buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (C) 2011 The Android Open Source Project |
| 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | */ |
| 16 | |
| 17 | /* |
| 18 | * This file contains codegen for the Thumb2 ISA and is intended to be |
| 19 | * includes by: |
| 20 | * |
| 21 | * Codegen-$(TARGET_ARCH_VARIANT).c |
| 22 | * |
| 23 | */ |
| 24 | |
| 25 | /* |
| 26 | * Construct an s4 from two consecutive half-words of switch data. |
| 27 | * This needs to check endianness because the DEX optimizer only swaps |
| 28 | * half-words in instruction stream. |
| 29 | * |
| 30 | * "switchData" must be 32-bit aligned. |
| 31 | */ |
| 32 | #if __BYTE_ORDER == __LITTLE_ENDIAN |
| 33 | static inline s4 s4FromSwitchData(const void* switchData) { |
| 34 | return *(s4*) switchData; |
| 35 | } |
| 36 | #else |
| 37 | static inline s4 s4FromSwitchData(const void* switchData) { |
| 38 | u2* data = switchData; |
| 39 | return data[0] | (((s4) data[1]) << 16); |
| 40 | } |
| 41 | #endif |
| 42 | |
| 43 | /* |
| 44 | * Generate a Thumb2 IT instruction, which can nullify up to |
| 45 | * four subsequent instructions based on a condition and its |
| 46 | * inverse. The condition applies to the first instruction, which |
| 47 | * is executed if the condition is met. The string "guide" consists |
| 48 | * of 0 to 3 chars, and applies to the 2nd through 4th instruction. |
| 49 | * A "T" means the instruction is executed if the condition is |
| 50 | * met, and an "E" means the instruction is executed if the condition |
| 51 | * is not met. |
| 52 | */ |
| 53 | static ArmLIR* genIT(CompilationUnit* cUnit, ArmConditionCode code, |
| 54 | const char* guide) |
| 55 | { |
| 56 | int mask; |
| 57 | int condBit = code & 1; |
| 58 | int altBit = condBit ^ 1; |
| 59 | int mask3 = 0; |
| 60 | int mask2 = 0; |
| 61 | int mask1 = 0; |
| 62 | |
| 63 | //Note: case fallthroughs intentional |
| 64 | switch(strlen(guide)) { |
| 65 | case 3: |
| 66 | mask1 = (guide[2] == 'T') ? condBit : altBit; |
| 67 | case 2: |
| 68 | mask2 = (guide[1] == 'T') ? condBit : altBit; |
| 69 | case 1: |
| 70 | mask3 = (guide[0] == 'T') ? condBit : altBit; |
| 71 | break; |
| 72 | case 0: |
| 73 | break; |
| 74 | default: |
| 75 | LOG(FATAL) << "OAT: bad case in genIT"; |
| 76 | } |
| 77 | mask = (mask3 << 3) | (mask2 << 2) | (mask1 << 1) | |
| 78 | (1 << (3 - strlen(guide))); |
| 79 | return newLIR2(cUnit, kThumb2It, code, mask); |
| 80 | } |
| 81 | |
| 82 | /* |
| 83 | * Insert a kArmPseudoCaseLabel at the beginning of the Dalvik |
| 84 | * offset vaddr. This label will be used to fix up the case |
| 85 | * branch table during the assembly phase. Be sure to set |
| 86 | * all resource flags on this to prevent code motion across |
| 87 | * target boundaries. KeyVal is just there for debugging. |
| 88 | */ |
| 89 | static ArmLIR* insertCaseLabel(CompilationUnit* cUnit, int vaddr, int keyVal) |
| 90 | { |
| 91 | ArmLIR* lir; |
| 92 | for (lir = (ArmLIR*)cUnit->firstLIRInsn; lir; lir = NEXT_LIR(lir)) { |
| 93 | if ((lir->opcode == kArmPseudoDalvikByteCodeBoundary) && |
| 94 | (lir->generic.dalvikOffset == vaddr)) { |
| 95 | ArmLIR* newLabel = (ArmLIR*)oatNew(sizeof(ArmLIR), true); |
| 96 | newLabel->generic.dalvikOffset = vaddr; |
| 97 | newLabel->opcode = kArmPseudoCaseLabel; |
| 98 | newLabel->operands[0] = keyVal; |
| 99 | oatInsertLIRAfter((LIR*)lir, (LIR*)newLabel); |
| 100 | return newLabel; |
| 101 | } |
| 102 | } |
| 103 | oatCodegenDump(cUnit); |
| 104 | LOG(FATAL) << "Error: didn't find vaddr 0x" << std::hex << vaddr; |
| 105 | return NULL; // Quiet gcc |
| 106 | } |
| 107 | |
| 108 | static void markPackedCaseLabels(CompilationUnit* cUnit, SwitchTable *tabRec) |
| 109 | { |
| 110 | const u2* table = tabRec->table; |
| 111 | int baseVaddr = tabRec->vaddr; |
| 112 | int *targets = (int*)&table[4]; |
| 113 | int entries = table[1]; |
| 114 | int lowKey = s4FromSwitchData(&table[2]); |
| 115 | for (int i = 0; i < entries; i++) { |
| 116 | tabRec->targets[i] = insertCaseLabel(cUnit, baseVaddr + targets[i], |
| 117 | i + lowKey); |
| 118 | } |
| 119 | } |
| 120 | |
| 121 | static void markSparseCaseLabels(CompilationUnit* cUnit, SwitchTable *tabRec) |
| 122 | { |
| 123 | const u2* table = tabRec->table; |
| 124 | int baseVaddr = tabRec->vaddr; |
| 125 | int entries = table[1]; |
| 126 | int* keys = (int*)&table[2]; |
| 127 | int* targets = &keys[entries]; |
| 128 | for (int i = 0; i < entries; i++) { |
| 129 | tabRec->targets[i] = insertCaseLabel(cUnit, baseVaddr + targets[i], |
| 130 | keys[i]); |
| 131 | } |
| 132 | } |
| 133 | |
| 134 | void oatProcessSwitchTables(CompilationUnit* cUnit) |
| 135 | { |
| 136 | GrowableListIterator iterator; |
| 137 | oatGrowableListIteratorInit(&cUnit->switchTables, &iterator); |
| 138 | while (true) { |
| 139 | SwitchTable *tabRec = (SwitchTable *) oatGrowableListIteratorNext( |
| 140 | &iterator); |
| 141 | if (tabRec == NULL) break; |
| 142 | if (tabRec->table[0] == kPackedSwitchSignature) |
| 143 | markPackedCaseLabels(cUnit, tabRec); |
| 144 | else if (tabRec->table[0] == kSparseSwitchSignature) |
| 145 | markSparseCaseLabels(cUnit, tabRec); |
| 146 | else { |
| 147 | LOG(FATAL) << "Invalid switch table"; |
| 148 | } |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | static void dumpSparseSwitchTable(const u2* table) |
| 153 | /* |
| 154 | * Sparse switch data format: |
| 155 | * ushort ident = 0x0200 magic value |
| 156 | * ushort size number of entries in the table; > 0 |
| 157 | * int keys[size] keys, sorted low-to-high; 32-bit aligned |
| 158 | * int targets[size] branch targets, relative to switch opcode |
| 159 | * |
| 160 | * Total size is (2+size*4) 16-bit code units. |
| 161 | */ |
| 162 | { |
| 163 | u2 ident = table[0]; |
| 164 | int entries = table[1]; |
| 165 | int* keys = (int*)&table[2]; |
| 166 | int* targets = &keys[entries]; |
| 167 | LOG(INFO) << "Sparse switch table - ident:0x" << std::hex << ident << |
| 168 | ", entries: " << std::dec << entries; |
| 169 | for (int i = 0; i < entries; i++) { |
| 170 | LOG(INFO) << " Key[" << keys[i] << "] -> 0x" << std::hex << |
| 171 | targets[i]; |
| 172 | } |
| 173 | } |
| 174 | |
| 175 | static void dumpPackedSwitchTable(const u2* table) |
| 176 | /* |
| 177 | * Packed switch data format: |
| 178 | * ushort ident = 0x0100 magic value |
| 179 | * ushort size number of entries in the table |
| 180 | * int first_key first (and lowest) switch case value |
| 181 | * int targets[size] branch targets, relative to switch opcode |
| 182 | * |
| 183 | * Total size is (4+size*2) 16-bit code units. |
| 184 | */ |
| 185 | { |
| 186 | u2 ident = table[0]; |
| 187 | int* targets = (int*)&table[4]; |
| 188 | int entries = table[1]; |
| 189 | int lowKey = s4FromSwitchData(&table[2]); |
| 190 | LOG(INFO) << "Packed switch table - ident:0x" << std::hex << ident << |
| 191 | ", entries: " << std::dec << entries << ", lowKey: " << lowKey; |
| 192 | for (int i = 0; i < entries; i++) { |
| 193 | LOG(INFO) << " Key[" << (i + lowKey) << "] -> 0x" << std::hex << |
| 194 | targets[i]; |
| 195 | } |
| 196 | } |
| 197 | |
| 198 | /* |
| 199 | * The sparse table in the literal pool is an array of <key,displacement> |
| 200 | * pairs. For each set, we'll load them as a pair using ldmia. |
| 201 | * This means that the register number of the temp we use for the key |
| 202 | * must be lower than the reg for the displacement. |
| 203 | * |
| 204 | * The test loop will look something like: |
| 205 | * |
| 206 | * adr rBase, <table> |
| 207 | * ldr rVal, [rSP, vRegOff] |
| 208 | * mov rIdx, #tableSize |
| 209 | * lp: |
| 210 | * ldmia rBase!, {rKey, rDisp} |
| 211 | * sub rIdx, #1 |
| 212 | * cmp rVal, rKey |
| 213 | * ifeq |
| 214 | * add rPC, rDisp ; This is the branch from which we compute displacement |
| 215 | * cbnz rIdx, lp |
| 216 | */ |
| 217 | static void genSparseSwitch(CompilationUnit* cUnit, MIR* mir, |
| 218 | RegLocation rlSrc) |
| 219 | { |
| 220 | const u2* table = cUnit->insns + mir->offset + mir->dalvikInsn.vB; |
| 221 | if (cUnit->printMe) { |
| 222 | dumpSparseSwitchTable(table); |
| 223 | } |
| 224 | // Add the table to the list - we'll process it later |
| 225 | SwitchTable *tabRec = (SwitchTable *)oatNew(sizeof(SwitchTable), |
| 226 | true); |
| 227 | tabRec->table = table; |
| 228 | tabRec->vaddr = mir->offset; |
| 229 | int size = table[1]; |
| 230 | tabRec->targets = (ArmLIR* *)oatNew(size * sizeof(ArmLIR*), true); |
| 231 | oatInsertGrowableList(&cUnit->switchTables, (intptr_t)tabRec); |
| 232 | |
| 233 | // Get the switch value |
| 234 | rlSrc = loadValue(cUnit, rlSrc, kCoreReg); |
| 235 | int rBase = oatAllocTemp(cUnit); |
| 236 | /* Allocate key and disp temps */ |
| 237 | int rKey = oatAllocTemp(cUnit); |
| 238 | int rDisp = oatAllocTemp(cUnit); |
| 239 | // Make sure rKey's register number is less than rDisp's number for ldmia |
| 240 | if (rKey > rDisp) { |
| 241 | int tmp = rDisp; |
| 242 | rDisp = rKey; |
| 243 | rKey = tmp; |
| 244 | } |
| 245 | // Materialize a pointer to the switch table |
| 246 | newLIR3(cUnit, kThumb2AdrST, rBase, 0, (intptr_t)tabRec); |
| 247 | // Set up rIdx |
| 248 | int rIdx = oatAllocTemp(cUnit); |
| 249 | loadConstant(cUnit, rIdx, size); |
| 250 | // Establish loop branch target |
| 251 | ArmLIR* target = newLIR0(cUnit, kArmPseudoTargetLabel); |
| 252 | target->defMask = ENCODE_ALL; |
| 253 | // Load next key/disp |
| 254 | newLIR2(cUnit, kThumb2LdmiaWB, rBase, (1 << rKey) | (1 << rDisp)); |
| 255 | opRegReg(cUnit, kOpCmp, rKey, rlSrc.lowReg); |
| 256 | // Go if match. NOTE: No instruction set switch here - must stay Thumb2 |
| 257 | genIT(cUnit, kArmCondEq, ""); |
| 258 | ArmLIR* switchBranch = newLIR1(cUnit, kThumb2AddPCR, rDisp); |
| 259 | tabRec->bxInst = switchBranch; |
| 260 | // Needs to use setflags encoding here |
| 261 | newLIR3(cUnit, kThumb2SubsRRI12, rIdx, rIdx, 1); |
| 262 | ArmLIR* branch = opCondBranch(cUnit, kArmCondNe); |
| 263 | branch->generic.target = (LIR*)target; |
| 264 | } |
| 265 | |
| 266 | |
| 267 | static void genPackedSwitch(CompilationUnit* cUnit, MIR* mir, |
| 268 | RegLocation rlSrc) |
| 269 | { |
| 270 | const u2* table = cUnit->insns + mir->offset + mir->dalvikInsn.vB; |
| 271 | if (cUnit->printMe) { |
| 272 | dumpPackedSwitchTable(table); |
| 273 | } |
| 274 | // Add the table to the list - we'll process it later |
| 275 | SwitchTable *tabRec = (SwitchTable *)oatNew(sizeof(SwitchTable), |
| 276 | true); |
| 277 | tabRec->table = table; |
| 278 | tabRec->vaddr = mir->offset; |
| 279 | int size = table[1]; |
| 280 | tabRec->targets = (ArmLIR* *)oatNew(size * sizeof(ArmLIR*), true); |
| 281 | oatInsertGrowableList(&cUnit->switchTables, (intptr_t)tabRec); |
| 282 | |
| 283 | // Get the switch value |
| 284 | rlSrc = loadValue(cUnit, rlSrc, kCoreReg); |
| 285 | int tableBase = oatAllocTemp(cUnit); |
| 286 | // Materialize a pointer to the switch table |
| 287 | newLIR3(cUnit, kThumb2AdrST, tableBase, 0, (intptr_t)tabRec); |
| 288 | int lowKey = s4FromSwitchData(&table[2]); |
| 289 | int keyReg; |
| 290 | // Remove the bias, if necessary |
| 291 | if (lowKey == 0) { |
| 292 | keyReg = rlSrc.lowReg; |
| 293 | } else { |
| 294 | keyReg = oatAllocTemp(cUnit); |
| 295 | opRegRegImm(cUnit, kOpSub, keyReg, rlSrc.lowReg, lowKey); |
| 296 | } |
| 297 | // Bounds check - if < 0 or >= size continue following switch |
| 298 | opRegImm(cUnit, kOpCmp, keyReg, size-1); |
| 299 | ArmLIR* branchOver = opCondBranch(cUnit, kArmCondHi); |
| 300 | |
| 301 | // Load the displacement from the switch table |
| 302 | int dispReg = oatAllocTemp(cUnit); |
| 303 | loadBaseIndexed(cUnit, tableBase, keyReg, dispReg, 2, kWord); |
| 304 | |
| 305 | // ..and go! NOTE: No instruction set switch here - must stay Thumb2 |
| 306 | ArmLIR* switchBranch = newLIR1(cUnit, kThumb2AddPCR, dispReg); |
| 307 | tabRec->bxInst = switchBranch; |
| 308 | |
| 309 | /* branchOver target here */ |
| 310 | ArmLIR* target = newLIR0(cUnit, kArmPseudoTargetLabel); |
| 311 | target->defMask = ENCODE_ALL; |
| 312 | branchOver->generic.target = (LIR*)target; |
| 313 | } |
| 314 | |
| 315 | /* |
| 316 | * Array data table format: |
| 317 | * ushort ident = 0x0300 magic value |
| 318 | * ushort width width of each element in the table |
| 319 | * uint size number of elements in the table |
| 320 | * ubyte data[size*width] table of data values (may contain a single-byte |
| 321 | * padding at the end) |
| 322 | * |
| 323 | * Total size is 4+(width * size + 1)/2 16-bit code units. |
| 324 | */ |
| 325 | static void genFillArrayData(CompilationUnit* cUnit, MIR* mir, |
| 326 | RegLocation rlSrc) |
| 327 | { |
| 328 | const u2* table = cUnit->insns + mir->offset + mir->dalvikInsn.vB; |
| 329 | // Add the table to the list - we'll process it later |
| 330 | FillArrayData *tabRec = (FillArrayData *) |
| 331 | oatNew(sizeof(FillArrayData), true); |
| 332 | tabRec->table = table; |
| 333 | tabRec->vaddr = mir->offset; |
| 334 | u2 width = tabRec->table[1]; |
| 335 | u4 size = tabRec->table[2] | (((u4)tabRec->table[3]) << 16); |
| 336 | tabRec->size = (size * width) + 8; |
| 337 | |
| 338 | oatInsertGrowableList(&cUnit->fillArrayData, (intptr_t)tabRec); |
| 339 | |
| 340 | // Making a call - use explicit registers |
| 341 | oatFlushAllRegs(cUnit); /* Everything to home location */ |
| 342 | loadValueDirectFixed(cUnit, rlSrc, r0); |
| 343 | loadWordDisp(cUnit, rSELF, |
| 344 | OFFSETOF_MEMBER(Thread, pArtHandleFillArrayDataNoThrow), rLR); |
| 345 | // Materialize a pointer to the switch table |
| 346 | newLIR3(cUnit, kThumb2AdrST, r1, 0, (intptr_t)tabRec); |
| 347 | opReg(cUnit, kOpBlx, rLR); |
| 348 | oatClobberCallRegs(cUnit); |
| 349 | } |
| 350 | |
| 351 | /* |
| 352 | * Mark garbage collection card. Skip if the value we're storing is null. |
| 353 | */ |
| 354 | static void markGCCard(CompilationUnit* cUnit, int valReg, int tgtAddrReg) |
| 355 | { |
| 356 | int regCardBase = oatAllocTemp(cUnit); |
| 357 | int regCardNo = oatAllocTemp(cUnit); |
| 358 | ArmLIR* branchOver = genCmpImmBranch(cUnit, kArmCondEq, valReg, 0); |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 359 | loadWordDisp(cUnit, rSELF, Thread::CardTableOffset().Int32Value(), |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 360 | regCardBase); |
| 361 | opRegRegImm(cUnit, kOpLsr, regCardNo, tgtAddrReg, GC_CARD_SHIFT); |
| 362 | storeBaseIndexed(cUnit, regCardBase, regCardNo, regCardBase, 0, |
| 363 | kUnsignedByte); |
| 364 | ArmLIR* target = newLIR0(cUnit, kArmPseudoTargetLabel); |
| 365 | target->defMask = ENCODE_ALL; |
| 366 | branchOver->generic.target = (LIR*)target; |
| 367 | oatFreeTemp(cUnit, regCardBase); |
| 368 | oatFreeTemp(cUnit, regCardNo); |
| 369 | } |
| 370 | |
| 371 | static void genIGetX(CompilationUnit* cUnit, MIR* mir, OpSize size, |
| 372 | RegLocation rlDest, RegLocation rlObj) |
| 373 | { |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 374 | Field* fieldPtr = cUnit->method->GetDeclaringClass()->GetDexCache()-> |
| 375 | GetResolvedField(mir->dalvikInsn.vC); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 376 | if (fieldPtr == NULL) { |
| 377 | /* |
| 378 | * With current scheme, we should never be in a situation |
| 379 | * in which the fieldPtr is null here. If something changes |
| 380 | * and we need to handle it, generate code to load the field |
| 381 | * pointer at run-time. |
| 382 | */ |
| 383 | LOG(FATAL) << "Unexpected null field pointer"; |
| 384 | } |
| 385 | #if ANDROID_SMP != 0 |
| 386 | bool isVolatile = dvmIsVolatileField(fieldPtr); |
| 387 | #else |
| 388 | bool isVolatile = false; |
| 389 | #endif |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 390 | int fieldOffset = fieldPtr->GetOffset(); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 391 | RegLocation rlResult; |
| 392 | RegisterClass regClass = oatRegClassBySize(size); |
| 393 | rlObj = loadValue(cUnit, rlObj, kCoreReg); |
| 394 | rlResult = oatEvalLoc(cUnit, rlDest, regClass, true); |
| 395 | genNullCheck(cUnit, rlObj.sRegLow, rlObj.lowReg, mir->offset, |
| 396 | NULL);/* null object? */ |
| 397 | loadBaseDisp(cUnit, mir, rlObj.lowReg, fieldOffset, rlResult.lowReg, |
| 398 | size, rlObj.sRegLow); |
| 399 | if (isVolatile) { |
| 400 | oatGenMemBarrier(cUnit, kSY); |
| 401 | } |
| 402 | |
| 403 | storeValue(cUnit, rlDest, rlResult); |
| 404 | } |
| 405 | |
| 406 | static void genIPutX(CompilationUnit* cUnit, MIR* mir, OpSize size, |
| 407 | RegLocation rlSrc, RegLocation rlObj, bool isObject) |
| 408 | { |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 409 | Field* fieldPtr = cUnit->method->GetDeclaringClass()->GetDexCache()-> |
| 410 | GetResolvedField(mir->dalvikInsn.vC); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 411 | if (fieldPtr == NULL) { |
| 412 | /* |
| 413 | * With current scheme, we should never be in a situation |
| 414 | * in which the fieldPtr is null here. If something changes |
| 415 | * and we need to handle it, generate code to load the field |
| 416 | * pointer at run-time. |
| 417 | */ |
| 418 | LOG(FATAL) << "Unexpected null field pointer"; |
| 419 | } |
| 420 | #if ANDROID_SMP != 0 |
| 421 | bool isVolatile = dvmIsVolatileField(fieldPtr); |
| 422 | #else |
| 423 | bool isVolatile = false; |
| 424 | #endif |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 425 | int fieldOffset = fieldPtr->GetOffset(); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 426 | RegisterClass regClass = oatRegClassBySize(size); |
| 427 | rlObj = loadValue(cUnit, rlObj, kCoreReg); |
| 428 | rlSrc = loadValue(cUnit, rlSrc, regClass); |
| 429 | genNullCheck(cUnit, rlObj.sRegLow, rlObj.lowReg, mir->offset, |
| 430 | NULL);/* null object? */ |
| 431 | |
| 432 | if (isVolatile) { |
| 433 | oatGenMemBarrier(cUnit, kSY); |
| 434 | } |
| 435 | storeBaseDisp(cUnit, rlObj.lowReg, fieldOffset, rlSrc.lowReg, size); |
| 436 | if (isObject) { |
| 437 | /* NOTE: marking card based on object head */ |
| 438 | markGCCard(cUnit, rlSrc.lowReg, rlObj.lowReg); |
| 439 | } |
| 440 | } |
| 441 | |
| 442 | static void genIGetWideX(CompilationUnit* cUnit, MIR* mir, RegLocation rlDest, |
| 443 | RegLocation rlObj) |
| 444 | { |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 445 | Field* fieldPtr = cUnit->method->GetDeclaringClass()->GetDexCache()-> |
| 446 | GetResolvedField(mir->dalvikInsn.vC); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 447 | if (fieldPtr == NULL) { |
| 448 | /* |
| 449 | * With current scheme, we should never be in a situation |
| 450 | * in which the fieldPtr is null here. If something changes |
| 451 | * and we need to handle it, generate code to load the field |
| 452 | * pointer at run-time. |
| 453 | */ |
| 454 | LOG(FATAL) << "Unexpected null field pointer"; |
| 455 | } |
| 456 | #if ANDROID_SMP != 0 |
| 457 | bool isVolatile = dvmIsVolatileField(fieldPtr); |
| 458 | #else |
| 459 | bool isVolatile = false; |
| 460 | #endif |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 461 | int fieldOffset = fieldPtr->GetOffset(); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 462 | RegLocation rlResult; |
| 463 | rlObj = loadValue(cUnit, rlObj, kCoreReg); |
| 464 | int regPtr = oatAllocTemp(cUnit); |
| 465 | |
| 466 | assert(rlDest.wide); |
| 467 | |
| 468 | genNullCheck(cUnit, rlObj.sRegLow, rlObj.lowReg, mir->offset, |
| 469 | NULL);/* null object? */ |
| 470 | opRegRegImm(cUnit, kOpAdd, regPtr, rlObj.lowReg, fieldOffset); |
| 471 | rlResult = oatEvalLoc(cUnit, rlDest, kAnyReg, true); |
| 472 | |
| 473 | loadPair(cUnit, regPtr, rlResult.lowReg, rlResult.highReg); |
| 474 | |
| 475 | if (isVolatile) { |
| 476 | oatGenMemBarrier(cUnit, kSY); |
| 477 | } |
| 478 | |
| 479 | oatFreeTemp(cUnit, regPtr); |
| 480 | storeValueWide(cUnit, rlDest, rlResult); |
| 481 | } |
| 482 | |
| 483 | static void genIPutWideX(CompilationUnit* cUnit, MIR* mir, RegLocation rlSrc, |
| 484 | RegLocation rlObj) |
| 485 | { |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 486 | Field* fieldPtr = cUnit->method->GetDeclaringClass()->GetDexCache()-> |
| 487 | GetResolvedField(mir->dalvikInsn.vC); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 488 | if (fieldPtr == NULL) { |
| 489 | /* |
| 490 | * With current scheme, we should never be in a situation |
| 491 | * in which the fieldPtr is null here. If something changes |
| 492 | * and we need to handle it, generate code to load the field |
| 493 | * pointer at run-time. |
| 494 | */ |
| 495 | LOG(FATAL) << "Unexpected null field pointer"; |
| 496 | } |
| 497 | #if ANDROID_SMP != 0 |
| 498 | bool isVolatile = dvmIsVolatileField(fieldPtr); |
| 499 | #else |
| 500 | bool isVolatile = false; |
| 501 | #endif |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 502 | int fieldOffset = fieldPtr->GetOffset(); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 503 | |
| 504 | rlObj = loadValue(cUnit, rlObj, kCoreReg); |
| 505 | int regPtr; |
| 506 | rlSrc = loadValueWide(cUnit, rlSrc, kAnyReg); |
| 507 | genNullCheck(cUnit, rlObj.sRegLow, rlObj.lowReg, mir->offset, |
| 508 | NULL);/* null object? */ |
| 509 | regPtr = oatAllocTemp(cUnit); |
| 510 | opRegRegImm(cUnit, kOpAdd, regPtr, rlObj.lowReg, fieldOffset); |
| 511 | |
| 512 | if (isVolatile) { |
| 513 | oatGenMemBarrier(cUnit, kSY); |
| 514 | } |
| 515 | storePair(cUnit, regPtr, rlSrc.lowReg, rlSrc.highReg); |
| 516 | |
| 517 | oatFreeTemp(cUnit, regPtr); |
| 518 | } |
| 519 | |
| 520 | static void genConstClass(CompilationUnit* cUnit, MIR* mir, |
| 521 | RegLocation rlDest, RegLocation rlSrc) |
| 522 | { |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 523 | Class* classPtr = cUnit->method->GetDeclaringClass()->GetDexCache()-> |
| 524 | GetResolvedClass(mir->dalvikInsn.vB); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 525 | |
| 526 | if (classPtr == NULL) { |
| 527 | LOG(FATAL) << "Unexpected null class pointer"; |
| 528 | } |
| 529 | |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 530 | UNIMPLEMENTED(WARNING) << "Not position independent. Fix"; |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 531 | RegLocation rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 532 | loadConstantNoClobber(cUnit, rlResult.lowReg, (int) classPtr ); |
| 533 | storeValue(cUnit, rlDest, rlResult); |
| 534 | } |
| 535 | |
| 536 | static void genConstString(CompilationUnit* cUnit, MIR* mir, |
| 537 | RegLocation rlDest, RegLocation rlSrc) |
| 538 | { |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 539 | String* strPtr = cUnit->method->GetDeclaringClass()->GetDexCache()-> |
| 540 | GetResolvedString(mir->dalvikInsn.vB); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 541 | |
| 542 | if (strPtr == NULL) { |
| 543 | /* Shouldn't happen */ |
| 544 | LOG(FATAL) << "Unexpected null const string pointer"; |
| 545 | } |
| 546 | |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 547 | UNIMPLEMENTED(WARNING) << "Not position indendent. Fix"; |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 548 | RegLocation rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 549 | loadConstantNoClobber(cUnit, rlResult.lowReg, (int) strPtr ); |
| 550 | storeValue(cUnit, rlDest, rlResult); |
| 551 | } |
| 552 | |
| 553 | static void genNewInstance(CompilationUnit* cUnit, MIR* mir, |
| 554 | RegLocation rlDest) |
| 555 | { |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 556 | Class* classPtr = cUnit->method->GetDeclaringClass()->GetDexCache()-> |
| 557 | GetResolvedClass(mir->dalvikInsn.vB); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 558 | |
| 559 | if (classPtr == NULL) { |
| 560 | /* Shouldn't happen */ |
| 561 | LOG(FATAL) << "Unexpected null class pointer"; |
| 562 | } |
| 563 | |
| 564 | // Verifier should have already rejected abstract/interface |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 565 | assert((classPtr->access_flags_ & |
| 566 | (art::kAccInterface|art::kAccAbstract)) == 0); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 567 | oatFlushAllRegs(cUnit); /* Everything to home location */ |
| 568 | loadWordDisp(cUnit, rSELF, |
| 569 | OFFSETOF_MEMBER(Thread, pArtAllocObjectNoThrow), rLR); |
| 570 | loadConstant(cUnit, r0, (int) classPtr); |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 571 | UNIMPLEMENTED(WARNING) << "Need NewWorld dvmAllocObject"; |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 572 | opReg(cUnit, kOpBlx, rLR); |
| 573 | oatClobberCallRegs(cUnit); |
| 574 | RegLocation rlResult = oatGetReturn(cUnit); |
| 575 | storeValue(cUnit, rlDest, rlResult); |
| 576 | } |
| 577 | |
| 578 | void genThrow(CompilationUnit* cUnit, MIR* mir, RegLocation rlSrc) |
| 579 | { |
| 580 | loadWordDisp(cUnit, rSELF, |
| 581 | OFFSETOF_MEMBER(Thread, pArtAllocObjectNoThrow), rLR); |
| 582 | loadValueDirectFixed(cUnit, rlSrc, r1); /* Exception object */ |
| 583 | genRegCopy(cUnit, r0, rSELF); |
| 584 | opReg(cUnit, kOpBlx, rLR); |
| 585 | } |
| 586 | |
| 587 | static void genInstanceof(CompilationUnit* cUnit, MIR* mir, RegLocation rlDest, |
| 588 | RegLocation rlSrc) |
| 589 | { |
| 590 | // May generate a call - use explicit registers |
| 591 | RegLocation rlResult; |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 592 | Class* classPtr = cUnit->method->GetDeclaringClass()->GetDexCache()-> |
| 593 | GetResolvedClass(mir->dalvikInsn.vC); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 594 | if (classPtr == NULL) { |
| 595 | /* Shouldn't happen */ |
| 596 | LOG(FATAL) << "Unexpected null class pointer"; |
| 597 | } |
| 598 | oatFlushAllRegs(cUnit); /* Everything to home location */ |
| 599 | loadValueDirectFixed(cUnit, rlSrc, r0); /* Ref */ |
| 600 | loadConstant(cUnit, r2, (int) classPtr ); |
| 601 | /* When taken r0 has NULL which can be used for store directly */ |
| 602 | ArmLIR* branch1 = genCmpImmBranch(cUnit, kArmCondEq, r0, 0); |
| 603 | /* r1 now contains object->clazz */ |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 604 | assert(OFFSETOF_MEMBER(Object, klass_) == 0); |
| 605 | loadWordDisp(cUnit, r0, OFFSETOF_MEMBER(Object, klass_), r1); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 606 | /* r1 now contains object->clazz */ |
| 607 | loadWordDisp(cUnit, rSELF, |
| 608 | OFFSETOF_MEMBER(Thread, pArtInstanceofNonTrivial), rLR); |
| 609 | loadConstant(cUnit, r0, 1); /* Assume true */ |
| 610 | opRegReg(cUnit, kOpCmp, r1, r2); |
| 611 | ArmLIR* branch2 = opCondBranch(cUnit, kArmCondEq); |
| 612 | genRegCopy(cUnit, r0, r1); |
| 613 | genRegCopy(cUnit, r1, r2); |
| 614 | opReg(cUnit, kOpBlx, rLR); |
| 615 | oatClobberCallRegs(cUnit); |
| 616 | /* branch target here */ |
| 617 | ArmLIR* target = newLIR0(cUnit, kArmPseudoTargetLabel); |
| 618 | target->defMask = ENCODE_ALL; |
| 619 | rlResult = oatGetReturn(cUnit); |
| 620 | storeValue(cUnit, rlDest, rlResult); |
| 621 | branch1->generic.target = (LIR*)target; |
| 622 | branch2->generic.target = (LIR*)target; |
| 623 | } |
| 624 | |
| 625 | static void genCheckCast(CompilationUnit* cUnit, MIR* mir, RegLocation rlSrc) |
| 626 | { |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 627 | Class* classPtr = cUnit->method->GetDeclaringClass()->GetDexCache()-> |
| 628 | GetResolvedClass(mir->dalvikInsn.vB); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 629 | if (classPtr == NULL) { |
| 630 | /* Shouldn't happen with our current model */ |
| 631 | LOG(FATAL) << "Unexpected null class pointer"; |
| 632 | } |
| 633 | oatFlushAllRegs(cUnit); /* Everything to home location */ |
| 634 | loadConstant(cUnit, r1, (int) classPtr ); |
| 635 | rlSrc = loadValue(cUnit, rlSrc, kCoreReg); |
| 636 | /* Null? */ |
| 637 | ArmLIR* branch1 = genCmpImmBranch(cUnit, kArmCondEq, |
| 638 | rlSrc.lowReg, 0); |
| 639 | /* |
| 640 | * rlSrc.lowReg now contains object->clazz. Note that |
| 641 | * it could have been allocated r0, but we're okay so long |
| 642 | * as we don't do anything desctructive until r0 is loaded |
| 643 | * with clazz. |
| 644 | */ |
| 645 | /* r0 now contains object->clazz */ |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 646 | loadWordDisp(cUnit, rlSrc.lowReg, OFFSETOF_MEMBER(Object, klass_), r0); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 647 | loadWordDisp(cUnit, rSELF, |
| 648 | OFFSETOF_MEMBER(Thread, pArtInstanceofNonTrivialNoThrow), rLR); |
| 649 | opRegReg(cUnit, kOpCmp, r0, r1); |
| 650 | ArmLIR* branch2 = opCondBranch(cUnit, kArmCondEq); |
| 651 | // Assume success - if not, artInstanceOfNonTrivial will handle throw |
| 652 | opReg(cUnit, kOpBlx, rLR); |
| 653 | oatClobberCallRegs(cUnit); |
| 654 | ArmLIR* target = newLIR0(cUnit, kArmPseudoTargetLabel); |
| 655 | target->defMask = ENCODE_ALL; |
| 656 | branch1->generic.target = (LIR*)target; |
| 657 | branch2->generic.target = (LIR*)target; |
| 658 | } |
| 659 | |
| 660 | static void genNegFloat(CompilationUnit* cUnit, RegLocation rlDest, |
| 661 | RegLocation rlSrc) |
| 662 | { |
| 663 | RegLocation rlResult; |
| 664 | rlSrc = loadValue(cUnit, rlSrc, kFPReg); |
| 665 | rlResult = oatEvalLoc(cUnit, rlDest, kFPReg, true); |
| 666 | newLIR2(cUnit, kThumb2Vnegs, rlResult.lowReg, rlSrc.lowReg); |
| 667 | storeValue(cUnit, rlDest, rlResult); |
| 668 | } |
| 669 | |
| 670 | static void genNegDouble(CompilationUnit* cUnit, RegLocation rlDest, |
| 671 | RegLocation rlSrc) |
| 672 | { |
| 673 | RegLocation rlResult; |
| 674 | rlSrc = loadValueWide(cUnit, rlSrc, kFPReg); |
| 675 | rlResult = oatEvalLoc(cUnit, rlDest, kFPReg, true); |
| 676 | newLIR2(cUnit, kThumb2Vnegd, S2D(rlResult.lowReg, rlResult.highReg), |
| 677 | S2D(rlSrc.lowReg, rlSrc.highReg)); |
| 678 | storeValueWide(cUnit, rlDest, rlResult); |
| 679 | } |
| 680 | |
| 681 | /* |
| 682 | * To avoid possible conflicts, we use a lot of temps here. Note that |
| 683 | * our usage of Thumb2 instruction forms avoids the problems with register |
| 684 | * reuse for multiply instructions prior to arm6. |
| 685 | */ |
| 686 | static void genMulLong(CompilationUnit* cUnit, RegLocation rlDest, |
| 687 | RegLocation rlSrc1, RegLocation rlSrc2) |
| 688 | { |
| 689 | RegLocation rlResult; |
| 690 | int resLo = oatAllocTemp(cUnit); |
| 691 | int resHi = oatAllocTemp(cUnit); |
| 692 | int tmp1 = oatAllocTemp(cUnit); |
| 693 | |
| 694 | rlSrc1 = loadValueWide(cUnit, rlSrc1, kCoreReg); |
| 695 | rlSrc2 = loadValueWide(cUnit, rlSrc2, kCoreReg); |
| 696 | |
| 697 | newLIR3(cUnit, kThumb2MulRRR, tmp1, rlSrc2.lowReg, rlSrc1.highReg); |
| 698 | newLIR4(cUnit, kThumb2Umull, resLo, resHi, rlSrc2.lowReg, rlSrc1.lowReg); |
| 699 | newLIR4(cUnit, kThumb2Mla, tmp1, rlSrc1.lowReg, rlSrc2.highReg, tmp1); |
| 700 | newLIR4(cUnit, kThumb2AddRRR, resHi, tmp1, resHi, 0); |
| 701 | oatFreeTemp(cUnit, tmp1); |
| 702 | |
| 703 | rlResult = oatGetReturnWide(cUnit); |
| 704 | rlResult.lowReg = resLo; |
| 705 | rlResult.highReg = resHi; |
| 706 | storeValueWide(cUnit, rlDest, rlResult); |
| 707 | } |
| 708 | |
| 709 | static void genLong3Addr(CompilationUnit* cUnit, MIR* mir, OpKind firstOp, |
| 710 | OpKind secondOp, RegLocation rlDest, |
| 711 | RegLocation rlSrc1, RegLocation rlSrc2) |
| 712 | { |
| 713 | RegLocation rlResult; |
| 714 | rlSrc1 = loadValueWide(cUnit, rlSrc1, kCoreReg); |
| 715 | rlSrc2 = loadValueWide(cUnit, rlSrc2, kCoreReg); |
| 716 | rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 717 | opRegRegReg(cUnit, firstOp, rlResult.lowReg, rlSrc1.lowReg, rlSrc2.lowReg); |
| 718 | opRegRegReg(cUnit, secondOp, rlResult.highReg, rlSrc1.highReg, |
| 719 | rlSrc2.highReg); |
| 720 | storeValueWide(cUnit, rlDest, rlResult); |
| 721 | } |
| 722 | |
| 723 | void oatInitializeRegAlloc(CompilationUnit* cUnit) |
| 724 | { |
| 725 | int numRegs = sizeof(coreRegs)/sizeof(*coreRegs); |
| 726 | int numReserved = sizeof(reservedRegs)/sizeof(*reservedRegs); |
| 727 | int numTemps = sizeof(coreTemps)/sizeof(*coreTemps); |
| 728 | int numFPRegs = sizeof(fpRegs)/sizeof(*fpRegs); |
| 729 | int numFPTemps = sizeof(fpTemps)/sizeof(*fpTemps); |
| 730 | RegisterPool *pool = (RegisterPool *)oatNew(sizeof(*pool), true); |
| 731 | cUnit->regPool = pool; |
| 732 | pool->numCoreRegs = numRegs; |
| 733 | pool->coreRegs = (RegisterInfo *) |
| 734 | oatNew(numRegs * sizeof(*cUnit->regPool->coreRegs), true); |
| 735 | pool->numFPRegs = numFPRegs; |
| 736 | pool->FPRegs = (RegisterInfo *) |
| 737 | oatNew(numFPRegs * sizeof(*cUnit->regPool->FPRegs), true); |
| 738 | oatInitPool(pool->coreRegs, coreRegs, pool->numCoreRegs); |
| 739 | oatInitPool(pool->FPRegs, fpRegs, pool->numFPRegs); |
| 740 | // Keep special registers from being allocated |
| 741 | for (int i = 0; i < numReserved; i++) { |
| 742 | oatMarkInUse(cUnit, reservedRegs[i]); |
| 743 | } |
| 744 | // Mark temp regs - all others not in use can be used for promotion |
| 745 | for (int i = 0; i < numTemps; i++) { |
| 746 | oatMarkTemp(cUnit, coreTemps[i]); |
| 747 | } |
| 748 | for (int i = 0; i < numFPTemps; i++) { |
| 749 | oatMarkTemp(cUnit, fpTemps[i]); |
| 750 | } |
| 751 | pool->nullCheckedRegs = |
| 752 | oatAllocBitVector(cUnit->numSSARegs, false); |
| 753 | } |
| 754 | |
| 755 | /* |
| 756 | * Handle simple case (thin lock) inline. If it's complicated, bail |
| 757 | * out to the heavyweight lock/unlock routines. We'll use dedicated |
| 758 | * registers here in order to be in the right position in case we |
| 759 | * to bail to dvm[Lock/Unlock]Object(self, object) |
| 760 | * |
| 761 | * r0 -> self pointer [arg0 for dvm[Lock/Unlock]Object |
| 762 | * r1 -> object [arg1 for dvm[Lock/Unlock]Object |
| 763 | * r2 -> intial contents of object->lock, later result of strex |
| 764 | * r3 -> self->threadId |
| 765 | * r12 -> allow to be used by utilities as general temp |
| 766 | * |
| 767 | * The result of the strex is 0 if we acquire the lock. |
| 768 | * |
| 769 | * See comments in Sync.c for the layout of the lock word. |
| 770 | * Of particular interest to this code is the test for the |
| 771 | * simple case - which we handle inline. For monitor enter, the |
| 772 | * simple case is thin lock, held by no-one. For monitor exit, |
| 773 | * the simple case is thin lock, held by the unlocking thread with |
| 774 | * a recurse count of 0. |
| 775 | * |
| 776 | * A minor complication is that there is a field in the lock word |
| 777 | * unrelated to locking: the hash state. This field must be ignored, but |
| 778 | * preserved. |
| 779 | * |
| 780 | */ |
| 781 | static void genMonitorEnter(CompilationUnit* cUnit, MIR* mir, |
| 782 | RegLocation rlSrc) |
| 783 | { |
| 784 | ArmLIR* target; |
| 785 | ArmLIR* hopTarget; |
| 786 | ArmLIR* branch; |
| 787 | ArmLIR* hopBranch; |
| 788 | |
| 789 | oatFlushAllRegs(cUnit); |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 790 | assert(art::Monitor::kLwShapeThin == 0); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 791 | loadValueDirectFixed(cUnit, rlSrc, r1); // Get obj |
| 792 | oatLockAllTemps(cUnit); // Prepare for explicit register usage |
| 793 | genNullCheck(cUnit, rlSrc.sRegLow, r1, mir->offset, NULL); |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 794 | loadWordDisp(cUnit, rSELF, Thread::IdOffset().Int32Value(), r3); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 795 | newLIR3(cUnit, kThumb2Ldrex, r2, r1, |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 796 | OFFSETOF_MEMBER(Object, monitor_) >> 2); // Get object->lock |
| 797 | // Align owner |
| 798 | opRegImm(cUnit, kOpLsl, r3, art::Monitor::kLwLockOwnerShift); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 799 | // Is lock unheld on lock or held by us (==threadId) on unlock? |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 800 | newLIR4(cUnit, kThumb2Bfi, r3, r2, 0, art::Monitor::kLwLockOwnerShift |
| 801 | - 1); |
| 802 | newLIR3(cUnit, kThumb2Bfc, r2, art::Monitor::kLwHashStateShift, |
| 803 | art::Monitor::kLwLockOwnerShift - 1); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 804 | hopBranch = newLIR2(cUnit, kThumb2Cbnz, r2, 0); |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 805 | newLIR4(cUnit, kThumb2Strex, r2, r3, r1, |
| 806 | OFFSETOF_MEMBER(Object, monitor_) >> 2); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 807 | oatGenMemBarrier(cUnit, kSY); |
| 808 | branch = newLIR2(cUnit, kThumb2Cbz, r2, 0); |
| 809 | |
| 810 | hopTarget = newLIR0(cUnit, kArmPseudoTargetLabel); |
| 811 | hopTarget->defMask = ENCODE_ALL; |
| 812 | hopBranch->generic.target = (LIR*)hopTarget; |
| 813 | |
| 814 | // Go expensive route - artLockObjectNoThrow(self, obj); |
| 815 | loadWordDisp(cUnit, rSELF, OFFSETOF_MEMBER(Thread, pArtLockObjectNoThrow), |
| 816 | rLR); |
| 817 | genRegCopy(cUnit, r0, rSELF); |
| 818 | newLIR1(cUnit, kThumbBlxR, rLR); |
| 819 | |
| 820 | // Resume here |
| 821 | target = newLIR0(cUnit, kArmPseudoTargetLabel); |
| 822 | target->defMask = ENCODE_ALL; |
| 823 | branch->generic.target = (LIR*)target; |
| 824 | } |
| 825 | |
| 826 | /* |
| 827 | * For monitor unlock, we don't have to use ldrex/strex. Once |
| 828 | * we've determined that the lock is thin and that we own it with |
| 829 | * a zero recursion count, it's safe to punch it back to the |
| 830 | * initial, unlock thin state with a store word. |
| 831 | */ |
| 832 | static void genMonitorExit(CompilationUnit* cUnit, MIR* mir, |
| 833 | RegLocation rlSrc) |
| 834 | { |
| 835 | ArmLIR* target; |
| 836 | ArmLIR* branch; |
| 837 | ArmLIR* hopTarget; |
| 838 | ArmLIR* hopBranch; |
| 839 | |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 840 | assert(art::Monitor::kLwShapeThin == 0); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 841 | oatFlushAllRegs(cUnit); |
| 842 | loadValueDirectFixed(cUnit, rlSrc, r1); // Get obj |
| 843 | oatLockAllTemps(cUnit); // Prepare for explicit register usage |
| 844 | genNullCheck(cUnit, rlSrc.sRegLow, r1, mir->offset, NULL); |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 845 | loadWordDisp(cUnit, r1, OFFSETOF_MEMBER(Object, monitor_), r2); // Get lock |
| 846 | loadWordDisp(cUnit, rSELF, Thread::IdOffset().Int32Value(), r3); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 847 | // Is lock unheld on lock or held by us (==threadId) on unlock? |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 848 | opRegRegImm(cUnit, kOpAnd, r12, r2, (art::Monitor::kLwHashStateMask << |
| 849 | art::Monitor::kLwHashStateShift)); |
| 850 | // Align owner |
| 851 | opRegImm(cUnit, kOpLsl, r3, art::Monitor::kLwLockOwnerShift); |
| 852 | newLIR3(cUnit, kThumb2Bfc, r2, art::Monitor::kLwHashStateShift, |
| 853 | art::Monitor::kLwLockOwnerShift - 1); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 854 | opRegReg(cUnit, kOpSub, r2, r3); |
| 855 | hopBranch = opCondBranch(cUnit, kArmCondNe); |
| 856 | oatGenMemBarrier(cUnit, kSY); |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 857 | storeWordDisp(cUnit, r1, OFFSETOF_MEMBER(Object, monitor_), r12); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 858 | branch = opNone(cUnit, kOpUncondBr); |
| 859 | |
| 860 | hopTarget = newLIR0(cUnit, kArmPseudoTargetLabel); |
| 861 | hopTarget->defMask = ENCODE_ALL; |
| 862 | hopBranch->generic.target = (LIR*)hopTarget; |
| 863 | |
| 864 | // Go expensive route - artUnlockObjectNoThrow(self, obj); |
| 865 | loadWordDisp(cUnit, rSELF, OFFSETOF_MEMBER(Thread, pArtUnlockObjectNoThrow), |
| 866 | rLR); |
| 867 | genRegCopy(cUnit, r0, rSELF); |
| 868 | newLIR1(cUnit, kThumbBlxR, rLR); |
| 869 | |
| 870 | // Resume here |
| 871 | target = newLIR0(cUnit, kArmPseudoTargetLabel); |
| 872 | target->defMask = ENCODE_ALL; |
| 873 | branch->generic.target = (LIR*)target; |
| 874 | } |
| 875 | |
| 876 | /* |
| 877 | * 64-bit 3way compare function. |
| 878 | * mov rX, #-1 |
| 879 | * cmp op1hi, op2hi |
| 880 | * blt done |
| 881 | * bgt flip |
| 882 | * sub rX, op1lo, op2lo (treat as unsigned) |
| 883 | * beq done |
| 884 | * ite hi |
| 885 | * mov(hi) rX, #-1 |
| 886 | * mov(!hi) rX, #1 |
| 887 | * flip: |
| 888 | * neg rX |
| 889 | * done: |
| 890 | */ |
| 891 | static void genCmpLong(CompilationUnit* cUnit, MIR* mir, |
| 892 | RegLocation rlDest, RegLocation rlSrc1, |
| 893 | RegLocation rlSrc2) |
| 894 | { |
| 895 | RegLocation rlTemp = LOC_C_RETURN; // Just using as template, will change |
| 896 | ArmLIR* target1; |
| 897 | ArmLIR* target2; |
| 898 | rlSrc1 = loadValueWide(cUnit, rlSrc1, kCoreReg); |
| 899 | rlSrc2 = loadValueWide(cUnit, rlSrc2, kCoreReg); |
| 900 | rlTemp.lowReg = oatAllocTemp(cUnit); |
| 901 | loadConstant(cUnit, rlTemp.lowReg, -1); |
| 902 | opRegReg(cUnit, kOpCmp, rlSrc1.highReg, rlSrc2.highReg); |
| 903 | ArmLIR* branch1 = opCondBranch(cUnit, kArmCondLt); |
| 904 | ArmLIR* branch2 = opCondBranch(cUnit, kArmCondGt); |
| 905 | opRegRegReg(cUnit, kOpSub, rlTemp.lowReg, rlSrc1.lowReg, rlSrc2.lowReg); |
| 906 | ArmLIR* branch3 = opCondBranch(cUnit, kArmCondEq); |
| 907 | |
| 908 | genIT(cUnit, kArmCondHi, "E"); |
| 909 | newLIR2(cUnit, kThumb2MovImmShift, rlTemp.lowReg, modifiedImmediate(-1)); |
| 910 | loadConstant(cUnit, rlTemp.lowReg, 1); |
| 911 | genBarrier(cUnit); |
| 912 | |
| 913 | target2 = newLIR0(cUnit, kArmPseudoTargetLabel); |
| 914 | target2->defMask = -1; |
| 915 | opRegReg(cUnit, kOpNeg, rlTemp.lowReg, rlTemp.lowReg); |
| 916 | |
| 917 | target1 = newLIR0(cUnit, kArmPseudoTargetLabel); |
| 918 | target1->defMask = -1; |
| 919 | |
| 920 | storeValue(cUnit, rlDest, rlTemp); |
| 921 | |
| 922 | branch1->generic.target = (LIR*)target1; |
| 923 | branch2->generic.target = (LIR*)target2; |
| 924 | branch3->generic.target = branch1->generic.target; |
| 925 | } |
| 926 | |
| 927 | static void genMultiplyByTwoBitMultiplier(CompilationUnit* cUnit, |
| 928 | RegLocation rlSrc, RegLocation rlResult, int lit, |
| 929 | int firstBit, int secondBit) |
| 930 | { |
| 931 | opRegRegRegShift(cUnit, kOpAdd, rlResult.lowReg, rlSrc.lowReg, rlSrc.lowReg, |
| 932 | encodeShift(kArmLsl, secondBit - firstBit)); |
| 933 | if (firstBit != 0) { |
| 934 | opRegRegImm(cUnit, kOpLsl, rlResult.lowReg, rlResult.lowReg, firstBit); |
| 935 | } |
| 936 | } |
| 937 | |
| 938 | static bool genConversionCall(CompilationUnit* cUnit, MIR* mir, int funcOffset, |
| 939 | int srcSize, int tgtSize) |
| 940 | { |
| 941 | /* |
| 942 | * Don't optimize the register usage since it calls out to support |
| 943 | * functions |
| 944 | */ |
| 945 | RegLocation rlSrc; |
| 946 | RegLocation rlDest; |
| 947 | oatFlushAllRegs(cUnit); /* Send everything to home location */ |
| 948 | loadWordDisp(cUnit, rSELF, funcOffset, rLR); |
| 949 | if (srcSize == 1) { |
| 950 | rlSrc = oatGetSrc(cUnit, mir, 0); |
| 951 | loadValueDirectFixed(cUnit, rlSrc, r0); |
| 952 | } else { |
| 953 | rlSrc = oatGetSrcWide(cUnit, mir, 0, 1); |
| 954 | loadValueDirectWideFixed(cUnit, rlSrc, r0, r1); |
| 955 | } |
| 956 | opReg(cUnit, kOpBlx, rLR); |
| 957 | oatClobberCallRegs(cUnit); |
| 958 | if (tgtSize == 1) { |
| 959 | RegLocation rlResult; |
| 960 | rlDest = oatGetDest(cUnit, mir, 0); |
| 961 | rlResult = oatGetReturn(cUnit); |
| 962 | storeValue(cUnit, rlDest, rlResult); |
| 963 | } else { |
| 964 | RegLocation rlResult; |
| 965 | rlDest = oatGetDestWide(cUnit, mir, 0, 1); |
| 966 | rlResult = oatGetReturnWide(cUnit); |
| 967 | storeValueWide(cUnit, rlDest, rlResult); |
| 968 | } |
| 969 | return false; |
| 970 | } |
| 971 | |
| 972 | static bool genArithOpFloatPortable(CompilationUnit* cUnit, MIR* mir, |
| 973 | RegLocation rlDest, RegLocation rlSrc1, |
| 974 | RegLocation rlSrc2) |
| 975 | { |
| 976 | RegLocation rlResult; |
| 977 | int funcOffset; |
| 978 | |
| 979 | switch (mir->dalvikInsn.opcode) { |
| 980 | case OP_ADD_FLOAT_2ADDR: |
| 981 | case OP_ADD_FLOAT: |
| 982 | funcOffset = OFFSETOF_MEMBER(Thread, pFadd); |
| 983 | break; |
| 984 | case OP_SUB_FLOAT_2ADDR: |
| 985 | case OP_SUB_FLOAT: |
| 986 | funcOffset = OFFSETOF_MEMBER(Thread, pFsub); |
| 987 | break; |
| 988 | case OP_DIV_FLOAT_2ADDR: |
| 989 | case OP_DIV_FLOAT: |
| 990 | funcOffset = OFFSETOF_MEMBER(Thread, pFdiv); |
| 991 | break; |
| 992 | case OP_MUL_FLOAT_2ADDR: |
| 993 | case OP_MUL_FLOAT: |
| 994 | funcOffset = OFFSETOF_MEMBER(Thread, pFmul); |
| 995 | break; |
| 996 | case OP_REM_FLOAT_2ADDR: |
| 997 | case OP_REM_FLOAT: |
| 998 | funcOffset = OFFSETOF_MEMBER(Thread, pFmodf); |
| 999 | break; |
| 1000 | case OP_NEG_FLOAT: { |
| 1001 | genNegFloat(cUnit, rlDest, rlSrc1); |
| 1002 | return false; |
| 1003 | } |
| 1004 | default: |
| 1005 | return true; |
| 1006 | } |
| 1007 | oatFlushAllRegs(cUnit); /* Send everything to home location */ |
| 1008 | loadWordDisp(cUnit, rSELF, funcOffset, rLR); |
| 1009 | loadValueDirectFixed(cUnit, rlSrc1, r0); |
| 1010 | loadValueDirectFixed(cUnit, rlSrc2, r1); |
| 1011 | opReg(cUnit, kOpBlx, rLR); |
| 1012 | oatClobberCallRegs(cUnit); |
| 1013 | rlResult = oatGetReturn(cUnit); |
| 1014 | storeValue(cUnit, rlDest, rlResult); |
| 1015 | return false; |
| 1016 | } |
| 1017 | |
| 1018 | static bool genArithOpDoublePortable(CompilationUnit* cUnit, MIR* mir, |
| 1019 | RegLocation rlDest, RegLocation rlSrc1, |
| 1020 | RegLocation rlSrc2) |
| 1021 | { |
| 1022 | RegLocation rlResult; |
| 1023 | int funcOffset; |
| 1024 | |
| 1025 | switch (mir->dalvikInsn.opcode) { |
| 1026 | case OP_ADD_DOUBLE_2ADDR: |
| 1027 | case OP_ADD_DOUBLE: |
| 1028 | funcOffset = OFFSETOF_MEMBER(Thread, pDadd); |
| 1029 | break; |
| 1030 | case OP_SUB_DOUBLE_2ADDR: |
| 1031 | case OP_SUB_DOUBLE: |
| 1032 | funcOffset = OFFSETOF_MEMBER(Thread, pDsub); |
| 1033 | break; |
| 1034 | case OP_DIV_DOUBLE_2ADDR: |
| 1035 | case OP_DIV_DOUBLE: |
| 1036 | funcOffset = OFFSETOF_MEMBER(Thread, pDdiv); |
| 1037 | break; |
| 1038 | case OP_MUL_DOUBLE_2ADDR: |
| 1039 | case OP_MUL_DOUBLE: |
| 1040 | funcOffset = OFFSETOF_MEMBER(Thread, pDmul); |
| 1041 | break; |
| 1042 | case OP_REM_DOUBLE_2ADDR: |
| 1043 | case OP_REM_DOUBLE: |
| 1044 | funcOffset = OFFSETOF_MEMBER(Thread, pFmod); |
| 1045 | break; |
| 1046 | case OP_NEG_DOUBLE: { |
| 1047 | genNegDouble(cUnit, rlDest, rlSrc1); |
| 1048 | return false; |
| 1049 | } |
| 1050 | default: |
| 1051 | return true; |
| 1052 | } |
| 1053 | oatFlushAllRegs(cUnit); /* Send everything to home location */ |
| 1054 | loadWordDisp(cUnit, rSELF, funcOffset, rLR); |
| 1055 | loadValueDirectWideFixed(cUnit, rlSrc1, r0, r1); |
| 1056 | loadValueDirectWideFixed(cUnit, rlSrc2, r2, r3); |
| 1057 | opReg(cUnit, kOpBlx, rLR); |
| 1058 | oatClobberCallRegs(cUnit); |
| 1059 | rlResult = oatGetReturnWide(cUnit); |
| 1060 | storeValueWide(cUnit, rlDest, rlResult); |
| 1061 | return false; |
| 1062 | } |
| 1063 | |
| 1064 | static bool genConversionPortable(CompilationUnit* cUnit, MIR* mir) |
| 1065 | { |
| 1066 | Opcode opcode = mir->dalvikInsn.opcode; |
| 1067 | |
| 1068 | switch (opcode) { |
| 1069 | case OP_INT_TO_FLOAT: |
| 1070 | return genConversionCall(cUnit, mir, OFFSETOF_MEMBER(Thread, pI2f), |
| 1071 | 1, 1); |
| 1072 | case OP_FLOAT_TO_INT: |
| 1073 | return genConversionCall(cUnit, mir, OFFSETOF_MEMBER(Thread, pF2iz), |
| 1074 | 1, 1); |
| 1075 | case OP_DOUBLE_TO_FLOAT: |
| 1076 | return genConversionCall(cUnit, mir, OFFSETOF_MEMBER(Thread, pD2f), |
| 1077 | 2, 1); |
| 1078 | case OP_FLOAT_TO_DOUBLE: |
| 1079 | return genConversionCall(cUnit, mir, OFFSETOF_MEMBER(Thread, pF2d), |
| 1080 | 1, 2); |
| 1081 | case OP_INT_TO_DOUBLE: |
| 1082 | return genConversionCall(cUnit, mir, OFFSETOF_MEMBER(Thread, pI2d), |
| 1083 | 1, 2); |
| 1084 | case OP_DOUBLE_TO_INT: |
| 1085 | return genConversionCall(cUnit, mir, OFFSETOF_MEMBER(Thread, pD2iz), |
| 1086 | 2, 1); |
| 1087 | case OP_FLOAT_TO_LONG: |
| 1088 | return genConversionCall(cUnit, mir, OFFSETOF_MEMBER(Thread, |
| 1089 | pArtF2l), 1, 2); |
| 1090 | case OP_LONG_TO_FLOAT: |
| 1091 | return genConversionCall(cUnit, mir, OFFSETOF_MEMBER(Thread, pL2f), |
| 1092 | 2, 1); |
| 1093 | case OP_DOUBLE_TO_LONG: |
| 1094 | return genConversionCall(cUnit, mir, OFFSETOF_MEMBER(Thread, |
| 1095 | pArtD2l), 2, 2); |
| 1096 | case OP_LONG_TO_DOUBLE: |
| 1097 | return genConversionCall(cUnit, mir, OFFSETOF_MEMBER(Thread, pL2d), |
| 1098 | 2, 2); |
| 1099 | default: |
| 1100 | return true; |
| 1101 | } |
| 1102 | return false; |
| 1103 | } |
| 1104 | |
| 1105 | /* Generate conditional branch instructions */ |
| 1106 | static ArmLIR* genConditionalBranch(CompilationUnit* cUnit, |
| 1107 | ArmConditionCode cond, |
| 1108 | ArmLIR* target) |
| 1109 | { |
| 1110 | ArmLIR* branch = opCondBranch(cUnit, cond); |
| 1111 | branch->generic.target = (LIR*) target; |
| 1112 | return branch; |
| 1113 | } |
| 1114 | |
| 1115 | /* Generate a unconditional branch to go to the interpreter */ |
| 1116 | static inline ArmLIR* genTrap(CompilationUnit* cUnit, int dOffset, |
| 1117 | ArmLIR* pcrLabel) |
| 1118 | { |
| 1119 | ArmLIR* branch = opNone(cUnit, kOpUncondBr); |
| 1120 | return genCheckCommon(cUnit, dOffset, branch, pcrLabel); |
| 1121 | } |
| 1122 | |
| 1123 | /* |
| 1124 | * Generate array store |
| 1125 | * |
| 1126 | */ |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 1127 | static void genArrayPut(CompilationUnit* cUnit, MIR* mir, |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 1128 | RegLocation rlArray, RegLocation rlIndex, |
| 1129 | RegLocation rlSrc, int scale) |
| 1130 | { |
| 1131 | RegisterClass regClass = oatRegClassBySize(kWord); |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 1132 | int lenOffset = Array::LengthOffset().Int32Value(); |
| 1133 | int dataOffset = Array::DataOffset().Int32Value(); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 1134 | |
| 1135 | /* Make sure it's a legal object Put. Use direct regs at first */ |
| 1136 | loadValueDirectFixed(cUnit, rlArray, r1); |
| 1137 | loadValueDirectFixed(cUnit, rlSrc, r0); |
| 1138 | |
| 1139 | /* null array object? */ |
| 1140 | ArmLIR* pcrLabel = NULL; |
| 1141 | |
| 1142 | if (!(mir->OptimizationFlags & MIR_IGNORE_NULL_CHECK)) { |
| 1143 | pcrLabel = genNullCheck(cUnit, rlArray.sRegLow, r1, |
| 1144 | mir->offset, NULL); |
| 1145 | } |
| 1146 | loadWordDisp(cUnit, rSELF, |
| 1147 | OFFSETOF_MEMBER(Thread, pArtCanPutArrayElementNoThrow), rLR); |
| 1148 | /* Get the array's clazz */ |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 1149 | loadWordDisp(cUnit, r1, OFFSETOF_MEMBER(Object, klass_), r1); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 1150 | /* Get the object's clazz */ |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 1151 | loadWordDisp(cUnit, r0, OFFSETOF_MEMBER(Object, klass_), r0); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 1152 | opReg(cUnit, kOpBlx, rLR); |
| 1153 | oatClobberCallRegs(cUnit); |
| 1154 | |
| 1155 | // Now, redo loadValues in case they didn't survive the call |
| 1156 | |
| 1157 | int regPtr; |
| 1158 | rlArray = loadValue(cUnit, rlArray, kCoreReg); |
| 1159 | rlIndex = loadValue(cUnit, rlIndex, kCoreReg); |
| 1160 | |
| 1161 | if (oatIsTemp(cUnit, rlArray.lowReg)) { |
| 1162 | oatClobber(cUnit, rlArray.lowReg); |
| 1163 | regPtr = rlArray.lowReg; |
| 1164 | } else { |
| 1165 | regPtr = oatAllocTemp(cUnit); |
| 1166 | genRegCopy(cUnit, regPtr, rlArray.lowReg); |
| 1167 | } |
| 1168 | |
| 1169 | if (!(mir->OptimizationFlags & MIR_IGNORE_RANGE_CHECK)) { |
| 1170 | int regLen = oatAllocTemp(cUnit); |
| 1171 | //NOTE: max live temps(4) here. |
| 1172 | /* Get len */ |
| 1173 | loadWordDisp(cUnit, rlArray.lowReg, lenOffset, regLen); |
| 1174 | /* regPtr -> array data */ |
| 1175 | opRegImm(cUnit, kOpAdd, regPtr, dataOffset); |
| 1176 | genBoundsCheck(cUnit, rlIndex.lowReg, regLen, mir->offset, |
| 1177 | pcrLabel); |
| 1178 | oatFreeTemp(cUnit, regLen); |
| 1179 | } else { |
| 1180 | /* regPtr -> array data */ |
| 1181 | opRegImm(cUnit, kOpAdd, regPtr, dataOffset); |
| 1182 | } |
| 1183 | /* at this point, regPtr points to array, 2 live temps */ |
| 1184 | rlSrc = loadValue(cUnit, rlSrc, regClass); |
| 1185 | storeBaseIndexed(cUnit, regPtr, rlIndex.lowReg, rlSrc.lowReg, |
| 1186 | scale, kWord); |
| 1187 | } |
| 1188 | |
| 1189 | /* |
| 1190 | * Generate array load |
| 1191 | */ |
| 1192 | static void genArrayGet(CompilationUnit* cUnit, MIR* mir, OpSize size, |
| 1193 | RegLocation rlArray, RegLocation rlIndex, |
| 1194 | RegLocation rlDest, int scale) |
| 1195 | { |
| 1196 | RegisterClass regClass = oatRegClassBySize(size); |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 1197 | int lenOffset = Array::LengthOffset().Int32Value(); |
| 1198 | int dataOffset = Array::DataOffset().Int32Value(); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 1199 | RegLocation rlResult; |
| 1200 | rlArray = loadValue(cUnit, rlArray, kCoreReg); |
| 1201 | rlIndex = loadValue(cUnit, rlIndex, kCoreReg); |
| 1202 | int regPtr; |
| 1203 | |
| 1204 | /* null object? */ |
| 1205 | ArmLIR* pcrLabel = NULL; |
| 1206 | |
| 1207 | if (!(mir->OptimizationFlags & MIR_IGNORE_NULL_CHECK)) { |
| 1208 | pcrLabel = genNullCheck(cUnit, rlArray.sRegLow, |
| 1209 | rlArray.lowReg, mir->offset, NULL); |
| 1210 | } |
| 1211 | |
| 1212 | regPtr = oatAllocTemp(cUnit); |
| 1213 | |
| 1214 | if (!(mir->OptimizationFlags & MIR_IGNORE_RANGE_CHECK)) { |
| 1215 | int regLen = oatAllocTemp(cUnit); |
| 1216 | /* Get len */ |
| 1217 | loadWordDisp(cUnit, rlArray.lowReg, lenOffset, regLen); |
| 1218 | /* regPtr -> array data */ |
| 1219 | opRegRegImm(cUnit, kOpAdd, regPtr, rlArray.lowReg, dataOffset); |
| 1220 | genBoundsCheck(cUnit, rlIndex.lowReg, regLen, mir->offset, |
| 1221 | pcrLabel); |
| 1222 | oatFreeTemp(cUnit, regLen); |
| 1223 | } else { |
| 1224 | /* regPtr -> array data */ |
| 1225 | opRegRegImm(cUnit, kOpAdd, regPtr, rlArray.lowReg, dataOffset); |
| 1226 | } |
| 1227 | if ((size == kLong) || (size == kDouble)) { |
| 1228 | if (scale) { |
| 1229 | int rNewIndex = oatAllocTemp(cUnit); |
| 1230 | opRegRegImm(cUnit, kOpLsl, rNewIndex, rlIndex.lowReg, scale); |
| 1231 | opRegReg(cUnit, kOpAdd, regPtr, rNewIndex); |
| 1232 | oatFreeTemp(cUnit, rNewIndex); |
| 1233 | } else { |
| 1234 | opRegReg(cUnit, kOpAdd, regPtr, rlIndex.lowReg); |
| 1235 | } |
| 1236 | rlResult = oatEvalLoc(cUnit, rlDest, regClass, true); |
| 1237 | |
| 1238 | loadPair(cUnit, regPtr, rlResult.lowReg, rlResult.highReg); |
| 1239 | |
| 1240 | oatFreeTemp(cUnit, regPtr); |
| 1241 | storeValueWide(cUnit, rlDest, rlResult); |
| 1242 | } else { |
| 1243 | rlResult = oatEvalLoc(cUnit, rlDest, regClass, true); |
| 1244 | |
| 1245 | loadBaseIndexed(cUnit, regPtr, rlIndex.lowReg, rlResult.lowReg, |
| 1246 | scale, size); |
| 1247 | |
| 1248 | oatFreeTemp(cUnit, regPtr); |
| 1249 | storeValue(cUnit, rlDest, rlResult); |
| 1250 | } |
| 1251 | } |
| 1252 | |
| 1253 | /* |
| 1254 | * Generate array store |
| 1255 | * |
| 1256 | */ |
| 1257 | static void genArrayPut(CompilationUnit* cUnit, MIR* mir, OpSize size, |
| 1258 | RegLocation rlArray, RegLocation rlIndex, |
| 1259 | RegLocation rlSrc, int scale) |
| 1260 | { |
| 1261 | RegisterClass regClass = oatRegClassBySize(size); |
buzbee | c143c55 | 2011-08-20 17:38:58 -0700 | [diff] [blame^] | 1262 | int lenOffset = Array::LengthOffset().Int32Value(); |
| 1263 | int dataOffset = Array::DataOffset().Int32Value(); |
buzbee | 67bf885 | 2011-08-17 17:51:35 -0700 | [diff] [blame] | 1264 | |
| 1265 | int regPtr; |
| 1266 | rlArray = loadValue(cUnit, rlArray, kCoreReg); |
| 1267 | rlIndex = loadValue(cUnit, rlIndex, kCoreReg); |
| 1268 | |
| 1269 | if (oatIsTemp(cUnit, rlArray.lowReg)) { |
| 1270 | oatClobber(cUnit, rlArray.lowReg); |
| 1271 | regPtr = rlArray.lowReg; |
| 1272 | } else { |
| 1273 | regPtr = oatAllocTemp(cUnit); |
| 1274 | genRegCopy(cUnit, regPtr, rlArray.lowReg); |
| 1275 | } |
| 1276 | |
| 1277 | /* null object? */ |
| 1278 | ArmLIR* pcrLabel = NULL; |
| 1279 | |
| 1280 | if (!(mir->OptimizationFlags & MIR_IGNORE_NULL_CHECK)) { |
| 1281 | pcrLabel = genNullCheck(cUnit, rlArray.sRegLow, rlArray.lowReg, |
| 1282 | mir->offset, NULL); |
| 1283 | } |
| 1284 | |
| 1285 | if (!(mir->OptimizationFlags & MIR_IGNORE_RANGE_CHECK)) { |
| 1286 | int regLen = oatAllocTemp(cUnit); |
| 1287 | //NOTE: max live temps(4) here. |
| 1288 | /* Get len */ |
| 1289 | loadWordDisp(cUnit, rlArray.lowReg, lenOffset, regLen); |
| 1290 | /* regPtr -> array data */ |
| 1291 | opRegImm(cUnit, kOpAdd, regPtr, dataOffset); |
| 1292 | genBoundsCheck(cUnit, rlIndex.lowReg, regLen, mir->offset, |
| 1293 | pcrLabel); |
| 1294 | oatFreeTemp(cUnit, regLen); |
| 1295 | } else { |
| 1296 | /* regPtr -> array data */ |
| 1297 | opRegImm(cUnit, kOpAdd, regPtr, dataOffset); |
| 1298 | } |
| 1299 | /* at this point, regPtr points to array, 2 live temps */ |
| 1300 | if ((size == kLong) || (size == kDouble)) { |
| 1301 | //TODO: need specific wide routine that can handle fp regs |
| 1302 | if (scale) { |
| 1303 | int rNewIndex = oatAllocTemp(cUnit); |
| 1304 | opRegRegImm(cUnit, kOpLsl, rNewIndex, rlIndex.lowReg, scale); |
| 1305 | opRegReg(cUnit, kOpAdd, regPtr, rNewIndex); |
| 1306 | oatFreeTemp(cUnit, rNewIndex); |
| 1307 | } else { |
| 1308 | opRegReg(cUnit, kOpAdd, regPtr, rlIndex.lowReg); |
| 1309 | } |
| 1310 | rlSrc = loadValueWide(cUnit, rlSrc, regClass); |
| 1311 | |
| 1312 | storePair(cUnit, regPtr, rlSrc.lowReg, rlSrc.highReg); |
| 1313 | |
| 1314 | oatFreeTemp(cUnit, regPtr); |
| 1315 | } else { |
| 1316 | rlSrc = loadValue(cUnit, rlSrc, regClass); |
| 1317 | |
| 1318 | storeBaseIndexed(cUnit, regPtr, rlIndex.lowReg, rlSrc.lowReg, |
| 1319 | scale, size); |
| 1320 | } |
| 1321 | } |
| 1322 | |
| 1323 | static bool genShiftOpLong(CompilationUnit* cUnit, MIR* mir, |
| 1324 | RegLocation rlDest, RegLocation rlSrc1, |
| 1325 | RegLocation rlShift) |
| 1326 | { |
| 1327 | /* |
| 1328 | * Don't mess with the regsiters here as there is a particular calling |
| 1329 | * convention to the out-of-line handler. |
| 1330 | */ |
| 1331 | RegLocation rlResult; |
| 1332 | |
| 1333 | loadValueDirectWideFixed(cUnit, rlSrc1, r0, r1); |
| 1334 | loadValueDirect(cUnit, rlShift, r2); |
| 1335 | switch( mir->dalvikInsn.opcode) { |
| 1336 | case OP_SHL_LONG: |
| 1337 | case OP_SHL_LONG_2ADDR: |
| 1338 | //genDispatchToHandler(cUnit, TEMPLATE_SHL_LONG); |
| 1339 | assert(0); // unimp |
| 1340 | break; |
| 1341 | case OP_SHR_LONG: |
| 1342 | case OP_SHR_LONG_2ADDR: |
| 1343 | //genDispatchToHandler(cUnit, TEMPLATE_SHR_LONG); |
| 1344 | assert(0); // unimp |
| 1345 | break; |
| 1346 | case OP_USHR_LONG: |
| 1347 | case OP_USHR_LONG_2ADDR: |
| 1348 | //genDispatchToHandler(cUnit, TEMPLATE_USHR_LONG); |
| 1349 | assert(0); // unimp |
| 1350 | break; |
| 1351 | default: |
| 1352 | return true; |
| 1353 | } |
| 1354 | rlResult = oatGetReturnWide(cUnit); |
| 1355 | storeValueWide(cUnit, rlDest, rlResult); |
| 1356 | return false; |
| 1357 | } |
| 1358 | |
| 1359 | static bool genArithOpLong(CompilationUnit* cUnit, MIR* mir, |
| 1360 | RegLocation rlDest, RegLocation rlSrc1, |
| 1361 | RegLocation rlSrc2) |
| 1362 | { |
| 1363 | RegLocation rlResult; |
| 1364 | OpKind firstOp = kOpBkpt; |
| 1365 | OpKind secondOp = kOpBkpt; |
| 1366 | bool callOut = false; |
| 1367 | int funcOffset; |
| 1368 | int retReg = r0; |
| 1369 | |
| 1370 | switch (mir->dalvikInsn.opcode) { |
| 1371 | case OP_NOT_LONG: |
| 1372 | rlSrc2 = loadValueWide(cUnit, rlSrc2, kCoreReg); |
| 1373 | rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 1374 | opRegReg(cUnit, kOpMvn, rlResult.lowReg, rlSrc2.lowReg); |
| 1375 | opRegReg(cUnit, kOpMvn, rlResult.highReg, rlSrc2.highReg); |
| 1376 | storeValueWide(cUnit, rlDest, rlResult); |
| 1377 | return false; |
| 1378 | break; |
| 1379 | case OP_ADD_LONG: |
| 1380 | case OP_ADD_LONG_2ADDR: |
| 1381 | firstOp = kOpAdd; |
| 1382 | secondOp = kOpAdc; |
| 1383 | break; |
| 1384 | case OP_SUB_LONG: |
| 1385 | case OP_SUB_LONG_2ADDR: |
| 1386 | firstOp = kOpSub; |
| 1387 | secondOp = kOpSbc; |
| 1388 | break; |
| 1389 | case OP_MUL_LONG: |
| 1390 | case OP_MUL_LONG_2ADDR: |
| 1391 | genMulLong(cUnit, rlDest, rlSrc1, rlSrc2); |
| 1392 | return false; |
| 1393 | case OP_DIV_LONG: |
| 1394 | case OP_DIV_LONG_2ADDR: |
| 1395 | callOut = true; |
| 1396 | retReg = r0; |
| 1397 | funcOffset = OFFSETOF_MEMBER(Thread, pLdivmod); |
| 1398 | break; |
| 1399 | /* NOTE - result is in r2/r3 instead of r0/r1 */ |
| 1400 | case OP_REM_LONG: |
| 1401 | case OP_REM_LONG_2ADDR: |
| 1402 | callOut = true; |
| 1403 | funcOffset = OFFSETOF_MEMBER(Thread, pLdivmod); |
| 1404 | retReg = r2; |
| 1405 | break; |
| 1406 | case OP_AND_LONG_2ADDR: |
| 1407 | case OP_AND_LONG: |
| 1408 | firstOp = kOpAnd; |
| 1409 | secondOp = kOpAnd; |
| 1410 | break; |
| 1411 | case OP_OR_LONG: |
| 1412 | case OP_OR_LONG_2ADDR: |
| 1413 | firstOp = kOpOr; |
| 1414 | secondOp = kOpOr; |
| 1415 | break; |
| 1416 | case OP_XOR_LONG: |
| 1417 | case OP_XOR_LONG_2ADDR: |
| 1418 | firstOp = kOpXor; |
| 1419 | secondOp = kOpXor; |
| 1420 | break; |
| 1421 | case OP_NEG_LONG: { |
| 1422 | //TUNING: can improve this using Thumb2 code |
| 1423 | int tReg = oatAllocTemp(cUnit); |
| 1424 | rlSrc2 = loadValueWide(cUnit, rlSrc2, kCoreReg); |
| 1425 | rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 1426 | loadConstantNoClobber(cUnit, tReg, 0); |
| 1427 | opRegRegReg(cUnit, kOpSub, rlResult.lowReg, |
| 1428 | tReg, rlSrc2.lowReg); |
| 1429 | opRegReg(cUnit, kOpSbc, tReg, rlSrc2.highReg); |
| 1430 | genRegCopy(cUnit, rlResult.highReg, tReg); |
| 1431 | storeValueWide(cUnit, rlDest, rlResult); |
| 1432 | return false; |
| 1433 | } |
| 1434 | default: |
| 1435 | LOG(FATAL) << "Invalid long arith op"; |
| 1436 | } |
| 1437 | if (!callOut) { |
| 1438 | genLong3Addr(cUnit, mir, firstOp, secondOp, rlDest, rlSrc1, rlSrc2); |
| 1439 | } else { |
| 1440 | // Adjust return regs in to handle case of rem returning r2/r3 |
| 1441 | oatFlushAllRegs(cUnit); /* Send everything to home location */ |
| 1442 | loadWordDisp(cUnit, rSELF, funcOffset, rLR); |
| 1443 | loadValueDirectWideFixed(cUnit, rlSrc1, r0, r1); |
| 1444 | loadValueDirectWideFixed(cUnit, rlSrc2, r2, r3); |
| 1445 | opReg(cUnit, kOpBlx, rLR); |
| 1446 | oatClobberCallRegs(cUnit); |
| 1447 | if (retReg == r0) |
| 1448 | rlResult = oatGetReturnWide(cUnit); |
| 1449 | else |
| 1450 | rlResult = oatGetReturnWideAlt(cUnit); |
| 1451 | storeValueWide(cUnit, rlDest, rlResult); |
| 1452 | } |
| 1453 | return false; |
| 1454 | } |
| 1455 | |
| 1456 | static bool genArithOpInt(CompilationUnit* cUnit, MIR* mir, |
| 1457 | RegLocation rlDest, RegLocation rlSrc1, |
| 1458 | RegLocation rlSrc2) |
| 1459 | { |
| 1460 | OpKind op = kOpBkpt; |
| 1461 | bool callOut = false; |
| 1462 | bool checkZero = false; |
| 1463 | bool unary = false; |
| 1464 | int retReg = r0; |
| 1465 | int funcOffset; |
| 1466 | RegLocation rlResult; |
| 1467 | bool shiftOp = false; |
| 1468 | |
| 1469 | switch (mir->dalvikInsn.opcode) { |
| 1470 | case OP_NEG_INT: |
| 1471 | op = kOpNeg; |
| 1472 | unary = true; |
| 1473 | break; |
| 1474 | case OP_NOT_INT: |
| 1475 | op = kOpMvn; |
| 1476 | unary = true; |
| 1477 | break; |
| 1478 | case OP_ADD_INT: |
| 1479 | case OP_ADD_INT_2ADDR: |
| 1480 | op = kOpAdd; |
| 1481 | break; |
| 1482 | case OP_SUB_INT: |
| 1483 | case OP_SUB_INT_2ADDR: |
| 1484 | op = kOpSub; |
| 1485 | break; |
| 1486 | case OP_MUL_INT: |
| 1487 | case OP_MUL_INT_2ADDR: |
| 1488 | op = kOpMul; |
| 1489 | break; |
| 1490 | case OP_DIV_INT: |
| 1491 | case OP_DIV_INT_2ADDR: |
| 1492 | callOut = true; |
| 1493 | checkZero = true; |
| 1494 | funcOffset = OFFSETOF_MEMBER(Thread, pIdiv); |
| 1495 | retReg = r0; |
| 1496 | break; |
| 1497 | /* NOTE: returns in r1 */ |
| 1498 | case OP_REM_INT: |
| 1499 | case OP_REM_INT_2ADDR: |
| 1500 | callOut = true; |
| 1501 | checkZero = true; |
| 1502 | funcOffset = OFFSETOF_MEMBER(Thread, pIdivmod); |
| 1503 | retReg = r1; |
| 1504 | break; |
| 1505 | case OP_AND_INT: |
| 1506 | case OP_AND_INT_2ADDR: |
| 1507 | op = kOpAnd; |
| 1508 | break; |
| 1509 | case OP_OR_INT: |
| 1510 | case OP_OR_INT_2ADDR: |
| 1511 | op = kOpOr; |
| 1512 | break; |
| 1513 | case OP_XOR_INT: |
| 1514 | case OP_XOR_INT_2ADDR: |
| 1515 | op = kOpXor; |
| 1516 | break; |
| 1517 | case OP_SHL_INT: |
| 1518 | case OP_SHL_INT_2ADDR: |
| 1519 | shiftOp = true; |
| 1520 | op = kOpLsl; |
| 1521 | break; |
| 1522 | case OP_SHR_INT: |
| 1523 | case OP_SHR_INT_2ADDR: |
| 1524 | shiftOp = true; |
| 1525 | op = kOpAsr; |
| 1526 | break; |
| 1527 | case OP_USHR_INT: |
| 1528 | case OP_USHR_INT_2ADDR: |
| 1529 | shiftOp = true; |
| 1530 | op = kOpLsr; |
| 1531 | break; |
| 1532 | default: |
| 1533 | LOG(FATAL) << "Invalid word arith op: " << |
| 1534 | (int)mir->dalvikInsn.opcode; |
| 1535 | } |
| 1536 | if (!callOut) { |
| 1537 | rlSrc1 = loadValue(cUnit, rlSrc1, kCoreReg); |
| 1538 | if (unary) { |
| 1539 | rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 1540 | opRegReg(cUnit, op, rlResult.lowReg, |
| 1541 | rlSrc1.lowReg); |
| 1542 | } else { |
| 1543 | rlSrc2 = loadValue(cUnit, rlSrc2, kCoreReg); |
| 1544 | if (shiftOp) { |
| 1545 | int tReg = oatAllocTemp(cUnit); |
| 1546 | opRegRegImm(cUnit, kOpAnd, tReg, rlSrc2.lowReg, 31); |
| 1547 | rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 1548 | opRegRegReg(cUnit, op, rlResult.lowReg, |
| 1549 | rlSrc1.lowReg, tReg); |
| 1550 | oatFreeTemp(cUnit, tReg); |
| 1551 | } else { |
| 1552 | rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 1553 | opRegRegReg(cUnit, op, rlResult.lowReg, |
| 1554 | rlSrc1.lowReg, rlSrc2.lowReg); |
| 1555 | } |
| 1556 | } |
| 1557 | storeValue(cUnit, rlDest, rlResult); |
| 1558 | } else { |
| 1559 | RegLocation rlResult; |
| 1560 | oatFlushAllRegs(cUnit); /* Send everything to home location */ |
| 1561 | loadValueDirectFixed(cUnit, rlSrc2, r1); |
| 1562 | loadWordDisp(cUnit, rSELF, funcOffset, rLR); |
| 1563 | loadValueDirectFixed(cUnit, rlSrc1, r0); |
| 1564 | if (checkZero) { |
| 1565 | genNullCheck(cUnit, rlSrc2.sRegLow, r1, mir->offset, NULL); |
| 1566 | } |
| 1567 | opReg(cUnit, kOpBlx, rLR); |
| 1568 | oatClobberCallRegs(cUnit); |
| 1569 | if (retReg == r0) |
| 1570 | rlResult = oatGetReturn(cUnit); |
| 1571 | else |
| 1572 | rlResult = oatGetReturnAlt(cUnit); |
| 1573 | storeValue(cUnit, rlDest, rlResult); |
| 1574 | } |
| 1575 | return false; |
| 1576 | } |
| 1577 | |
| 1578 | /* Generate unconditional branch instructions */ |
| 1579 | static ArmLIR* genUnconditionalBranch(CompilationUnit* cUnit, ArmLIR* target) |
| 1580 | { |
| 1581 | ArmLIR* branch = opNone(cUnit, kOpUncondBr); |
| 1582 | branch->generic.target = (LIR*) target; |
| 1583 | return branch; |
| 1584 | } |
| 1585 | |
| 1586 | /* |
| 1587 | * Fetch *self->info.breakFlags. If the breakFlags are non-zero, |
| 1588 | * punt to the interpreter. |
| 1589 | */ |
| 1590 | static void genSuspendPoll(CompilationUnit* cUnit, MIR* mir) |
| 1591 | { |
| 1592 | UNIMPLEMENTED(WARNING); |
| 1593 | #if 0 |
| 1594 | int rTemp = oatAllocTemp(cUnit); |
| 1595 | ArmLIR* ld; |
| 1596 | ld = loadBaseDisp(cUnit, NULL, rSELF, |
| 1597 | offsetof(Thread, interpBreak.ctl.breakFlags), |
| 1598 | rTemp, kUnsignedByte, INVALID_SREG); |
| 1599 | setMemRefType(ld, true /* isLoad */, kMustNotAlias); |
| 1600 | genRegImmCheck(cUnit, kArmCondNe, rTemp, 0, mir->offset, NULL); |
| 1601 | #endif |
| 1602 | } |
| 1603 | |
| 1604 | /* |
| 1605 | * The following are the first-level codegen routines that analyze the format |
| 1606 | * of each bytecode then either dispatch special purpose codegen routines |
| 1607 | * or produce corresponding Thumb instructions directly. |
| 1608 | */ |
| 1609 | |
| 1610 | static bool isPowerOfTwo(int x) |
| 1611 | { |
| 1612 | return (x & (x - 1)) == 0; |
| 1613 | } |
| 1614 | |
| 1615 | // Returns true if no more than two bits are set in 'x'. |
| 1616 | static bool isPopCountLE2(unsigned int x) |
| 1617 | { |
| 1618 | x &= x - 1; |
| 1619 | return (x & (x - 1)) == 0; |
| 1620 | } |
| 1621 | |
| 1622 | // Returns the index of the lowest set bit in 'x'. |
| 1623 | static int lowestSetBit(unsigned int x) { |
| 1624 | int bit_posn = 0; |
| 1625 | while ((x & 0xf) == 0) { |
| 1626 | bit_posn += 4; |
| 1627 | x >>= 4; |
| 1628 | } |
| 1629 | while ((x & 1) == 0) { |
| 1630 | bit_posn++; |
| 1631 | x >>= 1; |
| 1632 | } |
| 1633 | return bit_posn; |
| 1634 | } |
| 1635 | |
| 1636 | // Returns true if it added instructions to 'cUnit' to divide 'rlSrc' by 'lit' |
| 1637 | // and store the result in 'rlDest'. |
| 1638 | static bool handleEasyDivide(CompilationUnit* cUnit, Opcode dalvikOpcode, |
| 1639 | RegLocation rlSrc, RegLocation rlDest, int lit) |
| 1640 | { |
| 1641 | if (lit < 2 || !isPowerOfTwo(lit)) { |
| 1642 | return false; |
| 1643 | } |
| 1644 | int k = lowestSetBit(lit); |
| 1645 | if (k >= 30) { |
| 1646 | // Avoid special cases. |
| 1647 | return false; |
| 1648 | } |
| 1649 | bool div = (dalvikOpcode == OP_DIV_INT_LIT8 || |
| 1650 | dalvikOpcode == OP_DIV_INT_LIT16); |
| 1651 | rlSrc = loadValue(cUnit, rlSrc, kCoreReg); |
| 1652 | RegLocation rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 1653 | if (div) { |
| 1654 | int tReg = oatAllocTemp(cUnit); |
| 1655 | if (lit == 2) { |
| 1656 | // Division by 2 is by far the most common division by constant. |
| 1657 | opRegRegImm(cUnit, kOpLsr, tReg, rlSrc.lowReg, 32 - k); |
| 1658 | opRegRegReg(cUnit, kOpAdd, tReg, tReg, rlSrc.lowReg); |
| 1659 | opRegRegImm(cUnit, kOpAsr, rlResult.lowReg, tReg, k); |
| 1660 | } else { |
| 1661 | opRegRegImm(cUnit, kOpAsr, tReg, rlSrc.lowReg, 31); |
| 1662 | opRegRegImm(cUnit, kOpLsr, tReg, tReg, 32 - k); |
| 1663 | opRegRegReg(cUnit, kOpAdd, tReg, tReg, rlSrc.lowReg); |
| 1664 | opRegRegImm(cUnit, kOpAsr, rlResult.lowReg, tReg, k); |
| 1665 | } |
| 1666 | } else { |
| 1667 | int cReg = oatAllocTemp(cUnit); |
| 1668 | loadConstant(cUnit, cReg, lit - 1); |
| 1669 | int tReg1 = oatAllocTemp(cUnit); |
| 1670 | int tReg2 = oatAllocTemp(cUnit); |
| 1671 | if (lit == 2) { |
| 1672 | opRegRegImm(cUnit, kOpLsr, tReg1, rlSrc.lowReg, 32 - k); |
| 1673 | opRegRegReg(cUnit, kOpAdd, tReg2, tReg1, rlSrc.lowReg); |
| 1674 | opRegRegReg(cUnit, kOpAnd, tReg2, tReg2, cReg); |
| 1675 | opRegRegReg(cUnit, kOpSub, rlResult.lowReg, tReg2, tReg1); |
| 1676 | } else { |
| 1677 | opRegRegImm(cUnit, kOpAsr, tReg1, rlSrc.lowReg, 31); |
| 1678 | opRegRegImm(cUnit, kOpLsr, tReg1, tReg1, 32 - k); |
| 1679 | opRegRegReg(cUnit, kOpAdd, tReg2, tReg1, rlSrc.lowReg); |
| 1680 | opRegRegReg(cUnit, kOpAnd, tReg2, tReg2, cReg); |
| 1681 | opRegRegReg(cUnit, kOpSub, rlResult.lowReg, tReg2, tReg1); |
| 1682 | } |
| 1683 | } |
| 1684 | storeValue(cUnit, rlDest, rlResult); |
| 1685 | return true; |
| 1686 | } |
| 1687 | |
| 1688 | // Returns true if it added instructions to 'cUnit' to multiply 'rlSrc' by 'lit' |
| 1689 | // and store the result in 'rlDest'. |
| 1690 | static bool handleEasyMultiply(CompilationUnit* cUnit, |
| 1691 | RegLocation rlSrc, RegLocation rlDest, int lit) |
| 1692 | { |
| 1693 | // Can we simplify this multiplication? |
| 1694 | bool powerOfTwo = false; |
| 1695 | bool popCountLE2 = false; |
| 1696 | bool powerOfTwoMinusOne = false; |
| 1697 | if (lit < 2) { |
| 1698 | // Avoid special cases. |
| 1699 | return false; |
| 1700 | } else if (isPowerOfTwo(lit)) { |
| 1701 | powerOfTwo = true; |
| 1702 | } else if (isPopCountLE2(lit)) { |
| 1703 | popCountLE2 = true; |
| 1704 | } else if (isPowerOfTwo(lit + 1)) { |
| 1705 | powerOfTwoMinusOne = true; |
| 1706 | } else { |
| 1707 | return false; |
| 1708 | } |
| 1709 | rlSrc = loadValue(cUnit, rlSrc, kCoreReg); |
| 1710 | RegLocation rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 1711 | if (powerOfTwo) { |
| 1712 | // Shift. |
| 1713 | opRegRegImm(cUnit, kOpLsl, rlResult.lowReg, rlSrc.lowReg, |
| 1714 | lowestSetBit(lit)); |
| 1715 | } else if (popCountLE2) { |
| 1716 | // Shift and add and shift. |
| 1717 | int firstBit = lowestSetBit(lit); |
| 1718 | int secondBit = lowestSetBit(lit ^ (1 << firstBit)); |
| 1719 | genMultiplyByTwoBitMultiplier(cUnit, rlSrc, rlResult, lit, |
| 1720 | firstBit, secondBit); |
| 1721 | } else { |
| 1722 | // Reverse subtract: (src << (shift + 1)) - src. |
| 1723 | assert(powerOfTwoMinusOne); |
| 1724 | // TODO: rsb dst, src, src lsl#lowestSetBit(lit + 1) |
| 1725 | int tReg = oatAllocTemp(cUnit); |
| 1726 | opRegRegImm(cUnit, kOpLsl, tReg, rlSrc.lowReg, lowestSetBit(lit + 1)); |
| 1727 | opRegRegReg(cUnit, kOpSub, rlResult.lowReg, tReg, rlSrc.lowReg); |
| 1728 | } |
| 1729 | storeValue(cUnit, rlDest, rlResult); |
| 1730 | return true; |
| 1731 | } |
| 1732 | |
| 1733 | static bool genArithOpIntLit(CompilationUnit* cUnit, MIR* mir, |
| 1734 | RegLocation rlDest, RegLocation rlSrc, |
| 1735 | int lit) |
| 1736 | { |
| 1737 | Opcode dalvikOpcode = mir->dalvikInsn.opcode; |
| 1738 | RegLocation rlResult; |
| 1739 | OpKind op = (OpKind)0; /* Make gcc happy */ |
| 1740 | int shiftOp = false; |
| 1741 | bool isDiv = false; |
| 1742 | int funcOffset; |
| 1743 | |
| 1744 | switch (dalvikOpcode) { |
| 1745 | case OP_RSUB_INT_LIT8: |
| 1746 | case OP_RSUB_INT: { |
| 1747 | int tReg; |
| 1748 | //TUNING: add support for use of Arm rsub op |
| 1749 | rlSrc = loadValue(cUnit, rlSrc, kCoreReg); |
| 1750 | tReg = oatAllocTemp(cUnit); |
| 1751 | loadConstant(cUnit, tReg, lit); |
| 1752 | rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 1753 | opRegRegReg(cUnit, kOpSub, rlResult.lowReg, |
| 1754 | tReg, rlSrc.lowReg); |
| 1755 | storeValue(cUnit, rlDest, rlResult); |
| 1756 | return false; |
| 1757 | break; |
| 1758 | } |
| 1759 | |
| 1760 | case OP_ADD_INT_LIT8: |
| 1761 | case OP_ADD_INT_LIT16: |
| 1762 | op = kOpAdd; |
| 1763 | break; |
| 1764 | case OP_MUL_INT_LIT8: |
| 1765 | case OP_MUL_INT_LIT16: { |
| 1766 | if (handleEasyMultiply(cUnit, rlSrc, rlDest, lit)) { |
| 1767 | return false; |
| 1768 | } |
| 1769 | op = kOpMul; |
| 1770 | break; |
| 1771 | } |
| 1772 | case OP_AND_INT_LIT8: |
| 1773 | case OP_AND_INT_LIT16: |
| 1774 | op = kOpAnd; |
| 1775 | break; |
| 1776 | case OP_OR_INT_LIT8: |
| 1777 | case OP_OR_INT_LIT16: |
| 1778 | op = kOpOr; |
| 1779 | break; |
| 1780 | case OP_XOR_INT_LIT8: |
| 1781 | case OP_XOR_INT_LIT16: |
| 1782 | op = kOpXor; |
| 1783 | break; |
| 1784 | case OP_SHL_INT_LIT8: |
| 1785 | lit &= 31; |
| 1786 | shiftOp = true; |
| 1787 | op = kOpLsl; |
| 1788 | break; |
| 1789 | case OP_SHR_INT_LIT8: |
| 1790 | lit &= 31; |
| 1791 | shiftOp = true; |
| 1792 | op = kOpAsr; |
| 1793 | break; |
| 1794 | case OP_USHR_INT_LIT8: |
| 1795 | lit &= 31; |
| 1796 | shiftOp = true; |
| 1797 | op = kOpLsr; |
| 1798 | break; |
| 1799 | |
| 1800 | case OP_DIV_INT_LIT8: |
| 1801 | case OP_DIV_INT_LIT16: |
| 1802 | case OP_REM_INT_LIT8: |
| 1803 | case OP_REM_INT_LIT16: |
| 1804 | if (lit == 0) { |
| 1805 | UNIMPLEMENTED(FATAL); |
| 1806 | // FIXME: generate an explicit throw here |
| 1807 | return false; |
| 1808 | } |
| 1809 | if (handleEasyDivide(cUnit, dalvikOpcode, rlSrc, rlDest, lit)) { |
| 1810 | return false; |
| 1811 | } |
| 1812 | oatFlushAllRegs(cUnit); /* Everything to home location */ |
| 1813 | loadValueDirectFixed(cUnit, rlSrc, r0); |
| 1814 | oatClobber(cUnit, r0); |
| 1815 | if ((dalvikOpcode == OP_DIV_INT_LIT8) || |
| 1816 | (dalvikOpcode == OP_DIV_INT_LIT16)) { |
| 1817 | funcOffset = OFFSETOF_MEMBER(Thread, pIdiv); |
| 1818 | isDiv = true; |
| 1819 | } else { |
| 1820 | funcOffset = OFFSETOF_MEMBER(Thread, pIdivmod); |
| 1821 | isDiv = false; |
| 1822 | } |
| 1823 | loadWordDisp(cUnit, rSELF, funcOffset, rLR); |
| 1824 | loadConstant(cUnit, r1, lit); |
| 1825 | opReg(cUnit, kOpBlx, rLR); |
| 1826 | oatClobberCallRegs(cUnit); |
| 1827 | if (isDiv) |
| 1828 | rlResult = oatGetReturn(cUnit); |
| 1829 | else |
| 1830 | rlResult = oatGetReturnAlt(cUnit); |
| 1831 | storeValue(cUnit, rlDest, rlResult); |
| 1832 | return false; |
| 1833 | break; |
| 1834 | default: |
| 1835 | return true; |
| 1836 | } |
| 1837 | rlSrc = loadValue(cUnit, rlSrc, kCoreReg); |
| 1838 | rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 1839 | // Avoid shifts by literal 0 - no support in Thumb. Change to copy |
| 1840 | if (shiftOp && (lit == 0)) { |
| 1841 | genRegCopy(cUnit, rlResult.lowReg, rlSrc.lowReg); |
| 1842 | } else { |
| 1843 | opRegRegImm(cUnit, op, rlResult.lowReg, rlSrc.lowReg, lit); |
| 1844 | } |
| 1845 | storeValue(cUnit, rlDest, rlResult); |
| 1846 | return false; |
| 1847 | } |
| 1848 | |
| 1849 | /* Architectural-specific debugging helpers go here */ |
| 1850 | void oatArchDump(void) |
| 1851 | { |
| 1852 | /* Print compiled opcode in this VM instance */ |
| 1853 | int i, start, streak; |
| 1854 | char buf[1024]; |
| 1855 | |
| 1856 | streak = i = 0; |
| 1857 | buf[0] = 0; |
| 1858 | while (opcodeCoverage[i] == 0 && i < kNumPackedOpcodes) { |
| 1859 | i++; |
| 1860 | } |
| 1861 | if (i == kNumPackedOpcodes) { |
| 1862 | return; |
| 1863 | } |
| 1864 | for (start = i++, streak = 1; i < kNumPackedOpcodes; i++) { |
| 1865 | if (opcodeCoverage[i]) { |
| 1866 | streak++; |
| 1867 | } else { |
| 1868 | if (streak == 1) { |
| 1869 | sprintf(buf+strlen(buf), "%x,", start); |
| 1870 | } else { |
| 1871 | sprintf(buf+strlen(buf), "%x-%x,", start, start + streak - 1); |
| 1872 | } |
| 1873 | streak = 0; |
| 1874 | while (opcodeCoverage[i] == 0 && i < kNumPackedOpcodes) { |
| 1875 | i++; |
| 1876 | } |
| 1877 | if (i < kNumPackedOpcodes) { |
| 1878 | streak = 1; |
| 1879 | start = i; |
| 1880 | } |
| 1881 | } |
| 1882 | } |
| 1883 | if (streak) { |
| 1884 | if (streak == 1) { |
| 1885 | sprintf(buf+strlen(buf), "%x", start); |
| 1886 | } else { |
| 1887 | sprintf(buf+strlen(buf), "%x-%x", start, start + streak - 1); |
| 1888 | } |
| 1889 | } |
| 1890 | if (strlen(buf)) { |
| 1891 | LOG(INFO) << "dalvik.vm.oat.op = " << buf; |
| 1892 | } |
| 1893 | } |