArg0, kRef), TargetPtrReg(kSp), outs_offset); 1172 OpRegRegImm(kOpAdd, TargetReg(kArg1, kRef), TargetPtrReg(kSp), start_offset); 1173 if (cu_->target64) { 1174 CallRuntimeHelperRegRegImm(QUICK_ENTRYPOINT_OFFSET(8, pMemcpy), TargetReg(kArg0, kRef), 1175 TargetReg(kArg1, kRef), (info->num_arg_words - 3) * 4, false); 1176 } else { 1177 CallRuntimeHelperRegRegImm(QUICK_ENTRYPOINT_OFFSET(4, pMemcpy), TargetReg(kArg0, kRef), 1178 TargetReg(kArg1, kRef), (info->num_arg_words - 3) * 4, false); 1179 } 1180 } 1181 1182 call_state = LoadArgRegs(info, call_state, next_call_insn, 1183 target_method, vtable_idx, direct_code, direct_method, 1184 type, skip_this); 1185 1186 call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx, 1187 direct_code, direct_method, type); 1188 if (pcrLabel) { 1189 if (cu_->compiler_driver->GetCompilerOptions().GetExplicitNullChecks()) { 1190 *pcrLabel = GenExplicitNullCheck(TargetReg(kArg1, kRef), info->opt_flags); 1191 } else { 1192 *pcrLabel = nullptr; 1193 // In lieu of generating a check for kArg1 being null, we need to 1194 // perform a load when doing implicit checks. 1195 RegStorage tmp = AllocTemp(); 1196 Load32Disp(TargetReg(kArg1, kRef), 0, tmp); 1197 MarkPossibleNullPointerException(info->opt_flags); 1198 FreeTemp(tmp); 1199 } 1200 } 1201 return call_state; 1202} 1203 1204RegLocation Mir2Lir::InlineTarget(CallInfo* info) { 1205 RegLocation res; 1206 if (info->result.location == kLocInvalid) { 1207 res = GetReturn(LocToRegClass(info->result)); 1208 } else { 1209 res = info->result; 1210 } 1211 return res; 1212} 1213 1214RegLocation Mir2Lir::InlineTargetWide(CallInfo* info) { 1215 RegLocation res; 1216 if (info->result.location == kLocInvalid) { 1217 res = GetReturnWide(kCoreReg); 1218 } else { 1219 res = info->result; 1220 } 1221 return res; 1222} 1223 1224bool Mir2Lir::GenInlinedGet(CallInfo* info) { 1225 if (cu_->instruction_set == kMips) { 1226 // TODO - add Mips implementation 1227 return false; 1228 } 1229 1230 // the refrence class is stored in the image dex file which might not be the same as the cu's 1231 // dex file. Query the reference class for the image dex file then reset to starting dex file 1232 // in after loading class type. 1233 uint16_t type_idx = 0; 1234 const DexFile* ref_dex_file = nullptr; 1235 { 1236 ScopedObjectAccess soa(Thread::Current()); 1237 type_idx = mirror::Reference::GetJavaLangRefReference()->GetDexTypeIndex(); 1238 ref_dex_file = mirror::Reference::GetJavaLangRefReference()->GetDexCache()->GetDexFile(); 1239 } 1240 CHECK(LIKELY(ref_dex_file != nullptr)); 1241 1242 // address is either static within the image file, or needs to be patched up after compilation. 1243 bool unused_type_initialized; 1244 bool use_direct_type_ptr; 1245 uintptr_t direct_type_ptr; 1246 bool is_finalizable; 1247 const DexFile* old_dex = cu_->dex_file; 1248 cu_->dex_file = ref_dex_file; 1249 RegStorage reg_class = TargetPtrReg(kArg1); 1250 if (!cu_->compiler_driver->CanEmbedTypeInCode(*ref_dex_file, type_idx, &unused_type_initialized, 1251 &use_direct_type_ptr, &direct_type_ptr, 1252 &is_finalizable) || is_finalizable) { 1253 cu_->dex_file = old_dex; 1254 // address is not known and post-compile patch is not possible, cannot insert intrinsic. 1255 return false; 1256 } 1257 if (use_direct_type_ptr) { 1258 LoadConstant(reg_class, direct_type_ptr); 1259 } else { 1260 LoadClassType(type_idx, kArg1); 1261 } 1262 cu_->dex_file = old_dex; 1263 1264 // get the offset for flags in reference class. 1265 uint32_t slow_path_flag_offset = 0; 1266 uint32_t disable_flag_offset = 0; 1267 { 1268 ScopedObjectAccess soa(Thread::Current()); 1269 mirror::Class* reference_class = mirror::Reference::GetJavaLangRefReference(); 1270 slow_path_flag_offset = reference_class->GetSlowPathFlagOffset().Uint32Value(); 1271 disable_flag_offset = reference_class->GetDisableIntrinsicFlagOffset().Uint32Value(); 1272 } 1273 CHECK(slow_path_flag_offset && disable_flag_offset && 1274 (slow_path_flag_offset != disable_flag_offset)); 1275 1276 // intrinsic logic start. 1277 RegLocation rl_obj = info->args[0]; 1278 rl_obj = LoadValue(rl_obj); 1279 1280 RegStorage reg_slow_path = AllocTemp(); 1281 RegStorage reg_disabled = AllocTemp(); 1282 Load32Disp(reg_class, slow_path_flag_offset, reg_slow_path); 1283 Load32Disp(reg_class, disable_flag_offset, reg_disabled); 1284 OpRegRegReg(kOpOr, reg_slow_path, reg_slow_path, reg_disabled); 1285 FreeTemp(reg_disabled); 1286 1287 // if slow path, jump to JNI path target 1288 LIR* slow_path_branch = OpCmpImmBranch(kCondNe, reg_slow_path, 0, nullptr); 1289 FreeTemp(reg_slow_path); 1290 1291 // slow path not enabled, simply load the referent of the reference object 1292 RegLocation rl_dest = InlineTarget(info); 1293 RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true); 1294 GenNullCheck(rl_obj.reg, info->opt_flags); 1295 LoadRefDisp(rl_obj.reg, mirror::Reference::ReferentOffset().Int32Value(), rl_result.reg, 1296 kNotVolatile); 1297 MarkPossibleNullPointerException(info->opt_flags); 1298 StoreValue(rl_dest, rl_result); 1299 1300 LIR* intrinsic_finish = NewLIR0(kPseudoTargetLabel); 1301 AddIntrinsicSlowPath(info, slow_path_branch, intrinsic_finish); 1302 1303 return true; 1304} 1305 1306bool Mir2Lir::GenInlinedCharAt(CallInfo* info) { 1307 if (cu_->instruction_set == kMips) { 1308 // TODO - add Mips implementation 1309 return false; 1310 } 1311 // Location of reference to data array 1312 int value_offset = mirror::String::ValueOffset().Int32Value(); 1313 // Location of count 1314 int count_offset = mirror::String::CountOffset().Int32Value(); 1315 // Starting offset within data array 1316 int offset_offset = mirror::String::OffsetOffset().Int32Value(); 1317 // Start of char data with array_ 1318 int data_offset = mirror::Array::DataOffset(sizeof(uint16_t)).Int32Value(); 1319 1320 RegLocation rl_obj = info->args[0]; 1321 RegLocation rl_idx = info->args[1]; 1322 rl_obj = LoadValue(rl_obj, kRefReg); 1323 // X86 wants to avoid putting a constant index into a register. 1324 if (!((cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64)&& rl_idx.is_const)) { 1325 rl_idx = LoadValue(rl_idx, kCoreReg); 1326 } 1327 RegStorage reg_max; 1328 GenNullCheck(rl_obj.reg, info->opt_flags); 1329 bool range_check = (!(info->opt_flags & MIR_IGNORE_RANGE_CHECK)); 1330 LIR* range_check_branch = nullptr; 1331 RegStorage reg_off; 1332 RegStorage reg_ptr; 1333 if (cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64) { 1334 reg_off = AllocTemp(); 1335 reg_ptr = AllocTempRef(); 1336 if (range_check) { 1337 reg_max = AllocTemp(); 1338 Load32Disp(rl_obj.reg, count_offset, reg_max); 1339 MarkPossibleNullPointerException(info->opt_flags); 1340 } 1341 Load32Disp(rl_obj.reg, offset_offset, reg_off); 1342 MarkPossibleNullPointerException(info->opt_flags); 1343 LoadRefDisp(rl_obj.reg, value_offset, reg_ptr, kNotVolatile); 1344 if (range_check) { 1345 // Set up a slow path to allow retry in case of bounds violation */ 1346 OpRegReg(kOpCmp, rl_idx.reg, reg_max); 1347 FreeTemp(reg_max); 1348 range_check_branch = OpCondBranch(kCondUge, nullptr); 1349 } 1350 OpRegImm(kOpAdd, reg_ptr, data_offset); 1351 } else { 1352 if (range_check) { 1353 // On x86, we can compare to memory directly 1354 // Set up a launch pad to allow retry in case of bounds violation */ 1355 if (rl_idx.is_const) { 1356 range_check_branch = OpCmpMemImmBranch( 1357 kCondUlt, RegStorage::InvalidReg(), rl_obj.reg, count_offset, 1358 mir_graph_->ConstantValue(rl_idx.orig_sreg), nullptr); 1359 } else { 1360 OpRegMem(kOpCmp, rl_idx.reg, rl_obj.reg, count_offset); 1361 range_check_branch = OpCondBranch(kCondUge, nullptr); 1362 } 1363 } 1364 reg_off = AllocTemp(); 1365 reg_ptr = AllocTempRef(); 1366 Load32Disp(rl_obj.reg, offset_offset, reg_off); 1367 LoadRefDisp(rl_obj.reg, value_offset, reg_ptr, kNotVolatile); 1368 } 1369 if (rl_idx.is_const) { 1370 OpRegImm(kOpAdd, reg_off, mir_graph_->ConstantValue(rl_idx.orig_sreg)); 1371 } else { 1372 OpRegReg(kOpAdd, reg_off, rl_idx.reg); 1373 } 1374 FreeTemp(rl_obj.reg); 1375 if (rl_idx.location == kLocPhysReg) { 1376 FreeTemp(rl_idx.reg); 1377 } 1378 RegLocation rl_dest = InlineTarget(info); 1379 RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true); 1380 if (cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64) { 1381 LoadBaseIndexed(reg_ptr, reg_off, rl_result.reg, 1, kUnsignedHalf); 1382 } else { 1383 LoadBaseIndexedDisp(reg_ptr, reg_off, 1, data_offset, rl_result.reg, kUnsignedHalf); 1384 } 1385 FreeTemp(reg_off); 1386 FreeTemp(reg_ptr); 1387 StoreValue(rl_dest, rl_result); 1388 if (range_check) { 1389 DCHECK(range_check_branch != nullptr); 1390 info->opt_flags |= MIR_IGNORE_NULL_CHECK; // Record that we've already null checked. 1391 AddIntrinsicSlowPath(info, range_check_branch); 1392 } 1393 return true; 1394} 1395 1396// Generates an inlined String.is_empty or String.length. 1397bool Mir2Lir::GenInlinedStringIsEmptyOrLength(CallInfo* info, bool is_empty) { 1398 if (cu_->instruction_set == kMips) { 1399 // TODO - add Mips implementation 1400 return false; 1401 } 1402 // dst = src.length(); 1403 RegLocation rl_obj = info->args[0]; 1404 rl_obj = LoadValue(rl_obj, kRefReg); 1405 RegLocation rl_dest = InlineTarget(info); 1406 RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true); 1407 GenNullCheck(rl_obj.reg, info->opt_flags); 1408 Load32Disp(rl_obj.reg, mirror::String::CountOffset().Int32Value(), rl_result.reg); 1409 MarkPossibleNullPointerException(info->opt_flags); 1410 if (is_empty) { 1411 // dst = (dst == 0); 1412 if (cu_->instruction_set == kThumb2) { 1413 RegStorage t_reg = AllocTemp(); 1414 OpRegReg(kOpNeg, t_reg, rl_result.reg); 1415 OpRegRegReg(kOpAdc, rl_result.reg, rl_result.reg, t_reg); 1416 } else if (cu_->instruction_set == kArm64) { 1417 OpRegImm(kOpSub, rl_result.reg, 1); 1418 OpRegRegImm(kOpLsr, rl_result.reg, rl_result.reg, 31); 1419 } else { 1420 DCHECK(cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64); 1421 OpRegImm(kOpSub, rl_result.reg, 1); 1422 OpRegImm(kOpLsr, rl_result.reg, 31); 1423 } 1424 } 1425 StoreValue(rl_dest, rl_result); 1426 return true; 1427} 1428 1429bool Mir2Lir::GenInlinedReverseBytes(CallInfo* info, OpSize size) { 1430 if (cu_->instruction_set == kMips) { 1431 // TODO - add Mips implementation. 1432 return false; 1433 } 1434 RegLocation rl_src_i = info->args[0]; 1435 RegLocation rl_i = (size == k64) ? LoadValueWide(rl_src_i, kCoreReg) : LoadValue(rl_src_i, kCoreReg); 1436 RegLocation rl_dest = (size == k64) ? InlineTargetWide(info) : InlineTarget(info); // result reg 1437 RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true); 1438 if (size == k64) { 1439 if (cu_->instruction_set == kArm64 || cu_->instruction_set == kX86_64) { 1440 OpRegReg(kOpRev, rl_result.reg, rl_i.reg); 1441 StoreValueWide(rl_dest, rl_result); 1442 return true; 1443 } 1444 RegStorage r_i_low = rl_i.reg.GetLow(); 1445 if (rl_i.reg.GetLowReg() == rl_result.reg.GetLowReg()) { 1446 // First REV shall clobber rl_result.reg.GetReg(), save the value in a temp for the second REV. 1447 r_i_low = AllocTemp(); 1448 OpRegCopy(r_i_low, rl_i.reg); 1449 } 1450 OpRegReg(kOpRev, rl_result.reg.GetLow(), rl_i.reg.GetHigh()); 1451 OpRegReg(kOpRev, rl_result.reg.GetHigh(), r_i_low); 1452 if (rl_i.reg.GetLowReg() == rl_result.reg.GetLowReg()) { 1453 FreeTemp(r_i_low); 1454 } 1455 StoreValueWide(rl_dest, rl_result); 1456 } else { 1457 DCHECK(size == k32 || size == kSignedHalf); 1458 OpKind op = (size == k32) ? kOpRev : kOpRevsh; 1459 OpRegReg(op, rl_result.reg, rl_i.reg); 1460 StoreValue(rl_dest, rl_result); 1461 } 1462 return true; 1463} 1464 1465bool Mir2Lir::GenInlinedAbsInt(CallInfo* info) { 1466 if (cu_->instruction_set == kMips) { 1467 // TODO - add Mips implementation 1468 return false; 1469 } 1470 RegLocation rl_src = info->args[0]; 1471 rl_src = LoadValue(rl_src, kCoreReg); 1472 RegLocation rl_dest = InlineTarget(info); 1473 RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true); 1474 RegStorage sign_reg = AllocTemp(); 1475 // abs(x) = y<=x>>31, (x+y)^y. 1476 OpRegRegImm(kOpAsr, sign_reg, rl_src.reg, 31); 1477 OpRegRegReg(kOpAdd, rl_result.reg, rl_src.reg, sign_reg); 1478 OpRegReg(kOpXor, rl_result.reg, sign_reg); 1479 StoreValue(rl_dest, rl_result); 1480 return true; 1481} 1482 1483bool Mir2Lir::GenInlinedAbsLong(CallInfo* info) { 1484 if (cu_->instruction_set == kMips) { 1485 // TODO - add Mips implementation 1486 return false; 1487 } 1488 RegLocation rl_src = info->args[0]; 1489 rl_src = LoadValueWide(rl_src, kCoreReg); 1490 RegLocation rl_dest = InlineTargetWide(info); 1491 RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true); 1492 1493 // If on x86 or if we would clobber a register needed later, just copy the source first. 1494 if (cu_->instruction_set != kX86_64 && 1495 (cu_->instruction_set == kX86 || 1496 rl_result.reg.GetLowReg() == rl_src.reg.GetHighReg())) { 1497 OpRegCopyWide(rl_result.reg, rl_src.reg); 1498 if (rl_result.reg.GetLowReg() != rl_src.reg.GetLowReg() && 1499 rl_result.reg.GetLowReg() != rl_src.reg.GetHighReg() && 1500 rl_result.reg.GetHighReg() != rl_src.reg.GetLowReg() && 1501 rl_result.reg.GetHighReg() != rl_src.reg.GetHighReg()) { 1502 // Reuse source registers to avoid running out of temps. 1503 FreeTemp(rl_src.reg); 1504 } 1505 rl_src = rl_result; 1506 } 1507 1508 // abs(x) = y<=x>>31, (x+y)^y. 1509 RegStorage sign_reg; 1510 if (cu_->instruction_set == kX86_64) { 1511 sign_reg = AllocTempWide(); 1512 OpRegRegImm(kOpAsr, sign_reg, rl_src.reg, 63); 1513 OpRegRegReg(kOpAdd, rl_result.reg, rl_src.reg, sign_reg); 1514 OpRegReg(kOpXor, rl_result.reg, sign_reg); 1515 } else { 1516 sign_reg = AllocTemp(); 1517 OpRegRegImm(kOpAsr, sign_reg, rl_src.reg.GetHigh(), 31); 1518 OpRegRegReg(kOpAdd, rl_result.reg.GetLow(), rl_src.reg.GetLow(), sign_reg); 1519 OpRegRegReg(kOpAdc, rl_result.reg.GetHigh(), rl_src.reg.GetHigh(), sign_reg); 1520 OpRegReg(kOpXor, rl_result.reg.GetLow(), sign_reg); 1521 OpRegReg(kOpXor, rl_result.reg.GetHigh(), sign_reg); 1522 } 1523 FreeTemp(sign_reg); 1524 StoreValueWide(rl_dest, rl_result); 1525 return true; 1526} 1527 1528bool Mir2Lir::GenInlinedAbsFloat(CallInfo* info) { 1529 if (cu_->instruction_set == kMips) { 1530 // TODO - add Mips implementation 1531 return false; 1532 } 1533 RegLocation rl_src = info->args[0]; 1534 rl_src = LoadValue(rl_src, kCoreReg); 1535 RegLocation rl_dest = InlineTarget(info); 1536 RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true); 1537 OpRegRegImm(kOpAnd, rl_result.reg, rl_src.reg, 0x7fffffff); 1538 StoreValue(rl_dest, rl_result); 1539 return true; 1540} 1541 1542bool Mir2Lir::GenInlinedReverseBits(CallInfo* info, OpSize size) { 1543 // Currently implemented only for ARM64 1544 return false; 1545} 1546 1547bool Mir2Lir::GenInlinedMinMaxFP(CallInfo* info, bool is_min, bool is_double) { 1548 // Currently implemented only for ARM64 1549 return false; 1550} 1551 1552bool Mir2Lir::GenInlinedAbsDouble(CallInfo* info) { 1553 if (cu_->instruction_set == kMips) { 1554 // TODO - add Mips implementation 1555 return false; 1556 } 1557 RegLocation rl_src = info->args[0]; 1558 rl_src = LoadValueWide(rl_src, kCoreReg); 1559 RegLocation rl_dest = InlineTargetWide(info); 1560 RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true); 1561 1562 OpRegCopyWide(rl_result.reg, rl_src.reg); 1563 OpRegImm(kOpAnd, rl_result.reg.GetHigh(), 0x7fffffff); 1564 StoreValueWide(rl_dest, rl_result); 1565 return true; 1566} 1567 1568bool Mir2Lir::GenInlinedFloatCvt(CallInfo* info) { 1569 if (cu_->instruction_set == kMips) { 1570 // TODO - add Mips implementation 1571 return false; 1572 } 1573 RegLocation rl_src = info->args[0]; 1574 RegLocation rl_dest = InlineTarget(info); 1575 StoreValue(rl_dest, rl_src); 1576 return true; 1577} 1578 1579bool Mir2Lir::GenInlinedDoubleCvt(CallInfo* info) { 1580 if (cu_->instruction_set == kMips) { 1581 // TODO - add Mips implementation 1582 return false; 1583 } 1584 RegLocation rl_src = info->args[0]; 1585 RegLocation rl_dest = InlineTargetWide(info); 1586 StoreValueWide(rl_dest, rl_src); 1587 return true; 1588} 1589 1590bool Mir2Lir::GenInlinedArrayCopyCharArray(CallInfo* info) { 1591 return false; 1592} 1593 1594 1595/* 1596 * Fast String.indexOf(I) & (II). Tests for simple case of char <= 0xFFFF, 1597 * otherwise bails to standard library code. 1598 */ 1599bool Mir2Lir::GenInlinedIndexOf(CallInfo* info, bool zero_based) { 1600 if (cu_->instruction_set == kMips) { 1601 // TODO - add Mips implementation 1602 return false; 1603 } 1604 if (cu_->instruction_set == kX86_64) { 1605 // TODO - add kX86_64 implementation 1606 return false; 1607 } 1608 RegLocation rl_obj = info->args[0]; 1609 RegLocation rl_char = info->args[1]; 1610 if (rl_char.is_const && (mir_graph_->ConstantValue(rl_char) & ~0xFFFF) != 0) { 1611 // Code point beyond 0xFFFF. Punt to the real String.indexOf(). 1612 return false; 1613 } 1614 1615 ClobberCallerSave(); 1616 LockCallTemps(); // Using fixed registers 1617 RegStorage reg_ptr = TargetReg(kArg0, kRef); 1618 RegStorage reg_char = TargetReg(kArg1, kNotWide); 1619 RegStorage reg_start = TargetReg(kArg2, kNotWide); 1620 1621 LoadValueDirectFixed(rl_obj, reg_ptr); 1622 LoadValueDirectFixed(rl_char, reg_char); 1623 if (zero_based) { 1624 LoadConstant(reg_start, 0); 1625 } else { 1626 RegLocation rl_start = info->args[2]; // 3rd arg only present in III flavor of IndexOf. 1627 LoadValueDirectFixed(rl_start, reg_start); 1628 } 1629 RegStorage r_tgt = cu_->target64 ? 1630 LoadHelper(QUICK_ENTRYPOINT_OFFSET(8, pIndexOf)) : 1631 LoadHelper(QUICK_ENTRYPOINT_OFFSET(4, pIndexOf)); 1632 GenExplicitNullCheck(reg_ptr, info->opt_flags); 1633 LIR* high_code_point_branch = 1634 rl_char.is_const ? nullptr : OpCmpImmBranch(kCondGt, reg_char, 0xFFFF, nullptr); 1635 // NOTE: not a safepoint 1636 OpReg(kOpBlx, r_tgt); 1637 if (!rl_char.is_const) { 1638 // Add the slow path for code points beyond 0xFFFF. 1639 DCHECK(high_code_point_branch != nullptr); 1640 LIR* resume_tgt = NewLIR0(kPseudoTargetLabel); 1641 info->opt_flags |= MIR_IGNORE_NULL_CHECK; // Record that we've null checked. 1642 AddIntrinsicSlowPath(info, high_code_point_branch, resume_tgt); 1643 } else { 1644 DCHECK_EQ(mir_graph_->ConstantValue(rl_char) & ~0xFFFF, 0); 1645 DCHECK(high_code_point_branch == nullptr); 1646 } 1647 RegLocation rl_return = GetReturn(kCoreReg); 1648 RegLocation rl_dest = InlineTarget(info); 1649 StoreValue(rl_dest, rl_return); 1650 return true; 1651} 1652 1653/* Fast string.compareTo(Ljava/lang/string;)I. */ 1654bool Mir2Lir::GenInlinedStringCompareTo(CallInfo* info) { 1655 if (cu_->instruction_set == kMips) { 1656 // TODO - add Mips implementation 1657 return false; 1658 } 1659 ClobberCallerSave(); 1660 LockCallTemps(); // Using fixed registers 1661 RegStorage reg_this = TargetReg(kArg0, kRef); 1662 RegStorage reg_cmp = TargetReg(kArg1, kRef); 1663 1664 RegLocation rl_this = info->args[0]; 1665 RegLocation rl_cmp = info->args[1]; 1666 LoadValueDirectFixed(rl_this, reg_this); 1667 LoadValueDirectFixed(rl_cmp, reg_cmp); 1668 RegStorage r_tgt; 1669 if (cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64) { 1670 if (cu_->target64) { 1671 r_tgt = LoadHelper(QUICK_ENTRYPOINT_OFFSET(8, pStringCompareTo)); 1672 } else { 1673 r_tgt = LoadHelper(QUICK_ENTRYPOINT_OFFSET(4, pStringCompareTo)); 1674 } 1675 } else { 1676 r_tgt = RegStorage::InvalidReg(); 1677 } 1678 GenExplicitNullCheck(reg_this, info->opt_flags); 1679 info->opt_flags |= MIR_IGNORE_NULL_CHECK; // Record that we've null checked. 1680 // TUNING: check if rl_cmp.s_reg_low is already null checked 1681 LIR* cmp_null_check_branch = OpCmpImmBranch(kCondEq, reg_cmp, 0, nullptr); 1682 AddIntrinsicSlowPath(info, cmp_null_check_branch); 1683 // NOTE: not a safepoint 1684 if (cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64) { 1685 OpReg(kOpBlx, r_tgt); 1686 } else { 1687 if (cu_->target64) { 1688 OpThreadMem(kOpBlx, QUICK_ENTRYPOINT_OFFSET(8, pStringCompareTo)); 1689 } else { 1690 OpThreadMem(kOpBlx, QUICK_ENTRYPOINT_OFFSET(4, pStringCompareTo)); 1691 } 1692 } 1693 RegLocation rl_return = GetReturn(kCoreReg); 1694 RegLocation rl_dest = InlineTarget(info); 1695 StoreValue(rl_dest, rl_return); 1696 return true; 1697} 1698 1699bool Mir2Lir::GenInlinedCurrentThread(CallInfo* info) { 1700 RegLocation rl_dest = InlineTarget(info); 1701 1702 // Early exit if the result is unused. 1703 if (rl_dest.orig_sreg < 0) { 1704 return true; 1705 } 1706 1707 RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true); 1708 1709 switch (cu_->instruction_set) { 1710 case kArm: 1711 // Fall-through. 1712 case kThumb2: 1713 // Fall-through. 1714 case kMips: 1715 Load32Disp(TargetPtrReg(kSelf), Thread::PeerOffset<4>().Int32Value(), rl_result.reg); 1716 break; 1717 1718 case kArm64: 1719 LoadRefDisp(TargetPtrReg(kSelf), Thread::PeerOffset<8>().Int32Value(), rl_result.reg, 1720 kNotVolatile); 1721 break; 1722 1723 case kX86: 1724 reinterpret_cast<X86Mir2Lir*>(this)->OpRegThreadMem(kOpMov, rl_result.reg, 1725 Thread::PeerOffset<4>()); 1726 break; 1727 1728 case kX86_64: 1729 reinterpret_cast<X86Mir2Lir*>(this)->OpRegThreadMem(kOpMov, rl_result.reg, 1730 Thread::PeerOffset<8>()); 1731 break; 1732 1733 default: 1734 LOG(FATAL) << "Unexpected isa " << cu_->instruction_set; 1735 } 1736 StoreValue(rl_dest, rl_result); 1737 return true; 1738} 1739 1740bool Mir2Lir::GenInlinedUnsafeGet(CallInfo* info, 1741 bool is_long, bool is_volatile) { 1742 if (cu_->instruction_set == kMips) { 1743 // TODO - add Mips implementation 1744 return false; 1745 } 1746 // Unused - RegLocation rl_src_unsafe = info->args[0]; 1747 RegLocation rl_src_obj = info->args[1]; // Object 1748 RegLocation rl_src_offset = info->args[2]; // long low 1749 rl_src_offset = NarrowRegLoc(rl_src_offset); // ignore high half in info->args[3] 1750 RegLocation rl_dest = is_long ? InlineTargetWide(info) : InlineTarget(info); // result reg 1751 1752 RegLocation rl_object = LoadValue(rl_src_obj, kRefReg); 1753 RegLocation rl_offset = LoadValue(rl_src_offset, kCoreReg); 1754 RegLocation rl_result = EvalLoc(rl_dest, LocToRegClass(rl_dest), true); 1755 if (is_long) { 1756 if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64 1757 || cu_->instruction_set == kArm64) { 1758 LoadBaseIndexed(rl_object.reg, rl_offset.reg, rl_result.reg, 0, k64); 1759 } else { 1760 RegStorage rl_temp_offset = AllocTemp(); 1761 OpRegRegReg(kOpAdd, rl_temp_offset, rl_object.reg, rl_offset.reg); 1762 LoadBaseDisp(rl_temp_offset, 0, rl_result.reg, k64, kNotVolatile); 1763 FreeTemp(rl_temp_offset); 1764 } 1765 } else { 1766 if (rl_result.ref) { 1767 LoadRefIndexed(rl_object.reg, rl_offset.reg, rl_result.reg, 0); 1768 } else { 1769 LoadBaseIndexed(rl_object.reg, rl_offset.reg, rl_result.reg, 0, k32); 1770 } 1771 } 1772 1773 if (is_volatile) { 1774 GenMemBarrier(kLoadAny); 1775 } 1776 1777 if (is_long) { 1778 StoreValueWide(rl_dest, rl_result); 1779 } else { 1780 StoreValue(rl_dest, rl_result); 1781 } 1782 return true; 1783} 1784 1785bool Mir2Lir::GenInlinedUnsafePut(CallInfo* info, bool is_long, 1786 bool is_object, bool is_volatile, bool is_ordered) { 1787 if (cu_->instruction_set == kMips) { 1788 // TODO - add Mips implementation 1789 return false; 1790 } 1791 // Unused - RegLocation rl_src_unsafe = info->args[0]; 1792 RegLocation rl_src_obj = info->args[1]; // Object 1793 RegLocation rl_src_offset = info->args[2]; // long low 1794 rl_src_offset = NarrowRegLoc(rl_src_offset); // ignore high half in info->args[3] 1795 RegLocation rl_src_value = info->args[4]; // value to store 1796 if (is_volatile || is_ordered) { 1797 GenMemBarrier(kAnyStore); 1798 } 1799 RegLocation rl_object = LoadValue(rl_src_obj, kRefReg); 1800 RegLocation rl_offset = LoadValue(rl_src_offset, kCoreReg); 1801 RegLocation rl_value; 1802 if (is_long) { 1803 rl_value = LoadValueWide(rl_src_value, kCoreReg); 1804 if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64 1805 || cu_->instruction_set == kArm64) { 1806 StoreBaseIndexed(rl_object.reg, rl_offset.reg, rl_value.reg, 0, k64); 1807 } else { 1808 RegStorage rl_temp_offset = AllocTemp(); 1809 OpRegRegReg(kOpAdd, rl_temp_offset, rl_object.reg, rl_offset.reg); 1810 StoreBaseDisp(rl_temp_offset, 0, rl_value.reg, k64, kNotVolatile); 1811 FreeTemp(rl_temp_offset); 1812 } 1813 } else { 1814 rl_value = LoadValue(rl_src_value); 1815 if (rl_value.ref) { 1816 StoreRefIndexed(rl_object.reg, rl_offset.reg, rl_value.reg, 0); 1817 } else { 1818 StoreBaseIndexed(rl_object.reg, rl_offset.reg, rl_value.reg, 0, k32); 1819 } 1820 } 1821 1822 // Free up the temp early, to ensure x86 doesn't run out of temporaries in MarkGCCard. 1823 FreeTemp(rl_offset.reg); 1824 1825 if (is_volatile) { 1826 // Prevent reordering with a subsequent volatile load. 1827 // May also be needed to address store atomicity issues. 1828 GenMemBarrier(kAnyAny); 1829 } 1830 if (is_object) { 1831 MarkGCCard(rl_value.reg, rl_object.reg); 1832 } 1833 return true; 1834} 1835 1836void Mir2Lir::GenInvoke(CallInfo* info) { 1837 if ((info->opt_flags & MIR_INLINED) != 0) { 1838 // Already inlined but we may still need the null check. 1839 if (info->type != kStatic && 1840 ((cu_->disable_opt & (1 << kNullCheckElimination)) != 0 || 1841 (info->opt_flags & MIR_IGNORE_NULL_CHECK) == 0)) { 1842 RegLocation rl_obj = LoadValue(info->args[0], kRefReg); 1843 GenNullCheck(rl_obj.reg); 1844 } 1845 return; 1846 } 1847 DCHECK(cu_->compiler_driver->GetMethodInlinerMap() != nullptr); 1848 if (cu_->compiler_driver->GetMethodInlinerMap()->GetMethodInliner(cu_->dex_file) 1849 ->GenIntrinsic(this, info)) { 1850 return; 1851 } 1852 GenInvokeNoInline(info); 1853} 1854 1855template <size_t pointer_size> 1856static LIR* GenInvokeNoInlineCall(Mir2Lir* mir_to_lir, InvokeType type) { 1857 ThreadOffset<pointer_size> trampoline(-1); 1858 switch (type) { 1859 case kInterface: 1860 trampoline = QUICK_ENTRYPOINT_OFFSET(pointer_size, pInvokeInterfaceTrampolineWithAccessCheck); 1861 break; 1862 case kDirect: 1863 trampoline = QUICK_ENTRYPOINT_OFFSET(pointer_size, pInvokeDirectTrampolineWithAccessCheck); 1864 break; 1865 case kStatic: 1866 trampoline = QUICK_ENTRYPOINT_OFFSET(pointer_size, pInvokeStaticTrampolineWithAccessCheck); 1867 break; 1868 case kSuper: 1869 trampoline = QUICK_ENTRYPOINT_OFFSET(pointer_size, pInvokeSuperTrampolineWithAccessCheck); 1870 break; 1871 case kVirtual: 1872 trampoline = QUICK_ENTRYPOINT_OFFSET(pointer_size, pInvokeVirtualTrampolineWithAccessCheck); 1873 break; 1874 default: 1875 LOG(FATAL) << "Unexpected invoke type"; 1876 } 1877 return mir_to_lir->OpThreadMem(kOpBlx, trampoline); 1878} 1879 1880void Mir2Lir::GenInvokeNoInline(CallInfo* info) { 1881 int call_state = 0; 1882 LIR* null_ck; 1883 LIR** p_null_ck = NULL; 1884 NextCallInsn next_call_insn; 1885 FlushAllRegs(); /* Everything to home location */ 1886 // Explicit register usage 1887 LockCallTemps(); 1888 1889 const MirMethodLoweringInfo& method_info = mir_graph_->GetMethodLoweringInfo(info->mir); 1890 cu_->compiler_driver->ProcessedInvoke(method_info.GetInvokeType(), method_info.StatsFlags()); 1891 BeginInvoke(info); 1892 InvokeType original_type = static_cast<InvokeType>(method_info.GetInvokeType()); 1893 info->type = static_cast<InvokeType>(method_info.GetSharpType()); 1894 bool fast_path = method_info.FastPath(); 1895 bool skip_this; 1896 if (info->type == kInterface) { 1897 next_call_insn = fast_path ? NextInterfaceCallInsn : NextInterfaceCallInsnWithAccessCheck; 1898 skip_this = fast_path; 1899 } else if (info->type == kDirect) { 1900 if (fast_path) { 1901 p_null_ck = &null_ck; 1902 } 1903 next_call_insn = fast_path ? NextSDCallInsn : NextDirectCallInsnSP; 1904 skip_this = false; 1905 } else if (info->type == kStatic) { 1906 next_call_insn = fast_path ? NextSDCallInsn : NextStaticCallInsnSP; 1907 skip_this = false; 1908 } else if (info->type == kSuper) { 1909 DCHECK(!fast_path); // Fast path is a direct call. 1910 next_call_insn = NextSuperCallInsnSP; 1911 skip_this = false; 1912 } else { 1913 DCHECK_EQ(info->type, kVirtual); 1914 next_call_insn = fast_path ? NextVCallInsn : NextVCallInsnSP; 1915 skip_this = fast_path; 1916 } 1917 MethodReference target_method = method_info.GetTargetMethod(); 1918 if (!info->is_range) { 1919 call_state = GenDalvikArgsNoRange(info, call_state, p_null_ck, 1920 next_call_insn, target_method, method_info.VTableIndex(), 1921 method_info.DirectCode(), method_info.DirectMethod(), 1922 original_type, skip_this); 1923 } else { 1924 call_state = GenDalvikArgsRange(info, call_state, p_null_ck, 1925 next_call_insn, target_method, method_info.VTableIndex(), 1926 method_info.DirectCode(), method_info.DirectMethod(), 1927 original_type, skip_this); 1928 } 1929 // Finish up any of the call sequence not interleaved in arg loading 1930 while (call_state >= 0) { 1931 call_state = next_call_insn(cu_, info, call_state, target_method, method_info.VTableIndex(), 1932 method_info.DirectCode(), method_info.DirectMethod(), original_type); 1933 } 1934 LIR* call_inst; 1935 if (cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64) { 1936 call_inst = OpReg(kOpBlx, TargetPtrReg(kInvokeTgt)); 1937 } else { 1938 if (fast_path) { 1939 if (method_info.DirectCode() == static_cast<uintptr_t>(-1)) { 1940 // We can have the linker fixup a call relative. 1941 call_inst = 1942 reinterpret_cast<X86Mir2Lir*>(this)->CallWithLinkerFixup(target_method, info->type); 1943 } else { 1944 call_inst = OpMem(kOpBlx, TargetReg(kArg0, kRef), 1945 mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset().Int32Value()); 1946 } 1947 } else { 1948 // TODO: Extract? 1949 if (cu_->target64) { 1950 call_inst = GenInvokeNoInlineCall<8>(this, info->type); 1951 } else { 1952 call_inst = GenInvokeNoInlineCall<4>(this, info->type); 1953 } 1954 } 1955 } 1956 EndInvoke(info); 1957 MarkSafepointPC(call_inst); 1958 1959 ClobberCallerSave(); 1960 if (info->result.location != kLocInvalid) { 1961 // We have a following MOVE_RESULT - do it now. 1962 if (info->result.wide) { 1963 RegLocation ret_loc = GetReturnWide(LocToRegClass(info->result)); 1964 StoreValueWide(info->result, ret_loc); 1965 } else { 1966 RegLocation ret_loc = GetReturn(LocToRegClass(info->result)); 1967 StoreValue(info->result, ret_loc); 1968 } 1969 } 1970} 1971 1972} // namespace art 1973