| 1 | // |
| 2 | // Copyright (C) 2015 LunarG, Inc. |
| 3 | // |
| 4 | // All rights reserved. |
| 5 | // |
| 6 | // Redistribution and use in source and binary forms, with or without |
| 7 | // modification, are permitted provided that the following conditions |
| 8 | // are met: |
| 9 | // |
| 10 | // Redistributions of source code must retain the above copyright |
| 11 | // notice, this list of conditions and the following disclaimer. |
| 12 | // |
| 13 | // Redistributions in binary form must reproduce the above |
| 14 | // copyright notice, this list of conditions and the following |
| 15 | // disclaimer in the documentation and/or other materials provided |
| 16 | // with the distribution. |
| 17 | // |
| 18 | // Neither the name of 3Dlabs Inc. Ltd. nor the names of its |
| 19 | // contributors may be used to endorse or promote products derived |
| 20 | // from this software without specific prior written permission. |
| 21 | // |
| 22 | // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
| 23 | // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
| 24 | // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS |
| 25 | // FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE |
| 26 | // COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, |
| 27 | // INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, |
| 28 | // BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; |
| 29 | // LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER |
| 30 | // CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
| 31 | // LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN |
| 32 | // ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE |
| 33 | // POSSIBILITY OF SUCH DAMAGE. |
| 34 | // |
| 35 | |
| 36 | #include "SPVRemapper.h" |
| 37 | #include "doc.h" |
| 38 | |
| 39 | #include <algorithm> |
| 40 | #include <cassert> |
| 41 | |
| 42 | namespace spv { |
| 43 | |
| 44 | // By default, just abort on error. Can be overridden via RegisterErrorHandler |
| 45 | spirvbin_t::errorfn_t spirvbin_t::errorHandler = [](const std::string&) { exit(status: 5); }; |
| 46 | // By default, eat log messages. Can be overridden via RegisterLogHandler |
| 47 | spirvbin_t::logfn_t spirvbin_t::logHandler = [](const std::string&) { }; |
| 48 | |
| 49 | // This can be overridden to provide other message behavior if needed |
| 50 | void spirvbin_t::msg(int minVerbosity, int indent, const std::string& txt) const |
| 51 | { |
| 52 | if (verbose >= minVerbosity) |
| 53 | logHandler(std::string(indent, ' ') + txt); |
| 54 | } |
| 55 | |
| 56 | // hash opcode, with special handling for OpExtInst |
| 57 | std::uint32_t spirvbin_t::asOpCodeHash(unsigned word) |
| 58 | { |
| 59 | const spv::Op opCode = asOpCode(word); |
| 60 | |
| 61 | std::uint32_t offset = 0; |
| 62 | |
| 63 | switch (opCode) { |
| 64 | case spv::OpExtInst: |
| 65 | offset += asId(word: word + 4); break; |
| 66 | default: |
| 67 | break; |
| 68 | } |
| 69 | |
| 70 | return opCode * 19 + offset; // 19 = small prime |
| 71 | } |
| 72 | |
| 73 | spirvbin_t::range_t spirvbin_t::literalRange(spv::Op opCode) const |
| 74 | { |
| 75 | static const int maxCount = 1<<30; |
| 76 | |
| 77 | switch (opCode) { |
| 78 | case spv::OpTypeFloat: // fall through... |
| 79 | case spv::OpTypePointer: return range_t(2, 3); |
| 80 | case spv::OpTypeInt: return range_t(2, 4); |
| 81 | // TODO: case spv::OpTypeImage: |
| 82 | // TODO: case spv::OpTypeSampledImage: |
| 83 | case spv::OpTypeSampler: return range_t(3, 8); |
| 84 | case spv::OpTypeVector: // fall through |
| 85 | case spv::OpTypeMatrix: // ... |
| 86 | case spv::OpTypePipe: return range_t(3, 4); |
| 87 | case spv::OpConstant: return range_t(3, maxCount); |
| 88 | default: return range_t(0, 0); |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | spirvbin_t::range_t spirvbin_t::typeRange(spv::Op opCode) const |
| 93 | { |
| 94 | static const int maxCount = 1<<30; |
| 95 | |
| 96 | if (isConstOp(opCode)) |
| 97 | return range_t(1, 2); |
| 98 | |
| 99 | switch (opCode) { |
| 100 | case spv::OpTypeVector: // fall through |
| 101 | case spv::OpTypeMatrix: // ... |
| 102 | case spv::OpTypeSampler: // ... |
| 103 | case spv::OpTypeArray: // ... |
| 104 | case spv::OpTypeRuntimeArray: // ... |
| 105 | case spv::OpTypePipe: return range_t(2, 3); |
| 106 | case spv::OpTypeStruct: // fall through |
| 107 | case spv::OpTypeFunction: return range_t(2, maxCount); |
| 108 | case spv::OpTypePointer: return range_t(3, 4); |
| 109 | default: return range_t(0, 0); |
| 110 | } |
| 111 | } |
| 112 | |
| 113 | spirvbin_t::range_t spirvbin_t::constRange(spv::Op opCode) const |
| 114 | { |
| 115 | static const int maxCount = 1<<30; |
| 116 | |
| 117 | switch (opCode) { |
| 118 | case spv::OpTypeArray: // fall through... |
| 119 | case spv::OpTypeRuntimeArray: return range_t(3, 4); |
| 120 | case spv::OpConstantComposite: return range_t(3, maxCount); |
| 121 | default: return range_t(0, 0); |
| 122 | } |
| 123 | } |
| 124 | |
| 125 | // Return the size of a type in 32-bit words. This currently only |
| 126 | // handles ints and floats, and is only invoked by queries which must be |
| 127 | // integer types. If ever needed, it can be generalized. |
| 128 | unsigned spirvbin_t::typeSizeInWords(spv::Id id) const |
| 129 | { |
| 130 | const unsigned typeStart = idPos(id); |
| 131 | const spv::Op opCode = asOpCode(word: typeStart); |
| 132 | |
| 133 | if (errorLatch) |
| 134 | return 0; |
| 135 | |
| 136 | switch (opCode) { |
| 137 | case spv::OpTypeInt: // fall through... |
| 138 | case spv::OpTypeFloat: return (spv[typeStart+2]+31)/32; |
| 139 | default: |
| 140 | return 0; |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | // Looks up the type of a given const or variable ID, and |
| 145 | // returns its size in 32-bit words. |
| 146 | unsigned spirvbin_t::idTypeSizeInWords(spv::Id id) const |
| 147 | { |
| 148 | const auto tid_it = idTypeSizeMap.find(x: id); |
| 149 | if (tid_it == idTypeSizeMap.end()) { |
| 150 | error(txt: "type size for ID not found" ); |
| 151 | return 0; |
| 152 | } |
| 153 | |
| 154 | return tid_it->second; |
| 155 | } |
| 156 | |
| 157 | // Is this an opcode we should remove when using --strip? |
| 158 | bool spirvbin_t::isStripOp(spv::Op opCode, unsigned start) const |
| 159 | { |
| 160 | switch (opCode) { |
| 161 | case spv::OpSource: |
| 162 | case spv::OpSourceExtension: |
| 163 | case spv::OpName: |
| 164 | case spv::OpMemberName: |
| 165 | case spv::OpLine : |
| 166 | { |
| 167 | const std::string name = literalString(word: start + 2); |
| 168 | |
| 169 | std::vector<std::string>::const_iterator it; |
| 170 | for (it = stripWhiteList.begin(); it < stripWhiteList.end(); it++) |
| 171 | { |
| 172 | if (name.find(str: *it) != std::string::npos) { |
| 173 | return false; |
| 174 | } |
| 175 | } |
| 176 | |
| 177 | return true; |
| 178 | } |
| 179 | default : |
| 180 | return false; |
| 181 | } |
| 182 | } |
| 183 | |
| 184 | // Return true if this opcode is flow control |
| 185 | bool spirvbin_t::isFlowCtrl(spv::Op opCode) const |
| 186 | { |
| 187 | switch (opCode) { |
| 188 | case spv::OpBranchConditional: |
| 189 | case spv::OpBranch: |
| 190 | case spv::OpSwitch: |
| 191 | case spv::OpLoopMerge: |
| 192 | case spv::OpSelectionMerge: |
| 193 | case spv::OpLabel: |
| 194 | case spv::OpFunction: |
| 195 | case spv::OpFunctionEnd: return true; |
| 196 | default: return false; |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | // Return true if this opcode defines a type |
| 201 | bool spirvbin_t::isTypeOp(spv::Op opCode) const |
| 202 | { |
| 203 | switch (opCode) { |
| 204 | case spv::OpTypeVoid: |
| 205 | case spv::OpTypeBool: |
| 206 | case spv::OpTypeInt: |
| 207 | case spv::OpTypeFloat: |
| 208 | case spv::OpTypeVector: |
| 209 | case spv::OpTypeMatrix: |
| 210 | case spv::OpTypeImage: |
| 211 | case spv::OpTypeSampler: |
| 212 | case spv::OpTypeArray: |
| 213 | case spv::OpTypeRuntimeArray: |
| 214 | case spv::OpTypeStruct: |
| 215 | case spv::OpTypeOpaque: |
| 216 | case spv::OpTypePointer: |
| 217 | case spv::OpTypeFunction: |
| 218 | case spv::OpTypeEvent: |
| 219 | case spv::OpTypeDeviceEvent: |
| 220 | case spv::OpTypeReserveId: |
| 221 | case spv::OpTypeQueue: |
| 222 | case spv::OpTypeSampledImage: |
| 223 | case spv::OpTypePipe: return true; |
| 224 | default: return false; |
| 225 | } |
| 226 | } |
| 227 | |
| 228 | // Return true if this opcode defines a constant |
| 229 | bool spirvbin_t::isConstOp(spv::Op opCode) const |
| 230 | { |
| 231 | switch (opCode) { |
| 232 | case spv::OpConstantSampler: |
| 233 | error(txt: "unimplemented constant type" ); |
| 234 | return true; |
| 235 | |
| 236 | case spv::OpConstantNull: |
| 237 | case spv::OpConstantTrue: |
| 238 | case spv::OpConstantFalse: |
| 239 | case spv::OpConstantComposite: |
| 240 | case spv::OpConstant: |
| 241 | return true; |
| 242 | |
| 243 | default: |
| 244 | return false; |
| 245 | } |
| 246 | } |
| 247 | |
| 248 | const auto inst_fn_nop = [](spv::Op, unsigned) { return false; }; |
| 249 | const auto op_fn_nop = [](spv::Id&) { }; |
| 250 | |
| 251 | // g++ doesn't like these defined in the class proper in an anonymous namespace. |
| 252 | // Dunno why. Also MSVC doesn't like the constexpr keyword. Also dunno why. |
| 253 | // Defining them externally seems to please both compilers, so, here they are. |
| 254 | const spv::Id spirvbin_t::unmapped = spv::Id(-10000); |
| 255 | const spv::Id spirvbin_t::unused = spv::Id(-10001); |
| 256 | const int spirvbin_t:: = 5; |
| 257 | |
| 258 | spv::Id spirvbin_t::nextUnusedId(spv::Id id) |
| 259 | { |
| 260 | while (isNewIdMapped(newId: id)) // search for an unused ID |
| 261 | ++id; |
| 262 | |
| 263 | return id; |
| 264 | } |
| 265 | |
| 266 | spv::Id spirvbin_t::localId(spv::Id id, spv::Id newId) |
| 267 | { |
| 268 | //assert(id != spv::NoResult && newId != spv::NoResult); |
| 269 | |
| 270 | if (id > bound()) { |
| 271 | error(txt: std::string("ID out of range: " ) + std::to_string(val: id)); |
| 272 | return spirvbin_t::unused; |
| 273 | } |
| 274 | |
| 275 | if (id >= idMapL.size()) |
| 276 | idMapL.resize(new_size: id+1, x: unused); |
| 277 | |
| 278 | if (newId != unmapped && newId != unused) { |
| 279 | if (isOldIdUnused(oldId: id)) { |
| 280 | error(txt: std::string("ID unused in module: " ) + std::to_string(val: id)); |
| 281 | return spirvbin_t::unused; |
| 282 | } |
| 283 | |
| 284 | if (!isOldIdUnmapped(oldId: id)) { |
| 285 | error(txt: std::string("ID already mapped: " ) + std::to_string(val: id) + " -> " |
| 286 | + std::to_string(val: localId(id))); |
| 287 | |
| 288 | return spirvbin_t::unused; |
| 289 | } |
| 290 | |
| 291 | if (isNewIdMapped(newId)) { |
| 292 | error(txt: std::string("ID already used in module: " ) + std::to_string(val: newId)); |
| 293 | return spirvbin_t::unused; |
| 294 | } |
| 295 | |
| 296 | msg(minVerbosity: 4, indent: 4, txt: std::string("map: " ) + std::to_string(val: id) + " -> " + std::to_string(val: newId)); |
| 297 | setMapped(newId); |
| 298 | largestNewId = std::max(a: largestNewId, b: newId); |
| 299 | } |
| 300 | |
| 301 | return idMapL[id] = newId; |
| 302 | } |
| 303 | |
| 304 | // Parse a literal string from the SPIR binary and return it as an std::string |
| 305 | // Due to C++11 RValue references, this doesn't copy the result string. |
| 306 | std::string spirvbin_t::literalString(unsigned word) const |
| 307 | { |
| 308 | std::string literal; |
| 309 | const spirword_t * pos = spv.data() + word; |
| 310 | |
| 311 | literal.reserve(res: 16); |
| 312 | |
| 313 | do { |
| 314 | spirword_t word = *pos; |
| 315 | for (int i = 0; i < 4; i++) { |
| 316 | char c = word & 0xff; |
| 317 | if (c == '\0') |
| 318 | return literal; |
| 319 | literal += c; |
| 320 | word >>= 8; |
| 321 | } |
| 322 | pos++; |
| 323 | } while (true); |
| 324 | } |
| 325 | |
| 326 | void spirvbin_t::applyMap() |
| 327 | { |
| 328 | msg(minVerbosity: 3, indent: 2, txt: std::string("Applying map: " )); |
| 329 | |
| 330 | // Map local IDs through the ID map |
| 331 | process(inst_fn_nop, // ignore instructions |
| 332 | [this](spv::Id& id) { |
| 333 | id = localId(id); |
| 334 | |
| 335 | if (errorLatch) |
| 336 | return; |
| 337 | |
| 338 | assert(id != unused && id != unmapped); |
| 339 | } |
| 340 | ); |
| 341 | } |
| 342 | |
| 343 | // Find free IDs for anything we haven't mapped |
| 344 | void spirvbin_t::mapRemainder() |
| 345 | { |
| 346 | msg(minVerbosity: 3, indent: 2, txt: std::string("Remapping remainder: " )); |
| 347 | |
| 348 | spv::Id unusedId = 1; // can't use 0: that's NoResult |
| 349 | spirword_t maxBound = 0; |
| 350 | |
| 351 | for (spv::Id id = 0; id < idMapL.size(); ++id) { |
| 352 | if (isOldIdUnused(oldId: id)) |
| 353 | continue; |
| 354 | |
| 355 | // Find a new mapping for any used but unmapped IDs |
| 356 | if (isOldIdUnmapped(oldId: id)) { |
| 357 | localId(id, newId: unusedId = nextUnusedId(id: unusedId)); |
| 358 | if (errorLatch) |
| 359 | return; |
| 360 | } |
| 361 | |
| 362 | if (isOldIdUnmapped(oldId: id)) { |
| 363 | error(txt: std::string("old ID not mapped: " ) + std::to_string(val: id)); |
| 364 | return; |
| 365 | } |
| 366 | |
| 367 | // Track max bound |
| 368 | maxBound = std::max(a: maxBound, b: localId(id) + 1); |
| 369 | |
| 370 | if (errorLatch) |
| 371 | return; |
| 372 | } |
| 373 | |
| 374 | bound(b: maxBound); // reset header ID bound to as big as it now needs to be |
| 375 | } |
| 376 | |
| 377 | // Mark debug instructions for stripping |
| 378 | void spirvbin_t::stripDebug() |
| 379 | { |
| 380 | // Strip instructions in the stripOp set: debug info. |
| 381 | process( |
| 382 | [&](spv::Op opCode, unsigned start) { |
| 383 | // remember opcodes we want to strip later |
| 384 | if (isStripOp(opCode, start)) |
| 385 | stripInst(start); |
| 386 | return true; |
| 387 | }, |
| 388 | op_fn_nop); |
| 389 | } |
| 390 | |
| 391 | // Mark instructions that refer to now-removed IDs for stripping |
| 392 | void spirvbin_t::stripDeadRefs() |
| 393 | { |
| 394 | process( |
| 395 | [&](spv::Op opCode, unsigned start) { |
| 396 | // strip opcodes pointing to removed data |
| 397 | switch (opCode) { |
| 398 | case spv::OpName: |
| 399 | case spv::OpMemberName: |
| 400 | case spv::OpDecorate: |
| 401 | case spv::OpMemberDecorate: |
| 402 | if (idPosR.find(x: asId(word: start+1)) == idPosR.end()) |
| 403 | stripInst(start); |
| 404 | break; |
| 405 | default: |
| 406 | break; // leave it alone |
| 407 | } |
| 408 | |
| 409 | return true; |
| 410 | }, |
| 411 | op_fn_nop); |
| 412 | |
| 413 | strip(); |
| 414 | } |
| 415 | |
| 416 | // Update local maps of ID, type, etc positions |
| 417 | void spirvbin_t::buildLocalMaps() |
| 418 | { |
| 419 | msg(minVerbosity: 2, indent: 2, txt: std::string("build local maps: " )); |
| 420 | |
| 421 | mapped.clear(); |
| 422 | idMapL.clear(); |
| 423 | // preserve nameMap, so we don't clear that. |
| 424 | fnPos.clear(); |
| 425 | fnCalls.clear(); |
| 426 | typeConstPos.clear(); |
| 427 | idPosR.clear(); |
| 428 | entryPoint = spv::NoResult; |
| 429 | largestNewId = 0; |
| 430 | |
| 431 | idMapL.resize(new_size: bound(), x: unused); |
| 432 | |
| 433 | int fnStart = 0; |
| 434 | spv::Id fnRes = spv::NoResult; |
| 435 | |
| 436 | // build local Id and name maps |
| 437 | process( |
| 438 | [&](spv::Op opCode, unsigned start) { |
| 439 | unsigned word = start+1; |
| 440 | spv::Id typeId = spv::NoResult; |
| 441 | |
| 442 | if (spv::InstructionDesc[opCode].hasType()) |
| 443 | typeId = asId(word: word++); |
| 444 | |
| 445 | // If there's a result ID, remember the size of its type |
| 446 | if (spv::InstructionDesc[opCode].hasResult()) { |
| 447 | const spv::Id resultId = asId(word: word++); |
| 448 | idPosR[resultId] = start; |
| 449 | |
| 450 | if (typeId != spv::NoResult) { |
| 451 | const unsigned idTypeSize = typeSizeInWords(id: typeId); |
| 452 | |
| 453 | if (errorLatch) |
| 454 | return false; |
| 455 | |
| 456 | if (idTypeSize != 0) |
| 457 | idTypeSizeMap[resultId] = idTypeSize; |
| 458 | } |
| 459 | } |
| 460 | |
| 461 | if (opCode == spv::Op::OpName) { |
| 462 | const spv::Id target = asId(word: start+1); |
| 463 | const std::string name = literalString(word: start+2); |
| 464 | nameMap[name] = target; |
| 465 | |
| 466 | } else if (opCode == spv::Op::OpFunctionCall) { |
| 467 | ++fnCalls[asId(word: start + 3)]; |
| 468 | } else if (opCode == spv::Op::OpEntryPoint) { |
| 469 | entryPoint = asId(word: start + 2); |
| 470 | } else if (opCode == spv::Op::OpFunction) { |
| 471 | if (fnStart != 0) { |
| 472 | error(txt: "nested function found" ); |
| 473 | return false; |
| 474 | } |
| 475 | |
| 476 | fnStart = start; |
| 477 | fnRes = asId(word: start + 2); |
| 478 | } else if (opCode == spv::Op::OpFunctionEnd) { |
| 479 | assert(fnRes != spv::NoResult); |
| 480 | if (fnStart == 0) { |
| 481 | error(txt: "function end without function start" ); |
| 482 | return false; |
| 483 | } |
| 484 | |
| 485 | fnPos[fnRes] = range_t(fnStart, start + asWordCount(word: start)); |
| 486 | fnStart = 0; |
| 487 | } else if (isConstOp(opCode)) { |
| 488 | if (errorLatch) |
| 489 | return false; |
| 490 | |
| 491 | assert(asId(start + 2) != spv::NoResult); |
| 492 | typeConstPos.insert(x: start); |
| 493 | } else if (isTypeOp(opCode)) { |
| 494 | assert(asId(start + 1) != spv::NoResult); |
| 495 | typeConstPos.insert(x: start); |
| 496 | } |
| 497 | |
| 498 | return false; |
| 499 | }, |
| 500 | |
| 501 | [this](spv::Id& id) { localId(id, newId: unmapped); } |
| 502 | ); |
| 503 | } |
| 504 | |
| 505 | // Validate the SPIR header |
| 506 | void spirvbin_t::validate() const |
| 507 | { |
| 508 | msg(minVerbosity: 2, indent: 2, txt: std::string("validating: " )); |
| 509 | |
| 510 | if (spv.size() < header_size) { |
| 511 | error(txt: "file too short: " ); |
| 512 | return; |
| 513 | } |
| 514 | |
| 515 | if (magic() != spv::MagicNumber) { |
| 516 | error(txt: "bad magic number" ); |
| 517 | return; |
| 518 | } |
| 519 | |
| 520 | // field 1 = version |
| 521 | // field 2 = generator magic |
| 522 | // field 3 = result <id> bound |
| 523 | |
| 524 | if (schemaNum() != 0) { |
| 525 | error(txt: "bad schema, must be 0" ); |
| 526 | return; |
| 527 | } |
| 528 | } |
| 529 | |
| 530 | int spirvbin_t::processInstruction(unsigned word, instfn_t instFn, idfn_t idFn) |
| 531 | { |
| 532 | const auto instructionStart = word; |
| 533 | const unsigned wordCount = asWordCount(word: instructionStart); |
| 534 | const int nextInst = word++ + wordCount; |
| 535 | spv::Op opCode = asOpCode(word: instructionStart); |
| 536 | |
| 537 | if (nextInst > int(spv.size())) { |
| 538 | error(txt: "spir instruction terminated too early" ); |
| 539 | return -1; |
| 540 | } |
| 541 | |
| 542 | // Base for computing number of operands; will be updated as more is learned |
| 543 | unsigned numOperands = wordCount - 1; |
| 544 | |
| 545 | if (instFn(opCode, instructionStart)) |
| 546 | return nextInst; |
| 547 | |
| 548 | // Read type and result ID from instruction desc table |
| 549 | if (spv::InstructionDesc[opCode].hasType()) { |
| 550 | idFn(asId(word: word++)); |
| 551 | --numOperands; |
| 552 | } |
| 553 | |
| 554 | if (spv::InstructionDesc[opCode].hasResult()) { |
| 555 | idFn(asId(word: word++)); |
| 556 | --numOperands; |
| 557 | } |
| 558 | |
| 559 | // Extended instructions: currently, assume everything is an ID. |
| 560 | // TODO: add whatever data we need for exceptions to that |
| 561 | if (opCode == spv::OpExtInst) { |
| 562 | |
| 563 | idFn(asId(word)); // Instruction set is an ID that also needs to be mapped |
| 564 | |
| 565 | word += 2; // instruction set, and instruction from set |
| 566 | numOperands -= 2; |
| 567 | |
| 568 | for (unsigned op=0; op < numOperands; ++op) |
| 569 | idFn(asId(word: word++)); // ID |
| 570 | |
| 571 | return nextInst; |
| 572 | } |
| 573 | |
| 574 | // Circular buffer so we can look back at previous unmapped values during the mapping pass. |
| 575 | static const unsigned idBufferSize = 4; |
| 576 | spv::Id idBuffer[idBufferSize]; |
| 577 | unsigned idBufferPos = 0; |
| 578 | |
| 579 | // Store IDs from instruction in our map |
| 580 | for (int op = 0; numOperands > 0; ++op, --numOperands) { |
| 581 | // SpecConstantOp is special: it includes the operands of another opcode which is |
| 582 | // given as a literal in the 3rd word. We will switch over to pretending that the |
| 583 | // opcode being processed is the literal opcode value of the SpecConstantOp. See the |
| 584 | // SPIRV spec for details. This way we will handle IDs and literals as appropriate for |
| 585 | // the embedded op. |
| 586 | if (opCode == spv::OpSpecConstantOp) { |
| 587 | if (op == 0) { |
| 588 | opCode = asOpCode(word: word++); // this is the opcode embedded in the SpecConstantOp. |
| 589 | --numOperands; |
| 590 | } |
| 591 | } |
| 592 | |
| 593 | switch (spv::InstructionDesc[opCode].operands.getClass(op)) { |
| 594 | case spv::OperandId: |
| 595 | case spv::OperandScope: |
| 596 | case spv::OperandMemorySemantics: |
| 597 | idBuffer[idBufferPos] = asId(word); |
| 598 | idBufferPos = (idBufferPos + 1) % idBufferSize; |
| 599 | idFn(asId(word: word++)); |
| 600 | break; |
| 601 | |
| 602 | case spv::OperandVariableIds: |
| 603 | for (unsigned i = 0; i < numOperands; ++i) |
| 604 | idFn(asId(word: word++)); |
| 605 | return nextInst; |
| 606 | |
| 607 | case spv::OperandVariableLiterals: |
| 608 | // for clarity |
| 609 | // if (opCode == spv::OpDecorate && asDecoration(word - 1) == spv::DecorationBuiltIn) { |
| 610 | // ++word; |
| 611 | // --numOperands; |
| 612 | // } |
| 613 | // word += numOperands; |
| 614 | return nextInst; |
| 615 | |
| 616 | case spv::OperandVariableLiteralId: { |
| 617 | if (opCode == OpSwitch) { |
| 618 | // word-2 is the position of the selector ID. OpSwitch Literals match its type. |
| 619 | // In case the IDs are currently being remapped, we get the word[-2] ID from |
| 620 | // the circular idBuffer. |
| 621 | const unsigned literalSizePos = (idBufferPos+idBufferSize-2) % idBufferSize; |
| 622 | const unsigned literalSize = idTypeSizeInWords(id: idBuffer[literalSizePos]); |
| 623 | const unsigned numLiteralIdPairs = (nextInst-word) / (1+literalSize); |
| 624 | |
| 625 | if (errorLatch) |
| 626 | return -1; |
| 627 | |
| 628 | for (unsigned arg=0; arg<numLiteralIdPairs; ++arg) { |
| 629 | word += literalSize; // literal |
| 630 | idFn(asId(word: word++)); // label |
| 631 | } |
| 632 | } else { |
| 633 | assert(0); // currentely, only OpSwitch uses OperandVariableLiteralId |
| 634 | } |
| 635 | |
| 636 | return nextInst; |
| 637 | } |
| 638 | |
| 639 | case spv::OperandLiteralString: { |
| 640 | const int stringWordCount = literalStringWords(str: literalString(word)); |
| 641 | word += stringWordCount; |
| 642 | numOperands -= (stringWordCount-1); // -1 because for() header post-decrements |
| 643 | break; |
| 644 | } |
| 645 | |
| 646 | case spv::OperandVariableLiteralStrings: |
| 647 | return nextInst; |
| 648 | |
| 649 | // Execution mode might have extra literal operands. Skip them. |
| 650 | case spv::OperandExecutionMode: |
| 651 | return nextInst; |
| 652 | |
| 653 | case spv::OperandMemoryAccess: |
| 654 | { |
| 655 | uint32_t mask = spv[word]; |
| 656 | if (mask & uint32_t(spv::MemoryAccessMask::MemoryAccessAlignedMask)) { |
| 657 | ++word; |
| 658 | --numOperands; |
| 659 | } |
| 660 | if (mask & uint32_t(spv::MemoryAccessMask::MemoryAccessMakePointerAvailableMask | |
| 661 | spv::MemoryAccessMask::MemoryAccessMakePointerVisibleMask)) { |
| 662 | idFn(asId(word: word+1)); |
| 663 | ++word; |
| 664 | --numOperands; |
| 665 | } |
| 666 | ++word; |
| 667 | } |
| 668 | break; |
| 669 | |
| 670 | case spv::OperandTensorAddressingOperands: |
| 671 | { |
| 672 | uint32_t mask = spv[word]; |
| 673 | if (mask & uint32_t(spv::TensorAddressingOperandsMask::TensorAddressingOperandsTensorViewMask)) { |
| 674 | idFn(asId(word: word+1)); |
| 675 | ++word; |
| 676 | --numOperands; |
| 677 | } |
| 678 | if (mask & uint32_t(spv::TensorAddressingOperandsMask::TensorAddressingOperandsDecodeFuncMask)) { |
| 679 | idFn(asId(word: word+1)); |
| 680 | ++word; |
| 681 | --numOperands; |
| 682 | } |
| 683 | ++word; |
| 684 | } |
| 685 | break; |
| 686 | |
| 687 | // Single word operands we simply ignore, as they hold no IDs |
| 688 | case spv::OperandLiteralNumber: |
| 689 | case spv::OperandSource: |
| 690 | case spv::OperandExecutionModel: |
| 691 | case spv::OperandAddressing: |
| 692 | case spv::OperandMemory: |
| 693 | case spv::OperandStorage: |
| 694 | case spv::OperandDimensionality: |
| 695 | case spv::OperandSamplerAddressingMode: |
| 696 | case spv::OperandSamplerFilterMode: |
| 697 | case spv::OperandSamplerImageFormat: |
| 698 | case spv::OperandImageChannelOrder: |
| 699 | case spv::OperandImageChannelDataType: |
| 700 | case spv::OperandImageOperands: |
| 701 | case spv::OperandFPFastMath: |
| 702 | case spv::OperandFPRoundingMode: |
| 703 | case spv::OperandLinkageType: |
| 704 | case spv::OperandAccessQualifier: |
| 705 | case spv::OperandFuncParamAttr: |
| 706 | case spv::OperandDecoration: |
| 707 | case spv::OperandBuiltIn: |
| 708 | case spv::OperandSelect: |
| 709 | case spv::OperandLoop: |
| 710 | case spv::OperandFunction: |
| 711 | case spv::OperandGroupOperation: |
| 712 | case spv::OperandKernelEnqueueFlags: |
| 713 | case spv::OperandKernelProfilingInfo: |
| 714 | case spv::OperandCapability: |
| 715 | case spv::OperandCooperativeMatrixOperands: |
| 716 | ++word; |
| 717 | break; |
| 718 | |
| 719 | default: |
| 720 | assert(0 && "Unhandled Operand Class" ); |
| 721 | break; |
| 722 | } |
| 723 | } |
| 724 | |
| 725 | return nextInst; |
| 726 | } |
| 727 | |
| 728 | // Make a pass over all the instructions and process them given appropriate functions |
| 729 | spirvbin_t& spirvbin_t::process(instfn_t instFn, idfn_t idFn, unsigned begin, unsigned end) |
| 730 | { |
| 731 | // For efficiency, reserve name map space. It can grow if needed. |
| 732 | nameMap.reserve(n: 32); |
| 733 | |
| 734 | // If begin or end == 0, use defaults |
| 735 | begin = (begin == 0 ? header_size : begin); |
| 736 | end = (end == 0 ? unsigned(spv.size()) : end); |
| 737 | |
| 738 | // basic parsing and InstructionDesc table borrowed from SpvDisassemble.cpp... |
| 739 | unsigned nextInst = unsigned(spv.size()); |
| 740 | |
| 741 | for (unsigned word = begin; word < end; word = nextInst) { |
| 742 | nextInst = processInstruction(word, instFn, idFn); |
| 743 | |
| 744 | if (errorLatch) |
| 745 | return *this; |
| 746 | } |
| 747 | |
| 748 | return *this; |
| 749 | } |
| 750 | |
| 751 | // Apply global name mapping to a single module |
| 752 | void spirvbin_t::mapNames() |
| 753 | { |
| 754 | static const std::uint32_t softTypeIdLimit = 3011; // small prime. TODO: get from options |
| 755 | static const std::uint32_t firstMappedID = 3019; // offset into ID space |
| 756 | |
| 757 | for (const auto& name : nameMap) { |
| 758 | std::uint32_t hashval = 1911; |
| 759 | for (const char c : name.first) |
| 760 | hashval = hashval * 1009 + c; |
| 761 | |
| 762 | if (isOldIdUnmapped(oldId: name.second)) { |
| 763 | localId(id: name.second, newId: nextUnusedId(id: hashval % softTypeIdLimit + firstMappedID)); |
| 764 | if (errorLatch) |
| 765 | return; |
| 766 | } |
| 767 | } |
| 768 | } |
| 769 | |
| 770 | // Map fn contents to IDs of similar functions in other modules |
| 771 | void spirvbin_t::mapFnBodies() |
| 772 | { |
| 773 | static const std::uint32_t softTypeIdLimit = 19071; // small prime. TODO: get from options |
| 774 | static const std::uint32_t firstMappedID = 6203; // offset into ID space |
| 775 | |
| 776 | // Initial approach: go through some high priority opcodes first and assign them |
| 777 | // hash values. |
| 778 | |
| 779 | spv::Id fnId = spv::NoResult; |
| 780 | std::vector<unsigned> instPos; |
| 781 | instPos.reserve(n: unsigned(spv.size()) / 16); // initial estimate; can grow if needed. |
| 782 | |
| 783 | // Build local table of instruction start positions |
| 784 | process( |
| 785 | instFn: [&](spv::Op, unsigned start) { instPos.push_back(x: start); return true; }, |
| 786 | idFn: op_fn_nop); |
| 787 | |
| 788 | if (errorLatch) |
| 789 | return; |
| 790 | |
| 791 | // Window size for context-sensitive canonicalization values |
| 792 | // Empirical best size from a single data set. TODO: Would be a good tunable. |
| 793 | // We essentially perform a little convolution around each instruction, |
| 794 | // to capture the flavor of nearby code, to hopefully match to similar |
| 795 | // code in other modules. |
| 796 | static const unsigned windowSize = 2; |
| 797 | |
| 798 | for (unsigned entry = 0; entry < unsigned(instPos.size()); ++entry) { |
| 799 | const unsigned start = instPos[entry]; |
| 800 | const spv::Op opCode = asOpCode(word: start); |
| 801 | |
| 802 | if (opCode == spv::OpFunction) |
| 803 | fnId = asId(word: start + 2); |
| 804 | |
| 805 | if (opCode == spv::OpFunctionEnd) |
| 806 | fnId = spv::NoResult; |
| 807 | |
| 808 | if (fnId != spv::NoResult) { // if inside a function |
| 809 | if (spv::InstructionDesc[opCode].hasResult()) { |
| 810 | const unsigned word = start + (spv::InstructionDesc[opCode].hasType() ? 2 : 1); |
| 811 | const spv::Id resId = asId(word); |
| 812 | std::uint32_t hashval = fnId * 17; // small prime |
| 813 | |
| 814 | for (unsigned i = entry-1; i >= entry-windowSize; --i) { |
| 815 | if (asOpCode(word: instPos[i]) == spv::OpFunction) |
| 816 | break; |
| 817 | hashval = hashval * 30103 + asOpCodeHash(word: instPos[i]); // 30103 = semiarbitrary prime |
| 818 | } |
| 819 | |
| 820 | for (unsigned i = entry; i <= entry + windowSize; ++i) { |
| 821 | if (asOpCode(word: instPos[i]) == spv::OpFunctionEnd) |
| 822 | break; |
| 823 | hashval = hashval * 30103 + asOpCodeHash(word: instPos[i]); // 30103 = semiarbitrary prime |
| 824 | } |
| 825 | |
| 826 | if (isOldIdUnmapped(oldId: resId)) { |
| 827 | localId(id: resId, newId: nextUnusedId(id: hashval % softTypeIdLimit + firstMappedID)); |
| 828 | if (errorLatch) |
| 829 | return; |
| 830 | } |
| 831 | |
| 832 | } |
| 833 | } |
| 834 | } |
| 835 | |
| 836 | spv::Op thisOpCode(spv::OpNop); |
| 837 | std::unordered_map<int, int> opCounter; |
| 838 | int idCounter(0); |
| 839 | fnId = spv::NoResult; |
| 840 | |
| 841 | process( |
| 842 | instFn: [&](spv::Op opCode, unsigned start) { |
| 843 | switch (opCode) { |
| 844 | case spv::OpFunction: |
| 845 | // Reset counters at each function |
| 846 | idCounter = 0; |
| 847 | opCounter.clear(); |
| 848 | fnId = asId(word: start + 2); |
| 849 | break; |
| 850 | |
| 851 | case spv::OpImageSampleImplicitLod: |
| 852 | case spv::OpImageSampleExplicitLod: |
| 853 | case spv::OpImageSampleDrefImplicitLod: |
| 854 | case spv::OpImageSampleDrefExplicitLod: |
| 855 | case spv::OpImageSampleProjImplicitLod: |
| 856 | case spv::OpImageSampleProjExplicitLod: |
| 857 | case spv::OpImageSampleProjDrefImplicitLod: |
| 858 | case spv::OpImageSampleProjDrefExplicitLod: |
| 859 | case spv::OpDot: |
| 860 | case spv::OpCompositeExtract: |
| 861 | case spv::OpCompositeInsert: |
| 862 | case spv::OpVectorShuffle: |
| 863 | case spv::OpLabel: |
| 864 | case spv::OpVariable: |
| 865 | |
| 866 | case spv::OpAccessChain: |
| 867 | case spv::OpLoad: |
| 868 | case spv::OpStore: |
| 869 | case spv::OpCompositeConstruct: |
| 870 | case spv::OpFunctionCall: |
| 871 | ++opCounter[opCode]; |
| 872 | idCounter = 0; |
| 873 | thisOpCode = opCode; |
| 874 | break; |
| 875 | default: |
| 876 | thisOpCode = spv::OpNop; |
| 877 | } |
| 878 | |
| 879 | return false; |
| 880 | }, |
| 881 | |
| 882 | idFn: [&](spv::Id& id) { |
| 883 | if (thisOpCode != spv::OpNop) { |
| 884 | ++idCounter; |
| 885 | const std::uint32_t hashval = |
| 886 | // Explicitly cast operands to unsigned int to avoid integer |
| 887 | // promotion to signed int followed by integer overflow, |
| 888 | // which would result in undefined behavior. |
| 889 | static_cast<unsigned int>(opCounter[thisOpCode]) |
| 890 | * thisOpCode |
| 891 | * 50047 |
| 892 | + idCounter |
| 893 | + static_cast<unsigned int>(fnId) * 117; |
| 894 | |
| 895 | if (isOldIdUnmapped(oldId: id)) |
| 896 | localId(id, newId: nextUnusedId(id: hashval % softTypeIdLimit + firstMappedID)); |
| 897 | } |
| 898 | }); |
| 899 | } |
| 900 | |
| 901 | // EXPERIMENTAL: forward IO and uniform load/stores into operands |
| 902 | // This produces invalid Schema-0 SPIRV |
| 903 | void spirvbin_t::forwardLoadStores() |
| 904 | { |
| 905 | idset_t fnLocalVars; // set of function local vars |
| 906 | idmap_t idMap; // Map of load result IDs to what they load |
| 907 | |
| 908 | // EXPERIMENTAL: Forward input and access chain loads into consumptions |
| 909 | process( |
| 910 | instFn: [&](spv::Op opCode, unsigned start) { |
| 911 | // Add inputs and uniforms to the map |
| 912 | if ((opCode == spv::OpVariable && asWordCount(word: start) == 4) && |
| 913 | (spv[start+3] == spv::StorageClassUniform || |
| 914 | spv[start+3] == spv::StorageClassUniformConstant || |
| 915 | spv[start+3] == spv::StorageClassInput)) |
| 916 | fnLocalVars.insert(x: asId(word: start+2)); |
| 917 | |
| 918 | if (opCode == spv::OpAccessChain && fnLocalVars.count(x: asId(word: start+3)) > 0) |
| 919 | fnLocalVars.insert(x: asId(word: start+2)); |
| 920 | |
| 921 | if (opCode == spv::OpLoad && fnLocalVars.count(x: asId(word: start+3)) > 0) { |
| 922 | idMap[asId(word: start+2)] = asId(word: start+3); |
| 923 | stripInst(start); |
| 924 | } |
| 925 | |
| 926 | return false; |
| 927 | }, |
| 928 | |
| 929 | idFn: [&](spv::Id& id) { if (idMap.find(x: id) != idMap.end()) id = idMap[id]; } |
| 930 | ); |
| 931 | |
| 932 | if (errorLatch) |
| 933 | return; |
| 934 | |
| 935 | // EXPERIMENTAL: Implicit output stores |
| 936 | fnLocalVars.clear(); |
| 937 | idMap.clear(); |
| 938 | |
| 939 | process( |
| 940 | instFn: [&](spv::Op opCode, unsigned start) { |
| 941 | // Add inputs and uniforms to the map |
| 942 | if ((opCode == spv::OpVariable && asWordCount(word: start) == 4) && |
| 943 | (spv[start+3] == spv::StorageClassOutput)) |
| 944 | fnLocalVars.insert(x: asId(word: start+2)); |
| 945 | |
| 946 | if (opCode == spv::OpStore && fnLocalVars.count(x: asId(word: start+1)) > 0) { |
| 947 | idMap[asId(word: start+2)] = asId(word: start+1); |
| 948 | stripInst(start); |
| 949 | } |
| 950 | |
| 951 | return false; |
| 952 | }, |
| 953 | idFn: op_fn_nop); |
| 954 | |
| 955 | if (errorLatch) |
| 956 | return; |
| 957 | |
| 958 | process( |
| 959 | instFn: inst_fn_nop, |
| 960 | idFn: [&](spv::Id& id) { if (idMap.find(x: id) != idMap.end()) id = idMap[id]; } |
| 961 | ); |
| 962 | |
| 963 | if (errorLatch) |
| 964 | return; |
| 965 | |
| 966 | strip(); // strip out data we decided to eliminate |
| 967 | } |
| 968 | |
| 969 | // optimize loads and stores |
| 970 | void spirvbin_t::optLoadStore() |
| 971 | { |
| 972 | idset_t fnLocalVars; // candidates for removal (only locals) |
| 973 | idmap_t idMap; // Map of load result IDs to what they load |
| 974 | blockmap_t blockMap; // Map of IDs to blocks they first appear in |
| 975 | int blockNum = 0; // block count, to avoid crossing flow control |
| 976 | |
| 977 | // Find all the function local pointers stored at most once, and not via access chains |
| 978 | process( |
| 979 | instFn: [&](spv::Op opCode, unsigned start) { |
| 980 | const int wordCount = asWordCount(word: start); |
| 981 | |
| 982 | // Count blocks, so we can avoid crossing flow control |
| 983 | if (isFlowCtrl(opCode)) |
| 984 | ++blockNum; |
| 985 | |
| 986 | // Add local variables to the map |
| 987 | if ((opCode == spv::OpVariable && spv[start+3] == spv::StorageClassFunction && asWordCount(word: start) == 4)) { |
| 988 | fnLocalVars.insert(x: asId(word: start+2)); |
| 989 | return true; |
| 990 | } |
| 991 | |
| 992 | // Ignore process vars referenced via access chain |
| 993 | if ((opCode == spv::OpAccessChain || opCode == spv::OpInBoundsAccessChain) && fnLocalVars.count(x: asId(word: start+3)) > 0) { |
| 994 | fnLocalVars.erase(x: asId(word: start+3)); |
| 995 | idMap.erase(x: asId(word: start+3)); |
| 996 | return true; |
| 997 | } |
| 998 | |
| 999 | if (opCode == spv::OpLoad && fnLocalVars.count(x: asId(word: start+3)) > 0) { |
| 1000 | const spv::Id varId = asId(word: start+3); |
| 1001 | |
| 1002 | // Avoid loads before stores |
| 1003 | if (idMap.find(x: varId) == idMap.end()) { |
| 1004 | fnLocalVars.erase(x: varId); |
| 1005 | idMap.erase(x: varId); |
| 1006 | } |
| 1007 | |
| 1008 | // don't do for volatile references |
| 1009 | if (wordCount > 4 && (spv[start+4] & spv::MemoryAccessVolatileMask)) { |
| 1010 | fnLocalVars.erase(x: varId); |
| 1011 | idMap.erase(x: varId); |
| 1012 | } |
| 1013 | |
| 1014 | // Handle flow control |
| 1015 | if (blockMap.find(x: varId) == blockMap.end()) { |
| 1016 | blockMap[varId] = blockNum; // track block we found it in. |
| 1017 | } else if (blockMap[varId] != blockNum) { |
| 1018 | fnLocalVars.erase(x: varId); // Ignore if crosses flow control |
| 1019 | idMap.erase(x: varId); |
| 1020 | } |
| 1021 | |
| 1022 | return true; |
| 1023 | } |
| 1024 | |
| 1025 | if (opCode == spv::OpStore && fnLocalVars.count(x: asId(word: start+1)) > 0) { |
| 1026 | const spv::Id varId = asId(word: start+1); |
| 1027 | |
| 1028 | if (idMap.find(x: varId) == idMap.end()) { |
| 1029 | idMap[varId] = asId(word: start+2); |
| 1030 | } else { |
| 1031 | // Remove if it has more than one store to the same pointer |
| 1032 | fnLocalVars.erase(x: varId); |
| 1033 | idMap.erase(x: varId); |
| 1034 | } |
| 1035 | |
| 1036 | // don't do for volatile references |
| 1037 | if (wordCount > 3 && (spv[start+3] & spv::MemoryAccessVolatileMask)) { |
| 1038 | fnLocalVars.erase(x: asId(word: start+3)); |
| 1039 | idMap.erase(x: asId(word: start+3)); |
| 1040 | } |
| 1041 | |
| 1042 | // Handle flow control |
| 1043 | if (blockMap.find(x: varId) == blockMap.end()) { |
| 1044 | blockMap[varId] = blockNum; // track block we found it in. |
| 1045 | } else if (blockMap[varId] != blockNum) { |
| 1046 | fnLocalVars.erase(x: varId); // Ignore if crosses flow control |
| 1047 | idMap.erase(x: varId); |
| 1048 | } |
| 1049 | |
| 1050 | return true; |
| 1051 | } |
| 1052 | |
| 1053 | return false; |
| 1054 | }, |
| 1055 | |
| 1056 | // If local var id used anywhere else, don't eliminate |
| 1057 | idFn: [&](spv::Id& id) { |
| 1058 | if (fnLocalVars.count(x: id) > 0) { |
| 1059 | fnLocalVars.erase(x: id); |
| 1060 | idMap.erase(x: id); |
| 1061 | } |
| 1062 | } |
| 1063 | ); |
| 1064 | |
| 1065 | if (errorLatch) |
| 1066 | return; |
| 1067 | |
| 1068 | process( |
| 1069 | instFn: [&](spv::Op opCode, unsigned start) { |
| 1070 | if (opCode == spv::OpLoad && fnLocalVars.count(x: asId(word: start+3)) > 0) |
| 1071 | idMap[asId(word: start+2)] = idMap[asId(word: start+3)]; |
| 1072 | return false; |
| 1073 | }, |
| 1074 | idFn: op_fn_nop); |
| 1075 | |
| 1076 | if (errorLatch) |
| 1077 | return; |
| 1078 | |
| 1079 | // Chase replacements to their origins, in case there is a chain such as: |
| 1080 | // 2 = store 1 |
| 1081 | // 3 = load 2 |
| 1082 | // 4 = store 3 |
| 1083 | // 5 = load 4 |
| 1084 | // We want to replace uses of 5 with 1. |
| 1085 | for (const auto& idPair : idMap) { |
| 1086 | spv::Id id = idPair.first; |
| 1087 | while (idMap.find(x: id) != idMap.end()) // Chase to end of chain |
| 1088 | id = idMap[id]; |
| 1089 | |
| 1090 | idMap[idPair.first] = id; // replace with final result |
| 1091 | } |
| 1092 | |
| 1093 | // Remove the load/store/variables for the ones we've discovered |
| 1094 | process( |
| 1095 | instFn: [&](spv::Op opCode, unsigned start) { |
| 1096 | if ((opCode == spv::OpLoad && fnLocalVars.count(x: asId(word: start+3)) > 0) || |
| 1097 | (opCode == spv::OpStore && fnLocalVars.count(x: asId(word: start+1)) > 0) || |
| 1098 | (opCode == spv::OpVariable && fnLocalVars.count(x: asId(word: start+2)) > 0)) { |
| 1099 | |
| 1100 | stripInst(start); |
| 1101 | return true; |
| 1102 | } |
| 1103 | |
| 1104 | return false; |
| 1105 | }, |
| 1106 | |
| 1107 | idFn: [&](spv::Id& id) { |
| 1108 | if (idMap.find(x: id) != idMap.end()) id = idMap[id]; |
| 1109 | } |
| 1110 | ); |
| 1111 | |
| 1112 | if (errorLatch) |
| 1113 | return; |
| 1114 | |
| 1115 | strip(); // strip out data we decided to eliminate |
| 1116 | } |
| 1117 | |
| 1118 | // remove bodies of uncalled functions |
| 1119 | void spirvbin_t::dceFuncs() |
| 1120 | { |
| 1121 | msg(minVerbosity: 3, indent: 2, txt: std::string("Removing Dead Functions: " )); |
| 1122 | |
| 1123 | // TODO: There are more efficient ways to do this. |
| 1124 | bool changed = true; |
| 1125 | |
| 1126 | while (changed) { |
| 1127 | changed = false; |
| 1128 | |
| 1129 | for (auto fn = fnPos.begin(); fn != fnPos.end(); ) { |
| 1130 | if (fn->first == entryPoint) { // don't DCE away the entry point! |
| 1131 | ++fn; |
| 1132 | continue; |
| 1133 | } |
| 1134 | |
| 1135 | const auto call_it = fnCalls.find(x: fn->first); |
| 1136 | |
| 1137 | if (call_it == fnCalls.end() || call_it->second == 0) { |
| 1138 | changed = true; |
| 1139 | stripRange.push_back(x: fn->second); |
| 1140 | |
| 1141 | // decrease counts of called functions |
| 1142 | process( |
| 1143 | instFn: [&](spv::Op opCode, unsigned start) { |
| 1144 | if (opCode == spv::Op::OpFunctionCall) { |
| 1145 | const auto call_it = fnCalls.find(x: asId(word: start + 3)); |
| 1146 | if (call_it != fnCalls.end()) { |
| 1147 | if (--call_it->second <= 0) |
| 1148 | fnCalls.erase(position: call_it); |
| 1149 | } |
| 1150 | } |
| 1151 | |
| 1152 | return true; |
| 1153 | }, |
| 1154 | idFn: op_fn_nop, |
| 1155 | begin: fn->second.first, |
| 1156 | end: fn->second.second); |
| 1157 | |
| 1158 | if (errorLatch) |
| 1159 | return; |
| 1160 | |
| 1161 | fn = fnPos.erase(position: fn); |
| 1162 | } else ++fn; |
| 1163 | } |
| 1164 | } |
| 1165 | } |
| 1166 | |
| 1167 | // remove unused function variables + decorations |
| 1168 | void spirvbin_t::dceVars() |
| 1169 | { |
| 1170 | msg(minVerbosity: 3, indent: 2, txt: std::string("DCE Vars: " )); |
| 1171 | |
| 1172 | std::unordered_map<spv::Id, int> varUseCount; |
| 1173 | |
| 1174 | // Count function variable use |
| 1175 | process( |
| 1176 | instFn: [&](spv::Op opCode, unsigned start) { |
| 1177 | if (opCode == spv::OpVariable) { |
| 1178 | ++varUseCount[asId(word: start+2)]; |
| 1179 | return true; |
| 1180 | } else if (opCode == spv::OpEntryPoint) { |
| 1181 | const int wordCount = asWordCount(word: start); |
| 1182 | for (int i = 4; i < wordCount; i++) { |
| 1183 | ++varUseCount[asId(word: start+i)]; |
| 1184 | } |
| 1185 | return true; |
| 1186 | } else |
| 1187 | return false; |
| 1188 | }, |
| 1189 | |
| 1190 | idFn: [&](spv::Id& id) { if (varUseCount[id]) ++varUseCount[id]; } |
| 1191 | ); |
| 1192 | |
| 1193 | if (errorLatch) |
| 1194 | return; |
| 1195 | |
| 1196 | // Remove single-use function variables + associated decorations and names |
| 1197 | process( |
| 1198 | instFn: [&](spv::Op opCode, unsigned start) { |
| 1199 | spv::Id id = spv::NoResult; |
| 1200 | if (opCode == spv::OpVariable) |
| 1201 | id = asId(word: start+2); |
| 1202 | if (opCode == spv::OpDecorate || opCode == spv::OpName) |
| 1203 | id = asId(word: start+1); |
| 1204 | |
| 1205 | if (id != spv::NoResult && varUseCount[id] == 1) |
| 1206 | stripInst(start); |
| 1207 | |
| 1208 | return true; |
| 1209 | }, |
| 1210 | idFn: op_fn_nop); |
| 1211 | } |
| 1212 | |
| 1213 | // remove unused types |
| 1214 | void spirvbin_t::dceTypes() |
| 1215 | { |
| 1216 | std::vector<bool> isType(bound(), false); |
| 1217 | |
| 1218 | // for speed, make O(1) way to get to type query (map is log(n)) |
| 1219 | for (const auto typeStart : typeConstPos) |
| 1220 | isType[asTypeConstId(word: typeStart)] = true; |
| 1221 | |
| 1222 | std::unordered_map<spv::Id, int> typeUseCount; |
| 1223 | |
| 1224 | // This is not the most efficient algorithm, but this is an offline tool, and |
| 1225 | // it's easy to write this way. Can be improved opportunistically if needed. |
| 1226 | bool changed = true; |
| 1227 | while (changed) { |
| 1228 | changed = false; |
| 1229 | strip(); |
| 1230 | typeUseCount.clear(); |
| 1231 | |
| 1232 | // Count total type usage |
| 1233 | process(instFn: inst_fn_nop, |
| 1234 | idFn: [&](spv::Id& id) { if (isType[id]) ++typeUseCount[id]; } |
| 1235 | ); |
| 1236 | |
| 1237 | if (errorLatch) |
| 1238 | return; |
| 1239 | |
| 1240 | // Remove single reference types |
| 1241 | for (const auto typeStart : typeConstPos) { |
| 1242 | const spv::Id typeId = asTypeConstId(word: typeStart); |
| 1243 | if (typeUseCount[typeId] == 1) { |
| 1244 | changed = true; |
| 1245 | --typeUseCount[typeId]; |
| 1246 | stripInst(start: typeStart); |
| 1247 | } |
| 1248 | } |
| 1249 | |
| 1250 | if (errorLatch) |
| 1251 | return; |
| 1252 | } |
| 1253 | } |
| 1254 | |
| 1255 | #ifdef NOTDEF |
| 1256 | bool spirvbin_t::matchType(const spirvbin_t::globaltypes_t& globalTypes, spv::Id lt, spv::Id gt) const |
| 1257 | { |
| 1258 | // Find the local type id "lt" and global type id "gt" |
| 1259 | const auto lt_it = typeConstPosR.find(lt); |
| 1260 | if (lt_it == typeConstPosR.end()) |
| 1261 | return false; |
| 1262 | |
| 1263 | const auto typeStart = lt_it->second; |
| 1264 | |
| 1265 | // Search for entry in global table |
| 1266 | const auto gtype = globalTypes.find(gt); |
| 1267 | if (gtype == globalTypes.end()) |
| 1268 | return false; |
| 1269 | |
| 1270 | const auto& gdata = gtype->second; |
| 1271 | |
| 1272 | // local wordcount and opcode |
| 1273 | const int wordCount = asWordCount(typeStart); |
| 1274 | const spv::Op opCode = asOpCode(typeStart); |
| 1275 | |
| 1276 | // no type match if opcodes don't match, or operand count doesn't match |
| 1277 | if (opCode != opOpCode(gdata[0]) || wordCount != opWordCount(gdata[0])) |
| 1278 | return false; |
| 1279 | |
| 1280 | const unsigned numOperands = wordCount - 2; // all types have a result |
| 1281 | |
| 1282 | const auto cmpIdRange = [&](range_t range) { |
| 1283 | for (int x=range.first; x<std::min(range.second, wordCount); ++x) |
| 1284 | if (!matchType(globalTypes, asId(typeStart+x), gdata[x])) |
| 1285 | return false; |
| 1286 | return true; |
| 1287 | }; |
| 1288 | |
| 1289 | const auto cmpConst = [&]() { return cmpIdRange(constRange(opCode)); }; |
| 1290 | const auto cmpSubType = [&]() { return cmpIdRange(typeRange(opCode)); }; |
| 1291 | |
| 1292 | // Compare literals in range [start,end) |
| 1293 | const auto cmpLiteral = [&]() { |
| 1294 | const auto range = literalRange(opCode); |
| 1295 | return std::equal(spir.begin() + typeStart + range.first, |
| 1296 | spir.begin() + typeStart + std::min(range.second, wordCount), |
| 1297 | gdata.begin() + range.first); |
| 1298 | }; |
| 1299 | |
| 1300 | assert(isTypeOp(opCode) || isConstOp(opCode)); |
| 1301 | |
| 1302 | switch (opCode) { |
| 1303 | case spv::OpTypeOpaque: // TODO: disable until we compare the literal strings. |
| 1304 | case spv::OpTypeQueue: return false; |
| 1305 | case spv::OpTypeEvent: // fall through... |
| 1306 | case spv::OpTypeDeviceEvent: // ... |
| 1307 | case spv::OpTypeReserveId: return false; |
| 1308 | // for samplers, we don't handle the optional parameters yet |
| 1309 | case spv::OpTypeSampler: return cmpLiteral() && cmpConst() && cmpSubType() && wordCount == 8; |
| 1310 | default: return cmpLiteral() && cmpConst() && cmpSubType(); |
| 1311 | } |
| 1312 | } |
| 1313 | |
| 1314 | // Look for an equivalent type in the globalTypes map |
| 1315 | spv::Id spirvbin_t::findType(const spirvbin_t::globaltypes_t& globalTypes, spv::Id lt) const |
| 1316 | { |
| 1317 | // Try a recursive type match on each in turn, and return a match if we find one |
| 1318 | for (const auto& gt : globalTypes) |
| 1319 | if (matchType(globalTypes, lt, gt.first)) |
| 1320 | return gt.first; |
| 1321 | |
| 1322 | return spv::NoType; |
| 1323 | } |
| 1324 | #endif // NOTDEF |
| 1325 | |
| 1326 | // Return start position in SPV of given Id. error if not found. |
| 1327 | unsigned spirvbin_t::idPos(spv::Id id) const |
| 1328 | { |
| 1329 | const auto tid_it = idPosR.find(x: id); |
| 1330 | if (tid_it == idPosR.end()) { |
| 1331 | error(txt: "ID not found" ); |
| 1332 | return 0; |
| 1333 | } |
| 1334 | |
| 1335 | return tid_it->second; |
| 1336 | } |
| 1337 | |
| 1338 | // Hash types to canonical values. This can return ID collisions (it's a bit |
| 1339 | // inevitable): it's up to the caller to handle that gracefully. |
| 1340 | std::uint32_t spirvbin_t::hashType(unsigned typeStart) const |
| 1341 | { |
| 1342 | const unsigned wordCount = asWordCount(word: typeStart); |
| 1343 | const spv::Op opCode = asOpCode(word: typeStart); |
| 1344 | |
| 1345 | switch (opCode) { |
| 1346 | case spv::OpTypeVoid: return 0; |
| 1347 | case spv::OpTypeBool: return 1; |
| 1348 | case spv::OpTypeInt: return 3 + (spv[typeStart+3]); |
| 1349 | case spv::OpTypeFloat: return 5; |
| 1350 | case spv::OpTypeVector: |
| 1351 | return 6 + hashType(typeStart: idPos(id: spv[typeStart+2])) * (spv[typeStart+3] - 1); |
| 1352 | case spv::OpTypeMatrix: |
| 1353 | return 30 + hashType(typeStart: idPos(id: spv[typeStart+2])) * (spv[typeStart+3] - 1); |
| 1354 | case spv::OpTypeImage: |
| 1355 | return 120 + hashType(typeStart: idPos(id: spv[typeStart+2])) + |
| 1356 | spv[typeStart+3] + // dimensionality |
| 1357 | spv[typeStart+4] * 8 * 16 + // depth |
| 1358 | spv[typeStart+5] * 4 * 16 + // arrayed |
| 1359 | spv[typeStart+6] * 2 * 16 + // multisampled |
| 1360 | spv[typeStart+7] * 1 * 16; // format |
| 1361 | case spv::OpTypeSampler: |
| 1362 | return 500; |
| 1363 | case spv::OpTypeSampledImage: |
| 1364 | return 502; |
| 1365 | case spv::OpTypeArray: |
| 1366 | return 501 + hashType(typeStart: idPos(id: spv[typeStart+2])) * spv[typeStart+3]; |
| 1367 | case spv::OpTypeRuntimeArray: |
| 1368 | return 5000 + hashType(typeStart: idPos(id: spv[typeStart+2])); |
| 1369 | case spv::OpTypeStruct: |
| 1370 | { |
| 1371 | std::uint32_t hash = 10000; |
| 1372 | for (unsigned w=2; w < wordCount; ++w) |
| 1373 | hash += w * hashType(typeStart: idPos(id: spv[typeStart+w])); |
| 1374 | return hash; |
| 1375 | } |
| 1376 | |
| 1377 | case spv::OpTypeOpaque: return 6000 + spv[typeStart+2]; |
| 1378 | case spv::OpTypePointer: return 100000 + hashType(typeStart: idPos(id: spv[typeStart+3])); |
| 1379 | case spv::OpTypeFunction: |
| 1380 | { |
| 1381 | std::uint32_t hash = 200000; |
| 1382 | for (unsigned w=2; w < wordCount; ++w) |
| 1383 | hash += w * hashType(typeStart: idPos(id: spv[typeStart+w])); |
| 1384 | return hash; |
| 1385 | } |
| 1386 | |
| 1387 | case spv::OpTypeEvent: return 300000; |
| 1388 | case spv::OpTypeDeviceEvent: return 300001; |
| 1389 | case spv::OpTypeReserveId: return 300002; |
| 1390 | case spv::OpTypeQueue: return 300003; |
| 1391 | case spv::OpTypePipe: return 300004; |
| 1392 | case spv::OpConstantTrue: return 300007; |
| 1393 | case spv::OpConstantFalse: return 300008; |
| 1394 | case spv::OpConstantComposite: |
| 1395 | { |
| 1396 | std::uint32_t hash = 300011 + hashType(typeStart: idPos(id: spv[typeStart+1])); |
| 1397 | for (unsigned w=3; w < wordCount; ++w) |
| 1398 | hash += w * hashType(typeStart: idPos(id: spv[typeStart+w])); |
| 1399 | return hash; |
| 1400 | } |
| 1401 | case spv::OpConstant: |
| 1402 | { |
| 1403 | std::uint32_t hash = 400011 + hashType(typeStart: idPos(id: spv[typeStart+1])); |
| 1404 | for (unsigned w=3; w < wordCount; ++w) |
| 1405 | hash += w * spv[typeStart+w]; |
| 1406 | return hash; |
| 1407 | } |
| 1408 | case spv::OpConstantNull: |
| 1409 | { |
| 1410 | std::uint32_t hash = 500009 + hashType(typeStart: idPos(id: spv[typeStart+1])); |
| 1411 | return hash; |
| 1412 | } |
| 1413 | case spv::OpConstantSampler: |
| 1414 | { |
| 1415 | std::uint32_t hash = 600011 + hashType(typeStart: idPos(id: spv[typeStart+1])); |
| 1416 | for (unsigned w=3; w < wordCount; ++w) |
| 1417 | hash += w * spv[typeStart+w]; |
| 1418 | return hash; |
| 1419 | } |
| 1420 | |
| 1421 | default: |
| 1422 | error(txt: "unknown type opcode" ); |
| 1423 | return 0; |
| 1424 | } |
| 1425 | } |
| 1426 | |
| 1427 | void spirvbin_t::mapTypeConst() |
| 1428 | { |
| 1429 | globaltypes_t globalTypeMap; |
| 1430 | |
| 1431 | msg(minVerbosity: 3, indent: 2, txt: std::string("Remapping Consts & Types: " )); |
| 1432 | |
| 1433 | static const std::uint32_t softTypeIdLimit = 3011; // small prime. TODO: get from options |
| 1434 | static const std::uint32_t firstMappedID = 8; // offset into ID space |
| 1435 | |
| 1436 | for (auto& typeStart : typeConstPos) { |
| 1437 | const spv::Id resId = asTypeConstId(word: typeStart); |
| 1438 | const std::uint32_t hashval = hashType(typeStart); |
| 1439 | |
| 1440 | if (errorLatch) |
| 1441 | return; |
| 1442 | |
| 1443 | if (isOldIdUnmapped(oldId: resId)) { |
| 1444 | localId(id: resId, newId: nextUnusedId(id: hashval % softTypeIdLimit + firstMappedID)); |
| 1445 | if (errorLatch) |
| 1446 | return; |
| 1447 | } |
| 1448 | } |
| 1449 | } |
| 1450 | |
| 1451 | // Strip a single binary by removing ranges given in stripRange |
| 1452 | void spirvbin_t::strip() |
| 1453 | { |
| 1454 | if (stripRange.empty()) // nothing to do |
| 1455 | return; |
| 1456 | |
| 1457 | // Sort strip ranges in order of traversal |
| 1458 | std::sort(first: stripRange.begin(), last: stripRange.end()); |
| 1459 | |
| 1460 | // Allocate a new binary big enough to hold old binary |
| 1461 | // We'll step this iterator through the strip ranges as we go through the binary |
| 1462 | auto strip_it = stripRange.begin(); |
| 1463 | |
| 1464 | int strippedPos = 0; |
| 1465 | for (unsigned word = 0; word < unsigned(spv.size()); ++word) { |
| 1466 | while (strip_it != stripRange.end() && word >= strip_it->second) |
| 1467 | ++strip_it; |
| 1468 | |
| 1469 | if (strip_it == stripRange.end() || word < strip_it->first || word >= strip_it->second) |
| 1470 | spv[strippedPos++] = spv[word]; |
| 1471 | } |
| 1472 | |
| 1473 | spv.resize(new_size: strippedPos); |
| 1474 | stripRange.clear(); |
| 1475 | |
| 1476 | buildLocalMaps(); |
| 1477 | } |
| 1478 | |
| 1479 | // Strip a single binary by removing ranges given in stripRange |
| 1480 | void spirvbin_t::remap(std::uint32_t opts) |
| 1481 | { |
| 1482 | options = opts; |
| 1483 | |
| 1484 | // Set up opcode tables from SpvDoc |
| 1485 | spv::Parameterize(); |
| 1486 | |
| 1487 | validate(); // validate header |
| 1488 | buildLocalMaps(); // build ID maps |
| 1489 | |
| 1490 | msg(minVerbosity: 3, indent: 4, txt: std::string("ID bound: " ) + std::to_string(val: bound())); |
| 1491 | |
| 1492 | if (options & STRIP) stripDebug(); |
| 1493 | if (errorLatch) return; |
| 1494 | |
| 1495 | strip(); // strip out data we decided to eliminate |
| 1496 | if (errorLatch) return; |
| 1497 | |
| 1498 | if (options & OPT_LOADSTORE) optLoadStore(); |
| 1499 | if (errorLatch) return; |
| 1500 | |
| 1501 | if (options & OPT_FWD_LS) forwardLoadStores(); |
| 1502 | if (errorLatch) return; |
| 1503 | |
| 1504 | if (options & DCE_FUNCS) dceFuncs(); |
| 1505 | if (errorLatch) return; |
| 1506 | |
| 1507 | if (options & DCE_VARS) dceVars(); |
| 1508 | if (errorLatch) return; |
| 1509 | |
| 1510 | if (options & DCE_TYPES) dceTypes(); |
| 1511 | if (errorLatch) return; |
| 1512 | |
| 1513 | strip(); // strip out data we decided to eliminate |
| 1514 | if (errorLatch) return; |
| 1515 | |
| 1516 | stripDeadRefs(); // remove references to things we DCEed |
| 1517 | if (errorLatch) return; |
| 1518 | |
| 1519 | // after the last strip, we must clean any debug info referring to now-deleted data |
| 1520 | |
| 1521 | if (options & MAP_TYPES) mapTypeConst(); |
| 1522 | if (errorLatch) return; |
| 1523 | |
| 1524 | if (options & MAP_NAMES) mapNames(); |
| 1525 | if (errorLatch) return; |
| 1526 | |
| 1527 | if (options & MAP_FUNCS) mapFnBodies(); |
| 1528 | if (errorLatch) return; |
| 1529 | |
| 1530 | if (options & MAP_ALL) { |
| 1531 | mapRemainder(); // map any unmapped IDs |
| 1532 | if (errorLatch) return; |
| 1533 | |
| 1534 | applyMap(); // Now remap each shader to the new IDs we've come up with |
| 1535 | if (errorLatch) return; |
| 1536 | } |
| 1537 | } |
| 1538 | |
| 1539 | // remap from a memory image |
| 1540 | void spirvbin_t::remap(std::vector<std::uint32_t>& in_spv, const std::vector<std::string>& whiteListStrings, |
| 1541 | std::uint32_t opts) |
| 1542 | { |
| 1543 | stripWhiteList = whiteListStrings; |
| 1544 | spv.swap(x&: in_spv); |
| 1545 | remap(opts); |
| 1546 | spv.swap(x&: in_spv); |
| 1547 | } |
| 1548 | |
| 1549 | // remap from a memory image - legacy interface without white list |
| 1550 | void spirvbin_t::remap(std::vector<std::uint32_t>& in_spv, std::uint32_t opts) |
| 1551 | { |
| 1552 | stripWhiteList.clear(); |
| 1553 | spv.swap(x&: in_spv); |
| 1554 | remap(opts); |
| 1555 | spv.swap(x&: in_spv); |
| 1556 | } |
| 1557 | |
| 1558 | } // namespace SPV |
| 1559 | |
| 1560 | |