| 1 | // Licensed to the .NET Foundation under one or more agreements. |
| 2 | // The .NET Foundation licenses this file to you under the MIT license. |
| 3 | // See the LICENSE file in the project root for more information. |
| 4 | |
| 5 | /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 6 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 7 | XX XX |
| 8 | XX UnwindInfo XX |
| 9 | XX XX |
| 10 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 11 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 12 | */ |
| 13 | |
| 14 | #include "jitpch.h" |
| 15 | #ifdef _MSC_VER |
| 16 | #pragma hdrstop |
| 17 | #endif |
| 18 | |
| 19 | #if defined(_TARGET_ARM_) && defined(_TARGET_UNIX_) |
| 20 | int Compiler::mapRegNumToDwarfReg(regNumber reg) |
| 21 | { |
| 22 | int dwarfReg = DWARF_REG_ILLEGAL; |
| 23 | |
| 24 | switch (reg) |
| 25 | { |
| 26 | case REG_R0: |
| 27 | dwarfReg = 0; |
| 28 | break; |
| 29 | case REG_R1: |
| 30 | dwarfReg = 1; |
| 31 | break; |
| 32 | case REG_R2: |
| 33 | dwarfReg = 2; |
| 34 | break; |
| 35 | case REG_R3: |
| 36 | dwarfReg = 3; |
| 37 | break; |
| 38 | case REG_R4: |
| 39 | dwarfReg = 4; |
| 40 | break; |
| 41 | case REG_R5: |
| 42 | dwarfReg = 5; |
| 43 | break; |
| 44 | case REG_R6: |
| 45 | dwarfReg = 6; |
| 46 | break; |
| 47 | case REG_R7: |
| 48 | dwarfReg = 7; |
| 49 | break; |
| 50 | case REG_R8: |
| 51 | dwarfReg = 8; |
| 52 | break; |
| 53 | case REG_R9: |
| 54 | dwarfReg = 9; |
| 55 | break; |
| 56 | case REG_R10: |
| 57 | dwarfReg = 10; |
| 58 | break; |
| 59 | case REG_R11: |
| 60 | dwarfReg = 11; |
| 61 | break; |
| 62 | case REG_R12: |
| 63 | dwarfReg = 12; |
| 64 | break; |
| 65 | case REG_R13: |
| 66 | dwarfReg = 13; |
| 67 | break; |
| 68 | case REG_R14: |
| 69 | dwarfReg = 14; |
| 70 | break; |
| 71 | case REG_R15: |
| 72 | dwarfReg = 15; |
| 73 | break; |
| 74 | case REG_F0: |
| 75 | dwarfReg = 64; |
| 76 | break; |
| 77 | case REG_F1: |
| 78 | dwarfReg = 65; |
| 79 | break; |
| 80 | case REG_F2: |
| 81 | dwarfReg = 66; |
| 82 | break; |
| 83 | case REG_F3: |
| 84 | dwarfReg = 67; |
| 85 | break; |
| 86 | case REG_F4: |
| 87 | dwarfReg = 68; |
| 88 | break; |
| 89 | case REG_F5: |
| 90 | dwarfReg = 69; |
| 91 | break; |
| 92 | case REG_F6: |
| 93 | dwarfReg = 70; |
| 94 | break; |
| 95 | case REG_F7: |
| 96 | dwarfReg = 71; |
| 97 | break; |
| 98 | case REG_F8: |
| 99 | dwarfReg = 72; |
| 100 | break; |
| 101 | case REG_F9: |
| 102 | dwarfReg = 73; |
| 103 | break; |
| 104 | case REG_F10: |
| 105 | dwarfReg = 74; |
| 106 | break; |
| 107 | case REG_F11: |
| 108 | dwarfReg = 75; |
| 109 | break; |
| 110 | case REG_F12: |
| 111 | dwarfReg = 76; |
| 112 | break; |
| 113 | case REG_F13: |
| 114 | dwarfReg = 77; |
| 115 | break; |
| 116 | case REG_F14: |
| 117 | dwarfReg = 78; |
| 118 | break; |
| 119 | case REG_F15: |
| 120 | dwarfReg = 79; |
| 121 | break; |
| 122 | case REG_F16: |
| 123 | dwarfReg = 80; |
| 124 | break; |
| 125 | case REG_F17: |
| 126 | dwarfReg = 81; |
| 127 | break; |
| 128 | case REG_F18: |
| 129 | dwarfReg = 82; |
| 130 | break; |
| 131 | case REG_F19: |
| 132 | dwarfReg = 83; |
| 133 | break; |
| 134 | case REG_F20: |
| 135 | dwarfReg = 84; |
| 136 | break; |
| 137 | case REG_F21: |
| 138 | dwarfReg = 85; |
| 139 | break; |
| 140 | case REG_F22: |
| 141 | dwarfReg = 86; |
| 142 | break; |
| 143 | case REG_F23: |
| 144 | dwarfReg = 87; |
| 145 | break; |
| 146 | case REG_F24: |
| 147 | dwarfReg = 88; |
| 148 | break; |
| 149 | case REG_F25: |
| 150 | dwarfReg = 89; |
| 151 | break; |
| 152 | case REG_F26: |
| 153 | dwarfReg = 90; |
| 154 | break; |
| 155 | case REG_F27: |
| 156 | dwarfReg = 91; |
| 157 | break; |
| 158 | case REG_F28: |
| 159 | dwarfReg = 92; |
| 160 | break; |
| 161 | case REG_F29: |
| 162 | dwarfReg = 93; |
| 163 | break; |
| 164 | case REG_F30: |
| 165 | dwarfReg = 94; |
| 166 | break; |
| 167 | case REG_F31: |
| 168 | dwarfReg = 95; |
| 169 | break; |
| 170 | default: |
| 171 | noway_assert(!"unexpected REG_NUM" ); |
| 172 | } |
| 173 | |
| 174 | return dwarfReg; |
| 175 | } |
| 176 | #endif // _TARGET_ARM_ && _TARGET_UNIX_ |
| 177 | |
| 178 | #ifdef _TARGET_ARMARCH_ |
| 179 | |
| 180 | /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 181 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 182 | XX XX |
| 183 | XX Unwind APIs XX |
| 184 | XX XX |
| 185 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 186 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 187 | */ |
| 188 | |
| 189 | void Compiler::unwindBegProlog() |
| 190 | { |
| 191 | assert(compGeneratingProlog); |
| 192 | |
| 193 | #if defined(_TARGET_UNIX_) |
| 194 | if (generateCFIUnwindCodes()) |
| 195 | { |
| 196 | unwindBegPrologCFI(); |
| 197 | return; |
| 198 | } |
| 199 | #endif // _TARGET_UNIX_ |
| 200 | |
| 201 | FuncInfoDsc* func = funCurrentFunc(); |
| 202 | |
| 203 | // There is only one prolog for a function/funclet, and it comes first. So now is |
| 204 | // a good time to initialize all the unwind data structures. |
| 205 | |
| 206 | emitLocation* startLoc; |
| 207 | emitLocation* endLoc; |
| 208 | unwindGetFuncLocations(func, true, &startLoc, &endLoc); |
| 209 | |
| 210 | func->uwi.InitUnwindInfo(this, startLoc, endLoc); |
| 211 | func->uwi.CaptureLocation(); |
| 212 | |
| 213 | func->uwiCold = NULL; // No cold data yet |
| 214 | } |
| 215 | |
| 216 | void Compiler::unwindEndProlog() |
| 217 | { |
| 218 | assert(compGeneratingProlog); |
| 219 | } |
| 220 | |
| 221 | void Compiler::unwindBegEpilog() |
| 222 | { |
| 223 | assert(compGeneratingEpilog); |
| 224 | |
| 225 | #if defined(_TARGET_UNIX_) |
| 226 | if (generateCFIUnwindCodes()) |
| 227 | { |
| 228 | return; |
| 229 | } |
| 230 | #endif // _TARGET_UNIX_ |
| 231 | |
| 232 | funCurrentFunc()->uwi.AddEpilog(); |
| 233 | } |
| 234 | |
| 235 | void Compiler::unwindEndEpilog() |
| 236 | { |
| 237 | assert(compGeneratingEpilog); |
| 238 | } |
| 239 | |
| 240 | #if defined(_TARGET_ARM_) |
| 241 | |
| 242 | void Compiler::unwindPushPopMaskInt(regMaskTP maskInt, bool useOpsize16) |
| 243 | { |
| 244 | // floating point registers cannot be specified in 'maskInt' |
| 245 | assert((maskInt & RBM_ALLFLOAT) == 0); |
| 246 | |
| 247 | UnwindInfo* pu = &funCurrentFunc()->uwi; |
| 248 | |
| 249 | if (useOpsize16) |
| 250 | { |
| 251 | // The 16-bit opcode only encode R0-R7 and LR |
| 252 | assert((maskInt & ~(RBM_R0 | RBM_R1 | RBM_R2 | RBM_R3 | RBM_R4 | RBM_R5 | RBM_R6 | RBM_R7 | RBM_LR)) == 0); |
| 253 | |
| 254 | bool shortFormat = false; |
| 255 | BYTE val = 0; |
| 256 | |
| 257 | if ((maskInt & (RBM_R0 | RBM_R1 | RBM_R2 | RBM_R3)) == 0) |
| 258 | { |
| 259 | regMaskTP matchMask = maskInt & (RBM_R4 | RBM_R5 | RBM_R6 | RBM_R7); |
| 260 | regMaskTP valMask = RBM_R4; |
| 261 | while (val < 4) |
| 262 | { |
| 263 | if (matchMask == valMask) |
| 264 | { |
| 265 | shortFormat = true; |
| 266 | break; |
| 267 | } |
| 268 | |
| 269 | valMask <<= 1; |
| 270 | valMask |= RBM_R4; |
| 271 | |
| 272 | val++; |
| 273 | } |
| 274 | } |
| 275 | |
| 276 | if (shortFormat) |
| 277 | { |
| 278 | // D0-D7 : pop {r4-rX,lr} (X=4-7) (opsize 16) |
| 279 | pu->AddCode(0xD0 | ((maskInt >> 12) & 0x4) | val); |
| 280 | } |
| 281 | else |
| 282 | { |
| 283 | // EC-ED : pop {r0-r7,lr} (opsize 16) |
| 284 | pu->AddCode(0xEC | ((maskInt >> 14) & 0x1), (BYTE)maskInt); |
| 285 | } |
| 286 | } |
| 287 | else |
| 288 | { |
| 289 | assert((maskInt & |
| 290 | ~(RBM_R0 | RBM_R1 | RBM_R2 | RBM_R3 | RBM_R4 | RBM_R5 | RBM_R6 | RBM_R7 | RBM_R8 | RBM_R9 | RBM_R10 | |
| 291 | RBM_R11 | RBM_R12 | RBM_LR)) == 0); |
| 292 | |
| 293 | bool shortFormat = false; |
| 294 | BYTE val = 0; |
| 295 | |
| 296 | if (((maskInt & (RBM_R0 | RBM_R1 | RBM_R2 | RBM_R3)) == 0) && |
| 297 | ((maskInt & (RBM_R4 | RBM_R5 | RBM_R6 | RBM_R7 | RBM_R8)) == (RBM_R4 | RBM_R5 | RBM_R6 | RBM_R7 | RBM_R8))) |
| 298 | { |
| 299 | regMaskTP matchMask = maskInt & (RBM_R4 | RBM_R5 | RBM_R6 | RBM_R7 | RBM_R8 | RBM_R9 | RBM_R10 | RBM_R11); |
| 300 | regMaskTP valMask = RBM_R4 | RBM_R5 | RBM_R6 | RBM_R7 | RBM_R8; |
| 301 | while (val < 4) |
| 302 | { |
| 303 | if (matchMask == valMask) |
| 304 | { |
| 305 | shortFormat = true; |
| 306 | break; |
| 307 | } |
| 308 | |
| 309 | valMask <<= 1; |
| 310 | valMask |= RBM_R4; |
| 311 | |
| 312 | val++; |
| 313 | } |
| 314 | } |
| 315 | |
| 316 | if (shortFormat) |
| 317 | { |
| 318 | // D8-DF : pop {r4-rX,lr} (X=8-11) (opsize 32) |
| 319 | pu->AddCode(0xD8 | ((maskInt >> 12) & 0x4) | val); |
| 320 | } |
| 321 | else |
| 322 | { |
| 323 | // 80-BF : pop {r0-r12,lr} (opsize 32) |
| 324 | pu->AddCode(0x80 | ((maskInt >> 8) & 0x1F) | ((maskInt >> 9) & 0x20), (BYTE)maskInt); |
| 325 | } |
| 326 | } |
| 327 | } |
| 328 | |
| 329 | void Compiler::unwindPushPopMaskFloat(regMaskTP maskFloat) |
| 330 | { |
| 331 | // Only floating pointer registers can be specified in 'maskFloat' |
| 332 | assert((maskFloat & ~RBM_ALLFLOAT) == 0); |
| 333 | |
| 334 | // If the maskFloat is zero there is no unwind code to emit |
| 335 | // |
| 336 | if (maskFloat == RBM_NONE) |
| 337 | { |
| 338 | return; |
| 339 | } |
| 340 | |
| 341 | UnwindInfo* pu = &funCurrentFunc()->uwi; |
| 342 | |
| 343 | BYTE val = 0; |
| 344 | regMaskTP valMask = (RBM_F16 | RBM_F17); |
| 345 | |
| 346 | while (maskFloat != valMask) |
| 347 | { |
| 348 | valMask <<= 2; |
| 349 | valMask |= (RBM_F16 | RBM_F17); |
| 350 | |
| 351 | val++; |
| 352 | |
| 353 | if (val == 8) |
| 354 | { |
| 355 | noway_assert(!"Illegal maskFloat" ); |
| 356 | } |
| 357 | } |
| 358 | |
| 359 | // E0-E7 : vpop {d8-dX} (X=8-15) (opsize 32) |
| 360 | assert(0 <= val && val <= 7); |
| 361 | pu->AddCode(0xE0 | val); |
| 362 | } |
| 363 | |
| 364 | void Compiler::unwindPushMaskInt(regMaskTP maskInt) |
| 365 | { |
| 366 | // Only r0-r12 and lr are supported |
| 367 | assert((maskInt & |
| 368 | ~(RBM_R0 | RBM_R1 | RBM_R2 | RBM_R3 | RBM_R4 | RBM_R5 | RBM_R6 | RBM_R7 | RBM_R8 | RBM_R9 | RBM_R10 | |
| 369 | RBM_R11 | RBM_R12 | RBM_LR)) == 0); |
| 370 | |
| 371 | #if defined(_TARGET_UNIX_) |
| 372 | if (generateCFIUnwindCodes()) |
| 373 | { |
| 374 | // If we are pushing LR, we should give unwind codes in terms of caller's PC |
| 375 | if (maskInt & RBM_LR) |
| 376 | { |
| 377 | maskInt = (maskInt & ~RBM_LR) | RBM_PC; |
| 378 | } |
| 379 | unwindPushPopMaskCFI(maskInt, false); |
| 380 | return; |
| 381 | } |
| 382 | #endif // _TARGET_UNIX_ |
| 383 | |
| 384 | bool useOpsize16 = ((maskInt & (RBM_LOW_REGS | RBM_LR)) == maskInt); // Can PUSH use the 16-bit encoding? |
| 385 | unwindPushPopMaskInt(maskInt, useOpsize16); |
| 386 | } |
| 387 | |
| 388 | void Compiler::unwindPushMaskFloat(regMaskTP maskFloat) |
| 389 | { |
| 390 | // Only floating point registers should be in maskFloat |
| 391 | assert((maskFloat & RBM_ALLFLOAT) == maskFloat); |
| 392 | |
| 393 | #if defined(_TARGET_UNIX_) |
| 394 | if (generateCFIUnwindCodes()) |
| 395 | { |
| 396 | unwindPushPopMaskCFI(maskFloat, true); |
| 397 | return; |
| 398 | } |
| 399 | #endif // _TARGET_UNIX_ |
| 400 | |
| 401 | unwindPushPopMaskFloat(maskFloat); |
| 402 | } |
| 403 | |
| 404 | void Compiler::unwindPopMaskInt(regMaskTP maskInt) |
| 405 | { |
| 406 | #if defined(_TARGET_UNIX_) |
| 407 | if (generateCFIUnwindCodes()) |
| 408 | { |
| 409 | return; |
| 410 | } |
| 411 | #endif // _TARGET_UNIX_ |
| 412 | |
| 413 | // Only r0-r12 and lr and pc are supported (pc is mapped to lr when encoding) |
| 414 | assert((maskInt & |
| 415 | ~(RBM_R0 | RBM_R1 | RBM_R2 | RBM_R3 | RBM_R4 | RBM_R5 | RBM_R6 | RBM_R7 | RBM_R8 | RBM_R9 | RBM_R10 | |
| 416 | RBM_R11 | RBM_R12 | RBM_LR | RBM_PC)) == 0); |
| 417 | |
| 418 | bool useOpsize16 = ((maskInt & (RBM_LOW_REGS | RBM_PC)) == maskInt); // Can POP use the 16-bit encoding? |
| 419 | |
| 420 | // If we are popping PC, then we'll return from the function. In this case, we assume |
| 421 | // the first thing the prolog did was push LR, so give the unwind codes in terms of |
| 422 | // the LR that was pushed. Note that the epilog unwind codes are meant to reverse |
| 423 | // the effect of the prolog. For "pop {pc}", the prolog had "push {lr}", so we need |
| 424 | // an epilog code to model the reverse of that. |
| 425 | if (maskInt & RBM_PC) |
| 426 | { |
| 427 | maskInt = (maskInt & ~RBM_PC) | RBM_LR; |
| 428 | } |
| 429 | unwindPushPopMaskInt(maskInt, useOpsize16); |
| 430 | } |
| 431 | |
| 432 | void Compiler::unwindPopMaskFloat(regMaskTP maskFloat) |
| 433 | { |
| 434 | #if defined(_TARGET_UNIX_) |
| 435 | if (generateCFIUnwindCodes()) |
| 436 | { |
| 437 | return; |
| 438 | } |
| 439 | #endif // _TARGET_UNIX_ |
| 440 | |
| 441 | // Only floating point registers should be in maskFloat |
| 442 | assert((maskFloat & RBM_ALLFLOAT) == maskFloat); |
| 443 | unwindPushPopMaskFloat(maskFloat); |
| 444 | } |
| 445 | |
| 446 | void Compiler::unwindAllocStack(unsigned size) |
| 447 | { |
| 448 | #if defined(_TARGET_UNIX_) |
| 449 | if (generateCFIUnwindCodes()) |
| 450 | { |
| 451 | if (compGeneratingProlog) |
| 452 | { |
| 453 | unwindAllocStackCFI(size); |
| 454 | } |
| 455 | return; |
| 456 | } |
| 457 | #endif // _TARGET_UNIX_ |
| 458 | |
| 459 | UnwindInfo* pu = &funCurrentFunc()->uwi; |
| 460 | |
| 461 | assert(size % 4 == 0); |
| 462 | size /= 4; |
| 463 | |
| 464 | if (size <= 0x7F) |
| 465 | { |
| 466 | // 00-7F : add sp, sp, #X*4 (opsize 16) |
| 467 | pu->AddCode((BYTE)size); |
| 468 | } |
| 469 | else if (size <= 0x3FF) |
| 470 | { |
| 471 | // E8-EB : addw sp, sp, #X*4 (opsize 32) |
| 472 | pu->AddCode(0xE8 | (BYTE)(size >> 8), (BYTE)size); |
| 473 | } |
| 474 | else if (size <= 0xFFFF) |
| 475 | { |
| 476 | // F7 : add sp, sp, #X*4 (opsize 16) |
| 477 | // F9 : add sp, sp, #X*4 (opsize 32) |
| 478 | // |
| 479 | // For large stack size, the most significant bits |
| 480 | // are stored first (and next to the opCode (F9)) per the unwind spec. |
| 481 | unsigned instrSizeInBytes = pu->GetInstructionSize(); |
| 482 | BYTE b1 = (instrSizeInBytes == 2) ? 0xF7 : 0xF9; |
| 483 | pu->AddCode(b1, |
| 484 | (BYTE)(size >> 8), // msb |
| 485 | (BYTE)size); // lsb |
| 486 | } |
| 487 | else |
| 488 | { |
| 489 | // F8 : add sp, sp, #X*4 (opsize 16) |
| 490 | // FA : add sp, sp, #X*4 (opsize 32) |
| 491 | // |
| 492 | // For large stack size, the most significant bits |
| 493 | // are stored first (and next to the opCode (FA)) per the unwind spec. |
| 494 | unsigned instrSizeInBytes = pu->GetInstructionSize(); |
| 495 | BYTE b1 = (instrSizeInBytes == 2) ? 0xF8 : 0xFA; |
| 496 | pu->AddCode(b1, (BYTE)(size >> 16), (BYTE)(size >> 8), (BYTE)size); |
| 497 | } |
| 498 | } |
| 499 | |
| 500 | void Compiler::unwindSetFrameReg(regNumber reg, unsigned offset) |
| 501 | { |
| 502 | #if defined(_TARGET_UNIX_) |
| 503 | if (generateCFIUnwindCodes()) |
| 504 | { |
| 505 | if (compGeneratingProlog) |
| 506 | { |
| 507 | unwindSetFrameRegCFI(reg, offset); |
| 508 | } |
| 509 | return; |
| 510 | } |
| 511 | #endif // _TARGET_UNIX_ |
| 512 | |
| 513 | UnwindInfo* pu = &funCurrentFunc()->uwi; |
| 514 | |
| 515 | // Arm unwind info does not allow offset |
| 516 | assert(offset == 0); |
| 517 | assert(0 <= reg && reg <= 15); |
| 518 | |
| 519 | // C0-CF : mov sp, rX (opsize 16) |
| 520 | pu->AddCode((BYTE)(0xC0 + reg)); |
| 521 | } |
| 522 | |
| 523 | void Compiler::unwindSaveReg(regNumber reg, unsigned offset) |
| 524 | { |
| 525 | unreached(); |
| 526 | } |
| 527 | |
| 528 | void Compiler::unwindBranch16() |
| 529 | { |
| 530 | #if defined(_TARGET_UNIX_) |
| 531 | if (generateCFIUnwindCodes()) |
| 532 | { |
| 533 | return; |
| 534 | } |
| 535 | #endif // _TARGET_UNIX_ |
| 536 | |
| 537 | UnwindInfo* pu = &funCurrentFunc()->uwi; |
| 538 | |
| 539 | // TODO-CQ: need to handle changing the exit code from 0xFF to 0xFD. Currently, this will waste an extra 0xFF at the |
| 540 | // end, automatically added. |
| 541 | pu->AddCode(0xFD); |
| 542 | } |
| 543 | |
| 544 | void Compiler::unwindNop(unsigned codeSizeInBytes) // codeSizeInBytes is 2 or 4 bytes for Thumb2 instruction |
| 545 | { |
| 546 | #if defined(_TARGET_UNIX_) |
| 547 | if (generateCFIUnwindCodes()) |
| 548 | { |
| 549 | return; |
| 550 | } |
| 551 | #endif // _TARGET_UNIX_ |
| 552 | |
| 553 | UnwindInfo* pu = &funCurrentFunc()->uwi; |
| 554 | |
| 555 | #ifdef DEBUG |
| 556 | if (verbose) |
| 557 | { |
| 558 | printf("unwindNop: adding NOP for %d byte instruction\n" , codeSizeInBytes); |
| 559 | } |
| 560 | #endif |
| 561 | |
| 562 | INDEBUG(pu->uwiAddingNOP = true); |
| 563 | |
| 564 | if (codeSizeInBytes == 2) |
| 565 | { |
| 566 | // FB : nop (opsize 16) |
| 567 | pu->AddCode(0xFB); |
| 568 | } |
| 569 | else |
| 570 | { |
| 571 | noway_assert(codeSizeInBytes == 4); |
| 572 | |
| 573 | // FC : nop (opsize 32) |
| 574 | pu->AddCode(0xFC); |
| 575 | } |
| 576 | |
| 577 | INDEBUG(pu->uwiAddingNOP = false); |
| 578 | } |
| 579 | |
| 580 | #endif // defined(_TARGET_ARM_) |
| 581 | |
| 582 | // The instructions between the last captured "current state" and the current instruction |
| 583 | // are in the prolog but have no effect for unwinding. Emit the appropriate NOP unwind codes |
| 584 | // for them. |
| 585 | void Compiler::unwindPadding() |
| 586 | { |
| 587 | #if defined(_TARGET_UNIX_) |
| 588 | if (generateCFIUnwindCodes()) |
| 589 | { |
| 590 | return; |
| 591 | } |
| 592 | #endif // _TARGET_UNIX_ |
| 593 | |
| 594 | UnwindInfo* pu = &funCurrentFunc()->uwi; |
| 595 | genEmitter->emitUnwindNopPadding(pu->GetCurrentEmitterLocation(), this); |
| 596 | } |
| 597 | |
| 598 | // Ask the VM to reserve space for the unwind information for the function and |
| 599 | // all its funclets. |
| 600 | void Compiler::unwindReserve() |
| 601 | { |
| 602 | assert(!compGeneratingProlog); |
| 603 | assert(!compGeneratingEpilog); |
| 604 | |
| 605 | assert(compFuncInfoCount > 0); |
| 606 | for (unsigned funcIdx = 0; funcIdx < compFuncInfoCount; funcIdx++) |
| 607 | { |
| 608 | unwindReserveFunc(funGetFunc(funcIdx)); |
| 609 | } |
| 610 | } |
| 611 | |
| 612 | void Compiler::unwindReserveFunc(FuncInfoDsc* func) |
| 613 | { |
| 614 | BOOL isFunclet = (func->funKind == FUNC_ROOT) ? FALSE : TRUE; |
| 615 | bool funcHasColdSection = false; |
| 616 | |
| 617 | #if defined(_TARGET_UNIX_) |
| 618 | if (generateCFIUnwindCodes()) |
| 619 | { |
| 620 | DWORD unwindCodeBytes = 0; |
| 621 | if (fgFirstColdBlock != nullptr) |
| 622 | { |
| 623 | eeReserveUnwindInfo(isFunclet, true /*isColdCode*/, unwindCodeBytes); |
| 624 | } |
| 625 | unwindCodeBytes = (DWORD)(func->cfiCodes->size() * sizeof(CFI_CODE)); |
| 626 | eeReserveUnwindInfo(isFunclet, false /*isColdCode*/, unwindCodeBytes); |
| 627 | |
| 628 | return; |
| 629 | } |
| 630 | #endif // _TARGET_UNIX_ |
| 631 | |
| 632 | // If there is cold code, split the unwind data between the hot section and the |
| 633 | // cold section. This needs to be done before we split into fragments, as each |
| 634 | // of the hot and cold sections can have multiple fragments. |
| 635 | |
| 636 | if (fgFirstColdBlock != NULL) |
| 637 | { |
| 638 | assert(!isFunclet); // TODO-CQ: support hot/cold splitting with EH |
| 639 | |
| 640 | emitLocation* startLoc; |
| 641 | emitLocation* endLoc; |
| 642 | unwindGetFuncLocations(func, false, &startLoc, &endLoc); |
| 643 | |
| 644 | func->uwiCold = new (this, CMK_UnwindInfo) UnwindInfo(); |
| 645 | func->uwiCold->InitUnwindInfo(this, startLoc, endLoc); |
| 646 | func->uwiCold->HotColdSplitCodes(&func->uwi); |
| 647 | |
| 648 | funcHasColdSection = true; |
| 649 | } |
| 650 | |
| 651 | // First we need to split the function or funclet into fragments that are no larger |
| 652 | // than 512K, so the fragment size will fit in the unwind data "Function Length" field. |
| 653 | // The ARM Exception Data specification "Function Fragments" section describes this. |
| 654 | func->uwi.Split(); |
| 655 | |
| 656 | func->uwi.Reserve(isFunclet, true); |
| 657 | |
| 658 | // After the hot section, split and reserve the cold section |
| 659 | |
| 660 | if (funcHasColdSection) |
| 661 | { |
| 662 | assert(func->uwiCold != NULL); |
| 663 | |
| 664 | func->uwiCold->Split(); |
| 665 | func->uwiCold->Reserve(isFunclet, false); |
| 666 | } |
| 667 | } |
| 668 | |
| 669 | // unwindEmit: Report all the unwind information to the VM. |
| 670 | // Arguments: |
| 671 | // pHotCode: Pointer to the beginning of the memory with the function and funclet hot code |
| 672 | // pColdCode: Pointer to the beginning of the memory with the function and funclet cold code. |
| 673 | |
| 674 | void Compiler::unwindEmit(void* pHotCode, void* pColdCode) |
| 675 | { |
| 676 | assert(compFuncInfoCount > 0); |
| 677 | for (unsigned funcIdx = 0; funcIdx < compFuncInfoCount; funcIdx++) |
| 678 | { |
| 679 | unwindEmitFunc(funGetFunc(funcIdx), pHotCode, pColdCode); |
| 680 | } |
| 681 | } |
| 682 | |
| 683 | void Compiler::unwindEmitFunc(FuncInfoDsc* func, void* pHotCode, void* pColdCode) |
| 684 | { |
| 685 | // Verify that the JIT enum is in sync with the JIT-EE interface enum |
| 686 | static_assert_no_msg(FUNC_ROOT == (FuncKind)CORJIT_FUNC_ROOT); |
| 687 | static_assert_no_msg(FUNC_HANDLER == (FuncKind)CORJIT_FUNC_HANDLER); |
| 688 | static_assert_no_msg(FUNC_FILTER == (FuncKind)CORJIT_FUNC_FILTER); |
| 689 | |
| 690 | #if defined(_TARGET_UNIX_) |
| 691 | if (generateCFIUnwindCodes()) |
| 692 | { |
| 693 | unwindEmitFuncCFI(func, pHotCode, pColdCode); |
| 694 | return; |
| 695 | } |
| 696 | #endif // _TARGET_UNIX_ |
| 697 | |
| 698 | func->uwi.Allocate((CorJitFuncKind)func->funKind, pHotCode, pColdCode, true); |
| 699 | |
| 700 | if (func->uwiCold != NULL) |
| 701 | { |
| 702 | func->uwiCold->Allocate((CorJitFuncKind)func->funKind, pHotCode, pColdCode, false); |
| 703 | } |
| 704 | } |
| 705 | |
| 706 | #if defined(_TARGET_ARM_) |
| 707 | |
| 708 | /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 709 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 710 | XX XX |
| 711 | XX Unwind Info Debug helpers XX |
| 712 | XX XX |
| 713 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 714 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 715 | */ |
| 716 | |
| 717 | #ifdef DEBUG |
| 718 | |
| 719 | // Return the opcode size of an instruction, in bytes, given the first byte of |
| 720 | // its corresponding unwind code. |
| 721 | |
| 722 | unsigned GetOpcodeSizeFromUnwindHeader(BYTE b1) |
| 723 | { |
| 724 | static BYTE s_UnwindOpsize[256] = { |
| 725 | // array of opsizes, in bytes (as specified in the ARM unwind specification) |
| 726 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // 00-0F |
| 727 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // 10-1F |
| 728 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // 20-2F |
| 729 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // 30-3F |
| 730 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // 40-4F |
| 731 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // 50-5F |
| 732 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // 60-6F |
| 733 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // 70-7F |
| 734 | 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, // 80-8F |
| 735 | 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, // 90-9F |
| 736 | 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, // A0-AF |
| 737 | 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, // B0-BF |
| 738 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // C0-CF |
| 739 | 2, 2, 2, 2, 2, 2, 2, 2, 4, 4, 4, 4, 4, 4, 4, 4, // D0-DF |
| 740 | 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 2, 2, 2, 4, // E0-EF |
| 741 | 0, 0, 0, 0, 0, 4, 4, 2, 2, 4, 4, 2, 4, 2, 4, 0 // F0-FF |
| 742 | }; |
| 743 | |
| 744 | BYTE opsize = s_UnwindOpsize[b1]; |
| 745 | assert(opsize == 2 || |
| 746 | opsize == 4); // We shouldn't get a code with no opsize (the 0xFF end code is handled specially) |
| 747 | return opsize; |
| 748 | } |
| 749 | |
| 750 | // Return the size of the unwind code (from 1 to 4 bytes), given the first byte of the unwind bytes |
| 751 | |
| 752 | unsigned GetUnwindSizeFromUnwindHeader(BYTE b1) |
| 753 | { |
| 754 | static BYTE s_UnwindSize[256] = { |
| 755 | // array of unwind sizes, in bytes (as specified in the ARM unwind specification) |
| 756 | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, // 00-0F |
| 757 | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, // 10-1F |
| 758 | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, // 20-2F |
| 759 | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, // 30-3F |
| 760 | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, // 40-4F |
| 761 | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, // 50-5F |
| 762 | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, // 60-6F |
| 763 | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, // 70-7F |
| 764 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // 80-8F |
| 765 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // 90-9F |
| 766 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // A0-AF |
| 767 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // B0-BF |
| 768 | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, // C0-CF |
| 769 | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, // D0-DF |
| 770 | 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, // E0-EF |
| 771 | 1, 1, 1, 1, 1, 2, 2, 3, 4, 3, 4, 1, 1, 1, 1, 1 // F0-FF |
| 772 | }; |
| 773 | |
| 774 | unsigned size = s_UnwindSize[b1]; |
| 775 | assert(1 <= size && size <= 4); |
| 776 | return size; |
| 777 | } |
| 778 | |
| 779 | #endif // DEBUG |
| 780 | |
| 781 | /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 782 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 783 | XX XX |
| 784 | XX Unwind Info Support Classes XX |
| 785 | XX XX |
| 786 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 787 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 788 | */ |
| 789 | |
| 790 | /////////////////////////////////////////////////////////////////////////////// |
| 791 | // |
| 792 | // UnwindCodesBase |
| 793 | // |
| 794 | /////////////////////////////////////////////////////////////////////////////// |
| 795 | |
| 796 | #ifdef DEBUG |
| 797 | |
| 798 | // Walk the prolog codes and calculate the size of the prolog or epilog, in bytes. |
| 799 | // The 0xFD and 0xFE "end + NOP" codes need to be handled differently between |
| 800 | // the prolog and epilog. They count as pure "end" codes in a prolog, but they |
| 801 | // count as 16 and 32 bit NOPs (respectively), as well as an "end", in an epilog. |
| 802 | unsigned UnwindCodesBase::GetCodeSizeFromUnwindCodes(bool isProlog) |
| 803 | { |
| 804 | BYTE* pCodesStart = GetCodes(); |
| 805 | BYTE* pCodes = pCodesStart; |
| 806 | unsigned size = 0; |
| 807 | for (;;) |
| 808 | { |
| 809 | BYTE b1 = *pCodes; |
| 810 | if (b1 >= 0xFD) |
| 811 | { |
| 812 | // 0xFD, 0xFE, 0xFF are "end" codes |
| 813 | |
| 814 | if (!isProlog && (b1 == 0xFD || b1 == 0xFE)) |
| 815 | { |
| 816 | // Count the special "end + NOP" code size in the epilog |
| 817 | size += GetOpcodeSizeFromUnwindHeader(b1); |
| 818 | } |
| 819 | |
| 820 | break; // We hit an "end" code; we're done |
| 821 | } |
| 822 | size += GetOpcodeSizeFromUnwindHeader(b1); |
| 823 | pCodes += GetUnwindSizeFromUnwindHeader(b1); |
| 824 | assert(pCodes - pCodesStart < 256); // 255 is the absolute maximum number of code bytes allowed |
| 825 | } |
| 826 | return size; |
| 827 | } |
| 828 | |
| 829 | #endif // DEBUG |
| 830 | |
| 831 | #endif // defined(_TARGET_ARM_) |
| 832 | |
| 833 | /////////////////////////////////////////////////////////////////////////////// |
| 834 | // |
| 835 | // UnwindPrologCodes |
| 836 | // |
| 837 | /////////////////////////////////////////////////////////////////////////////// |
| 838 | |
| 839 | // We're going to use the prolog codes memory to store the final unwind data. |
| 840 | // Ensure we have enough memory to store everything. If 'epilogBytes' > 0, then |
| 841 | // move the prolog codes so there are 'epilogBytes' bytes after the prolog codes. |
| 842 | // Set the header pointer for future use, adding the header bytes (this pointer |
| 843 | // is updated when a header byte is added), and remember the index that points |
| 844 | // to the beginning of the header. |
| 845 | |
| 846 | void UnwindPrologCodes::SetFinalSize(int , int epilogBytes) |
| 847 | { |
| 848 | #ifdef DEBUG |
| 849 | // We're done adding codes. Check that we didn't accidentally create a bigger prolog. |
| 850 | unsigned codeSize = GetCodeSizeFromUnwindCodes(true); |
| 851 | assert(codeSize <= MAX_PROLOG_SIZE_BYTES); |
| 852 | #endif // DEBUG |
| 853 | |
| 854 | int prologBytes = Size(); |
| 855 | |
| 856 | EnsureSize(headerBytes + prologBytes + epilogBytes + 3); // 3 = padding bytes for alignment |
| 857 | |
| 858 | upcUnwindBlockSlot = upcCodeSlot - headerBytes - epilogBytes; // Index of the first byte of the unwind header |
| 859 | |
| 860 | assert(upcMemSize == upcUnwindBlockSlot + headerBytes + prologBytes + epilogBytes + 3); |
| 861 | |
| 862 | upcHeaderSlot = upcUnwindBlockSlot - 1; // upcHeaderSlot is always incremented before storing |
| 863 | assert(upcHeaderSlot >= -1); |
| 864 | |
| 865 | if (epilogBytes > 0) |
| 866 | { |
| 867 | // The prolog codes that are already at the end of the array need to get moved to the middle, |
| 868 | // with space for the non-matching epilog codes to follow. |
| 869 | |
| 870 | memmove_s(&upcMem[upcUnwindBlockSlot + headerBytes], upcMemSize - (upcUnwindBlockSlot + headerBytes), |
| 871 | &upcMem[upcCodeSlot], prologBytes); |
| 872 | |
| 873 | // Note that the three UWC_END padding bytes still exist at the end of the array. |
| 874 | CLANG_FORMAT_COMMENT_ANCHOR; |
| 875 | |
| 876 | #ifdef DEBUG |
| 877 | // Zero out the epilog codes memory, to ensure we've copied the right bytes. Don't zero the padding bytes. |
| 878 | memset(&upcMem[upcUnwindBlockSlot + headerBytes + prologBytes], 0, epilogBytes); |
| 879 | #endif // DEBUG |
| 880 | |
| 881 | upcEpilogSlot = |
| 882 | upcUnwindBlockSlot + headerBytes + prologBytes; // upcEpilogSlot points to the next epilog location to fill |
| 883 | |
| 884 | // Update upcCodeSlot to point at the new beginning of the prolog codes |
| 885 | upcCodeSlot = upcUnwindBlockSlot + headerBytes; |
| 886 | } |
| 887 | } |
| 888 | |
| 889 | // Add a header word. Header words are added starting at the beginning, in order: first to last. |
| 890 | // This is in contrast to the prolog unwind codes, which are added in reverse order. |
| 891 | void UnwindPrologCodes::(DWORD d) |
| 892 | { |
| 893 | assert(-1 <= upcHeaderSlot); |
| 894 | assert(upcHeaderSlot + 4 < upcCodeSlot); // Don't collide with the unwind codes that are already there! |
| 895 | |
| 896 | // Store it byte-by-byte in little-endian format. We've already ensured there is enough space |
| 897 | // in SetFinalSize(). |
| 898 | upcMem[++upcHeaderSlot] = (BYTE)d; |
| 899 | upcMem[++upcHeaderSlot] = (BYTE)(d >> 8); |
| 900 | upcMem[++upcHeaderSlot] = (BYTE)(d >> 16); |
| 901 | upcMem[++upcHeaderSlot] = (BYTE)(d >> 24); |
| 902 | } |
| 903 | |
| 904 | // AppendEpilog: copy the epilog bytes to the next epilog bytes slot |
| 905 | void UnwindPrologCodes::AppendEpilog(UnwindEpilogInfo* pEpi) |
| 906 | { |
| 907 | assert(upcEpilogSlot != -1); |
| 908 | |
| 909 | int epiSize = pEpi->Size(); |
| 910 | memcpy_s(&upcMem[upcEpilogSlot], upcMemSize - upcEpilogSlot - 3, pEpi->GetCodes(), |
| 911 | epiSize); // -3 to avoid writing to the alignment padding |
| 912 | assert(pEpi->GetStartIndex() == |
| 913 | upcEpilogSlot - upcCodeSlot); // Make sure we copied it where we expected to copy it. |
| 914 | |
| 915 | upcEpilogSlot += epiSize; |
| 916 | assert(upcEpilogSlot <= upcMemSize - 3); |
| 917 | } |
| 918 | |
| 919 | // GetFinalInfo: return a pointer to the final unwind info to hand to the VM, and the size of this info in bytes |
| 920 | void UnwindPrologCodes::GetFinalInfo(/* OUT */ BYTE** ppUnwindBlock, /* OUT */ ULONG* pUnwindBlockSize) |
| 921 | { |
| 922 | assert(upcHeaderSlot + 1 == upcCodeSlot); // We better have filled in the header before asking for the final data! |
| 923 | |
| 924 | *ppUnwindBlock = &upcMem[upcUnwindBlockSlot]; |
| 925 | |
| 926 | // We put 4 'end' codes at the end for padding, so we can ensure we have an |
| 927 | // unwind block that is a multiple of 4 bytes in size. Subtract off three 'end' |
| 928 | // codes (leave one), and then align the size up to a multiple of 4. |
| 929 | *pUnwindBlockSize = AlignUp((UINT)(upcMemSize - upcUnwindBlockSlot - 3), sizeof(DWORD)); |
| 930 | } |
| 931 | |
| 932 | // Do the argument unwind codes match our unwind codes? |
| 933 | // If they don't match, return -1. If they do, return the offset into |
| 934 | // our codes at which they match. Note that this means that the |
| 935 | // argument codes can match a subset of our codes. The subset needs to be at |
| 936 | // the end, for the "end" code to match. |
| 937 | // |
| 938 | // This is similar to UnwindEpilogInfo::Match(). |
| 939 | // |
| 940 | #if defined(_TARGET_ARM_) |
| 941 | // Note that if we wanted to handle 0xFD and 0xFE codes, by converting |
| 942 | // an existing 0xFF code to one of those, we might do that here. |
| 943 | #endif // defined(_TARGET_ARM_) |
| 944 | |
| 945 | int UnwindPrologCodes::Match(UnwindEpilogInfo* pEpi) |
| 946 | { |
| 947 | if (Size() < pEpi->Size()) |
| 948 | { |
| 949 | return -1; |
| 950 | } |
| 951 | |
| 952 | int matchIndex = Size() - pEpi->Size(); |
| 953 | |
| 954 | if (0 == memcmp(GetCodes() + matchIndex, pEpi->GetCodes(), pEpi->Size())) |
| 955 | { |
| 956 | return matchIndex; |
| 957 | } |
| 958 | |
| 959 | return -1; |
| 960 | } |
| 961 | |
| 962 | // Copy the prolog codes from another prolog. The only time this is legal is |
| 963 | // if we are at the initial state and no prolog codes have been added. |
| 964 | // This is used to create the 'phantom' prolog for non-first fragments. |
| 965 | |
| 966 | void UnwindPrologCodes::CopyFrom(UnwindPrologCodes* pCopyFrom) |
| 967 | { |
| 968 | assert(uwiComp == pCopyFrom->uwiComp); |
| 969 | assert(upcMem == upcMemLocal); |
| 970 | assert(upcMemSize == UPC_LOCAL_COUNT); |
| 971 | assert(upcHeaderSlot == -1); |
| 972 | assert(upcEpilogSlot == -1); |
| 973 | |
| 974 | // Copy the codes |
| 975 | EnsureSize(pCopyFrom->upcMemSize); |
| 976 | assert(upcMemSize == pCopyFrom->upcMemSize); |
| 977 | memcpy_s(upcMem, upcMemSize, pCopyFrom->upcMem, pCopyFrom->upcMemSize); |
| 978 | |
| 979 | // Copy the other data |
| 980 | upcCodeSlot = pCopyFrom->upcCodeSlot; |
| 981 | upcHeaderSlot = pCopyFrom->upcHeaderSlot; |
| 982 | upcEpilogSlot = pCopyFrom->upcEpilogSlot; |
| 983 | upcUnwindBlockSlot = pCopyFrom->upcUnwindBlockSlot; |
| 984 | } |
| 985 | |
| 986 | void UnwindPrologCodes::EnsureSize(int requiredSize) |
| 987 | { |
| 988 | if (requiredSize > upcMemSize) |
| 989 | { |
| 990 | // Reallocate, and copy everything to a new array. |
| 991 | |
| 992 | // Choose the next power of two size. This may or may not be the best choice. |
| 993 | noway_assert((requiredSize & 0xC0000000) == 0); // too big! |
| 994 | int newSize; |
| 995 | for (newSize = upcMemSize << 1; newSize < requiredSize; newSize <<= 1) |
| 996 | { |
| 997 | // do nothing |
| 998 | } |
| 999 | |
| 1000 | BYTE* newUnwindCodes = new (uwiComp, CMK_UnwindInfo) BYTE[newSize]; |
| 1001 | memcpy_s(newUnwindCodes + newSize - upcMemSize, upcMemSize, upcMem, |
| 1002 | upcMemSize); // copy the existing data to the end |
| 1003 | #ifdef DEBUG |
| 1004 | // Clear the old unwind codes; nobody should be looking at them |
| 1005 | memset(upcMem, 0xFF, upcMemSize); |
| 1006 | #endif // DEBUG |
| 1007 | upcMem = newUnwindCodes; // we don't free anything that used to be there since we have a no-release allocator |
| 1008 | upcCodeSlot += newSize - upcMemSize; |
| 1009 | upcMemSize = newSize; |
| 1010 | } |
| 1011 | } |
| 1012 | |
| 1013 | #ifdef DEBUG |
| 1014 | void UnwindPrologCodes::Dump(int indent) |
| 1015 | { |
| 1016 | printf("%*sUnwindPrologCodes @0x%08p, size:%d:\n" , indent, "" , dspPtr(this), sizeof(*this)); |
| 1017 | printf("%*s uwiComp: 0x%08p\n" , indent, "" , dspPtr(uwiComp)); |
| 1018 | printf("%*s &upcMemLocal[0]: 0x%08p\n" , indent, "" , dspPtr(&upcMemLocal[0])); |
| 1019 | printf("%*s upcMem: 0x%08p\n" , indent, "" , dspPtr(upcMem)); |
| 1020 | printf("%*s upcMemSize: %d\n" , indent, "" , upcMemSize); |
| 1021 | printf("%*s upcCodeSlot: %d\n" , indent, "" , upcCodeSlot); |
| 1022 | printf("%*s upcHeaderSlot: %d\n" , indent, "" , upcHeaderSlot); |
| 1023 | printf("%*s upcEpilogSlot: %d\n" , indent, "" , upcEpilogSlot); |
| 1024 | printf("%*s upcUnwindBlockSlot: %d\n" , indent, "" , upcUnwindBlockSlot); |
| 1025 | |
| 1026 | if (upcMemSize > 0) |
| 1027 | { |
| 1028 | printf("%*s codes:" , indent, "" ); |
| 1029 | for (int i = 0; i < upcMemSize; i++) |
| 1030 | { |
| 1031 | printf(" %02x" , upcMem[i]); |
| 1032 | if (i == upcCodeSlot) |
| 1033 | printf(" <-C" ); |
| 1034 | else if (i == upcHeaderSlot) |
| 1035 | printf(" <-H" ); |
| 1036 | else if (i == upcEpilogSlot) |
| 1037 | printf(" <-E" ); |
| 1038 | else if (i == upcUnwindBlockSlot) |
| 1039 | printf(" <-U" ); |
| 1040 | } |
| 1041 | printf("\n" ); |
| 1042 | } |
| 1043 | } |
| 1044 | #endif // DEBUG |
| 1045 | |
| 1046 | /////////////////////////////////////////////////////////////////////////////// |
| 1047 | // |
| 1048 | // UnwindEpilogCodes |
| 1049 | // |
| 1050 | /////////////////////////////////////////////////////////////////////////////// |
| 1051 | |
| 1052 | void UnwindEpilogCodes::EnsureSize(int requiredSize) |
| 1053 | { |
| 1054 | if (requiredSize > uecMemSize) |
| 1055 | { |
| 1056 | // Reallocate, and copy everything to a new array. |
| 1057 | |
| 1058 | // Choose the next power of two size. This may or may not be the best choice. |
| 1059 | noway_assert((requiredSize & 0xC0000000) == 0); // too big! |
| 1060 | int newSize; |
| 1061 | for (newSize = uecMemSize << 1; newSize < requiredSize; newSize <<= 1) |
| 1062 | { |
| 1063 | // do nothing |
| 1064 | } |
| 1065 | |
| 1066 | BYTE* newUnwindCodes = new (uwiComp, CMK_UnwindInfo) BYTE[newSize]; |
| 1067 | memcpy_s(newUnwindCodes, newSize, uecMem, uecMemSize); |
| 1068 | #ifdef DEBUG |
| 1069 | // Clear the old unwind codes; nobody should be looking at them |
| 1070 | memset(uecMem, 0xFF, uecMemSize); |
| 1071 | #endif // DEBUG |
| 1072 | uecMem = newUnwindCodes; // we don't free anything that used to be there since we have a no-release allocator |
| 1073 | // uecCodeSlot stays the same |
| 1074 | uecMemSize = newSize; |
| 1075 | } |
| 1076 | } |
| 1077 | |
| 1078 | #ifdef DEBUG |
| 1079 | void UnwindEpilogCodes::Dump(int indent) |
| 1080 | { |
| 1081 | printf("%*sUnwindEpilogCodes @0x%08p, size:%d:\n" , indent, "" , dspPtr(this), sizeof(*this)); |
| 1082 | printf("%*s uwiComp: 0x%08p\n" , indent, "" , dspPtr(uwiComp)); |
| 1083 | printf("%*s &uecMemLocal[0]: 0x%08p\n" , indent, "" , dspPtr(&uecMemLocal[0])); |
| 1084 | printf("%*s uecMem: 0x%08p\n" , indent, "" , dspPtr(uecMem)); |
| 1085 | printf("%*s uecMemSize: %d\n" , indent, "" , uecMemSize); |
| 1086 | printf("%*s uecCodeSlot: %d\n" , indent, "" , uecCodeSlot); |
| 1087 | printf("%*s uecFinalized: %s\n" , indent, "" , dspBool(uecFinalized)); |
| 1088 | |
| 1089 | if (uecMemSize > 0) |
| 1090 | { |
| 1091 | printf("%*s codes:" , indent, "" ); |
| 1092 | for (int i = 0; i < uecMemSize; i++) |
| 1093 | { |
| 1094 | printf(" %02x" , uecMem[i]); |
| 1095 | if (i == uecCodeSlot) |
| 1096 | printf(" <-C" ); // Indicate the current pointer |
| 1097 | } |
| 1098 | printf("\n" ); |
| 1099 | } |
| 1100 | } |
| 1101 | #endif // DEBUG |
| 1102 | |
| 1103 | /////////////////////////////////////////////////////////////////////////////// |
| 1104 | // |
| 1105 | // UnwindEpilogInfo |
| 1106 | // |
| 1107 | /////////////////////////////////////////////////////////////////////////////// |
| 1108 | |
| 1109 | // Do the current unwind codes match those of the argument epilog? |
| 1110 | // If they don't match, return -1. If they do, return the offset into |
| 1111 | // our codes at which the argument codes match. Note that this means that |
| 1112 | // the argument codes can match a subset of our codes. The subset needs to be at |
| 1113 | // the end, for the "end" code to match. |
| 1114 | // |
| 1115 | // Note that if we wanted to handle 0xFD and 0xFE codes, by converting |
| 1116 | // an existing 0xFF code to one of those, we might do that here. |
| 1117 | |
| 1118 | int UnwindEpilogInfo::Match(UnwindEpilogInfo* pEpi) |
| 1119 | { |
| 1120 | if (Matches()) |
| 1121 | { |
| 1122 | // We are already matched to someone else, and won't provide codes to the final layout |
| 1123 | return -1; |
| 1124 | } |
| 1125 | |
| 1126 | if (Size() < pEpi->Size()) |
| 1127 | { |
| 1128 | return -1; |
| 1129 | } |
| 1130 | |
| 1131 | int matchIndex = Size() - pEpi->Size(); |
| 1132 | |
| 1133 | if (0 == memcmp(GetCodes() + matchIndex, pEpi->GetCodes(), pEpi->Size())) |
| 1134 | { |
| 1135 | return matchIndex; |
| 1136 | } |
| 1137 | |
| 1138 | return -1; |
| 1139 | } |
| 1140 | |
| 1141 | void UnwindEpilogInfo::CaptureEmitLocation() |
| 1142 | { |
| 1143 | noway_assert(epiEmitLocation == NULL); // This function is only called once per epilog |
| 1144 | epiEmitLocation = new (uwiComp, CMK_UnwindInfo) emitLocation(); |
| 1145 | epiEmitLocation->CaptureLocation(uwiComp->genEmitter); |
| 1146 | } |
| 1147 | |
| 1148 | void UnwindEpilogInfo::FinalizeOffset() |
| 1149 | { |
| 1150 | epiStartOffset = epiEmitLocation->CodeOffset(uwiComp->genEmitter); |
| 1151 | } |
| 1152 | |
| 1153 | #ifdef DEBUG |
| 1154 | void UnwindEpilogInfo::Dump(int indent) |
| 1155 | { |
| 1156 | printf("%*sUnwindEpilogInfo @0x%08p, size:%d:\n" , indent, "" , dspPtr(this), sizeof(*this)); |
| 1157 | printf("%*s uwiComp: 0x%08p\n" , indent, "" , dspPtr(uwiComp)); |
| 1158 | printf("%*s epiNext: 0x%08p\n" , indent, "" , dspPtr(epiNext)); |
| 1159 | printf("%*s epiEmitLocation: 0x%08p\n" , indent, "" , dspPtr(epiEmitLocation)); |
| 1160 | printf("%*s epiStartOffset: 0x%x\n" , indent, "" , epiStartOffset); |
| 1161 | printf("%*s epiMatches: %s\n" , indent, "" , dspBool(epiMatches)); |
| 1162 | printf("%*s epiStartIndex: %d\n" , indent, "" , epiStartIndex); |
| 1163 | |
| 1164 | epiCodes.Dump(indent + 2); |
| 1165 | } |
| 1166 | #endif // DEBUG |
| 1167 | |
| 1168 | /////////////////////////////////////////////////////////////////////////////// |
| 1169 | // |
| 1170 | // UnwindFragmentInfo |
| 1171 | // |
| 1172 | /////////////////////////////////////////////////////////////////////////////// |
| 1173 | |
| 1174 | UnwindFragmentInfo::UnwindFragmentInfo(Compiler* comp, emitLocation* emitLoc, bool hasPhantomProlog) |
| 1175 | : UnwindBase(comp) |
| 1176 | , ufiNext(NULL) |
| 1177 | , ufiEmitLoc(emitLoc) |
| 1178 | , ufiHasPhantomProlog(hasPhantomProlog) |
| 1179 | , ufiPrologCodes(comp) |
| 1180 | , ufiEpilogFirst(comp) |
| 1181 | , ufiEpilogList(NULL) |
| 1182 | , ufiEpilogLast(NULL) |
| 1183 | , ufiCurCodes(&ufiPrologCodes) |
| 1184 | , ufiSize(0) |
| 1185 | , ufiStartOffset(UFI_ILLEGAL_OFFSET) |
| 1186 | { |
| 1187 | #ifdef DEBUG |
| 1188 | ufiNum = 1; |
| 1189 | ufiInProlog = true; |
| 1190 | ufiInitialized = UFI_INITIALIZED_PATTERN; |
| 1191 | #endif // DEBUG |
| 1192 | } |
| 1193 | |
| 1194 | void UnwindFragmentInfo::FinalizeOffset() |
| 1195 | { |
| 1196 | if (ufiEmitLoc == NULL) |
| 1197 | { |
| 1198 | // NULL emit location means the beginning of the code. This is to handle the first fragment prolog. |
| 1199 | ufiStartOffset = 0; |
| 1200 | } |
| 1201 | else |
| 1202 | { |
| 1203 | ufiStartOffset = ufiEmitLoc->CodeOffset(uwiComp->genEmitter); |
| 1204 | } |
| 1205 | |
| 1206 | for (UnwindEpilogInfo* pEpi = ufiEpilogList; pEpi != NULL; pEpi = pEpi->epiNext) |
| 1207 | { |
| 1208 | pEpi->FinalizeOffset(); |
| 1209 | } |
| 1210 | } |
| 1211 | |
| 1212 | void UnwindFragmentInfo::AddEpilog() |
| 1213 | { |
| 1214 | assert(ufiInitialized == UFI_INITIALIZED_PATTERN); |
| 1215 | |
| 1216 | #ifdef DEBUG |
| 1217 | if (ufiInProlog) |
| 1218 | { |
| 1219 | assert(ufiEpilogList == NULL); |
| 1220 | ufiInProlog = false; |
| 1221 | } |
| 1222 | else |
| 1223 | { |
| 1224 | assert(ufiEpilogList != NULL); |
| 1225 | } |
| 1226 | #endif // DEBUG |
| 1227 | |
| 1228 | // Either allocate a new epilog object, or, for the first one, use the |
| 1229 | // preallocated one that is a member of the UnwindFragmentInfo class. |
| 1230 | |
| 1231 | UnwindEpilogInfo* newepi; |
| 1232 | |
| 1233 | if (ufiEpilogList == NULL) |
| 1234 | { |
| 1235 | // Use the epilog that's in the class already. Be sure to initialize it! |
| 1236 | newepi = ufiEpilogList = &ufiEpilogFirst; |
| 1237 | } |
| 1238 | else |
| 1239 | { |
| 1240 | newepi = new (uwiComp, CMK_UnwindInfo) UnwindEpilogInfo(uwiComp); |
| 1241 | } |
| 1242 | |
| 1243 | // Put the new epilog at the end of the epilog list |
| 1244 | |
| 1245 | if (ufiEpilogLast != NULL) |
| 1246 | { |
| 1247 | ufiEpilogLast->epiNext = newepi; |
| 1248 | } |
| 1249 | |
| 1250 | ufiEpilogLast = newepi; |
| 1251 | |
| 1252 | // What is the starting code offset of the epilog? Store an emitter location |
| 1253 | // so we can ask the emitter later, after codegen. |
| 1254 | |
| 1255 | newepi->CaptureEmitLocation(); |
| 1256 | |
| 1257 | // Put subsequent unwind codes in this new epilog |
| 1258 | |
| 1259 | ufiCurCodes = &newepi->epiCodes; |
| 1260 | } |
| 1261 | |
| 1262 | // Copy the prolog codes from the 'pCopyFrom' fragment. These prolog codes will |
| 1263 | // become 'phantom' prolog codes in this fragment. Note that this fragment should |
| 1264 | // not have any prolog codes currently; it is at the initial state. |
| 1265 | |
| 1266 | void UnwindFragmentInfo::CopyPrologCodes(UnwindFragmentInfo* pCopyFrom) |
| 1267 | { |
| 1268 | ufiPrologCodes.CopyFrom(&pCopyFrom->ufiPrologCodes); |
| 1269 | #ifdef _TARGET_ARM64_ |
| 1270 | ufiPrologCodes.AddCode(UWC_END_C); |
| 1271 | #endif |
| 1272 | } |
| 1273 | |
| 1274 | // Split the epilog codes that currently exist in 'pSplitFrom'. The ones that represent |
| 1275 | // epilogs that start at or after the location represented by 'emitLoc' are removed |
| 1276 | // from 'pSplitFrom' and moved to this fragment. Note that this fragment should not have |
| 1277 | // any epilog codes currently; it is at the initial state. |
| 1278 | |
| 1279 | void UnwindFragmentInfo::SplitEpilogCodes(emitLocation* emitLoc, UnwindFragmentInfo* pSplitFrom) |
| 1280 | { |
| 1281 | UnwindEpilogInfo* pEpiPrev; |
| 1282 | UnwindEpilogInfo* pEpi; |
| 1283 | |
| 1284 | UNATIVE_OFFSET splitOffset = emitLoc->CodeOffset(uwiComp->genEmitter); |
| 1285 | |
| 1286 | for (pEpiPrev = NULL, pEpi = pSplitFrom->ufiEpilogList; pEpi != NULL; pEpiPrev = pEpi, pEpi = pEpi->epiNext) |
| 1287 | { |
| 1288 | pEpi->FinalizeOffset(); // Get the offset of the epilog from the emitter so we can compare it |
| 1289 | if (pEpi->GetStartOffset() >= splitOffset) |
| 1290 | { |
| 1291 | // This epilog and all following epilogs, which must be in order of increasing offsets, |
| 1292 | // get moved to this fragment. |
| 1293 | |
| 1294 | // Splice in the epilogs to this fragment. Set the head of the epilog |
| 1295 | // list to this epilog. |
| 1296 | ufiEpilogList = pEpi; // In this case, don't use 'ufiEpilogFirst' |
| 1297 | ufiEpilogLast = pSplitFrom->ufiEpilogLast; |
| 1298 | |
| 1299 | // Splice out the tail of the list from the 'pSplitFrom' epilog list |
| 1300 | pSplitFrom->ufiEpilogLast = pEpiPrev; |
| 1301 | if (pSplitFrom->ufiEpilogLast == NULL) |
| 1302 | { |
| 1303 | pSplitFrom->ufiEpilogList = NULL; |
| 1304 | } |
| 1305 | else |
| 1306 | { |
| 1307 | pSplitFrom->ufiEpilogLast->epiNext = NULL; |
| 1308 | } |
| 1309 | |
| 1310 | // No more codes should be added once we start splitting |
| 1311 | pSplitFrom->ufiCurCodes = NULL; |
| 1312 | ufiCurCodes = NULL; |
| 1313 | |
| 1314 | break; |
| 1315 | } |
| 1316 | } |
| 1317 | } |
| 1318 | |
| 1319 | // Is this epilog at the end of an unwind fragment? Ask the emitter. |
| 1320 | // Note that we need to know this before all code offsets are finalized, |
| 1321 | // so we can determine whether we can omit an epilog scope word for a |
| 1322 | // single matching epilog. |
| 1323 | |
| 1324 | bool UnwindFragmentInfo::IsAtFragmentEnd(UnwindEpilogInfo* pEpi) |
| 1325 | { |
| 1326 | return uwiComp->genEmitter->emitIsFuncEnd(pEpi->epiEmitLocation, (ufiNext == NULL) ? NULL : ufiNext->ufiEmitLoc); |
| 1327 | } |
| 1328 | |
| 1329 | // Merge the unwind codes as much as possible. |
| 1330 | // This function is called before all offsets are final. |
| 1331 | // Also, compute the size of the final unwind block. Store this |
| 1332 | // and some other data for later, when we actually emit the |
| 1333 | // unwind block. |
| 1334 | |
| 1335 | void UnwindFragmentInfo::MergeCodes() |
| 1336 | { |
| 1337 | assert(ufiInitialized == UFI_INITIALIZED_PATTERN); |
| 1338 | |
| 1339 | unsigned epilogCount = 0; |
| 1340 | unsigned epilogCodeBytes = 0; // The total number of unwind code bytes used by epilogs that don't match the |
| 1341 | // prolog codes |
| 1342 | unsigned epilogIndex = ufiPrologCodes.Size(); // The "Epilog Start Index" for the next non-matching epilog codes |
| 1343 | UnwindEpilogInfo* pEpi; |
| 1344 | |
| 1345 | for (pEpi = ufiEpilogList; pEpi != NULL; pEpi = pEpi->epiNext) |
| 1346 | { |
| 1347 | ++epilogCount; |
| 1348 | |
| 1349 | pEpi->FinalizeCodes(); |
| 1350 | |
| 1351 | // Does this epilog match the prolog? |
| 1352 | // NOTE: for the purpose of matching, we don't handle the 0xFD and 0xFE end codes that allow slightly unequal |
| 1353 | // prolog and epilog codes. |
| 1354 | |
| 1355 | int matchIndex; |
| 1356 | |
| 1357 | matchIndex = ufiPrologCodes.Match(pEpi); |
| 1358 | if (matchIndex != -1) |
| 1359 | { |
| 1360 | pEpi->SetMatches(); |
| 1361 | pEpi->SetStartIndex(matchIndex); // Prolog codes start at zero, so matchIndex is exactly the start index |
| 1362 | } |
| 1363 | else |
| 1364 | { |
| 1365 | // The epilog codes don't match the prolog codes. Do they match any of the epilogs |
| 1366 | // we've seen so far? |
| 1367 | |
| 1368 | bool matched = false; |
| 1369 | for (UnwindEpilogInfo* pEpi2 = ufiEpilogList; pEpi2 != pEpi; pEpi2 = pEpi2->epiNext) |
| 1370 | { |
| 1371 | matchIndex = pEpi2->Match(pEpi); |
| 1372 | if (matchIndex != -1) |
| 1373 | { |
| 1374 | // Use the same epilog index as the one we matched, as it has already been set. |
| 1375 | pEpi->SetMatches(); |
| 1376 | pEpi->SetStartIndex(pEpi2->GetStartIndex() + matchIndex); // We might match somewhere inside pEpi2's |
| 1377 | // codes, in which case matchIndex > 0 |
| 1378 | matched = true; |
| 1379 | break; |
| 1380 | } |
| 1381 | } |
| 1382 | |
| 1383 | if (!matched) |
| 1384 | { |
| 1385 | pEpi->SetStartIndex(epilogIndex); // We'll copy these codes to the next available location |
| 1386 | epilogCodeBytes += pEpi->Size(); |
| 1387 | epilogIndex += pEpi->Size(); |
| 1388 | } |
| 1389 | } |
| 1390 | } |
| 1391 | |
| 1392 | DWORD codeBytes = ufiPrologCodes.Size() + epilogCodeBytes; |
| 1393 | codeBytes = AlignUp(codeBytes, sizeof(DWORD)); |
| 1394 | |
| 1395 | DWORD codeWords = |
| 1396 | codeBytes / sizeof(DWORD); // This is how many words we need to store all the unwind codes in the unwind block |
| 1397 | |
| 1398 | // Do we need the 2nd header word for "Extended Code Words" or "Extended Epilog Count"? |
| 1399 | |
| 1400 | bool needExtendedCodeWordsEpilogCount = |
| 1401 | (codeWords > UW_MAX_CODE_WORDS_COUNT) || (epilogCount > UW_MAX_EPILOG_COUNT); |
| 1402 | |
| 1403 | // How many epilog scope words do we need? |
| 1404 | |
| 1405 | bool setEBit = false; // do we need to set the E bit? |
| 1406 | unsigned epilogScopes = epilogCount; // Note that this could be zero if we have no epilogs! |
| 1407 | |
| 1408 | if (epilogCount == 1) |
| 1409 | { |
| 1410 | assert(ufiEpilogList != NULL); |
| 1411 | assert(ufiEpilogList->epiNext == NULL); |
| 1412 | |
| 1413 | if (ufiEpilogList->Matches() && (ufiEpilogList->GetStartIndex() == 0) && // The match is with the prolog |
| 1414 | !needExtendedCodeWordsEpilogCount && IsAtFragmentEnd(ufiEpilogList)) |
| 1415 | { |
| 1416 | epilogScopes = 0; // Don't need any epilog scope words |
| 1417 | setEBit = true; |
| 1418 | } |
| 1419 | } |
| 1420 | |
| 1421 | DWORD = (1 // Always need first header DWORD |
| 1422 | + (needExtendedCodeWordsEpilogCount ? 1 : 0) // Do we need the 2nd DWORD for Extended Code |
| 1423 | // Words or Extended Epilog Count? |
| 1424 | + epilogScopes // One DWORD per epilog scope, for EBit = 0 |
| 1425 | ) * |
| 1426 | sizeof(DWORD); // convert it to bytes |
| 1427 | |
| 1428 | DWORD finalSize = headerBytes + codeBytes; // Size of actual unwind codes, aligned up to 4-byte words, |
| 1429 | // including end padding if necessary |
| 1430 | |
| 1431 | // Construct the final unwind information. |
| 1432 | |
| 1433 | // We re-use the memory for the prolog unwind codes to construct the full unwind data. If all the epilogs |
| 1434 | // match the prolog, this is easy: we just prepend the header. If there are epilog codes that don't match |
| 1435 | // the prolog, we still use the prolog codes memory, but it's a little more complicated, since the |
| 1436 | // unwind info is ordered as: (a) header, (b) prolog codes, (c) non-matching epilog codes. And, the prolog |
| 1437 | // codes array is filled in from end-to-beginning. So, we compute the size of memory we need, ensure we |
| 1438 | // have that much memory, and then copy the prolog codes to the right place, appending the non-matching |
| 1439 | // epilog codes and prepending the header. |
| 1440 | |
| 1441 | ufiPrologCodes.SetFinalSize(headerBytes, epilogCodeBytes); |
| 1442 | |
| 1443 | if (epilogCodeBytes != 0) |
| 1444 | { |
| 1445 | // We need to copy the epilog code bytes to their final memory location |
| 1446 | |
| 1447 | for (pEpi = ufiEpilogList; pEpi != NULL; pEpi = pEpi->epiNext) |
| 1448 | { |
| 1449 | if (!pEpi->Matches()) |
| 1450 | { |
| 1451 | ufiPrologCodes.AppendEpilog(pEpi); |
| 1452 | } |
| 1453 | } |
| 1454 | } |
| 1455 | |
| 1456 | // Save some data for later |
| 1457 | |
| 1458 | ufiSize = finalSize; |
| 1459 | ufiSetEBit = setEBit; |
| 1460 | ufiNeedExtendedCodeWordsEpilogCount = needExtendedCodeWordsEpilogCount; |
| 1461 | ufiCodeWords = codeWords; |
| 1462 | ufiEpilogScopes = epilogScopes; |
| 1463 | } |
| 1464 | |
| 1465 | // Finalize: Prepare the unwind information for the VM. Compute and prepend the unwind header. |
| 1466 | |
| 1467 | void UnwindFragmentInfo::Finalize(UNATIVE_OFFSET functionLength) |
| 1468 | { |
| 1469 | assert(ufiInitialized == UFI_INITIALIZED_PATTERN); |
| 1470 | |
| 1471 | #ifdef DEBUG |
| 1472 | if (0 && uwiComp->verbose) |
| 1473 | { |
| 1474 | printf("*************** Before fragment #%d finalize\n" , ufiNum); |
| 1475 | Dump(); |
| 1476 | } |
| 1477 | #endif |
| 1478 | |
| 1479 | // Compute the header |
| 1480 | |
| 1481 | #if defined(_TARGET_ARM_) |
| 1482 | noway_assert((functionLength & 1) == 0); |
| 1483 | DWORD headerFunctionLength = functionLength / 2; |
| 1484 | #elif defined(_TARGET_ARM64_) |
| 1485 | noway_assert((functionLength & 3) == 0); |
| 1486 | DWORD = functionLength / 4; |
| 1487 | #endif // _TARGET_ARM64_ |
| 1488 | |
| 1489 | DWORD = 0; // Version of the unwind info is zero. No other version number is currently defined. |
| 1490 | DWORD = 0; // We never generate "exception data", but the VM might add some. |
| 1491 | DWORD ; |
| 1492 | #if defined(_TARGET_ARM_) |
| 1493 | DWORD headerFBit = ufiHasPhantomProlog ? 1 : 0; // Is this data a fragment in the sense of the unwind data |
| 1494 | // specification? That is, do the prolog codes represent a real |
| 1495 | // prolog or not? |
| 1496 | #endif // defined(_TARGET_ARM_) |
| 1497 | DWORD ; // This depends on how we set headerEBit. |
| 1498 | DWORD ; |
| 1499 | DWORD = 0; // This depends on how we set headerEBit. |
| 1500 | DWORD = 0; |
| 1501 | |
| 1502 | if (ufiSetEBit) |
| 1503 | { |
| 1504 | headerEBit = 1; |
| 1505 | headerEpilogCount = ufiEpilogList->GetStartIndex(); // probably zero -- the start of the prolog codes! |
| 1506 | headerCodeWords = ufiCodeWords; |
| 1507 | } |
| 1508 | else |
| 1509 | { |
| 1510 | headerEBit = 0; |
| 1511 | |
| 1512 | if (ufiNeedExtendedCodeWordsEpilogCount) |
| 1513 | { |
| 1514 | headerEpilogCount = 0; |
| 1515 | headerCodeWords = 0; |
| 1516 | headerExtendedEpilogCount = ufiEpilogScopes; |
| 1517 | headerExtendedCodeWords = ufiCodeWords; |
| 1518 | } |
| 1519 | else |
| 1520 | { |
| 1521 | headerEpilogCount = ufiEpilogScopes; |
| 1522 | headerCodeWords = ufiCodeWords; |
| 1523 | } |
| 1524 | } |
| 1525 | |
| 1526 | // Start writing the header |
| 1527 | |
| 1528 | noway_assert(headerFunctionLength <= |
| 1529 | 0x3FFFFU); // We create fragments to prevent this from firing, so if it hits, we have an internal error |
| 1530 | |
| 1531 | if ((headerEpilogCount > UW_MAX_EPILOG_COUNT) || (headerCodeWords > UW_MAX_CODE_WORDS_COUNT)) |
| 1532 | { |
| 1533 | IMPL_LIMITATION("unwind data too large" ); |
| 1534 | } |
| 1535 | |
| 1536 | #if defined(_TARGET_ARM_) |
| 1537 | DWORD header = headerFunctionLength | (headerVers << 18) | (headerXBit << 20) | (headerEBit << 21) | |
| 1538 | (headerFBit << 22) | (headerEpilogCount << 23) | (headerCodeWords << 28); |
| 1539 | #elif defined(_TARGET_ARM64_) |
| 1540 | DWORD = headerFunctionLength | (headerVers << 18) | (headerXBit << 20) | (headerEBit << 21) | |
| 1541 | (headerEpilogCount << 22) | (headerCodeWords << 27); |
| 1542 | #endif // defined(_TARGET_ARM64_) |
| 1543 | |
| 1544 | ufiPrologCodes.AddHeaderWord(header); |
| 1545 | |
| 1546 | // Construct the second header word, if needed |
| 1547 | |
| 1548 | if (ufiNeedExtendedCodeWordsEpilogCount) |
| 1549 | { |
| 1550 | noway_assert(headerEBit == 0); |
| 1551 | noway_assert(headerEpilogCount == 0); |
| 1552 | noway_assert(headerCodeWords == 0); |
| 1553 | noway_assert((headerExtendedEpilogCount > UW_MAX_EPILOG_COUNT) || |
| 1554 | (headerExtendedCodeWords > UW_MAX_CODE_WORDS_COUNT)); |
| 1555 | |
| 1556 | if ((headerExtendedEpilogCount > UW_MAX_EXTENDED_EPILOG_COUNT) || |
| 1557 | (headerExtendedCodeWords > UW_MAX_EXTENDED_CODE_WORDS_COUNT)) |
| 1558 | { |
| 1559 | IMPL_LIMITATION("unwind data too large" ); |
| 1560 | } |
| 1561 | |
| 1562 | DWORD = headerExtendedEpilogCount | (headerExtendedCodeWords << 16); |
| 1563 | |
| 1564 | ufiPrologCodes.AddHeaderWord(header2); |
| 1565 | } |
| 1566 | |
| 1567 | // Construct the epilog scope words, if needed |
| 1568 | |
| 1569 | if (!ufiSetEBit) |
| 1570 | { |
| 1571 | for (UnwindEpilogInfo* pEpi = ufiEpilogList; pEpi != NULL; pEpi = pEpi->epiNext) |
| 1572 | { |
| 1573 | #if defined(_TARGET_ARM_) |
| 1574 | DWORD headerCondition = 0xE; // The epilog is unconditional. We don't have epilogs under the IT instruction. |
| 1575 | #endif // defined(_TARGET_ARM_) |
| 1576 | |
| 1577 | // The epilog must strictly follow the prolog. The prolog is in the first fragment of |
| 1578 | // the hot section. If this epilog is at the start of a fragment, it can't be the |
| 1579 | // first fragment in the hot section. We actually don't know if we're processing |
| 1580 | // the hot or cold section (or a funclet), so we can't distinguish these cases. Thus, |
| 1581 | // we just assert that the epilog starts within the fragment. |
| 1582 | assert(pEpi->GetStartOffset() >= GetStartOffset()); |
| 1583 | |
| 1584 | // We report the offset of an epilog as the offset from the beginning of the function/funclet fragment, |
| 1585 | // NOT the offset from the beginning of the main function. |
| 1586 | DWORD = pEpi->GetStartOffset() - GetStartOffset(); |
| 1587 | |
| 1588 | #if defined(_TARGET_ARM_) |
| 1589 | noway_assert((headerEpilogStartOffset & 1) == 0); |
| 1590 | headerEpilogStartOffset /= 2; // The unwind data stores the actual offset divided by 2 (since the low bit of |
| 1591 | // the actual offset is always zero) |
| 1592 | #elif defined(_TARGET_ARM64_) |
| 1593 | noway_assert((headerEpilogStartOffset & 3) == 0); |
| 1594 | headerEpilogStartOffset /= 4; // The unwind data stores the actual offset divided by 4 (since the low 2 bits |
| 1595 | // of the actual offset is always zero) |
| 1596 | #endif // defined(_TARGET_ARM64_) |
| 1597 | |
| 1598 | DWORD = pEpi->GetStartIndex(); |
| 1599 | |
| 1600 | if ((headerEpilogStartOffset > UW_MAX_EPILOG_START_OFFSET) || |
| 1601 | (headerEpilogStartIndex > UW_MAX_EPILOG_START_INDEX)) |
| 1602 | { |
| 1603 | IMPL_LIMITATION("unwind data too large" ); |
| 1604 | } |
| 1605 | |
| 1606 | #if defined(_TARGET_ARM_) |
| 1607 | DWORD epilogScopeWord = headerEpilogStartOffset | (headerCondition << 20) | (headerEpilogStartIndex << 24); |
| 1608 | #elif defined(_TARGET_ARM64_) |
| 1609 | DWORD epilogScopeWord = headerEpilogStartOffset | (headerEpilogStartIndex << 22); |
| 1610 | #endif // defined(_TARGET_ARM64_) |
| 1611 | |
| 1612 | ufiPrologCodes.AddHeaderWord(epilogScopeWord); |
| 1613 | } |
| 1614 | } |
| 1615 | |
| 1616 | // The unwind code words are already here, following the header, so we're done! |
| 1617 | } |
| 1618 | |
| 1619 | void UnwindFragmentInfo::Reserve(BOOL isFunclet, bool isHotCode) |
| 1620 | { |
| 1621 | assert(isHotCode || !isFunclet); // TODO-CQ: support hot/cold splitting in functions with EH |
| 1622 | |
| 1623 | MergeCodes(); |
| 1624 | |
| 1625 | BOOL isColdCode = isHotCode ? FALSE : TRUE; |
| 1626 | |
| 1627 | ULONG unwindSize = Size(); |
| 1628 | |
| 1629 | #ifdef DEBUG |
| 1630 | if (uwiComp->verbose) |
| 1631 | { |
| 1632 | if (ufiNum != 1) |
| 1633 | printf("reserveUnwindInfo: fragment #%d:\n" , ufiNum); |
| 1634 | } |
| 1635 | #endif |
| 1636 | |
| 1637 | uwiComp->eeReserveUnwindInfo(isFunclet, isColdCode, unwindSize); |
| 1638 | } |
| 1639 | |
| 1640 | // Allocate the unwind info for a fragment with the VM. |
| 1641 | // Arguments: |
| 1642 | // funKind: funclet kind |
| 1643 | // pHotCode: hot section code buffer |
| 1644 | // pColdCode: cold section code buffer |
| 1645 | // funcEndOffset: offset of the end of this function/funclet. Used if this fragment is the last one for a |
| 1646 | // function/funclet. |
| 1647 | // isHotCode: are we allocating the unwind info for the hot code section? |
| 1648 | |
| 1649 | void UnwindFragmentInfo::Allocate( |
| 1650 | CorJitFuncKind funKind, void* pHotCode, void* pColdCode, UNATIVE_OFFSET funcEndOffset, bool isHotCode) |
| 1651 | { |
| 1652 | UNATIVE_OFFSET startOffset; |
| 1653 | UNATIVE_OFFSET endOffset; |
| 1654 | UNATIVE_OFFSET codeSize; |
| 1655 | |
| 1656 | // We don't support hot/cold splitting with EH, so if there is cold code, this |
| 1657 | // better not be a funclet! |
| 1658 | // TODO-CQ: support funclets in cold code |
| 1659 | |
| 1660 | noway_assert(isHotCode || funKind == CORJIT_FUNC_ROOT); |
| 1661 | |
| 1662 | // Compute the final size, and start and end offsets of the fragment |
| 1663 | |
| 1664 | startOffset = GetStartOffset(); |
| 1665 | |
| 1666 | if (ufiNext == NULL) |
| 1667 | { |
| 1668 | // This is the last fragment, so the fragment extends to the end of the function/fragment. |
| 1669 | assert(funcEndOffset != 0); |
| 1670 | endOffset = funcEndOffset; |
| 1671 | } |
| 1672 | else |
| 1673 | { |
| 1674 | // The fragment length is all the code between the beginning of this fragment |
| 1675 | // and the beginning of the next fragment. Note that all fragments have had their |
| 1676 | // offsets computed before any fragment is allocated. |
| 1677 | endOffset = ufiNext->GetStartOffset(); |
| 1678 | } |
| 1679 | |
| 1680 | assert(endOffset > startOffset); |
| 1681 | codeSize = endOffset - startOffset; |
| 1682 | |
| 1683 | // Finalize the fragment unwind block to hand to the VM |
| 1684 | |
| 1685 | Finalize(codeSize); |
| 1686 | |
| 1687 | // Get the final unwind information and hand it to the VM |
| 1688 | |
| 1689 | ULONG unwindBlockSize; |
| 1690 | BYTE* pUnwindBlock; |
| 1691 | |
| 1692 | GetFinalInfo(&pUnwindBlock, &unwindBlockSize); |
| 1693 | |
| 1694 | #ifdef DEBUG |
| 1695 | if (uwiComp->opts.dspUnwind) |
| 1696 | { |
| 1697 | DumpUnwindInfo(uwiComp, isHotCode, startOffset, endOffset, pUnwindBlock, unwindBlockSize); |
| 1698 | } |
| 1699 | #endif // DEBUG |
| 1700 | |
| 1701 | // Adjust for cold or hot code: |
| 1702 | // 1. The VM doesn't want the cold code pointer unless this is cold code. |
| 1703 | // 2. The startOffset and endOffset need to be from the base of the hot section for hot code |
| 1704 | // and from the base of the cold section for cold code |
| 1705 | |
| 1706 | if (isHotCode) |
| 1707 | { |
| 1708 | assert(endOffset <= uwiComp->info.compTotalHotCodeSize); |
| 1709 | pColdCode = NULL; |
| 1710 | } |
| 1711 | else |
| 1712 | { |
| 1713 | assert(startOffset >= uwiComp->info.compTotalHotCodeSize); |
| 1714 | startOffset -= uwiComp->info.compTotalHotCodeSize; |
| 1715 | endOffset -= uwiComp->info.compTotalHotCodeSize; |
| 1716 | } |
| 1717 | |
| 1718 | #ifdef DEBUG |
| 1719 | if (uwiComp->verbose) |
| 1720 | { |
| 1721 | if (ufiNum != 1) |
| 1722 | printf("unwindEmit: fragment #%d:\n" , ufiNum); |
| 1723 | } |
| 1724 | #endif // DEBUG |
| 1725 | |
| 1726 | uwiComp->eeAllocUnwindInfo((BYTE*)pHotCode, (BYTE*)pColdCode, startOffset, endOffset, unwindBlockSize, pUnwindBlock, |
| 1727 | funKind); |
| 1728 | } |
| 1729 | |
| 1730 | #ifdef DEBUG |
| 1731 | void UnwindFragmentInfo::Dump(int indent) |
| 1732 | { |
| 1733 | unsigned count; |
| 1734 | UnwindEpilogInfo* pEpi; |
| 1735 | |
| 1736 | count = 0; |
| 1737 | for (pEpi = ufiEpilogList; pEpi != NULL; pEpi = pEpi->epiNext) |
| 1738 | { |
| 1739 | ++count; |
| 1740 | } |
| 1741 | |
| 1742 | printf("%*sUnwindFragmentInfo #%d, @0x%08p, size:%d:\n" , indent, "" , ufiNum, dspPtr(this), sizeof(*this)); |
| 1743 | printf("%*s uwiComp: 0x%08p\n" , indent, "" , dspPtr(uwiComp)); |
| 1744 | printf("%*s ufiNext: 0x%08p\n" , indent, "" , dspPtr(ufiNext)); |
| 1745 | printf("%*s ufiEmitLoc: 0x%08p\n" , indent, "" , dspPtr(ufiEmitLoc)); |
| 1746 | printf("%*s ufiHasPhantomProlog: %s\n" , indent, "" , dspBool(ufiHasPhantomProlog)); |
| 1747 | printf("%*s %d epilog%s\n" , indent, "" , count, (count != 1) ? "s" : "" ); |
| 1748 | printf("%*s ufiEpilogList: 0x%08p\n" , indent, "" , dspPtr(ufiEpilogList)); |
| 1749 | printf("%*s ufiEpilogLast: 0x%08p\n" , indent, "" , dspPtr(ufiEpilogLast)); |
| 1750 | printf("%*s ufiCurCodes: 0x%08p\n" , indent, "" , dspPtr(ufiCurCodes)); |
| 1751 | printf("%*s ufiSize: %u\n" , indent, "" , ufiSize); |
| 1752 | printf("%*s ufiSetEBit: %s\n" , indent, "" , dspBool(ufiSetEBit)); |
| 1753 | printf("%*s ufiNeedExtendedCodeWordsEpilogCount: %s\n" , indent, "" , dspBool(ufiNeedExtendedCodeWordsEpilogCount)); |
| 1754 | printf("%*s ufiCodeWords: %u\n" , indent, "" , ufiCodeWords); |
| 1755 | printf("%*s ufiEpilogScopes: %u\n" , indent, "" , ufiEpilogScopes); |
| 1756 | printf("%*s ufiStartOffset: 0x%x\n" , indent, "" , ufiStartOffset); |
| 1757 | printf("%*s ufiInProlog: %s\n" , indent, "" , dspBool(ufiInProlog)); |
| 1758 | printf("%*s ufiInitialized: 0x%08x\n" , indent, "" , ufiInitialized); |
| 1759 | |
| 1760 | ufiPrologCodes.Dump(indent + 2); |
| 1761 | |
| 1762 | for (pEpi = ufiEpilogList; pEpi != NULL; pEpi = pEpi->epiNext) |
| 1763 | { |
| 1764 | pEpi->Dump(indent + 2); |
| 1765 | } |
| 1766 | } |
| 1767 | #endif // DEBUG |
| 1768 | |
| 1769 | /////////////////////////////////////////////////////////////////////////////// |
| 1770 | // |
| 1771 | // UnwindInfo |
| 1772 | // |
| 1773 | /////////////////////////////////////////////////////////////////////////////// |
| 1774 | |
| 1775 | void UnwindInfo::InitUnwindInfo(Compiler* comp, emitLocation* startLoc, emitLocation* endLoc) |
| 1776 | { |
| 1777 | uwiComp = comp; |
| 1778 | |
| 1779 | // The first fragment is a member of UnwindInfo, so it doesn't need to be allocated. |
| 1780 | // However, its constructor needs to be explicitly called, since the constructor for |
| 1781 | // UnwindInfo is not called. |
| 1782 | |
| 1783 | uwiFragmentFirst.UnwindFragmentInfo::UnwindFragmentInfo(comp, startLoc, false); |
| 1784 | |
| 1785 | uwiFragmentLast = &uwiFragmentFirst; |
| 1786 | |
| 1787 | uwiEndLoc = endLoc; |
| 1788 | |
| 1789 | // Allocate an emitter location object. It is initialized to something |
| 1790 | // invalid: it has a null 'ig' that needs to get set before it can be used. |
| 1791 | // Note that when we create an UnwindInfo for the cold section, this never |
| 1792 | // gets initialized with anything useful, since we never add unwind codes |
| 1793 | // to the cold section; we simply distribute the existing (previously added) codes. |
| 1794 | uwiCurLoc = new (uwiComp, CMK_UnwindInfo) emitLocation(); |
| 1795 | |
| 1796 | #ifdef DEBUG |
| 1797 | uwiInitialized = UWI_INITIALIZED_PATTERN; |
| 1798 | uwiAddingNOP = false; |
| 1799 | #endif // DEBUG |
| 1800 | } |
| 1801 | |
| 1802 | // Split the unwind codes in 'puwi' into those that are in the hot section (leave them in 'puwi') |
| 1803 | // and those that are in the cold section (move them to 'this'). There is exactly one fragment |
| 1804 | // in each UnwindInfo; the fragments haven't been split for size, yet. |
| 1805 | |
| 1806 | void UnwindInfo::HotColdSplitCodes(UnwindInfo* puwi) |
| 1807 | { |
| 1808 | // Ensure that there is exactly a single fragment in both the hot and the cold sections |
| 1809 | assert(&uwiFragmentFirst == uwiFragmentLast); |
| 1810 | assert(&puwi->uwiFragmentFirst == puwi->uwiFragmentLast); |
| 1811 | assert(uwiFragmentLast->ufiNext == NULL); |
| 1812 | assert(puwi->uwiFragmentLast->ufiNext == NULL); |
| 1813 | |
| 1814 | // The real prolog is in the hot section, so this, cold, section has a phantom prolog |
| 1815 | uwiFragmentLast->ufiHasPhantomProlog = true; |
| 1816 | uwiFragmentLast->CopyPrologCodes(puwi->uwiFragmentLast); |
| 1817 | |
| 1818 | // Now split the epilog codes |
| 1819 | uwiFragmentLast->SplitEpilogCodes(uwiFragmentLast->ufiEmitLoc, puwi->uwiFragmentLast); |
| 1820 | } |
| 1821 | |
| 1822 | // Split the function or funclet into fragments that are no larger than 512K, |
| 1823 | // so the fragment size will fit in the unwind data "Function Length" field. |
| 1824 | // The ARM Exception Data specification "Function Fragments" section describes this. |
| 1825 | // We split the function so that it is no larger than 512K bytes, or the value of |
| 1826 | // the COMPlus_JitSplitFunctionSize value, if defined (and smaller). We must determine |
| 1827 | // how to split the function/funclet before we issue the instructions, so we can |
| 1828 | // reserve the unwind space with the VM. The instructions issued may shrink (but not |
| 1829 | // expand!) during issuing (although this is extremely rare in any case, and may not |
| 1830 | // actually occur on ARM), so we don't finalize actual sizes or offsets. |
| 1831 | // |
| 1832 | // ARM64 has very similar limitations, except functions can be up to 1MB. TODO-ARM64-Bug?: make sure this works! |
| 1833 | // |
| 1834 | // We don't split any prolog or epilog. Ideally, we might not split an instruction, |
| 1835 | // although that doesn't matter because the unwind at any point would still be |
| 1836 | // well-defined. |
| 1837 | |
| 1838 | void UnwindInfo::Split() |
| 1839 | { |
| 1840 | UNATIVE_OFFSET maxFragmentSize; // The maximum size of a code fragment in bytes |
| 1841 | |
| 1842 | maxFragmentSize = UW_MAX_FRAGMENT_SIZE_BYTES; |
| 1843 | |
| 1844 | #ifdef DEBUG |
| 1845 | // Consider COMPlus_JitSplitFunctionSize |
| 1846 | unsigned splitFunctionSize = (unsigned)JitConfig.JitSplitFunctionSize(); |
| 1847 | |
| 1848 | if (splitFunctionSize != 0) |
| 1849 | if (splitFunctionSize < maxFragmentSize) |
| 1850 | maxFragmentSize = splitFunctionSize; |
| 1851 | #endif // DEBUG |
| 1852 | |
| 1853 | // Now, there should be exactly one fragment. |
| 1854 | |
| 1855 | assert(uwiFragmentLast != NULL); |
| 1856 | assert(uwiFragmentLast == &uwiFragmentFirst); |
| 1857 | assert(uwiFragmentLast->ufiNext == NULL); |
| 1858 | |
| 1859 | // Find the code size of this function/funclet. |
| 1860 | |
| 1861 | UNATIVE_OFFSET startOffset; |
| 1862 | UNATIVE_OFFSET endOffset; |
| 1863 | UNATIVE_OFFSET codeSize; |
| 1864 | |
| 1865 | if (uwiFragmentLast->ufiEmitLoc == NULL) |
| 1866 | { |
| 1867 | // NULL emit location means the beginning of the code. This is to handle the first fragment prolog. |
| 1868 | startOffset = 0; |
| 1869 | } |
| 1870 | else |
| 1871 | { |
| 1872 | startOffset = uwiFragmentLast->ufiEmitLoc->CodeOffset(uwiComp->genEmitter); |
| 1873 | } |
| 1874 | |
| 1875 | if (uwiEndLoc == NULL) |
| 1876 | { |
| 1877 | // Note that compTotalHotCodeSize and compTotalColdCodeSize are computed before issuing instructions |
| 1878 | // from the emitter instruction group offsets, and will be accurate unless the issued code shrinks. |
| 1879 | // compNativeCodeSize is precise, but is only set after instructions are issued, which is too late |
| 1880 | // for us, since we need to decide how many fragments we need before the code memory is allocated |
| 1881 | // (which is before instruction issuing). |
| 1882 | UNATIVE_OFFSET estimatedTotalCodeSize = |
| 1883 | uwiComp->info.compTotalHotCodeSize + uwiComp->info.compTotalColdCodeSize; |
| 1884 | assert(estimatedTotalCodeSize != 0); |
| 1885 | endOffset = estimatedTotalCodeSize; |
| 1886 | } |
| 1887 | else |
| 1888 | { |
| 1889 | endOffset = uwiEndLoc->CodeOffset(uwiComp->genEmitter); |
| 1890 | } |
| 1891 | |
| 1892 | assert(endOffset > startOffset); // there better be at least 1 byte of code |
| 1893 | codeSize = endOffset - startOffset; |
| 1894 | |
| 1895 | // Now that we know the code size for this section (main function hot or cold, or funclet), |
| 1896 | // figure out how many fragments we're going to need. |
| 1897 | |
| 1898 | UNATIVE_OFFSET numberOfFragments = (codeSize + maxFragmentSize - 1) / maxFragmentSize; // round up |
| 1899 | assert(numberOfFragments > 0); |
| 1900 | |
| 1901 | if (numberOfFragments == 1) |
| 1902 | { |
| 1903 | // No need to split; we're done |
| 1904 | return; |
| 1905 | } |
| 1906 | |
| 1907 | // Now, we're going to commit to splitting the function into "numberOfFragments" fragments, |
| 1908 | // for the purpose of unwind information. We need to do the actual splits so we can figure out |
| 1909 | // the size of each piece of unwind data for the call to reserveUnwindInfo(). We won't know |
| 1910 | // the actual offsets of the splits since we haven't issued the instructions yet, so store |
| 1911 | // an emitter location instead of an offset, and "finalize" the offset in the unwindEmit() phase, |
| 1912 | // like we do for the function length and epilog offsets. |
| 1913 | CLANG_FORMAT_COMMENT_ANCHOR; |
| 1914 | |
| 1915 | #ifdef DEBUG |
| 1916 | if (uwiComp->verbose) |
| 1917 | { |
| 1918 | printf("Split unwind info into %d fragments (function/funclet size: %d, maximum fragment size: %d)\n" , |
| 1919 | numberOfFragments, codeSize, maxFragmentSize); |
| 1920 | } |
| 1921 | #endif // DEBUG |
| 1922 | |
| 1923 | // Call the emitter to do the split, and call us back for every split point it chooses. |
| 1924 | uwiComp->genEmitter->emitSplit(uwiFragmentLast->ufiEmitLoc, uwiEndLoc, maxFragmentSize, (void*)this, |
| 1925 | EmitSplitCallback); |
| 1926 | |
| 1927 | #ifdef DEBUG |
| 1928 | // Did the emitter split the function/funclet into as many fragments as we asked for? |
| 1929 | // It might be fewer if the COMPlus_JitSplitFunctionSize was used, but it better not |
| 1930 | // be fewer if we're splitting into 512K blocks! |
| 1931 | |
| 1932 | unsigned fragCount = 0; |
| 1933 | for (UnwindFragmentInfo* pFrag = &uwiFragmentFirst; pFrag != NULL; pFrag = pFrag->ufiNext) |
| 1934 | { |
| 1935 | ++fragCount; |
| 1936 | } |
| 1937 | if (fragCount < numberOfFragments) |
| 1938 | { |
| 1939 | if (uwiComp->verbose) |
| 1940 | { |
| 1941 | printf("WARNING: asked the emitter for %d fragments, but only got %d\n" , numberOfFragments, fragCount); |
| 1942 | } |
| 1943 | |
| 1944 | // If this fires, then we split into fewer fragments than we asked for, and we are using |
| 1945 | // the default, unwind-data-defined 512K maximum fragment size. We won't be able to fit |
| 1946 | // this fragment into the unwind data! If you set COMPlus_JitSplitFunctionSize to something |
| 1947 | // small, we might not be able to split into as many fragments as asked for, because we |
| 1948 | // can't split prologs or epilogs. |
| 1949 | assert(maxFragmentSize != UW_MAX_FRAGMENT_SIZE_BYTES); |
| 1950 | } |
| 1951 | #endif // DEBUG |
| 1952 | } |
| 1953 | |
| 1954 | /*static*/ void UnwindInfo::EmitSplitCallback(void* context, emitLocation* emitLoc) |
| 1955 | { |
| 1956 | UnwindInfo* puwi = (UnwindInfo*)context; |
| 1957 | puwi->AddFragment(emitLoc); |
| 1958 | } |
| 1959 | |
| 1960 | // Reserve space for the unwind info for all fragments |
| 1961 | |
| 1962 | void UnwindInfo::Reserve(BOOL isFunclet, bool isHotCode) |
| 1963 | { |
| 1964 | assert(uwiInitialized == UWI_INITIALIZED_PATTERN); |
| 1965 | assert(isHotCode || !isFunclet); |
| 1966 | |
| 1967 | for (UnwindFragmentInfo* pFrag = &uwiFragmentFirst; pFrag != NULL; pFrag = pFrag->ufiNext) |
| 1968 | { |
| 1969 | pFrag->Reserve(isFunclet, isHotCode); |
| 1970 | } |
| 1971 | } |
| 1972 | |
| 1973 | // Allocate and populate VM unwind info for all fragments |
| 1974 | |
| 1975 | void UnwindInfo::Allocate(CorJitFuncKind funKind, void* pHotCode, void* pColdCode, bool isHotCode) |
| 1976 | { |
| 1977 | assert(uwiInitialized == UWI_INITIALIZED_PATTERN); |
| 1978 | |
| 1979 | UnwindFragmentInfo* pFrag; |
| 1980 | |
| 1981 | // First, finalize all the offsets (the location of the beginning of fragments, and epilogs), |
| 1982 | // so a fragment can use the finalized offset of the subsequent fragment to determine its code size. |
| 1983 | |
| 1984 | UNATIVE_OFFSET endOffset; |
| 1985 | |
| 1986 | if (uwiEndLoc == NULL) |
| 1987 | { |
| 1988 | assert(uwiComp->info.compNativeCodeSize != 0); |
| 1989 | endOffset = uwiComp->info.compNativeCodeSize; |
| 1990 | } |
| 1991 | else |
| 1992 | { |
| 1993 | endOffset = uwiEndLoc->CodeOffset(uwiComp->genEmitter); |
| 1994 | } |
| 1995 | |
| 1996 | for (pFrag = &uwiFragmentFirst; pFrag != NULL; pFrag = pFrag->ufiNext) |
| 1997 | { |
| 1998 | pFrag->FinalizeOffset(); |
| 1999 | } |
| 2000 | |
| 2001 | for (pFrag = &uwiFragmentFirst; pFrag != NULL; pFrag = pFrag->ufiNext) |
| 2002 | { |
| 2003 | pFrag->Allocate(funKind, pHotCode, pColdCode, endOffset, isHotCode); |
| 2004 | } |
| 2005 | } |
| 2006 | |
| 2007 | void UnwindInfo::AddEpilog() |
| 2008 | { |
| 2009 | assert(uwiInitialized == UWI_INITIALIZED_PATTERN); |
| 2010 | assert(uwiFragmentLast != NULL); |
| 2011 | uwiFragmentLast->AddEpilog(); |
| 2012 | CaptureLocation(); |
| 2013 | } |
| 2014 | |
| 2015 | #if defined(_TARGET_ARM_) |
| 2016 | |
| 2017 | unsigned UnwindInfo::GetInstructionSize() |
| 2018 | { |
| 2019 | assert(uwiInitialized == UWI_INITIALIZED_PATTERN); |
| 2020 | return uwiComp->genEmitter->emitGetInstructionSize(uwiCurLoc); |
| 2021 | } |
| 2022 | |
| 2023 | #endif // defined(_TARGET_ARM_) |
| 2024 | |
| 2025 | void UnwindInfo::CaptureLocation() |
| 2026 | { |
| 2027 | assert(uwiInitialized == UWI_INITIALIZED_PATTERN); |
| 2028 | assert(uwiCurLoc != NULL); |
| 2029 | uwiCurLoc->CaptureLocation(uwiComp->genEmitter); |
| 2030 | } |
| 2031 | |
| 2032 | void UnwindInfo::AddFragment(emitLocation* emitLoc) |
| 2033 | { |
| 2034 | assert(uwiInitialized == UWI_INITIALIZED_PATTERN); |
| 2035 | assert(uwiFragmentLast != NULL); |
| 2036 | |
| 2037 | UnwindFragmentInfo* newFrag = new (uwiComp, CMK_UnwindInfo) UnwindFragmentInfo(uwiComp, emitLoc, true); |
| 2038 | |
| 2039 | #ifdef DEBUG |
| 2040 | newFrag->ufiNum = uwiFragmentLast->ufiNum + 1; |
| 2041 | #endif // DEBUG |
| 2042 | |
| 2043 | newFrag->CopyPrologCodes(&uwiFragmentFirst); |
| 2044 | newFrag->SplitEpilogCodes(emitLoc, uwiFragmentLast); |
| 2045 | |
| 2046 | // Link the new fragment in at the end of the fragment list |
| 2047 | uwiFragmentLast->ufiNext = newFrag; |
| 2048 | uwiFragmentLast = newFrag; |
| 2049 | } |
| 2050 | |
| 2051 | #ifdef DEBUG |
| 2052 | |
| 2053 | #if defined(_TARGET_ARM_) |
| 2054 | |
| 2055 | // Given the first byte of the unwind code, check that its opsize matches |
| 2056 | // the last instruction added in the emitter. |
| 2057 | void UnwindInfo::CheckOpsize(BYTE b1) |
| 2058 | { |
| 2059 | // Adding NOP padding goes through the same path, but doesn't update the location to indicate |
| 2060 | // the correct location of the instruction for which we are adding a NOP, so just skip the |
| 2061 | // assert. Should be ok, because the emitter is telling us the size of the instruction for |
| 2062 | // which we are adding the NOP. |
| 2063 | if (uwiAddingNOP) |
| 2064 | return; |
| 2065 | |
| 2066 | unsigned opsizeInBytes = GetOpcodeSizeFromUnwindHeader(b1); |
| 2067 | unsigned instrSizeInBytes = GetInstructionSize(); |
| 2068 | assert(opsizeInBytes == instrSizeInBytes); |
| 2069 | } |
| 2070 | |
| 2071 | #endif // defined(_TARGET_ARM_) |
| 2072 | |
| 2073 | void UnwindInfo::Dump(bool isHotCode, int indent) |
| 2074 | { |
| 2075 | unsigned count; |
| 2076 | UnwindFragmentInfo* pFrag; |
| 2077 | |
| 2078 | count = 0; |
| 2079 | for (pFrag = &uwiFragmentFirst; pFrag != NULL; pFrag = pFrag->ufiNext) |
| 2080 | { |
| 2081 | ++count; |
| 2082 | } |
| 2083 | |
| 2084 | printf("%*sUnwindInfo %s@0x%08p, size:%d:\n" , indent, "" , isHotCode ? "" : "COLD " , dspPtr(this), sizeof(*this)); |
| 2085 | printf("%*s uwiComp: 0x%08p\n" , indent, "" , dspPtr(uwiComp)); |
| 2086 | printf("%*s %d fragment%s\n" , indent, "" , count, (count != 1) ? "s" : "" ); |
| 2087 | printf("%*s uwiFragmentLast: 0x%08p\n" , indent, "" , dspPtr(uwiFragmentLast)); |
| 2088 | printf("%*s uwiEndLoc: 0x%08p\n" , indent, "" , dspPtr(uwiEndLoc)); |
| 2089 | printf("%*s uwiInitialized: 0x%08x\n" , indent, "" , uwiInitialized); |
| 2090 | |
| 2091 | for (pFrag = &uwiFragmentFirst; pFrag != NULL; pFrag = pFrag->ufiNext) |
| 2092 | { |
| 2093 | pFrag->Dump(indent + 2); |
| 2094 | } |
| 2095 | } |
| 2096 | |
| 2097 | #endif // DEBUG |
| 2098 | |
| 2099 | #if defined(_TARGET_ARM_) |
| 2100 | |
| 2101 | /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 2102 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 2103 | XX XX |
| 2104 | XX Debug dumpers XX |
| 2105 | XX XX |
| 2106 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 2107 | XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX |
| 2108 | */ |
| 2109 | |
| 2110 | #ifdef DEBUG |
| 2111 | |
| 2112 | // start is 0-based index from LSB, length is number of bits |
| 2113 | DWORD ExtractBits(DWORD dw, DWORD start, DWORD length) |
| 2114 | { |
| 2115 | return (dw >> start) & ((1 << length) - 1); |
| 2116 | } |
| 2117 | |
| 2118 | // Dump an integer register set. 'x' is an array of bits where bit 0 = r0, bit 1 = r1, etc. |
| 2119 | // The highest register considered is r12. |
| 2120 | // If 'lr' is non-zero, the "lr" register is emitted last. |
| 2121 | // Returns the number of characters printed. |
| 2122 | DWORD DumpIntRegSet(DWORD x, DWORD lr) |
| 2123 | { |
| 2124 | assert(x != 0 || lr != 0); // we must have one |
| 2125 | assert((x & 0xE000) == 0); // don't handle r13 (sp), r14 (lr), r15 (pc) in 'x' |
| 2126 | DWORD printed = 0; |
| 2127 | |
| 2128 | printf("{" ); |
| 2129 | ++printed; |
| 2130 | bool first = true; |
| 2131 | DWORD bitMask = 1; |
| 2132 | for (DWORD bitNum = 0; bitNum < 12; bitNum++) |
| 2133 | { |
| 2134 | if (x & bitMask) |
| 2135 | { |
| 2136 | if (!first) |
| 2137 | { |
| 2138 | printf("," ); |
| 2139 | ++printed; |
| 2140 | } |
| 2141 | printf("r%u" , bitNum); |
| 2142 | printed += (bitNum < 10) ? 2 : 3; |
| 2143 | first = false; |
| 2144 | } |
| 2145 | bitMask <<= 1; |
| 2146 | } |
| 2147 | if (lr) |
| 2148 | { |
| 2149 | if (!first) |
| 2150 | { |
| 2151 | printf("," ); |
| 2152 | ++printed; |
| 2153 | } |
| 2154 | printf("lr" ); |
| 2155 | printed += 2; |
| 2156 | } |
| 2157 | printf("}" ); |
| 2158 | ++printed; |
| 2159 | |
| 2160 | return printed; |
| 2161 | } |
| 2162 | |
| 2163 | // Dump a register set range from register 'start' to register 'end'. |
| 2164 | // rtype should be "r" or "d" to indicate register type. |
| 2165 | // If 'lr' is non-zero, the "lr" register is emitted last. (Note that |
| 2166 | // 'lr' should be zero for rtype == "d".) |
| 2167 | // Returns the number of characters printed. |
| 2168 | DWORD DumpRegSetRange(const char* const rtype, DWORD start, DWORD end, DWORD lr) |
| 2169 | { |
| 2170 | assert(start <= end); |
| 2171 | DWORD printed = 0; |
| 2172 | DWORD rtypeLen = (DWORD)strlen(rtype); |
| 2173 | |
| 2174 | printf("{" ); |
| 2175 | ++printed; |
| 2176 | bool first = true; |
| 2177 | for (DWORD reg = start; reg <= end; reg++) |
| 2178 | { |
| 2179 | if (!first) |
| 2180 | { |
| 2181 | printf("," ); |
| 2182 | ++printed; |
| 2183 | } |
| 2184 | printf("%s%u" , rtype, reg); |
| 2185 | printed += rtypeLen + ((reg < 10) ? 1 : 2); |
| 2186 | first = false; |
| 2187 | } |
| 2188 | if (lr) |
| 2189 | { |
| 2190 | assert(!first); // If 'lr' is set, it can't be first, since we require a non-empty range |
| 2191 | printf(",lr" ); |
| 2192 | printed += 3; |
| 2193 | } |
| 2194 | printf("}" ); |
| 2195 | ++printed; |
| 2196 | |
| 2197 | return printed; |
| 2198 | } |
| 2199 | |
| 2200 | // Dump the opsize. |
| 2201 | // Returns the number of characters printed. |
| 2202 | DWORD DumpOpsize(DWORD padding, DWORD opsize) |
| 2203 | { |
| 2204 | if (padding > 100) // underflow? |
| 2205 | padding = 4; |
| 2206 | DWORD printed = padding; |
| 2207 | for (; padding > 0; padding--) |
| 2208 | printf(" " ); |
| 2209 | printf("; opsize %d\n" , opsize); |
| 2210 | return printed + 11; // assumes opsize is always 2 digits |
| 2211 | } |
| 2212 | |
| 2213 | // Dump the unwind data. |
| 2214 | // Arguments: |
| 2215 | // isHotCode: true if this unwind data is for the hot section |
| 2216 | // startOffset: byte offset of the code start that this unwind data represents |
| 2217 | // endOffset: byte offset of the code end that this unwind data represents |
| 2218 | // pHeader: pointer to the unwind data blob |
| 2219 | // unwindBlockSize: size in bytes of the unwind data blob |
| 2220 | |
| 2221 | void DumpUnwindInfo(Compiler* comp, |
| 2222 | bool isHotCode, |
| 2223 | UNATIVE_OFFSET startOffset, |
| 2224 | UNATIVE_OFFSET endOffset, |
| 2225 | const BYTE* const pHeader, |
| 2226 | ULONG unwindBlockSize) |
| 2227 | { |
| 2228 | printf("Unwind Info%s:\n" , isHotCode ? "" : " COLD" ); |
| 2229 | |
| 2230 | // pHeader is not guaranteed to be aligned. We put four 0xFF end codes at the end |
| 2231 | // to provide padding, and round down to get a multiple of 4 bytes in size. |
| 2232 | DWORD UNALIGNED* pdw = (DWORD UNALIGNED*)pHeader; |
| 2233 | DWORD dw; |
| 2234 | |
| 2235 | dw = *pdw++; |
| 2236 | |
| 2237 | DWORD codeWords = ExtractBits(dw, 28, 4); |
| 2238 | DWORD epilogCount = ExtractBits(dw, 23, 5); |
| 2239 | DWORD FBit = ExtractBits(dw, 22, 1); |
| 2240 | DWORD EBit = ExtractBits(dw, 21, 1); |
| 2241 | DWORD XBit = ExtractBits(dw, 20, 1); |
| 2242 | DWORD Vers = ExtractBits(dw, 18, 2); |
| 2243 | DWORD functionLength = ExtractBits(dw, 0, 18); |
| 2244 | |
| 2245 | printf(" >> Start offset : 0x%06x (not in unwind data)\n" , comp->dspOffset(startOffset)); |
| 2246 | printf(" >> End offset : 0x%06x (not in unwind data)\n" , comp->dspOffset(endOffset)); |
| 2247 | printf(" Code Words : %u\n" , codeWords); |
| 2248 | printf(" Epilog Count : %u\n" , epilogCount); |
| 2249 | printf(" F bit : %u\n" , FBit); |
| 2250 | printf(" E bit : %u\n" , EBit); |
| 2251 | printf(" X bit : %u\n" , XBit); |
| 2252 | printf(" Vers : %u\n" , Vers); |
| 2253 | printf(" Function Length : %u (0x%05x) Actual length = %u (0x%06x)\n" , functionLength, functionLength, |
| 2254 | functionLength * 2, functionLength * 2); |
| 2255 | |
| 2256 | assert(functionLength * 2 == endOffset - startOffset); |
| 2257 | |
| 2258 | if (codeWords == 0 && epilogCount == 0) |
| 2259 | { |
| 2260 | // We have an extension word specifying a larger number of Code Words or Epilog Counts |
| 2261 | // than can be specified in the header word. |
| 2262 | |
| 2263 | dw = *pdw++; |
| 2264 | |
| 2265 | codeWords = ExtractBits(dw, 16, 8); |
| 2266 | epilogCount = ExtractBits(dw, 0, 16); |
| 2267 | assert((dw & 0xF0000000) == 0); // reserved field should be zero |
| 2268 | |
| 2269 | printf(" ---- Extension word ----\n" ); |
| 2270 | printf(" Extended Code Words : %u\n" , codeWords); |
| 2271 | printf(" Extended Epilog Count : %u\n" , epilogCount); |
| 2272 | } |
| 2273 | |
| 2274 | bool epilogStartAt[256] = {}; // One byte per possible epilog start index; initialized to false |
| 2275 | |
| 2276 | if (EBit == 0) |
| 2277 | { |
| 2278 | // We have an array of epilog scopes |
| 2279 | |
| 2280 | printf(" ---- Epilog scopes ----\n" ); |
| 2281 | if (epilogCount == 0) |
| 2282 | { |
| 2283 | printf(" No epilogs\n" ); |
| 2284 | } |
| 2285 | else |
| 2286 | { |
| 2287 | for (DWORD scope = 0; scope < epilogCount; scope++) |
| 2288 | { |
| 2289 | dw = *pdw++; |
| 2290 | |
| 2291 | DWORD epilogStartOffset = ExtractBits(dw, 0, 18); |
| 2292 | DWORD res = ExtractBits(dw, 18, 2); |
| 2293 | DWORD condition = ExtractBits(dw, 20, 4); |
| 2294 | DWORD epilogStartIndex = ExtractBits(dw, 24, 8); |
| 2295 | |
| 2296 | // Note that epilogStartOffset for a funclet is the offset from the beginning |
| 2297 | // of the current funclet, not the offset from the beginning of the main function. |
| 2298 | // To help find it when looking through JitDump output, also show the offset from |
| 2299 | // the beginning of the main function. |
| 2300 | DWORD epilogStartOffsetFromMainFunctionBegin = epilogStartOffset * 2 + startOffset; |
| 2301 | |
| 2302 | assert(res == 0); |
| 2303 | |
| 2304 | printf(" ---- Scope %d\n" , scope); |
| 2305 | printf(" Epilog Start Offset : %u (0x%05x) Actual offset = %u (0x%06x) Offset from main " |
| 2306 | "function begin = %u (0x%06x)\n" , |
| 2307 | comp->dspOffset(epilogStartOffset), comp->dspOffset(epilogStartOffset), |
| 2308 | comp->dspOffset(epilogStartOffset * 2), comp->dspOffset(epilogStartOffset * 2), |
| 2309 | comp->dspOffset(epilogStartOffsetFromMainFunctionBegin), |
| 2310 | comp->dspOffset(epilogStartOffsetFromMainFunctionBegin)); |
| 2311 | printf(" Condition : %u (0x%x)%s\n" , condition, condition, |
| 2312 | (condition == 0xE) ? " (always)" : "" ); |
| 2313 | printf(" Epilog Start Index : %u (0x%02x)\n" , epilogStartIndex, epilogStartIndex); |
| 2314 | |
| 2315 | epilogStartAt[epilogStartIndex] = true; // an epilog starts at this offset in the unwind codes |
| 2316 | } |
| 2317 | } |
| 2318 | } |
| 2319 | else |
| 2320 | { |
| 2321 | printf(" --- One epilog, unwind codes at %u\n" , epilogCount); |
| 2322 | assert(epilogCount < _countof(epilogStartAt)); |
| 2323 | epilogStartAt[epilogCount] = true; // the one and only epilog starts its unwind codes at this offset |
| 2324 | } |
| 2325 | |
| 2326 | if (FBit) |
| 2327 | { |
| 2328 | printf(" ---- Note: 'F' bit is set. Prolog codes are for a 'phantom' prolog.\n" ); |
| 2329 | } |
| 2330 | |
| 2331 | // Dump the unwind codes |
| 2332 | |
| 2333 | printf(" ---- Unwind codes ----\n" ); |
| 2334 | |
| 2335 | DWORD countOfUnwindCodes = codeWords * 4; |
| 2336 | PBYTE pUnwindCode = (PBYTE)pdw; |
| 2337 | BYTE b1, b2, b3, b4; |
| 2338 | DWORD x, y; |
| 2339 | DWORD opsize; |
| 2340 | DWORD opCol = 52; |
| 2341 | DWORD printed; |
| 2342 | for (DWORD i = 0; i < countOfUnwindCodes; i++) |
| 2343 | { |
| 2344 | // Does this byte start an epilog sequence? If so, note that fact. |
| 2345 | if (epilogStartAt[i]) |
| 2346 | { |
| 2347 | printf(" ---- Epilog start at index %u ----\n" , i); |
| 2348 | } |
| 2349 | |
| 2350 | b1 = *pUnwindCode++; |
| 2351 | |
| 2352 | if ((b1 & 0x80) == 0) |
| 2353 | { |
| 2354 | // 00-7F : add sp, sp, #X*4 (opsize 16) |
| 2355 | x = b1 & 0x7F; |
| 2356 | printf(" %02X add sp, sp, #%-8d" , b1, x * 4); |
| 2357 | DumpOpsize(opCol - 37, 16); |
| 2358 | } |
| 2359 | else if ((b1 & 0xC0) == 0x80) |
| 2360 | { |
| 2361 | // 80-BF : pop {r0-r12,lr} (X = bitmask) (opsize 32) |
| 2362 | assert(i + 1 < countOfUnwindCodes); |
| 2363 | b2 = *pUnwindCode++; |
| 2364 | i++; |
| 2365 | |
| 2366 | DWORD LBit = ExtractBits(b1, 5, 1); |
| 2367 | x = ((DWORD)(b1 & 0x1F) << 8) | (DWORD)b2; |
| 2368 | |
| 2369 | printf(" %02X %02X pop " , b1, b2); |
| 2370 | printed = 20; |
| 2371 | printed += DumpIntRegSet(x, LBit); |
| 2372 | DumpOpsize(opCol - printed, 32); |
| 2373 | } |
| 2374 | else if ((b1 & 0xF0) == 0xC0) |
| 2375 | { |
| 2376 | // C0-CF : mov sp, rX (X=0-15) (opsize 16) |
| 2377 | x = b1 & 0xF; |
| 2378 | printf(" %02X mov sp, r%u" , b1, x); |
| 2379 | printed = 25 + ((x > 10) ? 2 : 1); |
| 2380 | DumpOpsize(opCol - printed, 16); |
| 2381 | } |
| 2382 | else if ((b1 & 0xF8) == 0xD0) |
| 2383 | { |
| 2384 | // D0-D7 : pop {r4-rX,lr} (X=4-7) (opsize 16) |
| 2385 | x = b1 & 0x3; |
| 2386 | DWORD LBit = b1 & 0x4; |
| 2387 | printf(" %02X pop " , b1); |
| 2388 | printed = 20; |
| 2389 | printed += DumpRegSetRange("r" , 4, x + 4, LBit); |
| 2390 | DumpOpsize(opCol - printed, 16); |
| 2391 | } |
| 2392 | else if ((b1 & 0xF8) == 0xD8) |
| 2393 | { |
| 2394 | // D8-DF : pop {r4-rX,lr} (X=8-11) (opsize 32) |
| 2395 | x = b1 & 0x3; |
| 2396 | DWORD LBit = b1 & 0x4; |
| 2397 | printf(" %02X pop " , b1); |
| 2398 | printed = 20; |
| 2399 | printed += DumpRegSetRange("r" , 4, x + 8, LBit); |
| 2400 | DumpOpsize(opCol - printed, 32); |
| 2401 | } |
| 2402 | else if ((b1 & 0xF8) == 0xE0) |
| 2403 | { |
| 2404 | // E0-E7 : vpop {d8-dX} (X=8-15) (opsize 32) |
| 2405 | x = b1 & 0x7; |
| 2406 | printf(" %02X vpop " , b1); |
| 2407 | printed = 21; |
| 2408 | printed += DumpRegSetRange("d" , 8, x + 8, 0); |
| 2409 | DumpOpsize(opCol - printed, 32); |
| 2410 | } |
| 2411 | else if ((b1 & 0xFC) == 0xE8) |
| 2412 | { |
| 2413 | // E8-EB : addw sp, sp, #X*4 (opsize 32) |
| 2414 | assert(i + 1 < countOfUnwindCodes); |
| 2415 | b2 = *pUnwindCode++; |
| 2416 | i++; |
| 2417 | |
| 2418 | x = ((DWORD)(b1 & 0x3) << 8) | (DWORD)b2; |
| 2419 | |
| 2420 | printf(" %02X %02X addw sp, sp, #%-8u" , b1, b2, x * 4); |
| 2421 | DumpOpsize(opCol - 38, 32); |
| 2422 | } |
| 2423 | else if ((b1 & 0xFE) == 0xEC) |
| 2424 | { |
| 2425 | // EC-ED : pop {r0-r7,lr} (X = bitmask) (opsize 16) |
| 2426 | assert(i + 1 < countOfUnwindCodes); |
| 2427 | b2 = *pUnwindCode++; |
| 2428 | i++; |
| 2429 | |
| 2430 | DWORD LBit = ExtractBits(b1, 0, 1); |
| 2431 | x = (DWORD)b2; |
| 2432 | |
| 2433 | printf(" %02X %02X pop " , b1, b2); |
| 2434 | printed = 20; |
| 2435 | printed += DumpIntRegSet(x, LBit); |
| 2436 | DumpOpsize(opCol - printed, 16); |
| 2437 | } |
| 2438 | else if (b1 == 0xEE) |
| 2439 | { |
| 2440 | assert(i + 1 < countOfUnwindCodes); |
| 2441 | b2 = *pUnwindCode++; |
| 2442 | i++; |
| 2443 | |
| 2444 | if ((b2 & 0xF0) == 0) |
| 2445 | { |
| 2446 | // EE/0x (opsize 16) |
| 2447 | x = b2 & 0xF; |
| 2448 | printf(" %02X %02X Microsoft-specific (x = %02X)" , b1, b2, x); |
| 2449 | DumpOpsize(4, 16); |
| 2450 | } |
| 2451 | else |
| 2452 | { |
| 2453 | // EE/xy (opsize 16) |
| 2454 | x = ExtractBits(b2, 4, 4); |
| 2455 | y = ExtractBits(b2, 0, 4); |
| 2456 | printf(" %02X %02X Available (x = %02X, y = %02X)" , b1, b2, x, y); |
| 2457 | DumpOpsize(4, 16); |
| 2458 | } |
| 2459 | } |
| 2460 | else if (b1 == 0xEF) |
| 2461 | { |
| 2462 | assert(i + 1 < countOfUnwindCodes); |
| 2463 | b2 = *pUnwindCode++; |
| 2464 | i++; |
| 2465 | |
| 2466 | if ((b2 & 0xF0) == 0) |
| 2467 | { |
| 2468 | // EF/0x : ldr lr, [sp], #X*4 (opsize 32) |
| 2469 | x = b2 & 0xF; |
| 2470 | printf(" %02X %02X ldr lr, [sp], #%-8u" , b1, b2, x * 4); |
| 2471 | DumpOpsize(opCol - 39, 32); |
| 2472 | } |
| 2473 | else |
| 2474 | { |
| 2475 | // EF/xy (opsize 32) |
| 2476 | x = ExtractBits(b2, 4, 4); |
| 2477 | y = ExtractBits(b2, 0, 4); |
| 2478 | printf(" %02X %02X Available (x = %02X, y = %02X)" , b1, b2, x, y); |
| 2479 | DumpOpsize(4, 32); |
| 2480 | } |
| 2481 | } |
| 2482 | else if ((b1 & 0xF7) == 0xF0) |
| 2483 | { |
| 2484 | // F0-F4 |
| 2485 | x = b1 & 0x7; |
| 2486 | printf(" %02X Available (x = %02X)\n" , b1, x); |
| 2487 | } |
| 2488 | else if (b1 == 0xF5) |
| 2489 | { |
| 2490 | // F5 : vpop {dS-dE} (opsize 32) |
| 2491 | |
| 2492 | assert(i + 1 < countOfUnwindCodes); |
| 2493 | b2 = *pUnwindCode++; |
| 2494 | i++; |
| 2495 | |
| 2496 | DWORD s = ExtractBits(b2, 4, 4); |
| 2497 | DWORD e = ExtractBits(b2, 0, 4); |
| 2498 | |
| 2499 | printf(" %02X %02X vpop " , b1, b2); |
| 2500 | printed = 21; |
| 2501 | printed += DumpRegSetRange("d" , s, e, 0); |
| 2502 | DumpOpsize(opCol - printed, 32); |
| 2503 | } |
| 2504 | else if (b1 == 0xF6) |
| 2505 | { |
| 2506 | // F6 : vpop {d(S+16)-d(E+16)} (opsize 32) |
| 2507 | |
| 2508 | assert(i + 1 < countOfUnwindCodes); |
| 2509 | b2 = *pUnwindCode++; |
| 2510 | i++; |
| 2511 | |
| 2512 | DWORD s = ExtractBits(b2, 4, 4); |
| 2513 | DWORD e = ExtractBits(b2, 0, 4); |
| 2514 | |
| 2515 | printf(" %02X %02X vpop " , b1, b2); |
| 2516 | printed = 21; |
| 2517 | printed += DumpRegSetRange("d" , s + 16, e + 16, 0); |
| 2518 | DumpOpsize(opCol - printed, 32); |
| 2519 | } |
| 2520 | else if (b1 == 0xF7 || b1 == 0xF9) |
| 2521 | { |
| 2522 | // F7, F9 : add sp, sp, #X*4 |
| 2523 | // 0xF7 has opsize 16, 0xF9 has opsize 32 |
| 2524 | |
| 2525 | assert(i + 2 < countOfUnwindCodes); |
| 2526 | b2 = *pUnwindCode++; |
| 2527 | b3 = *pUnwindCode++; |
| 2528 | i += 2; |
| 2529 | |
| 2530 | x = ((DWORD)b2 << 8) | (DWORD)b3; |
| 2531 | |
| 2532 | opsize = (b1 == 0xF7) ? 16 : 32; |
| 2533 | |
| 2534 | printf(" %02X %02X %02X add sp, sp, #%-8u" , b1, b2, b3, x * 4, opsize); |
| 2535 | DumpOpsize(opCol - 37, opsize); |
| 2536 | } |
| 2537 | else if (b1 == 0xF8 || b1 == 0xFA) |
| 2538 | { |
| 2539 | // F8, FA : add sp, sp, #X*4 |
| 2540 | // 0xF8 has opsize 16, 0xFA has opsize 32 |
| 2541 | |
| 2542 | assert(i + 3 < countOfUnwindCodes); |
| 2543 | b2 = *pUnwindCode++; |
| 2544 | b3 = *pUnwindCode++; |
| 2545 | b4 = *pUnwindCode++; |
| 2546 | i += 3; |
| 2547 | |
| 2548 | x = ((DWORD)b2 << 16) | ((DWORD)b3 << 8) | (DWORD)b4; |
| 2549 | |
| 2550 | opsize = (b1 == 0xF8) ? 16 : 32; |
| 2551 | |
| 2552 | printf(" %02X %02X %02X %02X add sp, sp, #%-8u" , b1, b2, b3, b4, x * 4, opsize); |
| 2553 | DumpOpsize(opCol - 37, opsize); |
| 2554 | } |
| 2555 | else if (b1 == 0xFB || b1 == 0xFC) |
| 2556 | { |
| 2557 | // FB, FC : nop |
| 2558 | // 0xFB has opsize 16, 0xFC has opsize 32 |
| 2559 | |
| 2560 | opsize = (b1 == 0xFB) ? 16 : 32; |
| 2561 | |
| 2562 | printf(" %02X nop" , b1, opsize); |
| 2563 | DumpOpsize(opCol - 19, opsize); |
| 2564 | } |
| 2565 | else if (b1 == 0xFD || b1 == 0xFE) |
| 2566 | { |
| 2567 | // FD, FE : end + nop |
| 2568 | // 0xFD has opsize 16, 0xFE has opsize 32 |
| 2569 | |
| 2570 | opsize = (b1 == 0xFD) ? 16 : 32; |
| 2571 | |
| 2572 | printf(" %02X end + nop" , b1, opsize); |
| 2573 | DumpOpsize(opCol - 25, opsize); |
| 2574 | } |
| 2575 | else if (b1 == 0xFF) |
| 2576 | { |
| 2577 | // FF : end |
| 2578 | |
| 2579 | printf(" %02X end\n" , b1); |
| 2580 | } |
| 2581 | else |
| 2582 | { |
| 2583 | assert(!"Internal error decoding unwind codes" ); |
| 2584 | } |
| 2585 | } |
| 2586 | |
| 2587 | pdw += codeWords; |
| 2588 | assert((PBYTE)pdw == pUnwindCode); |
| 2589 | assert((PBYTE)pdw == pHeader + unwindBlockSize); |
| 2590 | |
| 2591 | assert(XBit == 0); // We don't handle the case where exception data is present, such as the Exception Handler RVA |
| 2592 | |
| 2593 | printf("\n" ); |
| 2594 | } |
| 2595 | |
| 2596 | #endif // DEBUG |
| 2597 | |
| 2598 | #endif // defined(_TARGET_ARM_) |
| 2599 | |
| 2600 | #endif // _TARGET_ARMARCH_ |
| 2601 | |