buzbee | e3acd07 | 2012-02-25 17:03:10 -0800 | [diff] [blame^] | 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 | /* |
| 18 | * This file contains codegen for the Mips ISA and is intended to be |
| 19 | * includes by: |
| 20 | * |
| 21 | * Codegen-$(TARGET_ARCH_VARIANT).c |
| 22 | * |
| 23 | */ |
| 24 | |
| 25 | namespace art { |
| 26 | |
| 27 | // FIXME: need the following: |
| 28 | void genSuspendTest(CompilationUnit* cUnit, MIR* mir) {} |
| 29 | void genMonitorEnter(CompilationUnit* cUnit, MIR* mir, RegLocation rlSrc) {} |
| 30 | void genMonitorExit(CompilationUnit* cUnit, MIR* mir, RegLocation rlSrc) {} |
| 31 | void genCheckCast(CompilationUnit* cUnit, MIR* mir, RegLocation rlSrc) {} |
| 32 | void genInstanceof(CompilationUnit* cUnit, MIR* mir, RegLocation rlDest, |
| 33 | RegLocation rlSrc) {} |
| 34 | void genNewInstance(CompilationUnit* cUnit, MIR* mir, |
| 35 | RegLocation rlDest) {} |
| 36 | void genThrow(CompilationUnit* cUnit, MIR* mir, RegLocation rlSrc) {} |
| 37 | void genConstString(CompilationUnit* cUnit, MIR* mir, |
| 38 | RegLocation rlDest, RegLocation rlSrc) {} |
| 39 | void genConstClass(CompilationUnit* cUnit, MIR* mir, |
| 40 | RegLocation rlDest, RegLocation rlSrc) {} |
| 41 | void genArrayGet(CompilationUnit* cUnit, MIR* mir, OpSize size, |
| 42 | RegLocation rlArray, RegLocation rlIndex, |
| 43 | RegLocation rlDest, int scale) {} |
| 44 | void genArrayPut(CompilationUnit* cUnit, MIR* mir, OpSize size, |
| 45 | RegLocation rlArray, RegLocation rlIndex, |
| 46 | RegLocation rlSrc, int scale) {} |
| 47 | void genArrayObjPut(CompilationUnit* cUnit, MIR* mir, |
| 48 | RegLocation rlArray, RegLocation rlIndex, |
| 49 | RegLocation rlSrc, int scale) {} |
| 50 | void genIPut(CompilationUnit* cUnit, MIR* mir, OpSize size, |
| 51 | RegLocation rlSrc, RegLocation rlObj, |
| 52 | bool isLongOrDouble, bool isObject) {} |
| 53 | bool genArithOpInt(CompilationUnit* cUnit, MIR* mir, |
| 54 | RegLocation rlDest, RegLocation rlSrc1, |
| 55 | RegLocation rlSrc2) { return 0; } |
| 56 | bool genArithOpLong(CompilationUnit* cUnit, MIR* mir, |
| 57 | RegLocation rlDest, RegLocation rlSrc1, |
| 58 | RegLocation rlSrc2) { return 0; } |
| 59 | bool genShiftOpLong(CompilationUnit* cUnit, MIR* mir, |
| 60 | RegLocation rlDest, RegLocation rlSrc1, |
| 61 | RegLocation rlShift) { return 0; } |
| 62 | bool genArithOpIntLit(CompilationUnit* cUnit, MIR* mir, |
| 63 | RegLocation rlDest, RegLocation rlSrc, |
| 64 | int lit) { return 0; } |
| 65 | |
| 66 | |
| 67 | |
| 68 | |
| 69 | |
| 70 | |
| 71 | |
| 72 | |
| 73 | STATIC RegLocation getRetLoc(CompilationUnit* cUnit); |
| 74 | |
| 75 | void warnIfUnresolved(CompilationUnit* cUnit, int fieldIdx, Field* field) { |
| 76 | if (field == NULL) { |
| 77 | const DexFile::FieldId& field_id = cUnit->dex_file->GetFieldId(fieldIdx); |
| 78 | std::string class_name(cUnit->dex_file->GetFieldDeclaringClassDescriptor(field_id)); |
| 79 | std::string field_name(cUnit->dex_file->GetFieldName(field_id)); |
| 80 | LOG(INFO) << "Field " << PrettyDescriptor(class_name) << "." << field_name |
| 81 | << " unresolved at compile time"; |
| 82 | } else { |
| 83 | // We also use the slow path for wide volatile fields. |
| 84 | } |
| 85 | } |
| 86 | |
| 87 | /* |
| 88 | * Construct an s4 from two consecutive half-words of switch data. |
| 89 | * This needs to check endianness because the DEX optimizer only swaps |
| 90 | * half-words in instruction stream. |
| 91 | * |
| 92 | * "switchData" must be 32-bit aligned. |
| 93 | */ |
| 94 | #if __BYTE_ORDER == __LITTLE_ENDIAN |
| 95 | STATIC inline s4 s4FromSwitchData(const void* switchData) { |
| 96 | return *(s4*) switchData; |
| 97 | } |
| 98 | #else |
| 99 | STATIC inline s4 s4FromSwitchData(const void* switchData) { |
| 100 | u2* data = switchData; |
| 101 | return data[0] | (((s4) data[1]) << 16); |
| 102 | } |
| 103 | #endif |
| 104 | /* |
| 105 | * Insert a kMipsPseudoCaseLabel at the beginning of the Dalvik |
| 106 | * offset vaddr. This label will be used to fix up the case |
| 107 | * branch table during the assembly phase. Be sure to set |
| 108 | * all resource flags on this to prevent code motion across |
| 109 | * target boundaries. KeyVal is just there for debugging. |
| 110 | */ |
| 111 | STATIC MipsLIR* insertCaseLabel(CompilationUnit* cUnit, int vaddr, int keyVal) |
| 112 | { |
| 113 | std::map<unsigned int, LIR*>::iterator it; |
| 114 | it = cUnit->boundaryMap.find(vaddr); |
| 115 | if (it == cUnit->boundaryMap.end()) { |
| 116 | LOG(FATAL) << "Error: didn't find vaddr 0x" << std::hex << vaddr; |
| 117 | } |
| 118 | MipsLIR* newLabel = (MipsLIR*)oatNew(cUnit, sizeof(MipsLIR), true, kAllocLIR); |
| 119 | newLabel->generic.dalvikOffset = vaddr; |
| 120 | newLabel->opcode = kMipsPseudoCaseLabel; |
| 121 | newLabel->operands[0] = keyVal; |
| 122 | oatInsertLIRAfter(it->second, (LIR*)newLabel); |
| 123 | return newLabel; |
| 124 | } |
| 125 | |
| 126 | STATIC void markPackedCaseLabels(CompilationUnit* cUnit, SwitchTable *tabRec) |
| 127 | { |
| 128 | const u2* table = tabRec->table; |
| 129 | int baseVaddr = tabRec->vaddr; |
| 130 | int *targets = (int*)&table[4]; |
| 131 | int entries = table[1]; |
| 132 | int lowKey = s4FromSwitchData(&table[2]); |
| 133 | for (int i = 0; i < entries; i++) { |
| 134 | tabRec->targets[i] = insertCaseLabel(cUnit, baseVaddr + targets[i], |
| 135 | i + lowKey); |
| 136 | } |
| 137 | } |
| 138 | |
| 139 | STATIC void markSparseCaseLabels(CompilationUnit* cUnit, SwitchTable *tabRec) |
| 140 | { |
| 141 | const u2* table = tabRec->table; |
| 142 | int baseVaddr = tabRec->vaddr; |
| 143 | int entries = table[1]; |
| 144 | int* keys = (int*)&table[2]; |
| 145 | int* targets = &keys[entries]; |
| 146 | for (int i = 0; i < entries; i++) { |
| 147 | tabRec->targets[i] = insertCaseLabel(cUnit, baseVaddr + targets[i], |
| 148 | keys[i]); |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | void oatProcessSwitchTables(CompilationUnit* cUnit) |
| 153 | { |
| 154 | GrowableListIterator iterator; |
| 155 | oatGrowableListIteratorInit(&cUnit->switchTables, &iterator); |
| 156 | while (true) { |
| 157 | SwitchTable *tabRec = (SwitchTable *) oatGrowableListIteratorNext( |
| 158 | &iterator); |
| 159 | if (tabRec == NULL) break; |
| 160 | if (tabRec->table[0] == kPackedSwitchSignature) |
| 161 | markPackedCaseLabels(cUnit, tabRec); |
| 162 | else if (tabRec->table[0] == kSparseSwitchSignature) |
| 163 | markSparseCaseLabels(cUnit, tabRec); |
| 164 | else { |
| 165 | LOG(FATAL) << "Invalid switch table"; |
| 166 | } |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | STATIC void dumpSparseSwitchTable(const u2* table) |
| 171 | /* |
| 172 | * Sparse switch data format: |
| 173 | * ushort ident = 0x0200 magic value |
| 174 | * ushort size number of entries in the table; > 0 |
| 175 | * int keys[size] keys, sorted low-to-high; 32-bit aligned |
| 176 | * int targets[size] branch targets, relative to switch opcode |
| 177 | * |
| 178 | * Total size is (2+size*4) 16-bit code units. |
| 179 | */ |
| 180 | { |
| 181 | u2 ident = table[0]; |
| 182 | int entries = table[1]; |
| 183 | int* keys = (int*)&table[2]; |
| 184 | int* targets = &keys[entries]; |
| 185 | LOG(INFO) << "Sparse switch table - ident:0x" << std::hex << ident << |
| 186 | ", entries: " << std::dec << entries; |
| 187 | for (int i = 0; i < entries; i++) { |
| 188 | LOG(INFO) << " Key[" << keys[i] << "] -> 0x" << std::hex << |
| 189 | targets[i]; |
| 190 | } |
| 191 | } |
| 192 | |
| 193 | STATIC void dumpPackedSwitchTable(const u2* table) |
| 194 | /* |
| 195 | * Packed switch data format: |
| 196 | * ushort ident = 0x0100 magic value |
| 197 | * ushort size number of entries in the table |
| 198 | * int first_key first (and lowest) switch case value |
| 199 | * int targets[size] branch targets, relative to switch opcode |
| 200 | * |
| 201 | * Total size is (4+size*2) 16-bit code units. |
| 202 | */ |
| 203 | { |
| 204 | u2 ident = table[0]; |
| 205 | int* targets = (int*)&table[4]; |
| 206 | int entries = table[1]; |
| 207 | int lowKey = s4FromSwitchData(&table[2]); |
| 208 | LOG(INFO) << "Packed switch table - ident:0x" << std::hex << ident << |
| 209 | ", entries: " << std::dec << entries << ", lowKey: " << lowKey; |
| 210 | for (int i = 0; i < entries; i++) { |
| 211 | LOG(INFO) << " Key[" << (i + lowKey) << "] -> 0x" << std::hex << |
| 212 | targets[i]; |
| 213 | } |
| 214 | } |
| 215 | |
| 216 | /* |
| 217 | * The sparse table in the literal pool is an array of <key,displacement> |
| 218 | * pairs. For each set, we'll load them as a pair using ldmia. |
| 219 | * This means that the register number of the temp we use for the key |
| 220 | * must be lower than the reg for the displacement. |
| 221 | * |
| 222 | * The test loop will look something like: |
| 223 | * |
| 224 | * adr rBase, <table> |
| 225 | * ldr rVal, [rSP, vRegOff] |
| 226 | * mov rIdx, #tableSize |
| 227 | * lp: |
| 228 | * ldmia rBase!, {rKey, rDisp} |
| 229 | * sub rIdx, #1 |
| 230 | * cmp rVal, rKey |
| 231 | * ifeq |
| 232 | * add rPC, rDisp ; This is the branch from which we compute displacement |
| 233 | * cbnz rIdx, lp |
| 234 | */ |
| 235 | STATIC void genSparseSwitch(CompilationUnit* cUnit, MIR* mir, |
| 236 | RegLocation rlSrc) |
| 237 | { |
| 238 | UNIMPLEMENTED(FATAL) << "Need Mips implementation"; |
| 239 | #if 0 |
| 240 | const u2* table = cUnit->insns + mir->offset + mir->dalvikInsn.vB; |
| 241 | if (cUnit->printMe) { |
| 242 | dumpSparseSwitchTable(table); |
| 243 | } |
| 244 | // Add the table to the list - we'll process it later |
| 245 | SwitchTable *tabRec = (SwitchTable *)oatNew(cUnit, sizeof(SwitchTable), |
| 246 | true, kAllocData); |
| 247 | tabRec->table = table; |
| 248 | tabRec->vaddr = mir->offset; |
| 249 | int size = table[1]; |
| 250 | tabRec->targets = (MipsLIR* *)oatNew(cUnit, size * sizeof(MipsLIR*), true, |
| 251 | kAllocLIR); |
| 252 | oatInsertGrowableList(cUnit, &cUnit->switchTables, (intptr_t)tabRec); |
| 253 | |
| 254 | // Get the switch value |
| 255 | rlSrc = loadValue(cUnit, rlSrc, kCoreReg); |
| 256 | int rBase = oatAllocTemp(cUnit); |
| 257 | /* Allocate key and disp temps */ |
| 258 | int rKey = oatAllocTemp(cUnit); |
| 259 | int rDisp = oatAllocTemp(cUnit); |
| 260 | // Make sure rKey's register number is less than rDisp's number for ldmia |
| 261 | if (rKey > rDisp) { |
| 262 | int tmp = rDisp; |
| 263 | rDisp = rKey; |
| 264 | rKey = tmp; |
| 265 | } |
| 266 | // Materialize a pointer to the switch table |
| 267 | newLIR3(cUnit, kThumb2Adr, rBase, 0, (intptr_t)tabRec); |
| 268 | // Set up rIdx |
| 269 | int rIdx = oatAllocTemp(cUnit); |
| 270 | loadConstant(cUnit, rIdx, size); |
| 271 | // Establish loop branch target |
| 272 | MipsLIR* target = newLIR0(cUnit, kMipsPseudoTargetLabel); |
| 273 | target->defMask = ENCODE_ALL; |
| 274 | // Load next key/disp |
| 275 | newLIR2(cUnit, kThumb2LdmiaWB, rBase, (1 << rKey) | (1 << rDisp)); |
| 276 | opRegReg(cUnit, kOpCmp, rKey, rlSrc.lowReg); |
| 277 | // Go if match. NOTE: No instruction set switch here - must stay Thumb2 |
| 278 | genIT(cUnit, kMipsCondEq, ""); |
| 279 | MipsLIR* switchBranch = newLIR1(cUnit, kThumb2AddPCR, rDisp); |
| 280 | tabRec->bxInst = switchBranch; |
| 281 | // Needs to use setflags encoding here |
| 282 | newLIR3(cUnit, kThumb2SubsRRI12, rIdx, rIdx, 1); |
| 283 | MipsLIR* branch = opCondBranch(cUnit, kMipsCondNe); |
| 284 | branch->generic.target = (LIR*)target; |
| 285 | #endif |
| 286 | } |
| 287 | |
| 288 | STATIC void genPackedSwitch(CompilationUnit* cUnit, MIR* mir, |
| 289 | RegLocation rlSrc) |
| 290 | { |
| 291 | UNIMPLEMENTED(FATAL) << "Need Mips implementation"; |
| 292 | #if 0 |
| 293 | const u2* table = cUnit->insns + mir->offset + mir->dalvikInsn.vB; |
| 294 | if (cUnit->printMe) { |
| 295 | dumpPackedSwitchTable(table); |
| 296 | } |
| 297 | // Add the table to the list - we'll process it later |
| 298 | SwitchTable *tabRec = (SwitchTable *)oatNew(cUnit, sizeof(SwitchTable), |
| 299 | true, kAllocData); |
| 300 | tabRec->table = table; |
| 301 | tabRec->vaddr = mir->offset; |
| 302 | int size = table[1]; |
| 303 | tabRec->targets = (MipsLIR* *)oatNew(cUnit, size * sizeof(MipsLIR*), true, |
| 304 | kAllocLIR); |
| 305 | oatInsertGrowableList(cUnit, &cUnit->switchTables, (intptr_t)tabRec); |
| 306 | |
| 307 | // Get the switch value |
| 308 | rlSrc = loadValue(cUnit, rlSrc, kCoreReg); |
| 309 | int tableBase = oatAllocTemp(cUnit); |
| 310 | // Materialize a pointer to the switch table |
| 311 | newLIR3(cUnit, kThumb2Adr, tableBase, 0, (intptr_t)tabRec); |
| 312 | int lowKey = s4FromSwitchData(&table[2]); |
| 313 | int keyReg; |
| 314 | // Remove the bias, if necessary |
| 315 | if (lowKey == 0) { |
| 316 | keyReg = rlSrc.lowReg; |
| 317 | } else { |
| 318 | keyReg = oatAllocTemp(cUnit); |
| 319 | opRegRegImm(cUnit, kOpSub, keyReg, rlSrc.lowReg, lowKey); |
| 320 | } |
| 321 | // Bounds check - if < 0 or >= size continue following switch |
| 322 | opRegImm(cUnit, kOpCmp, keyReg, size-1); |
| 323 | MipsLIR* branchOver = opCondBranch(cUnit, kMipsCondHi); |
| 324 | |
| 325 | // Load the displacement from the switch table |
| 326 | int dispReg = oatAllocTemp(cUnit); |
| 327 | loadBaseIndexed(cUnit, tableBase, keyReg, dispReg, 2, kWord); |
| 328 | |
| 329 | // ..and go! NOTE: No instruction set switch here - must stay Thumb2 |
| 330 | MipsLIR* switchBranch = newLIR1(cUnit, kThumb2AddPCR, dispReg); |
| 331 | tabRec->bxInst = switchBranch; |
| 332 | |
| 333 | /* branchOver target here */ |
| 334 | MipsLIR* target = newLIR0(cUnit, kMipsPseudoTargetLabel); |
| 335 | target->defMask = ENCODE_ALL; |
| 336 | branchOver->generic.target = (LIR*)target; |
| 337 | #endif |
| 338 | } |
| 339 | |
| 340 | /* |
| 341 | * Array data table format: |
| 342 | * ushort ident = 0x0300 magic value |
| 343 | * ushort width width of each element in the table |
| 344 | * uint size number of elements in the table |
| 345 | * ubyte data[size*width] table of data values (may contain a single-byte |
| 346 | * padding at the end) |
| 347 | * |
| 348 | * Total size is 4+(width * size + 1)/2 16-bit code units. |
| 349 | */ |
| 350 | STATIC void genFillArrayData(CompilationUnit* cUnit, MIR* mir, |
| 351 | RegLocation rlSrc) |
| 352 | { |
| 353 | UNIMPLEMENTED(FATAL) << "Need Mips implementation"; |
| 354 | #if 0 |
| 355 | const u2* table = cUnit->insns + mir->offset + mir->dalvikInsn.vB; |
| 356 | // Add the table to the list - we'll process it later |
| 357 | FillArrayData *tabRec = (FillArrayData *) |
| 358 | oatNew(cUnit, sizeof(FillArrayData), true, kAllocData); |
| 359 | tabRec->table = table; |
| 360 | tabRec->vaddr = mir->offset; |
| 361 | u2 width = tabRec->table[1]; |
| 362 | u4 size = tabRec->table[2] | (((u4)tabRec->table[3]) << 16); |
| 363 | tabRec->size = (size * width) + 8; |
| 364 | |
| 365 | oatInsertGrowableList(cUnit, &cUnit->fillArrayData, (intptr_t)tabRec); |
| 366 | |
| 367 | // Making a call - use explicit registers |
| 368 | oatFlushAllRegs(cUnit); /* Everything to home location */ |
| 369 | loadValueDirectFixed(cUnit, rlSrc, r0); |
| 370 | loadWordDisp(cUnit, rSELF, |
| 371 | OFFSETOF_MEMBER(Thread, pHandleFillArrayDataFromCode), rLR); |
| 372 | // Materialize a pointer to the fill data image |
| 373 | newLIR3(cUnit, kThumb2Adr, r1, 0, (intptr_t)tabRec); |
| 374 | callRuntimeHelper(cUnit, rLR); |
| 375 | #endif |
| 376 | } |
| 377 | |
| 378 | STATIC void genIGet(CompilationUnit* cUnit, MIR* mir, OpSize size, |
| 379 | RegLocation rlDest, RegLocation rlObj, |
| 380 | bool isLongOrDouble, bool isObject) |
| 381 | { |
| 382 | UNIMPLEMENTED(FATAL) << "Need Mips implementation"; |
| 383 | #if 0 |
| 384 | int fieldOffset; |
| 385 | bool isVolatile; |
| 386 | uint32_t fieldIdx = mir->dalvikInsn.vC; |
| 387 | bool fastPath = |
| 388 | cUnit->compiler->ComputeInstanceFieldInfo(fieldIdx, cUnit, |
| 389 | fieldOffset, isVolatile, false); |
| 390 | if (fastPath && !SLOW_FIELD_PATH) { |
| 391 | RegLocation rlResult; |
| 392 | RegisterClass regClass = oatRegClassBySize(size); |
| 393 | DCHECK_GE(fieldOffset, 0); |
| 394 | rlObj = loadValue(cUnit, rlObj, kCoreReg); |
| 395 | if (isLongOrDouble) { |
| 396 | DCHECK(rlDest.wide); |
| 397 | genNullCheck(cUnit, rlObj.sRegLow, rlObj.lowReg, mir);/* null obj? */ |
| 398 | int regPtr = oatAllocTemp(cUnit); |
| 399 | opRegRegImm(cUnit, kOpAdd, regPtr, rlObj.lowReg, fieldOffset); |
| 400 | rlResult = oatEvalLoc(cUnit, rlDest, regClass, true); |
| 401 | loadPair(cUnit, regPtr, rlResult.lowReg, rlResult.highReg); |
| 402 | if (isVolatile) { |
| 403 | oatGenMemBarrier(cUnit, kSY); |
| 404 | } |
| 405 | oatFreeTemp(cUnit, regPtr); |
| 406 | storeValueWide(cUnit, rlDest, rlResult); |
| 407 | } else { |
| 408 | rlResult = oatEvalLoc(cUnit, rlDest, regClass, true); |
| 409 | genNullCheck(cUnit, rlObj.sRegLow, rlObj.lowReg, mir);/* null object? */ |
| 410 | loadBaseDisp(cUnit, mir, rlObj.lowReg, fieldOffset, rlResult.lowReg, |
| 411 | kWord, rlObj.sRegLow); |
| 412 | if (isVolatile) { |
| 413 | oatGenMemBarrier(cUnit, kSY); |
| 414 | } |
| 415 | storeValue(cUnit, rlDest, rlResult); |
| 416 | } |
| 417 | } else { |
| 418 | int getterOffset = isLongOrDouble ? OFFSETOF_MEMBER(Thread, pGet64Instance) : |
| 419 | (isObject ? OFFSETOF_MEMBER(Thread, pGetObjInstance) |
| 420 | : OFFSETOF_MEMBER(Thread, pGet32Instance)); |
| 421 | loadWordDisp(cUnit, rSELF, getterOffset, rLR); |
| 422 | loadValueDirect(cUnit, rlObj, r1); |
| 423 | loadConstant(cUnit, r0, fieldIdx); |
| 424 | callRuntimeHelper(cUnit, rLR); |
| 425 | if (isLongOrDouble) { |
| 426 | RegLocation rlResult = oatGetReturnWide(cUnit); |
| 427 | storeValueWide(cUnit, rlDest, rlResult); |
| 428 | } else { |
| 429 | RegLocation rlResult = oatGetReturn(cUnit); |
| 430 | storeValue(cUnit, rlDest, rlResult); |
| 431 | } |
| 432 | } |
| 433 | #endif |
| 434 | } |
| 435 | |
| 436 | /* |
| 437 | * Perform a "reg cmp imm" operation and jump to the PCR region if condition |
| 438 | * satisfies. |
| 439 | */ |
| 440 | STATIC void genNegFloat(CompilationUnit *cUnit, RegLocation rlDest, |
| 441 | RegLocation rlSrc) |
| 442 | { |
| 443 | RegLocation rlResult; |
| 444 | rlSrc = loadValue(cUnit, rlSrc, kCoreReg); |
| 445 | rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 446 | opRegRegImm(cUnit, kOpAdd, rlResult.lowReg, |
| 447 | rlSrc.lowReg, 0x80000000); |
| 448 | storeValue(cUnit, rlDest, rlResult); |
| 449 | } |
| 450 | |
| 451 | STATIC void genNegDouble(CompilationUnit *cUnit, RegLocation rlDest, |
| 452 | RegLocation rlSrc) |
| 453 | { |
| 454 | RegLocation rlResult; |
| 455 | rlSrc = loadValueWide(cUnit, rlSrc, kCoreReg); |
| 456 | rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 457 | opRegRegImm(cUnit, kOpAdd, rlResult.highReg, rlSrc.highReg, |
| 458 | 0x80000000); |
| 459 | genRegCopy(cUnit, rlResult.lowReg, rlSrc.lowReg); |
| 460 | storeValueWide(cUnit, rlDest, rlResult); |
| 461 | } |
| 462 | |
| 463 | STATIC void genMulLong(CompilationUnit *cUnit, RegLocation rlDest, |
| 464 | RegLocation rlSrc1, RegLocation rlSrc2) |
| 465 | { |
| 466 | UNIMPLEMENTED(FATAL) << "Need Mips implementation"; |
| 467 | #if 0 |
| 468 | RegLocation rlResult; |
| 469 | loadValueDirectWideFixed(cUnit, rlSrc1, r_ARG0, r_ARG1); |
| 470 | loadValueDirectWideFixed(cUnit, rlSrc2, r_ARG2, r_ARG3); |
| 471 | genDispatchToHandler(cUnit, TEMPLATE_MUL_LONG); |
| 472 | rlResult = oatGetReturnWide(cUnit); |
| 473 | storeValueWide(cUnit, rlDest, rlResult); |
| 474 | #endif |
| 475 | } |
| 476 | |
| 477 | STATIC bool partialOverlap(int sreg1, int sreg2) |
| 478 | { |
| 479 | return abs(sreg1 - sreg2) == 1; |
| 480 | } |
| 481 | |
| 482 | STATIC void withCarryHelper(CompilationUnit *cUnit, MipsOpCode opc, |
| 483 | RegLocation rlDest, RegLocation rlSrc1, |
| 484 | RegLocation rlSrc2, int sltuSrc1, int sltuSrc2) |
| 485 | { |
| 486 | int tReg = oatAllocTemp(cUnit); |
| 487 | newLIR3(cUnit, opc, rlDest.lowReg, rlSrc1.lowReg, rlSrc2.lowReg); |
| 488 | newLIR3(cUnit, kMipsSltu, tReg, sltuSrc1, sltuSrc2); |
| 489 | newLIR3(cUnit, opc, rlDest.highReg, rlSrc1.highReg, rlSrc2.highReg); |
| 490 | newLIR3(cUnit, opc, rlDest.highReg, rlDest.highReg, tReg); |
| 491 | oatFreeTemp(cUnit, tReg); |
| 492 | } |
| 493 | |
| 494 | STATIC void genLong3Addr(CompilationUnit *cUnit, MIR *mir, OpKind firstOp, |
| 495 | OpKind secondOp, RegLocation rlDest, |
| 496 | RegLocation rlSrc1, RegLocation rlSrc2) |
| 497 | { |
| 498 | UNIMPLEMENTED(FATAL) << "Need Mips implementation"; |
| 499 | #if 0 |
| 500 | RegLocation rlResult; |
| 501 | int carryOp = (secondOp == kOpAdc || secondOp == kOpSbc); |
| 502 | |
| 503 | if (partialOverlap(rlSrc1.sRegLow,rlSrc2.sRegLow) || |
| 504 | partialOverlap(rlSrc1.sRegLow,rlDest.sRegLow) || |
| 505 | partialOverlap(rlSrc2.sRegLow,rlDest.sRegLow)) { |
| 506 | // Rare case - not enough registers to properly handle |
| 507 | genInterpSingleStep(cUnit, mir); |
| 508 | } else if (rlDest.sRegLow == rlSrc1.sRegLow) { |
| 509 | rlResult = loadValueWide(cUnit, rlDest, kCoreReg); |
| 510 | rlSrc2 = loadValueWide(cUnit, rlSrc2, kCoreReg); |
| 511 | if (!carryOp) { |
| 512 | opRegRegReg(cUnit, firstOp, rlResult.lowReg, rlResult.lowReg, rlSrc2.lowReg); |
| 513 | opRegRegReg(cUnit, secondOp, rlResult.highReg, rlResult.highReg, rlSrc2.highReg); |
| 514 | } else if (secondOp == kOpAdc) { |
| 515 | withCarryHelper(cUnit, kMipsAddu, rlResult, rlResult, rlSrc2, |
| 516 | rlResult.lowReg, rlSrc2.lowReg); |
| 517 | } else { |
| 518 | int tReg = oatAllocTemp(cUnit); |
| 519 | newLIR2(cUnit, kMipsMove, tReg, rlResult.lowReg); |
| 520 | withCarryHelper(cUnit, kMipsSubu, rlResult, rlResult, rlSrc2, |
| 521 | tReg, rlResult.lowReg); |
| 522 | oatFreeTemp(cUnit, tReg); |
| 523 | } |
| 524 | storeValueWide(cUnit, rlDest, rlResult); |
| 525 | } else if (rlDest.sRegLow == rlSrc2.sRegLow) { |
| 526 | rlResult = loadValueWide(cUnit, rlDest, kCoreReg); |
| 527 | rlSrc1 = loadValueWide(cUnit, rlSrc1, kCoreReg); |
| 528 | if (!carryOp) { |
| 529 | opRegRegReg(cUnit, firstOp, rlResult.lowReg, rlSrc1.lowReg, rlResult.lowReg); |
| 530 | opRegRegReg(cUnit, secondOp, rlResult.highReg, rlSrc1.highReg, rlResult.highReg); |
| 531 | } else if (secondOp == kOpAdc) { |
| 532 | withCarryHelper(cUnit, kMipsAddu, rlResult, rlSrc1, rlResult, |
| 533 | rlResult.lowReg, rlSrc1.lowReg); |
| 534 | } else { |
| 535 | withCarryHelper(cUnit, kMipsSubu, rlResult, rlSrc1, rlResult, |
| 536 | rlSrc1.lowReg, rlResult.lowReg); |
| 537 | } |
| 538 | storeValueWide(cUnit, rlDest, rlResult); |
| 539 | } else { |
| 540 | rlSrc1 = loadValueWide(cUnit, rlSrc1, kCoreReg); |
| 541 | rlSrc2 = loadValueWide(cUnit, rlSrc2, kCoreReg); |
| 542 | rlResult = oatEvalLoc(cUnit, rlDest, kCoreReg, true); |
| 543 | if (!carryOp) { |
| 544 | opRegRegReg(cUnit, firstOp, rlResult.lowReg, rlSrc1.lowReg, rlSrc2.lowReg); |
| 545 | opRegRegReg(cUnit, secondOp, rlResult.highReg, rlSrc1.highReg, rlSrc2.highReg); |
| 546 | } else if (secondOp == kOpAdc) { |
| 547 | withCarryHelper(cUnit, kMipsAddu, rlResult, rlSrc1, rlSrc2, |
| 548 | rlResult.lowReg, rlSrc1.lowReg); |
| 549 | } else { |
| 550 | withCarryHelper(cUnit, kMipsSubu, rlResult, rlSrc1, rlSrc2, |
| 551 | rlSrc1.lowReg, rlResult.lowReg); |
| 552 | } |
| 553 | storeValueWide(cUnit, rlDest, rlResult); |
| 554 | } |
| 555 | #endif |
| 556 | } |
| 557 | |
| 558 | void oatInitializeRegAlloc(CompilationUnit* cUnit) |
| 559 | { |
| 560 | int numRegs = sizeof(coreRegs)/sizeof(*coreRegs); |
| 561 | int numReserved = sizeof(reservedRegs)/sizeof(*reservedRegs); |
| 562 | int numTemps = sizeof(coreTemps)/sizeof(*coreTemps); |
| 563 | #ifdef __mips_hard_float |
| 564 | int numFPRegs = sizeof(fpRegs)/sizeof(*fpRegs); |
| 565 | int numFPTemps = sizeof(fpTemps)/sizeof(*fpTemps); |
| 566 | #else |
| 567 | int numFPRegs = 0; |
| 568 | int numFPTemps = 0; |
| 569 | #endif |
| 570 | RegisterPool *pool = (RegisterPool *)oatNew(cUnit, sizeof(*pool), true, |
| 571 | kAllocRegAlloc); |
| 572 | cUnit->regPool = pool; |
| 573 | pool->numCoreRegs = numRegs; |
| 574 | pool->coreRegs = (RegisterInfo *) |
| 575 | oatNew(cUnit, numRegs * sizeof(*cUnit->regPool->coreRegs), |
| 576 | true, kAllocRegAlloc); |
| 577 | pool->numFPRegs = numFPRegs; |
| 578 | pool->FPRegs = numFPRegs == 0 ? NULL : (RegisterInfo *) |
| 579 | oatNew(cUnit, numFPRegs * sizeof(*cUnit->regPool->FPRegs), true, |
| 580 | kAllocRegAlloc); |
| 581 | oatInitPool(pool->coreRegs, coreRegs, pool->numCoreRegs); |
| 582 | oatInitPool(pool->FPRegs, fpRegs, pool->numFPRegs); |
| 583 | // Keep special registers from being allocated |
| 584 | for (int i = 0; i < numReserved; i++) { |
| 585 | if (NO_SUSPEND && !cUnit->genDebugger && |
| 586 | (reservedRegs[i] == rSUSPEND)) { |
| 587 | //To measure cost of suspend check |
| 588 | continue; |
| 589 | } |
| 590 | oatMarkInUse(cUnit, reservedRegs[i]); |
| 591 | } |
| 592 | // Mark temp regs - all others not in use can be used for promotion |
| 593 | for (int i = 0; i < numTemps; i++) { |
| 594 | oatMarkTemp(cUnit, coreTemps[i]); |
| 595 | } |
| 596 | for (int i = 0; i < numFPTemps; i++) { |
| 597 | oatMarkTemp(cUnit, fpTemps[i]); |
| 598 | } |
| 599 | // Construct the alias map. |
| 600 | cUnit->phiAliasMap = (int*)oatNew(cUnit, cUnit->numSSARegs * |
| 601 | sizeof(cUnit->phiAliasMap[0]), false, |
| 602 | kAllocDFInfo); |
| 603 | for (int i = 0; i < cUnit->numSSARegs; i++) { |
| 604 | cUnit->phiAliasMap[i] = i; |
| 605 | } |
| 606 | for (MIR* phi = cUnit->phiList; phi; phi = phi->meta.phiNext) { |
| 607 | int defReg = phi->ssaRep->defs[0]; |
| 608 | for (int i = 0; i < phi->ssaRep->numUses; i++) { |
| 609 | for (int j = 0; j < cUnit->numSSARegs; j++) { |
| 610 | if (cUnit->phiAliasMap[j] == phi->ssaRep->uses[i]) { |
| 611 | cUnit->phiAliasMap[j] = defReg; |
| 612 | } |
| 613 | } |
| 614 | } |
| 615 | } |
| 616 | } |
| 617 | |
| 618 | STATIC void genMonitor(CompilationUnit *cUnit, MIR *mir) |
| 619 | { |
| 620 | UNIMPLEMENTED(FATAL) << "Need Mips implementation"; |
| 621 | #if 0 |
| 622 | genMonitorPortable(cUnit, mir); |
| 623 | #endif |
| 624 | } |
| 625 | |
| 626 | STATIC void genCmpLong(CompilationUnit *cUnit, MIR *mir, RegLocation rlDest, |
| 627 | RegLocation rlSrc1, RegLocation rlSrc2) |
| 628 | { |
| 629 | UNIMPLEMENTED(FATAL) << "Need Mips implementation"; |
| 630 | #if 0 |
| 631 | RegLocation rlResult; |
| 632 | loadValueDirectWideFixed(cUnit, rlSrc1, r_ARG0, r_ARG1); |
| 633 | loadValueDirectWideFixed(cUnit, rlSrc2, r_ARG2, r_ARG3); |
| 634 | genDispatchToHandler(cUnit, TEMPLATE_CMP_LONG); |
| 635 | rlResult = oatGetReturn(cUnit); |
| 636 | storeValue(cUnit, rlDest, rlResult); |
| 637 | #endif |
| 638 | } |
| 639 | |
| 640 | STATIC void genMultiplyByTwoBitMultiplier(CompilationUnit *cUnit, |
| 641 | RegLocation rlSrc, RegLocation rlResult, int lit, |
| 642 | int firstBit, int secondBit) |
| 643 | { |
| 644 | // We can't implement "add src, src, src, lsl#shift" on Thumb, so we have |
| 645 | // to do a regular multiply. |
| 646 | opRegRegImm(cUnit, kOpMul, rlResult.lowReg, rlSrc.lowReg, lit); |
| 647 | } |
| 648 | |
| 649 | } // namespace art |